Study And Implementation Of Face Recognition Methods

Face recognition is one of the challenging subjects in the areas of pattern recognition and computer vision, which has a wide range of potential applications in public security, identification of certificate, entrance control and video surveillance. Although it is easy for most human observers to identify different human faces, automatic face recognition is a very difficult problem. This dissertation presents the studies of feature extraction and the design method of classifier. The main points are as follows.Based on algebraic features of the images, this paper first introduces the PCA-based (principal component analysis) face recognition algorithm. The conventional PCA for image feature extraction is usually based on vectors, which makes it very time-consuming. To overcome the drawback of PCA, a novel and efficient PCA method based on original image matrices directly, termed image PCA (MPCA) is proposed. Experimental results show the proposed method is more efficient than the classical Eigenfaces method.Secondly, in order to eliminate some illumination or expression effects in PCA-based algorithm, this thesis introduces linear discriminant analysis (LDA) algorithm based on Fisher criteria. The primary purpose of LDA is to separate samples of distinct groups by maximizing the between-class scatter while minimizing the within-class scatter. Consequently, the performance of LDA is superior to the classical PCA-based algorithm. There are at least two critical drawbacks in the traditional LDA-based methods: the small sample size (SSS) problem; the Fisher criterion is not optimal with respect to minimizing the classification error rate. A novel LDA-based technique is proposed in this paper, which could effectively deal with the two problems. First, the Fisher criterion is redefined by introducing a weighting function of the contributions of individual class pairs to the overall criterion. Then, the weighted Fisher criterion is optimized under the conjugated orthogonal constrain, which can guarantee the derived projection directions are statistically uncorrelated. Experimental results show the effectiveness of the proposed algorithm and its insensitivity to the variants of face expression and illumination.At last, in order to integrate face features better for efficient classification, a new approach for face recognition using an embedded Hidden Markov Model (E-HMM) is proposed. This model inherits the partial size invariance of the standard HMM, and, due to its pseudo two-dimensional structure, is able to model two-dimensional data such as face images, better than the standard HMM. The observation vectors used to characterize the states of the E-HMM are obtained by using the coefficients of two-dimensional Discrete Cosine Transform (2D-DCT). Experimental results show the proposed model has better performance than previous HMM-based methods, with reduced computational complexity.