Research On Key-Technology Of Face Recognition

Face recognition is an active research area. It has a good applied prospect as an identity validation technology, such as information security, control to come and go. At the same time, face recognition is a classic high dimensional data and small sample size problem. It is harsh on the performance of algorithms. We deeply research some key-technology among them, for example D-LDA incremental algorithm, SVM multi-class classification algorithm.As a variety of LDA, D-LDA (direct LDA) has well theoretical foundation and experimental result. But it is baffled with a problem in practical application. The problem is incremental computing. This thesis proposes a D-LDA incremental algorithm (ID-LDA) by analyzing the within-class scatter matrix, the between-class scatter matrix and the algorithmic flow of D-LDA. Then this thesis proposes a regularization kernel D-LDA and realizes the kernel D-LDA incremental algorithm (IKD-LDA) based on the kernel D-LDA. In the experiments, we separately compare the ID-LDA algorithm with the traditional D-LDA batch algorithm, the IKD-LDA with the KDDA algorithm. Experimental results show, whether the ID-LDA algorithm or the IKD-LDA algorithm, the recognition accuracy is much closed to the traditional batch algorithm. It testified the proposed incremental algorithms are effective.Face recognition is a multi-class problem and the support vector machine (SVM) is a two-class method. We also research the SVM multi-class classification. The research work is consisted of two parts. The first part is about binary tree SVM. We propose a generating algorithm of binary tree structure based on clustering. The algorithm takes the separability between classes as metric and generates the node of binary tree by the clustering method. The second part transforms the multi-class problem to the two-class problem through varying the description of the problem. This thesis proposes a SVM multi-class classification method based on boosting in subspace. This method transforms the multi-class problem to the two-class problem by the intra-class difference subspace and extra-class difference subspace. And then we deal with the imbalance among the samples amount of two subspaces by boosting method. Experimental results show that the above-mentioned two algorithms both achieve the well result. And the method based on clustering is the best.