Coal mine belt foreign object detection by lightweight YOLO and super-resolution fusion

To address the challenges of low brightness and high dust concentrations in coal mine belt conveyor systems, which degrade image quality and hinder detection accuracy, we propose an integrated framework combining image enhancement with an optimized deep-learning-based detection algorithm. The pipeline includes adaptive brightness adjustment, SwinIR-based denoising, and super-resolution, validated by no-reference image quality assessment (IQA) metrics. On the CUMT_BelT dataset, applying the preprocessing pipeline to the baseline YOLOv11n improves mAP@0.5 from 67.1% (raw images) to 70.0%. The baseline YOLOv11n is further optimized via a DWConv-GhostConv hybrid module and C3Ghost/GhostBottleneck integration, achieving a 30% parameter reduction while maintaining competitive detection accuracy. Experimental results on the CUMT_BelT dataset show that the proposed method achieves 73.8% mAP@0.5, representing a 3.8% improvement over the enhanced baseline and a 6.7% improvement over raw-data detection, with real-time inference at 72.99 FPS. Benchmark evaluations confirm its robustness in harsh mining environments, demonstrating practical viability for safety-critical foreign object detection.