Deep image restoration and enhancement

This thesis mainly focuses on image denoising, an important part of image restoration and enhancement which attempts to recover a noise-free image from a noisy version. Recently, deep learning denoising methods have outperformed many traditional model-based denoising methods. These methods handle th...

Full description

Saved in:
Bibliographic Details
Main Author: Fu, Zixuan
Other Authors: Wen Bihan
Format: Thesis-Master by Coursework
Language:English
Published: Nanyang Technological University 2022
Subjects:
Online Access:https://hdl.handle.net/10356/158695
Tags: Add Tag
No Tags, Be the first to tag this record!
Institution: Nanyang Technological University
Language: English
Description
Summary:This thesis mainly focuses on image denoising, an important part of image restoration and enhancement which attempts to recover a noise-free image from a noisy version. Recently, deep learning denoising methods have outperformed many traditional model-based denoising methods. These methods handle the denoising problem by training a deep convolutional neural network in a supervised-learning manner, given a large dataset consisting of paired noisy and clean images. However, this scheme fails in some circumstances since well-aligned noisy and clean images sometime are hard to obtain. To solve this problem, this thesis considers a more general and practical unsupervised-learning setting for image denoising, which is achieving image denoising by utilizing unpaired noisy and clean images. However, training a denoising network with unpaired images directly is almost impossible. Thus, we separate the unsupervised denoising problem into an unsupervised noise generation problem and a supervised denoising problem. To be more specific, a generative model is first applied to learn the noise distribution from the noisy images, and synthesizes paired clean and noisy images. This stage is called the noise generation stage. Then the unpaired denoising problem degrades to a paired denoising problem, and a denoising network can be easily trained in a supervised-learning manner, called the denoising stage. Several experiments on synthetic noise dataset show our proposed method is promising to solve this unpaired denoising problem.