Beyond one-meter resolution: A comparative study on super-resolution methods

Author:  Clara Garcia Moll

Virtual Room: https://eu.bbcollab.com/guest/6592e39c79684714a3c5cbadf2e9c83f

Date & time: 22/09/2021 – 8:45 h

Session name: Aerial Images and low level vision

Company name: Satellogic Solutions SL

Supervisors: Felipe Lumbreras Ruiz, Javier Marin Tur

Abstract:

Super Resolution (SR) offers a great opportunity to improve certain remote sensing applications since these techniques deal with increasing the resolution providing much more details in the images. Moreover, further techniques were developed due to the recent important breakthrough in deep convolutional neural networks (CNNs).
In this work, we conduct a comparative study of the most recent SR methods on High Resolution satellite images. Specifically, we compare them at different scale factors, one-meter resolution being our starting point. Among the different methods we assess a contestant clearly stands above the rest qualitatively and quantitatively (PSNR, SSIM, SWD, FID). Additionally, aiming at improving the PSF and by doing so easing the resolving task, we incorporate a deblurring layer as a pre-stage. The resolved images yielded by the model when using such a layer outperform the original ones. A full detail of the experimental setup and data preparation are provided. It is our hope that this study provides a better insight on the latest advances on Super Resolution when using satellite imagery.

Committee:

– President: Xavier Otazu(UAB)
– Secretary: Antonio Agudo(UPF)
– Vocal: Javier Vazquez Corral(UAB)