Research On Illumination Invariant Algorithms For Single Sample Face Recognition

Face recognition, a new and high-end bio-metric identification technology in the twenty-first century, has broad application prospects in many areas with the development of society. However, the performance of face recognition drop dramatically when there are unexpected factors in a practical application and its wide application is also facing with severe challenges. Illumination issue is a difficult problem to handle among all the factors which have influence on face recogniton. Illumination has an influence on the face images inevitably in outdoor which makes the facial features unrecognizable. In this paper, we make an intensive study of illumination issue on face recognition and try to weaken illumination on face images to extract illumination invariant features. We mainly focus on the single sample face recognition under varying illumination conditions. Main works and research results are as follows:(1) Gradient feature is briefly introduced. Researches show that Gradient feature of face image is nearly an illumination insensitive measure. So in light of this viewpoint, we propose an improved version of gradientface algorithm which can be used for sparse image approximation by extracting gradient feature from eight directions and we also use LBP code scheme to make our proposed algorithm robust to illumination to a certain extent. Compared with other classic illumination invariant algorithms, our method has a significant improvement on single sample face recognition under varying illumination.(2) A geometric adaptive Haar-type wavelet transform especially designed for sparse image approximation — Tetrolet transform and a duplex nonlinear local contrast enhancement method based on the human retina model are introduced. On this basis, an effective illumination invariant feature extraction method is proposed. We combine the retina model processing as pre-processing with a suitable Tetrolet transform. The retina model processing is applied to the low frequency coefficient matrix for illumination removal which can increase the classification information.(3) Retinex theory and its modified version Single/Multi-scale Retinex are introduced. And in order to present the multi-scale features of human visual system effectively, we bring in Adaptive Nonsubsampled Laplacian Decomposition based on Nonsubsampled Pyramid and propose a new illumination invariant feature extraction method based on Retinex. We use local inhomogeneity to measure the degree of discontinuity of gray scale variation and set different weights for each pixel to evaluate illumination component.