東華大學圖書館 |

語系: 繁體中文

說明(常見問題)

回圖書館首頁

手機版館藏查詢

登入

回首頁

切換: 標籤 | MARC模式 | ISBD

Exploiting Common Layers among Heter...

Mansour, Iyad Faisal Ghazi.

FindBook

Google Book

Amazon

博客來

Exploiting Common Layers among Heterogeneous CNNs in Automotive Vision Systems.

紀錄類型:	書目-電子資源 : Monograph/item
正題名/作者:	Exploiting Common Layers among Heterogeneous CNNs in Automotive Vision Systems./
作者:	Mansour, Iyad Faisal Ghazi.
出版者:	Ann Arbor : ProQuest Dissertations & Theses, : 2019,
面頁冊數:	190 p.
附註:	Source: Dissertations Abstracts International, Volume: 80-12, Section: B.
Contained By:	Dissertations Abstracts International80-12B.
標題:	Computer Engineering. -
電子資源:	http://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=13810750
ISBN:	9781392244685

Exploiting Common Layers among Heterogeneous CNNs in Automotive Vision Systems.
Mansour, Iyad Faisal Ghazi.

Exploiting Common Layers among Heterogeneous CNNs in Automotive Vision Systems. - Ann Arbor : ProQuest Dissertations & Theses, 2019 - 190 p.

Source: Dissertations Abstracts International, Volume: 80-12, Section: B.

Thesis (Ph.D.)--Oakland University, 2019.

This item is not available from ProQuest Dissertations & Theses.

In recent years we witnessed massive improvements in the field of deep learning, neural networks, and especially in the sector of computer vision, speech recognition, character recognition, and chatbots, etc. With the extraordinary increase in the amount of labelled and ready to use data available on the internet for computer vision in the form of videos, web based data, etc., and with the increase in computational power of high performance hardware, the neural networks are surpassing their performance.The computer vision industry is relying more and more on Artificial Intelligence (AI) to process data generated by Cameras, LIDAR and Radar sensors. Neural Networks (NN) have been specifically successful in processing image data and to perform classification tasks in a way that now exceeds human classification capabilities. Traditional Convolutional Neural Networks (CNNs) based algorithms require significant use of Central Processing Units (CPUs) cycles and/or Graphics Processing Units (GPUs). The majority of these CNN models are built to operate on laptops or server like solutions that are not limited by the amount of CPU/GPU available for processing. In several industries (automotive, aerospace, medical, etc.) it is necessary to operate in an environment that is limited in processing power, current consumption, and thermal output and for that reason, server-based AI solutions are not always practical. In this work, I propose a new convolutional neural network model as a viable option to reduce the overall memory requirements and execution time while maintaining the required accuracy. The Efficient Multi-Function Convolutional Neural Network (EMF-CNN) is optimized for automotive grade embedded System on a Chip (SoC) solutions that limit the computational, current consumption, and thermal output of the system.In order to evaluate the proposed EMF-CNN model, a data collection setup is created, and a labeling tool is designed to annotate the data in the appropriate format. In this research, the EMF-CNN is focused on solving two essential machine vision tasks in autonomous vehicle vision systems, objects detection and lane line detection. A new training approach is also introduced to efficiently train the EMF-CNN model. Utilizing common Key Performance Indicators (KPIs), the EMF-CNN is evaluated and compared to the standalone implementations of the selected tasks.In this research, the concept of transfer learning is applied to enhance the training of EMF-CNN along with an optimization technique for providing a perception solution for autonomous vehicles ensuring the affluence of easy deployment on restricted resourced embedded System on Chips (SoC's). The results show that the final optimized EMF-CNN model is 53% smaller in size, and 40% faster on inference time as compared to the sequential object detection and lane line detection models, with 4% loss in the overall prediction accuracy. The loss in accuracy is very minimal considering the huge saving in the size and increase in the speed of the final optimized model.

ISBN: 9781392244685Subjects--Topical Terms:

1567821
Computer Engineering.

Exploiting Common Layers among Heterogeneous CNNs in Automotive Vision Systems.
LDR:04242nmm a2200349 4500 001 2264245
005 20200423112925.5
008 220629s2019 ||||||||||||||||| ||eng d
020 $a 9781392244685
035 $a (MiAaPQ)AAI13810750
035 $a (MiAaPQ)oakland:10137
035 $a AAI13810750
040 $a MiAaPQ $c MiAaPQ
100 1 $a Mansour, Iyad Faisal Ghazi. $3 3541346
245 1 0 $a Exploiting Common Layers among Heterogeneous CNNs in Automotive Vision Systems.
260 1 $a Ann Arbor : $b ProQuest Dissertations & Theses, $c 2019
300 $a 190 p.
500 $a Source: Dissertations Abstracts International, Volume: 80-12, Section: B.
500 $a Publisher info.: Dissertation/Thesis.
500 $a Advisor: Rawashdeh, Osamah A.
502 $a Thesis (Ph.D.)--Oakland University, 2019.
506 $a This item is not available from ProQuest Dissertations & Theses.
506 $a This item must not be sold to any third party vendors.
520 $a In recent years we witnessed massive improvements in the field of deep learning, neural networks, and especially in the sector of computer vision, speech recognition, character recognition, and chatbots, etc. With the extraordinary increase in the amount of labelled and ready to use data available on the internet for computer vision in the form of videos, web based data, etc., and with the increase in computational power of high performance hardware, the neural networks are surpassing their performance.The computer vision industry is relying more and more on Artificial Intelligence (AI) to process data generated by Cameras, LIDAR and Radar sensors. Neural Networks (NN) have been specifically successful in processing image data and to perform classification tasks in a way that now exceeds human classification capabilities. Traditional Convolutional Neural Networks (CNNs) based algorithms require significant use of Central Processing Units (CPUs) cycles and/or Graphics Processing Units (GPUs). The majority of these CNN models are built to operate on laptops or server like solutions that are not limited by the amount of CPU/GPU available for processing. In several industries (automotive, aerospace, medical, etc.) it is necessary to operate in an environment that is limited in processing power, current consumption, and thermal output and for that reason, server-based AI solutions are not always practical. In this work, I propose a new convolutional neural network model as a viable option to reduce the overall memory requirements and execution time while maintaining the required accuracy. The Efficient Multi-Function Convolutional Neural Network (EMF-CNN) is optimized for automotive grade embedded System on a Chip (SoC) solutions that limit the computational, current consumption, and thermal output of the system.In order to evaluate the proposed EMF-CNN model, a data collection setup is created, and a labeling tool is designed to annotate the data in the appropriate format. In this research, the EMF-CNN is focused on solving two essential machine vision tasks in autonomous vehicle vision systems, objects detection and lane line detection. A new training approach is also introduced to efficiently train the EMF-CNN model. Utilizing common Key Performance Indicators (KPIs), the EMF-CNN is evaluated and compared to the standalone implementations of the selected tasks.In this research, the concept of transfer learning is applied to enhance the training of EMF-CNN along with an optimization technique for providing a perception solution for autonomous vehicles ensuring the affluence of easy deployment on restricted resourced embedded System on Chips (SoC's). The results show that the final optimized EMF-CNN model is 53% smaller in size, and 40% faster on inference time as compared to the sequential object detection and lane line detection models, with 4% loss in the overall prediction accuracy. The loss in accuracy is very minimal considering the huge saving in the size and increase in the speed of the final optimized model.
590 $a School code: 0446.
650 4 $a Computer Engineering. $3 1567821
650 4 $a Artificial intelligence. $3 516317
650 4 $a Computer science. $3 523869
690 $a 0464
690 $a 0800
690 $a 0984
710 2 $a Oakland University. $b Engineering. $3 3288992
773 0 $t Dissertations Abstracts International $g 80-12B.
790 $a 0446
791 $a Ph.D.
792 $a 2019
793 $a English
856 4 0 $u http://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=13810750