Face Recognition Based On Retinex Under Varying Illumination

Face recognition is widely focused by many researchers for its stability, usability and Anti-counterfeiting in biology recognition technology. Many proposed face recognition algorithms can achieve high accuracy under simple environment, however complex environment, for example illumination, may cause accuracy falls sharply, so those algorithms can’t be applied actually. This paper researches face recognition under illumination. We mainly study illumination processing method based on Retinex with related work and finish some works about this theme. Our main work as follow.Firstly, we study main theories of illumination processing, and focus on the Retinex which come from illumination invariant theory. Besides, we learn and analyze the single-scale-Retinex (SSR), multi-scale-Retinex (MSR) and adaptive-smoothing-Retinex (ASR). We also discuss the features and compare the advantages and disadvantages of the three algorithms.Secondly, we implement an eye location approach under illumination, and propose a multi-level selection strategy which is used to obtain eye candidate detection. At beginning, the ASR is utilized to extract illumination invariant; then, we use the multi-level selection strategy to get eye regions. At last, we employ support vector machine (SVM) and eye probability map (EPM) to locate the position of eyes. The experiments suggest the proposed method can achieve high detection accuracy and low consumption time at the same time.Thirdly, we propose a face recognition based on ASR and wavelet transform. Low accuracy is obtained by the ASR because it ignores the contributions of low frequency of human face to the face recognition. We give an idea that wavelet transform can be used to extract low frequency to supply the illumination invariant. At first, illumination invariant and variant are generated by the ASR. Second, decompose the illumination component via wavelet transform and process its low-frequency coefficients. After that, the new illumination component is obtained by inverse transform. In the end, an image which is employed to recognize is acquired by restructuring the two components. Our experiment demonstrates that the proposed method get good performance in illumination environment.