Face Recognition Based On Weighted Kernel Sparse Representation

Face recognition research has attracted people's attention in recent years.However,current face recognition has poor recognition rate in non-limiting environments such as occlusion and lossy images,we still need to further study and improve recognition performance.Compared with other traditional recognition methods,sparse representation classification algorithm uses overcomplete dictionary to sparsely express data signals,and then classifies sparse representation coefficients by residuals.This method has a good classification effect in the partial occlusion environment of the face image.However,the traditional sparse representation has the problems of high computational complexity,large amount of calculation,low recognition rate of small sample data sets and linear inseparability in the high data dimension.The following improvements have been made in this thesis for these problems.The main research includes:Face recognition based on PCA and sparse representation classification is used to improve the computation speed.The algorithm first uses the principal component analysis algorithm to reduce the dimensions in face images data set.Then obtains the linear combination coefficient of the sparse representation by solving the minimum L₁norm of the reduced dimension data.At last,using the residuals of the test data and the training data to solve the classification result.Experiments show his method improves the recognition speed and recognition rate of the sparse representation algorithm.A face recognition method based on weighted nuclear sparse representation is proposed.For the nonlinear distribution of samples,the original samples are mapped to the high-dimensional kernel feature space,and the PCA dimension reduction method is used to reduce the computational complexity.A multi-scale Retinex is used to obtain a matrix representing the degree of similarity between the test sample and the training sample.The sparse coefficients of the test sample are solved by an optimization method,and the classification result is obtained by minimizing the error between the original sample and the reconstructed sample.The method presented in this thesis not only improves the computational complexity of traditional SRC methods,but also optimizes face recognition performance under non-limiting conditions such as illumination and occlusion.The test results on Yale B database and AR face database show that the algorithm can obtain high recognition rate and has good robustness.The maximum face recognition rate can reach 92.43%.