基于梯度积分投影和最大期望算法的人眼定位

针对传统积分投影方法易受眉毛、睫毛、阴影、遮挡及噪声等干扰的问题,提出了一种梯度积分投影与最大期望(EM)算法相结合的人眼精确定位方法,可以在人脸图像中分割出人眼区域,并精确定位人眼位置。首先,采用一种新的梯度算子计算人脸图像的行梯度积分投影粗略定位人眼区域;然后计算人眼区域的列梯度积分投影函数,用EM算法将所得列梯度积分投影函数曲线拟合成两个高斯曲线,并根据高斯曲线精确分割出人眼窗口;最后,利用我们提出的加权质心法在所得人眼窗口中精确定位双眼位置。在YaleB人脸数据库及自采数据库上的实验结果表明,本文方法不易受眉毛及噪声干扰,并能有效克服眼睑和睫毛的遮挡,对不同光照条件及头部姿态都有很好的鲁棒性。

Eye localization based on gradient integral projection and expectation-maximization algorithm

To overcome the effects of eyebrows,eyelashs,shadows,occlusion and noises in traditional integral projection methods,a precise eye localization method based on gradient integral projection and expectation-maximization(EM) algorithm is proposed.Both eye regions and precise eye positions can be detected in facial images by our method.Firstly,coarse eye regions are detected by row gradient integral projection function which is calculated by a novel gradient operator.Then,precise eye windows are detected by fitting two Gaussian distributions to the column gradient integral projection curve of the eye regions by EM algorithm.Finally,a weighted barycenter algorithm is proposed to locate the positions of eyes precisely.Experimental results on our image database and YaleB database demonstrate that our method is more robust to eyebrows,noises and occlusion due to eyelids and eyelashs,and performs well under different illumination conditions and head poses.