Research Of Face Image Analysis And Recognition Methods Based On,Local Features

Automatic facial image analysis and recognition has very large theoretic and practical values. How to simulate human visual characteristics, and use computer technology to get the wanted information through face image analysis and recognition, is attracting a large number of researchers to study from multiple fields such as computer vision, artificial intelligence and pattern recognition. This paper aims at several typical problems of face image analysis and recognition technology, such as face detection and location, face recognition, and facial expression recognition. The main research contents are as follows:In chapter 1, the related background and significance of the research are introduced. An overview of the basic components of face image analysis and recognition, as well as some research applications at home and abroad are interpretated. The framework of face recognition and facial expression recognition are analyzed, highlighting the typical methods of face detection, face recognition and facial expression recognition at present. The problems and development trends related to research are pointed out, and the main research contents of this thesis are also introduced.In chapter 2, aiming at the face detection and location problem in color images, a face detection algorithm combining skin Gaussian model in YCbCr color space and AdaBoost is proposed. First we discuss the characteristics of various color spaces. Then in YCbCr space we establish a single Gaussian model for Cb-Cr distribution of facial skin color through the statistical analysis of collected skin samples, and the candidate face regions are segmented from background by calculating the skin color similarity. We do some research on AdaBoost method which is on basis of Haar-like rectangle features and iterative training, and use it for precise positioning within the face candidate regions. Better results are achieved while the calculation amount is reduced significantly. Thus we realize fast and precise face detection by the combination of color information and local features.In chapter 3, aiming at the reseach of Scale Invariant Feature Transform (SIFT) for face recognition, a "local-holistic" face representation method based on SIFT and Space Pyramid Representation (SPR) is proposed. First we study the extraction of SIFT features for face images. In order to solve the problems of cardinality variance, spatial uncertainty and lack of meaningful order with SIFT features, which result in the difficulty of forming a global description vector for the face image, we propose to extend the spatial pyramid matching algorithm to SPR. The SIFT features are mapped into spatial and feature domain simultaneously, and then histograms of different levels are constructed to form a "local-holistic" representation for a face image. Finally classifiers such as Nearest Neighbor (NN) and Support Vector Machine (SVM) could be performed on this representation vector of equal length for face recognition. The experiments on the international generic ORL as well as Yale face databases verify the effectiveness of the proposed method.In chapter 4, aiming at the general problem of illumination change in face recognition, and started from the extraction of illumination invariants, a face recognition method based on the "Gradient-face" and local feature encoding is proposed. First the face image is transformed into gradient field and we discuss the theoretical basis of the two illumination invariant gradient components, Gradient-face magnitude (GFM) and phase (GFO). Then according to the characteristics of each component, LBP and LXP operators are introduced to encode local features from GFM and GFO at a higher level, which further reinforces the invariance against lighting changes, and increases the distinguish ability of features. Each GFM is transformed into a Local Gradient-face Binary Pattern (LGBP) map, while each GFO is transformed into a Local Gradient-face XOR Pattern (LGXP) map. Using statistics histogram technique, a weighted connection method at feature level is proposed, which further integrates illumination invariant characteristics of LGXBP for identification. Finally the Nearest Neighbor classifier is adopted for face recognition. The comparable experimental results on the generic Yale B as well as CMU-PIE lighting-change face databases confirm the validity of this method.In chapter 5, research of facial expression recognition on the basis of local features is investigated. A method based on local feature integration of "spatial-frequency" domain and automatically key sub-regions selection is proposed to get the critical LSPBP features for each expression. The principle analysis of LBP and Local Phase Quantization (LPQ) demonstrates the relevance and complementarity of these two efficient local features, and we propose a strategy to fuse them at feature level. Then AdaBoost training is introduced for automatic selecting the most critical sub-regions with high distinguish ability for different expressions. Finally, the learned AdaBoost strong classifier, as well as SVM classifier are utilized for the recognition of six basic facial expressions plus neutral expression. Experimental results on the JAFFE facial expression database are given to verify the effectiveness of the proposed algorithm.In chapter 6, a face analysis and recognition system based on single static image is built. The system can automatically complete face detection, face recognition and basic facial expression recognition for natural scenes pictures, which verifies the effectiveness of the above algorithms and creates a good platform for future experimental research.In chapter 7, the major work of this thesis is summarized. The conclusions and innovations of this thesis are introduced. Finally, some future development topics are presented in order to provide guidance for researchers, who are interested in such kind of projects.