A Robust Local Sparse Coding Method For Image Classification With Histogram Intersection Kernel

Machine learning now has been used in every regions. With the fastdevelopment of Internet, there exists billions of images that are being uploaded,downloaded every day, which brings a giant challenge to computer vision. Today thesearch engine consumes a lot of labor and money to annotate the images in thedataset and when uses search some type of images, they match the key strings withthe annotation, and obtain the candidate dataset, which returns to the users later. Ifwe can improve the accuracy of automatic image classification, and the machinescan classify the image dataset very well, it can return the candidate images for oneuser’s query. As a consequence, it will save the labors and meanwhile, improve thework of efficiency. We can adopt two methods to improve the accuracy of imageclassification:(1) Train brilliant classifier, like non-linear SVM, complicated neuron network,etc.(2) Learn salient features with simple classifier.This article aims to learn robust image features to improve the accuracy. Sparsecoding and local sparse coding are very effective methods to extract image features,but they encode the histogram in the histogram space. Histogram Intersection Kernel(HIK) is a kind of kernel function that widely has been used in image processing. Asa consequence, this kernel maps the histogram into a higher space, whichimplements the nonlinear transition of original features. This article combines localsparse coding with histogram intersection kernel, encodes every histogram in themapping space, builds the codes for one image and last verifies the method based onthe experiments. To complete this aim, we need:(1) Extract the histogram features of every image, like SIFT, HOG, etc.(2) Use EM algorithm to compute the dictionary in the mapping space.(3) Optimize the least square problem with a linear constraint to compute thelocal sparse code in the mapping space for every histogram.(4) Use pooling and spatial pyramid matching method to build the code of everyimage.(5) Classify the images with linear classifiers, and analyze the results of differentmethods.The experiments shows that the method proposed in this article notably canimprove the accuracy of image classification. Compared with traditional local sparsecode (done in the histogram space), the features built by our method is morediscriminant and can obtain higher accuracy.