Tunnel Lining Crack Detection Algorithm based on Improved YOLOv5s

To address the issues of false detection, poor detection performance, and leakage when applying the YOLOv5s model to detect tunnel lining crack images in complex backgrounds. We have improved the detection algorithm: the dynamic snake convolution is used in the backbone network instead of the standard convolution in the original C3 module so that the convolution kernel can fit the crack’s features in the image for learning. Add a large selective kernel module to the neck network to capture more crack features and localized information by expanding the receptive field. The experimental results indicate improvements of 3.63% in detection precision, 2.78% in recall, 3.27% in F1 value, 3.94% in mAP@50, 36.8% in mAP@75, and 21.3% in mAP@50-95, respectively, and the detection speed is up to 40 frames per second, which provides a new solution for detecting cracks in tunnel linings.