Deep Learning-based Surface Crack Detection

Abstract

Surface crack detection is paramount for ensuring the safety and longevity of civil infrastructure. Traditional manual inspection methods are time-consuming, costly, and operator-dependent. Prior automated methods based on traditional image processing showed promise, though a step forward, but their reliance on

handcrafted features and sensitivity to imaging conditions and real-world vari- ations in lighting and texture limited their global applicability. Deep learning

is a more powerful tool, but achieving reliable pixel-level remains a significant challenge separation of cracks especially fine, hairline cracks, usually only a few pixels thick with low contrast and easily discarded in the downsampling stages of standard encoder-decoder frameworks; and intricate, or lengthy crack networks. The current paper proposes a better deep learning technique for this task using the UNet++ architecture. This model was chosen specifically over others due to its design strategy. Although standard U-Net innovated vital skip connections to maintain detail, UNet++ enhances that by utilizing nested and dense skip pathways to bridge the semantic gap between shallow encoder and deep decoder. This architecture is especially well-suited to marrying high-resolution spatial data with deep semantic context, which is necessary for being capable of segmenting the fine-grained crack structure accurately. Compared to architectures such as SegNet, which are efficient in terms of computations but lose boundary detail, UNet++ is high-fidelity segmentation optimized and thus is particularly well-suited

to this task. The proposed UNet++ model achieved excellent quantitative re- sults: a Dice coefficient of 0.9338, an Intersection over Union (IoU) of 0.8805,

an F1-score of 0.9609, precision of 0.9603, and recall of 0.9614, outper forming a baseline U-Net and other contemporary semantic segmentation methods. A prototype system demonstrated the model’s applicability for real-world crack de tection tasks. These findings indicate that the UNet++ architecture, coupled with robust data augmentation, offers a highly accurate and reliable solution for pixel level surface crack segmentation, advancing automated structural health monitoring.