Research On Illumination Preprocessing Methods For Face Recognition

Face recognition,as a popular research in the field of biometric recognition,is also one of the most successful application in the field of computer vision.For the access control system,intelligent security,intelligent monitoring and national military and security fields,it plays an irreplaceable role and has significant potential.Through near 50 years of research,this technology has made great progress,and has been put into commercial use for a long time.However,there are still some problems.For example,the recognition performance may be influenced and will reduce rapidly if the system suffers the poor light condition or non-compliant user.So this dissertation first analyses the influence of illumination conditions on face recognition and gives a thorough research and then proposes two improved methods.This dissertation makes a detailed research on some basic image processing methods and the methods based on lambert model,and proposes two improved methods based on the theoretical analysis results.Firstly a self-quotient image methods based on Wiener filter is proposed.The self-quotient image methods can overcome the limitation of the traditional quotient image and can extract the inherent characteristics which is independent on the illumination,but it estimates the illumination component by using Gaussian filter which can smooth the face image and blurs the contour features at the same time.So the method proposed in this dissertation estimates the illumination component based on Wiener filter,and it can adjust the filtering intensity adaptively according to the local variance of the face image so as to improve the accuracy of face recognition effectively.Secondly the relative gradient methods based on multiple directions is proposed.Traditional gradient methods consider only the directions of X and Y,but the method proposed in this work considers four directions and combines the lambert model,and uses the Gaussian first derivative to calculate image convolution results.The weighted fusion of the relative gradient component of each direction is obtained so as to gain a better expression of the essential feature from face image.Experiments on two face illumination databases,including the Extended Yale B and CMU PIE database,have been conducted to test the methods proposed in this dissertation.They are compared with multiple illumination methods,and the nearest neighbor is used as the classification criterion.A large number of experiments show that the methods proposed in this dissertation can effectively obtain the illumination invariant component so as to improve the accuracy of face recognition.