利用肤色分割和对称变换实现鲁棒人脸特征定位

Robust facial feature location using skin color segmentation and symmetry transform

提出一种将肤色分割与灰度图像对称变换相结合的有效的人脸特征定位方法。为降低背景和人脸姿态的影响,在传统人脸肤色分割基础上,利用Blob分析及椭圆拟和方法判断并剔除无效区域,并将待选人脸区域初步校正为竖直;直接利用色度信息在待选人脸区域中分割并检测嘴唇;接着,采用灰度图像的对称变换检测双眼的待选位置;以嘴唇和待选双眼的位置为依据,提出了基于人脸几何模型知识的双眼匹配代价函数并利用组合优化方法检测真实双眼;最后,对人脸上其它器官的进行精确定位。实验结果表明,该方法对于头发遮挡、大倾斜角以及眼镜干扰具有较强的鲁棒性。

An efficient method for facial feature location was proposed based on skin color segmentation and gray image symmetry transform. To reduce the influence of background and variable facial poses, the BIob analysis and ellipse fitting methods were successively applied to the image blobs detected by conventional skin color segmenting method. Invalid blobs were removed and candidate face regions were adjusted vertically. Then, mouth region was extracted from each candidate face region by segmenting hue image. Afterwards, the Generalized Symmetry Transform （GST） of gray image was used to obtain candidate eyes and the real eyes were detected by combination optimization of the proposed cost function which was related to model of facial geometry and positions of eyes and mouth. Finally, other facial features such as chin and cheek could be precisely located. The test results show that the proposed approach is more effective and has better performance.