A Multi-Stage Feature Aggregation and Structure Awareness Network for Concrete Bridge Crack Detection

Area of Science:

• Civil Engineering
• Computer Vision
• Artificial Intelligence

Background:

• Concrete bridge cracks pose significant safety risks.
• Existing detection methods struggle with crack characteristics like slenderness, low contrast, and background noise.
• Conventional convolutional methods lack the ability to capture long-range dependencies and effectively suppress background interference.

Purpose of the Study:

• To propose a novel network, MFSA-Net, for accurate pixel-level concrete bridge crack detection.
• To enhance the perception of linear crack structures and long-range dependencies.
• To improve the precision and robustness of crack detection systems.

Main Methods:

• Developed a multi-stage feature aggregation and structure awareness network (MFSA-Net).
• Introduced a structure-aware convolution block combining square and strip convolutions.
• Implemented a feature attention fusion block for edge sharpening and feature fusion.
• Aggregated features from different stages for fine-grained segmentation.

Main Results:

• MFSA-Net achieved average precision of 73.74%, recall of 77.04%, F1 score of 75.30%, and IoU of 60.48% on a concrete bridge crack dataset.
• Demonstrated superior performance compared to existing methods.
• Showcased adaptability and optimal performance on concrete pavement crack datasets.

Conclusions:

• MFSA-Net effectively addresses the challenges in concrete bridge crack detection.
• The proposed network architecture enhances the ability to perceive linear structures and long-range dependencies.
• MFSA-Net demonstrates significant potential for crack detection in diverse infrastructure scenarios.