基于感知对抗网络的图像风格迁移方法研究  

Image style transfer based on perceptual adversarial networks

作　　者：李君艺[1] 尧雪娟[1] 李海林 LI Junyi;YAO Xuejuan;LI Hailin(School of Engineering Technology, Guangdong Polytechnic Institute, Guangzhou 510091, China)

机构地区：[1]广东理工职业学院工程技术学院,广东广州510091

出　　处：《合肥工业大学学报（自然科学版）》2020年第5期624-628,共5页Journal of Hefei University of Technology：Natural Science

基　　金：广东省普通高校特色创新类资助项目(2019KTSCX247);广东大学生科技创新培育专项资金资助项目(pdjh2019b0732);广东开放大学校级科研课题重点资助项目(20281935)。

摘　　要：基于有关图像风格迁移研究成果,结合生成式对抗网络(generative adversarial network,GAN)和感知损失函数,文章提出了基于感知对抗网络(perceptual adversarial network,PAN)的图像风格迁移方法。该方法利用感知损失函数进行对抗训练,通过图像转换网络与判别网络之间的交替优化,生成图与原图在多个网络层次的特征差异能被持续发掘,使生成图的内容和风格更接近原图。实验结果表明,基于PAN的图像风格迁移方法能取得更佳效果。By combining the method of generative adversarial network(GAN)and perception loss function,this paper proposes an improved image style transfer method based on perceptual adversarial network(PAN).Different from the fixed loss network in the existing study of image style transfer,the image transformation network and the discriminative network are in adversarial training using the perception loss function.As both networks are alternatively optimized,the differences between the generated graph and the original graph at multiple network levels can be continuously explored.The generated graph can be more similar to the original content image and the style image.By qualitative and quantitative comparison in the experiment,image style transfer based on PAN is proved to be more effective than previous methods.

关 键 词：图像风格迁移 生成式对抗网络(GAN) 感知损失 感知对抗网络(PAN) 

分 类 号：TP391.9[自动化与计算机技术—计算机应用技术]