A Comparative Study of U-Net-based Segmentation Architectures for Crack Detection

Automated crack detection is crucial for maintaining roads, buildings, and structures. Many models have been proposed, with U-Net being a common foundation. This study compares the performance of five architectures, including U-Net, and models derived from U-Net, namely, Unet-with-ResNet50, Residual Unet Crack Network (RUC-Net), CrackW-Net, and Mobile-Unet. In addition, the significance of the introduction of W-structure in CrackW-Net is also studied. We also introduce the W-structure in the encoder part of Mobile-Unet to check its effectiveness. Three datasets were adopted to train and test the models - CRACK500, CRACK500-C (a derived dataset of CRACK500), and our dataset formed by combining various established datasets. We observe that the basic U-Net shows the best performance, recording a test dice score of 0.6622, 0.6115, and 0.6196 on CRACK500, CRACK500-C, and our dataset, respectively. The W-structure with pre-trained encoders helped to improve the performance slightly, but it is overall observed to have no positive contribution to the improvement of the performance of Mobile-Unet. The performance ranking analysis showed that U-Net consistently outperforms all other models across all the datasets and performance measures followed by CrackW-Net-without-W. Further, based on the count of trainable parameters, U-Net ranks fourth while CrackW-Net-without-W has the fewest number of learning parameters. The comparison results highlight the robustness of U-Net and suggest exploring the performance of models on more large and diverse datasets.