TAD-YOLO11n: A Lightweight Network with Multiscale Feature Enhancement for Steel Surface Defect Inspection

Reliable and fast detection of steel surface defects is essential for quality control in intelligent manufacturing. In practical inspection scenarios, lightweight detectors still face challenges associated with low-contrast defect textures, the loss of fine edge information during downsampling, and insufficient interaction among features at different scales. To improve detection accuracy and computational efficiency, this study proposes TAD-YOLO11n, a lightweight multiscale feature-enhancement detector based on YOLO11n. In the backbone, C3k2-TFE-EMA is introduced to enhance the representation of weak textures and elongated defects by combining local texture enhancement, frequency-domain modeling, and EMA attention. ED-ADown is employed during downsampling to preserve edge details while reducing the parameter count and computational cost. In the neck, DynamicScalSeq+ASF is incorporated to promote adaptive interaction among P3, P4, and P5 features. Experiments on the NEU-DET dataset showed that TAD-YOLO11n achieved an mAP50 of 79.0%, exceeding the YOLO11n baseline by 4.8 percentage points. Meanwhile, the number of parameters decreased from 2.583 M to 2.106 M, and the computational cost decreased from 6.3 to 5.3 GFLOPs, while the inference speed reached 175.5 FPS. These results demonstrate that the proposed model improves detection performance under lightweight constraints and provides a practical solution for real-time steel surface defect inspection.