基于改进条件生成对抗网络的字体风格迁移算法

Font Style Transfer Method Based on Improved Conditional Generative Adversarial Network

为解决现有字体迁移风格网络难以快速收敛以及处理复杂字体结构能力较弱等问题,提出了一种基于条件生成对抗网络的汉字字体生成方法。通过条件生成对抗网络的方式训练生成器作为字体风格迁移网络,通过知识蒸馏技术将预训练的图像重建网络的特征信息引入网络,更好地将特征解码为目标风格字体,同时结合边缘平滑损失和感知损失提高目标字体的生成质量。与已有的字体生成算法进行定量分析与定性分析,在不同字体上进行的实验结果表明:该方法生成的目标字体更加真实并且文字的边缘更加清晰。

This paper introduces an approach for Chinese character font generating based on conditional generative adversarial networks,addressing the prevalent issues associated with sluggish convergence and intricate font structure handling found in conventional techniques.This method employs a generator trained by conditional generative adversarial networks as the core component of the font style transfer network.The method introduces a knowledge distillation technique that assimilates feature information from a pre-trained image reconstruction network to decode features with greater precision into target style fonts.Additionally,this method significantly bolsters the quality of generated target fonts by implementing a combination of edge smoothing loss and perceptual loss.This paper conducts a comprehensive analysis,both quantitatively and qualitatively,comparing various fonts to existing font generation algorithms.Experimental results conclusively demonstrate that the method generates more realistic target fonts with clearer defined character edges.