國立虎尾科技大學 |

Methods to Improve the Predictive Performance of Deep Learning Systems for Aerial Roof Building Detection for Multi-Spectral Images.

紀錄類型:	書目-語言資料,手稿 : Monograph/item
正題名/作者:	Methods to Improve the Predictive Performance of Deep Learning Systems for Aerial Roof Building Detection for Multi-Spectral Images./
作者:	Churi, Advait.
面頁冊數:	1 online resource (70 pages)
附註:	Source: Masters Abstracts International, Volume: 82-08.
Contained By:	Masters Abstracts International82-08.
標題:	Computer engineering. -
電子資源:	click for full text (PQDT)
ISBN:	9798569981007

Methods to Improve the Predictive Performance of Deep Learning Systems for Aerial Roof Building Detection for Multi-Spectral Images.
Churi, Advait.

Methods to Improve the Predictive Performance of Deep Learning Systems for Aerial Roof Building Detection for Multi-Spectral Images. - 1 online resource (70 pages)

Source: Masters Abstracts International, Volume: 82-08.

Thesis (M.S.)--University of Massachusetts Lowell, 2021.

Includes bibliographical references

In this thesis work, we study the challenges for multispectral visible-infrared aerial images. We present a dilated convolution neural network-based deep learning system. The proposed system utilizes an automatic multispectral band selection strategy and automatic histogram-based threshold finding for building roof segmentation to improve the deep learning system prediction accuracy. Current existing systems train one model on all the data at once. These systems usually use RGB (Red, Green, Blue) bands and avoid relevant post-processing of the predicted images. We present a novel method to solve some of these object detection and recognition challenges. In particular, we 1) implement separate models for each dataset, 2) propose an optimal automatic band selection strategy to choose the most informative bands from given multispectral images, 3) present a histogram-based automatic optimal threshold finding technique for segmentation. Our results show that having a separate model for each city rather than only one model for all the combined cities increases the prediction accuracy from by 72.54% to 87.61% (by 20.77%) for RGB (Red, Green, Blue) bands and from 82.99% to 83.87% (by 1.06%) for bands selected via the proposed band selection strategy. As our experimental results show, our proposed deep learning system has a 35.6% increase in building detection prediction accuracy F1-score(from 0.688% to 0.933%) compared to the winning implementation of SpaceNet Building Detection US National challenge-2 (2017) for the city of Rio.

Electronic reproduction.
Ann Arbor, Mich. :
ProQuest,
2024

Mode of access: World Wide Web

ISBN: 9798569981007Subjects--Topical Terms:

569006
Computer engineering.
Subjects--Index Terms:

Predictive performanceIndex Terms--Genre/Form:

554714
Electronic books.

Methods to Improve the Predictive Performance of Deep Learning Systems for Aerial Roof Building Detection for Multi-Spectral Images.
LDR:02942ntm a22003737 4500 001 1144208
005 20240531083759.5
006 m o d
007 cr mn ---uuuuu
008 250605s2021 xx obm 000 0 eng d
020 $a 9798569981007
035 $a (MiAaPQ)AAI28257624
035 $a AAI28257624
040 $a MiAaPQ $b eng $c MiAaPQ $d NTU
100 1 $a Churi, Advait. $3 1469157
245 1 0 $a Methods to Improve the Predictive Performance of Deep Learning Systems for Aerial Roof Building Detection for Multi-Spectral Images.
264 0 $c 2021
300 $a 1 online resource (70 pages)
336 $a text $b txt $2 rdacontent
337 $a computer $b c $2 rdamedia
338 $a online resource $b cr $2 rdacarrier
500 $a Source: Masters Abstracts International, Volume: 82-08.
500 $a Advisor: Megherbi, Dalila B.
502 $a Thesis (M.S.)--University of Massachusetts Lowell, 2021.
504 $a Includes bibliographical references
520 $a In this thesis work, we study the challenges for multispectral visible-infrared aerial images. We present a dilated convolution neural network-based deep learning system. The proposed system utilizes an automatic multispectral band selection strategy and automatic histogram-based threshold finding for building roof segmentation to improve the deep learning system prediction accuracy. Current existing systems train one model on all the data at once. These systems usually use RGB (Red, Green, Blue) bands and avoid relevant post-processing of the predicted images. We present a novel method to solve some of these object detection and recognition challenges. In particular, we 1) implement separate models for each dataset, 2) propose an optimal automatic band selection strategy to choose the most informative bands from given multispectral images, 3) present a histogram-based automatic optimal threshold finding technique for segmentation. Our results show that having a separate model for each city rather than only one model for all the combined cities increases the prediction accuracy from by 72.54% to 87.61% (by 20.77%) for RGB (Red, Green, Blue) bands and from 82.99% to 83.87% (by 1.06%) for bands selected via the proposed band selection strategy. As our experimental results show, our proposed deep learning system has a 35.6% increase in building detection prediction accuracy F1-score(from 0.688% to 0.933%) compared to the winning implementation of SpaceNet Building Detection US National challenge-2 (2017) for the city of Rio.
533 $a Electronic reproduction. $b Ann Arbor, Mich. : $c ProQuest, $d 2024
538 $a Mode of access: World Wide Web
650 4 $a Computer engineering. $3 569006
650 4 $a Artificial intelligence. $3 559380
653 $a Predictive performance
653 $a Deep learning systems
653 $a Aerial roof building detection
653 $a Multi-spectral images
655 7 $a Electronic books. $2 local $3 554714
690 $a 0464
690 $a 0800
710 2 $a ProQuest Information and Learning Co. $3 1178819
710 2 $a University of Massachusetts Lowell. $b Computer Engineering. $3 1469158
773 0 $t Masters Abstracts International $g 82-08.
856 4 0 $u http://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=28257624 $z click for full text (PQDT)