计算机工程与应用 ›› 2020, Vol. 56 ›› Issue (22): 217-223.DOI: 10.3778/j.issn.1002-8331.1910-0405

• 图形图像处理 • 上一篇    下一篇

引入特征损失对CycleGAN的影响研究

刘华超,张俊然,刘云飞   

  1. 四川大学 电气工程学院,成都 610065
  • 出版日期:2020-11-15 发布日期:2020-11-13

Influence of Identity Loss on CycleGAN

LIU Huachao, ZHANG Junran, LIU Yunfei   

  1. College of Electrical Engineering, Sichuan University, Chengdu 610065, China
  • Online:2020-11-15 Published:2020-11-13

摘要:

在图像生成领域,传统的图像风格迁移需要在两个配对的图像间进行转换。循环一致性生成对抗网络(Cycle Generative Adversarial Network,CycleGAN)在2017年被提出后,凭借其可以针对非配对图像进行图像生成的特点取得了良好的效果,迅速成为图像生成领域的研究热点。然而经典的CycleGAN由于生成器无法准确识别图像的特定转换域和无关域,从而存在图像无关域特征随意变换的缺点,使得生成图像失真。针对以上问题,通过引入特征损失来约束生成器的特征识别,利用L1损失保证转换后的图像与原图像的像素级别对应,可以有效改善该问题,并使得生成图像更清晰。通过调整特征损失的比例超参数[μ],进一步分析了选取不同[μ]值下的特征损失时,CycleGAN的各部分损失变化以及对生成图像的质量影响,最后给出了特征损失的比例超参数[μ]值的选取策略。

关键词: 图像生成, CycleGAN, 特征损失, 图像清晰度, 风格迁移

Abstract:

In the field of image generation, traditional image style transfer requires conversion between two paired images. CycleGAN(Cycle Generative Adversarial Network) has achieved good results by virtue of its ability to generate images for unpaired images, has become a research hotspot in the field of image generation. However, the classical CycleGAN generator cannot accurately identify the specific transformation domain and irrelevant domain of the image. Thus, it has the shortcoming of the arbitrary transformation of image irrelevant domain features, and the distortion of the generated image. For the above questions, identity loss is introduced to constrain the feature recognition of generator, L1 loss is used to guarantee that the converted image corresponds to the pixel level of the original image, effectively improve the problem, and make the generated image clearer. Meanwhile, by adjusting the ratio of identity loss [μ], the identity loss under different [μ] values is further analyzed, and the change of the loss of  CycleGAN and its effect on the quality of the generated image are discuessed. Finally, the selection strategy of proportional superparameter [μ] value of identity loss is given.

Key words: image generation, CycleGAN, identity loss, image sharpness, style transfer