Directional gradients integration image for illumination insensitive face representation

Abstract

The illumination problem is one of the main bottlenecks in a practical face recognition system. Illumination preprocessing is an effective way to handle lighting variations for robust face recognition. In this paper, we present a novel illumination insensitive image, namely directional gradients integration image (DGII), for illumination insensitive face recognition. Unlike the existing model-based methods, the DGII is generated directly from the decomposed gradient components of a logarithmic image, without involving any training procedure. Based on the Lambertian reflectance model, we first calculate the horizontal and vertical gradients in the logarithmic domain to eliminate the illumination component. Secondly, to utilize the gradient orientation information, the two gradients are further decomposed into four components along four directions. Then, the four directional gradients are integrated to reconstruct an illumination insensitive image using anisotropic diffusion. Finally, the reconstructed image is fused with the gradient magnitude image through weighted summing. For performance evaluation, we simply use principal component analysis for feature extraction, Euclidean distance as similarity measure and nearest-neighbor classifier for face recognition. Experiments on the Yale B, the extended Yale B and the CMU PIE (The Carnegie Mellon University pose, illumination and expression database) face databases show that the proposed method provides better results than some state-of-the-art methods, showing its effectiveness for illumination normalization.