Computer Engineering and Applications ›› 2022, Vol. 58 ›› Issue (9): 230-237.DOI: 10.3778/j.issn.1002-8331.2010-0139

• Graphics and Image Processing • Previous Articles     Next Articles

Research on Method of Animated Avatar Generation Based on Multi-Level Generative Adversarial Networks

GAO Wenchao, REN Shengbo, TIAN Chi, ZHAO Shanshan   

  1. School of Mechanical Electronic & Information Engineering, China University of Mining &Technology, Beijing, Beijing 100083, China
  • Online:2022-05-01 Published:2022-05-01



  1. 中国矿业大学(北京) 机电与信息工程学院,北京 100083

Abstract: The existing animation image generation methods have the problems of lack of diversity in synthetic images, unclear local textures, small sample variance, and difficulty in generating according to detailed descriptions. Based on the idea of ??StackGAN++, combined with auxiliary classifiers, this paper proposes an improved model ACM-GAN(auxiliary classification atteched multi-level generative adversial networks, a multi-level structure with auxiliary classifiers) for Anime character avatar generation. The network model is composed of two generators and two discriminators stacked, and the auxiliary classifier is used to constrain the generated results to increase the variance of the generated samples and increase the diversity of the generated samples. To ensure that the synthesized image is true, the loss of feature map space and the loss of image pixel space mean variance are introduced to minimize the distance between synthetic data and real data. The experimental results show that the multi-level structure can effectively stabilize the training process, increase the edge details and local texture of the image, and at the same time auxiliary classification effectively solve the pattern collapse problem, and improve the accuracy of generating images of the specified category. The FID score of the iamge generated by ACM-GAN reaches 27.96, which is an increase of 23.1% compared StackGAN++.

Key words: anime avatar generation, generative adversarial networks, multi-level structure, auxiliary classifier

摘要: 现有的动画图像生成方法存在合成图像多样性缺失、局部纹理不清晰、样本方差较小,难以根据细节描述进行生成的问题。基于堆叠式生成对抗网络(StackGAN++)的思想,结合辅助分类器,提出改进模型ACM-GAN(auxiliary classification atteched multi-level generative adversial networks,带有辅助分类器的多层次结构生成对抗网络)用于动画人物头像生成。该网络模型由两个生成器和两个判别器堆叠而成,并在判别器中嵌入辅助分类器对生成结果进行约束,使生成样本方差变大,增加生成样本的多样性。为保证合成图像真实度和清晰度,引入特征图空间损失和图像像素空间均值方差损失以最小化合成数据和真实数据的距离。实验结果表明,多层次结构能够有效稳定训练过程,增加图像的边缘细节和局部纹理,同时辅助分类器有效解决模式崩溃问题,提高生成指定类别图像的准确率。ACM-GAN生成图像的FID分数达到27.96,相比于StackGAN++提升23.1%。

关键词: 动画头像生成, 生成对抗网络, 多层次结构, 辅助分类器