基于残差扩散模型的遥感超分辨率图像生成研究
- 分享到:
摘要：
传统基于扩散的图像超分辨率方法通常需要大量采样步骤，并且优化功能强大的扩散模型需要耗费大量运算时间。为了在有限的计算资源上实现训练，现有的加速采样技术往往会牺牲部分图像质量，导致超分辨率结果模糊。为了解决这一问题，提出了一种改进的、高效的残差超分辨扩散模型。通过构建马尔可夫链，在高分辨率图像和低分辨率图像之间移动残差来实现图像之间的转移，有效减少扩散步骤的数量。该方法保证了超分辨率结果的质量和灵活性，同时也提高了转移效率，消除了推理过程中需要的后加速及其相关图像细节特征的退化。实验证明，即使只执行15个采样步骤，所提的方法在合成数据集和真实数据集上也可以获得优于或至少可以与当前最先进方法相当的图像质量。
Traditional diffusion-based image super-resolution methods usually require a large number of sampling stepsand a lot of computational time to optimize powerful diffusion models. In order to realize training on limited computational resources, existing accelerated sampling techniques often sacrifice image quality and lead to ambiguous super-resolution results. To address this problem, a improved and efficient residual super-resolution diffusion model is proposed in this paper. The transfer between images is achieved by constructing a Markov chain that shifts the residuals between high-resolution images and low-resolution images, significantly reducing the number of diffusion steps. The method ensures the quality and flexibility of the super-resolution results, while improving the transfer eficiency and eliminating the post-acceleration and its associated image detail feature degradation required in the inference process. lt is experimentally demonstrated that even with only 15 sampling steps, the method in this paper achieves image quality better than or at least comparable to current state-of-the-art methods on both synthetic and real datasets.
作者：
左宪禹，田展硕，殷梦晗，党兰学，乔保军，刘扬，谢毅
Zuo Xianyu，Tian Zhanshuo，Yin Menghan，Dang Lanxue，Qiao Baojun，Liu Yang，XieYia
机构地区：
河南大学计算机与信息工程学院；河南省大数据分析与处理重点实验室
引用本文：
左宪禹，田展硕，殷梦晗等。基于残差扩散模型的遥感超分辨率图像生成研究[J]. 学报（自然科学版）,2025, 53(3) : 58-65. (Zuo Xianyu, Tian Zhanshuo, Yin Menghan, et al. Remote sensing super-resolution image generation based on residual diffusion model[J] . Journal of Henan Normal University(Natural Science Edition) ,2025,53(3) :58-65.DOI:10.16366/j. cnki.1000-2367.2024.03.12.0002.)
基金：
国家自然科学基金；河南省高校科技创新团队支持计划；河南省科技攻关项目
关键词：
遥感；超分辨率；图像生成；残差移动；残差扩散模型
remote sensing; super-resolution; image generation; residual shifting; residual diffusion model
分类号：
TP751
