Abstract: In order to improve the face recognition rate under illumination variations, this paper proposes an improved retinex algorithm which Gauss filter is replaced by mean-shift filter to solve “halo” phenomenon in traditional retinex algorithm for face illumination image. The face image is enhanced by nonlinear method, and then mean-shift filter is used to estimate the illumination instead of gauss filter to solve the traditional retinex algorithm’s “halo” phenomenon problem. The algorithm’s performance is test by Yale B face database, the results show that the proposed algorithm can restrain the “halo” phenomenon and has illumination robustness to improve the face recognition rate.

Key words: illumination preprocessing, face recognition, Retinex algorithm, Mean-Shift filtering

摘要： 为了提高光照变化条件下的人脸识别率，针对Retinex算法处理人脸光照图像时易产生“光晕”难题，提出了一种基于Mean-Shift滤波的Retinex算法，并应用于人脸识别中的光照预处理。对人脸图像进行非线性增强；利用Mean-Shift滤波代替高斯滤波对光照估计，解决传统Retinex算法中存在的“光晕”难题。采用Yale B人脸库对算法性能进行测试，结果表明，该算法能够很好地抑制“光晕”现象的发生，具有光照鲁棒性，提高了人脸的识别率。

关键词: 光照预处理, 人脸识别, Retinex算法, Mean-Shift滤波