Support Vector Novelty Detection For Face Detection

In recent years, face detection has been concerned greatly by researchers because of its practical applications in security access control, visual surveillance, digital video processing, content-base retrieval and a new generation of human-machine interface. At present, many methods for face detection have been proposed. Commonly these methods have two steps. One step is about facial image feature extraction, or how to characterize a human face images. The other is about classification of the faces and non-faces by designing effective classifiers on extracted features. This thesis studies the application of support vector novelty detection (SVND) algorithm, and proposes some variations of SVND to improve the detection speed. It mainly includes the following three aspects.The standard SVND had been applied to human face detection and achieved good detection accuracy, but the detection speed is not satisfied. To solve this problem, we propose a cascade SVND method, which is inspired by cascade ensembles of classifiers. Experimental results show that our method is effective. Under the same condition of detection accuracy, cascade SVND can shorten the detection time compared with the standard SVND.In the cascade SVND method, we need much support vectors (SVs) to obtain a good detection accuracy. As it is well known, the number of SVs is the main reason which can affect the detection speed. Thus, we present a sparse SVND method in which two sparseness techniques are adopted, or the 1-norm regularization and the hinge loss. Both of them would lead to a sparsity. By using our method, we would get small number of sparse SVs, which can improve the detection speed. Our experimental results show that this method uses very small number of sparse SVs to achieve fast detection without loss of detection accuracy.In the practical detection process using SVND and its variation, we intentionally neglect the process of image search. We just save all test samples searched from images in advance, and use them directly in detection process. Thus our detection time is only about the test time of these samples. Since the image search is time-consuming, we can not take a real-time processing. In order to improve image search speed, we propose a fast searching image method based on linear filter. Since a two-dimensional linear filter can directly compute an inner product of two vectors by using convolution operation, we can take it as an implementation of the linear kernel. Here, only linear kernel can be employed. One-class SVM (SVND) could not get good detection accuracy if the linear kernel is used. Therefore, we also exploit sparse SVM for binary classification problem in our experiments. Experiments validate this method could greatly improve detection speed.