Radical Similarity Based Model Optimization and Post-correction for Chinese Character Recognition

Zhongyuan Han, Jun Du, Mobai Xue, Jiefeng Ma, Pengfei Hu, Zhenrong Zhang

Abstract: Radical-based methods for Chinese character recognition (CCR) have been proven effective and offer substantial advantages. Different from character-based methods, Chinese characters are described as combinations of structures and radicals, and character recognition is achieved by the proper identifications of these components. However, there are visual similarities among radicals, leading to the ambiguity problem for CCR, which is not fully utilized in previous work. Accordingly, in this study, we first employ the stroke order information of Chinese radicals to establish a radical similarity metric. Then we improve the radical-based CCR in two ways. During the training stage, we propose a new loss function called minimum Bayesian risk (MBR) based on the radical similarity metric to yield better performance. During the recognition stage, the radical similarity is adopted to post-correct the potential error recognition results, offering a low-cost yet effective solution. Experimental results on different radical-based CCR models and datasets demonstrate the effectiveness and robustness of our proposed method.