Image categorization by a classifier based on probabilistic topic model

With rapid increase of number of accessible images and videos, ability to recognize visual information is getting more and more important for content-based information retrieval. Recently, probabilistic topic models, which were originally developed for text analysis, have been used for image categorization successfully. Usually, “topics” which represent contents of an image is detected based on the underlying probabilistic model, then image categorization is carried out using topic distribution as the input feature. Typical method is to use k-nearest neighbor classifier based on L2-distance after topic discovery. In the method, topic distribution is just treated as a feature point. In this paper, we propose a categorization method based on more natural use of the topic distribution, which is derived by using pLSA model. Categorization is carried out by estimating conditional probability p(categoryjdata). We present two types of image categorization tasks, scene classification and document image segmentation, and show the proposed method performs very well. In addition, we also examine the performance of the proposed method under the situation where only the limited number of labeled examples are available. We show our method can perform quite well even in the circumstances.