Research On Face Recognition Based On Multi-features Representations

Nowadays,personal information security has been highly required and some traditional security measures such as passwords and keys have no longer met people's needs.As a result arises the biometric identification technology,which based on the structural feature information of a human body,consists of fingerprint recognition,voice recognition,facial recognition,iris recognition,retina recognition,and face recognition,etc.In recent years,the face recognition technology has become one of the most concerned research directions.Due to the uniqueness of an individual face,the technology has been an important tool to identify individual identity,and widely used in various fields.The widely-used face recognition technology is still facing a whole lot of challenges,such as the change of light,different degrees of occlusion,rich facial expressions,and other factors that will affect the result of face recognition.Consequently,how to better represent the face image and improve its recognition accuracy has been a hot issue in the academic research.Based on that,this paper aims at the representation method of face image and multi-feature fusion,and makes a profound research.The main research contents are contained in the following aspects:(1)In view of the fact that under the circumstance of a traditional single feature there is a lack of comprehensive description of sample information for the sparse classification method based on the error representation,this paper proposes a face recognition method using multiple features.Firstly,2 Dimensional Principal Component Analysis(2DPCA)method is used to extract the features of face image;secondly,Fast Fourier Transform(FFT)method is used to extract the spectrum features of face image;next,the regress ion error of the generated image features,spectrum features and original image under the collaborative sparse representation is calculated respectively as the classification score;finally,a new fusion mechanism is used to fuse the scores of the three classifications,and as the discrimination basis of the final category to improve the recognition accuracy.The experimental results show that the recognition performance of this method functioning in various datasets is better than that of the classic algorithm.(2)In order to solve the problem of insufficient training samples in face datasets with small samples,a virtual sample generation method based on closest neighboring samples is proposed by using the inherent symmetry of face image.This method designs corresponding segmentation and reorganization schemes for different face images of the same object to generate multiple virtual samples of the same kind.The advantages of this method are: On the one hand,the newly generated virtual sample set can not only expand the training samples,but also ensure that the training sample set used to represent the test samples contains sufficient useful information;on the other hand,the subspace features of all face images of each object can be distributed in the same subspace.After the sample expansion,the test samples are classified by the method of grouping sparse representation.A large number of experiments show that compared with the traditional method,this method achieves better classification effects on multiple open face datasets.(3)To solve the problem of insufficient training samples,a virtual sample generation method based on the residual samples is proposed to increase the number of training samples.In order to reduce the influence of noise on regression error,the method uses the residual images of the same kind of original training samples and their mean samples to generate virtual samples.This method effectively uses the intra-class difference as the data representation of part of noise,and simultaneously uses the intra-class difference and training samples to render the test sample a linear representation.Moreover,in the classification decision-making stage,the weighted score fusion strategy is used to integrate the results of virtual samples and original training samples as the final classification basis.A large number of experiments on multiple face datasets show that this method can effectively improve the accuracy of face recognition compared with the traditional algorithm.