SDS-YOLO11s: a lightweight detector for steel surface defects with enhanced feature interaction and content-aware upsampling

Due to manufacturing imperfections and environmental factors, surface defects frequently occur in steel materials, adversely affecting their reliability and service life. Therefore, accurate and efficient defect detection is critical for industrial quality control. We propose SDS-YOLO11s, a high-performance detection model based on YOLO11s. The backbone integrates a Depthwise Separable Space-to-Depth Convolution (DSSPDConv) module to strengthen spatial-channel feature interaction and enhance sensitivity to small defects. The standard C3K2 module has also been replaced with GhostC3K2_MSCA, incorporating multi-scale channel attention to strengthen feature representations. In the neck, the traditional upsampling operation is substituted with a dynamic upsampling module, DySample, to enhance resolution and localisation accuracy. Experimental results on the NEU-DET dataset show that SDS-YOLO11s attains a mAP50 of 80.6%, marking a 5.6% gain over the baseline while keeping computational cost almost unchanged. Moreover, it records a mAP50 of 72.6% on the GC10-DET dataset, further validating its robustness. These results indicate that SDS-YOLO11s delivers a dependable and high-performance solution for real-time, accurate surface defect inspection in industrial environments.