MBF-SR: A Single-Image Super-Resolution Method for Remote Sensing Images in Dynamic Imaging Scenarios

Single-image super-resolution is a common method for improving the resolution of remote sensing images. However, existing methods often suffer from artifacts and texture distortion when dealing with complex blur with significant directionality caused by the movement of dynamic imaging platforms. To address this, we constructed the RASSID dataset and proposed an adaptive remote sensing image super-resolution method, MBF-SR, based on the InVSR conditional diffusion framework. The core innovations of this article include the following three aspects: 1) dynamically generating diverse random motion blur kernels and applying random compression processes in training batches to simulate degradation samples closer to real physical scenarios; 2) analyzing the blur direction information of low-resolution images in the latent space and introducing this direction information as a conditional variable into the noise prediction network; 3) designing a Latent-space Motion-direction Aware BLur loss (LAMBL). LAMBL reuses the blur direction information extracted from low-resolution images to constrain the reconstructed images to align with real high-resolution images in terms of gradient similarity, direction consistency, and frequency-domain features. Experiments show that compared with current mainstream methods, the proposed method has achieved significant improvements in the detail fidelity and texture authenticity of reconstructed images, effectively alleviating the problems of artifacts and directional blur, and enhancing the stability of remote sensing image super-resolution reconstruction in dynamic imaging scenarios.