Files in this item



application/pdfLIU-DISSERTATION-2018.pdf (36MB)
(no description provided)PDF


Title:Deep learning for image restoration and enhancement
Author(s):Liu, Ding
Director of Research:Huang, Thomas S.
Doctoral Committee Chair(s):Huang, Thomas S.
Doctoral Committee Member(s):Hasegawa-Johnson, Mark A.; Do, Minh N.; Schiwing, Alexander G.
Department / Program:Electrical & Computer Eng
Discipline:Electrical & Computer Engr
Degree Granting Institution:University of Illinois at Urbana-Champaign
Subject(s):deep learning
neural network
image restoration
image enhancement
image super-resolution
video super-resolution
image denoising
Abstract:Image restoration is the process of recovering an original clean image from its degraded version, and image enhancement takes the goal of improving the image quality either objectively or subjectively. Both of them play a key part in computer vision and image processing and have broad applications in industry. The past few years have witnessed the revival of deep learning in computer vision, and substantial progress has been made due to the use of deep neural networks. In this dissertation, we use deep learning to address the problems of image restoration and enhancement, with the focus on the following topics: image and video super-resolution (SR), as well as image denoising. For these problems, deep neural networks are generally used as a regression model to predict the original clean image content from the input. However, designing a network structure that can effectively exploit the intrinsic image properties to achieve remarkable performances is not a trivial task. For image SR, several models based on deep neural networks have been recently proposed and attained superior performance that overshadows all previous handcrafted models. The question then arises whether large-capacity and data-driven models have become the dominant solution to this ill-posed problem. We argue that domain expertise represented by the conventional sparse coding model is still valuable, and it can be combined with the key ingredients of deep learning to achieve further improved results. We experimentally show that a sparse coding model particularly designed for SR can be incarnated as a neural network, which can be trained from end to end. The interpretation of the network based on sparse coding leads to much more efficient and effective training, as well as a reduced model size. In addition, we design a unified framework to learn a mixture of sub-networks for image SR so as to further boost SR accuracy. Video SR aims to generate a high-resolution (HR) frame from multiple low-resolution (LR) frames in a local temporal window. The inter-frame temporal relation is as crucial as the intra-frame spatial relation for tackling this problem. We design deep networks for utilizing the temporal relation from two aspects. First, we propose a temporal adaptive neural network that can adaptively determine the optimal scale of temporal dependency. Filters on various temporal scales are applied to the input LR sequence before their responses are adaptively aggregated. Second, we reduce the complexity of motion between neighboring frames using a spatial alignment network which is much more robust and efficient than competing alignment methods and can be jointly trained with the temporal adaptive network. Image denoising, as another important task of image restoration, is dedicated to recovering the underlying image signal from its noisy measurement. First we customize a convolutional neural network for image denoising. Second we investigate the mutual relation between image denoising and high-level vision tasks in a deep learning fashion when image denoising serves as a preprocessing step for high-level vision tasks. We design a network that cascades two modules for image denoising and various high-level tasks, and use the joint loss for updating only the denoising network via back-propagation. Self-similarity in natural images is widely used for image restoration by many classic approaches. We propose a non-local recurrent network as the first attempt to incorporate non-local operations into a recurrent neural network (RNN) for image restoration. Unlike existing methods that measure self-similarity in an isolated manner, the proposed non-local module can be flexibly integrated into existing deep networks for end-to-end training to capture deep feature correlation between each location and its neighborhood. We fully use an RNN for its parameter efficiency and allow deep feature correlation to be propagated along adjacent recurrent states. This design boosts robustness against inaccurate correlation estimation due to severely degraded images. Finally, we show that it is essential to choose a proper neighborhood size for computing deep feature correlation given degraded images, in order to obtain the best restoration performance.
Issue Date:2018-07-12
Rights Information:Copyright 2018 Ding Liu
Date Available in IDEALS:2018-09-27
Date Deposited:2018-08

This item appears in the following Collection(s)

Item Statistics