LTPLN: Automatic pavement distress detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Civil Engineering

Background:

• Traditional pavement disease detection methods lack efficiency and generalization.
• Deep learning methods face challenges with high-resolution images and diverse disease types.

Purpose of the Study:

• To propose a novel lightweight Transformer Patch Labeling Network (LTPLN) for enhanced automatic pavement disease detection.
• To improve efficiency and overcome limitations of existing methods in complex scenarios.

Main Methods:

• Histogram equalization preprocessing for image quality enhancement.
• Patch partitioning and input to an enhanced Transformer model with feature map labels.
• Integration of depthwise separable convolution and expectation maximization-based label distillation for weak supervision.

Main Results:

• Reduced computational complexity by 2.5G Flops and improved speed by 16% on a private dataset.
• Achieved a minor 1.2% decrease in AUC accuracy compared to the baseline model.
• Demonstrated balanced performance with improved accuracy and speed on a public dataset.

Conclusions:

• The LTPLN model offers significant performance advantages for automatic pavement disease detection.
• The proposed method is efficiently applicable in real-world pavement inspection scenarios.
• The model balances computational efficiency with detection accuracy.