Construction of a text digitization system for Nom historical documents

Abstract

This paper presents a text digitization system for Nom historical documents, employing image binarization, character segmentation and character recognition. It incorporates two versions of offline character recognition: one for automatic classification and the other for verification and correction by an operator. They employ the same recognition method but they are trained by two different sets of training patterns with 7,601 and 32,733 categories. For the recognition method, we use the Generalized Learning Vector Quantization (GLVQ) algorithm for coarse classification and the Modified Quadratic Discriminant Function (MQDF2) method for fine classification. Sample character patterns are generated artificially from 27 fonts of Chinese, Japanese and Nom characters since ground-truthed sample patterns are not available. Moreover, in order to accelerate large scale recognition, we use the kd-tree algorithm in the coarse classification process. The system also provides the interface through which an operator can verify and correct the results of image binarization, character segmentation and character recognition.