一种基于迁移学习的文本—图像特征映射算法-北京邮电大学学报.pdfVIP

2012年12月 北 京 邮 电 大 学 学 报 Dec.2012 第35卷 第6期 Journal of Beijing University of Posts and Telecommunications Vol.35 No.6 摇 摇 文章编号:1007鄄5321(2012)06鄄0001鄄05 一种基于迁移学习的文本—图像特征映射算法 刘摇 杰,摇 杜军平 (1郾 北京邮电大学 智能通信软件与多媒体北京市重点实验室,北京 100876;2郾 北京邮电大学 计算机学院,北京 100876) 摘要:提出一种事件约束下基于迁移学习的文本—图像特征映射算法. 通过潜在狄利克莱分配方法对事件文本数 据进行主题建模,并通过计算主题特征的信息增益选出最显著的文本特征;用视觉词袋模型和朴素贝叶斯方法对 事件图片进行主题建模;通过同事件下的文本数据特征分布和文本—图像共现数据特征分布,实现了对图像特征 分布的近似. 在包含15个主题事件的数据集上进行实验的结果证明了所提特征映射算法的有效性. 关摇 键摇 词:事件约束;迁移学习;文本—图像特征映射;共现数据 中图分类号:TN919摇 摇 摇 摇 文献标志码:A A Transfer Learning Based Text鄄Image Feature Mapping Algorithm LIUJie,摇 DUJun鄄ping (1郾Beijing Key Laboratory of Intelligent Telecommunication Software and Multimedia,Beijing University of Posts and Telecommunications,Beijing 100876,China; 2郾School of Computer,Beijing University of Posts and Telecommunications,Beijing 100876,China) Abstract:A transfer learning based text鄄image feature mapping algorithm under event constraint is pro鄄 posed. Firstly,the documents of each event are modeled by the latent dirichlet allocation,in which the most discriminating feature is obtained by computing the information gain of each topic. Secondly,the images of the corresponding event are modeled through the bag鄄of鄄visual鄄word model and the na觙ve bayes approach. Finally,the feature distributions of the target images are approximated by utilizing the feature distributions of the text data and the text鄄image co鄄occurrence datawithin the same event. Experiment is conducted on a dataset containing 15 categoriesof events. The effectivenessof theproposedfeature map鄄 ping algorithm is shown. Key words:event constraint;transfer learning;text鄄image feature mapping;co鄄occu