A Novel Feature Extraction Method for Face Recognition under Different Lighting Conditions

This paper develops a novel method named image decomposition based on locally adaptive regression kernels (ID-LARK) for feature extraction. ID-LARK is robust to variations of illumination, since it decomposes the local features into different sub-images. And they describe the structure information hidden in the unobserved space. More specially, ID-LARK first exploits local structure information by measuring geodesic distance between the central pixel and its neighbors in the local window with locally adaptive regression kernels. So, one image can be decomposed into several sub-images (structure images) according to the local feature vector of each pixel. We thus downsample every structure images and concatenate them to obtain the augmented feature vector. Finally, fisher linear discriminant analysis is used to provide powerful discriminative ID-LARK feature vector. The proposed method ID-LARK is evaluated using the Extended Yale B and CMU PIE face image databases. Experimental results show the significant advantages of our method over the state-of-art ones.