Face Recognition Reaserch Based On Improved Sparse Subspace Learning Algorithm

Face recognition is an important research direction in the field of pattern recognition and image processing.It has important theoretical research value and broad practical application prospect.At present,the research focus of face recognition is mainly concentrated on feature extraction and classification recognition.This paper mainly studied the face recognition feature extraction part.In the past few decades,the main method of feature extraction is subspace learning,various traditional subspace learning methods have been proposed which made the face recognition technology developed by leaps and bounds.The traditional subspace learning algorithms could be divided into two categories: supervised subspace learning methods and unsupervised subspace learning methods.Although the two categories have their own strengths,there still exists some problems that should be solved.In recent years,sparse subspace learning method based on sparse representation has achieved good recognition results and become an important research hotspot.However,the sparse subspace learning method belongs to the non-supervised subspace learning methods,it bear the problems which belong to non-supervised subspace learning methods.The global characteristics of sparse representation also caused some problems.The above problems of the sparse subspace learning method resulted in lower effectiveness.So the effectiveness of sparse subspace learning method should be improved urgently.Based on the related research at home and abroad,in this paper,the subspace learning method which based on sparse representation was improved and was expanded to semi-supervised subspace learning method.The main innovations and contents are as follows:1.The global characteristics of the traditional sparse subspace learning method would lead to some problems,such as the sample reconstruction errors and subspace learning errors.In order to solve the above problems,the sparsity neighboring correlation reconstruction method was proposed.By extracting the local structure information between all samples and sample label information of partial samples,the new reconstruction method could keep more accurate discriminant information and solve the problem of the traditional sparse subspace learning method.Based on the sparsity neighboring correlation reconstruction method,a subspace learning algorithm was proposed which is sparsity neighbouring preserving projections(SNPP).2.Although the SNPP algorithm could use label information of partial labeled samples,it can not extract the label information like MMC and other supervisor subspace learning methods,so it still belongs to unsupervised subspace learning method and its effectiveness is not high.In this paper,semi supervised technique was used to expand SNPP to semi-supervised subspace learning method.SNPP and MMC was combined to form a new algorithm which is the semi-supervised sparsity neighbouring preserving projection(SSNPP)algorithm.3.The performance of the SNPP algorithm and the SSNPP algorithm are tested in three standard face database: Extended Yale B,ORL and AR.The nearest neighbor classifier is used to classify the extracted features and compared the classification results with PCA,NPE,LPP,SPP algorithm.The experimental results show that the recognition accuracy of SNPP algorithm is significantly higher than that of PCA,NPE,LPP and SPP algorithm,which verified the validity and robustness of the sparsity neighboring correlation reconstruction method.In the three face database,the identifying rate of SSNPP achieved great improvement compared to the SNPP algorithm which proves the rationality of semi-supervised expansion.In addition,the experimental results also show that the sparsity neighboring correlation reconstruction method has good stability when extracting discriminant information.