Automatic Pavement Crack Detection Transformer Based on Convolutional and Sequential Feature Fusion.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Civil Engineering

Background:

• Pavement crack detection is crucial for infrastructure maintenance.
• Natural environmental factors often interfere with traditional detection methods, reducing accuracy.
• Existing deep learning models can be complex and computationally intensive.

Purpose of the Study:

• To develop a lightweight and efficient pavement crack detection framework.
• To improve the accuracy of pavement crack detection, especially under environmental interference.
• To reduce the implementation complexity of pavement crack detection systems.

Main Methods:

• Designed the Pavement Crack DEtection TRansformer (PCDETR) network.
• Fused convolutional features with sequence features using parallel Swin-Transformer and residual network backbones.
• Optimized the transformer's encoder-decoder for direct set prediction of crack location and category.

Main Results:

• Achieved an Average Precision (AP) of 45.8% on the COCO dataset, outperforming DETR (36.9%) and Conditional DETR (42.8%).
• Reached an AP of 45.6% on a custom pavement crack dataset, surpassing Mask R-CNN by 3.8% and Faster R-CNN by 8.8%.
• Demonstrated a low-code implementation, reducing complexity.

Conclusions:

• PCDETR offers an efficient and accurate solution for pavement crack detection.
• The fusion of convolutional and sequential features enhances detection performance.
• The proposed method effectively addresses challenges posed by environmental interference.