DEEP LEARNING FOR SINGLE IMAGE DEBLURRING

Abstract

Blind single image deblurring is a highly ill-posed problem. It often becomes even more challenging while blur is non-uniform. Since the rise of deep learning, many recent approaches are based on Convolutional Neural Networks (CNNs). These CNN-based approaches are diverse, in terms of their structures and components. However, existing methods have many disadvantages, for instance, they require intensive computation resources, and cannot restore sharp details when image blur is severe or non-uniform. In this thesis, a review the state-of-the-art methods in image restoration is firstly given, including image denoising, image dehazing, image super-resolution and image deblurring, especially of those learning based. Then various key elements and mechanisms in deblurring networks are analysed, including backbones, frameworks and conjecture that a good balance among receptive field, depth and efficiency. To achieve better performance, three networks are proposed in this research. By combining the strength of DenseNet and Inception-v4 to realize a balanced structure, a network is proposed and named as MixNet. A new network that uses dilated convolution and named DC-Deblur is also introduced as well as a Graph Convolutional Network (GCN) based method, termed as GCResNet. Further experiments in other image restoration tasks are given, in order to show the generalisation of the proposed methods. Quantitative evaluations in term of comprehensive image quality measures have been performed. Results show that the proposed MixNet, DS-Deblur, and GCResNet are able to elevate the state-of-the-art performance on end-to-end results for dynamic scene deblurring on all the benchmark datasets.

Date of Award	1 Aug 2022
Original language	English
Awarding Institution	The University of Manchester
Supervisor	Hujun Yin (Supervisor) & Alexandru Stancu (Supervisor)