Automatic quality assessment of highway tunnel linings: a deep learning framework enhanced by data augmentation techniques

Abstract Ground-penetrating radar is widely used for the non-invasive inspection of highway tunnel linings, yet manual interpretation of radar images is often inefficient and error-prone, as it relies heavily on expert knowledge and prior experience. Although artificial intelligence has been explored for lining quality assessment, developing reliable defect detection models remains challenging due to the limited availability of defect samples and imbalanced datasets. To address these issues, this study proposes YOLOv8-LQD, an enhanced detection model that integrates the YOLOv8 framework with deep convolutional generative adversarial networks to generate synthetic images, thereby augmenting data and balancing the training set. Experimental results demonstrate that YOLOv8-LQD improves detection accuracy by 7.4% compared to the original YOLOv8 model and significantly reduces data collection costs. Moreover, the framework outperforms existing approaches in identifying tunnel lining defects while maintaining low inference latency. These advancements highlight YOLOv8-LQD as a scalable and cost-effective solution for automated tunnel inspection, with broad potential for enhancing infrastructure safety and maintenance.