<p>Surface defect detection of strip steel is of great significance in industrial manufacturing, as its accuracy and efficiency directly affect product quality and production safety. To address the challenges of existing methods, including difficulties in small-object recognition, susceptibility to complex background interference, and limitations in embedded deployment, this paper proposes a lightweight detection model, RSE-YOLO, based on an improved YOLOv11 framework. The innovations of RSE-YOLO lie in three aspects. First, a C3k2-RVB module, which integrates structural re-parameterization with multi-scale attention, is introduced into the backbone to enhance contextual modeling and fine-grained feature extraction. Second, a lightweight SGFPN neck is designed, which combines Soft Nearest Interpolation (SNI) and efficient convolution (GSConvE) to improve both consistency and accuracy of cross-scale feature fusion. Third, an EfficientHead is proposed, which leverages partial convolution to substantially reduce redundant computation and optimize inference efficiency. Experimental results demonstrate that RSE-YOLO achieves 80.6% mAP on the NEU-DET dataset, surpassing the YOLOv11 baseline by 3.1%, with only 1.98M parameters. It attains 168.4 FPS on a PC and 60 FPS on the RK3588 embedded platform, meeting real-time detection requirements. Compared with other state-of-the-art models, RSE-YOLO achieves an excellent balance between detection accuracy, computational efficiency, and lightweight deployment, providing a feasible solution for industrial quality inspection scenarios that demand both real-time performance and high precision.</p>