Tunnel lining detection is an important element of quality management in tunnel construction and maintenance. Due to the variety of internal lining defects and unclear boundaries makes it challenging to identify these problems and train models effectively. Relying on manual detection or existing models, it is not possible to achieve fast and accurate defect detection. To address the above problems, A dataset consisted of 1 922 liner radar samples collected from Yunnan Tunnel B-scan was developed for training the model. A tunnel lining defect detection model YOLO-Tunnel based on YOLOv5 was proposed, which improved the model feature extraction ability, increased the receptive field, and improved the model localization ability by upgraded the Backbone and Neck. And further improved the model detection ability by selected the appropriate model size and balanced weight based on the dataset's scale and target size proportions. The results show that YOLO-Tunnel has better defect detection accuracy compared to YOLOv5s and also meets the real-time detection requirements, in which the precision, recall, and mAP are increased by 2.5, 9.0, and 8.1 percentage points, respectively, with the inference time increases by 2.7 ms to 21.8 ms. The research results provide a reference for further improving the performance of the detection of tunnel lining detection and the direction of optimization of the model reference.