Study on the Method of Progressive Restoration of Defective Chinese Characters

With the development of deep learning techniques, many methods have been applied to character recognition and generation. However, few studies have focused on the problem of Chinese character repair. Existing Chinese character repair methods usually produce pseudo-characters and fuzzy characters due to the lack of semantic information. Since Chinese characters contain both visual glyph representations and intrinsic semantic information, this paper proposes a novel progressive semantic repair network for Chinese characters, which aims to improve the accuracy and completeness of Chinese character repair. First, a deep semantic filtering network is proposed to perform deep pixel-by-pixel feature filtering on defective Chinese character images to refine the semantic information. Secondly, this paper proposes a progressive repair inference network to infer the defective content of a Chinese character image from different levels, which is then used as a clue for the next step to strengthen the constraints on the defective region, to generate a complete Chinese character with the correct glyph shape as well as clear semantics. Finally, the quantitative and qualitative experimental results show that the proposed method is more effective than the most advanced Chinese character restoration methods.