Enhancing deep learning-based models for no-reference image quality assessment: Innovations in loss function, model architecture and feature representation

Wang, Huasheng 2024. Enhancing deep learning-based models for no-reference image quality assessment: Innovations in loss function, model architecture and feature representation. PhD Thesis, Cardiff University. Item availability restricted.

	PDF (Huasheng Wang, PhD, Thesis) - Accepted Post-Print Version Restricted to Repository staff only until 5 June 2025 due to copyright restrictions. Available under License Creative Commons Attribution Non-commercial No Derivatives. Download (5MB)
	PDF (Cardiff University Electronic Thesis and Dissertation Form) - Supplemental Material Restricted to Repository staff only Download (793kB)

Abstract

The rapid advancement of deep learning has revolutionized no-reference image quality assessment (NR-IQA), offering significant progress in evaluating image quality without the need for reference images. Despite these advancements, there remains room for improvement. The core aspects of deep learning-based NR-IQA revolve around three primary components: loss function, model architecture and feature representation. To enhance deep learning-based NR-IQA from these three perspectives, this thesis proposes the following four innovative approaches: Initially, we reframe NR-IQA learning as an ordinal regression problem and propose a potent framework utilizing deep convolutional neural networks (DCNN) and Transformers. By incorporating a deep ordinal loss and soft ordinal inference, we convert predicted probabilities into a continuous variable for image quality. Subsequently, we introduce a novel framework for integrating saliency in NRIQA, inspired by the saliency-based visual search mechanism. This approach, embodied in BioSIQNet, a bio-inspired deep neural network within a multi-task learning framework, exploits the synergy between visual attention and image quality perception. Joint learning of these interconnected visual tasks enhances the feature representation of the primary IQA model. Furthermore, we present an original NR-IQA model employing a knowledge sharing (KS) strategy for quality score prediction. By utilizing a full-reference IQA (FRIQA) model as the teacher and an NR-IQA model as the student, our architecture involves shared decoding rather than independent models. Lastly, we devise an Adaptive Graph Attention (AGA) module to augment both local and contextual information simultaneously. This technique refines post-transformer features into an adaptive graph, enhancing local information and exploiting interactions among diverse feature channels. By combining these methodologies, we address the core aspects of deep learning-based NR-IQA, presenting four comprehensive approaches that enhance image quality assessment and contribute to more accurate and reliable image quality evaluation in diverse applications

Item Type:	Thesis (PhD)
Date Type:	Completion
Status:	Unpublished
Schools:	Computer Science & Informatics
Funders:	China Scholarship Council (CSC)
Date of First Compliant Deposit:	5 June 2024
Date of Acceptance:	5 June 2024
Last Modified:	07 Jun 2024 13:12
URI:	https://orca.cardiff.ac.uk/id/eprint/169500

Actions (repository staff only)

Edit Item