The Research On Face Authentication System And Related Techniques Based On Trace Transform

Identification technology based on biometrics plays more and more important roles in our society. Compared to other methods, the authentication system using human face characteristics draws more attentions, and it can be applied in a wide range of fields because of its low cost and good invisibility without any infringement. As one of the most difficult issues in pattern recognition field, it's also a very challenging task including several hard problems, such as face detection, face authentication, understanding and synthesizing of face expressions. This thesis mainly pays attention on following aspects.To increase the rate of face detection, an effective and robust image enhancement algorithm CLAHE (Contrast Limited Adaptive Histogram Equalization) is introduced for improving the contrast of digital images captured under varying conditions. Combined the contrast limiting approach with the adaptive histogram equalization, the CLAHE method gets a higher image contrast at a lower cost of visual appearance, which boost the performance of current Adaboost face detection algorithm evidently in our experiment implemented on the BAC2005-FDIB-B Face Database.To get a normalized face matching module for the later authentication using, a new normalization method is presented. Instead of traditional rectangles, a standard elliptical template was designed to get face area of input images. After geometry correction, template clipping and gray scale normalization, different faces in all images can be unified to a certain degree.To achieve higher performance in face authentication, a new system based on trace transform is explored and developed. The system converts the original face image to another "image" using trace transform. After image segmentation and edge detection, a shape context that expresses the original face image in an alternative measure can be generated. Then with Hausdorff context and a novel similarity measure which integrates the spatial information with structure information of the shape adopted, two shape contexts are compared to determine if they are matched. The optimal threshold for image segmentation can be searched by reinforcement learning. This authentication system is verified with experiments on the FDIB-B Face Database.