Scholarlink

<p>The automatic segmentation of the coal flow area plays a crucial role in coal mine monitoring. Through real-time and accurate segmentation results, the belt speed can be dynamically adjusted to achieve energy savings and reduce consumption. However, a challenge persists for existing segmentation methods to simultaneously balance accuracy, speed, and robustness within the harsh mine environment. To address this challenge, a lightweight real-time semantic segmentation network, termed MVitsU-net, is proposed in this study. The computational efficiency of the model is significantly improved through an optimized MobileVitV2 encoder and a network pruning strategy. To mitigate the information loss induced by the lightweight design, a Multi-scale Dynamic Attention Fusion Module (MD-AM) is designed for the skip connections to preserve critical detail features. Finally, an Efficient Multi-scale Attention (EMA) mechanism is integrated within the decoder to enhance edge feature characterization and model generalization. Experimental results demonstrate that MVitsU-net achieves an mIoU of 95.62% and an inference speed of 74.41 FPS on a self-constructed dataset. Moreover, its balance between accuracy and speed is shown to be superior to that of state-of-the-art real-time models, including Fast-SCNN and BiSeNetV2.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

MVitsU-net: A Real-time Semantic Segmentation Algorithm for Coal Flow in Mines

  • 期刊论文

  • 发表日期：

    2025年11月3日

查看全文
  • Fei Zhang,
  • Weimin Wu

摘要

The automatic segmentation of the coal flow area plays a crucial role in coal mine monitoring. Through real-time and accurate segmentation results, the belt speed can be dynamically adjusted to achieve energy savings and reduce consumption. However, a challenge persists for existing segmentation methods to simultaneously balance accuracy, speed, and robustness within the harsh mine environment. To address this challenge, a lightweight real-time semantic segmentation network, termed MVitsU-net, is proposed in this study. The computational efficiency of the model is significantly improved through an optimized MobileVitV2 encoder and a network pruning strategy. To mitigate the information loss induced by the lightweight design, a Multi-scale Dynamic Attention Fusion Module (MD-AM) is designed for the skip connections to preserve critical detail features. Finally, an Efficient Multi-scale Attention (EMA) mechanism is integrated within the decoder to enhance edge feature characterization and model generalization. Experimental results demonstrate that MVitsU-net achieves an mIoU of 95.62% and an inference speed of 74.41 FPS on a self-constructed dataset. Moreover, its balance between accuracy and speed is shown to be superior to that of state-of-the-art real-time models, including Fast-SCNN and BiSeNetV2.