Researches On Support Vector Machine And Its Application To Face Recognition

Support vector machine (SVM) is an important algorithm in machine learning, and is widely used in pattern recognition domain. Different from those traditional algorithms based on empirical risk minimization rule, SVM is based on structural risk minimization rule. Thereby it can achieve a good balance between empirical risk and classifier capacity and has better performance. Ever since SVM was proposed, lots of researchers have been attracted and a lot of developments have been made. However, there are still two open issues: parameter selection and multi-class SVMs construction. The former refers to how to obtain a classifier with satisfying generalization ability, the latter refers to how to deal with multi-class problems with SVMs. To address these two issues, we perform researches on the following subjects: Binary SVMs parameter selection, multi-class SVMs parameter selection and multi-class SVMs output construction.Face recognition (FR) is an important application in pattern recognition domain. It is essential to both machine intelligence and the understanding of human visual system. Due to its powerful ability, SVM has been widely used in FR researches and excellent performances have been reported. However, to our best knowledge, no existing commercial FR system is based on SVMs. In order to explore the feasibility of SVM-based FR systems for real-world tasks, we perform investigation on two relative large face databases, and implement a preliminary FR software named idTeller using Visual C++.The contributions of this dissertation include: Two novel SVMs parameter selection algorithms based on the VC bound, the fixed-C algorithm and the VC-CV algorithm, are proposed. Though the VC bound is considered ideal for parameter selection, some inherit shortcomings make it difficult to use. Through step-by-step analysis, we gradually conquer the shortcomings and make VC bound practical to use. Then two novel parameter selection algorithms are proposed: the fixed-C algorithm and the VC-CV algorithm. Experimental results on several benchmark datasets show that, compared with the cross validation algorithm, the VC-CV algorithm not only can obtain better classifiers but also has less computing complexity.