Dynamic character model generation for document keyword spotting

This paper proposes a novel method of generating statistical Korean Hangul character models in real time. From a set of grapheme average images we compose any character images, and then convert them to P2DHMMs. The nonlinear, 2D composition of letter models in Hangul is not straightforward and has not been tried for machine-print character recognition. It is obvious that the proposed method of character modeling is more advantageous than whole character or word HMMs in regard to the memory requirement as well as the training difficulty. In the proposed method individual character models are synthesized in real-time using the trained grapheme image templates. The proposed method has been applied to key character/word spotting in document images. In a series of preliminary experiments, we observed the performance of 86% and 84% in single and multiple word spotting respectively without language models. This performance, we believe, is adequate and the proposed method is effective for the real time keyword spotting applications.