Illumination Robust Face Detection And Recognition Research

With the improvement of computer technology, face recognition technology have made substantial progress till now. However, the existing face recognition systems usually have weak generality, flexibility and intelligence in dealing with uneven illumination and expression variation. Generally, a human face recognition system contains three modules as sample pre-process module, face detection module and face recognition module. In pre-process module, anti-illumination function is the most important component that keep the subsequent modules work properly.Among different illumination compensation algorithms, Retinex based algorithms draw a lot of attention since it can works only with the samples itself while some others need training procedure at the first step.In face detection, Viola-Jones detection algorithm, which takes advantage of haar-like feature and bases on Adaboost theory, is a typical cascaded algorithm. It can achieve quite high recognition rate, actually arbitrarily high rate in theory and get done in real time, thus become the best choice in many face recognition systems.Compare to the detection module, recognition module has much more algorithms to choose from, such as the classic PCA, LDA, support vector machine and hidden Markov model. In2006, Donoho and Candes proposed a new signal processing framework called compressed sensing theory, in which Nyquist limitation is not necessarily needed in sampling procedure. John Wright and Yang have taken advantage of compressed sensing theory in human face recognition filed. The sparse representation classification algorithm, which is proposed by them, has been verified as a robust solution in dealing with illumination variation, partly occlusion and expression changes. Moreover, SRC can directly use the picture with high dimensionality, thus has a stronger universality with regard to the images involved in face recognition procedure.The core algorithm proposed is based on sparse representation classification theory. Instead of introducing a sample solution of recognition, I integrate the face detection module and illumination compensation module with the face recognition module, which is a combination of PCA, Meta-face dictionary learning and SRC as well, into a complete face recognition system. The system has been test on the ORL database, Yale-B database and a self-build database and made comparison with the traditional solutions.