A lightweight coal-gangue detection model based on parallel deep residual networks

Area of Science:

• Computer Vision
• Artificial Intelligence
• Mining Engineering

Background:

• Accurate identification of coal-gangue is crucial for efficient underground coal transportation.
• Existing models face challenges with complex imaging conditions and high deployment costs.

Purpose of the Study:

• To develop a lightweight and accurate coal-gangue detection model for underground mining.
• To address limitations in feature extraction, upsampling, and optimization for real-world deployment.

Main Methods:

• Proposed a parallel depth residual network (P-RNet) for lightweight coal-gangue detection.
• Designed a feature extraction module (FEM) with decoupled training and inference.
• Optimized the feature fusion module (FFM) using a lightweight upsampling operator.
• Employed the Lookahead optimizer and extensive learning rate experiments for robust parameter optimization.

Main Results:

• The P-RNet model demonstrated effective improvement in coal-gangue recognition accuracy.
• The proposed methods successfully addressed issues with complex image conditions and upsampling artifacts.
• The model achieved a low deployment cost, making it suitable for practical applications.

Conclusions:

• P-RNet offers an effective solution for accurate coal-gangue identification in underground mining.
• The model's lightweight design and optimized components facilitate low-cost deployment.
• This research contributes to enhanced efficiency and safety in coal transportation processes.