中文摘要：

传统基于回归的人脸特征点定位算法存在忽略人脸局部结构信息、姿态偏转较大时定位精度差等问题。为此,提出一种基于模糊聚类回归的定位算法。利用人脸特征点之间的局部结构信息对人脸训练集进行聚类,并根据阈值判决结果适度扩充训练样本。分别训练所有子训练集的回归结构,在测试过程中加入多次形状约束以自动调整每次聚类的结果和回归结构的选择,由此提高人脸特征点定位的精度。在300-W数据库上的实验结果表明,与形状回归算法和鲁棒姿势回归算法相比,该算法明显提高了姿态偏转较大情况下的定位精度。

英文摘要：

There are such problems in traditional facial feature point localization algorithm based on regression that the local structure information is ignored and the localization accuracy is poor when the attitude deflection is large, so this paper proposes a localization algorithm based on fuzzy clustering regression. The face training set is clustered with the local structure information of the face feature points, and the training samples are extended according to the threshold decision. The regression structures for all the sub-training sets are trained separately, and the shape constraints are added for several times in the test process to automatically adjust the results of each clustering and the selection of the regression structure,which improves the location accuracy of the facial feature point localization. Experimental results on 300-W database show that compared with ESR and RCPR, the proposed algorithm can effectively improve the positioning accuracy in the condition of large attitude deflection.