Feature Description Based on LBP and Order Pooling

A novel local image descriptor named OCLBP(Order based Complete Local Binary Patterns) is proposed in this paper. It firstly extracts CLBP(Complete Local Binary Patterns) features pixel-by-pixel in the local image patch, and then pooled the CLBP features based on order relations of the pixel brightness in the patch. Specifically, the CLBP feature is based on CSLBP and adds the information of center pixel. The pattern number of the pixels is reduced compared to CSLBP. Furthermore, in order to better meet the scale invariance, the descriptors is extracted in multi-scales and the final descriptor is constructed by concatenating the histograms in multi-scales. The performance of the proposed descriptor is similar to the state-of-the-art descriptors SIFT and CSLBP in the condition of perspective transformation and light changes. In the condition of scaling, rotation, blurring, and JPEG compression, our descriptor has a better performance against the above descriptors using standard benchmarks.