Research On Methods Of Facial Image Analysis And Recognition

Automatic facial image analysis has very large theoretic and practical values. Topics in automatic facial image analysis include face recognition, expression recognition, mouse shape recognition, gaze tracking, facial image coding and facial image synthesis. The main content of this dissertation can be summarized as follow: 1.A global image feature extraction method based on sparse coding is proposed to extract the feature of human facial images. Based on unsupervised learning, sparse coding is suitable to describe images with non-Gaussian distribution and can get rid of the high order redundancy among the image pixels. Since the basis function of sparse coding has build-in clustering property, it increases the inter-class variations of the features. Thus, the recognition performance of sparse coding is better than traditional eigenface-based methods. 2.Based on the energy distribution of facial image, a DCT-based sparse coding method is proposed to reduce the computational complexity of the original sparse coding. Based on the clustering property of the basis function of sparse coding, a basis function initialization method using fuzzy C mean algorithm is proposed to help the energy function of sparse coding to converge to a better local minimum for recognition. Experimental results show that the classification and the sparseness of the features are both improved. 3.Since traditional sparse coding does not consider the intra-class variations of the features, two new sparse coding algorithms based on reinforcement learning are proposed to increase the classification property of the features. The class labels of the training samples are introduced during the training of the basis function to constrain the intra-class variations of the features. The features produced by the new sparse coding have large inter-class variations and small intra-class variations, thus the recognition performance of the reinforcement learning based sparse coding is better than that of traditional sparse coding. 4.An ICA and 2D EHMM based method is proposed to extract the feature of facial expressions. ICA is used to produce the independent observation vector of EHMM. Since EHMM can keep the states unchanged for a given range of the change of observation vector, the algorithm can effectively resist the impact of the shifting II error of the fiducial points in the face. 5.Developed a real time face recognition system, which can automatically perform face detection, facial feature extraction and recognition.