A Pavement Crack Translator for Data Augmentation and Pixel-Level Detection Based on Weakly Supervised Learning

Recent state-of-the-art pavement crack detection algorithms are data-driven and domain-sensitive due to their heavy reliance on datasets. Establishing a high-quality pavement crack dataset with various scenarios could help improve the robustness and generality of detection models. Therefore, this study presented the crack translator structure to achieve image style transfer, which can generate pavement crack images with different backgrounds for data augmentation. To get a better translation performance, we proposed two crack translators for comparison. The crack translator v2 was designed with attention blocks whereas crack translator v1 did not. The proposed crack translator v1 and v2 served as the weakly supervised method that was used to achieve pixel-level pavement crack detection and image transfer. Additionally, the crack translators were used to establish the dataset with various crack images and corresponding labels. Moreover, U-Net, SegNet, and W-segnet were compared to evaluate the effectiveness of data augmentation by using crack translators. Two open-source datasets were used to demonstrate the generalization of the proposed models. Results demonstrate that the proposed crack translator v2 outperforms crack translator v1 and other state-of-the-art methods in terms of evaluation metrics for image translation from normal background to shadow, wet, night, and grooved backgrounds. Crack translator (v2)-based augmentation contributes to a greater improvement than conventional augmentation for crack segmentation where F1 has increased by 5% for U-Net, SegNet, and 7% for W-segnet. The crack translator v2 yields equivalent results compared with U-Net on holdout EdmCrack600 and CRKWH100 datasets. Therefore, the proposed crack translator v2 can be effectively used to augment pavement crack images and it shows great potential for pixel-level pavement crack detection. Proposed crack translator v2 provides an integrated model for both data augmentation and pixel-level segmentation.