计算机应用 ›› 2013, Vol. 33 ›› Issue (07): 1988-1990.DOI: 10.11772/j.issn.1001-9081.2013.07.1988

• 多媒体技术 • 上一篇    下一篇

基于超球支持向量机的多姿态协同人脸检测

滕少华,陈海涛,张巍   

  1. 广东工业大学 计算机学院,广州 510006
  • 收稿日期:2013-01-10 修回日期:2013-03-06 出版日期:2013-07-01 发布日期:2013-07-06
  • 通讯作者: 滕少华
  • 作者简介:滕少华(1962-),男,江西南昌人,教授,博士,主要研究方向:协同计算、数据挖掘、图像分析与处理、网络安全;陈海涛(1987-),男,湖南邵阳人,硕士研究生,主要研究方向:图像分析与处理、数据挖掘、算法设计;张巍(1964-),女,江西南昌人,副教授,主要研究方向:协同计算、数据挖掘、网络安全。
  • 基金资助:

    教育部重点实验室基金资助项目(110411);广东省自然科学基金资助项目(10451009001004804,9151009001000007);广东省科技计划项目(2012B091000173);广州市科技计划项目(2012J5100054);韶关市科技计划项目(2010CXY/C05)

Multi-pose cooperative face detection based on hypersphere support vector machine

TENG Shaohua,CHEN Haitao,ZHANG Wei   

  1. School of Computer, Guangdong University of Technology, Guangzhou Guangdong 510006, China
  • Received:2013-01-10 Revised:2013-03-06 Online:2013-07-06 Published:2013-07-01
  • Contact: TENG Shaohua

摘要: 针对多姿态的人脸检测准确度差的问题,提出了一种多姿态的协同人脸检测模型。该模型由一组超球支持向量机组成,它们被分成三层:第一层1个、第二层3个、第三层9个,共13个支持向量机(SVM)。这些SVM按逐层精细化检测设计,协同完成人脸检测任务。因为一幅图像的大部分区域是非人脸,采用三层模型的设计一方面能提高人脸检测速度,另一方面也增强了检测的针对性,使得能逐层履行更精细的局部区域检测。另外,改进了k近邻(kNN)算法,使其能用于超球重叠样本的检测,并提高了人脸检测的准确度。实验结果表明,相对于传统基于SVM的人脸检测,所提算法在人脸检测的准确率上有5%左右的提升,通过逐层过滤,保证了人脸检测的速度。

关键词: 超球支持向量机, 协同人脸检测, 多姿态, k近邻, 超球重叠

Abstract: With regard to poor accuracy of multi-pose face detection, a hyper-sphere Support Vector Machine (SVM) was used to detect human faces. A model was proposed in this paper, which was composed by thirteen SVMs. These SVMs were divided into three levels, the first level had one SVM, the second level had three SVMs, and the third level had nine SVMs. Each SVM was a hyper-sphere support vector machine, which was exploited to detect multi-pose faces from various angles. The 3-tier model was applied to fast reduce detection area. On one hand, it accelerated the speed of detection; on the other hand it was favorable to make a careful detection in a small local area. In addition, the k-Nearest Neighbor (kNN) algorithm was improved in this paper. The improved kNN algorithm was applied to deal with the detection of hyper-sphere overlap samples. The experimental results show that the proposed algorithm can promote about 5% in the face detection accuracy than the traditional SVM-based face detection algorithm, but also ensure the speed of face detection.

Key words: Hyper-Sphere Support Vector Machine (HSSVM), cooperative face detection, multi-pose, k-Nearest Neighbor (kNN)

中图分类号: