Research on Local Self-Similarity Descriptors in Images

Local self-similarity descriptors use image patch as the basic unit for measuring internal self-similarities, which can capture more meaningful image patterns than individual pixel. But computation speed of the current method is slow. In order to improve the current method in extracting local self-similarity, we present a new approach. In our approach, we first obtain an image region centered at every pixel. Then every image region is divided into several patches and self-similarities are extracted from each patch. At last all of the self-similarities form a vector to describe image region. The results of simulation show feasibility of the approach.