多姿态小目标人脸检测算法研究

Algorithm study of face detection for pose-varied and small targets

由于人脸在不同的环境、背景等情况下有着不同的视觉效果,同时人脸本身有着细致复杂的模式变化,一般传统的检测方法都是针对正面五官清晰的人脸,而对于多姿态或小目标人脸检测效果不佳,由此提出了一种在光线补偿下基于人脸和头发几何约束的检测算法。该算法首先分别依据肤色和发色的色彩空间模型分割出目标区域,再利用几何约束特点检测出人脸。实验结果表明,相比于特征提取和模板匹配的人脸检测算法,该算法对于五官模糊的小目标人脸以及面部特征不齐全的多姿态人脸具有较高的检测精度。

As different backgrounds can make the different views of the human faces, together with some painstaking and complicated mode changes, the usual ways which are based on detecting and analyzing the five sense organs perform poorly when the faces are posevaried or small and fuzzy. A face detection algorithm based on geometry constraint using light compensation is proposed. Firstly, the target areas are segmented out respectively according to skin and hair color mode. Then the faces can be detected using the geometry constraints between face and hair. Experiments show that this method performed better on the small targeted and pose-varied faces when compared with the methods applying feature abstraction or template matching.