A study on image representation by hierarchical probabilistic models and its applications

本文主要研究了图像表示的层次概率模型及其在类别识别中的应用。在我们的方法中，我们使用了概率主题模型。提出了一种基于概率潜在语义分析的文档图像区域分类方法。根据基于MAP的EM算法估计的区域主题比例进行分类。该模型被进一步扩展到对较小的图像区域进行分类。提出了利用决策树桩对自然场景图像中的字符进行检测的方法。这项研究工作的一部分在2008年IAPR文件分析系统研讨会上荣获最佳论文奖。

We have mainly studied about hierarchical probabilistic model for image representation ant its application to category recognition. In our method, we used probabilistic topic model. We proposed a method for classifying regions in a document image based on pLSA (probabilistic latent semantic analysis). Classification is carried out based on topic proportion of the region that is estimated via MAP-based EM algorithm. The model is further extended to classify smaller image regions. We also presented the method to detect characters in a natural scene images by using decision stumps. A part of this research work received best paper award honorable mention at the IAPR workshop on Document Analysis Systems 2008.