Coal and Gangue Detection Networks with Compact and High-Performance Design.

Area of Science:

• Mining Engineering
• Computer Vision
• Artificial Intelligence

Background:

• Efficient coal and gangue separation is vital for energy, environment, and sustainability in mining.
• Machine vision methods struggle with dense scenes, leading to label rewriting issues and reduced performance.
• Accurate detection of closely distributed coal and gangue on conveyor belts remains a significant challenge.

Purpose of the Study:

• To develop a novel, compact convolutional neural network (CGDet) for accurate coal and gangue detection in dense mining environments.
• To address label rewriting problems and improve model performance in challenging visual conditions.
• To create an efficient object detection system suitable for real-time applications with limited computational resources.

Main Methods:

• Introduced Object Distribution Density Measurement (ODDM) to optimize input and feature map resolutions, mitigating label rewriting.
• Developed Relative Resolution Object Scale Measurement (RROSM) to guide a streamlined feature fusion structure, reducing redundancy.
• Designed CGDet, a compact convolutional neural network incorporating ODDM and RROSM for enhanced object detection.

Main Results:

• CGDet achieved high performance with AP50 of 96.7% and AR50 of 99.2%.
• Significantly reduced model parameters (46.76%), computational cost (47.94%), and inference time (31.50%) compared to traditional models.
• Demonstrated superior accuracy and efficiency in dense coal and gangue separation scenarios.

Conclusions:

• CGDet effectively overcomes challenges in dense scene object detection for coal and gangue separation.
• The proposed ODDM and RROSM methods offer a new approach to designing efficient and accurate detection networks.
• CGDet is well-suited for real-time sorting in resource-constrained underground mining environments, enhancing operational efficiency and safety.