الفهرس 
CHAPTER 1. IntroductionOnly 14 pages are availabe for public view
 |  | from | 118 |  |  |
| | | from | 118 | | |
Abstract
Images, today, constitute a basic corner stone in various domains, including photography, medical imaging, underwater imaging and surveillance. Digital images captured in the real world are susceptible to numerous kinds of distortion. Image distortion could be occurred due to certain reasons, such as camera shake, scene objects motion, missed focus, insufficient depth of field, or atmospheric turbulence. In other words, motion is the major reason for image distortion, which consequently leads to blur unclear images.
Then, image deblurring procedures are utilized to retrieve the clear image from the blurred one; therefore, the focus of this research is on dynamic scene restoration for a single image. Different image deblurring techniques are proposed to solve that concern, which would be classified mainly into: Blind, and Non-Blind image deblurring approaches. Blind deblurring approach is the process of deblurring the distorted image without any prior information about the blur kernel nevertheless, Non-blind deblurring assumes the prior knowledge such as the blur mask is available. The previous studies show that the blind deblurring process is highly ill-conditioned task in computer vision; as, when the blur is spatially variant, the task becomes more challenging. Then, various optimization-based and deep learning-based approaches are proposed to handle the blind images. Optimization-based techniques are based on minimizing the cost function by using iterative methods to obtain the optimal solution of deblurred images whereas, deep learning-based approaches implement the artificial neural systems to automatically learn the correspondence between the blurry and ground-truth images. Since the optimization-based methods are time consuming, our study utilizes the end-to-end deep learning techniques as non-uniform blind image restoration for dynamic scenes.
The thesis has proposed two multi-scale training-based approaches: The first approach exploits the traditional priors of optimization approaches (the dark and bright channel priors); it integrates them into the loss function of deep learning framework during the training process to constrain the multi-component loss function, which consequently leads to improve the recovered image quality.
The second proposed solution is a low complexity model that prevents the memory loss with partial sharing parameters and embedded attention mechanisms; utilizing the attention mechanism into the proposed framework assists it to generalize well, and accordingly, improve the image recovery process and maintain the model robustness.
The two proposed methods are evaluated using two standard datasets (GoPro and Kohler) and compared to the state-of-the-art image deblurring methods; the analytical results show that our proposed models outperform the previous works both quantitatively and qualitatively. The first proposed model outperforms other methods by a gain of about 0.28 - 4.59 dB PSNR and 0.0056 - 0.079 SSIM utilizing GoPro dataset whereas the second proposed model outperforms other methods by a gain of about 0.22 - 6.68 dB PSNR and 0.0027 - 0.1111 SSIM.