# MCF-Net: Personnel and machinery detection model for complex downhole drilling environments

## TL;DR

This paper introduces MCF-Net, a lightweight detection model for identifying personnel and machinery in underground coal mines, improving accuracy and speed in challenging conditions.

## Contribution

The novel MCF-Net model uses a lightweight attention mechanism and optimized architecture for better performance in low-light and occluded environments.

## Key findings

- MCF-Net improves mAP@0.5 by 0.013 and mAP@0.5:0.95 by 0.046 compared to the baseline model.
- The model reduces complexity by 2.27MB in Params and 8.4 in GLOPs while increasing inference speed by 11.83 FPS.

## Abstract

Detection of personnel and machinery in the drilling environment of underground coal mines is crucial to the safe production of coal. However, the existing detection models seriously affect the accuracy of the detection models due to the problems of insufficient light and mutual occlusion in the underground. To address these problems, this study proposes a lightweight downhole personnel and machinery detection model (MCF-Net), which aims to solve the above problems while improving the detection accuracy and speed of the model. In this study, YOLOv10 is used as the baseline model, and the PSA module in the backbone network is replaced by designing a lightweight attention mechanism MLBAM, which improves the model’s multi-scale feature extraction capability, enhances the model’s detection performance of mutually occluded objects and reduces the model’s complexity. In the neck network, C2f is reconstructed to get C2f-DualConv based on DualConv, and the group convolution technique is used to extract downhole image features, which can effectively overcome the influence of interference factors such as insufficient illumination in the downhole. Finally, Focaler-CIoU is introduced to reconstruct the IoU loss function, which can accurately locate the downhole objects. In addition, this study conducts experiments on a real downhole borehole dataset, and the results show that compared with the baseline model, MCF-Net improves the accuracy by 0.013 for mAP @ 0.5 and 0.046 for mAP @ 0.5:0.95, reduces the model complexity by 2.27MB for Params and 8.4 for GLOPs, and improves the inference speed by FPS is improved by 11.83 f/s.