NxV2net is a novel deep learning architecture for robust crack segmentation in complex real-world scenarios. It addresses the challenges of multiscale feature extraction, fusion, and generalization in the presence of lighting variations, water infiltration, and human interference.

Crack detection plays a crucial role in the safety assessment and maintenance of infrastructure. However, automatic crack segmentation remains challenging due to:

• Diverse crack patterns and complex topologies
• Variable imaging conditions (e.g., lighting, blur)
• Inadequate generalization of existing models
• Insufficiently diverse benchmark datasets

We introduce V2net, a nested multiscale segmentation network built upon cascaded VNet submodules, designed for improved generalization and feature representation.

• 🔗 V2net Architecture: Cascaded VNet-style modules that progressively enhance multiscale representation.
• 💡 Multichannel Fusion Attention (MCFA): A lightweight yet effective module for feature extraction and channel-wise fusion.
• 📦 SUES-CRACK Dataset: A real-world crack segmentation dataset featuring:
  • Lighting variation
  • Water-blurred boundaries
  • Human interference

Our model demonstrates superior performance in both quantitative and qualitative evaluations compared to existing crack segmentation methods.

• 📄 Paper: To be released upon publication
• 💻 Code & Dataset: https://github.com/nanxiang11/NxV2net

• Training instructions
• Inference demo
• Dataset format guide
• Model weights
• BibTeX citation

For any questions or collaborations, feel free to open an issue or contact the author via GitHub.