Crack-MsCGA: A Deep Learning Network with Multi-Scale Attention for Pavement Crack Detection.

Area of Science:

• Civil Engineering
• Computer Science
• Artificial Intelligence

Background:

• Pavement crack detection is vital for road safety and maintenance.
• Current convolutional neural network (CNN) methods struggle with small cracks due to low-level feature fusion interference.
• Existing attention mechanisms enhance global features but don't fully address local detail loss.

Purpose of the Study:

• To develop a computationally efficient deep learning network for accurate multi-scale pavement crack detection.
• To improve the detection of small-scale cracks with subtle local structures and varying global morphologies.
• To reduce noise interference by avoiding low-level feature fusion.

Main Methods:

• Proposed a novel deep learning network, Crack-MsCGA, utilizing CNNs and multi-scale attention.
• Introduced a multi-scale attention mechanism (MsCGA) focusing on high-level features for local and global information.
• Implemented local window attention for short-range dependencies and cascaded group attention for long-range dependencies, fused via Mixed Local Channel Attention (MLCA).

Main Results:

• Crack-MsCGA achieved improved detection accuracy across various scales on the DH807 dataset.
• Demonstrated significant improvements in AP@50: 11.3% for small-scale, 8.1% for medium-scale, and 5.9% for large-scale cracks.
• Outperformed five existing state-of-the-art methods in pavement crack detection.

Conclusions:

• The proposed Crack-MsCGA network effectively addresses the limitations of existing methods in multi-scale pavement crack detection.
• Avoiding low-level feature fusion and employing a novel multi-scale attention mechanism enhances detection accuracy, particularly for small cracks.
• The method offers a promising solution for improving road safety and optimizing maintenance through precise crack identification.