Tensor-Compensated Color Face Recognition

Abstract

Making face recognition more reliable under uncontrolled lighting conditions is one of the most important challenges for practical recognition systems. The reasons come from the need for automatic recognitions and security systems. To overcome this problem, we propose a novel illumination compensation method called adaptive high-order singular value decomposition to enhance face images at the preprocessing step of the face recognition system. First, we present an RGB color face image as a third-order tensor. Then, adaptive high-order singular value decomposition is proposed to adjust the core tensor automatically by multiplying three frontal slices of the core tensor with their corresponding compensation weight coefficients while keeping the third inverse factor fixed. The experiments performed on five of the most famous public color face databases, namely CMU-PIE, Color FERET, FEI, LFW, and IJB-C reveal that adaptive high-order singular value decomposition not only yields compensated images that are clear, natural, and smooth but also considerably improves the accuracy and computing time of face recognition.