An Improved Image Classification Method Based On Spatial Pyramid Matching Framework

Abstract

Spatial Pyramid Matching (SPM) and its variants have achieved a lot of success in image classification. However, the process of extraction of SIFT descriptors may cause texture information loss under the SPM framework. In addition, traditional Grey Level Co-occurrence Matrices (GLCM) method does not consider the spatial relationship of raw pixels. In this paper, to solve the issues, an improved GLCM method are proposed, which enhances the performance of original GLCM method for image texture feature representation. Firstly, the original images are subdivided to increasingly fine sub-regions and a weighted vector is generated by GLCM method inside each sub-regions. Secondly, the weighted vector of each sub-regions and the SIFT descriptors are concatenated. Thirdly, the descriptors are encoded by using the technique of sparse coding and the final features are obtained by max pooling. Finally, the features are fed to kernel classifiers to evaluate the effectiveness of our method. Comprehensive experimental results reveal that our method achieves comparable classification accuracy compared to the previous methods.