Face Recognition Algorithms Research Under Different Lighting Conditions

With the development of face recognition techniques, illumination problem becomes one of the obstacles. Sometimes, the difference between two face images caused by variant illumination is greater than that caused by individual difference. One important purpose of this paper is to improve the face recognition accuracy. To achieve this purpose, the focus of the work is on image preprocessing, feature extraction and feature recognition. The research starts from illumination problem, focuses on robust feature extraction and the main work includes:(1) Research of different illumination preprocessing methods. This part summarizes the current illumination preprocessing approaches which includes the method based on wavelet transformation, self-quotient image(single scale, multi-scale self-quotient image) method, Retinex(single scale, multi-scale, adaptive single scale Retinex) method, smoothing (isotropic, anisotropic) method, DCT normalization in logarithm domain method, homomorphic filter and local contrast enhancement(LCE) method. They are compared and analyzed that the same preprocessing approach in different face databases and different preprocessing approaches in the same face database through perfect experiment schemes.(2) Research of different feature extraction methods. First the two new means are proposed as follows: One is adaptive feature extraction based on curvelet transform in robust face recognition; another is 2-D PCA non-parametric subspace analysis face recognition algorithm. Finally, the experimental results show that the proposed methods are prior to the current methods which include 2DPCA, 2DLDA, (2D)2PCA, (2D)2LDA, 2DPCA+2DLDA, 2DNSA(two dimensional non-parametric subspace analysis), meanwhile the performance of the proposed methods is analyzed through different experiments.(3) Performance research of different feature extraction methods after different preprocessing. Most of current preprocessing methods have weak generalization. They are different to different face databases, so to achieve good recognition performance the specific preprocessing approach should match the specific feature extraction method in some face database.