Lightweight tunnel surface defect detection algorithm based on MDS-YOLO

To address the issues of severe interference in complex environments and the difficulty in accurately extracting and efficiently identifying multi-scale defect features in tunnel surface defect detection, a lightweight tunnel surface defect detection algorithm named MDS-YOLO was proposed based on an improved YOLOv8 model. The algorithm was built upon the YOLOv8n model. A multi-scale feature fusion (C2f_MSFA) module was designed in the backbone network to replace the original C2f feature extraction module. By using partial channel convolution and multi-scale feature fusion methods, the feature maps were effectively extracted and aggregated from different levels, enhancing the model's perception and representation ability for defect targets with significant size differences. A dynamic upsampling module (DySample) was introduced in the neck network to replace traditional upsampling methods. The module adaptively learned sampling parameters based on the input feature content, enhancing the feature restoration ability and spatial information preservation during upsampling and improving the accuracy and efficiency of feature fusion. A shared convolution detection head (SC_Detection) was constructed. By using shared convolution and group normalization strategies, the model's detection efficiency and stability were improved while reducing the number of parameters and computational complexity. Experimental results show that the MDS-YOLO model achieves accuracy improvements of 2.2%, 3.4%, and 4.4% in detecting three types of tunnel surface defects, including water leakage, cracks, and lining spalling, respectively. The average detection accuracy reaches 74.2%, which is 3.4% higher than that of the baseline model YOLOv8n. The number of model parameters is reduced from 3.00×10⁶ to 1.97×10⁶, with a decrease of 34.3%. The amount of computation is reduced from 8.1×10⁹ to 5.6×10⁹, with a reduction of 30.9%. The model size is compressed from 5.96 MB to 4.00 MB. The algorithm achieves a lightweight model while ensuring detection accuracy, meeting the application requirements of high accuracy and low computational resources in practical scenarios such as tunnel inspection and edge computing.