Face Detection Method Based On Improved Support Vector Machine

Face detection has become an important research direction of pattern recognition and computer vision, and it has practical values in many areas. Based on and overview of the current research on the problem of face detection, this thesis researches systematically the speed and the rebost of face detection. The main contents of the thesis are as follows:1. The classification speed of SVM is very slow, especially when the nonlinear kernel is used and the number of support vector is very large. In order to solve this problem, FLD/SVM hierarchical classification and a novel SVM decision tree is proposed according to the idea of coarse-to-fine. Firstly, the fisher linear classification is used to exclude large parts of background, and then, several linear SVMs are used for exclude part of background, lastly, a nonlinear SVM is used for verification. Experiment results show that the speed up factor is large and with no obvious loss in generalization performance.2. The performance of SVM highly depends on the parameters of model and the subset of feature, many methods optimize the parameter of SVM and choose the feature subset respectively, in fact, they affect with each other, so a novel method which optimizes the kernel parameters and feature subset simultaneously is proposed here. It use NGA,PCA and SVM to get the optimized eigenvector subset, and then the optimized feature subset is getted by the optimized eigenvector subset. Experimental results show that the new algorithm significantly improves the classification accuracy and reduce the number of feature.3. The approaches of face detection can be organized into two broad categories: the feature-based approach and the image-based approach. Each of them has some shortcomings and advantages, so the feature-base approach is introduced in the detectin phase. For an image which needs detected, firstly, the features of face are used for finding face candidates, and then, all the candidates are verified by the image-based approach. Experiment results show that this approach significant speeds up the detection, and there is no obvious decrease in the detection rate.