An Improved YOLOv11 Algorithm for Metal Surface Defect Detection

During the manufacturing of metal materials, various complex surface defects are inevitably generated, and complex background textures easily cause defect misclassification and missed detection. To boost the detection robustness and real‐time performance of lightweight detection networks, this paper designs a metal surface defect detection model named YOLO‐CTU based on YOLOv11n. Three targeted structural optimizations are proposed: a channel‐enhanced ADown (CED) module to mitigate fine‐grained feature loss during downsampling, a C3k2_UIB unit to strengthen extraction of tiny defect textures, and Triplet Attention embedded on the newly added P2 shallow feature fusion branch to excavate low‐level small defect features and suppress background noise. Experiments are carried out on GC10‐DET and NEU‐DET datasets. Compared with the baseline YOLOv11n, the proposed YOLO‐CTU achieves a 2.8% mAP@0.5 improvement on GC10‐DET and maintains competitive detection accuracy on NEU‐DET with fewer parameters and lower computational overhead, which satisfies the real‐time inspection demands of industrial production lines.