Investigation of Using RGB-Based Real-World Image Super-Resolution to Achieve Real-World Thermal Image Super-Resolution
Author
Allahham, Moaaz Mohamed Jamal
Term
4. term
Publication year
2021
Submitted on
2021-06-03
Pages
64
Abstract
Dette speciale undersøger, om real-world superopløsningsteknikker udviklet til RGB-fotos kan anvendes til at forbedre termiske billeder i praksis. Baggrunden er, at den høje produktionsomkostning for termiske kameraer ofte begrænser den rumlige opløsning. Arbejdet tilpasser et eksisterende RGB-baseret real-world superopløsningsframework, som estimerer realistiske billednedbrydninger for at syntetisere troværdige lav- og højopløsningspar. En superopløsningsmodel trænes til at lære kortlægningen mellem disse domæner og anvender den lærte kortlægning på nye lavopløsnings termiske billeder. Specialet gennemgår relateret arbejde, dataset- og billedkvalitetsvurderingsvalg samt trænings-, test- og ablationsstudier. Eksperimenterne indikerer tydelige forbedringer i den perceptuelle kvalitet af termiske billeder og en ydeevne, der overgår en aktuel state-of-the-art metode til termisk superopløsning. Afslutningsvis diskuteres muligheder for forbedringer og fremtidigt arbejde.
This thesis investigates whether real-world image super-resolution techniques developed for RGB photographs can be used to enhance real-world thermal imagery. The motivation is that the high manufacturing cost of thermal cameras often limits their spatial resolution. The work adapts an existing RGB-based real-world super-resolution framework that estimates realistic imaging degradations to synthesize credible low- and high-resolution pairs. A super-resolution model is trained to learn the mapping between these domains and then applies the learned mapping to new low-resolution thermal images. The thesis reviews related work, choices of datasets and image quality assessment, and details training, testing, and ablation studies. Experiments indicate clear improvements in the perceptual quality of thermal images and performance surpassing a current state-of-the-art method for thermal super-resolution. The thesis concludes with a discussion of possible improvements and future directions.
[This summary has been generated with the help of AI directly from the project (PDF)]
Keywords
Documents