The Study Of Face Recognition Method By Low Rank Recovery And Sparse Representation

Face recognition technique is an important research direction in the pattern recognition field at present,and it is applied in lots of social fields for its advantages such as non-contact, easy to accept, difficult tofind, and higher recognition rate. Face recognition method based on image sparse representation is aresearch hotspot in the area of computation vision and pattern recognition at the present stage. Whereas,recognition rate of the present sparse representation algorithm is lower when the training sample imagescontain errors of illumination, expression, pose, occlusion etc, or contaminated by noise. Furthermore,under the circumstances of less training samples, there is outlier appearing easily when the test samples areclassified using classifier, thus the performance of sparse representation algorithm is degraded. When thesample images are contaminated by noises and there are less training samples, the present algorithm has alower recognition rate, in order to solve this problem, a mount of work has been done in this paper, and themain work is as follows:1) In the case of training sample images are contaminated by noise, SRC algorithm adopt unit matrix asthe error dictionary can not describe the noise and error of the image well, therefore, in this paperemploys Low Rank Matrix Recovery algorithm (LR) to decompose the training samples into a lowrank approximation matrix and a sparse error matrix, And then, the low-rank approximation matrixand the error matrix compose a over-completed dictionary to obtain the sparse representation of thetest sample in this dictionary. Based on the coefficient of the sparsest representation and theover-completed dictionary, reconstruction testing samples associated with special class, and calculatethe reconstruction error associated with special class. Finally, based on the reconstruction error associated with special class, complete the classification of the test sample. The advantage of LRalgorithm is that it not only can solve the problem of face image affected by noise contamination, butalso effectively solves the problem of small sample.2) Features of image representation using Gabor wavelet in different directions and scales is similar tothe related characteristics of human visual system, therefore, image features extracted by Gaborwavelet is more suitable for image representation. On the basis of getting error dictionary using LRalgorithm, in this paper we can get Gabor feature vectors by Gabor transform to process each trainingsample images and their corresponding error images respectively, a dictionary is formed with newGabor feature vectors, the dictionary can improve the sparse coding ability further of test samplewhich has been processed by Gabor wavelet. Similarly, based on the sparse coding coefficient andGabor dictionary reconstruction test samples, complete the classification of the test sample.3) Under the condition that both the training samples and test samples are contaminated by noise, in thispaper,we propose a graph regularized low-rank sparse representation recovery algorithm (GLRSRR),the algorithm can efficiently recover out a clean face image sets and a strong robustness errordictionary from sample image sets affected by noise pollution, this clean face images not only hasstronger discriminated information, but also can keep the local geometric structure of the original data.After obtaining the clean face images, deal with it and get the projection space exploiting PCAalgorithm, and then put the training sample images and the error dictionary to this projection space,construct a dictionary using the projection data, and calculate the sparse representation of test samplesunder this dictionary, similarly, based on the sparsest coefficient of the test samples and the dictionary,to complete the classification of the test sample.Finally, large numbers of experiments of the database of CMU PIE, Extend Yale B and AR show that the face recognition method proposed in this paper has higher recognition rate and strongeranti-interference ability.