平面旋转人脸检测与特征定位方法研究

Study on Face Detection Under Rotation in Image Plane and Facial Features Localization

提出了一种基于角点检测、AdaBoost算法和C-V方法的平面旋转人脸检测及特征定位方法.方法首先根据AdaBoost算法训练样本得到脸、眼、鼻、嘴4个检测器;然后以角点作为眼睛的候选点,枚举任意两个角点构造可能的人脸区域,并在区域内运用人脸检测器进行检测;接着利用眼、鼻、嘴检测器检测出人脸特征所在的矩形区域;最后利用C-V方法从各个特征区域中分割出人脸特征的轮廓,进而得到人脸关键特征点的位置.在CMU平面旋转测试集上的检测率为94.6%,误报24个,提取出的特征点位置准确.实验结果表明方法是有效的.

In order to solve the problems of face detection under rotation in image plane and facial features localization, a hierarchical approach was presented. Comer detection,AdaBoost algorithm and C-V method were integrated into the approach.First, four kinds of detectors were trained by AdaBoost algorithm for detecting faces, eyes, noses and mouths. Second, the comers were extracted from image. The comers were regarded as candidate eyes and used to construct candidate face regions as input to the face derector. Third, the eye regions, nose regions and mouth regions were detected using the feature detectors, and the feature contours and feature points were extracted from the feature regions by C-V method. The experiments on CMU rotated face test set result a 94. 6% correct rate with 24 false alarms,and the position of the extracted feature points were accurate.Results show that the proposed approach is efficient.