基于级联形状回归的多视角人脸特征点定位  

Multi-view facial landmark location method based on cascade shape regression

作　　者：桑高丽 王国滨 朱蓉[1] 宋佳佳[1] SANG Gao-li;WANG Guo-bin;ZHU Rong;SONG Jia-jia(Department of Mathematics and Information Engineering,Jiaxing University,Jiaxing 314001,China)

机构地区：[1]嘉兴学院数理与信息工程学院

出　　处：《浙江大学学报（工学版）》2019年第7期1374-1379,1388,共7页Journal of Zhejiang University：Engineering Science

基　　金：浙江省自然科学基金资助项目(LY15F020039,LQ18F020006,LQ18F020007)

摘　　要：针对当前人脸特征点定位精度低、且当人脸图像存在较大姿态变化时不能在同一模型框架下实现任意姿态人脸图像的面部特征点精确定位问题,提出基于级联形状回归的对姿态、遮挡都鲁棒的人脸特征点定位方法.为了提高定位的准确度,提出按姿态偏转造成的遮挡程度对人脸区域进行分块,针对每一块分别训练形状估计回归器;为了能够在同一框架下实现任意姿态人脸的特征点定位,在特征点形状定义中引入特征点的可见/不可见属性;为了提高该算法的性能,在特征的计算方法和计算策略上分别进行改进.在Multi-PIE、AFLW、COFW和300-W数据库上的实验结果表明,提出算法不但对姿态、遮挡具有很强的鲁棒性,而且在其他不可控因素影响下取得很好的效果.A new cascade shape regression based facial landmark location method was proposed in order to improve the low landmark location accuracy and perform landmark location with large pose variations in a unified frame.The face area was divided according to the degree of occlusion in order to improve the accuracy of landmark location, and the shape regression was separately trained for each block.The visible/invisible attribute was introduced to the landmark's definition in order to locate the landmarks of any pose under the unified frame.The calculation and the strategy of feature extraction were improved in order to guarantee the performance of the proposed algorithm.The experimental results on the Multi-PIE,AFLW,COFW and 300-W databases show that the proposed method not only has strong robustness to pose and occlusion,but also achieves good results under other uncontrollable factors.

关 键 词：特征点定位 级联形状回归 区域划分 可见/不可见属性 多视角 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]