基于改进SinGAN的遥感图像数据增强方法  

Remote sensing image data enhancement based on improved SinGAN

作　　者：彭晏飞[1] 邓佳楠 王刚[2] PENG Yan-fei;DENG Jia-nan;WANG Gang(School of Electronic and Information Engineering,Liaoning Technical University,Huludao 125105,China;Bohai Shipbuilding Vocationla College,Huludao 125105,China)

机构地区：[1]辽宁工程技术大学电子与信息工程学院,辽宁葫芦岛125105 [2]渤海船舶职业学院,辽宁葫芦岛125105

出　　处：《液晶与显示》2023年第3期387-396,共10页Chinese Journal of Liquid Crystals and Displays

基　　金：国家自然科学基金(No.61772249);辽宁省高等学校基本科研项目(No.LJKZ0358);辽宁工程技术大学双一流学科创新团队资助项目(No.LNTU20TD-27)。

摘　　要：随着遥感技术的发展,遥感图像被大量应用于遥感图像识别、分割等领域。但遥感图像数量少、质量低以及多样性不足等问题阻碍了遥感解译等后续研究的性能提升,如何利用少量的遥感图像生成大量的数据集是目前亟待解决的问题。针对这一问题,本文以SinGAN网络为基础,将一种新的纯卷积网络ConvNeXt与之结合来构建遥感图像数据增强框架。实验结果表明,在遥感数据集NWPU-RESISC45 Dataset上,结合ConvNeXt卷积网络进行数据增强后的图像,其FID、SSIM和PSNR这3种GAN与图像质量评价指标分别提升了5.7%、6.2%和8.2%。基于改进SinGAN的遥感图像数据增强方法进行增强后的图像质量与多样性均优于SinGAN算法与传统图像增强方法,实际中可用于图像分割、目标检测等领域。With the development of remote sensing technology,remote sensing images have been applied to a large number of fields such as remote sensing image recognition and segmentation detection.However,the problems of lack of remote sensing images,low quality and insufficient diversity hinder the performance improvement of remote sensing interpretation and other subsequent researches,and how to use a small amount of remote sensing images to generate a large number of datasets is an urgent problem at present.To address this problem,this paper combines a new pure convolutional network,ConvNeXt,with SinGAN network to build a remote sensing image data enhancement framework.Combined with ConvNeXt convolution network,the three image quality evaluation indexes of FID,SSIM and PSNR are increased by 5.7%,6.2%and 8.2%,respectively,on the remote sensing dataset NWPU-RESISC45 Dataset after combining ConvNeXt convolutional network for data enhancement.The quality and diversity of the data enhanced images based on the improved SinGAN remote sensing image data enhancement method are better than the SinGAN algorithm and the traditional image enhancement method,which can be used in remote sensing interpretation,change detection and other fields in practice.

关 键 词：SinGAN 数据增强 遥感图像 ConvNeXt 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]