首页 | 本学科首页   官方微博 | 高级检索  
     

基于惩罚因子的PNMS算法的人脸检测和对齐
引用本文:李振东,钟勇,陈蔓,陶攀.基于惩罚因子的PNMS算法的人脸检测和对齐[J].四川大学学报(工程科学版),2018,50(6):225-231.
作者姓名:李振东  钟勇  陈蔓  陶攀
作者单位:中国科学院 成都计算机应用研究所, 四川 成都 610041;中国科学院大学, 北京 100049,中国科学院 成都计算机应用研究所, 四川 成都 610041;中国科学院大学, 北京 100049,中国科学院 成都计算机应用研究所, 四川 成都 610041;中国科学院大学, 北京 100049,中国科学院 成都计算机应用研究所, 四川 成都 610041;中国科学院大学, 北京 100049
基金项目:四川省科技支撑计划资助项目(2014GZ0013)
摘    要:针对人脸相互遮挡、人脸朝向等不确定性因素给人脸检测和对齐任务带来的困难问题,提出了基于惩罚因子的PNMS算法用以改进人脸检测和对齐的准确性。该算法首先根据人脸候选窗口相互之间的重叠度和候选窗口相应的检测得分,提出非连续的线性函数和基于高斯分布的连续函数,作为非极大值抑制算法的两种惩罚因子,用以改进并替代传统非极大值抑制算法对候选窗口的检测得分进行重分配。在此基础上,综合前两种惩罚因子的优缺点以及窗口之间的重叠度值,进一步提出连续非线性函数作为非极大值抑制算法的惩罚因子,使得窗口之间重叠度值越大则相应的惩罚权重越严重,且函数在整个重叠取值区间连续。将提出的算法在FDDB和WIDER FACE这2个人脸检测数据集上进行详尽的人脸检测实验验证,以及在AFLW人脸对齐数据集上进行人脸对齐实验验证。结果表明,提出的基于惩罚因子的PNMS算法相比于其他算法,在保持一定实时性的同时不仅有效地提高了人脸检测和对齐的准确率和可靠性,并且解决了一定程度的人脸相互遮挡被漏检的问题,降低了被遮挡人脸的漏检率。

关 键 词:人脸检测  人脸对齐  卷积神经网络  非极大值抑制
收稿时间:2017/12/22 0:00:00
修稿时间:2018/10/8 0:00:00

PNMS Algorithm Based on Penalty Factors for Face Detection and Alignment
LI Zhendong,ZHONG Yong,CHEN Man and TAO Pan.PNMS Algorithm Based on Penalty Factors for Face Detection and Alignment[J].Journal of Sichuan University (Engineering Science Edition),2018,50(6):225-231.
Authors:LI Zhendong  ZHONG Yong  CHEN Man and TAO Pan
Affiliation:Chengdu Inst. of Computer Application, Chinese Academy of Sciences, Chengdu 610041, China;Univ. of Chinese Academy of Sciences, Beijing 100084, China,Chengdu Inst. of Computer Application, Chinese Academy of Sciences, Chengdu 610041, China;Univ. of Chinese Academy of Sciences, Beijing 100084, China,Chengdu Inst. of Computer Application, Chinese Academy of Sciences, Chengdu 610041, China;Univ. of Chinese Academy of Sciences, Beijing 100084, China and Chengdu Inst. of Computer Application, Chinese Academy of Sciences, Chengdu 610041, China;Univ. of Chinese Academy of Sciences, Beijing 100084, China
Abstract:In order to solve the difficult problems of caused by mutual occlusion of face and face orientation, a PNMS algorithm based on penalty factors was proposed to improve the accuracy of face detection and alignment. Firstly, according to the overlap degree between face candidate windows and the corresponding detection scores of candidate windows, a non-continuous linear function and a continuous function based on Gaussian distribution were proposed and used as penalty factors for non-maximum suppression. Then the traditional non-maximum suppression algorithm was improved and replaced, and the detection score of the candidate window was redistributed. On this basis, combining the characteristics of the first two kinds of penalty factors, the continuous nonlinear function was further proposed as the penalty factor of the non-maximum suppression algorithm. Consequently, the greater the overlap value between windows, the more severe the penalty is, and the function is continuous throughout the overlapping value range. The proposed algorithm performed detailed face detection experiment verification on two face detection data sets of FDDB and WIDER FACE. The face alignment experiments were verified on the AFLW data set. The results showed that the proposed PNMS algorithms compared with other algorithms not only effectively improves the accuracy and reliability of face detection and alignment, but also solves a certain degree of face occlusion, and reduces the rate of detection failure of occluded faces.
Keywords:face detection  face alignment  convolutional neural network  non-maximum suppression
点击此处可从《四川大学学报(工程科学版)》浏览原始摘要信息
点击此处可从《四川大学学报(工程科学版)》下载全文
设为首页 | 免责声明 | 关于勤云 | 加入收藏

Copyright©北京勤云科技发展有限公司  京ICP备09084417号