Learning to Remove Noise from Astronomical Images Without Ground Truth Data
Year: 2024
Type: dissertation
Abstract: Contamination removal and denoising are inseparable parts of all sensors and imaging devices. Telescopes are no exception, as the low photon count of exposures necessitates comprehensive post-processing steps. This thesis reviews state-of-the-art image denoisers trained without access to ground truth for natural-scene images and examines the latest developments in denoisers specifically designed for astronomical images. Existing methods rely on the availability of cleaned datasets or aggregating and sampling of numerous frames, resulting in the loss of underlying signals and extensive observation time. By proposing a generalizable training pipeline, we enable deep neural networks to learn to remove noise from astronomical images without access to clean data. We show improvement in the signal-to-noise ratio of single exposures while keeping the distribution intact, reducing the number of exposures required to detect faint objects. We further investigate the effectiveness of denoising in detecting synthesized objects and real-world data.
Source: TSpace (University of Toronto)
Author: Vaheb, Omid
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7132924633