Recognition-based Segmentation for Digitization of Korean Historical Document Pages

摘要

We present a recognition-based digitization method for building digital library of large amount of historical archives. Digitization of historical document pages is essential for providing retrieval service and preventing from damages but needs laborious manual verification for accurate output. In this paper, split-merge approach is applied for segmenting overlapped and touched characters written by thick brushes. Character string images are split into primitive segments by nonlinear segmentation paths passing maximum curvature points. Split segments are merged in single probabilistic framework integrated by layout analysis, context information and recognition result. In experiment, our system achieved 96.4% character recognition rates on test data set, despite the obsolete characters and unique variants used in the archives. In conclusion, our method can be applied for digitizing Korean historical document pages and minimize manual verification.