Robust Face Recognition by Hierarchical Kernel Associative Memory Models Based on Spatial Domain Gabor Transforms

Abstract

Face recognition can be studied as an associative memory (AM) problem and kernel-based AM models have been proven efficient. In this paper, a hierarchical Kernel Associative Memory (KAM) face recognition scheme with a multiscale Gabor transform, is proposed. The pyramidal multiscale Gabor decomposition proposed by Nestares, Navarro, Portilla and Tabernero not only provides a very efficient implementation of the Gabor transform in the spatial domain, but also permits a fast reconstruction of images. In our method, face images of each person are first decomposed into their multiscale representations by a quasicomplete Gabor transform, which are then modelled by Kernel Associative Memories. In the recognition stage, a query face image is also represented by a Gabor multiresolution pyramid and the reconstructions from different KAM models corresponding to even Gabor channels are then simply summed to give the recall. The recognition scheme was thoroughly tested using several benchmarking face datasets, including the AR faces, UMIST faces, JAFFE faces and Yale A faces, which include different kind of face variations from occlusions, pose, expression and illumination. The experiment results show that the proposed method demonstrated strong robustness in recognizing faces under different conditions, particularly under occlusions, pose alterations and expression changes.