TL;DR
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.
Workplaces in construction and industry suffer from a significant number of workplace accidents because of a lack of safety mechanisms like helmets and high-visibility vests. In order to solve this issue, an automatic detection system for detecting the presence of the worker's helmet & safety vest using a deep learning model is developed. You Only Look Once – Neural Architecture Search (YOLO-NAS) algorithm was chosen to be used in the model because of its fast and high-quality detection process. A dataset with images containing healthy workers wearing helmets & safety vests (with some workers appearing in different safety vests) is collected and used as the training set. The resulting model is then saved for future usage in making predictions with new images. Then, this model is connected to Streamlit, which provides a convenient Web-based user interface through which a user can provide an image as an input to the model. The trained neural network is applied to the input image, and all detected objects that correspond to helmet & safety vest are put in bounding boxes on the input image. Moreover, each of the bounding boxes is labeled with the name of the object and confidence score of the detection. Thus, a clear visualization of the safety equipment worn by the worker is provided. Overall, the developed detection system significantly decreases the human intervention needed for visual verification of the safety helmet or vest. The described system is a user-friendly and cost-effective way of evaluating workplace safety compliance.
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.
To address the challenges of weak feature representation, background interference, and missed small-target detections in identifying subtle surface damages on highly reflective refrigerator doors during production, this paper proposes a lightweight detection algorithm, YOLOv8-FD, based on the YOLOv8n architecture. First, a lightweight GhostConv module replaces standard convolution layers within the backbone network. This is integrated with a modified lightweight CBAM attention mechanism (L-CBAM) to form the C2f-GC module, significantly enhancing fine-grained feature extraction under complex backgrounds while minimising parametric overhead. Furthermore, the SPPELAN module is incorporated to expand the receptive field and aggregate multi-scale features without compromising the lightweight design. To adapt to the geometric heterogeneity of defects, a novel Static-Asymmetry IoU (SAIoU) loss function based on a subtraction penalty mechanism is proposed, which optimises bounding box regression for irregular and asymmetric targets. Experimental evaluations on a strictly partitioned, independent hold-out test set demonstrate that YOLOv8-FD operates with only 5.3 M parameters and 6.0 GFLOPs of computational complexity, achieving reductions of 15.9% and 26.8%, respectively, compared to the baseline YOLOv8n model. Concurrently, it improves the mean average precision (mAP50) by 1.33 percentage points and the F1-score by 1.10 percentage points. These quantitative results empirically validate the efficacy of YOLOv8-FD in resolving fine-grained damage detection bottlenecks under resource-constrained edge-device scenarios.
Illegal waste dumping poses significant environmental and public health challenges worldwide, requiring automated surveillance systems for detection and prevention. This paper presents our solution for the IWDD 2026 Contest, addressing the dual challenge of detecting illegal dumping events in surveillance videos and localizing the exact moment of occurrence. We employ X3D-M, an efficient 3D convolutional network pretrained on Kinetics-400, combined with a sliding window inference strategy for temporal localization. Through systematic hyperparameter optimization across 96 configurations and ablation studies examining nine combinations of fine-tuning strategies and loss functions, we identify key design choices for this application domain. Our experiments reveal that differential learning rates-applying lower rates to the pretrained backbone while training the classifier more aggressively-outperform both frozen backbones and uniform fine-tuning. The optimal system achieves an F1-score of 0.8387 and a temporal F1score of 0.7742 on our test set, with 92.3% of correct detections within the temporal tolerance window. Operating at over 8 times real-time speed with only 2.97M parameters, our approach demonstrates that efficient video classification architectures can be effectively adapted for specialized surveillance applications through careful transfer learning and inference design.
Safety helmet wearing detection is a crucial component of safety management in construction sites. Traditional detection methods based on manual monitoring are inefficient, while existing deep learning models often suffer from high computational costs and poor detection performance for small or occluded targets. To address these issues, this paper proposes a lightweight detection algorithm named GCW-YOLOv8. Firstly, the Ghost Module is introduced into the backbone network to replace the conventional convolution layers, significantly reducing the number of parameters and floating-point operations (FLOPs) while maintaining feature extraction capability. Secondly, the Coordinate Attention (CA) mechanism is embedded into the neck network to enhance the model's sensitivity to spatial location and channel information, thereby improving the detection accuracy of small targets. Finally, the Wise-IoU (WIoU) loss function is adopted to replace the original CIoU loss, utilizing a dynamic non-monotonic focusing mechanism to optimize the gradient assignment for low-quality samples. Experimental results on the safety helmet dataset show that the proposed algorithm achieves a mean Average Precision (mAP@0.5) of 94.5%, while the inference speed reaches 108 FPS. Compared with the baseline YOLOv8n, our method improves detection accuracy by 1.4% while reducing parameter count by 34.4%, achieving a superior balance between accuracy and efficiency for smart construction site applications.
Large language models demonstrate high precision in power-system risk identification; however, their massive parameter counts and high resource consumption hinder real-time deployment on resource-constrained edge devices used in electromagnetic sensing systems, wearable monitoring terminals, and compact power-monitoring devices. Achieving low-latency processing in distributed sensing structures and edge-based electromagnetic monitoring devices requires a significant reduction in computational overhead to ensure immediate detection of electrical hazards, abnormal equipment states, and potential power-system risks. This paper proposes a lightweight compression method based on hierarchical supervised knowledge distillation. Experiments show that, after applying the proposed knowledge distillation method, the inference latency is reduced from 235 ms to a minimum of 26 ms, which is better than DistilBERT’s 35 ms. The number of student model parameters is reduced to 4.3% of the teacher model, namely 14.5M versus 340M, while the classification accuracy reaches 89.4%, close to the teacher model’s 92.7%. The F1 score for the equipment failure category reaches 90.3%, verifying the efficiency and practicality of the lightweight model in resource-constrained scenarios involving power-risk identification, electromagnetic sensing, and edge-based intelligent monitoring.