| Author: | Guo, Shi |
| Title: | Toward effective image restoration from raw domain |
| Advisors: | Zhang, Lei (COMP) |
| Degree: | Ph.D. |
| Year: | 2024 |
| Subject: | Image processing -- Digital techniques Image reconstruction Neural networks (Computer science) Hong Kong Polytechnic University -- Dissertations |
| Department: | Department of Computing |
| Pages: | xviii, 129 pages : color illustrations |
| Language: | English |
| Abstract: | With the rapid advancements of digital imaging technologies and the widespread use of social media, digital cameras such as smartphone cameras have become indispensable in our daily lives. Consequently, image and video restoration have become crucial in enhancing the visual quality of captured images and videos. The prevalence of smartphones as the primary means of digital imaging, however, poses significant challenges in image/video restoration because the captured data often exhibit higher level of noise. In this thesis, we embrace the recent advance of deep neural networks (DNNs) to address the challenges in restoration algorithms for single image, burst images and videos, with the goal of achieving high-quality image and video outputs. Firstly, we focus on developing effective single image denoising algorithms. Specifically, in Chapter 2, we introduce the Spatial-Frequency Attention Network (SFANet) to enhance the network’s capacity for exploiting long-range dependencies. To effectively incorporate frequency information into deep learning, we consider the frequency resolution and propose a Window-based Frequency Channel Attention (WFCA) block to adeptly model deep frequency features and their dependencies. The SFANet surpasses other methods in terms of denoising accuracy across multiple denoising benchmark datasets. Secondly, to obtain better and cleaner results for images with high noise level, such as those captured under low-light imaging conditions, we propose to perform joint denoising and demosaicking (JDD) using burst images. To design an appropriate network structure for processing real-world raw images, we leverage the green channel prior that the green channels exhibit higher Signal-to-Noise Ratio and have twice the sampling rate compared to the red/blue channels, and develop the Green Channel Prior Network (GCP-Net) in Chapter 3. In addition, alignment is a crucial step for multi-frame image restoration. However, previous methods often encounter difficulties in effectively compensating for large shifts caused by camera and object motion. In response to this issue, we develop in Chapter 4 a novel differentiable two-stage alignment scheme aimed at better utilizing the temporal information of burst images with large shift. The performance of our proposed methods surpasses existing approaches in the restoration of real-world burst images. Compared with single and burst image restoration, video restoration not only requires high restoration accuracy but also demands careful consideration of the temporal consistency between consecutive frames. Finally, in Chapter 5 we propose two temporal loss functions and a recurrent framework for video JDD. Our method achieves leading restoration performance in term of restoration accuracy, perceptual quality, and temporal consistency. In summary, this thesis contributes to the growing field of image and video restoration by presenting advancements in restoration algorithms for single image, burst images, and videos. Through comprehensive experiments and evaluations, we demonstrate the effectiveness of the proposed methods, providing valuable insights for practical applications in real-world digital imaging scenarios. |
| Rights: | All rights reserved |
| Access: | open access |
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/200/12922