Spatial pyramid face feature representation and weighted dissimilarity matching for improved face recognition

Abstract

In this paper, we present a novel face recognition (FR) algorithm based on multiresolution spatial pyramid. In our method, a face is subdivided into increasingly finer subregions (local regions) and represented at multiple levels of histogram representations. To address image misalignment problem, overlapped patch-based local descriptor extraction has been also developed in an effective way. To preserve multiple levels of detail in facial local characteristics and to encode holistic spatial configuration, face features obtained for concatenated histograms (coming from all levels of spatial pyramid) are integrated into a combined feature set, termed spatial pyramid face feature representation (SPFR). In addition, to perform recognition by matching between the pair of probe and gallery SPFR sets, we propose the use of a weighted sum of the dissimilarity scores computed at all spatial pyramid levels. For this purpose, we develop a novel weight determination solution based on class-wise discriminant power estimation for face feature at a specific pyramid level. We incorporate our proposed algorithm into general FR pipeline and achieve encouraging identification results on the CMU-PIE, FERET, and LFW datasets, compared to previously developed methods. In addition, the feasibility of our method has been successfully demonstrated by making comparisons with other state-of-the-art FR methods (including deep CNN based method) under the FERET and FRGC 2.0 evaluation protocols. Based on results, our method is advantageous in terms of high recognition accuracy and low complexity, as well as straightforward implementation.