RTLBP-AN Efficient Local Pattern For Facial Images Retrieval

Abstract

This paper introduces a new local descriptor called radial transition local binary pattern (RTLBP) that is designed to extract more discriminative information from images. Unlike existing descriptors that use a 3 × 3 window, RTLBP uses a 5 × 5 pixel window and separates pixels into two scales. At the first scale, pixels are compared to the central pixel to generate binary values, similar to local binary pattern (LBP). At the second scale, pixels at major and non-major directions are compared differently to create two 8-bit patterns. This information is then processed using four primary and secondary directional radial pixels to create the feature descriptor. Finally, histograms are generated for each of the transformed images. The proposed technique outperforms state-of-the-art descriptors on two publicly available databases, demonstrating its effectiveness.