Improvement of crack detectivity for concrete surface of subway tunnels with anti-corrosion coatings using deep learning and image processing

Abstract

The long-term survivability of subway tunnels heavily depends on the durability and stability of concrete structures. Cracks in concrete, caused by factors such as severe loading, environmental influences, and chemical effects etc., lead to a reduction in structural durability and may even result in a loss of stability. In this study, crack detection is achieved through deep learning and image processing. We design a novel crack locally ordered annotation method. Training the object detection model using the proposed annotation method can achieve more accurate crack localization. Subsequently, based on the proposed annotation method, we improve the You Only Look Once version 8 nano (YOLOv8n) model by incorporating Focal Efficient Intersection over Union (FEIoU) and a path aggregation feature pyramid network with dynamic snake convolution (PADFPN), resulting in a YOLOv8n model combined with FEIoU and PADFPN (YOLOv8n-FED). This model effectively integrates multi-scale information of cracks. Finally, we extract the detected crack regions and segment them using a region-growing algorithm. In terms of object detection, based on the proposed annotation method, YOLOv8n-FED, compared with the original model, achieves a detection precision of 95.0 %, an improvement of 3.7 %; and a mean Average Precision (mAP) 50–95 of 80.7 %, a gain of 6.2 %. For semantic segmentation, our method yielded satisfactory results without requiring laborious pixel-level annotations, achieving a precision and F1-score of 75.2 % and 80.9 %, respectively, both outperforming the comparison models. Moreover, it can capture finer crack edge features.