基于人眼定位的人脸检测与归一化算法

Face Detection and Normalization Based on Localization of Human Eyes

文中针对复杂背景下多姿态静态人脸图像,提出了一种通过对眼睛这一特征的自动检测与定位,从而实现对人脸检测的新方法。首先对原灰度图像做边缘灰度加强;然后结合根据人脸几何特征先验知识建立的人眼位置判定准则,在分割阈值递增的过程中,寻找能分割出双眼眼块的最优分割阈值;最后用两维相关系数作为对称相似度来检验检测出的双眼的真实性,并利用找到的双眼图像垂直方向的灰度积分投影,精确定位瞳孔中心。对彩色图像,利用肤色在YCbCr颜色空间的分布特性建立肤色模型,粗略找出肤色区域,进行灰度变换后,再采用上述方法检测人脸。最后提出一种脸相归一化方法,便于进一步的提取特征工作。实验结果证实了文中方法在速度和准确性方面具有良好的性能。

For still face images with complex backgrounds and multiple postures, presents a new algorithm for face detection based on autornatic detection and localization of human eyes. First, edge grayscales in all face images are enhanced. Second, based on the determination criterion of eye location established by the prior knowledge of geometrical facial features, the optimal threshold to separate the two eyes blocks can be found through progressive thresholding. Third, the 2 - D correlation coefficients are used as a symmetry similarity measure to check the truth of detected eyes. Finally, make vertical gray- level integration projection of two eye images to locate pupil centers. In color face images, the properties of YCbCr space can be utilized to build up skin model and find skin color district roughly. The images with found skin color districts can be transferred to grayscale images, then, the above algorithm can be used to detect faces. At last, a normalization approach for face images is introduced, which is convenient for further work of face feature extraction and face recognition. The experimental results demonstrate the good properties of the proposed algorithm in speed and accuracy.