Application And Research Of Feature Extraction And Pattern Classification On Face Recognition

From the narrow sense, face recognition is regarded as the process of extracting and detecting the features from the given human face, and then been compared to the known classes of samples, which is mainly related to two aspects of the feature extraction and classification. Feature extraction and classification are both the key techniques for face recognition, which directly impact on the performance of the entire system. However, the data sets collected from the real application problems are always with the characters of high-dimension. If processing on the original image directly, the algorithm has very high complexity, and the computer hardware performance is also a challenge, so in the field of face recognition feature extraction is the key to solve the problem. And the classifier should be selected according to the extracted features to get better performance.At present, face recognition as the target, a variety of relevant technical methods have been successfully applied to various fields, but it still faces many challenges. Our work is focusing on the feature detection and classification on the face image recognition, on this basis we put forward some more effective feature extraction methods and classification methods. Compared with the current mainstream methods, we verified the validity of our methods. Except the gray face image recognition, we also research on the color face image recognition. The main research works and contributions of this thesis are summarized as follows:(1) Inspired by sparse representation and dictionary learning, a dictionary learning based kernel sparse representation classification method is presented for face recognition.Dictionary learning and sparse representation have attracted many researchers. Inspired by Metafaces, a dictionary learning based on kernel sparse classification method is presented for face recognition. First, the kernel dictionary bases are learned based on Metafaces framework. Second, a kernel sparse representation classifier is proposed by extending sparse representation classifier to high dimensional space via kernel functions. At last, we reconstruct the samples by kernel dictionary and classify the face image according to the residual. The experimental results on AR, ORL and Yale face databases show that the proposed method works well.(2) Based on Gabor feature extraction method and linear regression classification method, we propose the method of Gabor kernel linear regression classification (GKLRC) for automatic face recognition. Gabor features are very effective in face recognition, which is robust to illumination and expression variations and has been successfully used in face recognition area. Our paper addresses the method of Gabor kernel linear regression (GKLR) for automatic face recognition. Based on the assumption that the patterns from a single-object class lie on a linear subspace, linear regression (LR) for classification is proposed and it is a regularized least-squares method to develop the linear dependency between a probe image and class-specific galleries. We employ the kernel technique and Gabor features in the linear regression classification to perform analysis in a high-dimensional feature space to extract the significant nonlinear Gabor features which have been widely used in pattern recognition. We demonstrate the Gabor kernel linear regression method on the face database. Experimental results on several standard face databases demonstrate that the proposed algorithm significantly outperforms LRC and so on. Therefore, the new method should be considered in recognition of face.(3) We have learned the Deconvolutional filters and the new method of face recognition with minimal convolution error measure oriented dictionary learning was proposed.Face images are modeled using a deconvolutional decomposition of images under a sparsity constraint, which can capture the mid-level cues spontaneously arising from image data. We also introduce a well learned dictionary matrix to achieve better FR performance with less dictionary atoms. So minimal convolution error oriented dictionary learning is proposed in this paper and the experimental results on face image databases demonstrated the effectiveness of our method.(4) We studied on different color spaces, and proposed the method of canonical correlation analysis on multiple color space.As opposed to gray face images, the color information of face images can be applied to improve the face image recognition rate. Canonical correlation analysis on multiple color spaces was proposed in this paper. First, the source images were transformed into contourlet domain and get the LP and HP images. Then we used CCA to get the projection matrix. At last, nearest neighbor classifier is selected to perform face recognition on multiple color spaces. Experimental results on color AR face database show that the proposed algorithm is more effective and faster than the method of Daubiches transformation.