Enhancing Object Detection in Underground Mines: UCM-Net and Self-Supervised Pre-Training.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Mining Engineering

Background:

• Underground coal mine monitoring faces challenges due to limited computational resources and harsh environments.
• Existing detection models struggle with recognition and computational demands in mine shafts.

Purpose of the Study:

• To develop an accurate and computationally efficient AI model for real-time monitoring of underground coal mining operations.
• To enhance feature capture stability and reduce model complexity for improved performance in challenging mining conditions.

Main Methods:

• Proposed ESFENet backbone with Global Response Normalization (GRN) and depthwise separable convolutions.
• Developed UCM-Net detection model based on YOLO architecture.
• Implemented a self-supervised pre-training method using an image-masking strategy for mine-specific feature acquisition.

Main Results:

• UCM-Net demonstrated superior accuracy and parameter efficiency compared to baseline and YOLOv12 models.
• Achieved 21.5% parameter reduction and 14.8% computational load decrease.
• Self-supervised pre-training enhanced training efficiency, yielding an average mAP50 of 94.4% across five datasets.

Conclusions:

• UCM-Net offers a robust solution for coal mine safety monitoring, improving detection capabilities.
• The proposed methods provide significant technical support for public safety in mining sectors.
• The study highlights the effectiveness of tailored AI models and self-supervised learning in specialized environments.