Jul 2026· 2026 6th International Conference on Inventive Computation and Information Technologies (ICICIT)· pp. 365-370· 0 citations· 20 references
Abstract
This paper presents a comprehensive study on deep learning-based personal protective equipment (PPE) detection for real-time construction safety monitoring by applying four recent YOLO-based object detection architectures on a unified dataset with 43,986 images and 14 PPE-related classes. The investigate the detection of not only PPE-compliance classes (Hardhat, Safety Vest, Gloves, Mask, Goggles) but also PPE-violation classes (NO-Hardhat, NO-Safety Vest, NO-Gloves, NO-Mask, NO-Goggles). Experimental results showed that minority PPE-violation classes such as NO-Safety Vest are consistently under-detected, owing to severe class imbalance and high visual resemblance to compliance classes. A class-specific weighted Binary Cross-Entropy (BCE) loss function is then proposed and applied under identical training conditions across the four evaluated architectures. Models are evaluated using precision, recall, mAP@0.50, mAP@0.50:0.95, F1-Score and inference latency. Our experiments show the weighted loss improved NO-Safety Vest performance by up to 35.5% (YOLO26-W: 0.6485 vs. YOLO26 baseline: 0.4787). The overall mAP decreased marginally. Among all considered models, YOLO11m exhibited the best trade-off between detection performance and inference speed. The results demonstrate that the proposed class-specific weighted loss strategy consistently improved minority-class detection performance across multiple YOLO architectures to improve minority PPE violation detection without the need for new data annotation and model redesign.
A production-line safety monitoring system that integrates You Only Look Once version 11 (YOLOv11) with a convolutional neural network (CNN) that demonstrated robust performance under complex conditions such as shading and varying lighting, supporting its effectiveness and practical application potential in production line safety monitoring.
Xinghua Miao, Chunlong Shen· Journal of Visualized Experi...· 0 citations
To address the challenges of diverse target poses, small object scales and complex background interference in laboratory hazardous behaviour detection, we propose BC-YOLOv10-S, an efficient visual recognition model for laboratory safety applications. The model introduces Shape-IoU to improve bounding-box regression, incorporates CBAM to enhance perception of critical hazardous regions, and adopts StarNet to reduce model complexity. Compared with the original YOLOv10 network, BC-YOLOv10-S improves Precision, Recall, F1-score and mAP by 7.42%, 6.94%, 7.18% and 4.62%, respectively. These results show that BC-YOLOv10-S delivers strong accuracy, robustness and lightweight deployment capability for laboratory safety warning, with clear engineering value and promising practical potential, thereby supporting the intelligent upgrading of laboratory safety management.
Qiang Zhao, Qingyang Zhang, Junhong Chen et al.· International Conference on...· 0 citations
Various pavements serving in different environments are vulnerable to numerous types of distress. It is crucial to promptly identify and repair fatal pavement distresses. Addressing the issues of time-consuming and labor-intensive traditional pavement distress detection, as well as its lack of precision, an adaptive and universal pavement distress detection method based on deep learning and transfer learning is proposed. First, by summarizing the primary distresses of a large number of different pavements, a classification method based on an improved vision transformer (ViT) deep learning model combined with transfer learning is proposed to categorize these distresses, achieving a Top-1 classification accuracy of more than 97% on four major distress types. Second, more than 3,000 images from multiple public datasets (CrackForest, RDD, CFD, etc.) are integrated using the improved ViT model to construct a universal pavement distress dataset covering local roads, highways, and rural roads. Finally, an improved You Only Look Once v8 (YOLOv8) deep learning model is proposed for multi-information pavement distresses detection. Experimental results demonstrate that while performance on single-pavement datasets is limited by data imbalance, training on the proposed universal pavement distress dataset significantly improves generalization, achieving 82%–89% mean average precision across multiple distress types and enabling robust detection under diverse real-world conditions. Accurate classification and detection of universal pavements have been achieved through the improved ViT and YOLOv8 deep learning models, providing a reference for pavement distress assessment and maintenance.
Wanrun Li, Tongtong Wang, Wenhai Zhao et al.· Journal of performance of co...· 0 citations
This work proposes a detection framework called distillation alignment YOLO (DA-YOLO) for PPE detection and introduces a teacher–student distillation framework with consistency constraints across predictions and high-order features extracted from baseline that enables the student model to achieve strong generalization while maintaining low computational cost.
Chonghua Zhou, Ruixuan Zhang, Yi-Xin Fu et al.· Multimedia Systems· 0 citations
Industrial pipeline inspection is a critical requirement in sectors such as water distribution, oil transportation, and manufacturing. Conventional inspection approaches remain largely manual, which leads to high operational costs, long inspection durations, and inconsistent defect detection due to human subjectivity. These limitations highlight the need for automated, accurate, and real-time inspection systems capable of reliable defect identification. In this work, we propose a robust real-time defect detection framework based on the YOLOv8 architecture enhanced with Oriented Bounding Boxes (YOLOv8-OBB). The proposed approach is specifically designed to handle elongated and arbitrarily oriented defects, such as cracks and leaks, which are common in industrial pipeline environments. The model is trained and evaluated on a custom dataset comprising 1,224 annotated images distributed across four defect categories: crack, dent, hole, and leak. Extensive experimental results demonstrate strong performance, achieving 83.92% precision, 86.06% recall, and 87.22% mAP@50, while maintaining real-time inference capability with processing times between 5 and 10 milliseconds per image. The results show the proposed system provides an efficient and scalable solution for intelligent industrial inspection and demonstrates strong potential for integration into real-world pipeline monitoring and maintenance platforms.
Manel Elleuchi, Ahmed Fakhfakh· Global Journal of Computer S...· 0 citations
Heating, Ventilation, and Air Conditioning (HVAC) systems play a critical role in ensuring energy efficiency, occupant comfort, and sustainability in modern buildings. Their complex operational dynamics,however, make them prone to faults that, if undetected, can result in higher energy consumption, reduced comfort, and costly maintenance. Existing approaches to fault detection and diagnosis (FDD) often face challenges related to scalability and interpretability, with many relying heavily on Transformer-or GNN-based models. This study introduces an enhanced TabNet-based semi-supervised framework for multi-class HVAC fault detection, incorporating rigorous pseudo-label generation via Label Propagation, cost-sensitive learning for imbalanced classes, and hybrid feature engineering. Validated on the 2022 LBNL FDD FCU dataset, the model achieves an accuracy of 82.1%, along with high ROC-AUC and F1-scores values of 96.1% and 82.2% respectively, while its attention-based feature importance provided insights into the most influential operational variables. These results demonstrate that the TabNet framework effectively balances predictive power and interpretability, making it a practical solution for semi-supervised FDD in HVAC systems. The proposed approach contributes to building automation by offering a transparent and reliable pathway for fault detection.
I. Samuel, B. Stephen, S. Ozuomba et al.· E3S Web of Conferences· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
