Face recognition using AMVP and WSRC under variable illumination and pose

Face recognition with variable illumination and pose is an important and challenging task in computer vision. In order to solve the problem that the accuracy of face recognition reduces with illumination and pose changes, this paper proposes a method via AMVP (AWULBP_MHOG_VGG_PCA) features and WSRC. In the proposed method, we need to extract AWULBP_MHOG and VGG_PCA features, respectively. As for AWULBP_MHOG features, firstly, variable illumination is normalized for face images. And uniform local binary pattern (ULBP) and multiple histogram of oriented gradient (MHOG) features are extracted from each block, which are called ULBP_MHOG features. Then, we use information entropy to obtain adaptively weighted ULBP_MHOG (AWULBP_MHOG) features. As for VGG_PCA features, we use the pre-trained VGG-Face model to extract VGG features from original face images. And PCA is used to reduce the dimension of VGG features to obtain VGG_PCA features. Then, AWULBP_MHOG and VGG_PCA features are combined to form AMVP (AWULBP_MHOG_VGG_PCA) features. Finally, test face images can be classified using weighted sparse represent (WSRC). The comparison experiments with different blocks, classifiers and features have been conducted on the ORL, Yale, Yale B and CMU-PIE databases. Experimental results prove that our method can improve the accuracy effectively for illumination and pose variable face recognition.