Research On Face Recognition Algorithm Under Complex Condition

Face recognition technology refers to the face discrimination and classification based on the application of collected face image information combined with the face image in the dictionary database,and then to achieve identity authentication.As one of the important branches of artificial intelligence applied to real life in the new era,this technology not only facilitates people's life,but also promotes the development of related industries.In practical applications,when the face image is influenced by illumination,noise,occlusion and other complex factors with the conditions of insufficient sampling,the recognition accuracy will be significantly decreased.Focusing on the face recognition algorithm based on sparse representation,this thesis focuses on the face recognition problem under the condition of insufficient sampling.The specific research contents are as follows:1.In view of the problem of lower recognition rate and lower efficiency of the traditional face recognition algorithms under the condition of insufficient samples,this thesis proposes a novel face recognition algorithm by utilizing l_p norm?0<p<1?and fusion dictionary.The training sample is decomposed into a fusion dictionary composed of a class-centered matrix and an intra-class variation matrix to increase the completeness of the dictionary under the condition of insufficient sampling in this thesis.Then the sparse representation of the test sample under the fusion dictionary is regularized by l_p norm to improve the sparsity and the solving speed of the solution.Extensive experiments on the benchmark databases,including Extended Yale B,ORL,AR and CMU PIE,show that the proposed algorithm runs fast and has high recognition accuracy.2.To improve the face recognition accuracy affected by complex factors,such as illumination and outliers under the condition of insufficient sampling,this thesis puts forward an adaptive robust sparse representation algorithm based on weighted and fusion dictionary.Primarily,the training sample is structured into a fusion dictionary.Then the weighted ₁l norm constraint is introduced to constraint to the outliers in the image on the basis of the fusion dictionary.Moreover,a low rank regularization is applied to coordinate the sparsity and the correlation between different images.Extensive experiments on the benchmark databases,including AR,ORL,Yale and CMU PIE,show that the proposed algorithm has strong stability and high recognition accuracy.