A graph neural network for heterogeneous multi-image super-resolution

Reconstructing a high-resolution image from multiple low-resolution observations, commonly known as multi-image super-resolution, is a challenging computer vision task. Although the approaches underpinned with convolutional neural networks have demonstrated outstanding results, they commonly exhibit limitations resulting from the tensor-based representation of the input data. These are concerned with the permutability and the pixel-wise alignment of the input images which requires them to be co-registered at whole-pixel precision—different scales or rotations must be compensated beforehand. In this paper, we address these shortcomings with a graph neural network that leverages permutation-invariant graph-based representation of multiple low-resolution images. Notably, we demonstrate the ability to reconstruct images from a set of heterogeneous images at different scales and various orientations, which cannot be easily achieved by the state-of-the-art techniques. This flexibility is crucial for practical applications, as confirmed both qualitatively and quantitatively through experiments with simulated and real-world input images.