國立高雄大學圖資館 |

Language: English

Back

Characterizing Surface Structures Captured in Scanning Electron Microscopy Images with Convolutional Neural Networks.

Record Type:	Electronic resources : Monograph/item
Title/Author:	Characterizing Surface Structures Captured in Scanning Electron Microscopy Images with Convolutional Neural Networks.
Author:	Ly, Nhat-Cuong Hoang.
Published:	Ann Arbor : ProQuest Dissertations & Theses, 2022
Description:	107 p.
Notes:	Source: Dissertations Abstracts International, Volume: 84-08, Section: B.
Notes:	Advisor: Tasdizen, Tolga.
Contained By:	Dissertations Abstracts International84-08B.
Subject:	Computer science.
Online resource:	http://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=29321489
ISBN:	9798371977557

Characterizing Surface Structures Captured in Scanning Electron Microscopy Images with Convolutional Neural Networks.
Ly, Nhat-Cuong Hoang.

Characterizing Surface Structures Captured in Scanning Electron Microscopy Images with Convolutional Neural Networks. - Ann Arbor : ProQuest Dissertations & Theses, 2022 - 107 p.

Source: Dissertations Abstracts International, Volume: 84-08, Section: B.

Thesis (Ph.D.)--The University of Utah, 2022.

This item must not be sold to any third party vendors.

Nuclear forensics analysis provides insights into the origin and processing history of an unknown interdicted nuclear material. Knowing the processing history of an illicitly trafficked material is key to mitigate illegal activities and add new safeguards preventing future recurrent of a similar incident. Recently, several studies have demonstrated that the surface structures captured in scanning electron microscopy (SEM) images can be used as signatures for discerning processing histories. However, the current quantitative analysis for characterizing surface structures is time-consuming and tedious due to the manual segmentation step. Furthermore, this analysis utilizes only a few morphological features of the surface structures to discern processing histories, which might not be sufficient to establish the statistical relevance for materials with extremely similar surface structures. Therefore, in this dissertation, machine learning models, convolutional neural networks (CNNs) in particular, is first proposed to 1) improve the analysis speed; and 2) discover more meaningful surface structure features, in turn providing a better tool for processing history discernment. Then, the second part of this dissertation focuses on improving the functionality of CNNs to further strengthen their use for discerning processing histories. Specifically, CNN is a data-driven approach. A labeled dataset with a reasonable sample size and representation is essential to obtain a good result with CNNs. Obtaining such a dataset is a cost-prohibitive and labor-intensive process. Another well-known shortcoming of CNNs is that they are prone to performance degradation when encountering out-of-distribution samples during inference. Thus, the second part of this dissertation addresses the above shortcomings by 3) leveraging synthetic samples to reduce the need for collecting necessary data for training CNNs, and 4) improving the robustness of CNNs via a domain adaptation framework.

ISBN: 9798371977557Subjects--Topical Terms:

199325
Computer science.
Subjects--Index Terms:

Computer vision

Characterizing Surface Structures Captured in Scanning Electron Microscopy Images with Convolutional Neural Networks.
LDR:03308nmm a2200409 4500 001 636159
005 20230501063914.5
006 m o d
007 cr#unu||||||||
008 230724s2022 ||||||||||||||||| ||eng d
020 $a 9798371977557
035 $a (MiAaPQ)AAI29321489
035 $a AAI29321489
040 $a MiAaPQ $c MiAaPQ
100 1 $a Ly, Nhat-Cuong Hoang. $3 942530
245 1 0 $a Characterizing Surface Structures Captured in Scanning Electron Microscopy Images with Convolutional Neural Networks.
260 1 $a Ann Arbor : $b ProQuest Dissertations & Theses, $c 2022
300 $a 107 p.
500 $a Source: Dissertations Abstracts International, Volume: 84-08, Section: B.
500 $a Advisor: Tasdizen, Tolga.
502 $a Thesis (Ph.D.)--The University of Utah, 2022.
506 $a This item must not be sold to any third party vendors.
520 $a Nuclear forensics analysis provides insights into the origin and processing history of an unknown interdicted nuclear material. Knowing the processing history of an illicitly trafficked material is key to mitigate illegal activities and add new safeguards preventing future recurrent of a similar incident. Recently, several studies have demonstrated that the surface structures captured in scanning electron microscopy (SEM) images can be used as signatures for discerning processing histories. However, the current quantitative analysis for characterizing surface structures is time-consuming and tedious due to the manual segmentation step. Furthermore, this analysis utilizes only a few morphological features of the surface structures to discern processing histories, which might not be sufficient to establish the statistical relevance for materials with extremely similar surface structures. Therefore, in this dissertation, machine learning models, convolutional neural networks (CNNs) in particular, is first proposed to 1) improve the analysis speed; and 2) discover more meaningful surface structure features, in turn providing a better tool for processing history discernment. Then, the second part of this dissertation focuses on improving the functionality of CNNs to further strengthen their use for discerning processing histories. Specifically, CNN is a data-driven approach. A labeled dataset with a reasonable sample size and representation is essential to obtain a good result with CNNs. Obtaining such a dataset is a cost-prohibitive and labor-intensive process. Another well-known shortcoming of CNNs is that they are prone to performance degradation when encountering out-of-distribution samples during inference. Thus, the second part of this dissertation addresses the above shortcomings by 3) leveraging synthetic samples to reduce the need for collecting necessary data for training CNNs, and 4) improving the robustness of CNNs via a domain adaptation framework.
590 $a School code: 0240.
650 4 $a Computer science. $3 199325
650 4 $a Information science. $3 190425
650 4 $a Information technology. $3 184390
653 $a Computer vision
653 $a Convolutional neural networks
653 $a Image analysis
653 $a Machine learning
653 $a Nuclear forensics
653 $a Scanning electron microscopy
690 $a 0984
690 $a 0723
690 $a 0489
710 2 $a The University of Utah. $b Electrical and Computer Engineering. $3 531049
773 0 $t Dissertations Abstracts International $g 84-08B.
790 $a 0240
791 $a Ph.D.
792 $a 2022
793 $a English
856 4 0 $u http://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=29321489