Page 131 - 《软件学报》2021年第9期

软件学报 ISSN 1000-9825, CODEN RUXUEW E-mail: jos@iscas.ac.cn
Journal of Software,2021,32(9):2755−2768 [doi: 10.13328/j.cnki.jos.005986] http://www.jos.org.cn
©中国科学院软件研究所版权所有. Tel: +86-10-62562563

∗
面向图像场景转换的改进型生成对抗网络

1
1
2
1
1
肖进胜 , 周景龙 , 雷俊锋 , 李 亮 , 丁 玲 , 杜治一 1
1
(武汉大学 电子信息学院,湖北 武汉 430072)
2 (湖北第二师范学院 计算机学院,湖北 武汉 430205)
通讯作者: 肖进胜, E-mail: xiaojs@whu.edu.cn

摘 要: 设计了新的生成器网络、判决器网络以及新的损失函数,用于图像场景转换.首先,生成器网络采用了带跨
层连接结构的深度卷积神经网络,其中,多个跨层连接以实现图像结构信息的共享;而判决器网络采用了多尺度全域
卷积网络,多尺度判决器可以区分不同尺寸下的真实和生成图像.同时,对于损失函数,该算法借鉴其他算法提出了 4
种损失函数的组合,并通过实验对比证明了新损失函数的有效性,包括 GAN 损失、L 1 损失、VGG 损失、FM 损失.
从实验结果显示,该算法能够实现多种转换,且转换后图像的细节保留较为完整,生成图像较为真实,明显消除了块
效应.
关键词: 图像生成;深度学习;生成对抗网络;跨层连接;场景转换
中图法分类号: TP183

中文引用格式: 肖进胜,周景龙,雷俊锋,李亮,丁玲,杜治一.面向图像场景转换的改进型生成对抗网络.软件学报,2021,32(9):
2755−2768. http://www.jos.org.cn/1000-9825/5986.htm
英文引用格式: Xiao JS, Zhou JL, Lei JF, Li L, Ding L, Du ZY. Improved generative adversarial network for image scene
transformation. Ruan Jian Xue Bao/Journal of Software, 2021,32(9):2755−2768 (in Chinese). http://www.jos.org.cn/1000-9825/
5986.htm
Improved Generative Adversarial Network for Image Scene Transformation

1
1
1
2
1
XIAO Jin-Sheng , ZHOU Jing-Long , LEI Jun-Feng , LI Liang , DING Ling , DU Zhi-Yi 1
1
(Electronic Information School, Wuhan University, Wuhan 430072, China)
2
(College of Computer, Hubei University of Education, Wuhan 430205, China)
Abstract: This study designs a new generator network, a new discriminator network, and a new loss function for image scene
conversion. First, the generator network uses a deep convolutional neural network with a skip connection structure, in which multi-skip
connection is used to share the structure information of the image. For the discriminator network, it uses a multi-scale global
convolutional network which can distinguish between real and generated images of different sizes. At the same time, the new loss function
is a combination of four loss functions referring to other algorithms, including GAN loss, L 1 loss, VGG loss, and feature matching loss.
Moreover, the validity of the new loss function is demonstrated through experimental comparisons. The experimental results show that the
proposed algorithm can achieve multi-image transformations, and the details of generated images are preserved completely, the generated
image is more realistic, and the block effect is obviously eliminated.
Key words: image generation; deep learning; generative adversarial networks; skip connection; scene conversion

许多计算机视觉问题可以被看作是一个图像到图像的翻译问题,是映射一个域中的映像到另一个域中的
对应映像,实际上都是像素到像素之间的映射.例如:超分辨率可以认为是将低分辨率图像映射到相应的高分辨

∗ 基金项目: 国家重点研发计划(2017YFB1302401); 国家自然科学基金(61471272)
Foundation item: National Key Research and Development Program of China (2017YFB1302401); National Natural Science
Foundation of China (61471272)
收稿时间: 2019-07-19; 修改时间: 2019-09-03, 2019-10-21; 采用时间: 2019-11-19