Study On Illumination Problem In Face Recognition

The target of research on face recognition is to endow computer with the ability of identifying people according faces. Face recognition has made great progress after more than forty years of research, but there still exist many unsatisfactory aspects. It needs more research on the problems of illumination variations, pose variations, expression variations and so on.The dissertation mainly focuses on illumination variations problem in face recognition. We analyze why illumination variations could affect face recognition. An overview of the state-of-the-art of illumination problem in face recognition is given. Research is conducted from three aspects: image preprocessing, extract illumination insensitive features, the combination of image preprocessing and illumination insensitive features. The main work of this dissertation is as follows:1) Block histogram equalization, block histogram match, and contrast limited adaptive histogram equalization are three illumination preprocessing methods based on traditional image processing techniques. Their performances are better than other methods of the same type. Experiments are conducted in order to compare them. The results show that they do not work well and enhance noise when there are shadows in face image.2) Illumination normalization methods based on Retinex theory have better capability of weakening shadows than illumination preprocessing methods based on traditional image processing techniques. When there are heavy shadows in face image, halo effect is produced at the edges of shadows after preprocessing by the existing illumination normalization methods based on Retinex theory. The halo effect could worsen the successive face recognition. To address this problem, an illumination normalization method named LogTVL2 based on Total Variation Model under L2 norm constraint is proposed. Because the TVL2 model has better capability of edge-preserving, the LogTVL2 could weaken halo effect well. In order to validate the performance of LogTVL2, experiments are conducted from three aspects: capability of weakening halo effect, time-consuming and performance of recognition. The LogTVL2 is compared with three representative methods MSR, LogDCT and SQI.3) At present, all the illumination preprocessing methods treat all face images in the same way without considering the specific illumination conditions. Although most of illumination preprocessing approaches can improve the recognition rate when probe images with large variations in lighting, they may have a bad effect on well-lit face images and lead to a decrease in recognition rate. To address this problem, illumination quality index of face image is proposed. The luminance images of probe face image and reference face image are first estimated. The reference face image is with normal illumination. The similarity score of the two estimated luminance images are calculated by normalized correlation. The similarity score obtained is defined as the illumination quality index of the probe image. Luminance image is estimated by only saving some low-frequency discrete cosine transform coefficients and then conducting inverse transform. Experimental results show that the illumination quality index of face image can measure the degree of uneven illumination of face image well. According to the IQI of input face image, we can decide whether illumination preprocessing should be used on it. This could avoid bringing bad effect on face images with good illumination conditions.4) The edges of facial objects are important cue for face recognition and are less sensitive to illumination changes. The illumination insensitivity of logarithmic gradient (LG), logarithmic vertical gradient (LVG) and logarithmic horizontal gradient (LHG) are compared based on Retinex theory and image processing techniques. By taking advantages of the LVG, LG and LHG, we fuse them at decision level. Experimental results demonstrate the effectiveness of the proposed method.5) Combining with the methods proposed above, a robust processing chain is proposed. Firstly the illumination quality index of probe image is calculated. According to the IQI, probe image with approximately normal illumination can be excluded. The excluded probe image needs no illumination normalization and directly step into recognition by normalized correlation. It not only decreases the possibility of misclassification but also saves times. In order to enhance robustness of face recognition to illumination variations, LogTVL2 is conducted on these non-excluded probe images. Then illumination normalized face images are fed to a new fusion algorithm combining gradient direction and magnitude at decision level. This preprocessing chain makes face recognition more reliable under varying illumination.