國立高雄大學圖資館 |

登入

回首頁

Efficient Small Tool Detection in Construction via Lightweight Deep Neural Networks /

紀錄類型:	書目-電子資源 : Monograph/item
正題名/作者:	Efficient Small Tool Detection in Construction via Lightweight Deep Neural Networks /Maryam Soleymani.
作者:	Soleymani, Maryam,
面頁冊數:	1 electronic resource (58 pages)
附註:	Source: Masters Abstracts International, Volume: 87-05.
附註:	Advisors: Turner, Bert S.
Contained By:	Masters Abstracts International87-05.
標題:	Fatalities.
電子資源:	http://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=32306973
ISBN:	9798265412980

Efficient Small Tool Detection in Construction via Lightweight Deep Neural Networks /
Soleymani, Maryam,

Efficient Small Tool Detection in Construction via Lightweight Deep Neural Networks /Maryam Soleymani. - 1 electronic resource (58 pages)

Source: Masters Abstracts International, Volume: 87-05.

Construction sites are dynamic and inherently hazardous environments, where small hand tools-although essential-pose serious safety risks due to their frequent use, portability, and tendency to be misplaced or dropped. This study introduces a novel and lightweight deep learning-based architecture, Lightweight Small Tool Detection (LSTD), specifically designed for fast detection of small tools in unstructured and challenging construction environments. Recognizing that small object detection remains a persistent limitation in existing computer vision models, particularly under poor lighting or cluttered backgrounds, LSTD integrates advanced modules for enhanced feature extraction, fusion, and classification. It achieves notable improvements in accuracy, recall, and computational efficiency compared to existing methods.The core architecture of LSTD is built upon YOLOv5 but is augmented with three significant components: Dynamic Feature Extraction (DFE), Integrated Feature Fusion (IFF), and Accurate Separated Head (ASH). These modules are tailored to improve the detection of small tools by capturing fine-grained details with fewer parameters. The DFE module focuses on preserving critical visual cues despite the limited size of target objects, while the IFF module ensures robust multi-scale feature representation without excessive computation. The ASH module decouples regression and classification to enhance convergence and precision. Together, these innovations lead to a substantial 73% reduction in model parameters and a 28% drop in computational load, while achieving a mean Average Precision (mAP) of 87.3%.To evaluate the robustness and generalizability of the LSTD model, the research utilized a comprehensive dataset of over 34,700 images of 12 commonly used construction tools.The dataset incorporated diverse conditions such as occlusions, varying illumination levels, and realistic construction site backgrounds. Experiments included ablation studies, comparisons with state-of-the-art models, and stress testing under misty and low-light scenarios. The LSTD model consistently outperformed other lightweight detectors like YOLOv6-Small, YOLOv7-Tiny,and YOLOv8-Small, demonstrating high detection accuracy across different environmental challenges. Moreover, the integration of Convolutional Block Attention Module (CBAM) proved critical in enabling the model to effectively distinguish tools from visually complex backgrounds.Ultimately, this research contributes a compact yet powerful detection model that supports real-time applications on edge devices, paving the way for smarter and safer construction practices. By enabling accurate detection of misplaced, dropped, or unauthorized tools, LSTD offers a proactive mechanism to reduce tripping hazards, improve inventory tracking, and support robotic monitoring systems. Its lightweight design makes it particularly suited for deployment in mobile platforms and embedded construction site monitoring systems. Future work may explore edge implementation, integration with autonomous robotics, and real-time adaptive learning in dynamic construction environments. This work marks a significant step forward in aligning modern computer vision technologies with the urgent safety needs of the construction industry.

English

ISBN: 9798265412980Subjects--Topical Terms:

966945
Fatalities.

Efficient Small Tool Detection in Construction via Lightweight Deep Neural Networks /
LDR:04586nmm a22003853i 4500 001 693792
005 20260511063025.5
006 m o d
007 cr|nu||||||||
008 260819s2025 miu||||||m |||||||eng d
020 $a 9798265412980
035 $a (MiAaPQD)AAI32306973
035 $a AAI32306973
040 $a MiAaPQD $b eng $c MiAaPQD $e rda
100 1 $a Soleymani, Maryam, $e author. $3 1010480
245 1 0 $a Efficient Small Tool Detection in Construction via Lightweight Deep Neural Networks / $c Maryam Soleymani.
264 1 $a Ann Arbor : $b ProQuest Dissertations & Theses, $c 2025
300 $a 1 electronic resource (58 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: 87-05.
500 $a Advisors: Turner, Bert S.
502 $b M.S. $c Louisiana State University and Agricultural & Mechanical College $d 2025.
520 $a Construction sites are dynamic and inherently hazardous environments, where small hand tools-although essential-pose serious safety risks due to their frequent use, portability, and tendency to be misplaced or dropped. This study introduces a novel and lightweight deep learning-based architecture, Lightweight Small Tool Detection (LSTD), specifically designed for fast detection of small tools in unstructured and challenging construction environments. Recognizing that small object detection remains a persistent limitation in existing computer vision models, particularly under poor lighting or cluttered backgrounds, LSTD integrates advanced modules for enhanced feature extraction, fusion, and classification. It achieves notable improvements in accuracy, recall, and computational efficiency compared to existing methods.The core architecture of LSTD is built upon YOLOv5 but is augmented with three significant components: Dynamic Feature Extraction (DFE), Integrated Feature Fusion (IFF), and Accurate Separated Head (ASH). These modules are tailored to improve the detection of small tools by capturing fine-grained details with fewer parameters. The DFE module focuses on preserving critical visual cues despite the limited size of target objects, while the IFF module ensures robust multi-scale feature representation without excessive computation. The ASH module decouples regression and classification to enhance convergence and precision. Together, these innovations lead to a substantial 73% reduction in model parameters and a 28% drop in computational load, while achieving a mean Average Precision (mAP) of 87.3%.To evaluate the robustness and generalizability of the LSTD model, the research utilized a comprehensive dataset of over 34,700 images of 12 commonly used construction tools.The dataset incorporated diverse conditions such as occlusions, varying illumination levels, and realistic construction site backgrounds. Experiments included ablation studies, comparisons with state-of-the-art models, and stress testing under misty and low-light scenarios. The LSTD model consistently outperformed other lightweight detectors like YOLOv6-Small, YOLOv7-Tiny,and YOLOv8-Small, demonstrating high detection accuracy across different environmental challenges. Moreover, the integration of Convolutional Block Attention Module (CBAM) proved critical in enabling the model to effectively distinguish tools from visually complex backgrounds.Ultimately, this research contributes a compact yet powerful detection model that supports real-time applications on edge devices, paving the way for smarter and safer construction practices. By enabling accurate detection of misplaced, dropped, or unauthorized tools, LSTD offers a proactive mechanism to reduce tripping hazards, improve inventory tracking, and support robotic monitoring systems. Its lightweight design makes it particularly suited for deployment in mobile platforms and embedded construction site monitoring systems. Future work may explore edge implementation, integration with autonomous robotics, and real-time adaptive learning in dynamic construction environments. This work marks a significant step forward in aligning modern computer vision technologies with the urgent safety needs of the construction industry.
546 $a English
590 $a School code: 0107
650 4 $a Fatalities. $3 966945
650 4 $a Deep learning. $3 915589
650 4 $a Computer vision. $3 200113
650 4 $a Real time. $3 990179
650 4 $a Neural networks. $3 915594
650 4 $a Electrocutions. $3 966944
650 4 $a Design. $3 204000
650 4 $a Tools. $3 806555
650 4 $a Semantics. $3 177636
650 4 $a Robotics. $3 181952
650 4 $a Computer science. $3 199325
690 $a 0771
690 $a 0389
690 $a 0800
690 $a 0984
710 2 $a Louisiana State University and Agricultural & Mechanical College. $3 942606
720 1 $a Turner, Bert S. $e degree supervisor.
773 0 $t Masters Abstracts International $g 87-05.
790 $a 0107
791 $a M.S.
792 $a 2025
856 4 0 $u http://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=32306973