摘要
利用形态学商图像对人眼区域和皮肤区域有明显不同响应值的特点,根据贝叶斯后验概率的优点,提出一种新的人眼定位的方法。算法分成人眼模板训练过程和人眼定位过程。训练过程首先计算人眼图像的形态学商图像,然后利用K-MEANS算法自动为不同角度或者戴眼镜的人眼建立模板。人眼定位过程首先计算人脸图像的形态学商图像,然后使用与模板相同大小的滑动窗口和离散余弦相似度判定每个窗口与模板匹配的程度获得后验概率图,最后根据后验概率图中各个连通区域的平均概率判定眼睛位置。在Caltech数据库、Labeled Faces in the Wild(LFW)数据库及Yale Face B数据库中的实验结果证明,该算法对不同角度,不同光照,低分辨率,甚至戴眼镜的人眼图片均有较高的定位率。
By utilising the feature of morphological quotient image that it has manifestly different response values on human eye areas and skin areas, and according to the advantages of Bayes posterior probability, we present a novel approach for human eye locatisation. The algorithm consists of two processes: the eye templates training and the eyes locating. In eye templates training, it first calculates the morphological quotient image of human eye image, and then automatically builds different templates for the eyes with different angles or wearing glasses by k-means algorithm. In eyes locating, the morphological quotient image of input face image is calculated first; then the sliding windows the same size as the templates and the discrete cosine similarity are employed to determine the extent of each window matching the template for obtaining the posterior probability graph; lastly, the locations Of eyes are determined by the average probability of the connected regions in posterior probability graph. The results of experiment on Caltech Database, Labelled Faces in the Wild Database and Yale Face Database B prove that our algorithm achieves high locating rate on the eyes image with different angles and illumination, low resolution, and even wearing glasses.
出处
《计算机应用与软件》
CSCD
2015年第1期194-198,共5页
Computer Applications and Software
基金
国家自然科学基金项目(60974019)
广东省自然科学基金项目(9451009001002686)