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

Sensors, Journal Year: 2025, Volume and Issue: 25(7), P. 2103 - 2103

Published: March 27, 2025

Accurate real-time monitoring of underground conditions in coal mines is crucial for effective production management. However, limited computational resources and complex environmental mine shafts significantly impact the recognition capabilities detection models. This study utilizes a comprehensive dataset containing 117,887 images from five common mining tasks: personnel detection, large lump identification, conveyor chain monitoring, miner behavior recognition, hydraulic support shield inspection. We propose ESFENet backbone network, incorporating Global Response Normalization (GRN) module to enhance feature capture stability while employing depthwise separable convolutions HGRNBlock modules reduce parameter volume complexity. Building upon this foundation, we UCM-Net, model based on YOLO architecture. Furthermore, self-supervised pre-training method introduced generate mine-specific pre-trained weights, providing with more semantic features. utilizing combined neck portions as encoder an image-masking structure strengthen acquisition improve performance small models learning. Experimental results demonstrate that UCM-Net outperforms both baseline state-of-the-art YOLOv12 terms accuracy efficiency across datasets. The proposed architecture achieves 21.5% reduction 14.8% load decrease compared showing notable improvements 1.3% (mAP50:95) 0.8% (mAP50) recognition. framework effectively enhances training efficiency, enabling attain average mAP50 94.4% all research outcomes can provide key technical safety offer valuable technological insights public sector.

Language: Английский

3D location of gangue by point cloud segmentation with RG-TCF

International Journal of Coal Preparation and Utilization, Journal Year: 2024, Volume and Issue: unknown, P. 1 - 24

Published: Dec. 30, 2024

The existing method of gangue location, primarily relying on 2D coordinates and simplified 3D coordinates, often results in distorted position information, leading to failures sorting. In this paper, we propose region growing with two-component feature (RG-TCF) algorithm segment the complete uncut point cloud coal for accurate location. Firstly, workflow RG-TCF was developed by advantage fast histograms (FPFH) over angle between two normal vectors used RG (region growing). Secondly, extraction, validation test sets were built based production annotation cloud. Thirdly, after eliminating noise redundant points proposed down-sampling key (DS-KP), segmentation thresholds also worked out histogram analysis. Finally, performance validated tested location experiments. It could be concluded that improved under-segmentation effectively; it increased Dice coefficient precision 10.8% 9.2% compared those popular algorithms, respectively.

Language: Английский