基于特征对比的循环生成对抗网络图像风格转换研究  

Research on image style transformation of cyclic generation adversarial network based on feature contrast

作　　者：闫娟[1] 康鹏帅 王士斌[2,3] 梅学术 李燕 刘栋 Yan Juan;Kang Pengshuai;Wang Shibin;Mei Xueshu;Li Yan;Liu Dong(Information Construction and Management Office,Henan Normal University,Xinxiang 453007,China;School of Computer and Information Engineering,Henan Normal University,Xinxiang 453007,China;Key Lab of"Artificial Intelligence and Personalized Learning in Education"in Henan Province,Henan Normal University,Xinxiang 453007,China)

机构地区：[1]河南师范大学信息化建设与管理办公室,河南新乡453007 [2]河南师范大学计算机与信息工程学院,河南新乡453007 [3]河南师范大学“教育人工智能与个性化学习”河南省重点实验室,河南新乡453007

出　　处：《河南师范大学学报（自然科学版）》2024年第6期73-79,共7页Journal of Henan Normal University(Natural Science Edition)

基　　金：国家自然科学基金(62072160);河南省科技攻关计划项目(222102210187);河南省普通本科高等学校智慧教学专项研究项目(202111)。

摘　　要：无监督图像到图像转换任务是在非配对训练数据的情况下学习源域图像到目标域图像的转换.但是,图像风格转换任务依然面临着图像内容丢失、模型坍塌等现象.为了解决上述问题,提出了一种局部特征对比来保持图像内容,通过特征提取器获得多层图像深层特征,使得图像编码器学习到高级语义信息,获得信息更加丰富的图像特征.同时,增加局部特征对比损失来引导特征提取器学习到有利于图像内容生成的特征.实验结果表明,在大多数情况下,所提方法在FID和KID分数方面优于之前的方法,图像生成质量有一定的提升.The unsupervised image-to-image translation task is to learn the transformation of source domain images to target domain images in the case of unpaired training data.However,the image style conversion task still faces phenomena such as image content loss and model collapse.In order to solve the above problems,we propose a local feature comparison to preserve image content,and obtain multi-layer image deep features through a feature extractor,allowing the image encoder to learn high-level semantic information and obtain more informative image features.At the same time,local feature contrast loss is added to guide the feature extractor to learn features that are beneficial to image content generation.Experimental results show that in most cases,our method outperforms previous methods in terms of FID and KID scores,and the quality of image generation is improved to a certain extent.

关 键 词：特征对比 图像风格转换 对比损失 

分 类 号：TP399[自动化与计算机技术—计算机应用技术]