Face Recognition Research Under Complex Illumination And Partial Occlusion

With the development and industrialization of face recognition technology,the shortcomings of existing technology begin to be exposed.Complex illumination and occlusion are two important factors.They affect the performance of face recognition system.It is inevitable to optimize face recognition system performance in the condition of complex illumination and occlusion.The research on these problems have theoretical significances and values of transforming to products.In this thesis,face recognition is studied from two aspects: face images with complex illumination and face images with occlusion.1.A face recognition method based on Retinex theory is proposed.The applications of single scale Retinex algorithm,multi-scale Retinex algorithm and Weber's Law are discussed respectively.The adaptive illumination multiscale Retinex algorithm is proposed.In order to enhance face images under complex illumination,the basic idea is to divide the gray values of face images by Weber's Law in to different illumination regions,and then,according to the illumination characteristics of different regions,we use Retinex method to get the facial features images with different surround coefficients.The ratio coefficients of adjacent pixels are used when combining different illumination regions to avoid uneven images and color blocks.Finally,the experiments are carried out on the extended Yale B face database and the ORL face database.The experimental result verify that the algorithm can effectively enhance face features information retained in face images under complex illumination,thus the performance of face recognition system has been improved.2.Adaptive occlusion face recognition method is proposed.The Haar-like features and sparse classifier with occlusion dictionary are discussed respectively,and a face recognition algorithm can adaptively deal with occlusion is proposed.The algorithm uses Haar-like features,convolution network and sparse classifier to improve face recognition system.The basic idea is to search image by using the Haarlike feature and the Adboost algorithm to calculate whether the face is occluded.Secondly,using convolution neural network to extract face features.The best sparse classifier is selected by whether the face is occluded to optimizing the recognition rate.Through the experiments on extended Yale B face database and AR face database,it is proved that the proposed algorithm can effectively improve the recognition performance of face images in the case of occlusion,and the new algorithm also ensures better real-time performance.