Topological Low-Light Image Denoising

Abstract

This thesis proposes novel approaches to image denoising that specifically address the challenges of noise from real low-light conditions. While numerous denoising algorithms have been developed, very few have been designed specifically for low-light imagery. The proposed methods utilize tools of topological data analysis such as invariants based on persistent homology groups.

The developed denoising loss functions were evaluated on state-of-the-art architectures and compared to a number of existing loss functions. The results demonstrated superior performance, as well as suppressing common artifacts and better adaptation to textured areas. Patch based loss function allowed the improvement of up to 7% in PSNR terms and 8 % in SSIM comparing to conventional loss functions. The use of Wavelet Transform allows to account for textured areas in the image, improving the results even further, with a significant performance gains of up to 25% in the Learned Perceptual Image Patch Similarity (LPIPS) metric.

The second scientific contribution of this thesis is a newly captured comprehensive dataset of natural images (BVI-Lowlight). The thesis includes chapters on the acquisition and preprocessing of the dataset, statistical and topological analysis of low-light images, and a summary of the findings. This work contributes to the field of image processing and provides a new approach to addressing low-light image denoising as well as new applications for topological data analysis.

Date of Award	16 Jan 2024
Original language	English
Awarding Institution	University of Bristol
Supervisor	David R Bull (Supervisor) & Pui Anantrasirichai (Supervisor)