Illumination Compensation In Face Recognition

Face recognition technology has great theoretical and practical value in social security. In fact, different face images have inherent similarity, and same face images often quite differ due to pose and expression variations. Meanwhile, the quality of face images are also vulnerable to light, shadows and other factors, face recognition accuracy in practical application still can not meet people’s requirements. Therefore, it is necessary to preprocess influence of illumination, shade and other effects before face recognition. This article focuses on the problems of illumination in the face recognition technology and eye location method of big angle rotation face.The research topics and the main contributions are as follows:Valid face recognition based on unsymmetrical characters under changed illumination is given in the paper. A new function named texture discriminability is proposed and a new method is put forward which could remove noise features and obtain valid features for recognition according to different changed illumination. Experiment results show that recognition speed is accelerated and recognition rate under changed illumination is improved.In the paper, a method on illumination direction estimation based on SVM is proposed to solve the problem of complex, nonlinear illumination direction classification in face recognition. By bifurcate tree, the complex classification becomes a simple and multi-level progressive classification, and only three classes are sorted at every branch of bifurcate tree. Moreover, different features of current scale are extracted for SVM according to their characteristics of every scale of light direction. The proposed method is efficient to solve multi-classes classification problem and the accuracy can be improved obviously.In this paper, we propose a new illumination normalization method based on inherent texture features within micro-neighborhood, which transforms the nonlinear process of grey level changes with illumination into a linear process in the micro-neighborhood. The inherent texture future is illumination insensitive and it weakens influence of face structure in illumination normalization. The effect of the algorithm is better and its’ computational complexity is low which is adapt to a real face recognition system.A new eyes location algorithm of big angle rotation face is proposed. A normal face rotation model is built based on circle-block submatrix, by which the estimation to angle rotation is transformed into computing the moving distance of circle-block matrix. The algorithm can be used to locate eyes in any rotation angle on FERET database.