Research On Feature Extraction And Face Recognition Based On Reconstruction

The original information from the model sample to extract the most effective information for the process of pattern classification is usually called the pattern feature extraction, feature extraction is a classical problem in pattern recognition. Therefore, the feature extraction process essentially can be seen as optimization problems in a certain criteria. At the same time, in the process of feature extraction, it can greatly reduce the dimension of the model. Therefore, feature extraction from the original mode of information not only in the best features for pattern classification, and these characteristics compared with the original sample information, greatly reducing the dimension model sample, so the feature extraction is an effective method for dimension reduction, this point in high dimensional pattern sample (such as face image) is the important recognition. Feature extraction of common criteria include two categories:Based on reconstruction and based on identification (classification). The former is oriented model representation, a typical representative of these methods are PCA and KPCA; the latter is the classification, the representative method is Fisher linear discriminant analysis method. Although feature extraction criteria to extract useful information for classification based on more, but is sensitive to noise; feature extraction method based on the criteria of the reconstruction in scalable and robust to noise than the feature extraction method based on authentication. This article through to the above two kinds of methods on the basis of comparative analysis, a thorough study of the theory and algorithm of feature extraction based on the criteria of the reconstruction, and verify the effectiveness of the proposed algorithm in more general face database.The main work is summarized as follows:1Design of the local feature reconstruction error discriminant projection method. Different from the minimum local characteristic error algorithm, the neighbor selection, select each sample k nearest neighbors, by minimizing the local within-class reconstruction error, local between-class reconstruction error while maximizing the correlation between class samples to reconstruct the generated to obtain the optimal projection axis. This method not only considers the local linear relationship between sample and extracted information discrimination, through the ORL library and YALE Library of the experiment shows that the method is effective, fast and robust.2Put forward the analysis method is proposed to the reconstruction error feature identification based on local correlation scale. By using the traditional k-nearest neighbor samples using Euclidean distance calculation of the distance, but the Euclidean distance of noise on the samples especially outlier sample (outliers) is more sensitive, do not have the robustness. The method to define the similarity between samples using scale correlation among samples, and used to calculate the similarity of each sample a minimum of similar samples and similar samples of different classes of the largest, followed by minimization of local within-class reconstruction error, while maximizing local reconstruction error to obtain the optimal projection axis. In the ORL face database and YALE face database verify the effectiveness of the experimental method of this chapter.3Design feature projection algorithm l2-norm least squares regression model based on the. Sparse representation method based on the l2-norm constraint is successfully applied to face recognition, but it also has an inherent defect is the computational complexity is too high. At the same time, recent studies show that the least squares regression model based on the l2-norm of the applications in face recognition with equal or even better recognition performance, and has higher computational efficiency. Therefore, this paper use norm least squares regression model is derived based on k-nearest neighbors of each sample, and then by minimizing the local within-class reconstruction error, while maximizing local reconstruction error to obtain the optimal projection axis. This method not only can extract more discriminative features, but also has higher computational efficiency.4.Put forward the feature extraction of2D image reconstruction based on sparse representation. The traditional sparse representation classifier is based on vector dimension, namely the need the original two-dimensional image is transformed into a one-dimensional vector, it has two disadvantages:One is the intrinsic geometric structure in the sample between pixels, two is added to increase the computational complexity, for example, vector dimension image transformation after the100*100is10000, the subsequent calculation must be carried out in such a high dimensional space. This paper proposes to construct sparse representation classifier using two-dimensional original image. The experimental results show that this method can not only improve the computational efficiency, but also greatly improve the classification performance.