Face Recognition Method Based On One-class SVM

With the development of the information security technology, face recognition which is an important branch of biometric identification technology has become a hot research topic in the field of artificial intelligence and pattern recognition. Rapidity, credibility and indirection are the most important characteristic of face recognition. But in the practical problem, face recognition technology is difficult to be extended to the specific application. We tend to come across some problems, such as the face samples can only get one class. Meanwhile, feature extraction and classifier selection are the most important parts in the process of face recognition. Aiming at this problem, this paper presents one-class support vector machine(OC-SVM) as a classifier to use in face recognition.One of the major benefits of OC-SVM is accurate description of training samples. When using OC-SVM to deal with a new sample, according to the similarity of the training samples, we could accept or reject it(the new sample). For the face samples can only get one class, namely one-class problem,we can use one OC-SVM to cope with the problem. For the face samples can get multi classes, namely multi-classification problem, we can use several OC-SVM to solve the problem. So OC-SVM can not only do with one-class problem but also handle the multi-classification problem. Through the research on the background and related theoretical knowledge of OC-SVM, we improve on rough one class support vector machine(ROCSVM) which is one of the OC-SVM. And we construct the study of the kernel function in OC-SVM. The main contributions of this paper are outlined as follows.(1)There is an obvious problem in traditional ROC-SVM, which ignores the inner-class structure of the training samples. In order to get better rough hyper-plane, our paper proposes a rough set one-class support vector machine based on within-class scatter(WSROC-SVM). This algorithm optimizes the inner-class structure of the training data by minimizing the within-class scatter of the training data. The method not only precipitates margin between the origin and the training data in a higher dimensional space as large as possible, but also makes full use of the inner-class structure of the training samples and makes the training data around the rough upper margin as tight as possible. What's more, our algorithm could deal with the over-fitting problem of training samples as well as consider the inner-class structure of training samples, which is one of important priori knowledge. Eventually, experimental results carried out on the UCI dataset indicate that the proposed method improves the accuracy as well as the generalization of the result. And the advantage could be reflected in the process of solving practical classification problems.(2)This paper studies the selection and structure of the kernel function in OC-SVM. Due to the training samples of OC-SVM just have one class, so the preprocessing of training samples has a great of effects on classification performance. Motivated by this observation, we construct a Gaussian kernel function based on feature weighted and selects a Fisher kernel with the gauss distribution, then take them as kernel of OC-SVM. Finally, from the experimental results on the UCI dataset, our algorithm proves the effectiveness of the two kernel functions and improves the recognition accuracy.(3)Based on WSROC-SVM, we design and implement a face recognition test system. Firstly, we use HOG to realize the image feature extraction and use WSROC-SVM to design an algorithm for face recognition. Then we select optimal parameters for the realization of face recognition test system by the experiments. Ultimately, experimental results indicate that WSROC-SVM is good at outlier detection in face recognition.