Automatic Pixel-Level Pavement Crack Recognition Using a Deep Feature Aggregation Segmentation Network with a scSE...

Area of Science:

• Computer Vision
• Artificial Intelligence
• Civil Engineering

Background:

• Manual pavement crack detection is subjective and time-consuming.
• Automated systems are needed but face challenges with complex crack topology and noise.
• Existing deep learning methods struggle to balance accuracy and speed.

Purpose of the Study:

• To propose a novel deep learning algorithm for accurate and efficient pavement crack detection.
• To address the limitations of current deep learning models in terms of parameters and detection speed.
• To develop a robust crack detection system resilient to various noise interferences.

Main Methods:

• Developed CrackDFANet, a deep feature aggregation network incorporating a spatial-channel squeeze & excitation (scSE) attention mechanism.
• Created a crack dataset by segmenting collected images into 512 × 512 pixel blocks.
• Iteratively optimized the model using training and validation sets for robustness.

Main Results:

• CrackDFANet demonstrated strong anti-interference capabilities against light, markings, water, plants, oil, and shadows.
• The model achieved superior accuracy and faster detection speeds compared to state-of-the-art algorithms.
• Significantly reduced model parameters and error rates were observed.

Conclusions:

• CrackDFANet offers improved robustness and generalization ability for pavement crack detection.
• The proposed algorithm effectively balances detection accuracy and speed, outperforming existing methods.
• This deep learning approach provides a promising solution for automated pavement inspection and maintenance.