The Study Of Eye Location Algorithm Under Non-uniform Illumination

Face recognition is a typical biometric identification technology. Face recognition relates to many subjects, such as image processing, physiology and pattern recognition. Meanwhile, face recognition for the national security, information security and other financial security also has very high social value and application prospect. The eyes are the most important facial features. Eye location is the prerequisite for face recognition and face analysis, but the accuracy of eye location is vulnerably affected by non-uniform illumination changes and noise. This paper carries out the research and proposes three kinds of eye location algorithm under non-uniform illumination:1. Combining Retinex theory and high discrimination features for eye location. Firstly, face images undergo illumination normalization by fully utilizing Retinex theory. Secondly, the LTP and LPQ features are extracted from the eye-candidate regions. Thirdly, the proposed method can obtain the SVM model by training the cascade eye features and then the trained SVM model can roughly locate the approximate position of the human eye. Fourthly, the position of the human eye is accurately determined by Gaussian fitting. Lastly, we verify the accuracy of eye location. The algorithm is tested on three well-known face databases: CMU PIE, Yale B and AR and has a comparison with the state-of-the-art approaches, such as the self-quotient image(SQI) and the fast logarithmic total variation(FLTV).2. Combining wavelet theory and high discrimination features for eye location. Firstly, face images undergo illumination normalization by fully utilizing wavelet theory. Secondly, the high discrimination features are extracted from the eye-candidate regions. Thirdly, we carry out the robust eye location and then verify the accuracy of eye location. The algorithm is also tested on three well-known face databases and compares with the state-of-the-art approaches and the above algorithm.3. Combining NSCT and high discrimination features for eye location. Firstly, face images undergo illumination normalization by fully utilizing NSCT(nonsubsampled contourlet transform). Secondly, the high discrimination features are extracted from the eye-candidate regions. Thirdly, we carry out the robust eye location and then verify the accuracy of eye location. The algorithm is also tested on three well-known face databases and compared with the state-of-the-art approaches and the two methods above.