Real-time damage detection network for mine conveyor belts based on knowledge distillation

Abstract

Current methods for damage detecting in mining conveyor belts face challenges. These include large model parameters that hinder real-time detection and imbalanced sample data, leading to low accuracy. This paper proposes a real-time damage detection system based on knowledge distillation to address these issues. Firstly, we introduce a GhostConv-based feature extraction block, replacing redundant convolution operations with linear transformations, significantly reducing model parameters and computation. Additionally, a Damage Shape Convolutional (DSC) module aligns the network’s receptive field with damage shapes, reducing missed detections. Furthermore, we employee a knowledge distillation framework allows the student model to learn from the teacher model, enhancing accuracy without increasing parameters. Finally, Class Focal Loss (CFL) addresses sample class imbalance with an inverse weighting strategy. Validation on YOLOv5 and YOLOv8 lightweight models achieves AP values of 98.9% and 98.7%, with parameter reductions of 36.0% and 48.8% respectively, AP in each category exceeded 95%.