idea and projects

1. plan to use GVF Snake Model to segment colon surface from colon volume. this is the 2D/3D GVF matlab implementation  http://www.mathworks.com/matlabcentral/fileexchange/28149 2. use method in visual event recognition in videos by learning from web data to do colon cancer detection:  http://videolectures.net/cvpr2010_duan_verv/ 3. develop curvature field as new feature for recognition and tracking 4. to deal with rotation affect or detect different direction object, need to use big gradient points as candidate points which has pretty stable direction as reference sample direction.Thus we can detect different pose objects once. Different size objects can be handled by pyramid methods 5. KD Tree can be used to speed up neural network learning. Refer to Omohundro "efficient algorithm with neural network behavior" journal of complex system, 1987. Hence KD Tree can be used to speed up Kenji's MTANN learning process. 6. KD Tree is a powerful tool for analyzing data set structure, especially for large scale database. For example, we can easily find each point's neighborhood. As we know, machine learning algorithm is mainly sensitive to those samples which lie on the boundary of classifier. However, these samples can be easily found out by KD-Tree. Hence KD-Tree can be expected to be used to speed up machine learning process by choosing the samples close to boundary with higher priority than samples far from boundary. 6. Machine learning can be used to predict the reference orientation of fingerprint based on core points. First,find out the core points in fingerprint image. Then,construct samples which are the windowed regions centered at the core points.At last, learn the model to predict the orientation. 7. Neural network can be used to model fingerprint orientation model. Using neural network to learn the map from 2D coordinates to the value of sin or cos functions over the corresponding angles. We can also try other machine learning models. 8. I am planning to test compressive sampling on fingerprint orientation field.It consist of three steps: computing orientation field; using CS to reconstruct orientation field; check the effect. By doing this, I hope to reconstruct a more smoother orientation field especially for noisy fingerprint images. 9. Detecting core point method: If draw a circle with the center of core point, the summation of all directions within the circular area would be zero. otherwise versa. how about delta point? 10. 3D component labeling method for colon segmentation: large intestine is a connected component which is biggest in volume in    abdomen. Thus avoiding other interference from other organs like small intestine, liver, lung etc. 11. polyp detection method: on colon surface, choosing any voxel and draw a sphere around the voxel. Thus the sphere must be divided into two parts by colon surface. Compute the volume ratio of colon lumen part over colon outside part. The ratio is large for polyp but small for flat area. 12. Choose one voxel on colon surface as a candidate point. Compute the inner product of its normal vector and its neighboring voxel's normal vector. Its neighboring voxels must be within a sphere with prescribed radius. For true polyp, the polyp must be closed by relatively low value of inner product. For false polyp, the point must not be closed by relatively low value of the inner product.It means there are some paths from the center point to the border along which the inner product is relatively big. So we call the polyp detection method "inner product method". 13. local diffusion based fingerprint matching method: 1. Using curvature feature descriptors to find out candidate corresponding pairs; 2. For each pair of candidate points, using them as diffusion centers. Within their respective neighborhood with radius R, searching matching point pairs by comparing their distance and orientation relative to the centers. 3. New added matching point pairs are used as new searching centers. Continue step 2 until no new added matching points. 4. Select the local diffusion map with the largest number of points as correct matching map. Done! 14.singular point detection based on Poincare Index image: 1. given the radius R, consider the circle as the integration path and computer Poincare Index for each pixel. Thus we get PI image. On PI image, singular points correspond to the center of circular area. Good point. 15. polyp detection method: choose any voxel with appropriate bright value(big enough to differentiate from lumen voxel) as an center of an sphere. Draw a sphere around the selected voxel. Compute the intensity histogram and gradient orientation histogram of the voxels inside the sphere as features. Then use the features to detect polyps. 16. Enlarging polyp method: given colon mesh surface, choose any voxel on the surface and find its neighbors. Try to displace all these voxels along the normal directions. If the distance from the center voxel to its neighbors is becoming bigger than that before displacement, the center voxel is displaced with certain distance along the normal direction. Thus, the potential polyp may become bigger and bigger. 17. polyp detection method: Choose any voxel on boundary of binary segmented colon volume. Compute its tangent plane. Consider the chosen voxel as center,draw a circle on the tangent plane. Compute the least distance from the voxels on the periphery of the circle to the voxels on surface as the feature for characterizing the chosen voxel. It can be assumed that for polyps, the least distance is the biggest, otherwise, smaller. We can use machine learning method to learn the suitable threshold to classify the two kinds of voxels. Good!http://childrensnational.org/careers-and-training/employee-resource-center 18. Fingerprint matching based on new information on minutiae: After fingerprint template alignment, check each pair of corresponding minutiae to see if their tails are overlapped. If they are not overlapped, it is a false match. Thus, for true match, it will not affected by this method. For false match, this method may find it and reduce false match. But what is the tail of one minutia? For ending point, its tail is its trunk. For bifurcation point, its tail is its two branches. We can sample a point on its tail when performing minutiae extraction process. Then each minutia has a tailing point accompanying them. It's convenient to check whether tails of matched points are overlapped. All we have to do is check the distance from one tailing point to another tailing point is less than a predefined threshold. Done! 19 AM-FM model can be used to enhance fingerprint image and detection singular points. 20. new fingerprint orientation estimation method: 1. For each pixel, compute the fourier transform of its rectangular neighborhood, i.e. 9*9 or 16*16; 2. fit the spectral data with mixing gauss model with two peaks. 3. the parameters of the found model convey the orientation of the central pixel. Thus it is hoped to get better orientation estimation. 21. fingerprint SP detection method: in high curvatured area, for each pixel, compute the fourier transformation of its rectangular neighborhood, consider the FT data as feature vector, use svm to classify them as SP or Non-SP 22. Define a new fingerprint reference point as the focus point of parabola obtained by interpolating the three points on the same ridge. Evaluate the robustness of the reference points. Using different reference points, a reference direction may be attained by connect them or fitting them linearly. Steps involved:1. choose three points in ridge; 2. interpolate them using parabola; 3. find out the parabolic focus point; 4. find out multiple of reference points from different ridges; 4. fitting different reference points linearly constituting reference line. I guess the reference line is robust to nonlinear transformation accompanying fingerprint images. Tell student to do it and submit to PR or PRL. 23. Inpainting algorithm can be used to predict the corrupted orientation field: 1. estimate good quality orientation blocks. 2. using inpainting algorithm to fill the bad quality blocks. 24. Learning Reconstruction Dictionary:  min_{a,D}\sum_{l=1}^{M}||x_{l}-Da_{l}||^2  s.t. ||a_l||_0<=L. K-SVD algorithm for this optimization problem. We propose to learn the most sparsest Dictionary:  min_{a,D}{max{||a_i||_{l_0}, i=1,2,...,N}}  s.t. ||X-Da||_F^2<=episilon. Steps: 1. initialize  D by selecting patch images randomly from training set. 2. find the sparsest  a using OMP algorithm. 3. find the most densest vector  a_i and corresponding  x_i, denote the bigest element value  a_i^j in vector  a_i, corresponding  D_j; 4. update  D_j=D_j+t(x_i-D_j), where parameter  t in  [0, 1]. Compute new a using new D. If the  ||a||_l_0 decreases, update is valid, otherwise cancel the update and find less denser vector  a_i, repeat step 3, 4. If no valid update, stop. otherwise repeat step 1-4 again. These two learning methods get different results. This can be shown in 2 dimension data.\ 25. Surprisingly, it seems possible to prove the sensing matrix \phy is Gaussian random distribution under orthogonal sparse representation matrix \psy. In order to minimize mutual  coherence of equivalent matrix D, D must be uniformly distributed in unit hypersphere surface. Thus D must be Gaussian random distributed according to the conclusions in http:  //mathworld.wolfram.com/HyperspherePointPicking.html. If we can prove that the product of Gaussian matrix D and the inverse of orthogonal matrix \psy is also Gaussian distributed. Then the sensing matrix \phy =D*inverse of \psy is Gaussian distributed.  It's amazing! 26.  To construct equivalent dictionary with K columns, we can construct 2K equally spaced points on unit hypersphere surface using the method in  http://www.rhinocerus.net/forum/lang-idl-pvwave/117743-equally-spaced-points-hypersphere.html. Then delete K points which has the largest inner products. The left  K points constitute th equivalent dictionary. It's cool. 27. How to construct M*N sensing matrix? we have a idea (force balance method) on that: 1)randomly generate N vectors in M dimenstional space. 2)Normalize to unit norm vectors. 3)These points distributed on the surface of unit ball. 4) Assume that each point exerts a force to other points, this force is proportional to the distance between any two points. The direction of force is along the line connecting these two points. 5) add all forces on each point to get the final force on the point. Move all points with the force on the ball surface. Repeat step 4) and 5) until no point moves. This is the method to construct optimal sensing matrix. 28. An easy way to construct sensing matrix with optimal coherence value using gaussian distribution. 1) for M*N sensing matrix, generate 2*N( probably more) unit random vectors in  M dimensional space. 2) exclude the vector which has the biggest correlation value. 29. Super-resolution: 1) use cubic interpolattion to magnify images to the desired degree, get magnified image y; 2)model high resolution image x with normal distribution; 3)model observation process as y=h*x+n, where h is a parameter matrix and n is noise. h can be learned using some samples. 4) using beyesian inference method to estimate x for given y, thus we get high resolution image. 30. measuring matrix \phi design: 1)for given k sparse signal in R^N, contruct final sensing matrix \phipsi by choosing N points in R^k uniformly distributed on unit sphere surface as the column vectors of final sensing matrix \phipsi. 2) choose orthonormal basis \psi under which signal has sparse representation. 3)Get measuring matrix \phi by multiplying \phipsi with the inverse of \psi. Thus we complete the measuring matrix design process with given orthonormal basis. It can be extended to the case where dictionary is learned. Excellent. 31. we can use existing sensing matrix learning method to learn a binary sensing matrix: 1)learn a nonbinary sensing matrix;2)find a binary sensing matrix closest to the learned matrix by finding a binary vector which is closest to the line determined by the nonbinary colomn vector of learned sensing matrix. 32. Marting suggest that using blurred version of image to do CS reconstruction works better than using the original image without blur.  Mathematically, y=\phi*S*x, where s is a smoothing filter. 33. On our PMMWI system, consider a subpixel movement when moving masks to obtain observations. Thus, we can get a different modulation method from the Hadamard matrix. We may get better results 34. SP Detection: Using convergence index filter for SP detection. refer to the paper "Convergence index filter for vector fields", 1999, IEEE TIP. 35. We can use error bar from Bayesian Compressive Sensing method to adaptively select the linear projection (i.e., the rows of the matrix \phi) to reduce uncertainty in the signal. 36. For compressive sensing model: y=\phi*x+n; we can estimate x with \phi known. we can try to estimate \phi and x simultaneously and compare the results from these two methods to see which x is better. Interesting. 37. For compressive sensing model: y=\phi*x+n; we can estimate x with \phi known. we can try to estimate h and x simultaneously for this new model y=\phi*h*x+n with \phi known and compare the results from these two methods to see which x is better and what the h is like. Interesting. 38. Sparse or smoothing assumption can be used to enhance image such as fluorescent microscopy cell images. This is Bayesian framework to enhance image.  We can refer to the reference "Variational Bayesian Blind Deconvolution Using a Total Variation Prior". This is new? We have test it on cell image and get nice results. 39. We can do compressive sensing and debluring simultaneously. the formula is y=\phi*h*x+n. To estimate h and x with y known, we will formulate this problem as a compressive sensing reconstruction problem alternatively( first give initial h, compute x; then with x, compute h. and so on) reference paper "blind motion debluring from a single image using sparse approximation. We will use the Bayesian framework since we have the codes of its implement. 40. Using graph cut to do  compressive sensing reconstruction. It is a graph cut with multilabel. The penalty for designating label is the error between the observation and the observation obtained from the label just assigned. Interaction potential is small for label zero, otherwise big. refer to the paper "graph cut with label cost". 41. Using haar wavelet as basis to do passive milimeter wave compressive imaging based on Bayesian reconstruction and Hardamard Matrix. This is for APL. 42. Using the technique of locating the important wavelet coefficients to reduce the number of measurements in CS. Or explore the structure pattern in wavelet domain. refer to papers from purdue University. For APL. 43. Since we have the set of  possible Hadamard sensing vectors \phi and some orthornormal basis \psi (Fourier basis, wavelet basis). We can choose the appropriate sensing vectors from the set \phi while making the incoherence between sensing vector and basis minimum. Or we can choose the sensing vector V  which maximizes the product V*\psi*S. For APL 44. Graph cut for fingerprint orientation field estimation. or reference to the other paper "Gauss-Markov measure field models" for Gauss Markov Measure Field Model.

Gauss-Markov Measure Field Models
for Low-Level Vision

45. Reference to the PAMI paper "Markov Random Field Model for Directional Field and Singularity Extraction in Fingerprint Images". Using graph cut to do orientation field estimation and using complex rational polynomial for standard singularity template to do singular point estimation. For PAMI. 46. We can do compressive sensing and super resolution simultaneously. Probably we should give the initial image estimation using first order derivative high pass filter matrix to regularize the psuedoinverse of the sensing matrix. reference to the paper "Sparse bayesian image restoration". TIP 47. we can do blind deconvolution Using Gaussian distribution prior on the output of high pass filter called first order derivative.  reference the papers " Sparse bayesian image restoration" and "Variational Bayesian Blind Deconvolution Using a Total Variation Prior". For TIP. 48. Fingerprint mathching, reference "A new algorithm for non rigid point matching" 49. Iterative Reweighted Least Square blind deconvolution used to do compressive sensing. using alternating optimization technique. refer to "Iterative reweighted least squares" 50. image segmentation model: min |g-f|^2+lamda*(|df/dx|^0+|df/dy|^0). refer to "Image smoothing via L0 Gradient Minimization" 51. using MMV model to do image compressive sensing: Y=\phi*X+V. X-input image, considered as multiple column signal using zhang zhilin MMV matlab model to recover the image. Thus avoid the super big size matrix (\phi) problem  for large size image. 52. compressive sensing and deconvolution based on L0 gradient minimization. Refer to "Image Smoothing via L0 gradient Minimization" 53. compressive sensing using objective function: \sum log(|x_i|^2). It approach L0 norm. we can solve the optimization problem using Iterative Reweighted Least Square. 54. Using geometric mean function as compressive sensing objective function:[ (x1+e)*(x2+e)*(x3+e)*...*(x_n+e)]^(1/n). subject to y=\phi*x;  That object function is concave. 55.  Using prior p(x)=e^{-a\sum_{i,k} p(g_{i,k}*x)} to do deconvolution or compressive sensing (where p(x) may be l1 norm) based on Bayessian framework using maximizing a posteriori probability. reference to "Sparse bayesian image restoration" and " deconvolution using natural imge priors". 56. graph based cell segmentation based on ellipse prior. reference to "Efficient Graph-based image segmentation". If data match ellipse model, the value of K in the mentioned paper is set to small value, vice versal. ellipse parameter can be computed using matlab command "regionprops", and reference the webpage  http://blogs.mathworks.com/steve/2010/07/30/visualizing-regionprops-ellipse-measurements/, and  http://www.mathworks.com/help/toolbox/images/ref/regionprops.html 57. Do compressive sensing or blind deconvolution with new regularizer L1/L2+L2/L1; This regularizer is better than L1/L2 and L2/L1. It will exclude the blurry solution and noisy solution. We can give the plot of the sparsity value for one image with different blur and noise, like paper "Blind Deconvolution Using a Normalized Sparsity Measure" (source code available). We can use the optimization method in that paper. TIP. 58. As the discussion with Xin, do compressive sensing or blind deconvolution with new regularizer L2+1/L2, L1+1/L1;we can prove that Lp+1/Lp is much better than Lp. This regularizer is also much better than L1/L2 and L2/L1. It will exclude the blurry solution and noisy solution. We can give the plot of the sparsity value for one image with different blur and noise, like paper "Blind Deconvolution Using a Normalized Sparsity Measure" (source code available). We can use the optimization method in that paper. TIP. 59. Connected component labeling based on the propagation of border pixels for any dimensional data and memory size. TIP. 60. Using Expectation Propergation  to do blind deconvolution according to Martin's variational inference method. 61. Use sampling method to evaluate the posterior probability p(A_k,Beta|y) in Derin's paper "Compressive passive millimeter wave imaging" rather than to use type ii maximum likelyhood method. The result would be expected to be better. 62. Subset selection algorithm for compressive sensing reconstruction, refer to Matrix Computation. 63. us spectral correspondence for point pattern matching to do fingerprint verification. refer to paper "spectral correspondence for point pattern matching". 64. Do point pattern matching by checking the corresponding adjacency matrices. compare each entry with each other and check the consistency between them. Great Idea.https://www.google.com/?gws_rd=ssl 65. Using martin's method of Bayesian based CS reconstruction  to do image enhancement ,denoise image, compute fingerprint orientation field, enhance nucleus image and segment nucleus image. 66. Using machine learning to estimate the principle of fingerprint image, and rotate query fingerprint image with respect template image, then Using spectral based point pattern matching method to do fingerprint minutiae matching. 67. For point pattern matching, choose any pair of candidate corresponding points, compute the transformation parameter value(DeltaX, DeltaY,Delta_Theta), Using meanshift method to find the densest cluster center which is the optimal transformation,Then transform the point set using the current found transformation, Then Using spectral point pattern matching method to match the transformed point set. 68. A new segmentation method based on specializing unary term in graph cut: 1 compute the gray (color) density for each objects; 2 assign unary term for each pixel in graph cut according likelyhood function value. We  expect get better segmentation results. 69. A new nuclei isolation method based on the combination of high gradient magnitude (genuine edges) with ellipsoid fitting: inspired from the tracing process in Canny edge detection method, the graph cut segmentation results mark the genuine ridges. Then we trace the ridges with hysteresis thresholding or using ellipsoid fitting to isolate individual nuclei. We can also use convex decomposition. 70. Nuclei segmentation method: 1. Get foreground nuclei by graph cut using appropriate parameter;2. Based on the segmented foreground, isolate nuclei using combination of graph cut with parameter focusing on fine ridges with ridge tracing skill which ensure the new added ridges connecting to 1. plan to use GVF Snake Model to segment colon surface from colon volume. this is the 2D/3D GVF matlab implementation  http://www.mathworks.com/matlabcentral/fileexchange/28149 2. use method in visual event recognition in videos by learning from web data to do colon cancer detection:  http://videolectures.net/cvpr2010_duan_verv/ 3. develop curvature field as new feature for recognition and tracking 4. to deal with rotation affect or detect different direction object, need to use big gradient points as candidate points which has pretty stable direction as reference sample direction.Thus we can detect different pose objects once. Different size objects can be handled by pyramid methods 5. KD Tree can be used to speed up neural network learning. Refer to Omohundro "efficient algorithm with neural network behavior" journal of complex system, 1987. Hence KD Tree can be used to speed up Kenji's MTANN learning process. 6. KD Tree is a powerful tool for analyzing data set structure, especially for large scale database. For example, we can easily find each point's neighborhood. As we know, machine learning algorithm is mainly sensitive to those samples which lie on the boundary of classifier. However, these samples can be easily found out by KD-Tree. Hence KD-Tree can be expected to be used to speed up machine learning process by choosing the samples close to boundary with higher priority than samples far from boundary. 6. Machine learning can be used to predict the reference orientation of fingerprint based on core points. First,find out the core points in fingerprint image. Then,construct samples which are the windowed regions centered at the core points.At last, learn the model to predict the orientation. 7. Neural network can be used to model fingerprint orientation model. Using neural network to learn the map from 2D coordinates to the value of sin or cos functions over the corresponding angles. We can also try other machine learning models. 8. I am planning to test compressive sampling on fingerprint orientation field.It consist of three steps: computing orientation field; using CS to reconstruct orientation field; check the effect. By doing this, I hope to reconstruct a more smoother orientation field especially for noisy fingerprint images. 9. Detecting core point method: If draw a circle with the center of core point, the summation of all directions within the circular area would be zero. otherwise versa. how about delta point? 10. 3D component labeling method for colon segmentation: large intestine is a connected component which is biggest in volume in    abdomen. Thus avoiding other interference from other organs like small intestine, liver, lung etc. 11. polyp detection method: on colon surface, choosing any voxel and draw a sphere around the voxel. Thus the sphere must be divided into two parts by colon surface. Compute the volume ratio of colon lumen part over colon outside part. The ratio is large for polyp but small for flat area. 12. Choose one voxel on colon surface as a candidate point. Compute the inner product of its normal vector and its neighboring voxel's normal vector. Its neighboring voxels must be within a sphere with prescribed radius. For true polyp, the polyp must be closed by relatively low value of inner product. For false polyp, the point must not be closed by relatively low value of the inner product.It means there are some paths from the center point to the border along which the inner product is relatively big. So we call the polyp detection method "inner product method". 13. local diffusion based fingerprint matching method: 1. Using curvature feature descriptors to find out candidate corresponding pairs; 2. For each pair of candidate points, using them as diffusion centers. Within their respective neighborhood with radius R, searching matching point pairs by comparing their distance and orientation relative to the centers. 3. New added matching point pairs are used as new searching centers. Continue step 2 until no new added matching points. 4. Select the local diffusion map with the largest number of points as correct matching map. Done! 14.singular point detection based on Poincare Index image: 1. given the radius R, consider the circle as the integration path and computer Poincare Index for each pixel. Thus we get PI image. On PI image, singular points correspond to the center of circular area. Good point. 15. polyp detection method: choose any voxel with appropriate bright value(big enough to differentiate from lumen voxel) as an center of an sphere. Draw a sphere around the selected voxel. Compute the intensity histogram and gradient orientation histogram of the voxels inside the sphere as features. Then use the features to detect polyps. 16. Enlarging polyp method: given colon mesh surface, choose any voxel on the surface and find its neighbors. Try to displace all these voxels along the normal directions. If the distance from the center voxel to its neighbors is becoming bigger than that before displacement, the center voxel is displaced with certain distance along the normal direction. Thus, the potential polyp may become bigger and bigger. 17. polyp detection method: Choose any voxel on boundary of binary segmented colon volume. Compute its tangent plane. Consider the chosen voxel as center,draw a circle on the tangent plane. Compute the least distance from the voxels on the periphery of the circle to the voxels on surface as the feature for characterizing the chosen voxel. It can be assumed that for polyps, the least distance is the biggest, otherwise, smaller. We can use machine learning method to learn the suitable threshold to classify the two kinds of voxels. Good! 18. Fingerprint matching based on new information on minutiae: After fingerprint template alignment, check each pair of corresponding minutiae to see if their tails are overlapped. If they are not overlapped, it is a false match. Thus, for true match, it will not affected by this method. For false match, this method may find it and reduce false match. But what is the tail of one minutia? For ending point, its tail is its trunk. For bifurcation point, its tail is its two branches. We can sample a point on its tail when performing minutiae extraction process. Then each minutia has a tailing point accompanying them. It's convenient to check whether tails of matched points are overlapped. All we have to do is check the distance from one tailing point to another tailing point is less than a predefined threshold. Done! 19 AM-FM model can be used to enhance fingerprint image and detection singular points. 20. new fingerprint orientation estimation method: 1. For each pixel, compute the fourier transform of its rectangular neighborhood, i.e. 9*9 or 16*16; 2. fit the spectral data with mixing gauss model with two peaks. 3. the parameters of the found model convey the orientation of the central pixel. Thus it is hoped to get better orientation estimation. 21. fingerprint SP detection method: in high curvatured area, for each pixel, compute the fourier transformation of its rectangular neighborhood, consider the FT data as feature vector, use svm to classify them as SP or Non-SP 22. Define a new fingerprint reference point as the focus point of parabola obtained by interpolating the three points on the same ridge. Evaluate the robustness of the reference points. Using different reference points, a reference direction may be attained by connect them or fitting them linearly. Steps involved:1. choose three points in ridge; 2. interpolate them using parabola; 3. find out the parabolic focus point; 4. find out multiple of reference points from different ridges; 4. fitting different reference points linearly constituting reference line. I guess the reference line is robust to nonlinear transformation accompanying fingerprint images. Tell student to do it and submit to PR or PRL. 23. Inpainting algorithm can be used to predict the corrupted orientation field: 1. estimate good quality orientation blocks. 2. using inpainting algorithm to fill the bad quality blocks. 24. Learning Reconstruction Dictionary:  min_{a,D}\sum_{l=1}^{M}||x_{l}-Da_{l}||^2  s.t. ||a_l||_0<=L. K-SVD algorithm for this optimization problem. We propose to learn the most sparsest Dictionary:  min_{a,D}{max{||a_i||_{l_0}, i=1,2,...,N}}  s.t. ||X-Da||_F^2<=episilon. Steps: 1. initialize  D by selecting patch images randomly from training set. 2. find the sparsest  a using OMP algorithm. 3. find the most densest vector  a_i and corresponding  x_i, denote the bigest element value  a_i^j in vector  a_i, corresponding  D_j; 4. update  D_j=D_j+t(x_i-D_j), where parameter  t in  [0, 1]. Compute new a using new D. If the  ||a||_l_0 decreases, update is valid, otherwise cancel the update and find less denser vector  a_i, repeat step 3, 4. If no valid update, stop. otherwise repeat step 1-4 again. These two learning methods get different results. This can be shown in 2 dimension data.\ 25. Surprisingly, it seems possible to prove the sensing matrix \phy is Gaussian random distribution under orthogonal sparse representation matrix \psy. In order to minimize mutual  coherence of equivalent matrix D, D must be uniformly distributed in unit hypersphere surface. Thus D must be Gaussian random distributed according to the conclusions in http:  //mathworld.wolfram.com/HyperspherePointPicking.html. If we can prove that the product of Gaussian matrix D and the inverse of orthogonal matrix \psy is also Gaussian distributed. Then the sensing matrix \phy =D*inverse of \psy is Gaussian distributed.  It's amazing! 26.  To construct equivalent dictionary with K columns, we can construct 2K equally spaced points on unit hypersphere surface using the method in  http://www.rhinocerus.net/forum/lang-idl-pvwave/117743-equally-spaced-points-hypersphere.html. Then delete K points which has the largest inner products. The left  K points constitute th equivalent dictionary. It's cool. 27. How to construct M*N sensing matrix? we have a idea (force balance method) on that: 1)randomly generate N vectors in M dimenstional space. 2)Normalize to unit norm vectors. 3)These points distributed on the surface of unit ball. 4) Assume that each point exerts a force to other points, this force is proportional to the distance between any two points. The direction of force is along the line connecting these two points. 5) add all forces on each point to get the final force on the point. Move all points with the force on the ball surface. Repeat step 4) and 5) until no point moves. This is the method to construct optimal sensing matrix. 28. An easy way to construct sensing matrix with optimal coherence value using gaussian distribution. 1) for M*N sensing matrix, generate 2*N( probably more) unit random vectors in  M dimensional space. 2) exclude the vector which has the biggest correlation value. 29. Super-resolution: 1) use cubic interpolattion to magnify images to the desired degree, get magnified image y; 2)model high resolution image x with normal distribution; 3)model observation process as y=h*x+n, where h is a parameter matrix and n is noise. h can be learned using some samples. 4) using beyesian inference method to estimate x for given y, thus we get high resolution image. 30. measuring matrix \phi design: 1)for given k sparse signal in R^N, contruct final sensing matrix \phipsi by choosing N points in R^k uniformly distributed on unit sphere surface as the column vectors of final sensing matrix \phipsi. 2) choose orthonormal basis \psi under which signal has sparse representation. 3)Get measuring matrix \phi by multiplying \phipsi with the inverse of \psi. Thus we complete the measuring matrix design process with given orthonormal basis. It can be extended to the case where dictionary is learned. Excellent. 31. we can use existing sensing matrix learning method to learn a binary sensing matrix: 1)learn a nonbinary sensing matrix;2)find a binary sensing matrix closest to the learned matrix by finding a binary vector which is closest to the line determined by the nonbinary colomn vector of learned sensing matrix. 32. Marting suggest that using blurred version of image to do CS reconstruction works better than using the original image without blur.  Mathematically, y=\phi*S*x, where s is a smoothing filter. 33. On our PMMWI system, consider a subpixel movement when moving masks to obtain observations. Thus, we can get a different modulation method from the Hadamard matrix. We may get better results 34. SP Detection: Using convergence index filter for SP detection. refer to the paper "Convergence index filter for vector fields", 1999, IEEE TIP. 35. We can use error bar from Bayesian Compressive Sensing method to adaptively select the linear projection (i.e., the rows of the matrix \phi) to reduce uncertainty in the signal. 36. For compressive sensing model: y=\phi*x+n; we can estimate x with \phi known. we can try to estimate \phi and x simultaneously and compare the results from these two methods to see which x is better. Interesting. 37. For compressive sensing model: y=\phi*x+n; we can estimate x with \phi known. we can try to estimate h and x simultaneously for this new model y=\phi*h*x+n with \phi known and compare the results from these two methods to see which x is better and what the h is like. Interesting. 38. Sparse or smoothing assumption can be used to enhance image such as fluorescent microscopy cell images. This is Bayesian framework to enhance image.  We can refer to the reference "Variational Bayesian Blind Deconvolution Using a Total Variation Prior". This is new? We have test it on cell image and get nice results. 39. We can do compressive sensing and debluring simultaneously. the formula is y=\phi*h*x+n. To estimate h and x with y known, we will formulate this problem as a compressive sensing reconstruction problem alternatively( first give initial h, compute x; then with x, compute h. and so on) reference paper "blind motion debluring from a single image using sparse approximation. We will use the Bayesian framework since we have the codes of its implement. 40. Using graph cut to do  compressive sensing reconstruction. It is a graph cut with multilabel. The penalty for designating label is the error between the observation and the observation obtained from the label just assigned. Interaction potential is small for label zero, otherwise big. refer to the paper "graph cut with label cost". 41. Using haar wavelet as basis to do passive milimeter wave compressive imaging based on Bayesian reconstruction and Hardamard Matrix. This is for APL. 42. Using the technique of locating the important wavelet coefficients to reduce the number of measurements in CS. Or explore the structure pattern in wavelet domain. refer to papers from purdue University. For APL. 43. Since we have the set of  possible Hadamard sensing vectors \phi and some orthornormal basis \psi (Fourier basis, wavelet basis). We can choose the appropriate sensing vectors from the set \phi while making the incoherence between sensing vector and basis minimum. Or we can choose the sensing vector V  which maximizes the product V*\psi*S. For APL 44. Graph cut for fingerprint orientation field estimation. or reference to the other paper "Gauss-Markov measure field models" for Gauss Markov Measure Field Model.

Gauss-Markov Measure Field Models
for Low-Level Vision

45. Reference to the PAMI paper "Markov Random Field Model for Directional Field and Singularity Extraction in Fingerprint Images". Using graph cut to do orientation field estimation and using complex rational polynomial for standard singularity template to do singular point estimation. For PAMI. 46. We can do compressive sensing and super resolution simultaneously. Probably we should give the initial image estimation using first order derivative high pass filter matrix to regularize the psuedoinverse of the sensing matrix. reference to the paper "Sparse bayesian image restoration". TIP 47. we can do blind deconvolution Using Gaussian distribution prior on the output of high pass filter called first order derivative.  reference the papers " Sparse bayesian image restoration" and "Variational Bayesian Blind Deconvolution Using a Total Variation Prior". For TIP. 48. Fingerprint mathching, reference "A new algorithm for non rigid point matching" 49. Iterative Reweighted Least Square blind deconvolution used to do compressive sensing. using alternating optimization technique. refer to "Iterative reweighted least squares" 50. image segmentation model: min |g-f|^2+lamda*(|df/dx|^0+|df/dy|^0). refer to "Image smoothing via L0 Gradient Minimization" 51. using MMV model to do image compressive sensing: Y=\phi*X+V. X-input image, considered as multiple column signal using zhang zhilin MMV matlab model to recover the image. Thus avoid the super big size matrix (\phi) problem  for large size image. 52. compressive sensing and deconvolution based on L0 gradient minimization. Refer to "Image Smoothing via L0 gradient Minimization" 53. compressive sensing using objective function: \sum log(|x_i|^2). It approach L0 norm. we can solve the optimization problem using Iterative Reweighted Least Square. 54. Using geometric mean function as compressive sensing objective function:[ (x1+e)*(x2+e)*(x3+e)*...*(x_n+e)]^(1/n). subject to y=\phi*x;  That object function is concave. 55.  Using prior p(x)=e^{-a\sum_{i,k} p(g_{i,k}*x)} to do deconvolution or compressive sensing (where p(x) may be l1 norm) based on Bayessian framework using maximizing a posteriori probability. reference to "Sparse bayesian image restoration" and " deconvolution using natural imge priors". 56. graph based cell segmentation based on ellipse prior. reference to "Efficient Graph-based image segmentation". If data match ellipse model, the value of K in the mentioned paper is set to small value, vice versal. ellipse parameter can be computed using matlab command "regionprops", and reference the webpage  http://blogs.mathworks.com/steve/2010/07/30/visualizing-regionprops-ellipse-measurements/, and  http://www.mathworks.com/help/toolbox/images/ref/regionprops.html 57. Do compressive sensing or blind deconvolution with new regularizer L1/L2+L2/L1; This regularizer is better than L1/L2 and L2/L1. It will exclude the blurry solution and noisy solution. We can give the plot of the sparsity value for one image with different blur and noise, like paper "Blind Deconvolution Using a Normalized Sparsity Measure" (source code available). We can use the optimization method in that paper. TIP. 58. As the discussion with Xin, do compressive sensing or blind deconvolution with new regularizer L2+1/L2, L1+1/L1;we can prove that Lp+1/Lp is much better than Lp. This regularizer is also much better than L1/L2 and L2/L1. It will exclude the blurry solution and noisy solution. We can give the plot of the sparsity value for one image with different blur and noise, like paper "Blind Deconvolution Using a Normalized Sparsity Measure" (source code available). We can use the optimization method in that paper. TIP. 59. Connected component labeling based on the propagation of border pixels for any dimensional data and memory size. TIP. 60. Using Expectation Propergation  to do blind deconvolution according to Martin's variational inference method. 61. Use sampling method to evaluate the posterior probability p(A_k,Beta|y) in Derin's paper "Compressive passive millimeter wave imaging" rather than to use type ii maximum likelyhood method. The result would be expected to be better. 62. Subset selection algorithm for compressive sensing reconstruction, refer to Matrix Computation. 63. us spectral correspondence for point pattern matching to do fingerprint verification. refer to paper "spectral correspondence for point pattern matching". 64. Do point pattern matching by checking the corresponding adjacency matrices. compare each entry with each other and check the consistency between them. Great Idea.https://www.google.com/?gws_rd=ssl 65. Using martin's method of Bayesian based CS reconstruction  to do image enhancement ,denoise image, compute fingerprint orientation field, enhance nucleus image and segment nucleus image. 66. Using machine learning to estimate the principle of fingerprint image, and rotate query fingerprint image with respect template image, then Using spectral based point pattern matching method to do fingerprint minutiae matching. 67. For point pattern matching, choose any pair of candidate corresponding points, compute the transformation parameter value(DeltaX, DeltaY,Delta_Theta), Using meanshift method to find the densest cluster center which is the optimal transformation,Then transform the point set using the current found transformation, Then Using spectral point pattern matching method to match the transformed point set. 68. A new segmentation method based on specializing unary term in graph cut: 1 compute the gray (color) density for each objects; 2 assign unary term for each pixel in graph cut according likelyhood function value. We  expect get better segmentation results. 69. A new nuclei isolation method based on the combination of high gradient magnitude (genuine edges) with ellipsoid fitting: inspired from the tracing process in Canny edge detection method, the graph cut segmentation results mark the genuine ridges. Then we trace the ridges with hysteresis thresholding or using ellipsoid fitting to isolate individual nuclei. We can also use convex decomposition. 70. Nuclei segmentation method: 1. Get foreground nuclei by graph cut using appropriate parameter;2. Based on the segmented foreground, isolate nuclei using combination of graph cut with parameter focusing on fine ridges with ridge tracing skill which ensure the new added ridges connecting to existing edges in segmented foreground. 71. edge based connected component labeling: 1. detect edges with binary image; 2. tracking each edge: begin from first pixel on the edge, scan the foreground pixels along the horizontal row to which the first edge pixel belongs. If a new edge is met, add this edge to the edge list to be scanned. When a background pixel is met, scan beginning with the edge pixel stops and the edge pixel is denoted with status "scanned". repeat above steps until all of edge pixels are searched. Thus achieve the connected component labeling task. This method is simple and efficient since it uses edge to begin with. the speed is fast and computation memory is low since we only keep tracking of edge pixels which is far less than the amount of foreground / image pixels. TIP. 72. edge dilation based connected component labeling: 1. detect edges with binary image;2. denote each connected edge with a id number. dilate each edge on foreground with structure element 3 by 3 neighborhood. 3. one region with different id numbers may appear due to one region consisting of many edges. For this case, label the region with one id number. 4. the performance should be good since fewer number of edges is used compare to large number of foreground / image pixels. memory requirement is slow and speed is fast. TIP 73. elliptic feature based watershed method for nuclei / polyp segmentation: 1. compute elliptic feature, refers to paper "Convex Hull of Elliptic Features for 3D Blob Detection"; 2. refine elliptic feature point by deleting small area;3.use those remaining feature points as foreground  to initiate watershed segmentation method;4. refine the results.TIP. (extended to 3D case) 74. elliptic feature based graph cut method for nuclei/polyp segmentation: 1. compute elliptic feature, refers to paper "Convex Hull of Elliptic Features for 3D Blob Detection";2. using the elliptic degree to set up the energy  function involved in graph cut;3. run graph cut;3. refine results.TIP. (extended to 3D case) 75. elliptic feature based graph cut with elliptic shape prior for nuclei / polyp segmentation:  1. compute elliptic feature points, refers to paper "Convex Hull of Elliptic Features for 3D Blob Detection"2. using the elliptic feature points as the initial elliptic shape to initiate graph cut with elliptic shape prior, referring to paper "Graph cut segmentation using an elliptical shape prior". 4. refine results. TIP. (extended to 3D case) 76. Watershed with gradient L0 smooth for nuclei segmentation: 1. gaussian filter and median  filter, Coherent filter; 2. L0 smooth filter;3. imregionalmax / imextenedmax to get   binary image as seed points for watershed. 4. graph cut to find background as seed points for watershed;5. watershed and refine. (makes sense by initial experiments) 77. graph cut based on ellipse shape  prior: 1. gaussian filter, median filter, coherent filter, L0 smooth filter;2. imregionalmax/imextendedmax to get seed points, fitting the seed points with ellipses. these ellipses can be used the initial ellipses in graph cut. refers to paper "Graph cuts segmentation using an elliptical shape prior". 78. using graph cut to select optimal rows of sensing vectors from our hadamard sensing matrix based on maximizing variance. refer to our paper "Optimized compressive sampling for passive millimeter wave imaging". interaction potential is large for label (0,0) or (1,1), otherwise inversa. 79. graph cut can be used in many optimization problems, such as denoise/ deblur, compressive sensing. refer to  http://www.math.ucla.edu/~wittman/Fields/Lec5.pdf. so many work can be done.^_^ 80. video compressive sensing using multitask compressive sensing with Dirichlet process prior for sharing information. refer to "multi task compressive sensing" "Analysis of sparse nayesian learning""A baysian analysis of some nonparametric problems" 81. simultaneous compressive sensing and denoising make sense with low percentage of measurements ,refer to paper CSDN(compressive sensing with denoising). 82. hybrid compressive sensing with fourier basis for low frequency signal and wavelet for high frequency signal. 83. combine Derin's gradient based reconstruction with sparse bayesian reconstruction to improve speed and do compressive sensing adaptively. 84. Sparse bayesian learning based signal deconvolution: h_1 is high pass  filter , F is fft transformation matrix, x is original signal, F*x is fft transformation of signal x, F*h_1 is fft transformation of filter h_1, y is convoluted signal of signal x with filter h_1. x can be written as: x=(F'./repmat(F*h_1,[N 1]))*F*theta, where theta is sparse coefficients. Thus, any signal can be represented as sparse combination of dictionary ((F'./repmat(F*h_1,[N 1]))*F). Any deconvolution or compressive sensing problem can be solved using fast sparse bayesian learning algorithm. 1) For deconvolution problem, h_2 low pass filter, y is convolution of signal x with filter h_2, so y=F'*(F*x.*F*h_2) (convolution equals dot product of fourier transformation). H_2 is the convolution matrix of filter h_2. hence y=H_2*x=H_2*(F'./repmat(F*h_1,[N 1]))*F*theta, theta has sparse prior. 2) for compressive sensing, y=phi*x=phi*(F'./repmat(F*h_1,[N 1]))*F*theta , theta has sparse prior. 85. set sensing matrix to zero in the place where corresponds to zero coefficients of sparse signal. The place is the high frequency place for common signal with fourier transformation. I have test it and prove that better performance can be gotten in terms of PSNR or less measurements. This is due to reduce the subspace needed to be searched. 86. Inspired by "learning with dynamic group sparsity" we know that many real sparse signal(sparse representation by basis) has clusters of nonzero coefficients. It means that normal sparse prior with diagonal co-variance matrix is not appropriate for real signal with clustered nonzero coefficients. We hope to use Markov random field to model the prior and develop new sparse bayesian learning method. Graph cut is appropriate to optimize the function and find the estimated sparse solution. 87. Following "fast marginal likelihood maximisation for sparse bayesian models", iteratively estimate reconstructed images and covariance and hyperparameters using adaptively added measurement. Derive new update formula for new added measurement. No need to do reconstruction from the scratch. 88. For group sparse compressive sensing, we propose model min |y-\Phi*x|+\lamda*|x|^1+\gamma*|D*H*x|^1, where H is low pass filter matrix, D is high pass derivative filter matrix, this model encourage structure sparse signal to be reconstructed. Inference refer to paper "minimization of the l1 norm of derivative" 89. structure sparse compressive sensing and edge detection simultaniously. Gradient based compressive sensing with structure sparsity + edge detection 90. variational Bayesian blind deconvolution based on the idea in the paper : http://www.aliahmed.org/code.html. Use this idea to do simultaneous compressive sensing and deconvolution. 91. Sparse template and graph cut based nuclei segmentation for 2D and 3D. two papers. refer to the paper "sparse template based variational image segmentation". First use predefined dictionary. Then learn the binary dictionary using inputting images. refer to papers "Learning Non-Negative Sparse Image codes by convex programming","Binary Matrix Factorization with applications". We can do sparse shape decomposition for binary images and gray scale images. 92. explore the structure of block circulant matrix with circulant block to do simultaneous compressive sensing and deconvolution, refer paper "Learning circulant sensing kernel"  ftp://ftp.math.ucla.edu/pub/camreport/cam12-49.pdf 93. propose binary NMF method based on the idea in the paper "Sparse template -based variational Image segmentation" 94. A new method for matrix inverse: given nonsingular matrix X, compute its inverse Y: Model: min |I-XY|*, where * is nuclear norm or other norm. this is a convex minimization problem. Hopefully solved. 95. A new Sparse Bayesian video compressive sensing method: Model: min ||Y-\Phi*(AB^t+E)||^2, where A,B are low rank and column sparse matrix, E is elementwise sparse.    these priors found in the paper "Sparse Bayesian Methods for Low-Rank Matrix Estimation" source code available. 96. Propose new denoising model based on Rudin-Osher-Fatemi(ROF) model: min (TV(U)+\lamda*(U+E-Y)+\omega*||E||^1), where U,E variable, Y given (i.e. image), E elementwise sparse (representing high noise or small objects), U representing smooth objects/signals (low Total variation value).  we can also formulate this problem with variational method or Bayesian method. quite interesting. refer papers " Algorithm for finding global minimizers of Image segmentation and denoising models" and "Sparse Bayesian Methods for Low-Rank Matrix Estimation" source code available. TIP 97. Propose new graph cut model: min \lamda*R(L)+B(L)+\omega*TV(L), where L labeling variable, R(L) and B(L) representing region and boundary terms, TV(L) penalize different labels among neighboring pixels. New term TV(L) representing label cost making L smooth. simultaneously,TV(L) make objects circular since circular object has smallest perimeter compared to non-circular objects with the same area.  similar to variational method to find smooth L. Refer to my paper " Fruit Fly eye nuclei segmentation based on graph cut and convex shape prior using fluorescent micropscopic images". 98. ROF denoising model:  min(wrs U) TV(U)+\lamda*||U-Y||^2 , or equivalent geometrical model: min(wrs U) Per(U)+\lamda*|U \xor Y|, where Y given binary region, U binary region, Per(U) perimeter, \xor symmetric difference between two sets, |.| Lebesgue measure. this problem can be solved by Graph Cut algorithm. 99. Propose TV based L1 reconstruction error compressive sensing using Bayesian framework. reference paper deran's paper "variational bayesian blind deconvolution using a total variational prior". 100. Image segmentation can be done using L1 minimization problem: Consider observed image I as observation measurement, using all kinds of regularization  (Bayesian sparse regularization/ determinated regularization). refer to paper "image smoothing via L0 gradient minimization","Bayesian gradient based compressive sensing" from derin. 101. sparse dictionary constrained transformation invariant sparse coding (motivated by paper "Transformation invariant sparse coding": in that paper, the learned dictionary is highly dependent on the value of sparse regularization parameter \lamda. When the regularization strength is weak or too strong, the dictionary solution is not sparse. To avoid this case, we proposed to constrain the dictionary solution with sparsity which can be imposed by any image sparsity measure, such as sparse gradient (from derian paper). 102. Binary TISC (transformation invariant sparse coding) mode: motivated by paper(P_TISC) "Transformation invariant sparse coding" which proposed TISC model for general image and paper (P_BMF)"Binary matrix factorization with applications" which proposed binary NMF. Our idea is to constrain TISC model with binary constraints by penalty method or thresholding method proposed in the paper P_BMF. 103. graph cut and convexity -concavity based splitting method: 3D graph cut for foreground segmentation. Convexity -concavity based splitting in xy plane and xz plane to split 3D  contacting nuclei. 104. Transformation invariant image segmentation: motivated by the paper "Sparse Templated-based variational image segmentation" which proposed a method to segment foreground with sparse template but the number of templates is potentially huge comparable to the number of image pixels. We propose to using convolution to reduce the number of all possible templates to solve template translation problem the idea as the paper "Transformation invariant sparse coding". we have implemented and tested the translation invariant image segmentation. (Pattern Recognition or others). In the future, we will extend the method to rotation invariant segmentation. It may be possible to learn the templates(dictionary) using the same optimization method. 105. propose greedy optimization algorithm to optimize the objective function in the paper "Sparse Templated-based variational image segmentation" similar to greedy compressive sensing method referring to greedy compressive sensing paper. It is pretty important to learn the binary templates(dictionary) using optimization method. we can refer to the relevant papers on dictionary learning in non-negative matrix factorization (NMF multiplicative update rule) or compressive sensing (alternative method) fields. 106. propose double sparsity in sparse coding and dictionary in NMF or TISC or Tensor Decomposition model. Using L1 norm to impose sparsity constraint on coding is easy and direct. But it is not easy to put sparsity constraint on dictionary. We propose the gradient sparsity constraint on dictionary. That means we assume the gradient of dictionary is sparse. The assumption is reasonable for natural images. refer to the paper " millimeter wave compressive imaging". 107. Hadamard matrix based compressive sensing without mask: inspired by the work "single exposure super-resolution compressive imaging by double phase encoding", we propose to generate a hadamard matrix as a random phase mask to modulate the Fourier output of signal. That is y=Fx.*Fm. where F,x,m are Fourier matrix,input image, hadamard matrix, respectively. In this way, we only have to generate one Hadamard matrix without needs to generate multiple masks and move masks. But we need to move detector in imaging plane. 108. Model based millimeter wave imaging compressive sensing: 1. using bayesian compressive sensing framework with the prior probability of derivative. 2. using thresholding solver to find first k basis. 3 using found k basis as node to construct graph, data term is the reduced error after project observed signal onto corresponding node (basis), smooth term is related to neighboring relationship which can be modeled by Bernouli distribution. 3. using graph cut 109. Dictionary Based sparse segmentation:1) learn different dictionaries for different object classes;2)nonoverlapping patch sparse representation;3) using sparse coefficients associating dictionary to segment patch. 110. Graph Cut Based compressive sensing: 1)Generate structured sparse signal to simulate sparse structured signal;2)Get observation using sensing matrix y=phi*x;3)using each entry in sparse signal x as node to construct graph G(V,E); t weights are the similarity  relationship between y and phi_i (column of sensing matrix phi); n weights are the potts model for zero label and non zeros label (different labels with neighboring pixels have large weights, encouraging structured signal.). label cost is larger with nonzero label in order to impose sparse constraint.3) using graph cut to solve compressive sensing reconstruction problem. 111. Fingerprint singular point detection: 1) detect singular points with low accuracy of position using our model based sliding filter response singular point detection method. Due to the low accuracy of orientation computation, there is relatively large position displacement between true singular points and detected singular points. 2) using ridge symmetry to locate singular points accurately. Draw circle with detected singular point as center. There are multiple intersecting points on the circle with ridges. Find one ridge inside the circle with two ending points on circle. Between the two ending points over the arc, there have to be multiple intersecting points. Beginning with the next ending point, trace another ridge which is as long as possible. Until no ending points left. The last traced ridge has the most curved point or ending point which is the singular point. 3)using convex hull to correct the position of singular points. Pattern Recognition, IEEE Signal Processing Letter.   111b. Vote for correction of singular point positions from convex hull geometry: 1) using our AMF features (in our CVPR 2014 paper) to detect candidate singular points; 2) to be robust to noise, Our CVPR2014 paper uses large Gaussian smooth to smooth orientation field. This large smooth highly smooths high curved areas (core area) and produces large displacement of core point position from true positions. for each core point, draw a circle with the core point as center; 3. begin with each ridge ri in the circle with ending points inside the circle or on the circle, compute its convex hull and detect new ridges intersecting with the convex hull, for each new ridge, compute its convex hull and detect new ridges intersecting with the convex hull. repeat the process with new ridges until no new ridges are added into convex hull. the vote for all ridges are initialized to zero. for ridge ri, votes for all new rides intersecting with convex hulls during the above process are increased by one. 4. repeat the process for all ridges inside the initial circle and increase the votes by one for all ridges intersecting with a sequence of convex hulls. 5. the ridge r with majority of votes are the ridge contain core point. detect the point with most curvature or top most ending point on the ridge r as core point. TFIS, Pattern Recognition 111c. Core correction for large displacement: 1) if orientation field is severely smoothed to resist noise, core points in our AMF method(our CVPR2014 paper) are far away from true position; 2) we can repeat the process "Vote for correction of core points from convex hull geometry" in above 111b until the process convergence. The final detected core points are considered as true core points. We show that the process must convergence at the top most inner ridge (core). TFIS, Pattern Recognition. 112. Estimate orientation of singular points based on zero-pole model: 1. using our convergence index filter based method to detect singular points;2. compute the background orientation based on standard pattern around single singular point. 113. image retrieval: 1. extract features (class of superpixels) refereing  to paper " Class Segmentation and Object Localization with Superpixel Neighborhoods";2. get the normalized histogram of features in one image and second order histogram of features (neighboring feature pairs);3. compute the distance between the histogram of querying image and the ones of database images. we are done. 114. graph cut based edge detection:1. compute gradient image of gray scale image;2. construct graph based on the orientation of gradient vectors:more penalty along the orientation of gradient vectors and less penalty along the direction orthogonal to the gradient direction. TIP. 115. scene understanding: 1. extract features (class of superpixels) refereing  to paper " Class Segmentation and Object Localization with Superpixel Neighborhoods"; 2. with labels associated with each superpixel, retrieve the images with most number of the same labels, and construct neighboring graph(weight=1)  reflecting the neighboring relationship. compare the similarity between two graphs with nodes labeled with different labels by subgraph isomorphism refering to paper " computing subgraph isomorphic queries using structural unification and minimum graph structures"; Note that node category (labels) should match between the two graphs when computing the similarity between two graph. By the way of comparing two graphs, the transformation between two image has no effect on the result. It is appealing. TIP 116. graph based fingerprint matching: 1. extract minutiae from fingerprint image;2. construct neighboring graphs by checking the neighboring relationship between ridges where minutiae are lying. 3. then compare two graphs by subgraph isomorphism refereing to paper " computing subgraph isomorphic queries using structural unification and minimum graph structures". 117. Transformation invariant dictionary based compressive sensing: 1. learn a transformation invariant dictionary (set of feature vectors) referring to the paper " transformation invariant sparse coding" 2. test compressive sensing with this dictionary. 118. binary matrix sparse factorization: 1. using the binary matrix factorization method in paper "Binary matrix factorization with application" with sparsity constraint. 2. replace normal matrix product with boolean matrix product.3. using convolution to make it translation invariant. 119. shape learning based sparse template segmentation: 1. learn shape dictionary by statistical shape learning method;2. use graph cut to obtain foreground binary image;3. sparse template based segmentation. TIP 120. sparse and low rank matrix recovery using convex programming: 1. blind deconvolution as low rank matrix recovery using convex programming, refer to the paper "Blind Deconvolution using convex programming"; 2 we put sparse constraint on the w and m, we have x=Cm, w,m\in R^L. C\in R^(L*L). Thus we extend w, m to L dimension space by sparsity constraint. w is sparse rather than its representative coefficients in low dimension subspace R^K in paper  "Blind Deconvolution using convex programming". 3. our optimization problem: min ||X||*+||X||_1, s.t. y^=A(X), where y^ is the FFT of y, A is linear operator in the paper "Blind Deconvolution using convex programming". TIP, IJCV special issue-sparse coding. looks like RPCA :  http://perception.csl.illinois.edu/matrix-rank/introduction.html; or optimize  min ||H||_1, max ||M||_2 , s.t. y^=A(HM*). the high frequency components are increased by maximizing ||M||_2; filter should be sparse by min ||H||_1. 121. Sparse convex blind deconvolution. We reformulate the problem min ||X||*, s.t. y=A(X) (min ||W||2+||H||2, s.t. y=A(WH*)) in  the paper "Blind Deconvolution using convex programming" as min||w||1+||h||1, s.t. y=A(wh*), 1. using augmented Lagrangian method (multiplier method) as in "Blind Deconvolution using convex programming" to solve the optimization problem to get W and H, then shrinkage the W and H.  Investigate the relationship between these two models by experiments.  questions: do they still find the global solution? TIP;  2. Optimize the problem with NMF method with sparsity constraint in the paper "Non -negative Matrix Factorization with Sparseness constraints" (source codes available). This method has non -negative constraints, TIP; 3. Min ||W||2+||H||2, s.t. y=A(WH*), 3.1) using Sparse Bayesian Methods in the paper "Sparse Bayesian Methods for Low Rank Matrix Estimation" with the change-\gamma_i different for all components of W and H. TIP. 3.2). The Gaussian prior p(A|gamma) and p(B|gamma) in the paper "Sparse Bayesian Methods for Low Rank Matrix Estimation" can be replaced with Laplace distribution or scale Laplace mixture distribution refer to  the paper "Bayesian Group-Sparse Modeling and Variational Inference(Source codes available)" and "Group Sparse Coding with a Laplacian Scale Mixture"(Reweighted iterative method). TIP. 122. Fast Iterative Shrinkage/Thresholding Algorithm: 1. min f(x)+||x||1, change the model as min f(x)+||x||2 (smooth regularization), optimize until convergence and get x0 as the initial point of the following model min f(x)+||x||1, we hope the convergence is faster compared to normal method.  Refer to the paper "A Fast Iterative Shrinkage-Thresholding Algorithm for Linear Inverse Problems". TSP. 123. sequential on line compressive sensing with automatic stoping rule:1)stachastical gradient descent method to learning x_(n+1)=x_n+\yita*(y_n-\phi*x_n)*x_n-\yita*\lamda*x_n;2) if the change rate is too small. then stop. AO. 124. content sparse based segmentation and scene understanding model.1. minimize ||I-\sum \dotprod[conv(D_i, E_i) . W_i]||_2+\sum ||E_i||_1+\sum||E_i||_1, where I is image, D_i are atom in dictionary, E_i are sparse coefficient images, W_i is sparse binary images. 2. dictionary can be learned by the transformation invariant sparse coding method " transformation invariant sparse coding". In the future, we plan to learn the dictionary from our above model. 3. optimization by alternating gradient descent method between E_i and W_i as the same as  " transformation invariant sparse coding" considering W_i is continuous real number value. we can also consider W_i is binary values and be optimized by greedy method. first initialize all W_i with one, then greedily optimize each entry in W_i by check the costs corresponding to when the entry is 0 or 1. 4. This is a universal and wide application model in segmentation and scene understanding incorporating global and structural information. It promote the idea that using the number of objects to interpret scene as less as possible. TIP, IJCV,PAMI 125. statistical features of regions and its application to scene categorization. 1. use superpixel to segment images into regions. Or use sparse gradient based segmentation method to divide images into regions; 2. statistical histograms on region area, region shape,region geometrical features as feature vectors (reference to matlab region shape properties such as ;3. train a classifier to predict scene categories. such as logistical regression or SVM.  code and paper reference to " Beyond Bags of Features: Spatial Pyramid Matching for Recognizing Natural Scene Categories" . This method is assumed to outperform spatial pyramid matching. CVPR, ICCV. 126. Logistic regression and graph cut for nuclei segmentation:1. use superpixel or sparse gradient method to divide image into regions, using region features and logistic regression to learn the data term in graph cut using training data; 2. with this data term using graph cut to segment nuclei. refer to " Finding Things: Image Parsing with Regions and Per-Exemplar Detectors". Pattern Recognition, Pattern Recognition Letter, Journal of microscopy. 130. fingerprint singular point detection based on the combination of Poincare Index and circular shape analysis:1. get Poincare index for fingerprint  images;2. detect circular regions which are considered as singular points with assumption that true singular point region has a perfect circle and noise area doesn't have a perfect circle. Pattern Recognition,Pattern Recognition Letter. 130. Recursive Point Pattern Matching in Fingerprint recognition and scene understanding:1. extract interesting points as point pattern for each image;2. construct kd tree for fast neighbor searching;3. start with local 2 nearest neighborhood as a sub-structure and match each sub-structure pairs between template and querying image by computing the distance difference of lines connecting feature points and relative angle difference between straight lines formed by connecting points; give a score for the sub-structure matching; 3. for each matching sub-structure, delete center point, find the 5 nearest neighborhood for each remaining points in the sub-structure from the other points without consideration previously; In the 5 nearest neighborhood, find the best 3-point sub-structure matching pair, scoring the best matching pair by weight which indicate what nearest neighborhood  at least contains the best substructure. (for example, for nearest 2, 3, 4,5, the associated weight becomes smaller indicate the degree of noise). 4. repeat the step 3 until all points are visited or the matching score is below some threshold. PAMI, IJCV. 131. convolution sparse representation neural network:1. learning sparse transformation invariant dictionary by convolution operation; " transformation invariant sparse coding" 2. get sparse codes for each image based on learned dictionary ; 3. the sparse coding step can be considered as the first layer of NN which is equivalent to the several layers of Deep Neural network functioning as convolution and max pooling;4. the last layer is the fully connected layer( linear /softmax). In this way, we propose 3 layer shallow NN compared to DNN . hopefully, they have the same performance, but the training is easy. PAMI,IJCV,NIPS,CVPR. 132. double stachastic process based bayesian compressed sensing: 1. using double stachastic process as the prior distribution of the sparse coefficients of images under 3D complex wavelet transformation; http://post.queensu.ca/~gazor/acsdocfinal.pdf 2. bayesian based method to esitmate sparse signal.  SP,SPL,Applied Optics, TIP. 133. sparse coefficient distribution based compressive sensing matrix create: 1. get the distribution of the sparse coefficient of natural images by histogram statisics.denoted by p(s_i) for ith coefficient. it is bernouli distribution of zero and nonzero.2. use the bernouli distribution as the distribution of each entry in the compressive sensing matrix and sample the distribution to create the desired sensing matrix. 3. This created matrix has better performance than general gaussian matrix. For another paper, this method can be combined with using the double stachastic process for sparse coefficient distribution prior in above 132. SP,SPL,Applied optics,TIP. 134. Sparse Bayesian based compressive sensing:1) observation model: N(y|\phi *w,\beta);2. signal model: N(w|,0,\alpha_i); 3. hyperprior \gamma(\alpha_i|a_i,b_i); 4. hyperhyperprior: \gamma(a_i|\lammda),\gamma(b_i|\lamda);5. reference paper " Bayesian Compressive Sensing Using Laplace Priors"  http://www.dbabacan.info/papers/babacan_CS.pdf 135. Three -Stage model for Sparse Bayesian Learning and the Relevance Vector Machine: 1) 1) observation model: N(y|\phi *w,\beta);2. signal model: N(w|,0,\alpha_i); 3. hyperprior \gamma(\alpha_i|a_i,b_i); 4. hyperhyperprior: \gamma(a_i|\lammda),\gamma(b_i|\lamda);5. reference paper " Bayesian Compressive Sensing Using Laplace Priors" and " Sparse Bayesian Learning and the Relevance Vector Machine ", TSP,PAMI,TIP. 136. Three -Stage model for fast marginal likelihood maximization for sparse bayesian models:  1) observation model: N(y|\phi *w,\beta);2. signal model: N(w|,0,\alpha_i); 3. hyperprior \gamma(\alpha_i|a_i,b_i); 4. hyperhyperprior: \gamma(a_i|\lammda),\gamma(b_i|\lamda);5. reference paper " Bayesian Compressive Sensing Using Laplace Priors" and " fast marginal likelihood maximization for bayesian models ", TSP,PAMI,TIP. 137. L1 norm of gradient based compressive sensing: 1) min ||y-\phi*x||^2+\lambda*||x_i-x_i+1||_1; 2)optimization based on FISTA algorithm. TIP,TSP. refer " A new algorithm for compressive sensing Based on Total variation Norm" 138. fast TV based Bayesian Compressive sensing: refer to paper : "variational Bayesian Blind Deconvolution Using a total Variation Prior" and "Bayesian compressive sensing using laplace priors", TSP. 139. nonlocal prior based Bayesian compressive sensing and debluring: refer to : https://www.ceremade.dauphine.fr/~cohen/mypapers/GabrielPeyreNLRECCV08.pdf 140. statistical frequency coefficient model for compressive sensing and sensing matrix design. 1. use the frequency magnitude model in http://research.microsoft.com/pubs/138795/tashev_acero_statistical_modeling_of_the_speech_signal.pdf  as the sparse prior model to do bayesian compressive sensing and compare its performance with the bayeisan method with parameter estimation. 2. using the scale parameter values (\chi) in paper   http://research.microsoft.com/pubs/138795/tashev_acero_statistical_modeling_of_the_speech_signal.pdf to generate the success probability p  proportional to \chi of bernoulli distribution to design sensing matrix. Signal Processing. TIP. 141.  learn double sparsity convolutional sparse coding for segmentation and object recognition: 1. learn a sparse dictionary by convolutional operation. refer to paper " Double Sparsity: Learning Sparse Dictionaries for Sparse Signal Approximation" and paper " Fast Convolutional Sparse Coding" ; 2. obtain sparse coding by learned dictionary for object recognition or segmentation.. TIP/PAMI 142. Sift feature and dictionary learning for object recognition and image classification/retrieve: 1. extract sift feature ; 2. learn dictionaries for each class by RICA (use sift position and histogram) and build sparse histogram(reference to our recent paper "Joint Dictionary learning for activity recognition"); 3. use svm classify or use histogram to finish retrieving. TMM/TIP. 143. fingerprint reference orientation definition and estimation: 1. use our AMI feature to detect singular points(CVPRW2014); 2. compute reference orientation o_\infinity according to equation 2*o(z)=2*o_\infinity+arg(z-z_c) for core point and  equation 2*o(z)=2*o_\infinity-arg(z-z_d) for delta point. equation (5) in our paper. IEEE Trans. Forensic and information security. 144. Down syndrome detection based on constrained NMF: 1. learn dictionary based on NMF using training data. 2. search landmark points based on local model match and global NMF model match. 3. extract features after obtaining landmark points. refer to the paper  published in Medical Image Analysis 2014 " Digital facial dysmorphology for genetic screening : Hierarchical constrained local model using ICA MedIA2014Dysmorphology." 145. Down syndrome detection based on pictorial model: 1. learning a pictorial model  from manual landmark points for normal people and down syndrome people; 2. match pictorial model to test image and classify based on posterior probability.  3. without the need to search landmark points; model match can finish landmark search and classification at the same time. It is global optimal other than local optimal. refer to the paper "pictorial structure for object recognition" source code http://www.d2.mpi-inf.mpg.de/andriluka_cvpr09 and http://people.csail.mit.edu/fergus/iccv2005/partsstructure.html 146. extend bag of words using transformation invariant sparse coding NMF: 1. learning dictionary based on transformation invariant NMF (compensate for object  transformation in image and invariant to object rotation and displacement, different from conventional dictionary which is sensitive to transformation and position); 2. compute sparse codes, get histogram of word frequency and statistics (center moments) for each word as feature vectors. do feature selection to select features (hopefully, choose the features corresponding to distinctive words). 3. train classifier using selected features. Advantage:dictionary is invariant to object rotation and displacement similar to visual eye.  part-based local words has strong discriminating power. feature selection process may select those distinctive words out as features. this is big contribution. refer to paper "Transformation invariant sparse coding" and "transformation invariant representation and NMF. 147. transformation invariant sparse coding for scene classification/object recognition: 1) using transformation invariant NMF to learn a dictionary for each class  (atoms are different objects or different parts owning high distinctive power); this class specific dictionary characterize the contents of object/scene; advantage is the dictionary is invariant to image transformation (displacement, scale, rotation); 2) for each training image, compute its sparse coding c, get the histogram h of c with each bin corresponding to one class. h  can be used as descriptive feature vectors to train a classifier.  3) for each test image; compute its h and classify it. Novelty: transformation invariant recognition consistent with human vision. refer to papers "transformation invariant sparse coding" and "robust face recognition via sparse representation".  PAMI 148. Shape retrieval by transformation invariant sparse coding for binary image:1. put basic shapes in dictionary; 2) for test shape, compute sparse coding and classify test shape based on sparse coding. 3) our algorithm also including displacement invariant. refer to TIP2013 paper " in plane rotation and scale invariant clustering using dictionary";   TIP 149. Improve accuracy of FISTA by locally replacing L1 with square of L2 around zero point: 1. ||x||_1 replaced by ||x||_2^2 around zero point; 2. compute gradient G and compute x_k=x_(k-1)-t*G ,  x_k=x_k +trend*(x_k-x_(k-1) and x_k=threshold(x_k, \theta)  based on gradient descent with box constrained to avoid vibration around zero point. 3. we get more accurate results than FISTA while keeping the same computation complexity. refer to paper " a fast iterative shrinkage thresholding algorithm for linear inverse problems",  TSP Other methods source code  http://www.eecs.berkeley.edu/~yang/software/l1benchmark/ 150. Improve accuracy of L1 optimization by hierarchical combination of convex norm L1 with non-convex norm Lp: 1. Use FISTA get non accurate results x with L1 norm; 2. Use FISTA get accurate results with Lp norm (0<p<1) by initializing x0=x which come from step 1; 3. trade off between accuracy and speed. Lp (p>=1) has fast speed but results are less accurate while Lp(p<1) has more accurate results but speed is slow. That's why we combine them appropiately. TSP Other methods source code  http://www.eecs.berkeley.edu/~yang/software/l1benchmark/ 151. Singular point search based on Zero-pole model: 1. choose center O of fingerprint image I as reference point and any point P in the image; compute the main argument  arg(OP) of complex vector \overarrow{OP} to get an argument map called as ArgO; 2. for any point Q in image; the argument map ArgQ wrt Q can be computed quickly as the translation of ArgO, i.e. ArgQ=(ArgO translated with vector Q-O); 3. therefore, Argument map only compute once for the whole image. ; 4. search operation is constrained within bad quality area to reduce the searching time and orientations in good quality area are used to satisfy Zero-Pole model since the orientation in bad quality area is not convincing;  5. The algorithm is parallelable and speedup  can be done with multicore or multi threads; 6. two papers can be produced by combining it with PoinCare Index or our AMI method in CVPRW2014 as initial candidate singular point detection. TFS and PAMI, TIP. 152. Extension of FISTA to general sparse coding model: 1. min ||Wx||_1 is equal to saddle point problem: min_{x} f_{\mu}(x)=min_{x} max_{u} <u,Wx>+\mu*p(u), where p(u) is a prox function such as l2 norm: p(u)=||u||_2^2; 2. The gradient \delta f_{\mu}(x) of saddle point problem can be computed as \delta f_{\mu}(x)=W^t*u^\star (x), where u^\star (x) is the maximum argument of max_{u} <u,Wx>+\mu*p(u). Therefore the non-smooth function is smoothed by saddle point problem and the gradient vector can be computed; 3. we plug the computed gradient of smooth function into FISTA by replacing the shrinking threshold operation with gradient descent with the computed gradient followed by threshold operation. Thus we extend FISTA to general sparse coding model such as ||Wx||_1, TV model. refer to the paper " NESTA: a fast and accurate first order method for sparse recovery". TSP,TIP. 153. Fingerprint orientation field enhancement based on low rank matrix and sparse representation: 1) compute initial orientation field Y using normal method (such as in our CVPR2014 paper.); 2) Y can be represented as Y=X+E+N. we solve the optimization problem: min ||X||*+\lambda*||E||_1+\gamma*||Y-X-E||_2, where ||X||* is nuclear norm of matrix. It can be solved by RPCA (refer to Yi Ma's paper). 3) Computed X is expected to be the enhanced orientation field estimation. TIP, PAMI, TFIS. 156. Adaptive orientation field smoothing based on singular area detection: 1) compute curvature map using method in our ICIP paper; 2) detect singular area based on curvature map by training some classifier (SVM); 3) For non-singular area, use large orientation smooth to resist noises such as scars, dryness. For singular area, we use small smooth to keep the high curvature area with small distortion as possible. In this way, the smoothed orientation field does not only resist noise, but also distort singular area as little as possible. IEEE SPL 157. Accurate method for Singular point detection: 1) use zero pole model to derive "Orientation Coherence Feature" with the analogy to our CVPR2014 paper in which we have a formula on the orientation difference between any two points z_1 and z_2  in fingerprint image. In that paper, we show that the orientation difference o(z_1)-o(z_2) is determined by all these straight lines which pass through z_1 and one of singular points or pass through z_2 and one of singular points. This demonstrate that orientation field is determined by singular points. The inverse is true that singular points are determined by orientation field. But it is hard to compute because it is highly nonlinear function. The simple way is to simplify the formula by constraining the sampling point pair z_1 and z_2. In CVPR2014, we propose a special sampling pattern and the sampling pattern can be elegantly used in convergence filter. 2) after compute the candidate singular points using the method in CVPR2014 paper, we borrow the idea that determine possible combinations of singular points in fingerprint image from the paper "A novel Algorithm for detecting singular points from fingerprint images" by computing Porn Care Index along the valid border around fingerprint image. Once determining the possible number of singular points, we can refine the positions of possible singular points by searching on the bad quality area and the small area around detected singular points from CVPR2014 paper. In this way singular point positions are not only refined, but also the missed singular points in bad quality area can be recovered. TFIS, THM,IEEESPL 158. Accurate Core Point detection based on geometrical shape analysis: 1) using our method in CVPR2014 to detection candidate singular points.; 2) remove small bridges in thinned images since they come from noise. Especially, in singular region, orientation smooth may produce many false bridges which affect our later geometrical shape analysis. It may make vote based convex hull analysis fail; 3) correct core point position by geometrical shape analysis. For each core point C, draw a circle around it, in this circular region,  find ridges with maximum number of votes from convex hull cover; find the intersecting points between circle and maximum vote ridges. compute the mass center m of the intersecting points. compute the intersecting points p closest to mass center m. the closest intersecting point p as the trace starting point. Assume that component L contains starting point p , remove branch points in component L results in several connected components from L. Find component r containing  starting point p. Detect ending point list epl of ridge r . If epl contains at least one ending point P_e which is not intersecting point, then P_e is corrected core point. otherwise find the point P_c with maximum curvature which is corrected core point; 4)repeat step three until correct core point converges, the final corrected core point can be found. In this way, far away core point can be corrected. Also, false core point may be removed if we require the core point must stay in area which is curved to some degree which vote difference should be large enough. TIP,TFIS,PR 159. Symmetry based fingerprint image enhancement: 1) use complex analytic function \sqrt(z) as coordinate transformation to transform fingerprint image f(x,y) to coordinate plane (\delta,\etta); 2) transformed image f(\delta \etta)=f(x,y) contains parallel lines at least in the neighborhood of core and delta points in fingerprint images; 3) we can enhance the parallel lines easily by gabor filter with one constant direction \omega which can be detected by Hough line detection method; 4) after enhance the parallel lines, then transform back to (x,y) plane to get enhanced fingerprint image; refer to paper " Pattern recognition in images by symmetries and coordinate transformations ". TIP,TSP,TFIS,PR 160. Singular area enhancement by symmetry: 1) use method in above 159 to enhance singular area. SPL 162. Poincare index and Source region based singular point detection: 1) compute poincare index image P without thresholding the accumulative direction change; 2) according to our CVPR2014 paper, singular area would appear as source region. take Mexican Hat wavelet transformation M of P; 3) assume that local extrema contain singular point. Detect local extrema as candidate singular points and thresholding them based on PI value.  As our CVPR2014 paper, shape is taken into account.   PR,TFIS,SPL 163. Naive Bayesian and sparse dictionary based activity recognition: 1, using RICA learn sparse dictionary as in our PLOS one 2014 paper; 2, Naive Bayesian Classifier is trained using training video; 3, test Naive Bayesian Classifier. PR, PRL 164. Kmeans deep learning: 1. first layer: input; 2. second layer: cluster input x by kmeans, centers as weight matrix; 3. third layer: cluster output h^1 from second layer by kmeans, centers as weight matrix. repeat the process to construct many layers; 4. output layer: fully connected classifier. PAMI 165. Neural network based compressive sensing learning: 1. suppose observation vector t=\phi*s, where \phi compressive sensing matrix , s sparse signal. purpose is to find s given t and \phi; 2. assume s=f(w,t)--a nonlinear function of t with respect to parameters w; assume f(w,t) is a neural network function form; 3. we have normal distribution N(t|\phi*f(w,t),\delta*I) and prior normal distribution N(s,\delta*I). this is two stage neural network with stage f(w,t) and stage  \phi*s (s=f(w,t)). the latter stage is known and fixed. The first stage needs training by BP method; 4. normally, we have only one sample t. using one sample t, we train the network to find w, then find solution s=f(w,t); 5. the feature is new method for compressive sensing; and for new observation, we may quickly get sparse solution by s=f(w,t) without computationally intensive process such as sparse recovery . 6. we can train the network with more samples and thus make the prediction more accurate . PAMI. 166. Auto-encoder based compressive sensing learning: 1. collect a set of sparse signal s with special prior distribution; 2. case 1: compressive sensing matrix \phi given, learn an auto-encoder to do CS reconstruction for some kind of sparse signal; case 2: CS matrix \phi unknown, learn an auto-encoder to do observation and sparse decoding; 3. advantage: fast and adaptive observation. PAMI 167. transformation invariant sparse coding based deep neural network: 1. learn a dictionary D1 by transformation invariant sparse coding; 2. using learned dictionary D1 to obtain sparse coefficients S !.  From S1 , repeat step 1 to get the second dictionary D2; 3. repeat step 1 until we get enough number of hierarchical  dictionaries, each dictionary is considered as one  layer. Thus we get deep NN and hierarchical representation. sparse coefficients may be binarized since the magnitude does not matter. Therefore, all dictionaries may be binary except the first D1;  4. the last layer is classifier. PAMI 168. Bayesian based transformation invariant sparse representation dictionary learning in big images: 1, model  X=\sum_i(D_i*S_i)+N, p(X|S)~gaussian, P(S_i)~sparse distribution, P(D_i)~direchilette distribution; 2. given X, we calculate the derivative of p(X) with respect to S and D, refer to paper " Optimization methods for convolutional sparse coding"  " Sparse coding and ICA" for derivative computation of convolution operation; 3.  we may use full bayesian model or hierarchical Bayesian model. PAMI 169. Sparse coding based temporal pyramid matching using body joint feature for human activity recognition: 1. reference our Plosone paper and IEEE THMS paper and paper " Linear spatial pyramid matching using sparse coding for image classification" . 2. main idea: densely sample joint volume; using sparse coding to learn dictionary; temporal pooling of sparse codes; concatenate pyramid histogram; training SVM by histogram; classify test video; This paper for THMS revision for correspondence; THMS 170. blind deconvolution based on convolution sparse coding: reference paper " Optimization methods for convolutional sparse coding" . ADMM method for filter and image computation. TIP. --------------------------------------------------------------------------------------------------------- Innovation 1. Image registration based on explicit function model: 1) use some explicit function to model data/image/video (such as interpolation reconstruct function from ADC, FFT basis function, polynomial function); 2) once obtaining the explicit function form which is a function of coordinate, we can plug transformation parameters(rotation \theta, displacement \deltaX,\deltaY)  into the function. 3)minimize the error: min e(\theta,\deltax,\deltay)=\int\int[f(x,y,\theta,\deltax,\deltay)-g(x,y)]^2 dxdy with respect to \theta,\deltax,\deltay. Thus we can get the transformation parameters. PAMI --------------------------------------------------------------------------------------------------------- project: 1. abnormal event recognition using visual event recognition in videos:  http://videolectures.net/cvpr2010_duan_verv/ and modeling mutual context of object and human pose :  http://videolectures.net/cvpr2010_fei_fei_mmco/ 2. fingerprint orientation field model; min \sum[arg(P(z))-arg(Q(z))-2*\theta], using gradient descent method. We develop rough gradient descent method: determine the sign of object function gradient by difference computation in two specific points,one of which is the initial known parameter and another is the perturbed parameters. The new gradient descending method is applicable in case of no object gradient available. 3. Create a new camera with binocular lens for the photogrammetry.  share one CCD sensor but taking two successive pictures in short time. Then compute two perspective centers using image processing method with collinearity constraints under image reference coordinate system. This camera only needs to estimate the 3d coordinates for the two perspective centers of two lens once. Later it can be used everywhere, anytime. Furthermore, it doesn't depend on the position and orientation of camera. Very convenient. refer to the paper"Toward a super imaging system" This is for new project or patent. Rapid paper: 1. graph cut+minimum path (experimental results ready), referring to "minimal path algorithms for the robust detection of linear features in gray images"; 2. shift+rotation+scale invariant sparse coding of image/volume (experimental results, formula ready), referring to "shift invariant sparse coding of image and music data". We can scale atoms. 3. shift invariant+3D sparse template based segmentation( simulated 3d nuclei results ready, algorithm can be changed to use "sparse binary matrix factorizaton" in existing paper). ----------------------------------------------------------------------------------------------------------- Business: 1. building a "fear conditioning test" web service for data/video process/analysis and provide intermediate results for research papers. refer to  http://www.jove.com/video/50871/contextual-cued-fear-conditioning-test-using-video-analyzing-system existing edges in segmented foreground. 71. edge based connected component labeling: 1. detect edges with binary image; 2. tracking each edge: begin from first pixel on the edge, scan the foreground pixels along the horizontal row to which the first edge pixel belongs. If a new edge is met, add this edge to the edge list to be scanned. When a background pixel is met, scan beginning with the edge pixel stops and the edge pixel is denoted with status "scanned". repeat above steps until all of edge pixels are searched. Thus achieve the connected component labeling task. This method is simple and efficient since it uses edge to begin with. the speed is fast and computation memory is low since we only keep tracking of edge pixels which is far less than the amount of foreground / image pixels. TIP. 72. edge dilation based connected component labeling: 1. detect edges with binary image;2. denote each connected edge with a id number. dilate each edge on foreground with structure element 3 by 3 neighborhood. 3. one region with different id numbers may appear due to one region consisting of many edges. For this case, label the region with one id number. 4. the performance should be good since fewer number of edges is used compare to large number of foreground / image pixels. memory requirement is slow and speed is fast. TIP 73. elliptic feature based watershed method for nuclei / polyp segmentation: 1. compute elliptic feature, refers to paper "Convex Hull of Elliptic Features for 3D Blob Detection"; 2. refine elliptic feature point by deleting small area;3.use those remaining feature points as foreground  to initiate watershed segmentation method;4. refine the results.TIP. (extended to 3D case) 74. elliptic feature based graph cut method for nuclei/polyp segmentation: 1. compute elliptic feature, refers to paper "Convex Hull of Elliptic Features for 3D Blob Detection";2. using the elliptic degree to set up the energy  function involved in graph cut;3. run graph cut;3. refine results.TIP. (extended to 3D case) 75. elliptic feature based graph cut with elliptic shape prior for nuclei / polyp segmentation:  1. compute elliptic feature points, refers to paper "Convex Hull of Elliptic Features for 3D Blob Detection"2. using the elliptic feature points as the initial elliptic shape to initiate graph cut with elliptic shape prior, referring to paper "Graph cut segmentation using an elliptical shape prior". 4. refine results. TIP. (extended to 3D case) 76. Watershed with gradient L0 smooth for nuclei segmentation: 1. gaussian filter and median  filter, Coherent filter; 2. L0 smooth filter;3. imregionalmax / imextenedmax to get   binary image as seed points for watershed. 4. graph cut to find background as seed points for watershed;5. watershed and refine. (makes sense by initial experiments) 77. graph cut based on ellipse shape  prior: 1. gaussian filter, median filter, coherent filter, L0 smooth filter;2. imregionalmax/imextendedmax to get seed points, fitting the seed points with ellipses. these ellipses can be used the initial ellipses in graph cut. refers to paper "Graph cuts segmentation using an elliptical shape prior". 78. using graph cut to select optimal rows of sensing vectors from our hadamard sensing matrix based on maximizing variance. refer to our paper "Optimized compressive sampling for passive millimeter wave imaging". interaction potential is large for label (0,0) or (1,1), otherwise inversa. 79. graph cut can be used in many optimization problems, such as denoise/ deblur, compressive sensing. refer to  http://www.math.ucla.edu/~wittman/Fields/Lec5.pdf. so many work can be done.^_^ 80. video compressive sensing using multitask compressive sensing with Dirichlet process prior for sharing information. refer to "multi task compressive sensing" "Analysis of sparse nayesian learning""A baysian analysis of some nonparametric problems" 81. simultaneous compressive sensing and denoising make sense with low percentage of measurements ,refer to paper CSDN(compressive sensing with denoising). 82. hybrid compressive sensing with fourier basis for low frequency signal and wavelet for high frequency signal. 83. combine Derin's gradient based reconstruction with sparse bayesian reconstruction to improve speed and do compressive sensing adaptively. 84. Sparse bayesian learning based signal deconvolution: h_1 is high pass  filter , F is fft transformation matrix, x is original signal, F*x is fft transformation of signal x, F*h_1 is fft transformation of filter h_1, y is convoluted signal of signal x with filter h_1. x can be written as: x=(F'./repmat(F*h_1,[N 1]))*F*theta, where theta is sparse coefficients. Thus, any signal can be represented as sparse combination of dictionary ((F'./repmat(F*h_1,[N 1]))*F). Any deconvolution or compressive sensing problem can be solved using fast sparse bayesian learning algorithm. 1) For deconvolution problem, h_2 low pass filter, y is convolution of signal x with filter h_2, so y=F'*(F*x.*F*h_2) (convolution equals dot product of fourier transformation). H_2 is the convolution matrix of filter h_2. hence y=H_2*x=H_2*(F'./repmat(F*h_1,[N 1]))*F*theta, theta has sparse prior. 2) for compressive sensing, y=phi*x=phi*(F'./repmat(F*h_1,[N 1]))*F*theta , theta has sparse prior. 85. set sensing matrix to zero in the place where corresponds to zero coefficients of sparse signal. The place is the high frequency place for common signal with fourier transformation. I have test it and prove that better performance can be gotten in terms of PSNR or less measurements. This is due to reduce the subspace needed to be searched. 86. Inspired by "learning with dynamic group sparsity" we know that many real sparse signal(sparse representation by basis) has clusters of nonzero coefficients. It means that normal sparse prior with diagonal co-variance matrix is not appropriate for real signal with clustered nonzero coefficients. We hope to use Markov random field to model the prior and develop new sparse bayesian learning method. Graph cut is appropriate to optimize the function and find the estimated sparse solution. 87. Following "fast marginal likelihood maximisation for sparse bayesian models", iteratively estimate reconstructed images and covariance and hyperparameters using adaptively added measurement. Derive new update formula for new added measurement. No need to do reconstruction from the scratch. 88. For group sparse compressive sensing, we propose model min |y-\Phi*x|+\lamda*|x|^1+\gamma*|D*H*x|^1, where H is low pass filter matrix, D is high pass derivative filter matrix, this model encourage structure sparse signal to be reconstructed. Inference refer to paper "minimization of the l1 norm of derivative" 89. structure sparse compressive sensing and edge detection simultaniously. Gradient based compressive sensing with structure sparsity + edge detection 90. variational Bayesian blind deconvolution based on the idea in the paper : http://www.aliahmed.org/code.html. Use this idea to do simultaneous compressive sensing and deconvolution. 91. Sparse template and graph cut based nuclei segmentation for 2D and 3D. two papers. refer to the paper "sparse template based variational image segmentation". First use predefined dictionary. Then learn the binary dictionary using inputting images. refer to papers "Learning Non-Negative Sparse Image codes by convex programming","Binary Matrix Factorization with applications". We can do sparse shape decomposition for binary images and gray scale images. 92. explore the structure of block circulant matrix with circulant block to do simultaneous compressive sensing and deconvolution, refer paper "Learning circulant sensing kernel"  ftp://ftp.math.ucla.edu/pub/camreport/cam12-49.pdf 93. propose binary NMF method based on the idea in the paper "Sparse template -based variational Image segmentation" 94. A new method for matrix inverse: given nonsingular matrix X, compute its inverse Y: Model: min |I-XY|*, where * is nuclear norm or other norm. this is a convex minimization problem. Hopefully solved. 95. A new Sparse Bayesian video compressive sensing method: Model: min ||Y-\Phi*(AB^t+E)||^2, where A,B are low rank and column sparse matrix, E is elementwise sparse.    these priors found in the paper "Sparse Bayesian Methods for Low-Rank Matrix Estimation" source code available. 96. Propose new denoising model based on Rudin-Osher-Fatemi(ROF) model: min (TV(U)+\lamda*(U+E-Y)+\omega*||E||^1), where U,E variable, Y given (i.e. image), E elementwise sparse (representing high noise or small objects), U representing smooth objects/signals (low Total variation value).  we can also formulate this problem with variational method or Bayesian method. quite interesting. refer papers " Algorithm for finding global minimizers of Image segmentation and denoising models" and "Sparse Bayesian Methods for Low-Rank Matrix Estimation" source code available. TIP 97. Propose new graph cut model: min \lamda*R(L)+B(L)+\omega*TV(L), where L labeling variable, R(L) and B(L) representing region and boundary terms, TV(L) penalize different labels among neighboring pixels. New term TV(L) representing label cost making L smooth. simultaneously,TV(L) make objects circular since circular object has smallest perimeter compared to non-circular objects with the same area.  similar to variational method to find smooth L. Refer to my paper " Fruit Fly eye nuclei segmentation based on graph cut and convex shape prior using fluorescent micropscopic images". 98. ROF denoising model:  min(wrs U) TV(U)+\lamda*||U-Y||^2 , or equivalent geometrical model: min(wrs U) Per(U)+\lamda*|U \xor Y|, where Y given binary region, U binary region, Per(U) perimeter, \xor symmetric difference between two sets, |.| Lebesgue measure. this problem can be solved by Graph Cut algorithm. 99. Propose TV based L1 reconstruction error compressive sensing using Bayesian framework. reference paper deran's paper "variational bayesian blind deconvolution using a total variational prior". 100. Image segmentation can be done using L1 minimization problem: Consider observed image I as observation measurement, using all kinds of regularization  (Bayesian sparse regularization/ determinated regularization). refer to paper "image smoothing via L0 gradient minimization","Bayesian gradient based compressive sensing" from derin. 101. sparse dictionary constrained transformation invariant sparse coding (motivated by paper "Transformation invariant sparse coding": in that paper, the learned dictionary is highly dependent on the value of sparse regularization parameter \lamda. When the regularization strength is weak or too strong, the dictionary solution is not sparse. To avoid this case, we proposed to constrain the dictionary solution with sparsity which can be imposed by any image sparsity measure, such as sparse gradient (from derian paper). 102. Binary TISC (transformation invariant sparse coding) mode: motivated by paper(P_TISC) "Transformation invariant sparse coding" which proposed TISC model for general image and paper (P_BMF)"Binary matrix factorization with applications" which proposed binary NMF. Our idea is to constrain TISC model with binary constraints by penalty method or thresholding method proposed in the paper P_BMF. 103. graph cut and convexity -concavity based splitting method: 3D graph cut for foreground segmentation. Convexity -concavity based splitting in xy plane and xz plane to split 3D  contacting nuclei. 104. Transformation invariant image segmentation: motivated by the paper "Sparse Templated-based variational image segmentation" which proposed a method to segment foreground with sparse template but the number of templates is potentially huge comparable to the number of image pixels. We propose to using convolution to reduce the number of all possible templates to solve template translation problem the idea as the paper "Transformation invariant sparse coding". we have implemented and tested the translation invariant image segmentation. (Pattern Recognition or others). In the future, we will extend the method to rotation invariant segmentation. It may be possible to learn the templates(dictionary) using the same optimization method. 105. propose greedy optimization algorithm to optimize the objective function in the paper "Sparse Templated-based variational image segmentation" similar to greedy compressive sensing method referring to greedy compressive sensing paper. It is pretty important to learn the binary templates(dictionary) using optimization method. we can refer to the relevant papers on dictionary learning in non-negative matrix factorization (NMF multiplicative update rule) or compressive sensing (alternative method) fields. 106. propose double sparsity in sparse coding and dictionary in NMF or TISC or Tensor Decomposition model. Using L1 norm to impose sparsity constraint on coding is easy and direct. But it is not easy to put sparsity constraint on dictionary. We propose the gradient sparsity constraint on dictionary. That means we assume the gradient of dictionary is sparse. The assumption is reasonable for natural images. refer to the paper " millimeter wave compressive imaging". 107. Hadamard matrix based compressive sensing without mask: inspired by the work "single exposure super-resolution compressive imaging by double phase encoding", we propose to generate a hadamard matrix as a random phase mask to modulate the Fourier output of signal. That is y=Fx.*Fm. where F,x,m are Fourier matrix,input image, hadamard matrix, respectively. In this way, we only have to generate one Hadamard matrix without needs to generate multiple masks and move masks. But we need to move detector in imaging plane. 108. Model based millimeter wave imaging compressive sensing: 1. using bayesian compressive sensing framework with the prior probability of derivative. 2. using thresholding solver to find first k basis. 3 using found k basis as node to construct graph, data term is the reduced error after project observed signal onto corresponding node (basis), smooth term is related to neighboring relationship which can be modeled by Bernouli distribution. 3. using graph cut 109. Dictionary Based sparse segmentation:1) learn different dictionaries for different object classes;2)nonoverlapping patch sparse representation;3) using sparse coefficients associating dictionary to segment patch. 110. Graph Cut Based compressive sensing: 1)Generate structured sparse signal to simulate sparse structured signal;2)Get observation using sensing matrix y=phi*x;3)using each entry in sparse signal x as node to construct graph G(V,E); t weights are the similarity  relationship between y and phi_i (column of sensing matrix phi); n weights are the potts model for zero label and non zeros label (different labels with neighboring pixels have large weights, encouraging structured signal.). label cost is larger with nonzero label in order to impose sparse constraint.3) using graph cut to solve compressive sensing reconstruction problem. 111. Fingerprint singular point detection: 1) detect singular points with low accuracy of position using our model based sliding filter response singular point detection method. Due to the low accuracy of orientation computation, there is relatively large position displacement between true singular points and detected singular points. 2) using ridge symmetry to locate singular points accurately. Draw circle with detected singular point as center. There are multiple intersecting points on the circle with ridges. Find one ridge inside the circle with two ending points on circle. Between the two ending points over the arc, there have to be multiple intersecting points. Beginning with the next ending point, trace another ridge which is as long as possible. Until no ending points left. The last traced ridge has the most curved point or ending point which is the singular point. 3)using convex hull to correct the position of singular points. Pattern Recognition, IEEE Signal Processing Letter.   111b. Vote for correction of singular point positions from convex hull geometry: 1) using our AMF features (in our CVPR 2014 paper) to detect candidate singular points; 2) to be robust to noise, Our CVPR2014 paper uses large Gaussian smooth to smooth orientation field. This large smooth highly smooths high curved areas (core area) and produces large displacement of core point position from true positions. for each core point, draw a circle with the core point as center; 3. begin with each ridge ri in the circle with ending points inside the circle or on the circle, compute its convex hull and detect new ridges intersecting with the convex hull, for each new ridge, compute its convex hull and detect new ridges intersecting with the convex hull. repeat the process with new ridges until no new ridges are added into convex hull. the vote for all ridges are initialized to zero. for ridge ri, votes for all new rides intersecting with convex hulls during the above process are increased by one. 4. repeat the process for all ridges inside the initial circle and increase the votes by one for all ridges intersecting with a sequence of convex hulls. 5. the ridge r with majority of votes are the ridge contain core point. detect the point with most curvature or top most ending point on the ridge r as core point. TFIS, Pattern Recognition 111c. Core correction for large displacement: 1) if orientation field is severely smoothed to resist noise, core points in our AMF method(our CVPR2014 paper) are far away from true position; 2) we can repeat the process "Vote for correction of core points from convex hull geometry" in above 111b until the process convergence. The final detected core points are considered as true core points. We show that the process must convergence at the top most inner ridge (core). TFIS, Pattern Recognition. 112. Estimate orientation of singular points based on zero-pole model: 1. using our convergence index filter based method to detect singular points;2. compute the background orientation based on standard pattern around single singular point. 113. image retrieval: 1. extract features (class of superpixels) refereing  to paper " Class Segmentation and Object Localization with Superpixel Neighborhoods";2. get the normalized histogram of features in one image and second order histogram of features (neighboring feature pairs);3. compute the distance between the histogram of querying image and the ones of database images. we are done. 114. graph cut based edge detection:1. compute gradient image of gray scale image;2. construct graph based on the orientation of gradient vectors:more penalty along the orientation of gradient vectors and less penalty along the direction orthogonal to the gradient direction. TIP. 115. scene understanding: 1. extract features (class of superpixels) refereing  to paper " Class Segmentation and Object Localization with Superpixel Neighborhoods"; 2. with labels associated with each superpixel, retrieve the images with most number of the same labels, and construct neighboring graph(weight=1)  reflecting the neighboring relationship. compare the similarity between two graphs with nodes labeled with different labels by subgraph isomorphism refering to paper " computing subgraph isomorphic queries using structural unification and minimum graph structures"; Note that node category (labels) should match between the two graphs when computing the similarity between two graph. By the way of comparing two graphs, the transformation between two image has no effect on the result. It is appealing. TIP 116. graph based fingerprint matching: 1. extract minutiae from fingerprint image;2. construct neighboring graphs by checking the neighboring relationship between ridges where minutiae are lying. 3. then compare two graphs by subgraph isomorphism refereing to paper " computing subgraph isomorphic queries using structural unification and minimum graph structures". 117. Transformation invariant dictionary based compressive sensing: 1. learn a transformation invariant dictionary (set of feature vectors) referring to the paper " transformation invariant sparse coding" 2. test compressive sensing with this dictionary. 118. binary matrix sparse factorization: 1. using the binary matrix factorization method in paper "Binary matrix factorization with application" with sparsity constraint. 2. replace normal matrix product with boolean matrix product.3. using convolution to make it translation invariant. 119. shape learning based sparse template segmentation: 1. learn shape dictionary by statistical shape learning method;2. use graph cut to obtain foreground binary image;3. sparse template based segmentation. TIP 120. sparse and low rank matrix recovery using convex programming: 1. blind deconvolution as low rank matrix recovery using convex programming, refer to the paper "Blind Deconvolution using convex programming"; 2 we put sparse constraint on the w and m, we have x=Cm, w,m\in R^L. C\in R^(L*L). Thus we extend w, m to L dimension space by sparsity constraint. w is sparse rather than its representative coefficients in low dimension subspace R^K in paper  "Blind Deconvolution using convex programming". 3. our optimization problem: min ||X||*+||X||_1, s.t. y^=A(X), where y^ is the FFT of y, A is linear operator in the paper "Blind Deconvolution using convex programming". TIP, IJCV special issue-sparse coding. looks like RPCA :  http://perception.csl.illinois.edu/matrix-rank/introduction.html; or optimize  min ||H||_1, max ||M||_2 , s.t. y^=A(HM*). the high frequency components are increased by maximizing ||M||_2; filter should be sparse by min ||H||_1. 121. Sparse convex blind deconvolution. We reformulate the problem min ||X||*, s.t. y=A(X) (min ||W||2+||H||2, s.t. y=A(WH*)) in  the paper "Blind Deconvolution using convex programming" as min||w||1+||h||1, s.t. y=A(wh*), 1. using augmented Lagrangian method (multiplier method) as in "Blind Deconvolution using convex programming" to solve the optimization problem to get W and H, then shrinkage the W and H.  Investigate the relationship between these two models by experiments.  questions: do they still find the global solution? TIP;  2. Optimize the problem with NMF method with sparsity constraint in the paper "Non -negative Matrix Factorization with Sparseness constraints" (source codes available). This method has non -negative constraints, TIP; 3. Min ||W||2+||H||2, s.t. y=A(WH*), 3.1) using Sparse Bayesian Methods in the paper "Sparse Bayesian Methods for Low Rank Matrix Estimation" with the change-\gamma_i different for all components of W and H. TIP. 3.2). The Gaussian prior p(A|gamma) and p(B|gamma) in the paper "Sparse Bayesian Methods for Low Rank Matrix Estimation" can be replaced with Laplace distribution or scale Laplace mixture distribution refer to  the paper "Bayesian Group-Sparse Modeling and Variational Inference(Source codes available)" and "Group Sparse Coding with a Laplacian Scale Mixture"(Reweighted iterative method). TIP. 122. Fast Iterative Shrinkage/Thresholding Algorithm: 1. min f(x)+||x||1, change the model as min f(x)+||x||2 (smooth regularization), optimize until convergence and get x0 as the initial point of the following model min f(x)+||x||1, we hope the convergence is faster compared to normal method.  Refer to the paper "A Fast Iterative Shrinkage-Thresholding Algorithm for Linear Inverse Problems". TSP. 123. sequential on line compressive sensing with automatic stoping rule:1)stachastical gradient descent method to learning x_(n+1)=x_n+\yita*(y_n-\phi*x_n)*x_n-\yita*\lamda*x_n;2) if the change rate is too small. then stop. AO. 124. content sparse based segmentation and scene understanding model.1. minimize ||I-\sum \dotprod[conv(D_i, E_i) . W_i]||_2+\sum ||E_i||_1+\sum||E_i||_1, where I is image, D_i are atom in dictionary, E_i are sparse coefficient images, W_i is sparse binary images. 2. dictionary can be learned by the transformation invariant sparse coding method " transformation invariant sparse coding". In the future, we plan to learn the dictionary from our above model. 3. optimization by alternating gradient descent method between E_i and W_i as the same as  " transformation invariant sparse coding" considering W_i is continuous real number value. we can also consider W_i is binary values and be optimized by greedy method. first initialize all W_i with one, then greedily optimize each entry in W_i by check the costs corresponding to when the entry is 0 or 1. 4. This is a universal and wide application model in segmentation and scene understanding incorporating global and structural information. It promote the idea that using the number of objects to interpret scene as less as possible. TIP, IJCV,PAMI 125. statistical features of regions and its application to scene categorization. 1. use superpixel to segment images into regions. Or use sparse gradient based segmentation method to divide images into regions; 2. statistical histograms on region area, region shape,region geometrical features as feature vectors (reference to matlab region shape properties such as ;3. train a classifier to predict scene categories. such as logistical regression or SVM.  code and paper reference to " Beyond Bags of Features: Spatial Pyramid Matching for Recognizing Natural Scene Categories" . This method is assumed to outperform spatial pyramid matching. CVPR, ICCV. 126. Logistic regression and graph cut for nuclei segmentation:1. use superpixel or sparse gradient method to divide image into regions, using region features and logistic regression to learn the data term in graph cut using training data; 2. with this data term using graph cut to segment nuclei. refer to " Finding Things: Image Parsing with Regions and Per-Exemplar Detectors". Pattern Recognition, Pattern Recognition Letter, Journal of microscopy. 130. fingerprint singular point detection based on the combination of Poincare Index and circular shape analysis:1. get Poincare index for fingerprint  images;2. detect circular regions which are considered as singular points with assumption that true singular point region has a perfect circle and noise area doesn't have a perfect circle. Pattern Recognition,Pattern Recognition Letter. 130. Recursive Point Pattern Matching in Fingerprint recognition and scene understanding:1. extract interesting points as point pattern for each image;2. construct kd tree for fast neighbor searching;3. start with local 2 nearest neighborhood as a sub-structure and match each sub-structure pairs between template and querying image by computing the distance difference of lines connecting feature points and relative angle difference between straight lines formed by connecting points; give a score for the sub-structure matching; 3. for each matching sub-structure, delete center point, find the 5 nearest neighborhood for each remaining points in the sub-structure from the other points without consideration previously; In the 5 nearest neighborhood, find the best 3-point sub-structure matching pair, scoring the best matching pair by weight which indicate what nearest neighborhood  at least contains the best substructure. (for example, for nearest 2, 3, 4,5, the associated weight becomes smaller indicate the degree of noise). 4. repeat the step 3 until all points are visited or the matching score is below some threshold. PAMI, IJCV. 131. convolution sparse representation neural network:1. learning sparse transformation invariant dictionary by convolution operation; " transformation invariant sparse coding" 2. get sparse codes for each image based on learned dictionary ; 3. the sparse coding step can be considered as the first layer of NN which is equivalent to the several layers of Deep Neural network functioning as convolution and max pooling;4. the last layer is the fully connected layer( linear /softmax). In this way, we propose 3 layer shallow NN compared to DNN . hopefully, they have the same performance, but the training is easy. PAMI,IJCV,NIPS,CVPR. 132. double stachastic process based bayesian compressed sensing: 1. using double stachastic process as the prior distribution of the sparse coefficients of images under 3D complex wavelet transformation; http://post.queensu.ca/~gazor/acsdocfinal.pdf 2. bayesian based method to esitmate sparse signal.  SP,SPL,Applied Optics, TIP. 133. sparse coefficient distribution based compressive sensing matrix create: 1. get the distribution of the sparse coefficient of natural images by histogram statisics.denoted by p(s_i) for ith coefficient. it is bernouli distribution of zero and nonzero.2. use the bernouli distribution as the distribution of each entry in the compressive sensing matrix and sample the distribution to create the desired sensing matrix. 3. This created matrix has better performance than general gaussian matrix. For another paper, this method can be combined with using the double stachastic process for sparse coefficient distribution prior in above 132. SP,SPL,Applied optics,TIP. 134. Sparse Bayesian based compressive sensing:1) observation model: N(y|\phi *w,\beta);2. signal model: N(w|,0,\alpha_i); 3. hyperprior \gamma(\alpha_i|a_i,b_i); 4. hyperhyperprior: \gamma(a_i|\lammda),\gamma(b_i|\lamda);5. reference paper " Bayesian Compressive Sensing Using Laplace Priors"  http://www.dbabacan.info/papers/babacan_CS.pdf 135. Three -Stage model for Sparse Bayesian Learning and the Relevance Vector Machine: 1) 1) observation model: N(y|\phi *w,\beta);2. signal model: N(w|,0,\alpha_i); 3. hyperprior \gamma(\alpha_i|a_i,b_i); 4. hyperhyperprior: \gamma(a_i|\lammda),\gamma(b_i|\lamda);5. reference paper " Bayesian Compressive Sensing Using Laplace Priors" and " Sparse Bayesian Learning and the Relevance Vector Machine ", TSP,PAMI,TIP. 136. Three -Stage model for fast marginal likelihood maximization for sparse bayesian models:  1) observation model: N(y|\phi *w,\beta);2. signal model: N(w|,0,\alpha_i); 3. hyperprior \gamma(\alpha_i|a_i,b_i); 4. hyperhyperprior: \gamma(a_i|\lammda),\gamma(b_i|\lamda);5. reference paper " Bayesian Compressive Sensing Using Laplace Priors" and " fast marginal likelihood maximization for bayesian models ", TSP,PAMI,TIP. 137. L1 norm of gradient based compressive sensing: 1) min ||y-\phi*x||^2+\lambda*||x_i-x_i+1||_1; 2)optimization based on FISTA algorithm. TIP,TSP. refer " A new algorithm for compressive sensing Based on Total variation Norm" 138. fast TV based Bayesian Compressive sensing: refer to paper : "variational Bayesian Blind Deconvolution Using a total Variation Prior" and "Bayesian compressive sensing using laplace priors", TSP. 139. nonlocal prior based Bayesian compressive sensing and debluring: refer to : https://www.ceremade.dauphine.fr/~cohen/mypapers/GabrielPeyreNLRECCV08.pdf 140. statistical frequency coefficient model for compressive sensing and sensing matrix design. 1. use the frequency magnitude model in http://research.microsoft.com/pubs/138795/tashev_acero_statistical_modeling_of_the_speech_signal.pdf  as the sparse prior model to do bayesian compressive sensing and compare its performance with the bayeisan method with parameter estimation. 2. using the scale parameter values (\chi) in paper   http://research.microsoft.com/pubs/138795/tashev_acero_statistical_modeling_of_the_speech_signal.pdf to generate the success probability p  proportional to \chi of bernoulli distribution to design sensing matrix. Signal Processing. TIP. 141.  learn double sparsity convolutional sparse coding for segmentation and object recognition: 1. learn a sparse dictionary by convolutional operation. refer to paper " Double Sparsity: Learning Sparse Dictionaries for Sparse Signal Approximation" and paper " Fast Convolutional Sparse Coding" ; 2. obtain sparse coding by learned dictionary for object recognition or segmentation.. TIP/PAMI 142. Sift feature and dictionary learning for object recognition and image classification/retrieve: 1. extract sift feature ; 2. learn dictionaries for each class by RICA (use sift position and histogram) and build sparse histogram(reference to our recent paper "Joint Dictionary learning for activity recognition"); 3. use svm classify or use histogram to finish retrieving. TMM/TIP. 143. fingerprint reference orientation definition and estimation: 1. use our AMI feature to detect singular points(CVPRW2014); 2. compute reference orientation o_\infinity according to equation 2*o(z)=2*o_\infinity+arg(z-z_c) for core point and  equation 2*o(z)=2*o_\infinity-arg(z-z_d) for delta point. equation (5) in our paper. IEEE Trans. Forensic and information security. 144. Down syndrome detection based on constrained NMF: 1. learn dictionary based on NMF using training data. 2. search landmark points based on local model match and global NMF model match. 3. extract features after obtaining landmark points. refer to the paper  published in Medical Image Analysis 2014 " Digital facial dysmorphology for genetic screening : Hierarchical constrained local model using ICA MedIA2014Dysmorphology." 145. Down syndrome detection based on pictorial model: 1. learning a pictorial model  from manual landmark points for normal people and down syndrome people; 2. match pictorial model to test image and classify based on posterior probability.  3. without the need to search landmark points; model match can finish landmark search and classification at the same time. It is global optimal other than local optimal. refer to the paper "pictorial structure for object recognition" source code http://www.d2.mpi-inf.mpg.de/andriluka_cvpr09 and http://people.csail.mit.edu/fergus/iccv2005/partsstructure.html 146. extend bag of words using transformation invariant sparse coding NMF: 1. learning dictionary based on transformation invariant NMF (compensate for object  transformation in image and invariant to object rotation and displacement, different from conventional dictionary which is sensitive to transformation and position); 2. compute sparse codes, get histogram of word frequency and statistics (center moments) for each word as feature vectors. do feature selection to select features (hopefully, choose the features corresponding to distinctive words). 3. train classifier using selected features. Advantage:dictionary is invariant to object rotation and displacement similar to visual eye.  part-based local words has strong discriminating power. feature selection process may select those distinctive words out as features. this is big contribution. refer to paper "Transformation invariant sparse coding" and "transformation invariant representation and NMF. 147. transformation invariant sparse coding for scene classification/object recognition: 1) using transformation invariant NMF to learn a dictionary for each class  (atoms are different objects or different parts owning high distinctive power); this class specific dictionary characterize the contents of object/scene; advantage is the dictionary is invariant to image transformation (displacement, scale, rotation); 2) for each training image, compute its sparse coding c, get the histogram h of c with each bin corresponding to one class. h  can be used as descriptive feature vectors to train a classifier.  3) for each test image; compute its h and classify it. Novelty: transformation invariant recognition consistent with human vision. refer to papers "transformation invariant sparse coding" and "robust face recognition via sparse representation".  PAMI 148. Shape retrieval by transformation invariant sparse coding for binary image:1. put basic shapes in dictionary; 2) for test shape, compute sparse coding and classify test shape based on sparse coding. 3) our algorithm also including displacement invariant. refer to TIP2013 paper " in plane rotation and scale invariant clustering using dictionary";   TIP 149. Improve accuracy of FISTA by locally replacing L1 with square of L2 around zero point: 1. ||x||_1 replaced by ||x||_2^2 around zero point; 2. compute gradient G and compute x_k=x_(k-1)-t*G ,  x_k=x_k +trend*(x_k-x_(k-1) and x_k=threshold(x_k, \theta)  based on gradient descent with box constrained to avoid vibration around zero point. 3. we get more accurate results than FISTA while keeping the same computation complexity. refer to paper " a fast iterative shrinkage thresholding algorithm for linear inverse problems",  TSP Other methods source code  http://www.eecs.berkeley.edu/~yang/software/l1benchmark/ 150. Improve accuracy of L1 optimization by hierarchical combination of convex norm L1 with non-convex norm Lp: 1. Use FISTA get non accurate results x with L1 norm; 2. Use FISTA get accurate results with Lp norm (0<p<1) by initializing x0=x which come from step 1; 3. trade off between accuracy and speed. Lp (p>=1) has fast speed but results are less accurate while Lp(p<1) has more accurate results but speed is slow. That's why we combine them appropiately. TSP Other methods source code  http://www.eecs.berkeley.edu/~yang/software/l1benchmark/ 151. Singular point search based on Zero-pole model: 1. choose center O of fingerprint image I as reference point and any point P in the image; compute the main argument  arg(OP) of complex vector \overarrow{OP} to get an argument map called as ArgO; 2. for any point Q in image; the argument map ArgQ wrt Q can be computed quickly as the translation of ArgO, i.e. ArgQ=(ArgO translated with vector Q-O); 3. therefore, Argument map only compute once for the whole image. ; 4. search operation is constrained within bad quality area to reduce the searching time and orientations in good quality area are used to satisfy Zero-Pole model since the orientation in bad quality area is not convincing;  5. The algorithm is parallelable and speedup  can be done with multicore or multi threads; 6. two papers can be produced by combining it with PoinCare Index or our AMI method in CVPRW2014 as initial candidate singular point detection. TFS and PAMI, TIP. 152. Extension of FISTA to general sparse coding model: 1. min ||Wx||_1 is equal to saddle point problem: min_{x} f_{\mu}(x)=min_{x} max_{u} <u,Wx>+\mu*p(u), where p(u) is a prox function such as l2 norm: p(u)=||u||_2^2; 2. The gradient \delta f_{\mu}(x) of saddle point problem can be computed as \delta f_{\mu}(x)=W^t*u^\star (x), where u^\star (x) is the maximum argument of max_{u} <u,Wx>+\mu*p(u). Therefore the non-smooth function is smoothed by saddle point problem and the gradient vector can be computed; 3. we plug the computed gradient of smooth function into FISTA by replacing the shrinking threshold operation with gradient descent with the computed gradient followed by threshold operation. Thus we extend FISTA to general sparse coding model such as ||Wx||_1, TV model. refer to the paper " NESTA: a fast and accurate first order method for sparse recovery". TSP,TIP. 153. Fingerprint orientation field enhancement based on low rank matrix and sparse representation: 1) compute initial orientation field Y using normal method (such as in our CVPR2014 paper.); 2) Y can be represented as Y=X+E+N. we solve the optimization problem: min ||X||*+\lambda*||E||_1+\gamma*||Y-X-E||_2, where ||X||* is nuclear norm of matrix. It can be solved by RPCA (refer to Yi Ma's paper). 3) Computed X is expected to be the enhanced orientation field estimation. TIP, PAMI, TFIS. 156. Adaptive orientation field smoothing based on singular area detection: 1) compute curvature map using method in our ICIP paper; 2) detect singular area based on curvature map by training some classifier (SVM); 3) For non-singular area, use large orientation smooth to resist noises such as scars, dryness. For singular area, we use small smooth to keep the high curvature area with small distortion as possible. In this way, the smoothed orientation field does not only resist noise, but also distort singular area as little as possible. IEEE SPL 157. Accurate method for Singular point detection: 1) use zero pole model to derive "Orientation Coherence Feature" with the analogy to our CVPR2014 paper in which we have a formula on the orientation difference between any two points z_1 and z_2  in fingerprint image. In that paper, we show that the orientation difference o(z_1)-o(z_2) is determined by all these straight lines which pass through z_1 and one of singular points or pass through z_2 and one of singular points. This demonstrate that orientation field is determined by singular points. The inverse is true that singular points are determined by orientation field. But it is hard to compute because it is highly nonlinear function. The simple way is to simplify the formula by constraining the sampling point pair z_1 and z_2. In CVPR2014, we propose a special sampling pattern and the sampling pattern can be elegantly used in convergence filter. 2) after compute the candidate singular points using the method in CVPR2014 paper, we borrow the idea that determine possible combinations of singular points in fingerprint image from the paper "A novel Algorithm for detecting singular points from fingerprint images" by computing Porn Care Index along the valid border around fingerprint image. Once determining the possible number of singular points, we can refine the positions of possible singular points by searching on the bad quality area and the small area around detected singular points from CVPR2014 paper. In this way singular point positions are not only refined, but also the missed singular points in bad quality area can be recovered. TFIS, THM,IEEESPL 158. Accurate Core Point detection based on geometrical shape analysis: 1) using our method in CVPR2014 to detection candidate singular points.; 2) remove small bridges in thinned images since they come from noise. Especially, in singular region, orientation smooth may produce many false bridges which affect our later geometrical shape analysis. It may make vote based convex hull analysis fail; 3) correct core point position by geometrical shape analysis. For each core point C, draw a circle around it, in this circular region,  find ridges with maximum number of votes from convex hull cover; find the intersecting points between circle and maximum vote ridges. compute the mass center m of the intersecting points. compute the intersecting points p closest to mass center m. the closest intersecting point p as the trace starting point. Assume that component L contains starting point p , remove branch points in component L results in several connected components from L. Find component r containing  starting point p. Detect ending point list epl of ridge r . If epl contains at least one ending point P_e which is not intersecting point, then P_e is corrected core point. otherwise find the point P_c with maximum curvature which is corrected core point; 4)repeat step three until correct core point converges, the final corrected core point can be found. In this way, far away core point can be corrected. Also, false core point may be removed if we require the core point must stay in area which is curved to some degree which vote difference should be large enough. TIP,TFIS,PR 159. Symmetry based fingerprint image enhancement: 1) use complex analytic function \sqrt(z) as coordinate transformation to transform fingerprint image f(x,y) to coordinate plane (\delta,\etta); 2) transformed image f(\delta \etta)=f(x,y) contains parallel lines at least in the neighborhood of core and delta points in fingerprint images; 3) we can enhance the parallel lines easily by gabor filter with one constant direction \omega which can be detected by Hough line detection method; 4) after enhance the parallel lines, then transform back to (x,y) plane to get enhanced fingerprint image; refer to paper " Pattern recognition in images by symmetries and coordinate transformations ". TIP,TSP,TFIS,PR 160. Singular area enhancement by symmetry: 1) use method in above 159 to enhance singular area. SPL 162. Poincare index and Source region based singular point detection: 1) compute poincare index image P without thresholding the accumulative direction change; 2) according to our CVPR2014 paper, singular area would appear as source region. take Mexican Hat wavelet transformation M of P; 3) assume that local extrema contain singular point. Detect local extrema as candidate singular points and thresholding them based on PI value.  As our CVPR2014 paper, shape is taken into account.   PR,TFIS,SPL 163. Naive Bayesian and sparse dictionary based activity recognition: 1, using RICA learn sparse dictionary as in our PLOS one 2014 paper; 2, Naive Bayesian Classifier is trained using training video; 3, test Naive Bayesian Classifier. PR, PRL 164. Kmeans deep learning: 1. first layer: input; 2. second layer: cluster input x by kmeans, centers as weight matrix; 3. third layer: cluster output h^1 from second layer by kmeans, centers as weight matrix. repeat the process to construct many layers; 4. output layer: fully connected classifier. PAMI 165. Neural network based compressive sensing learning: 1. suppose observation vector t=\phi*s, where \phi compressive sensing matrix , s sparse signal. purpose is to find s given t and \phi; 2. assume s=f(w,t)--a nonlinear function of t with respect to parameters w; assume f(w,t) is a neural network function form; 3. we have normal distribution N(t|\phi*f(w,t),\delta*I) and prior normal distribution N(s,\delta*I). this is two stage neural network with stage f(w,t) and stage  \phi*s (s=f(w,t)). the latter stage is known and fixed. The first stage needs training by BP method; 4. normally, we have only one sample t. using one sample t, we train the network to find w, then find solution s=f(w,t); 5. the feature is new method for compressive sensing; and for new observation, we may quickly get sparse solution by s=f(w,t) without computationally intensive process such as sparse recovery . 6. we can train the network with more samples and thus make the prediction more accurate . PAMI. 166. Auto-encoder based compressive sensing learning: 1. collect a set of sparse signal s with special prior distribution; 2. case 1: compressive sensing matrix \phi given, learn an auto-encoder to do CS reconstruction for some kind of sparse signal; case 2: CS matrix \phi unknown, learn an auto-encoder to do observation and sparse decoding; 3. advantage: fast and adaptive observation. PAMI 167. transformation invariant sparse coding based deep neural network: 1. learn a dictionary D1 by transformation invariant sparse coding; 2. using learned dictionary D1 to obtain sparse coefficients S !.  From S1 , repeat step 1 to get the second dictionary D2; 3. repeat step 1 until we get enough number of hierarchical  dictionaries, each dictionary is considered as one  layer. Thus we get deep NN and hierarchical representation. sparse coefficients may be binarized since the magnitude does not matter. Therefore, all dictionaries may be binary except the first D1;  4. the last layer is classifier. PAMI 168. Bayesian based transformation invariant sparse representation dictionary learning in big images: 1, model  X=\sum_i(D_i*S_i)+N, p(X|S)~gaussian, P(S_i)~sparse distribution, P(D_i)~direchilette distribution; 2. given X, we calculate the derivative of p(X) with respect to S and D, refer to paper " Optimization methods for convolutional sparse coding"  " Sparse coding and ICA" for derivative computation of convolution operation; 3.  we may use full bayesian model or hierarchical Bayesian model. PAMI 169. Sparse coding based temporal pyramid matching using body joint feature for human activity recognition: 1. reference our Plosone paper and IEEE THMS paper and paper " Linear spatial pyramid matching using sparse coding for image classification" . 2. main idea: densely sample joint volume; using sparse coding to learn dictionary; temporal pooling of sparse codes; concatenate pyramid histogram; training SVM by histogram; classify test video; This paper for THMS revision for correspondence; THMS 170. blind deconvolution based on convolution sparse coding: reference paper " Optimization methods for convolutional sparse coding" . ADMM method for filter and image computation. 1) image gradient sparse prior and filter with small support and summation to 1; TIP. 171. orthogonal sparse dictionary learning for real time human activity recognition using body joints: 1) learn an orthogonal dictionary based on sparse coding, refer to " fast sparsity-based orthogonal dictionary learning for image restoration"; 2) build bag of words histogram based on sparse coding (slow) by thresholding coefficients; 3)using svm to obtain classification results; 4) build bag of words histogram based on orthogonal projection (fast) by thresholding projection  coefficients and use SVM to obtain classification results; 5) compare the results from these two methods (hope the performance is comparable) to demonstrate the idea that both methods have the same accuracy performance but different speed. Therefore projection rather than sparse coding is preferable. 172. compare two sparse coding based orthogonal dictionary learning methods for joint based human activity recognition: SPCA, RICA and " fast sparsity-based orthogonal dictionary learning for image restoration". 1) obtain dictionaries and using orthogonal projection to get projection coefficients to build histograms based on thresholding; 2) compare results from three methods and hope to demonstrate the idea that they have the same performance and orthogonal dictionary learning is useful for fast recognition (due to fast coding by orthogonal projection). Orthogonal dictionary makes sparse coding problem equivalent to orthogonal projection problem. THMS 173. Sparse coding based dictionary learning (may not orthogonal) and non-orthogonal projection for fast human activity recognition using joints: 1)learning dictionary (may not orthogonal) based on sparse coding; 2) build bag of words histogram by sparse coding and thresholding coefficients; 3)build bag of works histogram by projection and thresholding coefficients; 4) compare the results from these two methods to demonstrate the idea that both methods have comparable recognition accuracy but different speed due to slow sparse coding and  fast projection operations. Therefore we prove the idea that for recognition accuracy, fast projection operation is as good as slow sparse coding. why? the idea of projection and thresholding is similar to the idea under greedy orthogonal matching persuit (OMP): OMP performs multiple times of  projections and each time choose one coefficient, however, our method performs one time of projection and choose all coefficients. We demonstrate the idea that coding coefficients may be different (we have to show how different they are by compute the difference between two kinds of coefficients) but the accuracy performance may be comparable. In a nutshell, we demonstrate that fast projection operation is comparable with slow sparse coding operation for general dictionary. Therefore, fast projection operation is ubiquitous for general dictionary. refer recent (2014) dictionary learning method to  http://www.ux.uis.no/~karlsk/dle/ and   http://www.math.nus.edu.sg/~matjh/research/research.htm THMS 174. If fast projection operation for general dictionary doesn't work (i.e., above 173 is wrong), then at least it works for orthogonal dictionary (above 172 is right). Therefore, orthogonal dictionary learning is preferable. We propose ADMM method for orthogonal dictionary learning, refer the formula to paper " Optimization methods for convolutional sparse coding" for dictionary update with extra orthogonal constraint W'W=I. TIP 175. Blind deconvolution based SAR image formulation: 1) in range direction, blind deconvolution is performed based on gradient sparse along range direction. refer to the formula in paper " Optimization methods for convolutional sparse coding"; 2) in azimuth direction, blind deconvolution is performed based on gradient sparse along range direction. refer to the formula in paper " Optimization methods for convolutional sparse coding"; 3) we get the final SAR image. TIP 176. learning robust orthonormal subspace for joint based human activity recognition : 1. learn an orthonormal subspace from joint; reference paper :" Robust Orthonormal Subspace Learning: Efficient Recovery of Corrupted Low-rank Matrices"; 2. build histogram and svm training. PR, PRL 1. learn a orthogonal dictionary to speed up sparse coding problem in classification; 2. explore the relationship between sparse coding  and first N maximum projection coefficients (latter is pretty fast) to see if  both of them achieve the same classification accuracy( for instance, joint based human activity recognition); refer to nice paper : " fast sparsity-based orthogonal dictionary learning for image restoration"; " Incoherent dictionary learning for sparse representation based image denoising" THMS 177. Filter based image whitening and dimension reduction simultaneously: 1. compute whitening filter f from average power spectrum of image patches, refer to paper " Natural Image Statistics"; 2. as shown in " Natural Image Statistics", filter f  is low pass filter and the power spectrum of the filtered data drop off sharply at high frequency. Therefore, we propose to detect the frequency point f1 which power spectrum drop off sharply and down sample filtered data to achieve dimension reduction. SP   When I tested my joint based human activity recognition system using dictionary learning method, I accidentally found that much better performance can be obtained if we use dictionary from other data except the data you are using. Specifically, I split MSRAction3D into 3 subsets:AS1, AS2 and AS3. When I used dictionary D_1 learned from AS1 to represent AS2 data, I got much better performance. it is true that I got much better performance when I use D_2 from AS2 to represent AS1 and AS2 or use D_3 from AS3 to represent AS1 and AS2. on the other hand, using dictionary from the dataset ASi (training samples in ASi) to represent the  same kind of dataset ASi (testing samples in ASi), the performance is worse. This is ridiculous. But It is an important discovery. I am thinking that at least we don't need to learn dictionary for feature representation, random dictionary is enough. therefore we have number 178 idea as following: 178. random dictionary and feature representation: 1.  for dimension reduction and feature representation, we can use a random dictionary (words normalized to unit length) to represent data by projection (dot product); 2. build histogram of word frequency as new feature vectors. 3. test the method in human activity recognition databases (joint based for fast speed) ( We already have results); 4. test the method in other general object classification databases( for example, imagenet). PAMI 179. rotation and translation invariant 3D shape matching based on polarization and phase correlation. 1. Fourier transformations of two 3D shapes/volumes (Signed distance volume); 2. one of the Fourier transformation is another Fourier transformation with appropriate rotation and phase addition. 3. polarizing the two phase volumes; 4. The relationship between the two polarized phase volumes is 2D translation (for 2D image, 1D translation); 5. compute 2D correlation (normalized correlation due to phase addition) of these two polarized phase volumes to get the optimal translation parameters; These optimal phase translation is the rotation between the original 3D volumes/shapes; 6. rotate the phase volume using the computed rotation and subtract one rotated phase volume with another phase volume, get the average phase difference, the phase difference is the translation between the original volume. TIP 180. Deep learning for joint based human activity recognition: 1) as our experience, treat 3D joint coordinates as volumes; 2) train multi-scale deep neural network based on multi-pyramid; rescale all training videos to the same size, then train deep convolution neural network.; 3) test trained deep network on new video based on sliding windows. THMS, CVPR. 181. Joint trajectory based human activity recognition: 1) extract features from RGBD images based on joint trajectories. following and refer to the paper   "Action Recognition with improved trajectories". 182. Down Syndrome Detection based on sparse representation: 1) collect training images and normalize them based on the distance between eye corners; 2) for test image, use viola and jones face detector to detect face; 3) using qian zhao's method to search lankmarks; and normalize detected face by the distance between landmarks (For example, distance between eye corners); 4) find sparse representation of this normalized test face in the training images; the sparse coefficients show the identity of the test face; refer to paper " Robust Face Recognition via sparse representation". MICCAI2015. 183.Transformation invariant  Down Syndrome Detection based on sparse coding: the same as above 182, but with alignment step; the test face may have large misalignment and illumination variance, sequentially align test face with each subject (here for two down syndrome and healthy groups) by using the method in paper " Towards a practical face recognition system: robust alignment and illumination by sparse representation"   MICCA2015 184. Personal closest normal shape search for craniosynostosis surgery based on sparse coding: 1)compute signed distance images (SDI) from all registered shapes; 2) using normal  SDI as dictionary D, represent test shape t as: t=Ds+\lambda_1*|e|_1+\lambda_2*|s|_1, the sparse coefficient vector s with maximum coefficient entry indicates the normal closest shape. refer to RPCA method; IEEE TMI or TBME. 185. Deep dictionary learning for object recognition:1) learning one dictionary D1 from raw patches x based on sparse coding (for instance, RICA); 2)  learn one more dictionary D2 from projection coefficients x1=D1*x ; 3) repeat above process until get enough number of dictionaries. 3) deep dictionary system is built by concatenate all of these dictionaries to form many layers of dictionaries; 4) training data flow through all dictionaries by direct projection then get the output as features; 5) the last layer is supervised layer (for example, SVM); this deep dictionary system is unsupervised and is suitable for unlabeled data (big data); simple and fast. TPAMI,CVPR 186. deep learning by layer wise back-propogation algorithm: 1) train a 3 layer shallow network using normal BP method, keep error relatively large by set the number of hidden units relatively small; 2) get the output of hidden layer as new raw data and train another 3 layer shallow network; 3) repeat above process until the validation error is small enough. It is similar to unsupervised layer wise training. But we use labeled data at each step to get better performance for supervised learning. TPAMI, CVPR --------------------------------------------------------------------------------------------------------- Innovation 1. Image registration based on explicit function model: 1) use some explicit function to model data/image/video (such as interpolation reconstruct function from ADC, FFT basis function, polynomial function); 2) once obtaining the explicit function form which is a function of coordinate, we can plug transformation parameters(rotation \theta, displacement \deltaX,\deltaY)  into the function. 3)minimize the error: min e(\theta,\deltax,\deltay)=\int\int[f(x,y,\theta,\deltax,\deltay)-g(x,y)]^2 dxdy with respect to \theta,\deltax,\deltay. Thus we can get the transformation parameters. PAMI --------------------------------------------------------------------------------------------------------- project: 1. abnormal event recognition using visual event recognition in videos:  http://videolectures.net/cvpr2010_duan_verv/ and modeling mutual context of object and human pose :  http://videolectures.net/cvpr2010_fei_fei_mmco/ 2. fingerprint orientation field model; min \sum[arg(P(z))-arg(Q(z))-2*\theta], using gradient descent method. We develop rough gradient descent method: determine the sign of object function gradient by difference computation in two specific points,one of which is the initial known parameter and another is the perturbed parameters. The new gradient descending method is applicable in case of no object gradient available. 3. Create a new camera with binocular lens for the photogrammetry.  share one CCD sensor but taking two successive pictures in short time. Then compute two perspective centers using image processing method with collinearity constraints under image reference coordinate system. This camera only needs to estimate the 3d coordinates for the two perspective centers of two lens once. Later it can be used everywhere, anytime. Furthermore, it doesn't depend on the position and orientation of camera. Very convenient. refer to the paper"Toward a super imaging system" This is for new project or patent. Rapid paper: 1. graph cut+minimum path (experimental results ready), referring to "minimal path algorithms for the robust detection of linear features in gray images"; 2. shift+rotation+scale invariant sparse coding of image/volume (experimental results, formula ready), referring to "shift invariant sparse coding of image and music data". We can scale atoms. 3. shift invariant+3D sparse template based segmentation( simulated 3d nuclei results ready, algorithm can be changed to use "sparse binary matrix factorizaton" in existing paper). ----------------------------------------------------------------------------------------------------------- Business: 1. building a "fear conditioning test" web service for data/video process/analysis and provide intermediate results for research papers. refer to  http://www.jove.com/video/50871/contextual-cued-fear-conditioning-test-using-video-analyzing-system

转载于:https://www.cnblogs.com/stoneresearch/archive/2012/01/06/4336449.html