Full metadata record
| DC Field | Value | Language |
|---|---|---|
| dc.contributor | Department of Computing | en_US |
| dc.contributor.advisor | Zhang, Lei (COMP) | en_US |
| dc.creator | Liang, Jie | - |
| dc.identifier.uri | https://theses.lib.polyu.edu.hk/handle/200/12573 | - |
| dc.language | English | en_US |
| dc.publisher | Hong Kong Polytechnic University | en_US |
| dc.rights | All rights reserved | en_US |
| dc.title | Toward effective real-world image restoration and enhancement | en_US |
| dcterms.abstract | Deep neural network-based image restoration and enhancement methods have become prevalent in producing high-quality and visually pleasing images. While existing works have shown remarkable improvements, most of them are developed by using synthetic data and overlook the practical requirements in real-world applications. In this thesis, we dive into the design of effective learning methods for real-world image restoration and enhancement tasks, as well as the construction of effective benchmarks to facilitate the research along this line. | en_US |
| dcterms.abstract | It is very challenging to balance the reconstruction accuracy and the perceptual quality in image super-resolution (SR) because these two objectives are contradictory in model optimization. To this end, in Chapter 2, we propose a locally discriminative learning (LDL) approach, where a generative adversarial network-based SR model is trained to stably generate perceptually realistic details while inhibiting visual artifacts. Then, considering the diversity of real-world images in terms of degradation, in Chapter 3, we design an efficient and degradation adaptive (DASR) method for real-world image super-resolution, whose parameters are adaptively specified by estimating the degradation of the input image. DASR is validated to be effective in handling real images with different degradation levels. Furthermore, in Chapter 4 we investigate a more challenging real-world task, i.e., joint demosaicking, denoising, and super-resolution (JDDSR), which aims to reconstruct full-color high-resolution high-quality images from sensor raw data. By analyzing the relationship of the three tasks in JDDSR, we propose a deep parallel network (DPN) that optimizes the tasks with conflict goals in parallel to improve the restoration performance. A large-scale and high-quality training dataset and a real-world benchmark test dataset are also established for use in the community. Finally, in Chapter 5, we study the portrait photo retouching (PPR) task, which is important to acquire a visually pleasing portrait photo with favorable tones. Inspired by the experience in real-world photography, we propose to optimize the human-region with high priority and keep the consistency of a group of photos. A large-scale PPR dataset is also constructed. | en_US |
| dcterms.abstract | In summary, in this thesis we present four works toward effective real-world image restoration and enhancement. Among them, LDL provides an effective learning strategy to stabilize the optimization of perceptual quality-oriented image SR tasks. DASR contributes an efficient yet effective SR method to enhance real-world images with diverse degradations in a unified model. DPN tackles the JDDSR task and presents an effective solution to handle multiple sub-tasks that have conflict goals. Finally, the PPR method handles the image retouching task and gives insights in how to design learning strategies to favor the requirement of human perceptions. Two large-scale datasets for the JDDSR and PPR tasks are also provided. Extensive experiments demonstrate the effectiveness of both the proposed methods and datasets in real-world image restoration and enhancement tasks. | en_US |
| dcterms.extent | xxiii, 148 pages : color illustrations | en_US |
| dcterms.isPartOf | PolyU Electronic Theses | en_US |
| dcterms.issued | 2023 | en_US |
| dcterms.educationalLevel | Ph.D. | en_US |
| dcterms.educationalLevel | All Doctorate | en_US |
| dcterms.LCSH | Image processing -- Digital techniques | en_US |
| dcterms.LCSH | Image reconstruction | 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/200/12573