A Novel Texture Descriptor for Interest Point Detectors

Feature descriptors combined with interest point detectors play an important role on image matching. The last decade, many descriptors, such as Scale Invariant Feature Transform (SIFT), Speed-up Robust Feature (SURF) and Fast Retina Keypoint (FREAK), come out to make it faster to compute, more compact while remaining robust to scale, rotation and noise. This paper introduces a novel texture descriptor computed on the regions detected by interest point detectors. The texture descriptor is 'local self-similarity (LSS) descriptor', which has been successfully used for generating a compact sketch best representing the common ensemble shared by all input images. While in this paper, LSS descriptor is to catch the texture features around the single interest point. Extensive comparisons demonstrate the LSS descriptor with Harris corner detector can achieve more robust and efficient correspondences.