Unsupervised Font Generation Networks for Joint Content and Style Representation

Generating Chinese fonts composed of a large number of characters is a challenging problem, and existing font generation methods are learned under the supervision of a large amount of paired data. Collecting such data is labor-intensive, and it is difficult to extend new styles of fonts. In order to assist font designers to improve the development efficiency of computer Chinese character fonts, an unsupervised font generation network that separates font content and style representation is proposed, and the two representations are established in the same domain. Dense semantic correspondence, and then used to guide the decoder to generate high-quality fonts Quality output. Deformable convolution is introduced in the skip connection, and the model is more focused on the structural features of fonts by learning the interdependence between offsets and channels. In addition, a multi-scale style discriminator is designed, which can be used on different scales Evaluate the style consistency of generated images. Extensive experiments show that the model in this paper achieves better results than the state-of-the-art font generation methods. In addition, the proposed model is able to learn font styles of other languages, and can generate font style.