Scholarlink

<p>The proper detection of fractures is critical to maintain structural integrity and operational safety in mining operations. Conventional crack detection systems, which rely on manual inspections or simple image processing algorithms, have limited scalability and accuracy. Although deep learning models, such as CNN, have showed promise in semantic segmentation tasks, they still fail to recognize fractures, understand long-range correlations between visual components, and detect minute crack details. To address these challenges, we developed a new architecture, SEFC_Net, that enhances DeeplabV3 + by integrating Squeeze-and-Excitation (S.E) Blocks with channel-wise attention methods, a Feature Fusion Module (FFM), and Channel Prior Convolutional Attention (CP-CA). The most important channels are dynamically prioritized using Channel Prior Convolutional Attention, low-level and high-level features are blended using Feature Fusion, and feature channels are recalibrated using S.E. Blocks with attention. Mine Cracks Dataset have been used to test the model’s performance. Our model outperformed DeeplabV3 + , according to the findings, with 87.42% of mIou, 89.93% of Precision, 95.15% of Recall, and 92.36% of F1-score. The outcomes also demonstrate how well the model handles noisy and missing data and how quickly it converges during training. Our model has real-world applications as it allows early fracture detection which can increase safety and prevent buildings from collapsing.</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

SEFC-Net: enhanced crack segmentation using attention mechanisms and channel prior convolutional attention in mining area

  • 期刊论文

  • 发表日期：

    2025年8月21日

查看全文
  • Tossou Claire Ingrid,
  • Geng-Kun Wu,
  • Mingchen Wei,
  • Letian Wang

摘要

The proper detection of fractures is critical to maintain structural integrity and operational safety in mining operations. Conventional crack detection systems, which rely on manual inspections or simple image processing algorithms, have limited scalability and accuracy. Although deep learning models, such as CNN, have showed promise in semantic segmentation tasks, they still fail to recognize fractures, understand long-range correlations between visual components, and detect minute crack details. To address these challenges, we developed a new architecture, SEFC_Net, that enhances DeeplabV3 + by integrating Squeeze-and-Excitation (S.E) Blocks with channel-wise attention methods, a Feature Fusion Module (FFM), and Channel Prior Convolutional Attention (CP-CA). The most important channels are dynamically prioritized using Channel Prior Convolutional Attention, low-level and high-level features are blended using Feature Fusion, and feature channels are recalibrated using S.E. Blocks with attention. Mine Cracks Dataset have been used to test the model’s performance. Our model outperformed DeeplabV3 + , according to the findings, with 87.42% of mIou, 89.93% of Precision, 95.15% of Recall, and 92.36% of F1-score. The outcomes also demonstrate how well the model handles noisy and missing data and how quickly it converges during training. Our model has real-world applications as it allows early fracture detection which can increase safety and prevent buildings from collapsing.