Scholarlink

<p>Pixel-level crack segmentation is vital for infrastructure health monitoring, yet achieving a balance between high accuracy and computational efficiency remains challenging, especially under complex backgrounds and on resource-constrained edge devices. Existing methods, including Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs), often struggle with either limited receptive fields or prohibitive computational costs. To address this, we introduce CrissMamba, a lightweight network designed specifically for efficient and precise crack segmentation. The model features a novel Cross-Dimensional Visual State Space (CDVSS) block within its encoder. This block incorporates a Parallel Visual Mamba (PVM) module with a Cyclic Shift mechanism to enable efficient cross-channel information exchange. Furthermore, it employs an enhanced multi-directional scanning strategy tailored to the linear and branching nature of cracks, significantly improving global feature capture. For the decoder, we propose a Lightweight Multi-layer Perceptron with a Dynamic upSampler (LMD), which uses learnable upsampling to achieve precise multi-scale feature alignment and sharp crack boundaries. Extensive evaluations on four benchmark datasets (TUT, DeepCrack, Crack500, CrackMap) demonstrate state-of-the-art performance. Notably, on the complex TUT dataset, CrissMamba achieves an mIoU of 84.79% and an F1-score of 83.65% while maintaining an exceptionally low parameter count of only 1.28M and a real-time inference speed of 37FPS at a <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(512 \times 512\)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>512</mn> <mo>×</mo> <mn>512</mn> </mrow> </math></EquationSource> </InlineEquation> resolution. This work provides a practical and high-performance solution for deploying automated crack detection systems in real-world, resource-limited engineering scenarios.</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

Crissmamba: a lightweight vision Mamba framework with cross-dimensional interaction for structural crack segmentation

  • 期刊论文

  • 发表日期：

    2026年6月23日

查看全文
  • Yilei Shu,
  • Peifang Xin,
  • Caixia Wang,
  • Junwei Guang,
  • Qing Zhang,
  • Ming Li,
  • Meizhong Liu,
  • Longwang Yue

摘要

Pixel-level crack segmentation is vital for infrastructure health monitoring, yet achieving a balance between high accuracy and computational efficiency remains challenging, especially under complex backgrounds and on resource-constrained edge devices. Existing methods, including Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs), often struggle with either limited receptive fields or prohibitive computational costs. To address this, we introduce CrissMamba, a lightweight network designed specifically for efficient and precise crack segmentation. The model features a novel Cross-Dimensional Visual State Space (CDVSS) block within its encoder. This block incorporates a Parallel Visual Mamba (PVM) module with a Cyclic Shift mechanism to enable efficient cross-channel information exchange. Furthermore, it employs an enhanced multi-directional scanning strategy tailored to the linear and branching nature of cracks, significantly improving global feature capture. For the decoder, we propose a Lightweight Multi-layer Perceptron with a Dynamic upSampler (LMD), which uses learnable upsampling to achieve precise multi-scale feature alignment and sharp crack boundaries. Extensive evaluations on four benchmark datasets (TUT, DeepCrack, Crack500, CrackMap) demonstrate state-of-the-art performance. Notably, on the complex TUT dataset, CrissMamba achieves an mIoU of 84.79% and an F1-score of 83.65% while maintaining an exceptionally low parameter count of only 1.28M and a real-time inference speed of 37FPS at a \(512 \times 512\) 512 × 512 resolution. This work provides a practical and high-performance solution for deploying automated crack detection systems in real-world, resource-limited engineering scenarios.