Automatic detection of tunnel lining crack based on mobile image acquisition system and deep learning ensemble model

Tunnel cracks are a crucial indicator of tunnel detection and performance evaluation. However, traditional manual inspection methods are time-consuming and dangerous. To address these problems, an automatic tunnel crack detection method based on a mobile image acquisition system and deep learning ensemble model is proposed. A novel mobile image acquisition system is proposed for tunnel data acquisition. A deep learningbased model, named You Only Look Once v8 enhanced by large separable kernel attention (LSKA) and dynamic snake convolution (DSC; YOLO-LD), is proposed to improve the crack detection performance. Collaborative learning is used to combine the YOLO-LD object detection and semantic segmentation models into an ensemble model to enhance the model's engineering adaptability. Edge computing technologies are used for ensemble model deployment and inference acceleration. The method is tested on the custom tunnel lining crack (TL-Crack), the open-access dataset LinkCrack, and highway tunnel field data. The results show that the mobile image acquisition system can rapidly acquire high-resolution images and form panoramic images. The YOLO-LD model outperforms other state-of-the-art models in terms of precision, recall, and F1-score on both TL-Crack and LinkCrack. The ensemble model fully exploits the YOLO-LD object detection model's crack localization capability and the YOLO-LD semantic segmentation model's crack extraction performance, improving the model's engineering adaptability. Edge computing techniques increase the inference speed of the ensemble model to 84 images/second. Parameters such as stake number, distribution, length, width, and type of cracks are calculated, and the crack distribution maps are prepared to assist inspectors in field verification.