Face Detection And Recognition Across Illumination

A robust face detection and recognition system should have the ability to dealwith the variance of the images caused by facial expressions, illumination conditions,poses, and ages. Our paper is mainly focused in the illumination variance. Some newmethods are proposed by means of image enhancement and construction ofillumination subspace respectively.The requirement for the surface normal of every pixel on the surface limits theapplication of spherical harmonic theory. Based on the planar model, we deduce alinear method to estimate the reflected light of the subject. To adjust the illuminationconditions of the images, we propose the inverse image to compensate the originalimage where the confidence matrix is determined by the illumination conditions of theimages. We prove the method in theory and test it in detection experiments as well.According to the reflection theory, the BRDF of Lambertian surface can beconsidered as the low-pass filter for the incident light. Thus the reflected light ismainly made up of low-frequency information. However, there is always somehigh-frequency information in the face images, such as the edges of the shadows, dueto the non-convex face or the non-continued light source. The wiener filter is appliedto extract the information from images to approximate the illumination, which cansmooth the low-frequency information while keep the high-frequency parts. Theinverse image and original image are compensated for each other in the procedure ofrendering image. The face detection experiments show the effectiveness of themethod.According to the illumination of the environment, the human vision system canadjust itself adaptively in order to get a better image. Based on the adaptation ofhuman vision system, we deduce a global adjustment function to adjust theillumination conditions of images. The parameter of the function is proposed to bedecided by the intensity distribution of the images. That is the global adjustmentfunction can adjust the images adaptively to its illumination conditions. For the underexposure image, the function can enhance its contrast and brightness; for the overexposure image, it can enlarge the contrast and suppress the absolute intensity valueof every pixel. As a result, the global adjustment function can alleviate the influence of the illumination.Due to the reason that global adjustment function may make some details lost, weimprove the global adjustment function with combining the local and globalinformation in the function and give a detailed adjustment function. We propose tooptimize the parameters based on the evaluation of the image quality, where theintensity entropy is applied to evaluate the image. The degree of adjustment isdetermined by the local conditions of every pixel and the global conditions of theimage, thus the details can be kept well. The entropy value of the image rendered withthe detailed adjustment function will increased. Also the appearance of the image willbe rich of details and with the proper brightness as well. Consequently the image willbecome more discriminative.The base images of 9 point light (9PL) subspace can be easily gotten in the realenvironment. We propose a Maximum A Posterior (MAP) estimation based algorithmto recover the base images of 9PL subspace from a single image (9PL-MAP).Compared with other methods based on statistics, the proposed method do not requirefor complex training samples, that is it can be practiced more easily. And it canreconstruct the 9PL subspace of a subject through an image taken under arbitraryillumination conditions, which shows its good generalization. In the recognitionexperiments, we probe the proposed algorithm on different standard face databases.The recovered 9PL subspace by the proposed algorithm shows its ability to representthe illumination variance of a subject under different illumination conditions.The image enhanced methods (including the linear estimation & compensationmethod, wiener filter estimation & compensation, global adjustment function anddetailed adjustment function) do not need any prior knowledge. All neededinformation used for adjustment can be extracted from images themselves. Althoughthe reconstruction of illumination subspace needs some training samples, it can givean exact description for the illumination variance. In practice, different methods canbe chosen for different environments.