由于井下煤矿环境存在能见度低、光照不均、频繁遮挡及高危作业等挑战，安全合规监测成为亟待解决的重要问题。为突破这些限制，本研究提出两种基于深度学习的先进检测方案，专门面向煤矿作业场景中的安全头盔自动识别任务。

   1)种方案针对弱光与遮挡场景，创新性地融合特征金字塔网络（Feature Pyramid Network, FPN）与视觉变换器（Vision Transformer, ViT）构建FPN+ViT模型，结合多尺度特征提取与全局上下文建模能力。其中，FPN模块增强对多尺度特征的感知能力，提升在光照变化条件下的适应性；ViT模块捕捉长距离依赖关系，提高识别的鲁棒性与准确性。实验结果表明，该模型以98.50%准确率、98.35%精确率、98.44%召回率及97.65%的F1值，全面优于Mask R-CNN、Faster R-CNN、Detectron2、EfficientNet、VGG16和DenseNet121等对比模型。

   2)种方案面向高速采矿过程中的实时检测需求，改进YOLOv8框架并引入多种注意力机制以强化特征选择与抗遮挡能力。在基准YOLOv8模型取得74%准确率的基础上，集成高效通道注意力模块（Efficient Channel Attention, ECA）与混洗注意力模块（Shuffle Attention, SA）的YOLOv8-ECA与YOLOv8-SA分别提升至77%与76%。进一步，集成卷积块注意力模块（Convolutional Block Attention Module, CBAM）的YOLOv8-ResCBAM模型在检测性能上表现最优，达到78%的准确率、80%的精确率及75%的F1值，显著提升目标定位能力与遮挡鲁棒性，验证了其在动态采矿环境中进行实时头盔检测的适用性。

   综上所述，FPN+ViT模型在极端弱光条件下展现出卓越性能，而YOLOv8-ResCBAM则为高速实时检测提供了最优解。本研究为井下煤矿中头盔自动检测与安全合规监控提供了可扩展、自适应且高效的技术框架。

Ensuring safety compliance in underground coal mines is a critical challenge due to poor visibility, uneven lighting, frequent occlusions, and hazardous working conditions. To address these limitations, this study proposes two advanced deep learning based approaches tailored for helmet detection in underground coal mining environments.  
      1) To address low-light conditions and occlusions, this research introduces a Fusion of Pyramid Network and Vision Transformer (FPN+ViT) model, integrating multi-scale feature extraction with global contextual learning. The FPN module enhances feature detection across different scales, improving robustness to variable lighting, while ViT captures long range dependencies, enhancing accuracy. Experimental results show that FPN+ViT achieves 98.50% accuracy, 98.35% precision, 98.44% recall, and a 97.65% F1-score, outperforming Mask R-CNN, Faster R-CNN, Detectron2, EfficientNet, VGG16, and DenseNet121.
      2) To improve real-time detection for high-speed mining operations, this study refines the YOLOv8 framework by integrating attention mechanisms for better feature selection and detection robustness. The baseline YOLOv8 model achieves 74% accuracy, while YOLOv8-ECA and YOLOv8-SA, incorporating Efficient Channel Attention (ECA) and Shuffle Attention (SA), improve accuracy to 77% and 76%, respectively. The YOLOv8-ResCBAM variant, integrating Convolutional Block Attention Modules (CBAM), achieves the highest performance (78% accuracy, 80% precision, 75% F1-score), demonstrating enhanced object localization and resilience to occlusions. These results confirm YOLOv8-ResCBAM's suitability for real-time helmet detection in dynamic mining environments requiring rapid processing.

      The findings of this study establish that FPN+ViT is highly effective for challenging low-light conditions, while YOLOv8-ResCBAM provides an optimized solution for high-speed, real-time applications. These approaches offer scalable, adaptable, and efficient frameworks for automated helmet detection and safety compliance monitoring in underground coal mines.