Character Type Classification via Probabilistic Topic Model
Takuma Yamaguchi, Minoru Maruyama · SOAR (Shinshu University) · 2012
In this paper, we propose a method for character type classification based on a probabilistic topic model.The topic model is originally developed for topic discovery in text analysis using bag-of-words representation.Recent studies have shown the model is also useful for image analysis.We adopt the probabilistic topic model for character type classification.In our method, character type classification is carried out by classifying image patches based on their topic proportions.Since the performance of the method depends on a visual vocabulary generated by image feature extraction, we compare several feature extraction and description methods, and examine the relations to classification performance.In addition, by extending the method, we propose a coarse-to-fine approach to achieve stable character type classification for a small image patch.For that purpose, firstly, we partition an image into several patches which contain enough information to estimate the model parameters via EM algorithm.Then, each patch is subdivided into smaller patches.Estimation on the small patch is carried out by MAP-technique with a prior reflecting topic proportion of its parent patch.Through the experiments, we show accurate character type classification is made possible by the probabilistic topic model.