Graph-Based Representation for Multi-image Super-Resolution

Tomasz Tarasiewicz, Michal Kawulok

Abstract: Multi-image super-resolution is a challenging computer vision problem that aims at recovering a high-resolution image from its multiple low-resolution counterparts. In recent years, deep learning-based approaches have shown promising results, however, they often lack the flexibility of modeling complex relations between pixels, permutability of the input data, or they were designed to process a specific number of input images. In this paper, we propose an improved version of our earlier graph neural network that benefits from permutation-invariant graph-based representation of multiple low-resolution images. Importantly, we demonstrate that our solution allows for performing reconstruction from a set of heterogeneous input images, which is not straightforward for other state-of-the-art techniques. Such flexibility is a crucial feature for practical applications, which is confirmed qualitatively and quantitatively for a set of real-world (rather than simulated) input images.