北京邮电大学学报

  • EI核心期刊

北京邮电大学学报 ›› 2024, Vol. 47 ›› Issue (4): 117-123.

• 论文 • 上一篇    下一篇

基于改进生成对抗网络的动漫图像风格迁移算法

李云红, 朱景坤, 刘杏瑞, 陈锦妮, 苏雪平   

  1. 西安工程大学 电子信息学院
  • 收稿日期:2023-07-11 修回日期:2023-09-13 出版日期:2024-08-28 发布日期:2024-08-26
  • 通讯作者: 李云红 E-mail:liyunhong@xpu.edu.cn
  • 基金资助:
    陕西省自然科学基础研究计划项目

Anime Image Style Transfer Algorithm Based on Improved Generative Adversarial Networks

LI Yunhong, ZHU Jingkun, LIU Xingrui, CHEN Jinni, SU Xueping   

  • Received:2023-07-11 Revised:2023-09-13 Online:2024-08-28 Published:2024-08-26
  • Contact: 云红 李 E-mail:liyunhong@xpu.edu.cn

摘要: 针对动漫风格迁移图像存在细节结构缺失、色彩偏移、语义内容伪影等问题,提出了一种改进型生成对抗网络的动漫风格迁移算法。首先,利用通道混洗操作结合改进后的反转残差块组成特征转换模块增强图像的局部特征属性,同时引用高效注意力机制进一步提升风格特征表达能力;其次,改进风格损失函数,抑制亮度和色彩对高频纹理信息学习的干扰;最后,将含有随机噪声的内容图像输入生成器,并在判别器中引入谱约束层限制谱半径,以解决训练过程中出现模式崩溃的问题。实验结果表明,所提算法生成的图像相比于其他算法细节刻画更为丰富,并且有效避免了伪影的出现与色彩的偏移,使生成图像具有更强的写实感,风格 Frechet 起始距离分别达到了154.61 和 115.64

关键词: 生成对抗网络, 风格迁移, 残差块, 风格损失

Abstract:  An improved anime style transfer algorithm for generative adversarial networks is proposed to address the issues of missing detail structure, color shifting, and semantic content artifacts. Firstly, a feature transformation module is constructed by combining channel shuffle operations with improved inverted residual blocks to enhance the local feature attributes of the image, and an efficient attention mechanism is incorporated to further improve the style feature representation capability. Secondly, the style loss function is modified to suppress the influence of brightness and color variations on high-frequency texture learning. Finally, content images containing random noise are fed into the generator and a spectral normalization constraint is applied to the discriminator network to address the issue of mode collapse. The experimental results demonstrate that the image generated by the proposed method is richer in detail than other algorithms, and effectively avoiding the occurrence of artifacts and color shifting, so that the generated image will have a greater sense of realism, achieving style Frechet inception distance of 154.61 and 115.64, respectively.

Key words: generative adversarial networks, style transfer, residual block, style loss

中图分类号: