Remote Sensing Image Blind Super-Resolution via a Convolutional Neural Network-Guided Conditional Diffusion Model
Fanen Meng, Sensen Wu, Laifu Zhang, Haoyu Jing, Yijun Chen, Yiming Yan, Tian Feng, Renyi Liu, Zhenhong Du · IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing · 2025
Conventional methods based on bicubic degradation assumption for single remote sensing image super-resolution (SRSISR) have achieved remarkable progress. However, they would possibly fail in plenty of real-world scenarios due to the emergence of complex and diverse degradation factors. Therefore, devising the blind super-resolution (SR) method for remote sensing images towards the degradation of unknown blur kernel is of significant value. In this paper, we propose DiffBSR, an SRSISR framework based on the Diffusion model. Specifically, the proposed conditional Diffusion model contains a forward diffusion process and a reverse generation process. The reverse process employs a convolutional neural network (CNN)-based SR network and a degradation representation network to help the model generate high-resolution images with clear vision and accurate semantics. The former controls the generation direction by generating conditional guidance images corresponding to low resolution images. The latter can represent the degradation information as a feature vector to assist the super-resolution network in flexibly adapting to different degradations. Compared with the state-of-the-art CNN-based methods, our DiffBSR improves the learned perceptual image patch similarity (LPIPS) metric by 32%-76% and achieves higher segmentation accuracy in classification evaluation experiments. In addition, more visual results and higher scores in evaluation metrics are obtained on the Gaofen satellite data, demonstrating the potential of the proposed framework for applications with real-world remote sensing images. Our code will be available at https://github.com/Meng-333/DiffBSR.