Hierarchical content classification and script determination for automatic document image processing

Page segmentation and image content classification play an important role in automatic image processing with applications to mixed-type document image compression, form and check reading, and automatic mail sorting. In this paper, we first present an enhanced background thinning based approach for fast page segmentation. After the analysis of three different methods individually, a hierarchical approach for document content classification is proposed, which classifies a sub-image into one of two categories: text and halftone. Our approach combines a neural network model, cross-correlation metric, and Kolmogorov complexity measure in a hierarchical structure. Considering the necessity of a recognition system, we also propose using a three-layer feedforward neural network to classify text regions into Chinese and English scripts. The classification accuracy on a number of document images reaches 100% and 97.1% for halftone region and text region, respectively. Meanwhile, the system can achieve a correct rate of 92.3% and 95.0% for Chinese and alphabetic script determination, respectively.