Face Detection And Recognition Based On Statistical Methods

Face recognition is one of the most challenging problems in the fields of pattern recognition and machine vision. It also becomes an active research topic recently. A robust face recognition system should contain detection and recognition. The detect rate and speed are two key problem of face detection, and multi-view face detection technique is still a challenging problem. Although face recognition technique develops quickly along with various applications, there are still many problems unsettled yet, for example, the robustness of recognition method and the influence of illumination.This thesis focuses on the problems presented above using statistical method and the main researches are as the follows.First, a new method about face detection using Census feature and Cost-Sensitive AdaBoost (CS-AdaBoost) is proposed. Census feature can present all the local spatial structures of an image, and it is robust to illumination. CS-AdaBoost is a new version of AdaBoost method, it can achieve higher detect rate during less iteration. Results show that this method can achieve higher detect rate than Froba's method, and it is faster than Viola's method because a multi-resolution image pyramid is used. Three view based face detectors are trained to solve the multi-view face detection.Second, a method that based on local singular value decomposition and HMM is presented. Local singular value decomposition is used to subtract the character of the face image. As a result, the recognition rate is increased and the computation is less then Liu's method.Finally, a method that based on phase-only reconstruction image and DiaPCA is presented. In the Fourier representing of signals, magnitude and phase tend to play different roles and in some situations phase contains more illumination information while magnitude contains texture information of an image. It is proved that DiaPCA is more accurate than original PCA. Phase spectrum combined with DiaPCA without any pre-processing is used to overcome the problem of illumination. Extensive experiment shows that the proposed method can achieve very encouraging performance in varying illumination conditions.