Real-time detection of highway tunnel lining cracks using YOLOv8-DTD with an android application

The heavy, memory-intensive nature of existing detection models limits their applicability for efficient crack recognition on mobile and embedded devices with constrained resources. To address this issue, this paper proposes YOLOv8-DTD, a real-time detection model for identifying tunnel lining cracks that integrates Deformable Convolutional Network v2 (DCNv2) and a Transformer Decoder. DCNv2 enhances precise and swift detection of crack deformation features, while the Transformer Decoder optimises the end-to-end process and eliminates computational costs associated with anchor-free methods. The model subsequently was deployed in an Android application for automatic real-time crack detection on smartphones. Results show that YOLOv8-DTD achieves 10.84 % and 9.31 % improvements in mAP and F1 score, respectively, while reducing parameters by 43.21 % and reaching 65.46 frames per second, evaluated on a dataset comprising lining cracks from 141 highway tunnels in Shaanxi Province, China. Moreover, detection efficiency is further validated via Jetson Nano acceleration and field feasibility testing.