Embankment surface crack pixel-wise identification in UAV images based on a lightweight U-Network with transfer learning

Abstract

Embankments in challenging environments experience cyclic external influences and loads that render their surfaces susceptible to a range of defects, with cracks particularly menacing. Therefore, accurate and efficient crack identification holds paramount significance for embankment safety. This paper addresses the challenges of low crack recognition rates on complex embankment surfaces and the high training costs associated with deep learning-based crack detection methods. We introduce a novel residual structure as the foundational innovation, followed by developing a streamlined U-shaped network (denoted as U²-Net_Aggregation) that amalgamates multi-scale insights by integrating skip connections. The newly devised U²-Net_Aggregation network demonstrates significant improvements in accuracy, intersection over union (IOU), and F1-score metrics through ablation tests on three distinct crack datasets. Comparisons with commonly employed crack detection methods (FCN, SegNet, U-Net, DeepCrack, and U²-Net) were conducted at the pixel level using an embankment crack dataset. Results unequivocally showcase enhancements in both accuracy and processing speed. To overcome the challenge of a high application threshold resulting from a scarcity of training data specific to embankment cracks, we compile an extensive dataset featuring diverse building crack scenarios, serving as a foundation for transfer learning. Following transfer learning, U2-Net_Aggregation attains recall, IOU, and F1-score values of 87.31%, 81.76%, and 90.06%, respectively. The proposed methodology enables swift and accurate crack detection by directly utilizing Unmanned Aerial Vehicle (UAV) images, offering a novel avenue for embankment crack detection in real-world field scenarios.