Locally Kernel-based Nonlinear Regression for Face Recognition

The variation of facial appearance due to the viewpoint or pose obviously degrades the accuracy of any face recognition systems. One solution is generating the virtual frontal view from any given non-frontal view to obtain a virtual gallery/probe face. As the state-of-the-art face recognition algorithm, linear regression computes a reconstruction matrix from the images of each subject and then approximates the probe face image by using the reconstruction matrix, but the performance of this linear algorithm is limited due to the nonlinear structure of the face images which is caused by variations in illumination, expression, pose and occlusion. Following this idea, in this paper, we propose an efficient and novel locally kernel-based nonlinear regression (LKNR) method, which generates the virtual frontal view from a given non-frontal face image. Because of the high (even infinite) dimensionality of the nonlinear transformation functions, it is infeasible to directly calculate the corresponding reconstruction matrix and therefore is unable to approximate explicitly the probe image. So, with the help of kernel functions, we overcome to this mentioned problem by embedding the nonlinear regression in the stage of computing the reconstruction matrix from the non-frontal input face and non-frontal face database. The comparison of the proposed method with locally linear regression (LLR) and eigen light-field (ELF) methods is also provided in terms of the face recognition accuracy. Experimental results show that the proposed method outperforms two other methods in terms of robustness and visual effects.