Full metadata record
| DC Field | Value | Language |
|---|---|---|
| dc.contributor | Department of Computing | en_US |
| dc.contributor.advisor | Wu, Xiao-ming (COMP) | en_US |
| dc.creator | Zhang, Wenlong | - |
| dc.identifier.uri | https://theses.lib.polyu.edu.hk/handle/200/13089 | - |
| dc.language | English | en_US |
| dc.publisher | Hong Kong Polytechnic University | en_US |
| dc.rights | All rights reserved | en_US |
| dc.title | Advancing real-world image super-resolution : comprehensive training and evaluation methods | en_US |
| dcterms.abstract | Real-world image super-resolution (real-SR) methods aim to reconstruct a high-quality (HQ) image from a low-quality (LQ) observation, which is affected by a diversity of degradation factors, such as blur, noise, and JPEG compression. However, there are still several challenges that need to be addressed in order for real-SR methods to effectively handle a broader range of degradation tasks. | en_US |
| dcterms.abstract | We first investigate the impact of degradation modeling on real-SR networks, exploring the possibility of modulating the degradation distribution to improve the performance of corner degradation tasks without sacrificing the overall performance. To develop an effective degradation modeling strategy, we propose a unified gated degradation model that can generate a wider range of degradation cases by leveraging a random gate controller to modulate the degradation distribution. Extensive experiments and comparisons show that the proposed degradation model can tackle not only complex degradation cases but also many corner cases that are ignored by existing real-SR methods. Through the adjusting of degradation distribution, we achieve improved performance on the corner cases for the real-SR network, closely approaching the performance upper bounds. | en_US |
| dcterms.abstract | To address the degradation task competition problem in real-SR, we study multi-task learning strategy for real-SR problem. Traditional approaches to real-SR often train an SR network on all degradation tasks, which are sampled equally in the training process. This poses a challenge known as task competition or task conflict in multi-task learning, where certain tasks dominate the learning process, resulting in poor performance on other tasks. To overcome this problem, we propose a task grouping approach, which can efficiently identify the degradation tasks where a real-SR model falls short and groups these unsatisfactory tasks into multiple task groups for further training. By grouping the unsatisfactory tasks together, our approach mitigates task competition and further improves the performance of the unsatisfactory tasks. | en_US |
| dcterms.abstract | Finally, to address the challenge of model evaluation in real-SR, this study proposes a systematic evaluation framework (SEAL) to provide a holistic view of overall performance for real-SR. Although the prevailing evaluation approaches to real-SR can offer quantitative performance, our study reveals that the mean performance on a limited set of degradations randomly chosen from a vast space often leads to inconsistent and potentially misleading results. To overcome this issue, we cluster the large degradation space to create a set of representative degradation cases, which serves as a comprehensive test set. Subsequently, we introduce a coarse-to-fine evaluation protocol to measure the distributed performance of real-SR methods. The protocol incorporates two new metrics: acceptance rate (AR) and relative performance ratio (RPR), derived from acceptance and excellence lines. Under SEAL, we benchmark existing real-SR methods, obtain new observations and insights into their performance, and develop a new strong baseline. | en_US |
| dcterms.abstract | In summary, this thesis presents comprehensive strategies to tackle the challenges encountered in real-world scenarios for real-SR. Our proposed methods offer novel insights and substantial improvements over current real-SR methods. The results have been accepted or published in various top AI conferences. | en_US |
| dcterms.extent | xvi, 125 pages : color illustrations | en_US |
| dcterms.isPartOf | PolyU Electronic Theses | en_US |
| dcterms.issued | 2024 | en_US |
| dcterms.educationalLevel | Ph.D. | en_US |
| dcterms.educationalLevel | All Doctorate | en_US |
| dcterms.LCSH | High resolution imaging | en_US |
| dcterms.LCSH | Image processing -- Digital techniques | en_US |
| dcterms.LCSH | Machine learning | en_US |
| dcterms.LCSH | Hong Kong Polytechnic University -- Dissertations | 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/200/13089