计算机应用 ›› 2018, Vol. 38 ›› Issue (5): 1289-1293.DOI: 10.11772/j.issn.1001-9081.2017102586

• 人工智能 • 上一篇    下一篇

改进的显式形状回归人脸特征点定位算法

贾项南, 于凤芹, 陈莹   

  1. 江南大学 物联网工程学院, 江苏 无锡 214100
  • 收稿日期:2017-10-31 修回日期:2017-12-13 出版日期:2018-05-10 发布日期:2018-05-24
  • 通讯作者: 贾项南
  • 作者简介:贾项南(1991-),女,江苏宿迁人,硕士研究生,主要研究方向:图像与视频信号处理;于凤芹(1962-),女,辽宁北镇人,教授,博士,主要研究方向:语音信号处理、非平稳信号时频分析;陈莹(1976-),女,浙江丽水人,副教授,博士,CCF会员,主要研究方向:计算机视觉、信息融合。
  • 基金资助:
    国家自然科学基金资助项目(61573168);中央高校基本科研业务费专项资金资助项目(JUSRP51733B)。

Improved explicit shape regression for face alignment algorithm

JIA Xiangnan, YU Fengqin, CHEN Ying   

  1. Internet of Things Engineering College, JiangNan University, Wuxi Jiangsu 214100, China
  • Received:2017-10-31 Revised:2017-12-13 Online:2018-05-10 Published:2018-05-24
  • Contact: 贾项南
  • Supported by:
    This work is partially supported by the National Natural Science Foundation of China (61573168), the Special Funds for Basic Scientific Research Operation of the Central Universities (JUSRP51733B).

摘要: 针对显式形状回归(ESR)人脸特征点定位精度低的问题,提出了改进的显式形状回归人脸特征点定位算法。首先定位出三点人脸形状代替人脸检测框作为初始形状的映射标准来得到更精确的初始人脸形状,然后采用像素块特征代替像素特征对抗光照变化来提高算法的鲁棒性,最后采用多假设融合策略代替平均法对多个定位结果进行最佳融合来进一步提高算法的定位精度。仿真实验结果表明,在LFPW、HELEN和300-W人脸库上,与显式形状回归算法相比,定位精度分别提高了7.96%、5.36%和1.94%。

关键词: 显式形状回归, 人脸特征点定位, 初始人脸形状, 像素块特征, 多假设融合策略

Abstract: To solve the problem that Explicit Shape Regression (ESR) has low precision in face alignment, an improved explicit shape regression for face alignment algorithm was proposed. Firstly, in order to get a more accurate initial shape, three-point face shape was used as an initial shape mapping standard to replace face rectangle. Then, pixel block feature was used against illumination variations instead of pixel feature, which improved the algorithm robustness. Finally, instead of average method, the accuracy of algorithm was further improved by multiple hypothesis fusion strategy which merged multiple estimations. Compared with explicit shape regression algorithm, the simulation experimental results show that the accuracy is improved by 7.96%, 5.36% and 1.94% respectively by using the proposed algorithm on LFPW, HELEN and 300-W face datasets.

Key words: Explicit Shape Regression (ESR), face alignment, initial face shape, pixel block feature, multiple hypothesis fusion strategy

中图分类号: