With the rapid development of artificial intelligence and big data technology,people have higher requirements for work efficiency and information security.And biometric identification has been more widely used in public security systems,banks,customs and other fields.Especially,face recognition technology has attracted the attention of many researchers due to its uniqueness,friendly process and simple operation.However,stable recognition has certain difficulties,because the face image itself is easily affected by factors such as facial expressions,lighting changes,and occlusion interference.Sparse representation and dictionary learning algorithms are the important theories in the field of face recognition,which can overcome the above problems and hold good robustness.However,these methods also need to be further improved when applied to face recognition.Environmental disturbances in acquisition process of the face image may damage the image quality and quantity,making face recognition a small size sample problem.In addition,most sparse representation algorithms require known sparseness or the number of nearest neighbor when solving sparse coding problem,which limits the practical application of these algorithms.Therefore,better solving sparse coding and using a more reasonable way to alleviate small size sample problem have important research significance for sparse representation to improving the recognition rate.In this paper,after deep study and research on the face recognition algorithm based on sparse representation,two improvement algorithms are proposed to further solve the above problems.Firstly,a new algorithm,sparsity adaptive matching pursuit for face recognition,is proposed in this paper.The algorithm adaptively explores the optimal solution of sparse coding and the valid training samples by iterative updating,which can avoid artificially setting the sparseness.In addition,the algorithm uses a two-phase classification strategy to expand the difference in class probability distributions,and supervisedly narrows the range of candidate labels.It reduces the interference of noise from the training samples to a certain extent,thereby improving the face recognition rate.Secondly,a discriminative dictionary learning algorithm based on sample diversity and locality of atoms for face recognition is proposed.The face recognition problem in practice is a small size sample problem.Insufficient training samples may make it difficult for the dictionary to learn all valid features or be easily disturbed by noise,which may lead to insufficient discrimination or poor robustness.Thus,the algorithm generates virtual samples in a variety of ways,and at the same time adds the error model to guarantee the rationality of the virtual samples,which properly enriches the diversity of sample.In addition,this paper uses the Laplacian matrix of atom to keep the local structure information from face image,and improves the discrimination of the atom in training to make the learned dictionary more discriminative.Finally,the two improvement algorithms proposed in this paper are compared and verified on several face databases.Experimental results prove that the above algorithms can achieve higher face recognition rates. |