Robust face recognition based on Kernel Reduced Rank Regression

In practical applications, face recognition will be influenced by a number of uncontrolled factors, such as varied facial expression, poses, illumination, etc. In our paper, we aim at reducing the impact brought by variations of head pose. Under ordinary conditions, there is only one frontal face of each person in the gallery, thus we augment the gallery by synthesizing images in other different poses by using an effective regression based approach. In this approach, the facial landmarks on non-frontal faces can be estimated from their frontal images by the learned mappings between frontal landmarks and non-frontal ones. The mappings are achieved offline via Kernel Reduced Rank Regression (KRRR). Then the non-frontal face images are synthesized by Piecewise Affine Warping (PAW) and used for gallery extension. To demonstrate the validation of this approach, a frontal recognition system based on Multi-Region Histograms is augmented, and the augmented recognition system is tested on Multi-PIE dataset. Compared with other state-of-the-art methods, the approach is more robust to pose variation, while less time consuming.