Learning based image enhancement and objective quality measure

Lyu, Hongjin 2024. Learning based image enhancement and objective quality measure. PhD Thesis, Cardiff University. Item availability restricted.

	PDF (Hongjin Lyu, PhD, Thesis) - Accepted Post-Print Version Restricted to Repository staff only until 15 January 2026 due to copyright restrictions. Available under License Creative Commons Attribution Non-commercial No Derivatives. Download (309MB)
	PDF (Cardiff University Publication Form) Restricted to Repository staff only Download (939kB)

Abstract

Image enhancement, such as image style transfer and image colourisation, aims to modify input images to achieve desired manipulations. Significant progress has been made recently, thanks to deep learning based generative models. In this thesis, we propose novel methods to improve generation quality and reduce training data demands. A fundamental problem for such generative tasks is the impossibility of obtaining ground truth results, and in many cases, there are no unique correct answers. We further develop a new metric to evaluate image colourisation quality. For image style transfer, we introduce WCGAN, a novel Generative Adversarial Network (GAN)-based model designed specifically for transferring high-quality watercolour style into portraits. Leveraging a localised Gram matrix loss and an adaptive discriminator architecture, WCGAN demonstrates superior performance in capturing intricate watercolor style characteristics in local regions and stably handling input of different scales. We further extend WCGAN to video style transfer by a novel video training data generation method. For image colourisation, we particularly address scribble-based colourisation where sparse user scribbles are used to resolve essential ambiguities. We propose the Local and Global Affinity network (LGA-Net), which regards scribble-based colourisation as an affinity propagation process. By explicitly integrating both local and global affinity relationships, LGA-Net achieves robustness against sparse scribbles while enhancing image structure understanding and color propagation accuracy. Furthermore, we present a novel objective metric, the Statistical Color Distribution (SCD), for evaluating re-colourised image quality based on statistical color information. By leveraging insights from the statistical perspective, SCD provides a systematic and reliable scoring mechanism for assessing the quality of colourised images. Through extensive experimentation and comparative analysis, we demonstrate the effectiveness and superiority of our proposed methods over existing state-of-the-art approaches. The insights from this research advance human aesthetic experience in real-world applications and guide future developments in related fields.

Item Type:	Thesis (PhD)
Date Type:	Completion
Status:	Unpublished
Schools:	Computer Science & Informatics
Subjects:	Q Science > QA Mathematics > QA75 Electronic computers. Computer science
Funders:	China Scholarship Council
Date of First Compliant Deposit:	15 January 2025
Date of Acceptance:	13 January 2025
Last Modified:	15 Jan 2025 14:33
URI:	https://orca.cardiff.ac.uk/id/eprint/175292

Actions (repository staff only)

Edit Item