DLR-YOLO: A High-Accuracy Lightweight Object Detector for Complex Underground Coal Mine Environments.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Mining Engineering

Background:

• Object detection in underground coal mines faces challenges like low illumination, dust, occlusion, and scale variation.
• These issues lead to missed detections and low confidence, impacting safety monitoring.

Purpose of the Study:

• To develop a high-performance, lightweight object detector for underground coal mines.
• To address limitations of existing methods in low-light and high-noise environments.

Main Methods:

• Proposed DLR-YOLO, a lightweight object detector based on YOLOv11n.
• Integrated Dynamic Multi-scale Global Perception Enhancement Module (DMGPEM) for adaptive feature extraction.
• Incorporated Lightweight Cross-Attention (LCA) module for feature fusion and noise suppression.
• Developed Reparameterized stem (RepStem) module for efficient initial feature extraction.

Main Results:

• DLR-YOLO achieved 94.4% mAP@50 and 66.7% mAP@50-95 on an in-house dataset, outperforming the YOLOv11n baseline.
• Demonstrated effectiveness of individual modules through ablation studies.
• Maintained a lightweight architecture (2.7M parameters, 6.6 GFLOPs) with high inference speed (157.1 FPS).
• Outperformed YOLOv12, YOLOv13, and RT-DETR in detection performance and speed.

Conclusions:

• DLR-YOLO offers a robust and efficient solution for object detection in complex underground coal mine environments.
• The proposed modules effectively enhance feature extraction, fusion, and noise suppression.
• Provides a strong technical foundation for real-time safety monitoring systems.