东北大学学报(自然科学版) ›› 2024, Vol. 45 ›› Issue (8): 1178-1184.DOI: 10.12068/j.issn.1005-3026.2024.08.014

• 资源与土木工程 • 上一篇    下一篇

多尺度特征融合的Transformer遥感影像超分辨率重建

王植, 王坤, 王梦晴   

  1. 东北大学 资源与土木工程学院,辽宁 沈阳 110819
  • 收稿日期:2023-04-11 出版日期:2024-08-15 发布日期:2024-11-12
  • 作者简介:王 植(1979-),男,辽宁沈阳人,东北大学副教授.
  • 基金资助:
    中央高校基本科研业务费专项资金资助项目(N170113027)

Super-resolution Reconstruction of Remote Sensing Image Based on Transformer of Multi-scale Feature Fusion

Zhi WANG, Kun WANG, Meng-qing WANG   

  1. School of Resources & Civil Engineering,Northeastern University,Shenyang 110819,China. Corresponding author: WANG Zhi,E-mail: wangzhi@ mail. neu. edu. cn
  • Received:2023-04-11 Online:2024-08-15 Published:2024-11-12

摘要:

针对现有遥感影像超分辨率重建算法,在处理复杂场景时,存在无法充分提取和利用特征,且计算复杂度高的问题,提出一种多尺度特征融合的Transformer遥感影像超分辨率重建网络模型.该模型引入了多尺度残差Swin Transformer模块,在充分提取特征的同时,减少用于提取浅层特征的模块冗余;建立了一个特征细化融合模块,可以充分提取图像特征来提高网络性能.基于UC Merced Land Use公开数据集进行实验,结果表明:提出的模型所需参数数量仅为目前主流超分辨率重建方法EDSR模型的61.6%,重建结果在不同尺度下的峰值信噪比和结构相似度相对EDSR分别平均提高了0.82 dB和0.024.通过对比分析,证明本文提出的模型在提高图像质量的同时,有效地减少了网络参数冗余,可明显提高重建图像质量,满足高分辨率遥感影像处理需要.

关键词: 遥感影像, 超分辨率重建, Transformer, 特征提取, 特征细化融合

Abstract:

To address the limitation of the existing super?resolution reconstruction of remote sensing image algorithms in fully extracting and utilizing features and coping with high computational complexity in complex scenes, a Transformer network model for super?resolution reconstruction of remote sensing image based on multi?scale feature fusion was proposed. The multi?scale residual Swin Transformer module was introduced to fully extract features and reduce the module redundancy used for flat feature extraction. A feature fusion refinement module was established that can fully extract image features to improve network performance. Based on the public UC Merced Land Use dataset, the experimental results show that the number of parameters required by the proposed model is only 61.6% of the parameters compared with the current mainstream super?resolution reconstruction method EDSR model. The peak signal?to?noise ratio and structural similarity of the reconstruction results at different scales are increased by 0.82 dB and 0.024 on average compared with the EDSR model. Through comparative analysis, it is proved that the model proposed can effectively reduce the redundancy of network parameters while improving the quality of the image. It can significantly improve the quality of the reconstructed image to meet the requirements of high?resolution remote sensing image processing.

Key words: remote sensing image, super?resolution reconstruction, Transformer, feature extraction, feature refinement fusion(FRF)

中图分类号: