Few-shot handwritten Chinese character generation based on generative adversarial network

Abstract

Chinese characters are different from the English alphabet that English consists of 26 simple words, whereas Chinese characters contain multiple strokes, complex structures and a variety of styles, which makes the study of Chinese character style transfer a challenging problem. In recent years, generative adversarial networks have been used to automatically generate handwritten Chinese characters, but these methods often produce Chinese characters with a variety of stroke errors, such as multiple strokes and missing strokes. Especially, few works focus on the problem of few-shot handwritten Chinese character generation. A further problem that is worth exploring is that in practice, people often want to use only a small number of handwritten Chinese characters, then let the model automatically generate a large number of handwritten Chinese characters with the same style. Therefore, in this paper, few-shot handwritten Chinese character generation is discussed in detail. The experimental results show that it is possible to use only a few tens of handwritten characters to automatically generate a library of handwritten characters with special style.