A system for document image classification

Abstract

In this thesis, the system for Document Image Classification is discussed in detail with three parts, i.e., document segmentation, and block classification and document classification. In fact, only document segmentation and document classification is used. The block classification does not need. The reason for it has been analyzed clearly. For the segmentation, the Fuzzy Segmentation Method (FSM) for the analysis of document images is developed. Based on the uses of the fuzzy set theory, the fuzzy blank and fuzzy black blocks are defined. The idea of 弇-cut sets is utilized when determining the vertical and horizontal thresholds in the segmentation process. With the definition of fuzzy sets, such a process can be made automatically in an adaptive manner. For the block classification, a system for automated pattern analysis and classification (APACS) method is applied to the block classification. The algorithm, features, and selection of the samples for using this method are specified. Furthermore, the part of affirming the feature of the blocks is omitted. The document classification algorithm in the third part employs the features of each block as parameters. This can reduce the error of recognition. The results are very satisfied. For the document classification, the Branch-and-Bound Technique is adopted to match two Attributed Random Graphs (ARGs) of a document. A new evaluation function of the Branch-and-Bound algorithm is proposed. ARG construction, decision tree construction and using branch-and-bound to match two ARGs are elaborated in detail. The document classification is successfully transferred to how to match the two ARGs. The experimental results are quite satisfied.