# Resampling and super-resolution of hexagonally sampled images using deep learning

**Authors:** Dylan Flaute, Russell C. Hardie, Hamed Elwarfalli  
**arXiv:** [2111.02520](https://arxiv.org/abs/2111.02520)  
**Date:** 2021-11-03  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2111.02520v1/

## Key points

- Super-resolution (SR) is important in fields like security, medical imaging, and object recognition for enhancing image resolution.
- The study "Resampling and super-resolution of hexagonally sampled images using deep learning" by Dylan Flaute, Russell C. Hardie, and Hamed Elwarfalli focuses on a deep learning-based SR system for hexagonally sampled low-resolution images.
- The researchers incorporate a realistic observation model to account for optical and sensor degradation during training and testing phases.
- The SR methodology involves non-uniform interpolation to upsample hexagonal imagery into a rectangular grid, followed by the use of Residual Channel Attention Network (RCAN) for further upscaling and restoration.
- The proposed system outperforms directly applying RCAN to rectangularly sampled low-resolution imagery with equivalent sample density.
- This research demonstrates practical benefits of hexagonal sampling combined with modern processing techniques like RCAN for SR tasks.
- By using a modified version of RCAN tailored for hexagonal SR, the researchers show significant improvement in image quality enhancement through hexagonally sampled imagery.

## Summary

Super-resolution (SR) is a crucial task in various fields such as security, medical imaging, and object recognition. It aims to enhance the resolution of images for improved visual quality and detail. In this study titled "Resampling and super-resolution of hexagonally sampled images using deep learning," authors Dylan Flaute, Russell C. Hardie, and Hamed Elwarfalli propose a novel deep learning-based SR system specifically designed for hexagonally sampled low-resolution images. The system takes these hexagonally sampled inputs and generates high-resolution rectangularly sampled images as outputs. To ensure the realism of their approach, the researchers incorporate a realistic observation model that accounts for optical degradation from diffraction and sensor degradation from detector integration during both training and testing phases. The SR methodology begins by employing non-uniform interpolation to partially upsample the observed hexagonal imagery and convert it into a rectangular grid. Subsequently, they leverage state-of-the-art convolutional neural network (CNN) architecture known as Residual Channel Attention Network (RCAN) to further upscale and restore the imagery, ultimately producing the final high-resolution image estimate. The study demonstrates that their proposed system outperforms directly applying RCAN to rectangularly sampled low-resolution imagery with equivalent sample density. While theoretical advantages of hexagonal sampling are well-documented, this research provides practical evidence of the benefits of hexagonal sampling when combined with modern processing techniques like RCAN for SR tasks. By utilizing a modified version of RCAN tailored for hexagonal SR, the researchers showcase a notable advantage in image quality enhancement through hexagonally sampled imagery. This comprehensive investigation published in Optical Engineering highlights the innovative use of deep learning algorithms in enhancing image resolution through hexagonal sampling techniques. The findings not only contribute to advancing super-resolution technology but also shed light on the potential advantages of alternative sampling methods in conjunction with sophisticated processing frameworks like RCAN.