A graph attention network for real-world multi-image super-resolution
Tomasz Tarasiewicz, Michał Kawulok · Information Fusion · 2025
Multi-image super-resolution is aimed at reconstructing high-frequency details relying on information fusion from multiple observations of the same scene, acquired at different times. Due to subpixel shifts between individual low-resolution images, each of them carries a different portion of high-resolution information. Even though a stack of input images can be effectively super-resolved using convolutional neural networks, they impose tensor-based representation which requires the input images to be aligned at whole-pixel precision, thus ignoring the valuable subpixel shifts. In this paper, we leverage graph-based representation of an input image stack which preserves spatial relation between the input images at subpixel precision. We propose the first end-to-end graph neural network with multi-level attention that processes such graphs to reconstruct a high-resolution image. The reported experimental results indicate that the elaborated network outperforms state-of-the-art techniques for simulated and real-world datasets, demonstrating outstanding robustness against temporal variability that is a common problem in real-world super-resolution.