Image processing - feature extractor



Image FeatureExtraction

• Edges
  • Sobel, roberts
  • LoG
  • Canny
• Interest point
  • Harris
  • SIFT
• Descriptors
  • SIFT
  • SURF
  • HOG

 

LoG:  add Gaussian weights to local box to smooth,and run Laplacian derivative to calculate edges.

 

Harris cornerdetector: calculate intensity gradient in both x and y direction, and multiplyto generate lamda1 and lamda2, from practice, lamda value comparison candetermine if it is edge, corner, or flat region.

 

HOG:  intuition behind this is to calculatehistogram for oriented gradient (group in orientation bin), which considersorientation, gradient, and histogram (global statistic info) together torepresent a feature. It finally to use a 1-D vector with thousands of featuresto represent the image. Practical parameters show that it works well for humandetection.

 

SIFT: Harris cornerdetector cannot work well for scaling cases. So SIFT adds a process to handlingscaling: image pyramid + sigma scaling (for DoG), while it also introducedorientation for keypoint selection.

 

SURF: improvement onSIFT for fast calculation.

 

Hough transform: line fitting algo(线拟合算法）

 

Stereo Image:  深度图像, image with depth info. (Depth sensor). Different with 3D image,which is to make differential images for 2 eyes to let brain has depth info.Here stereo image is the image contain depth data to let computer computespace.

 

Bag of features(words):  originated from texturerecognition. For multiple object recognition, it needs to train the model toabstract a group object descriptors (feature vector) for each object, and thenuse SVM or other ML methods to do classification.  Each object has its own Cluster descriptor.Using K-means algo (unsupervised) to train the model.  Why each object has multiple features descriptors?Considering size, rotation, illumination etc.

 

PCA = principlecomponent analysis

PCA for image: eigen vector and eigen value  Ax = lamdax , 特征矩阵.

Eigen vectors are orthonormal basis(正交基）, and each image can be represented by aserier of eigen vectors.

Idea behind that is 线性分解.

After get eigenvector, select few most significant eigen vectors for face recognition.