Chinese Font Generation Models Based On Chinese Character Prior Information

Chinese characters as a font with a long history,have various forms of expression that emphasize both imagery and emotional expression.They are widely used in practical life and application scenarios,which has led to increasing attention on Chinese character generation technology.The current mainstream models for generating Chinese characters are based on deep learning techniques,especially generative adversarial network(GAN).Although the existing GAN-based models have achieved good performance,they still suffer from mode collapse issues such as stroke redundancy and missing character styles,and are difficult to directly apply to few-shot Chinese character generation.To address these issues,this paper proposes two effective Chinese character generation models based on prior information such as stroke and structural information,respectively solving the mode collapse issues of existing GAN-based Chinese character generation models and the problem of generating few-shot Chinese character fonts.The main contributions of this paper are summarized as follows:(1)To address the stroke redundancy and incomplete character style issues of existing GAN-based deep Chinese character font generation models,this paper proposes a semi-supervised Chinese character generation model based on stroke encoding(referred to as Stroke GAN+),which introduces a small amount of paired dataset using semi-supervised learning to effectively guide the model in capturing Chinese character mode information.The small amount of paired dataset can provide important Chinese character priors without increasing the complexity of the model.Extensive experimental results demonstrate that the proposed model can significantly alleviate mode collapse issues such as stroke redundancy and incomplete character style,and outperforms existing methods in Chinese character font generation and has better generalization ability.(2)To improve the performance of the current mainstream few-shot Chinese character generation models based on style and content priors,this paper proposes a few-shot Chinese character generation method based on contrastive learning and square-block transformation(referred to as STCL-Font),which can adaptively capture font content and style information through contrastive learning and further enhance the structural prior of Chinese characters through square-block transformation data augmentation.The proposed contrastive learning can decouple Chinese character font style and content by adaptively capturing their similarity and difference.The squareblock transformation data augmentation method not only achieves data augmentation,but also effectively reflects the structural prior of Chinese character fonts.Extensive experimental results demonstrate that the proposed model can effectively decouple Chinese character font style and content,significantly improve few-shot Chinese character font generation performance,and outperform existing methods in both generation quality and robustness.