| 结构增强型生成对抗网络SAR图像超分辨率重建 |
| |
| 引用本文： | 闵锐,杨学志,董张玉,陈鲸.结构增强型生成对抗网络SAR图像超分辨率重建[J].地理与地理信息科学,2021,37(2):47-53. |
| |
| 作者姓名： | 闵锐 杨学志 董张玉 陈鲸 |
| |
| 作者单位： | 合肥工业大学计算机与信息学院,安徽 合肥230009;工业安全与应急技术安徽省重点实验室,安徽 合肥230009;工业安全与应急技术安徽省重点实验室,安徽 合肥230009;合肥工业大学软件学院,安徽 合肥230009;智能互联系统安徽省实验室,安徽 合肥230009;合肥工业大学计算机与信息学院,安徽 合肥230009;工业安全与应急技术安徽省重点实验室,安徽 合肥230009;智能互联系统安徽省实验室,安徽 合肥230009;合肥工业大学计算机与信息学院,安徽 合肥230009 |
| |
| 基金项目： | 安徽省科技攻关计划项目;中央高校基本科研业务费专项资金项目 |
| |
| 摘 要： | 针对利用生成对抗网络模型(Generative Adversarial Network,GAN)重建SAR(Synthetic Aperture Radar)图像存在边缘细节信息不足和“伪影”(artifacts)现象,该文基于增强型超分辨率生成对抗网络(Enhanced Super-Resolution Generative Adversarial Networks,ESRGAN)光学模型,重新设计生成网络上采样重建模块和结构损失函数,提出一种结构增强型生成对抗网络SAR图像超分辨率重建算法,包括特征提取、特征增强和上采样重建3个模块:在特征提取模块采用小尺度卷积层对输入SAR图像进行低层次特征提取;在特征增强模块采用多个级联残差密集块(Residual-in-Residual Dense Block,RRDB)和卷积层提取输入特征;在上采样重建模块交替使用最近邻插值(Nearest Neighbor Interpolation,NNI)和亚像素卷积(Sub-Pixel Convolution,SPC)对特征进行放大重建,使特征信息交互融合。与传统插值算法和经典深度学习重建算法相比,该算法在视觉效果和定量评价方面均有显著提升,能够在保持原网络模型重建图像内容信息不丢失的基础上,增强重建图像边缘细节信息和减缓“伪影”现象,有利于后续目标识别和灾害监测等工作开展。
|
| 关 键 词： | 生成对抗网络 超分辨率重建 合成孔径雷达图像 结构损失 残差密集块 |
SAR Image Super-Resolution Reconstruction Based on Enhanced Structural Generative Adversarial Network |
| |
| MIN Rui,YANG Xue-zhi,DONG Zhang-yu,CHEN Jing.SAR Image Super-Resolution Reconstruction Based on Enhanced Structural Generative Adversarial Network[J].Geography and Geo-Information Science,2021,37(2):47-53. |
| |
| Authors: | MIN Rui YANG Xue-zhi DONG Zhang-yu CHEN Jing |
| |
| Institution: | (School of Computer and Information,Hefei University of Technology,Hefei 230009;Anhui Province Key Laboratory of Industrial Safety and Emergency Technology,Hefei 230009;School of Software,Hefei University of Technology,Hefei 230009;Intelligent Interconnected System Laboratory of Anhui Province,Hefei 230009,China) |
| |
| Abstract: | Aiming at the insufficient edge detail information and"artifacts"in synthetic aperture radar(SAR)image super-resolution reconstruction by using the generative adversarial network(GAN)model,based on the ESRGAN(enhanced super-resolution generative adversarial networks)optical model,this paper redesigns the up-sampling reconstruction module and structure loss function in the generator network,and a SAR image super-resolution reconstruction algorithm based on enhanced structural generative adversarial network is proposed.In this algorithm,generator network is mainly divided into three modules:feature extraction,feature enhancement and up-sampling reconstruction.Firstly,the feature extraction module uses a small-scale convolution layer to extract low-level features from the input SAR images;then,the feature enhancement module uses multiple cascaded residual-in-residual dense blocks(RRDBs)and convolutional layer to extract the input features;finally,the up-sampling reconstruction module alternately uses nearest neighbor interpolation(NNI)and sub-pixel convolution(SPC)to enlarge and reconstruct the features,which improves the interactive fusion of feature information.Compared with traditional interpolation algorithms and typical deep learning reconstruction algorithms,the proposed algorithm has achieved significant improvements in visual effects and quantitative evaluation.It can enhance the edge detail information of the reconstructed images and reduce"artifacts"without losing image content information reconstructed by the original network model.This is conducive to further work such as target recognition and disaster monitoring,etc. |
| |
| Keywords: | generative adversarial network super-resolution reconstruction synthetic aperture radar image structure loss residual dense block |
| 本文献已被 CNKI 维普 万方数据 等数据库收录！ |
|
