Integrated pixel-level crack detection and quantification using an ensemble of advanced U-Net architectures

Abstract

Automated pavement crack detection faces significant challenges due to the complex shapes of crack patterns, their similarity to non-crack textures, and varying environmental conditions such as lighting and noise. Traditional methods often struggle to adapt, leading to inconsistent and less accurate results in real-world scenarios. This study introduces a hybrid framework that combines convolutional and transformer-based architectures, leveraging their strengths to achieve reliable crack segmentation and pixel-level quantification. The framework incorporates state-of-the-art deep learning models, including U-Net, Attention U-Net, Residual Attention U-Net (RAUNet), TransUNet, and Swin-Unet. U-Net variants, enhanced with attention mechanisms and residual connections, improve feature extraction and gradient flow, enabling precise delineation of crack boundaries. Transformer-based models like TransUNet and Swin-Unet use self-attention mechanisms to capture both local and global spatial relationships, enhancing robustness across diverse crack patterns. A key contribution of this study is the evaluation of loss functions, including Binary Cross-Entropy (BCE) Loss, Dice Loss, and Binary Focal Loss. Binary Focal Loss proved particularly effective in addressing class imbalance across four benchmark datasets. To further improve segmentation performance, two ensemble strategies were applied: stochastic reordering using logical operations (AND, OR, and averaging) and a weighted average ensemble optimized through grid search. The weighted average ensemble demonstrated superior performance, achieving mean Intersection over Union (mIoU) scores of 0.73, 0.70, 0.78, and 0.86 on the CFD, AgileRN, Crack500, and DeepCrack datasets, respectively. In addition to segmentation, this study developed a method for accurately quantifying crack length and width. By using Euclidean distance along skeletal paths, the algorithm minimized error rates in length and width estimation. This framework provides a scalable and efficient solution for automated pavement crack analysis. It addresses critical challenges in accuracy, adaptability, and reliability under diverse operational conditions, marking significant progress in crack detection technology.

Keywords

• Crack segmentation
• Crack quantification
• Deep learning
• U-Net
• TransUNet
• Swin-UNet