Face Recognition Based On The Hybrid Kernel Of SVM

Thanks to the urgent requirements of social safety, such as public security, identity authentication and so on, the biometric features recognition technology (BFRT) has obtained rapid development. The BFRT relies on the biological features that humans own, such as fingerprint, iris, face, sound, digital vein, etc. The BFRT has higher security and reliability than traditional identification methods because of the non-replication of biometric features of humans. Different from other biometric features, the parties could be identified in a non-cooperated condition, which makes face recognition friendlier and more direct.It is difficult to extract effective and complete face features due to the influences of light, decoration, posture, facial makeup and so on. The subsequent classification speed and recognition rate will also be affected. On the other hand, the similarity, mutability and non-rigidity of the face make face recognition one of the most challenging technology in biometric features recognition field. To solve these problems, the face recognition system based on the face database is researched and the whole process of face recognition system is analyzed in this paper. The feature extraction of face images and the classification process based on Support Vector Machine (SVM) are mainly focused on. In the feature extraction stage, Principal Component Analysis (PCA) and two-dimensional Principal Component Analysis (2DPCA) are fused. In the classification stage, an SVM combined with a hybrid kernel is proposed, and a cross validation parameters optimization algorithm is improved. The main work of this paper is as following:1. The thesis introduces the key issues statistical theory of the SVM, that is the minimum of VC dimension, the boundary of ability and the structural risk. Then, the theory of the SVM algorithm is detailed introduced. The conditions of the satisfaction and the preferences of the kernel functions are introduced as well. As a traditional binary classification machine, SVM could be applied to solve the multi-classification problems including the face recognition.2. Facial feature extraction. Firstly the PCA based on the Karhunen-Loeve conversion and the 2DPCA are introduced. The PCA is established on a one-dimensional vector of image covariance matrix, and the 2DPCA is directly established on the image covariance matrix. Then considering of the advantages and disadvantages of these two methods, an improved feature extraction method which combines 2DPCA with PCA is proposed. It is confirmed that this new method can improve the feature extraction and the classification speed and the recognition rate. The experiments on the ORL database show that the recognition rate can reach 93.0%.3. Face classification. The classification performance of the traditional kernel functions is studied using the above-mentioned SVM kernel method. A hybrid kernel combined the radial basis kernel with the polynomial kernel is proposed according to the learning ability and extension ability of the overall kernel and the partial kernel. The time-consuming problem of traditional leave-one-out cross validation is also improved, which raises the face recognition rate to 97.5%. Compared to the single kernel function, the hybrid kernel function shows better learning ability and extension ability.