稠密自适应生成对抗网络的爨体字风格迁移模型

Cuan Font Generation Model of Dense Adaptive Generation Adversarial Network

爨体字作为典型的衬线字体,不同于黑体、微软雅黑、等线这些非衬线字体,其字形结构十分多样.为了防止爨体字在生成过程中笔画弯折处出现伪影和模糊的现象,提出一种基于稠密自适应生成对抗网络的爨体字风格迁移模型.首先,生成器通过稠密自适应卷积块更加充分地提取风格特征和内容特征;然后,像素判别器对真实图片和生成图片进行分辨;再采用对抗损失、迁移损失、梯度损失和边缘损失对生成网络进行参数调节;最后,将自行采集的爨体字数据集作为训练集送入模型进行训练.实验结果证明,所提模型能够有效地学习到风格特征,达到更好的生成效果;其生成结果在字形大小保持上优于Zi-to-zi模型,在笔画细节特征的保留上优于StarGANv2以及CycleGAN模型,并在SSIM和L1 loss指标上得到了验证.

As a typical serif font,Cuan font is different from the sans serif fonts such as“Heiti”,"Microsoft"“Yahei”and“Isoline”,and its glyph structure is very diverse.In order to overcome the phenomenon of arti-facts and blurring at the bends of the strokes in the generation process of Cuan fonts,a style transfer model of Cuan fonts based on a dense adaptive generation adversarial network is proposed.Firstly,the generator extracts style features and content features more fully through dense adaptive convolution blocks.Secondly,the pixel discriminator distinguishes the real picture from the generated picture.Thirdly,use adversarial loss,migration loss,gradient loss and edge loss to adjust the parameters of the generation network.Finally,the self-collected Cuan font data set is used as a training set and sent to the model for training.The experimental results prove that the style transfer model of Cuan font can effectively learn style features to achieve better generation effect.The generated results are better than the Zi-to-zi model in terms of font size retention,and better than the StarGANv2 and CycleGAN models in terms of the retention of stroke detail features,and are verified on the SSIM and L1 loss indicators.