典型相关分析与多伯努利相关模型的图像标注  

Automatic image annotation method based on canonical correlation analysis and multiple Bernoulli relevance model

作　　者：周晓[1,2] 潘洁珠[2] 

机构地区：[1]合肥工业大学计算机与信息学院,安徽合肥230009 [2]合肥师范学院计算机科学与技术系,安徽合肥230061

出　　处：《合肥工业大学学报（自然科学版）》2010年第6期841-846,共6页Journal of Hefei University of Technology：Natural Science

基　　金：国家重点基础研究发展计划资助项目(2009CB326203);安徽高等学校省级自然科学研究资助项目(KJ2009B238Z)

摘　　要：文章提出一种基于图像的视觉词袋与文本标注的典型相关分析与分割无关的多伯努利相关模型的自动图像标注算法。在图像标注与分类任务中,矢量量化图像局部描述子得到的视觉词袋特征已显示了其鲁棒性与可区分性,文中对视觉词袋与文本特征作典型相关分析,确保投影变换后新的视觉特征与文本特征的相关性最大化,从而有效地在视觉与文本2种模态中建立联系,契合了自动图像标注的主旨。文中还提出了一种简化的多伯努利相关模型,实验结果证明了典型相关分析比概率潜藏语义分析更适合于图像自动标注,也证明了简化的多伯努利相关模型的有效性。A novel automatic image annotation method is presented on the basis of canonical correlation analysis between visual bag of words and textual annotation and non-segmentation multiple Bernoulli relevance model. In image annotation or classification tasks, visual bag of words via the vector quanti- zation of image local descriptors often possesses robust and distinct features. In the present method, canonical correlation analysis is made to maximize the correlation between the projected visual features and textual features, thereby to build the relationship between these two modalities. It is the exact intention of automatic image annotation. A simplified multiple Bernoulli relevance model is also proposed, which is another novelty in this paper. Experiment results prove that canonical correlation a- nalysis is better than probability latent semantic analysis in automatic image annotation, and also prove that the simplified multiple Bernoulli relevance model is effective as a new annotation model.

关 键 词：自动图像标注 尺度不变特征变换 层次化k-means 视觉词袋 典型相关分析 多伯努利相关模型 

分 类 号：TP391.3[自动化与计算机技术—计算机应用技术]