Scholarlink

<p>Road cracks are crucial indicators for evaluating pavement performance and ensuring traffic safety. While deep learning has become the primary approach for crack detection, existing high-precision models (e.g., DeepLabv3+) often rely on heavyweight backbones, leading to high computational costs that hinder efficient deployment in practical engineering scenarios. To bridge this gap, we present MDeepLab, an application-oriented lightweight redesign of semantic segmentation networks aimed at balancing structural complexity with predictive performance. Built upon the DeepLabv3+ framework, MDeepLab replaces the heavy Xception backbone with an optimized MobileNetV2, significantly reducing the model’s parameter count. To compensate for the potential information loss in the lightweight backbone, we integrate channel attention mechanisms with an Identity Mapping Initialization Strategy to enhance feature extraction and ensure stable convergence. Furthermore, a streamlined Atrous Spatial Pyramid Pooling (ASPP) module and a decoding-stage feature refinement module are developed to sharpen crack boundaries while minimizing computational redundancy. Consequently, the parameter count of the optimized MDeepLab is reduced from 54.77&#xa0;M in the original DeepLabv3+ to 3.92&#xa0;M, substantially improving computational efficiency and the feasibility of deployment. The proposed model has been comprehensively evaluated on three publicly available road crack datasets: Crack500, CrackForest, and GAPS384. Experimental results demonstrate that, compared with mainstream semantic segmentation models, MDeepLab significantly reduces the number of parameters while maintaining high segmentation accuracy, exhibiting promising engineering application value for automated road crack detection.</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

MDeepLab: a real-time lightweight network for accurate road crack segmentation

  • 期刊论文

  • 发表日期：

    2026年7月29日

查看全文
  • Guangling Sun,
  • Dongdong Wang,
  • Yanqiu Li,
  • Lei Yu,
  • Zuocai Wang

摘要

Road cracks are crucial indicators for evaluating pavement performance and ensuring traffic safety. While deep learning has become the primary approach for crack detection, existing high-precision models (e.g., DeepLabv3+) often rely on heavyweight backbones, leading to high computational costs that hinder efficient deployment in practical engineering scenarios. To bridge this gap, we present MDeepLab, an application-oriented lightweight redesign of semantic segmentation networks aimed at balancing structural complexity with predictive performance. Built upon the DeepLabv3+ framework, MDeepLab replaces the heavy Xception backbone with an optimized MobileNetV2, significantly reducing the model’s parameter count. To compensate for the potential information loss in the lightweight backbone, we integrate channel attention mechanisms with an Identity Mapping Initialization Strategy to enhance feature extraction and ensure stable convergence. Furthermore, a streamlined Atrous Spatial Pyramid Pooling (ASPP) module and a decoding-stage feature refinement module are developed to sharpen crack boundaries while minimizing computational redundancy. Consequently, the parameter count of the optimized MDeepLab is reduced from 54.77 M in the original DeepLabv3+ to 3.92 M, substantially improving computational efficiency and the feasibility of deployment. The proposed model has been comprehensively evaluated on three publicly available road crack datasets: Crack500, CrackForest, and GAPS384. Experimental results demonstrate that, compared with mainstream semantic segmentation models, MDeepLab significantly reduces the number of parameters while maintaining high segmentation accuracy, exhibiting promising engineering application value for automated road crack detection.