改进的单尺度Retinex及其在人脸识别中的应用*

为了提高在光照条件变化下的人脸识别率,提出一种改进的单尺度Retinex算法并用于人脸识别的光照预处理中.该算法通过非线性全局对比度增强对原图像增强,并利用Mean-Shift平滑滤波代替传统单尺度Retinex中的高斯滤波对光照估计,能够明显地消除单尺度Retinex算法中不能解决的光晕现象.在人脸库的实验表明,该算法不仅比直方图均衡化、Gamma校正、单尺度Retinex、多尺度Retinex算法具有更好的光照预处理效果,而且能够有效提高人脸识别率.

Improved monoscale Retinex and application in face recognition

In order to improve the face recognition rate in the change of illumination, this paper proposed a modified monoscale Retinex algorithm and applied to the illumination pretreatment of face recognition. This algorithm strengthened the original image through non-linear global contrast enhancement, and estimated the illumination by Mean-Shift smoothing filtering instead of Gaussian filter in traditional monoscale Retinex, and removed obviously the problem that monoscale Retinex could not solve halo formation. The experiment in face library shows that, this algorithm can not only has better illumination pretreatment effect than histogram equalization algorithm, Gamma correction algorithm, monoscale Retinex algorithm, multi-scale Retinex algorithm, but also improve the face recognition rate effectively.