Lightweight coal mine conveyor belt foreign object detection based on improved Yolov8n

Area of Science:

• Computer Vision
• Machine Learning
• Industrial Automation

Background:

• Deep learning models for conveyor belt foreign object detection face challenges with slow speeds, high parameter counts, and computational demands.
• Existing methods struggle with real-time processing and efficient deployment in industrial settings.

Purpose of the Study:

• To develop a lightweight and efficient foreign object detection algorithm for conveyor belts.
• To address the limitations of existing deep learning models in terms of speed, parameter count, and computational load.

Main Methods:

• An improved Yolov8n model incorporating a lightweight StarNet backbone, C2f.-EIEM module, Large Separable Kernel Attention (LSKA), C2f_MLCA, and a Detect-LSDECD head.
• Replacement of CIoU loss with MPDIoU for enhanced bounding box prediction accuracy.
• Integration of Detail Enhancement Convolution (DEConv) for improved recognition stability.

Main Results:

• Reduced model parameters by 80%, computational load by 60.49%, and storage size by 69.35%.
• Achieved a 1.9% increase in accuracy while maintaining the recall rate.
• Demonstrated enhanced feature learning, focus on key features, and improved recognition stability.

Conclusions:

• The proposed lightweight algorithm effectively overcomes the limitations of traditional methods for conveyor belt foreign object detection.
• The optimized model is suitable for real-time applications, particularly on coal conveyor belts in mining environments.
• Significant reductions in model size and computational requirements pave the way for efficient deployment.