| Author: | Wang, Xin |
| Title: | Towards efficient image revivification via deep generative networks |
| Advisors: | Li, Ping (COMP) |
| Degree: | Ph.D. |
| Year: | 2026 |
| Subject: | Image processing -- Digital techniques Image reconstruction Deep learning (Machine learning) Generative adversarial networks (Computer networks) Hong Kong Polytechnic University -- Dissertations |
| Department: | Department of Computing |
| Pages: | xxvi, 169 pages : color illustrations |
| Language: | English |
| Abstract: | Image revivification is a crucial research area in computer vision, focusing on reversing degradations and enhancing visual information to reconstruct high-quality images from given degraded inputs. A spectrum of challenging research tasks are involved in this area such as single image rain removal, denoising, inpainting, etc. Each task presents unique complexities, making it challenging to develop a unified framework for general revivification tasks. In this thesis, we analyze typical degradation to develop deep generative networks that construct latent space, where multiple features are sampled to eliminate degradations and restore object textures. This approach is applied to various image revivification tasks, including rain, snow, haze, shadow, and noise removal, as well as image retouching and inpainting. We first investigate image rain removal, which requires precise separation between the pixels of the rains and objects. However, the similar visual appearances of rains and objects often lead to the pixel misclassification, resulting in either residual rains or the loss of object details in the result. In this thesis, we propose SEIDNet equipped with the generative Status Estimation and Information Decoupling mechanisms to construct the status and kernel latent spaces, enabling multiple samplings of statuses and kernels for each pixel. The sampled statuses effectively capture the ambiguity between rain and object, while sampled pixel-wise kernels are utilized to remove the rain and restore the object appearance. The comparisons on image rain removal and other revivification tasks highlight the generalization capability of SEIDNet. We then investigate general image revivification. Gradually disseminating object features across multiple stages is a common approach to restore the degraded regions. However, the similarity in appearance between degradations and objects may mislead the feature dissemination process, accumulating the problematic features. We propose MSCNet, which comprises multiple stages of Status Hypothesis and Feature Dissemination, gathering as many correct features as possible for eliminating the accumulation of confusing degraded/object features during the dissemination process. At each stage, the status latent space built from status hypothesis enables multiple samplings of statuses for each pixel, which are used to compute kernels for disseminating reliable object features to the degraded regions. MSCNet retains the correct statuses from all previous stages for each subsequent stage, ensuring the network increasingly acquires useful features to enhance the final result. We evaluate MSCNet across various image revivification tasks, showcasing its state-of-the-art performance. Finally, we investigate image inpainting. Multi-domain image inpainting utilizes contextual information from various domains, where the pertinent details from auxiliary images help restore the corrupted pixels. Advanced methods reconstruct auxiliary images with incomplete visual details, ensuring the restored pixels provide additional information to repair damaged regions. However, these methods often encounter challenges: recovered pixels with intricate visual patterns may lack representative details, and those exhibiting simpler patterns might offer limited context. We propose HRC-Net equipped with three generative sub-networks, tailored to extract representative and pertinent object context from diverse auxiliary images. The Hypothesis Sub-network enables multiple samplings of pixel-wise hypotheses in the auxiliary image. The Representative and Collaboration Sub-networks collaborate to learn the distribution of representative scores and convolutional kernels, fostering a close interaction between the representative hypotheses from the auxiliary images and the corrupted pixels of the target image to meticulously repair the target image. We evaluate HRC-Net across various inpainting datasets, revealing its superiority. |
| 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/200/14541