Research On Illumination Normalization Technique Of Face Verification

Currently, it is difficult for face verification to overcome the impact of the ambient illumination. This is because the variation of face image, brought by the ambient illumination, is even bigger than the variation between the face images from different bodies. Meanwhile, the performance of the face verification system will dramatically drop if the ambient illumination of the verification process is different from that of the registration process. To enhance robustness of the system to illumination changes,a multiscale illumination invariant facial feature image extraction method is proposed in this thesis? based on anisotropic diffusion. To eliminate the halo effect of anisotropic diffusion algorithm, a novel anisotropic diffusion algorithm is proposed by defining new local inhomogeneity and conduction function. The final illumination invariant facial feature image extraction method is acquired by embedding the algorithm in the generalized quotient image method. The main research works in this thesis are as follows:①This thesis studies the gradient descriptor of the traditional anisotropic diffusion method which is used to describe the gradient of gray face image. Two kinds of gradient descriptors are introduced in this method for the drawback of its gradient describing. Spatial gradient is used to describe gradient of gray image, then the method improves the spatial gradient aiming at its drawback in describing gradient direction. To enhance the description of the inconsistency between neighboring areas, a new local inhomogeneity is introduced in the proposed method,and then gradient direction is obtained.②Conduction function in the traditional method may easily lead to sharpening of the edge of diffusion image. Although it is a good feature for edge extraction, it is detrimental to process the illumination changes of face image. And it will make the processed image mixed with a lot of noise. This thesis improves the transfer function for this situation, and thereby substantially reduces noise of the processed image. It significantly improves the treatment effect of the proposed method.③The implementation of traditional anisotropic diffusion is an iterative diffusion process, and the method is carried out without any iterative constraints. It is very likely to cause the overfit of the method if we directly use the method to process face image. This thesis introduces a specific diffusion constraint in the new anisotropic diffusion, which can make the method more suitable for processing illumination problem of face image.④To validate the superiority of the proposed method relative to other major illumination invariant feature image extract method, a large number of comparative experiments on Yale B and CMU PIE face database verified the validity of the proposed algorithm.⑤Comparative experiments of the proposed method relative to other major illumination invariant feature image extract method on Yale B and CMU PIE face database was conducted, and the results show the superiority of the proposed method.The proposed method, which has no modeling steps or training images required, is able to extract illumination invariant facial feature image from multiscale space. Meanwhile, the proposed method can preserve edge information in low frequency illumination fields, and can achieve promising results even under harsh illumination conditions.