Research Interests:
- Human Activity Recognition
- Fingerprint Recognition
- Cell/Nuclei Segmentation
- Compressed Sensing
- Colon Cancer Detection
- Defect Detection in Industrial Product
Research Projects:
Human Activity Recognition (on going)
Problem: human activity recognition based on RGB-D video is drawing more and more attentions with recent rapid proliferation of low-cost Microsoft Kinect cameras. Human activity recognition is still a challenging problem since it won't work well to directly translate conventional 2D RGB image based methods to 3D RGB-D videos.
Solution: Recently we proposed a new human activity recognition method based on dictionary learning and sparse coding. We only use joint position information (output of Microsoft Kinect camera) so that the scale of data is largely reduced the speed of our algorithm is fast. Experimental results across three public benchmark data sets demonstrate that our algorithm outperforms state-of-the-art methods. Matlab Source Codes for our IEEE THMS paper:
JinQiSourceCodesForTHMS2014Paper.rar Once download, you get the file with name "JinQiSourceCodesForTHMS2014Paper.rar.pdf", please remove ".pdf" and unzip the file to get source codes. Please refer to the readme file inside to run the program for reproducing the results in our paper.
Fingerprint Recognition (on going)
Problem: Singular point (SP) detection is essentially important for fingerprint classification. However, the performance of the SP detection methods in the first singular point detection competition, SPD2010, indicate that the state of the art SP detection methods are far from satisfying the requirements in fingerprint classification.
Solution: A new feature, called "Angle Matching" (AM) feature, is proposed by exploring the Zero-Pole model of fingerprint orientation field. The framework of using our AM features in the Convergence Index Filter is proposed to efficiently detect singular points in fingerprint images. The experimental results of our algorithm over the SPD2010 benchmark get the 1st ranking among all competitors in SPD2010 according to the recommended metrics. Matlab Codes available:
https://skydrive.live.com/redir?resid=C4E19F7E1DAF7C1F!24167
Our paper: 1. Jin Qi , "A Complex Rational Polynomial Model of Vector Field for Singular Point Detection",
IEEE Transaction on Pattern Analysis and Machine Intelligence (
PAMI)(Under review). Matlab Source Codes:
https://onedrive.live.com/redir?resid=C4E19F7E1DAF7C1F!24247&authkey=!APCr7j6qKRdaYA4&ithint=file%2crar 2. Jin Qi and Suxing Liu, "A general approach for singular point extraction based on complex polynomial model",
2014 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2014 Workshop on Biometrics.(PDF) 3. Jin Qi and Yangsheng Wang, "A robust fingerprint matching method",
Pattern Recognition, vol.38, no.10, pp.1665-1671, Dec. 2005 (
PDF) 4.
Jin Qi, Zhongchao Shi, Xuying Zhao, and Yangsheng Wang,“Fingerprint Matching Using its Orientation Field and Minutiae,”
Pattern Recognition Letter, vol.26, no.15, pp.2424-2430, Oct. 2005 (
PDF)
Cell/Nuclei Segmentation (on going):
Problem: It is a challenging problem to segment dense cells/nuclei in fluorescent microscopical images since the objects are tightly touching each other.
Solution: The modified Graph Cut method is proposed to segment the cell/nuclei foreground. A convexity-concavity based method is provided to split cell/nuclei clumps. and then a sparse template-based segmentation method is proposed to split cell/nuclei clumps with better performance.
Our paper: 1.
Jin Qi, “Dense Nuclei Segmentation Based on Graph Cut and Convexity-Concavity Analysis”,
Journal of Microscopy, Vol. 253, Issue 1, pp. 42-53, Jan. 2014.(
PDF) 2. Jin Qi, Bao Wang, N. Pelaez, I. Rebay, R. Carthew, A. K. Katsaggelos, L. Amaral, “Drosophila Eye Nuclei Segmentation Based on Graph Cut and Convex Shape Prior”,
Proceedings of the 19th IEEE International Conference on Image Processing (ICIP 2013), Sept. 2013.(
Potential Award Recipient, Top 10% of all accepted manuscripts, Top 4.5% of all submission to ICIP2013)
Compressed sensing (on going):
Problem: compressed sensing is to solve the following minimization puzzle:
min |x|_1, s.t. y=ϕψx, where ϕ is a sensing matrix. a simultaneous compressed and deblur model is proposed as following:
min |x|_1, s.t. y=ϕhψx, where h is a convolution matrix acting as an blurring operator.
Solution: a Bayesian method is proposed to solve our proposed problem i.e. the minimization problem mentioned above. Our Matlab source codes of Matrix implementation of Haar Wavelet Transformation:
http://www.mathworks.com/matlabcentral/fileexchange/33625-haar-wavelet-transformation-matrix-implementation
Our paper: 1. Jin Qi, "Variational Bayesian Compressive Non-Blind Deconvolution Using a Total Variation Prior",
IEEE Transaction on Image Processing (under review) 2. Leonidas Spinoulas,
Jin Qi, Aggelos K. Katsaggelos, etc. “Optimizaed Compuressive Sampling for Passive Millimeter-wave Imaging”,
Applied Optics, Vol. 51, Issue 26, pp. 6335-6342, 2012.
Colon Cancer Detection (2010-2011):
problem: Automatic colon tumor detection method is emerging as an critical computer aided diagnosis technique with the development of Computerized Tomography. However, the detection of flat lesion on the colon surface is challenging the current existing methods.
solution: A so called "spinning tangent" method is proposed by Kenji in the University of Chicago. A kd-tree based c/c++ implement of the algorithm makes the program as 12 times fast as the original implementation without any optimization.
Defect Detection in Industrial Product (2007-2009):
Problem: Automatic detection of the visual defects, such as color deviation, cracks, point/ line defect, on the surface of compact disc is highly desired in CD manufacturing industry.
Solution: A system consisting of industrial camera and led light are set up over the manufacturing line to capture the images of compact discs. A software has been developed in c/c++ language to analyze the contents of the captured images in real time and tell if a defect is appearing on the surface of the compact disc or not. The CD with detected defect will be removed automatically from the product line by a robotic arm.