Files in this item

FilesDescriptionFormat

application/pdf

application/pdfNIE-THESIS-2019.pdf (6MB)Restricted Access
(no description provided)PDF

Description

Title:Compression artifact suppression for color images with dual-domain SE-ARResNet
Author(s):Nie, Jiaxi
Advisor(s):Do, Minh N
Department / Program:Electrical & Computer Eng
Discipline:Electrical & Computer Engr
Degree Granting Institution:University of Illinois at Urbana-Champaign
Degree:M.S.
Genre:Thesis
Subject(s):JPEG compression
artifact removal
CNN
DCT
Abstract:JPEG compression has been a popular lossy image compression technique and is widely used in digital imaging. Restoring high-quality images from their compressed JPEG counterparts, however, is an ill-posed inverse problem but could be of great use in improving the visual quality of images. With the representational power that convolutional neural networks (CNNs) demon- strate, we show that it is possible to suppress JPEG compression artifacts and recover visually pleasing images. To recover original high-quality and high-resolution images from JPEG compressed images, we leverage prior knowledge of JPEG compression into consideration by exploiting frequency redundancies with the CNN in discrete cosine domain and constrain the quantization loss, in addition to exploiting spatial redundancies in the pixel domain. This data-driven approach tar- gets removing compression artifacts, including blocking, blurring, ringing and banding artifacts, and recovering high-frequency information for reconstruction. We design a deep CNN in each domain and fuse the outputs with an aggregation network to produce the output image. To improve the model performance, we leverage the robustness and ability to tackle vanishing gradient problems of ResNet to build a deep network, and utilize squeeze-and- excitation block, a technique typically found beneficial in classification tasks, to this regression problem to exploit global information in a larger scale. We refer to the module proposed in this work as squeeze-and-excitation artifact removal ResNet (SE-ARResNet). Prior work in this field mainly focuses on reconstructing a grayscale image or the luminance channel of the image. We demonstrate that we can reconstruct color images effectively and robustly with the dual-domain CNN approach.
Issue Date:2019-04-26
Type:Thesis
URI:http://hdl.handle.net/2142/105275
Rights Information:Copyright 2019 Jiaxi Jason Nie
Date Available in IDEALS:2019-08-23
Date Deposited:2019-05


This item appears in the following Collection(s)

Item Statistics