Research On Automatic Chinese Font Synthesis Based On Generative Adversarial Networks

Since the digital age,automatic Chinese font synthesis has been an important topic in the field of computer vision.The traditional methods generate new Chinese characters by disassembling the stroke structures of existing Chinese characters and then reorganizing these stroke structures.Rencently,with the continuous development of deep learning,most researchers have used methods based on convolutional neural networks or generative adversarial networks to generate Chinese character fonts.These end-to-end approaches have made good progress.However,for the task of generating Chinese fonts,it is necessary not only to ensure the consistency and authenticity of all Chinese character styles,but also to ensure the integrity of the stroke structure of each Chinese character.Faced with complex structure and diverse Chinese characters,the existing algorithms are not satisfactory entirely.This paper first constructs two datasets for the study of automatic font synthesis,and proposes two novel algorithms SAFont and DMFont to deal with the difficulties of the task of automatic Chinese font synthesis.The main work of this paper includes:1.Two datasets composed of multi-font Chinese characters and the calligraphy works of Wang Xizhi are used for the research of automatic Chinese font synthesis algorithms.The multi-font Chinese character dataset covers 100 commonly used fonts on social platforms,and each font contains about 4,000 Chinese character images.The images in the dataset of Wang Xizhi's calligraphy works are mainly from the calligraphy work of Lanting Collection,which contains about 320 images.2.Aiming at the problem that the results generated by the algorithms based on generative adversarial network tend to lack hierarchical structures between strokes,a novel automatic Chinese font synthesis algorithm based on self-attention mechanisms(SAFont)is proposed.The introduce of selfattention mechanisms enables the model to capture the hierarchical structure between strokes in Chinese characters during the image encoding and decoding stage,so that the model can generate Chinese characters with more distinct structural levels.At the same time,the edge loss is designed to force the model to pay more attention to the edge pixels in the Chinese character image.Extensive experimental results on the multi-font Chinese character dataset show that SAFont is able to generate Chinese characters with more vivid structure,richer details,and clearer outlines.3.Aiming at the problem that SAFont generates poor quality results on complex fonts and complex Chinese characters,a novel automatic Chinese font synthesis algorithm based on self-attention mechanisms(DMFont)is proposed.This network is a multi-model structure composed of multiple SAFont in series.Through massive bypass settings and feature maps reuse,DMFont is able to retain as much structure and position information as possible in complex Chinese characters or fonts,thereby generating results with better quality.At the same time,the strategy of phased training is used to optimize the parameters of DMFont.Extensive experiments on complex fonts and complex Chinese character datasets(both from the multi-font Chinese character dataset)and the dataset of Wang Xizhi's calligraphy works show that DMFont can generate Chinese characters with better quality.