Scholarlink

<p>Pavement crack detection is a critical task in road maintenance, yet traditional methods rely heavily on costly pixel-level annotations, which severely limit their scalability in real-world applications. To address this issue, we propose HelixNet (Intertwined Transformers), a novel weakly supervised crack segmentation framework that requires only image-level labels while achieving high segmentation accuracy. Built upon a Transformer backbone, HelixNet integrates three key modules: an Intertwined Feature Transformer (IFT) for multi-scale contextual representation, a Pixel Correlation Module (PCM) for structure-aware pseudo-label refinement, and an Attention-based Feature Affinity module (AFA) for enforcing global spatial consistency. Specifically, IFT captures long-range dependencies suitable for modeling elongated crack structures; PCM leverages semantic similarity to propagate activation and fill structural gaps in initial class activation maps; AFA constructs a global affinity matrix via multi-head self-attention to preserve region-level consistency. Experiments on the Crack500 and DeepCrack datasets demonstrate that HelixNet significantly outperforms existing weakly supervised methods, achieving mIoU gains of 7.7% and 4.7%, respectively. Despite using only image-level supervision, our method approaches the performance of fully supervised models, offering a practical and scalable solution for large-scale crack detection systems.</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

Helixnet: intertwined transformers for weakly supervised crack segmentation via pixel-affinity fusion

  • 期刊论文

  • 发表日期：

    2026年6月23日

查看全文
  • Yang Li,
  • Xinhang Li,
  • Ke Shen,
  • Dong Sui,
  • Yacong Li,
  • Maozu Guo

摘要

Pavement crack detection is a critical task in road maintenance, yet traditional methods rely heavily on costly pixel-level annotations, which severely limit their scalability in real-world applications. To address this issue, we propose HelixNet (Intertwined Transformers), a novel weakly supervised crack segmentation framework that requires only image-level labels while achieving high segmentation accuracy. Built upon a Transformer backbone, HelixNet integrates three key modules: an Intertwined Feature Transformer (IFT) for multi-scale contextual representation, a Pixel Correlation Module (PCM) for structure-aware pseudo-label refinement, and an Attention-based Feature Affinity module (AFA) for enforcing global spatial consistency. Specifically, IFT captures long-range dependencies suitable for modeling elongated crack structures; PCM leverages semantic similarity to propagate activation and fill structural gaps in initial class activation maps; AFA constructs a global affinity matrix via multi-head self-attention to preserve region-level consistency. Experiments on the Crack500 and DeepCrack datasets demonstrate that HelixNet significantly outperforms existing weakly supervised methods, achieving mIoU gains of 7.7% and 4.7%, respectively. Despite using only image-level supervision, our method approaches the performance of fully supervised models, offering a practical and scalable solution for large-scale crack detection systems.