Research On Discriminant Subspace Analysis And Application Based On Low-rank Representation

As human society moves towards the digital age,and with the development of Internet technology and computer technology,face recognition is one of the research hotspots in the field of pattern recognition,it has attracted the attention of domestic researchers in many biometric identification technologies.Because of its unique reliability,hard to replicate and initiative.After many years of exploration and research by many researchers at home and abroad,the face recognition has made a lot of important breakthroughs in technology and achieved many remarkable results.But in practical applications,face recognition technology still faces some challenges,for example,the most typical problem is that under the uncontrolled conditions,the variability of a face(change of expression),instability(variation of illumination)and difference(camouflage,stain)can affect the effect of face recognition.Therefore,how to effectively avoid or reduce the problems of the above face recognition technology is the key to improve the face recognition rate.In recent years,it has been shown that as a subspace segmentation method,low rank representation is not only more robust to noise,but also can better maintain the global structure of data.It is widely used in machine learning and computer vision.At the same time,the low rank representation also has some shortcomings and needs to be improved.In this paper,the low-rank representation learning as the core of face recognition algorithm is studied.And the main work is summarized as follows:1?Face Recognition Based on Discriminative Low-rank Matrix Recovery with Sparse Constraint.In consideration of the problem that the existing face recognition methods cannot handle the face recognition under unsatisfactory situations,such as shadows,occlusions,stains,which cause low recognition rate.Therefore,an algorithm based on discriminative low-rank matrix recovery with sparse constraint(DLRRSC)is proposed.First,A clean dictionary A and an error dictionary E*are recovered by using the method of discriminating low rank representation,and then it learns a low-rank projection matrix to correct the corrupted testing sample by projecting the sample onto its corresponding underlying subspace.Finally,the sparse representation method is used to classify the testing sample.Comparative experiments made on Yale B and AR Databases show that the performance of the method is better than other face recognition methods.2?Face Recognition and Analysis Method Based on Weighted Low-rank Representation and Adaptive Nearest Neighbor Selection.Considering the complicated problem of the discriminant projection method based on graph embedding must to select neighbor parameter k and insufficient existing sample category information,etc.,We propose a face recognition analysis method based on weighted low-rank representation and adaptive nearest neighbor selection.First,based on semi-supervised discriminant analysis(SDA)and ANSDA proposed by Shi Jun,In our method,using the unique composition of ANSDA,that is to say,we uses all the within-class samples.to construct the within-class graph and then adaptively chooses the between-class samples to construct the between-class graph.The problem of nearest neighbor parameter K selection is avoided effectively.On this basis,we use a regularization term by weighted low-rank represented to maintain the global similarity structure of samples.Finally,we carry out the experiments on FERET and yale_faceb databases,and compare this method with the traditional dimensionality reduction methods and the results demonstrate that ANSWLR-SDA method is effectiveness and robust to different types of noise than other state-of-art face recognition method.3?A Face Recognition Algorithm Based on Graph Regularized Low-rank Representation with Spatial Constraint.Considering the situation that the dictionary is incomplete in the face recognition,and the damage problem of the training samples caused by different facial expression changes,illumination and noise.In this paper,a more effective face recognition algorithm-SGRLRR_CRC is proposed.First,we using the given training samples to recovery the clean face dictionary A and error dictionary E*by the Graph Regularized Low-rank Representation with Spatial Constraint algorithm.And then based on the clean dictionary A,to the test samples y for collaborative representation.In the end,to demonstrate the effectiveness of the presented algorithm,and aiming at the changes of the training samples and dimensions,and we carried out with different levels of noise on samples,our comparative experiments were conducted using CMU-PIE and ORL face databases.The result shows that the effectiveness and robustness to noises are always than the SRC,CRC,LRR and GLRR methods.