Full metadata record
DC FieldValueLanguage
dc.contributorDepartment of Computingen_US
dc.contributor.advisorLi, Ping (COMP)en_US
dc.creatorLi, Zhecheng-
dc.identifier.urihttps://theses.lib.polyu.edu.hk/handle/200/11387-
dc.languageEnglishen_US
dc.publisherHong Kong Polytechnic Universityen_US
dc.rightsAll rights reserveden_US
dc.titleResearch on deep neural networks on image motion deblurringen_US
dcterms.abstractWith the widespread availability of all kinds of cameras and smartphones today, taking pictures has become one of the most common activities in people's daily lives. However, due to object movement, hand shake, etc., the pictures are often blurred. Therefore, an effective dynamic scene deblurring algorithm is of great research value. Dynamic scene deblurring has drawn growing attention in recent years. Traditional methods often observe various statistical patterns in the image to derive much a priori knowledge to build a mathematical model to construct an optimization problem, but the results are usually modest and time-consuming. The widespread use of convolutional neural networks in computer vision has made deep learning models the dominant solution in recent years. Although recent deep learning-based methods have made significant progress, the models also have a lot of room for improvement. Through our research, we found that recent deep learning models are usually focus on spatial features, and thus, lack the capability to deal with blurry images in various scenarios. To tackle this problem, We investigate the effect of receptive field and nonlinear mapping ability on network performance separately. For receptive field, we introduce ResInception module to achieve different scales of receptive field by processing images with different size of convolutional layers simultaneously. We also introduce TV(Total Variation For nonlinear mapping ability, we introduce the feature attention module to mine the channel features of the image, and use the feature transformation module with nested skip connections to enhance the nonlinear capability of the network. Through experimentation, our network has had comparable performance to state-of-the-art algorithms, not only qualitatively, but also quantitatively.en_US
dcterms.extentv, 50 pages : color illustrationsen_US
dcterms.isPartOfPolyU Electronic Thesesen_US
dcterms.issued2021en_US
dcterms.educationalLevelM.Sc.en_US
dcterms.educationalLevelAll Masteren_US
dcterms.LCSHImage processing -- Digital techniquesen_US
dcterms.LCSHNeural networks (Computer science)en_US
dcterms.LCSHHong Kong Polytechnic University -- Dissertationsen_US
dcterms.accessRightsrestricted accessen_US

Files in This Item:
File Description SizeFormat 
5825.pdfFor All Users (off-campus access for PolyU Staff & Students only)3.62 MBAdobe PDFView/Open


Copyright Undertaking

As a bona fide Library user, I declare that:

  1. I will abide by the rules and legal ordinances governing copyright regarding the use of the Database.
  2. I will use the Database for the purpose of my research or private study only and not for circulation or further reproduction or any other purpose.
  3. I agree to indemnify and hold the University harmless from and against any loss, damage, cost, liability or expenses arising from copyright infringement or unauthorized usage.

By downloading any item(s) listed above, you acknowledge that you have read and understood the copyright undertaking as stated above, and agree to be bound by all of its terms.

Show simple item record

Please use this identifier to cite or link to this item: https://theses.lib.polyu.edu.hk/handle/200/11387