This repository contains the results for "A Comparative Study on CNN-Based Single-Image Super-Resolution Techniques for Satellite Images". You can find the trained models in the Releases section of the repository. All experiments have been performed using the original implementations, which have been linked in the table below. Check out this english article or the گزارش فارسی for more details on the project.

Based on their novelty and reported performances, we have chosen the following techniques for this study, sorted by their earliest draft publication date:

• Zhang et al., Residual Dense Network (RDN) (repo)
• Zhang et al., Residual Channel Attention Network (RCAN) (repo)
• Li et al., Feedback Network for Image Super-Resolution (SRFBN) (repo)
• Anwar & Barnes, Densly Residual Laplacian Network (DRLN) (repo)
• Li et al., Gated Multiple Feedback Network (GMFN) (repo)
• Mei et al., Cross-Scale Non-Local Network (CSNLN) (repo)

Training and evaluation of the techniques has been done on a Tesla P100 GPU, using the PyTorch library, while the bicubic interpolation algorithm has been run on a Core i7-9500H CPU, with the tools provided by the Scikit-Image library. The results for the models marked with an * have been directly lifted from our baseline article.

The following shows a single image, being down-scaled and then reconstructed, first using the Bicubic interpolation, and then using the trained SISR models.