Chinese font generation based on deep learning

Abstract

Font generation is a challenging problem. To address the existing problems of poor font style conversion models, which have missing structure, blurred glyphs and require paired datasets, this paper proposes a Chinese font style migration algorithm based on the improved CycleGan. The model introduces deformable convolution in the encoder part of the generator, which can learn the font features adaptively. A skip connection module, which fuses global and local features, was added to the model, and the features in the encoder are projected to the decoder using this module to avoid the structural error problem by reducing the information loss of the decoder. Meanwhile, using the attention mechanism, we can quickly and efficiently obtain the key information of the target region. On this basis, we can further complete the local and global feature fusion. According to the research results, this method can better achieve font generation in practice, so it has high application value.