基于条件先验增强和扩散模型的遥感图像超分辨重建算法

打开文本图片集

关键词：图像超分辨率；扩散模型；特征增强；多尺度特征提取 中图分类号：TP391.4 文献标识码：A doi：10.37188/CJLCD.2025-0045 CSTR：32172.14.CJLCD.2025-0045

Abstract： A remote sensing image super-resolution reconstruction algorithm based on conditional prior enhancement and diffsion model is proposed to address the problems of blurry reconstruction effect of smalltargets in remote sensing images and loss of high-frequency details during the reconstruction process. Firstly，a shallow feature enhancement module that integrates multi branch standard convolution，dilated convolution，and coordinate attention is used to enhance the perception ability of smalltargets. Secondly， by stacking residual dense blocks，more representational features can be extracted while maintaining training stability；Subsequently，a multi-scale depth separable convolution module was designed to extract multi-scale prior information and prevent the lossof high-frequency details； Finally，the combination of the above modules is input as prior information into the difusion model，guiding it to iteratively refine and generate high-resolution images. The experimental results on the publicly available remote sensing image dataset RSCNN7 and NWPU-RESISC45 show that good performance is achieved when the scale factor is ×2，×4 ，and ×8 . Among them，on the RSCNN7，when the scale factor is ×4 ，compared with methods with different network architectures，the proposed model significantly reduces the PI and FID，compared to SOTA algorithm based on diffusion model，it reduces 1.43 and 20.56，respectively. In terms of subjective visual effects， it is closer to the true value compared to the comparison algorithm.

Key Words： image super-resolution; difusion model; feature enhancement；multi scale feature extraction

1引言

随着遥感技术的快速发展，遥感图像数据的获取变得日益便捷，应用范围也随之扩展到了农业[1-2]、城市规划[3]、环境监测[4]、防灾减灾[5]等多个领域。（剩余19391字）