Studies on support vector machines and applications to video object extraction.

摘要

Pattern classification is a fundamental problem under study in machine learning. During the past decade, Support Vector Machine (SVM), a learning scheme for classification, has drawn tremendous attention due to its theoretical merit and practical success. However, limitations still exist when SVM meets real-world applications. The major thesis of this dissertation is to introduce new formulations that are derived to overcome the limitations of SVM and thus extend its horizon in practice. Furthermore, based on SVM and the extensions a novel approach toward video object (VO) extraction is presented to add another practical dimension to this powerful learning machine.; The first extension to be introduced is psi-learning. It is motivated by the observation that the theory of SVM, which is well developed for separable cases, becomes less solid when extended to nonseparable cases. By replacing the hinge loss function in SVM with a designed psi function, psi-learning fully takes into account the generalization errors in nonseparable cases and consequently improves the classification accuracy in such situations.; The second limitation of SVM is the requirement of Boolean (or hard) memberships. To address this problem, we reformulate SVM to be a new learning machine named Soft SVM, which allows samples to belong to different classes by different degrees and adjust the classification boundary from them accordingly.; Thirdly, this dissertation considers the generalization of SVM from binary classification, which is the scenario the classifier is originally designed for, to multi-class as well as single-class scenarios. For the multi-class case, we introduce both static and dynamic reliability measures into the framework of the traditional one-against-all multi-class scheme, and then based on these reliability measures we propose a new decision strategy for a better one-against-all method. One-class classification, on the other hand, is one special problem that raises the issue of describing the target class rather than discriminating between classes as in the binary and multi-class problems. In the context of SVM, we propose a new one-class classifier named minimum enclosing and maximum excluding machine (MEMEM), which offers capabilities for both pattern description and discrimination.; In practice, run time is always a critical factor, and the problem of slow training of SVM has been a bottleneck. In this dissertation, we tackle this efficiency issue in the area of feature selection. Two steps are taken. First, a new criterion is proposed to effectively filter out non-essential features before each training step begins. Secondly, we dynamically maintain a subset of training samples and use them rather than all the available samples for every necessary training. As a result, the total computational load is significantly reduced.; Lastly, a novel approach toward VO extraction is presented. Each VO is considered as a class, and VO extraction is realized by classifying every pixel to one of the available classes. It is significantly different from the traditional approaches yet overcomes many of their shortcomings. SVM, psi-learning, and Soft SVM are employed as the classifier and experimental results demonstrate the great potential of machine learning in the area of VO extraction.