Research On Linear Regression Based Feature Extraction And Classification

With the rapid development of the information industry, human biological feature recognition technology has become a hot research field. In much biometric identification technology, because their unique initiatives, non-invasive and user-friendly advantages, face recognition technology has been widespread concern in academia and business community. Face recognition system is divided into four parts:face image acquisition, face image preprocessing, face image feature extraction and classification. Among these four parts, the most significant part is feature extraction. Feature extraction is mainly aimed at reducing dimension. The basic idea of feature extraction is project original sample to a low dimensional space, so as to obtain the most essential information of original sample. Under the unremitting efforts of many experts and scholars at home and abroad, many classical algorithms have been proposed, such as principal component analysis and linear discriminant analysis are the most famous method. Since then, the sparse representation method has gradually been known and been widely used in the field of pattern recognition. The basic idea of sparse representation is use a series of known samples to represent most information of an unknown sample. But these algorithms have some disadvantages, that is, in practical work;image acquisition is easy to be disturbed by the external environment (such as light, shade, etc.). So how to reduce the noise interference in image acquisition as much as possible is particularly important. We have a deep research on feature extraction and classification based on linear regression.1. Transfer Linear Representation ClassificationTraditional sparse representation methods use the linear combination of training samples to represent test samples. It ignores the distribution difference between the training set and test set. Besides, it only considers the information contained in training samples and ignores the information contained in test samples. In order to solve this problem, this paper proposes a method of transfer linear representation classification method. This method effectively considers information contained in test sample and decreases the distribution difference between training set and test set. In addition, it enhances the "compatible" between training set and test set and improves the recognition performance of the algorithm. This method can be used as an image preprocessing method. Experimental results on GT, AR, Extended Yale B and CMU PIE databases demonstrate the effectiveness of this method.2. Mirror Image Based Robust Minimum Squared Error AlgorithmThe minimum squared error method has been widely used in computer vision field. This method aims to find a mapping which can predict the label of sample. Then we can use this mapping to predict the class of test sample. But this method has less robustness, that is, the minimum squared error method can not obtain good recognition results for samples whose pose and expression are changed. In addition, lack of sufficient training samples is a classical problem in pattern recognition. Thus, some scholars have put forward the idea of mirror image, which can effectively expand the training set. In this paper, based on shortcomings of minimum squared error algorithm and mirror image theory, a mirror image based robust minimum squared error algorithm is presented. Experiments on FERET, Extended Yale B, ORL and AR face databases show that the proposed algorithm has stronger robustness and recognition performance compared with original minimum squared error method.3. Two Stages Face Recognition Based On Nonnegative Representation CoefficientUsually, we use the classification performance and algorithm complexity to evaluate face recognition algorithms. Traditional sparse representation method can indeed achieve good recognition performance, but it has high time complexity and often need to observe distance between training samples from each subject and test samples. This is much complicated. In fact, we do not need to consider some of training samples which have little contribution to the linear representation of the test sample. According to this idea, this paper proposes two stages face recognition based on nonnegative representation coefficient. The main idea of this method is: firstly, we use the linear combination of all the training samples to represent test sample and constraint representation coefficients to be nonnegative, removing training samples whose coefficient are much smaller than others; secondly, we classify the remaining training samples, then we can get reconstructed samples through represent test sample as a linear combination of the remaining training samples. Aiming at how to classify the test sample, we propose two schemes. The first scheme is that we classify the test sample by measuring residual between the test sample and each kind of reconstruction sample. The second scheme is that we use these reconstructed samples to represent test sample and constraint representation coefficients to be nonnegative, then we calculate coefficient and assign the test sample into the class that produces the minimum representation residual. Experimental results on ORL, FERET and GT face databases show that this method is very effective and practical.4. l2-norm Reconstruction Discriminant Projection Based On Transfer Linear RepresentationIn face recognition, due to traditional sparse representation methods do not give a mechanism for noise removing, besides, when the number of extracted feature is very small, the classification performance of different feature extraction methods also exist significant differences. Thus, Yang et al proposed sparse representation classifier steered discriminant projections to get a high recognition rate with a small number of features when use sparse representation classification to recognize samples. Aiming at this method is more time consuming, Cui et al proposed a collaborative expression discriminant projection. But these methods usually used training samples represent test samples linearly. This is a one-way process and these algorithms ignore information in test sample. When distribution difference between training set and test set is very large, recognition rate of these methods aren't enough. Thus, this paper proposes l2-norm reconstruction discriminant projection based on transfer linear representation. This algorithm use transfer linear representation to enhance the "compatible" between training set and test set, then use collaborative expression discriminant projection to do classification. Experiments on GT, ORL and AR face databases demonstrate the effectiveness of the proposed method.