%0 Journal Article %T 融合频域增强的遥感图像超分辨率重建方法
Remote Sensing Image Super-Resolution Reconstruction Method Enhanced by Frequency Domain Fusion %A 卢昌婷 %A 梁玉琦 %J Software Engineering and Applications %P 445-455 %@ 2325-2278 %D 2026 %I Hans Publishing %R 10.12677/sea.2026.153042 %X 现有遥感图像超分辨率重建方法在主体结构恢复方面取得了一定效果，但仍存在高频信息利用不足、边缘易模糊以及复杂纹理区域细节恢复不充分等问题。提出一种融合频域增强的遥感图像超分辨率重建方法(Frequency-Enhanced Prior Reconstruction, FEPR)。该方法以RRDBNet为主干网络，在深层特征融合阶段引入频域增强模块，通过统计特征频谱中的高频响应并进行通道自适应重标定，增强网络对边缘、纹理及局部高频结构的表征能力；同时结合像素重建损失与频域一致性损失进行联合监督，以提升重建结果在空间域和频域上的一致性。在NWPU-RESISC45、AID、UC Merced Land Use和WHU-RS19等公开遥感数据集上的实验结果表明，所提方法在不同数据集和放大倍率条件下均取得了较为稳定的重建效果。其中，在NWPU-RESISC45数据集×4任务上，FEPR取得25.52 dB的PSNR和0.6345的SSIM，SSIM优于EDSR和RCAN。消融实验表明，频域增强模块与频域一致性损失对提升重建性能具有一定作用。
Existing remote sensing image super-resolution reconstruction methods have achieved certain effects in the restoration of main structures, but there are still problems such as insufficient utilisation of high-frequency information, edge blurring, and inadequate detail recovery in complex texture areas. A remote sensing image super-resolution reconstruction method with frequency-domain enhancement (Frequency-Enhanced Prior Reconstruction, FEPR) is proposed. This method uses RRDBNet as the backbone network and introduces a frequency-domain enhancement module during the deep feature fusion stage to enhance the network’s ability to represent edges, textures, and local high-frequency structures by statistically analysing high-frequency responses in the feature spectrum and performing channel adaptive recalibration. Meanwhile, a combination of pixel reconstruction loss and frequency-domain consistency loss is used for joint supervision to improve the consistency of the reconstruction results in both spatial and frequency domains. Experimental results on publicly available remote sensing datasets such as NWPU-RESISC45, AID, UC Merced Land Use, and WHU-RS19 show that the proposed method achieves relatively stable reconstruction performance under different datasets and scale factors. Notably, in the ×4 task on the NWPU-RESISC45 dataset, FEPR achieved a PSNR of 25.52 dB and an SSIM of 0.6345, with SSIM outperforming EDSR and RCAN. Ablation experiments indicate that the frequency-domain enhancement module and frequency-domain consistency loss play a certain role in improving reconstruction performance. %K 遥感图像超分辨率重建， %K 频域增强， %K RRDBNet， %K 高频信息， %K 频域一致性损失
Remote Sensing Image Super-Resolution Reconstruction %K Frequency-Domain Enhancement %K RRDBNet %K High-Frequency Information %K Frequency-Domain Consistency Loss %U http://www.hanspub.org/journal/PaperInformation.aspx?PaperID=145447