Kernel class specific centralized dictionary learning for face recognition

As a fundamental and effective method, sparse representation based classification (SRC) has been applied to computer vision field for many years. However, SRC assumes that the training samples in each class contribute equally to the dictionary which will cause high residual errors and instability. In order to solve the problem and improve classification performance further, class specific centralized dictionary learning (CSCDL) algorithm was proposed. CSCDL considers the concentration of sparse codes in the same class and shows good recognition performance but the fact that CSCDL is only suitable to linear structures limits its applications. To address the limitation, in this paper, we expand the CSCDL algorithm into the kernel space and turn the nonlinear problem into linear ones. Kernel functions and some nonlinear mapping, are used to map original data into a high-dimensional kernel feature space. Effective experimental results on face recognition benchmark databases prove that the performance of kernel class specific centralized dictionary learning (KCSCDL) algorithm is superior to that of CSCDL.