A shared parts model for document image recognition

• International Conference on Document Analysis and Recognition (ICDAR 2007)

We address document image classification by visual appearance.  An image is represented by a variable-length list of visually salient features. A hierarchical Bayesian network is used to model the joint density of these features.  This model promotes generalization from a few samples by sharing component probability distributions among different categories, and by factoring out a common displacement vector shared by all features within an image. The  Bayesian network is implemented as a factor graph, and  parameter estimation and inference are both done by loopy  belief propagation. We explain and illustrate our model on  a simple shape classification task. We obtain close to 90%  accuracy on classifying journal articles from memos in the  UWASH-II dataset, as well as on other classification tasks  on a home-grown data set of technical articles.

citation

Das Gupta, M.; Sarkar, P. A shared parts model for document image recognition. 9th International Conference on Document Analysis and Recognition (ICDAR); 2007 September 23-26; Curitiba; Brazil. NY: IEEE; 2007; 2: 1163-1172.