Research On Face Recognition With Illumination Variation

Face recognition is to identify or verify one or more faces in still images or video sequences based on learned face images. As one of the most challenging problems in pattern recognition and machine vision, face recognition has a wide range of potential applications in the areas of public security, information security and human-computer interaction. After more than 30 years, face recognition technology under well-controlled environment is going to practical stage, while the recognition performance dramatically degrade under uncontrolled environment such as variant illumination, head poses, facial expressions, occlusion on face and so on. Therefore, one of key issues in face recognition is the study of robust recognition algorithm.In this thesis, our research focuses on face recognition with illumination variation, and two algorithms are proposed as follows:1. A Face recognition algorithm based on improved BP neural network is proposed. First, wavelet transform is utilized to reduce both the dimension of face images and the data amount. Then face images are normalized by illumination ratio image to overcome the influence of illumination variation. Finally, BP neural network is improved for face recognition, the problem on choice of parameters is discussed, the Sigmoid function and weight adjustment are improved for higher convergence speed.2. A face recognition algorithm based on estimation of reflectance is proposed. First, the reflectance at symmetric points on face surface are estimated by symmetry of human face using two face images under known illumination. Then face images in different illumination are generated by the reflectance. Illumination normalization is implemented using homomorphic filtering and amendment based on symmetry for both testing and synthesized images. Finally, face recognition is implemented using normalized images.The above two algorithms are tested using Yale Face Database B. The experiment results compared with other algorithms confirm the efficiency of proposed algorithms.