The rebar mesh inside the tunnel lining introduces significant interference in detecting defects, reducing their visibility in GPR images. This study proposes a global-to-local secondary recognition method based on an improved YOLOv8 model to address this challenge. Two datasets—global and local GPR images—were used, with an attention mechanism integrated into the YOLOv8 architecture. The improved YOLOv8 structure has been shown to increase the mean Average Precision (mAP) by 9.36 % and 3.86 % for the two datasets, respectively. Optimal performance was achieved with a rebar spacing of 0.4 m and a secondary recognition confidence of 0.867, while a rebar-defect distance of 1.20 m reached a confidence of 0.858. The model accurately identified the void defect shapes. Compared to traditional rebar signal suppression methods, this approach simplifies data processing, enhances accuracy, and reduces training costs. A tunnel field case further validated the method, boosting GPR image recognition confidence from 0.37 to 0.73, significantly improving the automated detection of tunnel lining defects.

中文翻译：

---

一种基于迁移学习和改进的 YOLOv8 的钢筋影响区域空隙检测新方法


隧道衬砌内的钢筋网在检测缺陷时引入了重大干扰，从而降低了它们在 GPR 图像中的可见性。本研究提出了一种基于改进的 YOLOv8 模型的全局到局部二级识别方法来应对这一挑战。使用了两个数据集——全局和局部 GPR 图像——并将注意力机制集成到 YOLOv8 架构中。改进的 YOLOv8 结构已被证明可以将两个数据集的平均精度 （mAP） 分别提高 9.36% 和 3.86%。钢筋间距为 0.4 m，二次识别置信度为 0.867，而钢筋缺陷距离为 1.20 m 时，置信度为 0.858，实现了最佳性能。该模型准确地识别了空隙缺陷的形状。与传统的钢筋信号抑制方法相比，这种方法简化了数据处理，提高了准确性，并降低了培训成本。隧道现场案例进一步验证了该方法，将 GPR 图像识别置信度从 0.37 提高到 0.73，显着提高了隧道衬砌缺陷的自动检测。