Low-Dimensional Local Self-Similarity Descriptor for Image Matching

Abstract

The advanced descriptor, i.e., Local Self-Similarities (LSS), is successfully adopted for object classification and scene recognition. However, it is only invariance against small geometric and photometric transformations. In this paper, two low-dimensional descriptors for extracting distinctive invariant features from interest regions are presented, i.e., PCA and Local Self-Similarities feature based descriptors, namely PCA-LSS and PLSS. They are achieved by applying PCA on LSS feature field and the improved LSS descriptors of normalized patches, respectively. The performance of these proposed PCA-LSS and PLSS descriptors to image matching is studied through extensive experiments on the INRIA Oxford Affine dataset. Empirical results indicate that the proposed PLSS achieves more transformation invariance, significantly outperforms the original LSS, and also outperforms the SIFT.