基于注意力机制的人脸漫画生成方法

A method of generating facial cartoons based on attention modules

在漫画生成领域,针对现有的基于生成对抗网络(generative adversarial networks,GAN)的方法无法捕获漫画和人脸之间高级对应关系的问题,提出了一个基于注意力机制的生成对抗网络模型(Attentions-based GAN)。在生成器和多尺度鉴别器上创造性地利用不同的注意力模块以保留重要的通道特征和空间特征,保证了鉴别器更准确地区分真实漫画和生成漫画,并促使生成器学习漫画域的分布。此外,构建了一个高分辨率的人脸漫画数据集,并在其上验证了本文模型的有效性。大量实验表明,本文方法生成的漫画更清晰真实,可以保留人物面部的细节并延续漫画的风格。

In the field of caricature generation,the existing methods based on generative adversarial networks(GAN)cannot capture the high-level correspondence between caricatures and faces.In order to solve the problem,a new model named attention-based generative adversarial network model(Attentions-based GAN)was proposed.Different attention modules were creatively used on the generator and multi-scale discriminators respectively to preserve the important channel features and spatial features,to ensure that the discriminator can more accurately distinguish the real and generative caricatures,and to encourage the generator to learn the distribution of the caricature domain.In addition,a high-resolution face caricature dataset was constructed and the effectiveness of the proposed model was verified on this dataset.Extensive experiments show that the proposed method can generate clearer and more realistic caricatures,retaining the details of the characters’faces and keeping the style of the caricatures.