Local-Global Feature Adaptive Fusion Network for Building Crack Detection

Area of Science:

• Civil Engineering
• Computer Vision
• Artificial Intelligence

Background:

• Cracks are a common structural damage, necessitating timely detection for building safety.
• Effective crack detection requires integrating both local details and global context.
• Existing methods may struggle with varying crack complexities and background similarities.

Purpose of the Study:

• To propose a novel network for accurate crack detection in building structures.
• To develop a method that effectively fuses local and global features for enhanced performance.
• To introduce a new dataset for evaluating crack detection algorithms.

Main Methods:

• A local-global feature adaptive fusion network (LGFAF-Net) was developed.
• The network utilizes a VMamba encoder for global feature extraction and a residual network for local feature extraction.
• A multi-feature adaptive fusion (MFAF) module integrates features from both branches.

Main Results:

• The proposed LGFAF-Net demonstrated superior performance in crack detection.
• Experiments on the new Building Exterior Wall Crack (BEWC) dataset and public datasets confirmed the method's effectiveness.
• The dual-encoding network and adaptive fusion module significantly improved detection accuracy.

Conclusions:

• The LGFAF-Net effectively integrates local and global information for robust crack detection.
• The proposed method offers a significant advancement in automated structural health monitoring.
• The BEWC dataset provides a valuable resource for future research in crack detection.