基于两阶段残差条件扩散网络的遥感图像超分辨重建

打开文本图片集

关键词：扩散模型；遥感超分辨率重建；残差网络；先验条件增强 中图分类号：TP391.4；TP751 文献标识码：A doi：10.37188/CJLCD.2025-0158 CSTR：32172.14.CJLCD.2025-0158

Abstract：When the traditional difusion model is used for super-resolution reconstruction of remote sensing images，there are difficulties such as insuffcient utilization of a priori conditions，lengthy sampling steps， and poor recovery of high frequency details.In this paper，we propose a two-stage residual conditional diffusion super-resolution network（TRCDSR）. In the first stage generates preliminary super-resolution results with a pre-trained lightweight CNN model to provide a high-quality structural priori for the difusion model. In the second stage，we introduce the residual conditional difusion mechanism，which takes the residual signals as the input to let the noise prediction network to focus on the high-frequency detail reconstruction. By improving the DDIM inverse sampling formula， the residual correction process is decoupled into a deterministic prediction term and a random noise term，and the high-quality reconstruction is completed in 20～50 steps. The multi-scale prior condition enhancement module（PCEM） and the fusion of spatial and channel atention mechanism（FAN）are further introduced to enhance the model’s adaptability to complex remote sensing scenes.Experiments on several remote sensing datasets，such as AID，SECOND，RSSCN， etc.，show that TRCDSR outperforms other diffsion models，GAN and Transformer-like methods in terms of reconstruction quality，computational efficiency and generalization ability.

Key words： difusion model； remote sensing super-resolution reconstruction；residual network；enhancement of a priori conditions

1引言

遥感（RS）图像因其覆盖范围广、数据采集及时且多样化，在农业病情监测、军事侦察、地质灾害监测等领域具有广泛的应用[1]，在RS数据分析中，分辨率直接决定了地物特征的清晰度和可辨识性，因此高分辨率卫星影像的需求越来越高。（剩余18173字）