Research On Single-view And Multi-pose Face Recognition

In With the improvement of social information and automation, face recognition has attracted growing attention as a kind of biometric personal identification technology. The face recognition under ideal conditions has been properly sovled, while in a real world application of face recognition technology, when facing the uncontrollable factors like view angle, pose, illumination, emotion etc., existing methods can not well eliminate their influences. Among all these factors, single-view and multi-pose in face recognition application usually exist at the same time, but there is a conflict between the single-view which offers too few samples and the multi-pose which need complete samples. So this thesis is dedicated to overcome the probelms above and improve the level of single-view and multi-pose face recognition. To this question, this thesis has studied the problems in the existing methods, proposing a solution aimed at the single-view and multi-pose face recognition.First we anlysised the existing Automatic Face Recognition System, proposing utilizing the recogniton preprocess methods of pose estimation and generating virtual faces to sovle this problem: we filter the faces not for recognition according to the results of pose estimation, then we extend the training face database by generating virtual faces to sovle the multi-pose probelm. And according to this solution we designed the single-view and multi-pose face recognition system based on pose estimation and virtual face generating.Then, we studied the pose estimation methods, and proposed a pose estimation method based on Within-Class Covariance Normalization which combined the feature extraction algorithm HOG and the data processing algorithm WCCN and utilized SVM with generalized linear kernel to complete pose classification. The experiments on CMU-PIE database have demostated that this method can well sovle pose estimation problem.Next, we studied virtual face gernerating methods, improve the results of the methods based on polynomials by mend the existing ones. Considering the weakness of methods based on polynomials in fitting errors, we proposed a virtual face gernerating method based on RBF neural network model with much lower fitting error and more approximation to real face. Finally, we compare the results of the face recognition experiments, concluding with the effectiveness of our methods in the single-view and multi-pose face recognition.