基于Gabor滤波和K-medoid聚类分析的人眼检测和瞳孔定位  

Eye detection and pupil location algorithm based on Gabor filters and K-medoid algorithm

作　　者：刘春生[1] 常发亮[1] 陈振学[1] 李爽[1] 

机构地区：[1]山东大学控制科学与工程学院,山东济南250061

出　　处：《光电子．激光》2013年第8期1606-1612,共7页Journal of Optoelectronics·Laser

基　　金：国家自然科学基金(60975025;61273277;61203261);山东省自然科学基金(ZR2011FM032);教育部留学回国人员科研启动基金(20101174)资助项目

摘　　要：针对人眼检测过程中存在的表情、光照和眼镜遮挡等干扰因素的影响,提出一种基于Gabor滤波和K-medoid聚类分析的人眼检测和瞳孔定位方法。首先采用鲁棒性较好的眼部横向特征作为检测对象来设计了Gabor滤波器,以突出眼部的横向特征;然后根据Gabor滤波后的眼部特征,并结合K-medoid算法设计了聚类算法检测人眼;在人眼检测基础上,结合灰度分布特征和熵函数设计了瞳孔定位方法。在BioID人脸库和FERET彩色人脸库上进行了实验,结果表明,本文方法在两个图库的3470幅人脸图像上能够达到97.8%的检测率,并且在设置误差阈值较小(0.15)情况下仍能达到95.5%的瞳孔定位准确率。Eye detection is an essential initial step in many face processing applications, such as race tracking, expression analysis and face recognition. In the process of eye detection,in order to overcome the factors of expression changes,illumination changes and glassesf shelters,the eye detection method based on Gabor filters and K-medoid algorithm is proposed. There are mainly three steps to detect eyes and locate pupils. Irlitially, the proposed method highlights the eyes~ position with differem scale Gabor filters, which is robust to the changes of eye,occlusion and illumination changes. It then combines Gabor filter method with the improved K-medoid algorithm to conduct cluster analysis to detect eyesI position. Since the position located by K-medoid algorithm is not accurate enough, the last pupil locating algorithm is designed on the location method that combines gray level distribution with entropy {unction. Experiments with BiolD database and FERET color database are performed to evaluate this method. The experimental results demonstrate the consistent robustness and efficiency of the proposed method, in which the detection rate is 97.8% in all the 3470 images and the pupil location rate is 95.5% in the case of low error threshold of 0.15.

关 键 词：人眼检测 瞳孔定位 GABOR滤波 K—medoid算法 人脸检测 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]