Research On Feature Extraction Of Human Face And Classifier Design

Human face recognition technique is an important research field of pattern recognition. It has been widely concerning about by international machine learning community. As human face recognition systems are widely used in business and security and other fields, the human face recognition technique known by more and more people. The problems it mainly considers can be described as:for a static or video image under a given screen, how to recognize one or more person in this screen with the stored face database. Feature extraction and classifier design are two important parts of face recognition. How to extract more effective features and to design a classifier with excellent performance are always hot topics in human face recognition technique. In this dissertation, this two topics are further discussed, in which some proposed algorithms work well for human face recognition.The main research contents and achievements are as follows:1. An improved learning method named Locality Embedding Projection (LEP) algorithm is proposed to overcome the defect of LDA:local geometric features are not considered. The local idea of manifold learning and Fisher criterion function are combined in this method, which makes it not only extracts the best discriminant features, but also better preserves the sample’s local geometric structures. More specifically, this method firstly uses the neighborhood relationship and class label information to classify the training sample set. For each training sample, there are two classes which are called neighbor class and non-neighbor class; Then, the inter-class scatter and intra-class scatter are defined for each training sample; Finally, the ratio of total inter-class scatter and total intra-class scatter is maximized to make the nearby samples with the same label are more compact, and nearby classes are better separated. Experiment results on ORL and FERET face databases verify the effectiveness of our proposed algorithm. 2. In this paper, a graph based semi-supervised learning algorithm named semi-supervised neighborhood discriminant analysis (SSNDA) is proposed. It combines the semi-supervised classification method and manifold learning method. This method firstly uses the self-training method to label the unlabeled samples under the guidance of a few labeled samples; Then, manifold-based learning method is adopted to extract features, so that the sample’s local geometric structures are better preserved, and the discriminant features are better extracted at the same time. More specifically, this method firstly uses the self-training method to label the unlabeled samples. In the process of self-training, the cut edge weight statistic method is employed to modify the self-labeled sample set before each iteration to reduce the influence from the noise samples, the mislabeled samples will be removed from the self-labeled sample set; Then, manifold-based learning method will be adopted for dimensional reduction: according to the labels of the self-labeled sample set and the similarity between each sample, different weights will be assigned for the edges in the neighborhood graph, so that the nearby samples with the same class label will be closer, meanwhile, the nearby samples with different class labels will be far away. Experimental results on the ORL face database and AR face database verify the effectiveness of the proposed algorithm.3. An improved classification method named fast sparse representation classification is proposed to solve the problem that the sparse representation based classification method faced:computational complexity. In the improved method, the idea of neighborhood class is adopted, so that the time that the non-neighborhood class samples are used to compute the sparse representation of the test sample are saved. More specifically, for each test sample, the most likely classes to which it would belong is firstly found; Then, only the training samples in these classes rather than all of the training samples will be used to compute the test sample’s sparse representation. The experiment results on ORL face database and FERET face database show that the proposed method not only largely reduces the computational complexity, but also increases the performance of recognization as the influence from the noise classes are eliminated.