SD-YOLO: A lightweight steel surface defect detection model with dynamic parameterisation for adaptive feature modulation

The production and manufacturing processes of steel inevitably generate various types of surface defects. The real-time and accurate detection of these surface defects is of great practical significance. To realise real-time detection of steel surface defects with significant differences in shape and size on resource constrained edge computing equipment, this paper proposes a lightweight real-time steel surface defect detection model SD-YOLO based on a dynamic parameterisation strategy. Firstly, a Dynamic Parameterised Enhancement Module is proposed, which dynamically assigns routing weights to parallel convolutional kernels based on input features, thereby enhancing the representation of defect features in the feature map and improving the network's ability to capture rich and detailed features. Secondly, the Efficient Intersection over Union loss function is employed to optimise the regression process of the prediction boxes. This enhances the model's fitting performance on bounding boxes with significant aspect ratio differences and improves the accuracy of detecting defects of various scales. Experimental results indicate that for the NEU-DET and GC10-DET datasets, SD-YOLO achieves a mean average precision of 83.1% and 74.1% respectively, with a stronger focus on defective regions, and detection speeds of 169.5 Frames Per Second (FPS) and 178.6 FPS, respectively. When SD-YOLO is deployed on the NVIDIA Jetson Orin NANO, the detection speed reaches 33.9 FPS and 66.7 FPS respectively, and maintains the same detection accuracy as the server-side, which realises real-time, accurate, and automatic detection of steel surface defects on edge computing devices with limited computational resources. Furthermore, SD-YOLO also demonstrates excellent generalisation ability and accuracy on images of steel surface defects collected in real industrial environments. In conclusion, SD-YOLO provides a practical and effective solution for real-time steel surface defect detection in resource-constrained environments, making it highly suitable for deployment in industrial applications. Source code is available at https://github.com/Xcy0512/SD-YOLO .