REAL-WORLD SUPER-RESOLUTION: HOW TO OVERCOME RECONSTRUCTION ARTIFACTS?

Multi-image super-resolution, which aims at reconstructing high-frequency spatial information from multiple images of the same area of interest, can be effectively achieved using deep neural networks. These networks require large training sets to learn both the process of information fusion and the relation between low-resolution and high-resolution appearances. While the training sets can be easily generated by simulating low-resolution images through downsampling original images, models trained from such simulated data often produce significant artifacts when applied to enhance real-world images at their original resolution. Conversely, collecting real-world datasets is both costly and challenging. In this paper, we propose a new approach to address the domain gap between super-resolution models trained on simulated and real-world data. Through an extensive experimental study, we demonstrate that the reconstruction artifacts can be mitigated with a properly trained noise removal algorithm. The proposed approach represents an important step toward the practical application of simulation-based training for real-world image reconstruction tasks.