Feature Extraction Based On LDA For Face Verification

Face verification is one of the most challenging problems in the fields of pattern recognition and computer vision. It has a wide range of promising applications, such as credentials verification, entry and exit control, automatic log-on for personal devices and so on. Feature extraction is a very important step in face verification, also, it is a hot topic in the face verification area. Linear discriminant analysis (LDA) based on the Fisher criterion is a successful and widely used method among the various feature extraction methods.However, LDA suffers a problem of extracting insufficient features when it is directly applied in face verification. Therefore, we focus on this problem to conduct our research work. The main contents can be summarized as follows:(1) Firstly, we proposed a new LDA-based feature extraction method, called orLDA (LDA for over-reducing problem). orLDA obtains more class-between information through defining a new class-between scatter matrix, and it is able to get enough features to effectively separate different face images. Experimental results on ORL, LFW and CK+ data sets demonstrate that the effectiveness of orLDA.(2) Based on orLDA, we presented another novel feature extraction method, called SSDA (Separability-Oriented Subclass Discriminant Analysis), to solve the above problem of LDA. SSDA applies the idea of subclass and uses the information of subclass. Firstly, SSDA uses a subclass separability criterion proposed in the paper and combines the criterion with agglomerative hierarchical clustering method to find the optimal number of subclasses for each class. Then a new class-between scatter matrix and a new class-within scatter matrix are defined through using the information of subclasses. Finally sufficient face features which are beneficial to classification can be obtained by SSDA. Experiments demonstrate that the features extracted by SSDA can improve classification accuracy effectively.