CrackAdaptNet: End-to-end domain adaptation for crack detection and quantification

Abstract

Crack detection and quantification are essential for ensuring the maintenance and safety of infrastructure, as cracks serve as early indicators of structural degradation. This paper introduces CrackAdaptNet, an end-to-end domain adaptation and semantic segmentation framework designed to address the complexities of crack detection in diverse engineering environments. Unlike existing approaches that face challenges in generalizing to practical applications, CrackAdaptNet utilizes extensive annotated vertical datasets to enhance detection accuracy in oblique imagery from uncontrolled settings. The framework comprises three core components: the Alignment Generator (AG), Segmentation Generator (SG), and Alignment-Segmentation Discriminator (ASD). AG mitigates the domain gap between controlled datasets and field-collected imagery. SG performs precise crack segmentation using Generative Adversarial Networks, while ASD assesses both alignment and segmentation quality. Experimental results indicate that CrackAdaptNet surpasses state-of-the-art models such as Mask2Former, K-Net, and SegFormer, achieving notable improvements in segmentation performance. Specifically, CrackAdaptNet achieves an IoU improvement of over 36% and an F1-score increase of more than 40% compared to these methods, demonstrating its superior generalization capability. Furthermore, field experiments conducted on Qingshuiting West Road in Nanjing, China, reveal strong correlations between model predictions and manual measurements. These results demonstrate the framework’s ability to minimize false positives and enhance the reliability of crack detection across diverse environments. CrackAdaptNet provides a robust solution for the monitoring and assessment of infrastructure conditions.