SIRI-YOLO: A Foreign Object Detection Method for Belt Conveyors in High-Entropy Underground Scenes.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Mining Engineering

Background:

• Poor detection in low-light coal mines due to noise and entropy degradation.
• Challenges in multi-scale object extraction with uneven entropy distribution.

Purpose of the Study:

• Propose an efficient foreign object detection model (SIRI-YOLO) for low-light coal mine environments.
• Enhance detection performance by addressing information entropy degradation and background noise.
• Improve multi-scale target extraction capabilities.

Main Methods:

• Developed SIRI-YOLO based on an improved YOLOv11n architecture.
• Integrated a Self-Calibrating Illumination Network (SCINet) for low-light image enhancement.
• Enhanced feature extraction with C2PSA-IRMB and RepGFPN modules.
• Utilized Inner-MPDIoU loss for accurate entropy minimization.

Main Results:

• SIRI-YOLO achieved 92.8% mAP@50 and 87.2% recall on coal mine datasets.
• Demonstrated superior performance over mainstream YOLO models with 2.92M parameters and 70.01 FPS.
• Improved mAP@50 by 4.2% on the ExDark low-light dataset, showing strong generalization.

Conclusions:

• SIRI-YOLO effectively handles low-light conditions, scale variations, and complex backgrounds in coal mines.
• The model offers a practical solution for enhancing coal mine safety through reduced system entropy.
• The proposed methods significantly improve foreign object detection in challenging industrial environments.