YOLO_Bolt: a lightweight network model for bolt detection.

Area of Science:

• Computer Vision
• Machine Learning
• Industrial Automation

Background:

• Accurate workpiece identification is crucial for industrial production efficiency.
• Existing models like YOLOv5 face limitations with hardware resource constraints in factory settings.
• Need for lightweight yet accurate object detection models for real-time industrial applications.

Purpose of the Study:

• To develop a lightweight object detection model, YOLO_Bolt, by optimizing YOLOv5 for industrial environments.
• To enhance feature utilization and detection accuracy while minimizing model size.
• To provide effective auxiliary technical support for automated workpiece detection.

Main Methods:

• Implemented ghost bottleneck lightweight deep convolution in backbone and neck modules of YOLOv5.
• Introduced an asymptotic feature pyramid network to improve feature utilization and reduce interference.
• Redesigned the decoupling head layers based on task-specific requirements and loss function analysis.

Main Results:

• YOLO_Bolt achieved a model size of 6.8 M parameters, halving the original YOLOv5s size.
• mAP increased by 2.4% on the MSCOCO 2017 dataset, with a 104 frames/s FPS improvement.
• On a homemade bolt dataset, mAP 0.5 increased by 4.2%, outperforming YOLOv8s by 1.2%.

Conclusions:

• YOLO_Bolt offers a significant reduction in model parameters and improved detection performance.
• The lightweight design and enhanced feature utilization make it suitable for resource-constrained industrial environments.
• The proposed model provides effective technical support for intelligent workpiece identification in manufacturing.