RUC-Net: A Residual-Unet-Based Convolutional Neural Network for Pixel-Level Pavement Crack Segmentation

Area of Science:

• Computer Vision
• Artificial Intelligence
• Materials Science

Background:

• Automatic crack detection is crucial for infrastructure maintenance.
• Challenges include complex backgrounds, poor illumination, and noise.
• Existing methods struggle with pixel-level pavement crack segmentation.

Purpose of the Study:

• To develop an effective deep learning model for pixel-level pavement crack segmentation.
• To improve the accuracy and robustness of automatic crack detection systems.
• To address the class imbalance issue in pavement crack datasets.

Main Methods:

• Proposed RUC-Net, a U-shaped encoder-decoder semantic segmentation network combining Unet and Resnet.
• Integrated spatial-channel squeeze and excitation (scSE) attention module.
• Utilized focal loss function to handle class imbalance.

Main Results:

• RUC-Net achieved superior performance compared to FCN, Unet, and SegNet on CFD, Crack500, and DeepCrack datasets.
• Ablation studies demonstrated the effectiveness of scSE modules in enhancing crack detection.
• The proposed method shows significant improvements in pixel-level pavement crack segmentation.

Conclusions:

• RUC-Net offers a robust and accurate solution for pavement crack segmentation.
• The integration of scSE attention and focal loss effectively tackles detection challenges.
• This work advances the field of automated infrastructure inspection and maintenance.