Research On Illumination Invariant Feature Image Extraction Method Of Face Verification

Face verification technology has got great progress in the past three decades.Illumination is one of the main factors affecting the performance of face verification allthe time. The complexity and uncertainty of illumination environment make theillumination invariant face verification to be difficult. Illumination invariant featureimage extraction method is the main research method about face illumination problem,it can get better result and has no require about training samples. Anisotropic diffusionalgorithm is a classic illumination invariant feature image extraction method. Itsperformance is mainly decided by its image gradient descriptor and conduction function.To eliminate the illumination’s impact to face verification, an illumination invariantfeature image extraction method based on Weber local descriptor is proposed in thethesis. The proposed method introduced the Weber local descriptor into anisotropicdiffusion algorithm to improve the descriptive property of gradient descriptor andproposed a centre-symmetric logarithmic transform to preprocess face image to reducethe impact of sidelight to face image. The main research works in this thesis are asfollows:①Based on the survey of face illumination processing, an in-depth study has beencarried on illumination invariant feature methods while focusing on the anisotropicdiffusion algorithm. The theory of anisotropic diffusion algorithm and the main factorsaffecting its performance are analyzed in detail.②The image gradient descriptor has been researched further. We found that theexisting gradient descriptors are good at describing the grey changes of local image butthey cannot describe the changes relative to background, and the relative changes tendto be more able to reflect the change’s degree of local image. Hence, Weber localdescriptor is introduced into anisotropic diffusion algorithm to describe the relativechanges of local image. The ability of processing illumination by anisotropic diffusionalgorithm using Weber local descriptor and space gradient to describe the gradientchanges of face image is improved.③The gradient descriptor is sensitive to changes of local image, and it cannotdistinguish which factor caused the changes, by the facial feature or the illuminationchanges. Especially when sidelight impacting face, there are obvious dark and brightregions on face image. Dramatic changes of gradient descriptor will happen near the boundary between dark and bright regions. To solve the impact of sidelight, acentre-symmetric logarithmic transform is proposed to preprocess face image. Thetransform combining the merits of logarithmic transform and exponential transform canimprove the grayscale dynamic range of face image effectively. Using the transform topreprocess face image, and then embedding the anisotropic diffusion algorithm utilizingWeber local descriptor into the framework of generalized quotient image, in this way animproved illumination invariant feature image extraction method has been formed.④Visual feeling, cosine similarity, ROC curve and face verification rate have beenintroduced to evaluate the performance of illumination invariant face feature imageextraction methods in this thesis. The experiments are carried out on CMU PIE andCAS-PEAL databases about the four evaluating indicators. The comparisons of theexperimental results have verified the effectiveness of the proposed method.