A Research Of Face Feature Extraction And Recognition Algorithms

Face identification is one of the important fields and hot research topics in pattern recognition. It involves a very wide area. Human face images are highly variable depending on influence factors such as circumstances and expressions, which make face identification a challenging and complex research task. Many problems and key technologies need to be further addressed and improved including the completeness of the feature extraction stage, that is, taking a full consideration of local and global features and image invariance under translation, stretching and rotation; the performance in the classification and identification stage with a high-precision recognition rate and fast classification algorithm.The focus of this thesis is the recognition algorithm of the facial feature extraction and classifier design for human face identification. New methods of feature extraction and recognition algorithms of classification and recognition are proposed. The results based on the large number of experiments carried out on the ORL and YALE face database show that the new methods and recognition algorithms enhance the performance in the rate and speed of recognition.The main work and contributions of this thesis are as follows:1) Considering the rate and speed of the wavelet recognition algorithms, a large number of experiments have been carried out to compare the performance of face recognition with the different choice of wavelet bases, the determination of the number of decomposition levels, and the selection of decomposition coefficients.2) The concept of block wavelet coefficients is given. A new face recognition algorithm base on the blocking wavelet transform is proposed. Combining the PCA+LDA methods, the algorithm reduced the characteristics dimension. Further more, in order to sparser eigenvector, this new face recognition algorithm is improved using the best local characteristics of curvelet transform. The experiments show that, compared with the traditional wavelet transform, the recognition rate is improved by at least more than1%on YALE, especially the recognition time is reduced by nearly50%on ORL. 3) A new face recognition algorithm of contourlet transform based on wavelet domain is proposed. Because of multi-scale analysis and direction analysis separately of contourlet transform, the combination of two transform can overcome the shortcomings of wavelet transform such as poor directional selectivity and unsuitable for describing the singularity structural characteristics such as the edge and the contour line of an image, so it is more conductive to extract and analyze the feature.4) Inspired by the embedded zero-tree coding, the concept of quad-tree of characteristics, the method of constructing quad-tree of characteristics for face image are proposed, and the sparse wavelet moment is obtained by combining the sparsity of the characteristics quad-tree vector and the rotational invariance of the wavelet moment. Moreover, an improved sparse wavelet moment algorithm based on quad-tree of characteristics for face recognition is proposed. Experiments show the good performance of the new algorithm.5) By the detailed analysis and study of the several classifiers. A new idea of classification named category characteristic method (CCM) is presented. The random forest has been put forward for the first time in face recognition. Two new classifier combination algorithms are proposed:firstly, SVM algorithm based on the KNN filter, secondly, optimal fusion classifier combination algorithm (KNN-CCM-SVM). A larger number of experiments show that the new algorithms have good ability to identify the face image.