Research On Statistical Face Recognition

Face Recognition Technology (FRT) is emerging as an active research area in the field of pattern recognition and artificial intelligence. As a biometric technology, FRT has numerous applications such as access control, law enforcement, e-commerce, video surveillance and so on. The process of face recognition can be divided into three steps, i.e., preprocessing, feature extraction, classification. With different application environment, different preprocessing method should be adopted, the corresponding algorithmic difficulties will change accordingly. Feature extraction is the core of recognition task, which directly impact on classification velocity and recognition ability. In this thesis, 5 methods were analyzed and discussed, include: PCA method based on statistical representation feature; LDA method based on statistical discrimination feature; method base on Singular Value feature; method based on Gabor feature in transformed domain and method based on Bayes rule.In the part of Principal Component Analysis (PCA) method, firstly the KL transform is introduced and its reconstruction ability is studied. Also research is conducted on the influence of those small things surrounding human faces, such as hair, beard. Experiments are carried out on several standard face database (includes ORL, Yale, Manchester face base) and good results are generated. Further more the recognition ability of class-meanvalue face is researched definitely, which has a good computational stability. We also find that PCA on class-meanvaluecan has comparable ability with traditional PCA with fewer features.PCA is a classic method in image processing, but not a method especially for classfication. As a class technique in pattern recognition, Fisher linear Discriminant Analysis takes different apart best by minimizing within-class scatter matrix S_w and maximizing between-class scatter matrix S_b.But Fisher rule can't be directly introduced into the problem of face recognition for there exist a small sample problem. Two methods are discussed in this thesis: by applying KL transform on face data, we project faces into a PCA space and carry out Fisher rule in the space directly. Also we find out that the null space of S_w is useful and we can get the project matrix using its null space.In the processing of recognition, the most important point is to extract those faetues that are most beneficial to recognition. Gabor transform is introduced in to face recognition for its likeness with human visual system. After discussing sampling question in space frequency domain, we suggest that different discrimition features lie in different frequency domain so that a non-equally sample (more samples in low frequency domain and less samples in high frequency domain) can get a good result vs equally sampling method. The experiments testify our suggest.The Singular Value Vector has good characteristic in both algebra and geometry, and it can also be taken as a classification feature. In experiments, singular value vectors of face pictures are taken into computation instead pixel values of face pictures, PCA and LDA method are carried out then and results show their advantages on traditional PCA and LDA.Bayes rule is the best method in pattern recognition. Pentland firstly introduced it into face recognition in his paper. By constructing difference picture database of same people and difference picture database of different people, the problem of recognition turns into judging which database the difference picture belongs to. In this paper the score formula is simplified and two new formulas are set up. Experiments on these formulas show the advantages of this method.