Study On Sparse Representation Based Classification For Face Recognition

Face recognition has become one of the most popular biometric recognition methods because of its rich facial information and wide application prospects.However,the quality of the image taken by the equipment is often affected in the real environment.Various faeial expressions,gestures and illumination conditions will affect the quality of face images,resulting in occlusion,translation and scale errors in normalized face images,which will reduce the robustness and recognition accuracy of face recognition algorithms.Among many known algorithms,Sparse Representation based Classification(SRC)algorithm has achieved amazing good performance in face recognition.The algorithm is robust to noise and partial occlusion,but it is still challenging to face recognition in the case of facial expression change,posture change and small sample size.Face recognition based on SRC consists of the following three steps:initializing the over-complete dictionary,solving the sparse reconstruction coefficients and discriminating criteria.This paper mainly improves the face recognition algorithm based on SRC from two aspects:the initialization of a complete dictionary and the solution of sparse reconstruction coefficients,so that the improved algorithm has better robustness in the case of image contamination and small samples.The main research contents are as follows:1)Face recognition is seriously disturbed by the quality of samples collected due to illumination,occlusion and expression changes.Based on this,this paper presents an efficient face recognition algorithm based on discriminant nonconvex low-rank matrix decomposition with superposed linear sparse representation.This algorithm efficiently eliminates the errors caused by occlusion and other unavoidable factors.Due to two major limitations of Robust Principal Component Analysis(RPCA):without structural incoherence and shrinking all the singular values equally,we utilize non-convex rank approximation(? norm)to replace nuclear norm.It overcomes the problem that the singular value of the matrix is scaled to the same multiple when solving the kernel norm by traditional RPCA method,which may lead to errors in the recognition results.In order to minimize the between-class scatter,we add to the theory of structural irrelevance in low-rank decomposition at the same time.And this method increases the incoherence among the low rank dictionaries,and thus improves the discrimination ability of low rank matrices.After getting the low-rank matrix,we divide into prototype dictionary and variation dictionary according to superposed linear.Then the two dictionaries are combined as the training dictionary in SRC.Finally,the classification is completed by Superposed Linear Sparse Representation Classification(SLRC).This paper eliminates the interference of intra-class correlation and inter-class correlation,and it is also universal on the under-sampled database.This study select AR database and CMU PIE database for experiments.In the AR database,the recognition rate of our algorithm is 98.67±0.57%in 10 experiments,which is better than SRC,ESRC,RPCA+SRC,LRSI,SLRC-l1 and so on.In addition,we choose different proportion of occluded pictures with the scarf or the sunglasses from 0 to 3/7.This means that the number of sunglasses or scarf increased from 0 to 3 and the total number of train images per class is 7.Compared to other algorithms,our algorithm has better robustness to occluded images and has higher recognition rate.As for CMU PIE database,we add salt and pepper noise from 0 to 40%in every image.The recognition rate of our algorithm reaches 90.1%?85.5%,77.8%,65.3%and 46.1%respectively and is the highest in the contrast algorithm.The algorithm proposed in this paper has a high recognition rate in different face databases,especially in the case of occlusion and noise pollution.And it can also maintain good robustness and efficiency,so it has a good application value in practice.2)Because of the limitation of sparse representation algorithm in the case of few samples and single samples,a face recognition algorithm based on semi-supervised Gauss Mixture Model(GMM)and SLRC is proposed in this paper.Both training and testing samples are considered in the process of constructing the prototype dictionary.The prototype dictionary is optimized by using the GMM and mutation dictionary to eliminate the non-linear noise of the face image and obtain more accurate centroid matrix.Specifically,all images are considered to obey the Gauss distribution.Then the improved semi-supervised expectation maximization(EM)algorithm is used to optimize the parameters of the Gauss mixture model.Its mean value can be used as the prototype Dictionary of training samples.Finally,the optimized prototype dictionary and mutation dictionary are used to superimpose the linear sparse representation.Face recognition.Because this algorithm makes full use of the image information of unlabeled samples/test samples to make the prototype dictionary more accurate,it has better robustness in the case of few samples or single samples.This conclusion can also be proved by AR face database and CAS-PEAL face database.In summary,this paper studies the face recognition algorithm based on superimposed linear sparse representation from two aspects:low rank matrix decomposition and Gaussian mixture model.The experimental results show that the proposed algorithm has good robustness for face recognition and has certain practical application value.