Face Recognition Based On Semi-supervised Dimensionality Reduction

Face recognition is a hot issue in pattern recognition and machine vision which plays animportant role in human-computer interaction, security, authentication and other fields. Faceimage has a complex structure and is vulnerable by various factors, such as illumination, age,expression, pose, occlusion, attitude, so that face recognition is still a challeng. A largenumber of face recognition algorithms emerge in recent years. In the face recognition system,its performance depends on the effectiveness of face image feature extraction. Because of theface data with high dimensionality, it easily causes the "curse of dimensionality”problem.Semi-supervised dimensionality reduction is a fast and effective method,butsemi-supervised dimensionality reduction algorithm has some problems when process faceimage containing various illumination.In order to overcome this kind of problems, we propose a method based on totalvariation (TV) and semi-supervised dimensionality reduction (SSDR) which is used to dealwith face images. The main research work is by the following aspects:(1)Experiment shows the total variation model has better capability of edge-preservingand can weak shadows.It has better capability in dealing with face image than traditionalmethods.(2)Experiment on other dimensionality reduction algorithm with semi-superviseddimensionality reduction (SSDR) shows that the semi-supervised dimensionality reduction(SSDR) can have better recognition rate than other algorithms.(3)Combining with the methods proposed above, a robust face recognition system isproposed.The experimental results demonstrated the superiority of the proposed method whendealing with face images that is under uncontrolled illumination conditions and have highdimensionality.Experiment on YaleB database，CAS-PEAL database and ORL database shows ouralgorithm is suit to deal with face image which is under various illumination condition andhave high dimensionality.