Traditional Chinese Character Recognition Based on Unsupervised Domain Adaptation

Chinese characters have a long history, and each evolution has a certain pattern. Simplified Chinese characters are the mainstream at present, traditional Chinese characters are only appears in a small number of regions. For the research of traditional Chinese characters, it is fundamental to identify traditional Chinese characters with their corresponding simplified Chinese characters. However, there is no suitable dataset for labeling traditional characters with their corresponding simplified characters. Therefore, we hope to use unsupervised domain adaptation method, with simplified characters as the source domain which have labels, and traditional characters as the target domain which have no label. In this paper, the network was designed using DANN as the backbone. Unlike DANN, the network combined the information of Chinese characters and utilized two feature extractor. The final accuracy of the network incorporating the information of Chinese characters is 95.65%. Therefore, this paper provides a method to achieve traditional Chinese character recognition without labeling traditional Chinese characters and the technique of simplifying traditional Chinese characters is used to improve the recognition success rate.