polito.it
Politecnico di Torino (logo)

A generative adversarial network approach to single image super-resolution of open-source satellite imagery

Filippo Barba

A generative adversarial network approach to single image super-resolution of open-source satellite imagery.

Rel. Roberto Fontana. Politecnico di Torino, Corso di laurea magistrale in Data Science and Engineering, 2021

[img]
Preview
PDF (Tesi_di_laurea) - Tesi
Licenza: Creative Commons Attribution Non-commercial No Derivatives.

Download (41MB) | Preview
Abstract:

Image super-resolution is a widely studied ill-posed problem in computer vision, where the objective is to convert a low-resolution image to a high-resolution one. Conventional methods for achieving super-resolution, such as interpolation-based methods, require a lot of pre/post-processing and optimization. Thanks to the rise in popularity of Deep Learning methods over recent years, several studies have shown how learning methods such as convolutional neural networks and generative adversarial networks can be used to perform super-resolution tasks with competitive results when compared to prior state of the art methods. This thesis proposes a focus on the application of super-resolution methods to open-source low-resolution satellite imagery gathered from the Sentinel-2 ESA’s satellite in the RGB domain. The open data policy plays an important role in the choice of this dataset, alongside other key characteristics of the Sentinel-2 mission, most notably the high revisitation frequency of the global covered area from 56°S to 84°N, which happens every 10 days under the same viewing angles. The design of a selection of renowned neural network based models for super-resolution will be subject to analysis, along with a study of the models’ applications to the dataset of choice and of the respective performances when using different upscaling factors (2x, 4x, 8x). Furthermore, this dissertation proposes a generative adversarial network architecture with multiple discriminators. The goal of this multi-discriminator model is to optimize the training process of the generative network over different scaling factors in a single training procedure. Super-resolution applications to the satellite imagery domain are of notable relevance given the impact that well-performing methods can have on existing algorithms relying on this kind of data, such as image classification, object detection, and environmental monitoring among others.

Relators: Roberto Fontana
Academic year: 2021/22
Publication type: Electronic
Number of Pages: 87
Subjects:
Corso di laurea: Corso di laurea magistrale in Data Science and Engineering
Classe di laurea: New organization > Master science > LM-32 - COMPUTER SYSTEMS ENGINEERING
Aziende collaboratrici: FONDAZIONE LINKS-LEADING INNOVATION & KNOWLEDGE
URI: http://webthesis.biblio.polito.it/id/eprint/20427
Modify record (reserved for operators) Modify record (reserved for operators)