第15卷第1期 智能系统学报 Vol.15 No.1 2020年1月 CAAI Transactions on Intelligent Systems Jan.2020 D0L:10.11992tis.202002002 基于生成对抗网络的机载遥感图像超分辨率重建 毕晓君，潘梦迪 (1.中夹民族大学信息工程学院，北京100081；2.哈尔滨工程大学信息与通信工程学院，黑龙江哈尔滨150001) 摘要：为解决机载遥感图像质量易受环境影响的问题，对其进行超分辨率重建，对现有深度学习机载遥感图 像超分辨率重建方法存在的特征提取能力差、重建图像边缘平滑、模型训练困难的问题进行改进，增强图像重 建效果。将生成对抗网络作为模型的整体框架，使用密集剩余残差块增强模型特征提取能力，增加跳跃连接， 有效提取机载遥感图像的浅层和深层特征，引入沃瑟斯坦式生成对抗网络优化模型训练。该方法能够有效对 机载遥感图像进行4倍重建，在峰值信噪比评价上较对比方法约有2B增益，重建出的机载遥感图像在视觉 上更清晰、细节更丰富、边缘更锐利。实验结果表明，该方法有效提升了模型特征提取能力，优化了训练过程， 重建的机载遥感图像效果较好。 关键词：机载遥感；超分辨率重建；深度学习：密集剩余残差块；特征提取；跳跃链接；沃瑟斯坦；生成对抗网络 中图分类号：TP751.1文献标志码：A文章编号：1673-4785(2020)01-0074-10 中文引用格式：毕晓君，潘梦迪.基于生成对抗网络的机载遥感图像超分辨率重建.智能系统学报，2020,15(1)：74-83. 英文引用格式：BI Xiaojun,PAN Mengdi.Super--resolution reconstruction of airborne remote sensing images based on the generat- ive adversarial networks[Jl.CAAI transactions on intelligent systems,2020,15(1):74-83. Super-resolution reconstruction of airborne remote sensing images based on the generative adversarial networks BI Xiaojun',PAN Mengdi? (1.School of Information Engineering,Minzu University of China,Beijing 100081,China;2.Department of Information and Com- munication Engineering,Harbin Engineering University,Harbin 150001,China) Abstract:To solve the problem that the quality of airborne remote sensing images is susceptible to environmental im- pacts,super-resolution reconstruction is carried out.The existing super-resolution reconstruction methods for deep learn- ing airborne remote sensing images has the problems of poor feature extraction capability,smooth edges of reconstruc- ted images and difficulty in model training,the image reconstruction effect is enhanced to solve the above problems. The generative adversarial network is taken as the overall framework of the model.The dense residual block is used to enhance the feature extraction capability of the model,and jump connection is added to effectively extract the shallow and deep features of airborne remote sensing images.The Wasserstein-type generative adversarial network optimization model training is introduced.The method can effectively reconstruct airborne remote sensing images by 4 times,and has a gain of 2 dB or so in peak signal-to-noise ratio evaluation compared with other methods for comparison.The recon- structed airborne remote sensing images are clearer in vision,richer in details and sharper in edges.The experimental results show that the method effectively improves the model feature extraction ability,optimizes the training process, and the reconstructed airborne remote sensing image has better effect. Keywords:airborne remote sensing;super-resolution reconstruction;deep learning;residual in residual dense block; feature extraction;jump connection:Wasserstein;generative adversarial network 近年来，随着人们对航空航天领域开发的重 收稿日期：2020-02-04. 基金项目：国家自然科学基金项目(51779050). 视，卫星、无人机等设备被应用到日常信息获取 通信作者：毕晓君.E-mail:bixiaojun@hrbeu.edu.cn 中来，由它们获取到的遥感图像具有面积大、范DOI: 10.11992/tis.202002002 基于生成对抗网络的机载遥感图像超分辨率重建 毕晓君1 ，潘梦迪2 （1. 中央民族大学 信息工程学院，北京 100081; 2. 哈尔滨工程大学 信息与通信工程学院，黑龙江 哈尔滨 150001） 摘 要：为解决机载遥感图像质量易受环境影响的问题，对其进行超分辨率重建，对现有深度学习机载遥感图 像超分辨率重建方法存在的特征提取能力差、重建图像边缘平滑、模型训练困难的问题进行改进，增强图像重 建效果。将生成对抗网络作为模型的整体框架，使用密集剩余残差块增强模型特征提取能力，增加跳跃连接， 有效提取机载遥感图像的浅层和深层特征，引入沃瑟斯坦式生成对抗网络优化模型训练。该方法能够有效对 机载遥感图像进行 4 倍重建，在峰值信噪比评价上较对比方法约有 2 dB 增益，重建出的机载遥感图像在视觉 上更清晰、细节更丰富、边缘更锐利。实验结果表明，该方法有效提升了模型特征提取能力，优化了训练过程， 重建的机载遥感图像效果较好。 关键词：机载遥感；超分辨率重建；深度学习；密集剩余残差块；特征提取；跳跃链接；沃瑟斯坦；生成对抗网络 中图分类号：TP751.1 文献标志码：A 文章编号：1673−4785(2020)01−0074−10 中文引用格式：毕晓君, 潘梦迪. 基于生成对抗网络的机载遥感图像超分辨率重建 [J]. 智能系统学报, 2020, 15(1): 74–83. 英文引用格式：BI Xiaojun, PAN Mengdi. Super-resolution reconstruction of airborne remote sensing images based on the generat￾ive adversarial networks[J]. CAAI transactions on intelligent systems, 2020, 15(1): 74–83. Super-resolution reconstruction of airborne remote sensing images based on the generative adversarial networks BI Xiaojun1 ，PAN Mengdi2 (1. School of Information Engineering, Minzu University of China, Beijing 100081, China; 2. Department of Information and Com￾munication Engineering, Harbin Engineering University, Harbin 150001, China) Abstract: To solve the problem that the quality of airborne remote sensing images is susceptible to environmental im￾pacts, super-resolution reconstruction is carried out. The existing super-resolution reconstruction methods for deep learn￾ing airborne remote sensing images has the problems of poor feature extraction capability, smooth edges of reconstruc￾ted images and difficulty in model training, the image reconstruction effect is enhanced to solve the above problems. The generative adversarial network is taken as the overall framework of the model. The dense residual block is used to enhance the feature extraction capability of the model, and jump connection is added to effectively extract the shallow and deep features of airborne remote sensing images. The Wasserstein-type generative adversarial network optimization model training is introduced. The method can effectively reconstruct airborne remote sensing images by 4 times, and has a gain of 2 dB or so in peak signal-to-noise ratio evaluation compared with other methods for comparison. The recon￾structed airborne remote sensing images are clearer in vision, richer in details and sharper in edges. The experimental results show that the method effectively improves the model feature extraction ability, optimizes the training process, and the reconstructed airborne remote sensing image has better effect. Keywords: airborne remote sensing; super-resolution reconstruction; deep learning; residual in residual dense block; feature extraction; jump connection; Wasserstein; generative adversarial network 近年来，随着人们对航空航天领域开发的重 视，卫星、无人机等设备被应用到日常信息获取 中来，由它们获取到的遥感图像具有面积大、范 收稿日期：2020−02−04. 基金项目：国家自然科学基金项目 (51779050). 通信作者：毕晓君. E-mail：bixiaojun@hrbeu.edu.cn. 第 15 卷第 1 期 智 能 系 统 学 报 Vol.15 No.1 2020 年 1 月 CAAI Transactions on Intelligent Systems Jan. 2020