面向图像场景转换的改进型生成对抗网络  

Improved Generative Adversarial Network for Image Scene Transformation

作　　者：肖进胜[1] 周景龙 雷俊锋[1] 李亮 丁玲 杜治一 XIAO Jin-Sheng;ZHOU Jing-Long;LEI Jun-Feng;LI Liang;DING Ling;DU Zhi-Yi(Electronic Information School,Wuhan University,Wuhan 430072,China;College of Computer,Hubei University of Education,Wuhan 430205,China)

机构地区：[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

出　　处：《软件学报》2021年第9期2755-2768,共14页Journal of Software

基　　金：国家重点研发计划(2017YFB1302401);国家自然科学基金(61471272)。

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.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.

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 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]