Handwritten chinese character recognition based on domain-specific knowledge

Abstract

Although some encouraging progress has been achieved in handwritten Chinese character recognition (HCCR), handwritten Chinese address recognition (HCAR) remains an ongoing challenge. Few methods achieve satisfying performance on it due to more irregular distortion and overlapping between characters. In this paper, we first extract keywords from the address by a specially designed key character classifier. Then we use a single character network to recognize the place names. In order to take advantage of hierarchical relationships among place names, we construct an address database from the Chinese administrative divisions and design an error-correction method to improve the recognition of place names. Experiments on handwritten Chinese address datasets demonstrate the effectiveness of the proposed method.