Study On Face Recognition Under Varying Illumination Based On Multi-scale Analysis

Face recognition is an important part of biometrics research. In public security, real-time monitoring, authentication and human-computer interaction fields, the face recognition has an extensive application prospect. Nowadays, there are some face recognition systems used for security applications, such as entrance guard of sensitive public places, attendance check using face and human track in passenger accumulation area. In 2008 Beijing Olympic Games, the face recognition system was used at ticket gate of the National Stadium, known as the Bird's Nest.But for the reason of improving the right recognition rate, in those face recognition systems used in practical application, the person to be recognized is always demanded to cooperate with the system's work. For instance, staying at controlled illumination environment, avoiding any exaggerated face expression, having no any occlusion object on face, etc. In these interferential factors, the varying illumination, including varying intensity and varying orientation, is most important. In outdoor and varying illumination environment, the right recognition rates will remarkably decrease. The variation in the face appearance caused by illumination variations can be much larger than the variation caused by personal identity.Facing the challenges of face recognition under varying illumination, this dissertation makes contributions on the following works, based on multi-scale analysis method.1. This dissertation proposes a calculating framework for fusing multi-scale detail images. The variance of high frequency signal caused by varying illumination is slighter than that of low frequency signal. So, in varying illumination condition, high frequency signal component has the ability of representing the face essential characteristic. According to that, there are many recognition methods that transform the image to high frequency information image and use this image to do some further feature extraction and classification work. However, whatever method is used, a question must be answered,"Whether the entire essential characteristic is represented in this high frequency detail image?"The answer is NO. But it is not possible to figure out that how much high frequency information is enough to represent the face essential characteristic. There is no exact definition of high frequency information and it is blurry. The face essential characteristic is dispersed in multi-scale detail images which contain much high frequency information. The purpose of the fusing calculating framework is adopting much high frequency information in multi layer detail images and concatenating the multi-scale frequency features to represent the face essential characteristic under varying illumination.2. Considering the energy of wavelet transform coefficient, energy of frequency domain in Fourier transform, image extremum points and contrast in local image region, this dissertation proposes 4 measurements to measure the detail information in one image. The fusing calculating framework takes account into several feature distances between the same layer detail images of different face images. And these distances make different contribution to the global feature distance. Although having much detail and texture information, the upper layer detail image has little structure and contour information. According to the cognitive view, structure and contour information is much helpful to recognize objects, so the feature distance between lower layer detail images makes more contribution than that between upper layer detail images. There is no same amount of high frequency information in each detail image. It is necessary to produce the measurement to measure the amount of detail information in detail image and to be transformed the weight value to measure the contribution ratio. Although not precise, the produced measure values can represent the relative amount of detail information in multi layers detail image. So, the measure value of the first layer detail image is the most and that of the last layer detail image is the least. Converting these measure values to weight values using a method proposed in this dissertation, the contribution ratio of constructing global face feature for each detail image can be calculated approximately and it is effective to improve the right recognition rate in face recognition under vary illumination adopting by fusing calculating framework.3. This dissertation proposes an improved LMCP(Local Multi-layer Contrast Pattern) based on LBP. The LBP method, proposed by Ahenon, only records the sign (positive or negative) of neighbor pixel gray value in local area, not considering the contrast of the neighbor pixel. But unfortunately, the losing contrast is important feature to distinguish one texture from another. According to the new idea, the face image is illumination normalized in preprocessing step so that the illumination variance can be controlled in a narrow range. And then, the contrast value between two neighbor pixels in local image region is mapped to a layer value. If the illumination varies slightly, the LMCP feature can get the illumination-invariant property and strengthened ability of represent the texture feature. Furthermore, the statistic and mapping method is produced to reduce the LMCP feature dimension.This dissertation also shows much experiment data based on open face databases. The experiment results prove above methods correct and effective.