Illumination Invariant Eye Detection In Facial Images Based On The Retinex Theory

Eye detection plays an important role in face recognition, because eye features provide a high recognition rate. However, illumination effects such as heavy shadows and drastic lighting change make it difficult to detect eyes well in actual faces. In this paper, we do the research to find the solution of the illumination invariant eye detection. The mainly content of this thesis is summarized as follows.(1) We proposed a new framework for eye detection on front face image. The core idea of our method is using Retinex and Adaptive smoothing based illumination normalization as a pre-processing for eye detection. In the illumination normalized image (IM image), there is a significantly reduce in the effect of gray contrast, illumination intensity inhomogeneity, high lights and cast shadows. Then, the ordinary eye detection procedures are performed on the IN images. The experimental results show the superiority of the illumination normalization in illumination invariant eye detection system.(2) We proposed an eye candidate detection method using the edge histogram descriptor (EHD). Due to the distinctive edge distribution of eye feature, the eye candidate region is effectively extract and verified by Support Vector Machine (SVM). After that, an Eye Probability Map (EPM) is calculated according the output of SVM, in order to determine the eye position. The experimental results show that our method achieved a high detection accuracy and that, the computing speed of our method is much more faster than using only SVM.(3) We proposed a post-processing method for eye detection base on symmetry axis detection. In most case, the illumination affect can be avoided by the illumination normalization presented in chapter1. However, in some illumination inhomogeneous situation, the eye feature failed to be preserved by adaptive smoothing. To solve this problem, a post-processing is carried out on EPM result. A rectification is given to the EPM according to the symmetry characteristic of front face. After the post-processing, a considerable improve is gain in the detection accuracy. The experimental results show that the proposed post-processing method is effective in most illumination situations when combined with the previous proposed method. Furthermore, a high detection accuracy and fast computing are achieved by this framework. The research is supported by the National Natural Science Foundation of China (No.610501101, No.60970067,No.61050110144).