【作者】 王杰；

【导师】 李宏伟；

【作者基本信息】 郑州大学 ， 测绘科学与技术， 2025， 硕士

【摘要】 高分辨率遥感影像在变化检测、资源勘探、目标检测等方面有广泛应用。遥感影像中丰富的纹理细节、结构特征、多样的地物信息,使得其超分辨率重建需要更高强度的特征表达能力。近年来,受益于深度学习技术和神经网络模型,对不同尺度精细特征学习的方法都得到了广泛研究。然而,不同尺度特征之间贡献的权衡,以及在大数据模型快速发展的背景下,边缘设备的应用限制问题仍待进一步优化解决。针对以上问题,本文重点研究了针对多光谱遥感影像的轻量级超分辨率重建技术,提出了两种轻量级网络。目的在于探索更适合边缘设备的轻量化超分辨率重建方法,发掘更具细节丰富、视觉舒适特质的高分辨率遥感影像。本文的研究主要内容如下:(1)针对如何融合transformer的长距离依赖(全局特征)学习能力以及卷积神经网络对局部特征提取优势的问题。本文展开构建轻量化遥感影像超分辨率重建方法研究,提出了 RepCHAT网络。构建了一个可以融合空间域和频率域特征的多尺度特征提取模块,利用结构重参数理论提升模型推理效率。同时,将深度可分离卷积引入基于混合注意力的transformer块中,提升transformer的局部特征学习能力。实现以较小的参数量和计算量获得优于常规的transformer和高性能卷积神经网络的超分重建结果。在UC Merced和AID数据集上4倍超分辨率重建任务中取得最优表现。实验结果表明,RepCHAT在确保优异性能的前提下实现了相对较低的复杂度,有利于边缘设备部署和使用。(2)针对RepCHAT模型的频率域特征提取颗粒度细化问题,及面对长序列输入transformer的计算瓶颈问题。本文展开轻量化超分辨率模型的优化研究,提出了 FM-SNRNet网络。利用傅里叶变换对图像解构,针对频谱特征设计了相位谱和振幅谱的差异化多尺度特征提取模块。嵌入基于窗口交叉扫描方式的选择性状态空间模型强化全局特征学习,基于信噪比图对全局和局部特征做交流融合。设计级联损失函数,正式训练阶段引入感知损失强化全局特征的学习。基于包含21类场景的遥感影像数据集的对比实验结果中,2倍和4倍超分辨分别取得PSNR:0.14dB,0.62dB的进步。并且2倍超分辨率任务中19个类别上取得进步,4倍重建任务中14个类别上取得进步。实验结果表明,FM-SNRNet网络在像素级评价指标及视觉效果两个方面都取得了较好重建表现。另外,通过遥感场景分类与语义分割两个实践应用的验证,证明本研究所提出的遥感超分辨率重建方法具有较强泛化性和实用性。更多还原

【Abstract】 High-resolution remote sensing images have extensive applications in change detection,resource exploration,and target detection.The rich texture details,structural features,and diverse ground object information in remote sensing images require a higher intensity of feature expression ability for their super-resolution reconstruction.In recent years,benefiting from deep learning technology and neural network models,methods for learning fine features at different scales have been widely studied.However,the trade-off of contributions among features at different scales and the application limitations of edge devices in the context of the rapid development of big data models still need to be further optimized and solved.To address these issues,this paper focuses on the research of lightweight super-resolution reconstruction technology for multispectral remote sensing images and proposes two lightweight networks.The aim is to explore more suitable lightweight super-resolution reconstruction methods for edge devices and to discover high-resolution remote sensing images with richer details and more visually comfortable characteristics.The main research contents of this paper are as follows:(1)How to integrate transformer’s long-distance dependent(global feature)learning ability and the advantages of convolutional neural network for local feature extraction.In this paper,a super-resolution reconstruction method for lightweight remote sensing images is developed,and a RepCHAT network is proposed.Specifically,a multi-scale feature extraction module which can integrate spatial domain and frequency domain features is constructed,and the structure heavy parameter theory is used to improve the inference efficiency of the model.At the same time,deep separable convolution is introduced into transformer block based on mixed attention to improve transformer’s local feature learning ability.The results of hyperfractional reconstruction can be obtained with less parameters and less computation than conventional transformer and high performance convolutional neural networks.Achieved optimal performance in 4x super-resolution reconstruction tasks on UC Merced and AID datasets.The experimental results show that RepCHAT achieves relatively low complexity while ensuring excellent performance,which is conducive to the deployment and use of edge devices.(2)Granularity refinement of frequency domain feature extraction for RepCHAT model and computational bottleneck in long sequence input transformer.In this paper,the optimization of lightweight super-resolution model is studied,and the FM-SNRNet network is proposed.Specifically,Fourier transform is used to deconstruct the image,and a differentiated multi-scale feature extraction module of phase spectrum and amplitude spectrum is designed.The selective state space model based on window cross-scanning is embedded to enhance the global feature learning,and the global and local features are fused based on the SNR graph.Cascade loss function is designed,and perceptual loss is introduced in formal training stage to enhance the learning of global features.Based on the comparative experimental results of remote sensing image dataset containing 21 types of scenes,the PSNR of 0.14dB and 0.62dB can be achieved by 2x and 4x super resolution,respectively.Progress was also made in 19 categories on the 2x super-resolution task and in 14 categories on the 4x reconstruction task.The experimental results show that the FM-SNRNet network has achieved good reconstruction performance in pixel level evaluation index and visual effect.In addition,through two practical applications of remote sensing scene classification and semantic segmentation,it is proved that the remote sensing super resolution reconstruction method proposed in this study has strong generalization and practicability.更多还原