Face Recognition Algorithms Research Under Varying Illumination Conditions

Face recognition system is used more and more widely,however,face recognition system still have some important issue to deal with,like the illumination issue.According to the FERET project evaluation(Face Recognition Technology Test)and Face Recognition FRVT supplier evaluation(Face Recognition Vendor Test)which organized by the US military,researchers find that the difference of face images caused by the illumination might be greater than the difference caused by different individuals' images.So the varying illumination is the important factor influencing face recognition,and the face recognition algorithm under varying illumination condition is the key point of our study.In the field of face recognition,the solution to overcome the varying illumination conditions are the methods based on image processing or the method based on the illumination invariant features.Our research is based on those methods,and the main content of this paper is as follows:1.The illumination normalization methods have been studied,such as gamma correction,logarithmic transformation,histogram equalization and homomorphic filtering method,then we applied those methods on the YaleB database,and obtain the Classification results by PCA method.Theoretical analysis and experimental study show that illumination normalization methods can reduce the influence of varying illumination,but under the condition of serious illumination disturbance,the improvement of face recognition is limited.2.A face recognition method is proposed based on local normalized method and LBP feature.The LBP descriptor lacks the ability of character description under serious varying illumination condition.In view of this flaw,we employ Local normalization method on the original images to reduce the influence of illumination,then extract classified texture feature by LBP descriptor.Experiments indicate that our method can improve the feature extract ability of LBP,and the face recognition rate.3.Verify and analysis three kind of Illumination invariant feature extraction method: MSR,SQI and WT,we hold experiments on Yale B and CMU-PIE database,then we analysis the hole effect caused by MSR method,and how SQI method improve the hole effect,at last,we analysis the influence caused by different wavelet bases.4.we proposed a new face recognition method based on Curvelet transform and Retinex theory,this method employ Curvelet transform to perform image's multiscale decomposition,then estimate illumination component in low frequency Curvelet coefficient and high frequency Curvelet coefficient by bilateral filter and threshold denoising separately,finally the illumination component of the original image is estimated by the revised Curvelet transform,and the illumination invariant feature is extracted by Retinex theory.Experiments on YaleB and CMU-PIE indicate that the method we proposed can achieve better performance than SSR,MSR,SQI and WT.