计算机工程 ›› 2013, Vol. 39 ›› Issue (8): 190-195.doi: 10.3969/j.issn.1000-3428.2013.08.041

• 人工智能及识别技术 • 上一篇    下一篇

基于自适应SVM的半监督主动学习视频标注

张建明,孙春梅,闫 婷   

  1. (江苏大学计算机科学与通信工程学院,江苏 镇江 212013)
  • 收稿日期:2012-04-20 出版日期:2013-08-15 发布日期:2013-08-13
  • 作者简介:张建明(1964-),男,教授、博士、CCF会员,主研方向:虚拟现实,图像处理,模式识别;孙春梅、闫 婷,硕士研究生
  • 基金项目:
    国家自然科学基金资助项目(61170126)

Video Annotation for Semi-supervised Active Learning Based on Adaptive SVM

ZHANG Jian-ming, SUN Chun-mei, YAN Ting   

  1. (School of Computer Science and Telecommunication Engineering, Jiangsu University, Zhenjiang 212013, China)
  • Received:2012-04-20 Online:2013-08-15 Published:2013-08-13

摘要: 具有不同分布特性的视频包含相同的语义概念,会表现出不同的视觉特征,从而导致标注正确率下降。为解决该问题,提出一种基于自适应支持向量机(SVM)的半监督主动学习视频标注算法。通过引入Δ函数和优化模型参数将现有分类器转换为自适应支持向量(A-SVM)分类器,将基于高斯调和函数的半监督学习融合到基于A-SVM的主动学习中,得出相关性评价函数,根据评价函数对视频数据进行标注。实验结果表明,该算法在跨域视频概念检测问题上的平均标准率为68.1%,平均标全率为60%,与支持向量机半监督主动学习和基于直推式支持向量机半监督主动学习相比有所提高。

关键词: 视频标注, 语义检测, 半监督学习, 主动学习, 支持向量机, 高斯调和函数

Abstract: In order to solve the problem which the rate of label accuracy sharp declined due to video shots of same concepts exhibit dissimilar visual features when they are from different domains, a semi-supervised active learning algorithm based on adaptive Support Vector Machine(SVM) for video annotation is proposed. This algorithm migrates the existing classifier into adaptive SVM classifier by the introduction of function Δ and the optimization of model parameters, and obtains the evaluation function by comprising a fusion of the semi-supervised learning based on Gaussian fields and harmonic functions and the active learning based on adaptive SVM, and labels videos through the evaluation function. Experimental results show that the Mean Average Precision(MAP) and Mean Average Recall(MAR) of the proposed algorithm are respectively 68.1%, 60% in cross-domain video concept detection, the results are significantly improved compared with Support Vector Machine Semi-supervised Active Learning(SVM-SAL) and Transducitve Support Vector Machine Semi- supervised Active Learning(TSVM-SAL).

Key words: video annotation, semantic detection, semi-supervised learning, active learning, Support Vector Machine(SVM, Gaussian harmonic function

中图分类号: