基于相对距离分布聚类的人脸特征点定位算法

An Algorithm of Locating the Feature Points of Human Face Landmarks Based on Relative Distance Distribution

针对目前主流算法对所有特征点采用整体回归而忽略人脸局部结构信息的问题,提出了一种新的回归流程结构,通过基于特征点相对距离分布直方图和K均值聚类,结合了人脸结构信息对特征点聚类并分别进行回归,可以更准确地进行人脸特征点定位。另外,通过对回归迭代方法进行优化,可以实现鲁棒的参数更新,大大提高运算速度。在有遮挡标识的人脸数据库(COFW)上进行了深入的实验,结果表明:论文算法对于人脸特征点定位效果显著,相较于鲁棒姿势级联回归(RCPR)等算法训练时间大幅减少,其定位准确度也有了一定提高,而且算法计算效率高,测试速度则达到220fps,能够满足实时处理的要求。

In the light of the problems that at present in the main stream algorithms the integral regression method is still applied to all the feature points whereas the local structure information of human face is ig- nored, a novel cascaded regression structure is presented based on relative distance distribution and K- means clustering combined with face structure information on clustering facial landmarks and performing regression for each part respectively, the feature points locating of human face can be performed more accurately. In addition, the regression method is optimized to make robust parameter updated with efficien- cy. The paper carries out a thorough experiment on a face database （COFW） with block identification. The experiments demonstrate that the algorithm is notable in effect with regard to the application of feature points of a human face to the location, and the algorithm is greatly short in training time compared with Robust Cascaded Pose Regression and other state-of-the-art methods and the testing speed is up to 220 fps, thus realizing real-time processing.