摘要
基于深度学习的超分辨率重建方法多数采用已知的模糊核训练网络,在实际应用中模糊核通常未知,在此情况下这类方法的重建效果将显著下降。零样本超分方法利用图像自身构建训练集,能够改善由于模糊核未知所带来的性能下降,但由于仅利用图像自身信息,对重建效果的提升有一定的局限性。本文提出增强少样本学习方法解决模糊核未知时的超分重建问题,一方面,选取与低分图像类似的示例图像构建训练集;另一方面,扩大网络规模并优化网络结构。在UCMerced_LandUse数据集上的实验结果表明,与零样本超分方法相比,本文所提方法具有更好的超分重建效果。
Most of Super-Resolution(SR)methods based on deep learning train the network with a known blur kernel.However,the blur kernels are usually unknown in realistic applications,resulting in severe performance drop for these SR methods.Zero-Shot SR constructs the training set with the input image itself,improving the performance while the blur kernels are unknown.But the information of the input image itself is very limited,so the improvement is limited.This paper proposes a method called Enhanced Few-Shot Super-Resolution to solve the problem of the unknown blur kernels in this paper.On the one hand,it constructs the training set with the low-res image itself and images similar to it.On the other hand,it enlarges the network and optimizes the network structure.The result of the experiment on UCMerced_LandUse shows that the method achieves better performance than ZSSR.
作者
李盛
潘宗序
雷斌
丁赤飚
LI Sheng;PAN Zong-xu;LEI Bin;DING Chi-biao(Aerospace Information Research Institute,Chinese Academy of Sciences,Beijing 100190 China;School of Electronic,Electrical and Communication Engineering,University of ChineseAcademy of Sciences,Beijing 100049 China;Key Laboratory of Technology in Geo-spatial Information Processing and Application System,Chinese Academy of Sciences,Beijing 100190 China;National Key Lab of Microwave Imaging Technology,Chinese Academy of Sciences,Beijing 100190 China)
出处
《自动化技术与应用》
2021年第6期1-5,共5页
Techniques of Automation and Applications
关键词
遥感图像
深度学习
增强少样本超分
盲超分
Remote Sensing Images
deep learning
Enhanced Few-Shot Super-Resolution
blind super-resolution
