SteelYOLO11 for strip steel defect detection via multi-scale sampling optimization

Abstract Aiming at the low efficiency of manual quality inspection for steel strips and the inability of existing lightweight detection algorithms to balance the recognition accuracy of tiny defects and anti-interference capability under complex industrial environments, this paper proposes SteelYOLO11, a steel strip surface defect detection model optimized based on multi-scale sampling, with YOLO11n as the baseline network. The overall sampling pipeline is collaboratively improved by retaining high-resolution feature layers, replacing bilinear upsampling, adding 1 × 1 convolution-based feature alignment modules, and optimizing the loss function. Experimental results on two mainstream steel strip defect datasets, NEU-DET and GC10-DET, demonstrate that compared with the baseline YOLO11n, the proposed model achieves mAP50 improvements of 1.5% and 3.8% on the two datasets respectively. In contrast with large-scale models such as Mamba-YOLO and RT-DETR-R50, SteelYOLO11 presents fewer parameters and higher inference speed. Under various single and combined industrial interference scenarios, the average performance degradation rate of the model is reduced to 12.99%, which is 3.74% lower than that of the baseline model. The proposed model simultaneously realizes lightweight design and strong anti-interference performance, and can provide an effective reference for intelligent surface quality inspection of various metal materials.