基于StarGANv2的多风格字体生成研究

Research of multi-style font generation based on StarGANv2

目的:实现多种风格域的无监督字体风格迁移。方法:使用StarGANv2网络模型,对生成器进行改进。引入IBN-Net,将实例标准化(Instance normalization,IN)和批标准化(Batch normalization,BN)结合,形成残差网络的基础模块。然后在此模型的风格编码网络中增加注意力机制,进一步对不同风格字体挖掘关键性区域特征,增强目标域的差异化特征分布,实现对小样本字体图像抽取更加丰富的特征信息。结果:在不同风格字体的数据集上来验证改进模型的有效性,研究表明,在像素级、感知级等评价指标中均优于其他算法。相比原始模型,本文方法生成的汉字图像,FID下降5.12,LPIPS下降0.0621,SSIM增加0.0411。并且视觉上有更好的生成效果,保留了更多的细节信息。结论:改进的模型能够有效的生成高质量的多风格字体汉字。

Aims:This paper aims to realize the unsupervised font style transfer of multiple style domains to get rid of the shackles of data set labels.Methods:We improved the generator with the StarGANv2 net model.IBN-Net was introduced to combine instance normalization(IN)and batch normalization(BN)to form the basic module of the residual network.Then,an attention mechanism was added to the style encoding network of this model to further mine the key regional features of the different styles of fonts,to enhance the differentiated feature distribution of the target domain,and to realize the extraction of more abundant feature information from small sample font images.Results:The effectiveness of the improved model was verified on the datasets of the different styles of fonts;and the results showed that it was superior to other algorithms in the evaluation indicators such as the pixel level and the perception level.Compared with the original model,the Chinese character images generated by this method had a decrease of 5.12 in FID,a decrease of 0.0621 in LPIPS,and an increase of 0.0411 in SSIM.And it had a better visual effect and retained more detailed information.Conclusions:The improved model can effectively generate high-quality multi-style Chinese characters.