Full metadata record
| DC Field | Value | Language |
|---|---|---|
| dc.contributor | Department of Computing | en_US |
| dc.creator | Chen, Du | - |
| dc.identifier.uri | https://theses.lib.polyu.edu.hk/handle/200/14062 | - |
| dc.language | English | en_US |
| dc.publisher | Hong Kong Polytechnic University | en_US |
| dc.rights | All rights reserved | en_US |
| dc.title | Research on quality enhancement and evaluation of image super-resolution | en_US |
| dcterms.abstract | Image super-resolution (ISR) aims to reconstruct high-resolution (HR) images from their low-resolution (LR) counterparts. While significant milestones have been made in this field, there remain many formidable challenges due to many knotty obstacles, such as the complex real-world degradations, the unsatisfied perceptual quality of existing training images, the unpleasing visual artifacts in the reconstruction results, the limitations of existing evaluation systems, and so on. Among these challenges, enhancing the quality of ISR results and developing robust evaluation methods remain central to low-level vision tasks. In this thesis, we aim to address these two challenges from multiple perspectives through four interconnected studies. | en_US |
| dcterms.abstract | Firstly, we propose a human-guided ground-truth (GT) generation scheme to improve realistic ISR results. Traditional methods rely on HR images with synthetic degradations, often yielding over-smoothed or artifact-prone results. Our approach enhances HR images using multiple enhancement models, enabling one LR image to have multiple HR versions. Human annotators label high-quality regions as positive samples and artifact-prone regions as negative samples. A novel dataset and loss function are designed, encouraging ISR models to produce perceptually realistic results with fewer artifacts. | en_US |
| dcterms.abstract | Secondly, we introduce a self-similarity loss (SSL) to optimize generative ISR models like GANs and Diffusion Models (DM), which often produce visual artifacts. Leveraging the inherent non-local self-similarity of natural images, SSL enforces alignment between the self-similarity graphs (SSGs) of GT and ISR results. By focusing on edge areas and reducing computational costs, SSL serves as a plug-and-play penalty, significantly improving generative-based SR models in terms of detail preservation and artifact reduction. | en_US |
| dcterms.abstract | Thirdly, we tackle arbitrary-scale super-resolution (ASR) by proposing GSASR (Gaussian Splatting based Arbitrary-scale Super-resolution), a framework motivated from 2D Gaussian Splatting (GS). While implicit neural representation (INR) struggles with limited receptive fields and high computational costs, GSASR predicts image-conditioned Gaussians from LR inputs and employs a differentiable 2D GPU/CUDA-based scale-aware rasterization for efficient super-resolved (SR) image rendering. This approach achieves superior visual quality and speed across arbitrary scales than state-of-the-art INR-based ASR models, advancing the quality of ASR results. | en_US |
| dcterms.abstract | Finally, we address limitations in full-reference image quality assessment (FR-IQA) by developing A-FINE, an adaptive fidelity-naturalness evaluator. Traditional FR-IQA assumes perfect reference quality, which is often invalid. To break the perfect reference quality assumption, we propose a large-scale dataset, DiffIQA, with 180,000 images generated by a diffusion-based enhancer. With human annotations, we train A-FINE on it through adaptively combining fidelity and naturalness terms. We further set up an SRIQA-Bench, the images of which are composed of ten state-of-the-art ISR models to prove the effectiveness of A-FINE. Validated on the established DiffIQA datasets and SRIQA-Bench, A-FINE outperforms existing models, providing a more reliable evaluation framework for ISR tasks. | en_US |
| dcterms.abstract | In summary, our works advance ISR through human-guided GT generation, self-similarity loss optimization, generalized and efficient ASR frameworks, and adaptive perceptual quality assessment, paving the way for future research in ISR model development and evaluation. | en_US |
| dcterms.extent | xxviii, 180 pages : color illustrations | en_US |
| dcterms.isPartOf | PolyU Electronic Theses | en_US |
| dcterms.issued | 2025 | en_US |
| dcterms.educationalLevel | Ph.D. | en_US |
| dcterms.educationalLevel | All Doctorate | en_US |
| dcterms.accessRights | open access | en_US |
Copyright Undertaking
As a bona fide Library user, I declare that:
- I will abide by the rules and legal ordinances governing copyright regarding the use of the Database.
- I will use the Database for the purpose of my research or private study only and not for circulation or further reproduction or any other purpose.
- I agree to indemnify and hold the University harmless from and against any loss, damage, cost, liability or expenses arising from copyright infringement or unauthorized usage.
By downloading any item(s) listed above, you acknowledge that you have read and understood the copyright undertaking as stated above, and agree to be bound by all of its terms.
Please use this identifier to cite or link to this item:
https://theses.lib.polyu.edu.hk/handle/200/14062