To address the challenge of balancing efficiency and accuracy in metal surface defect detection under industrial computing resource constraints, where traditional models struggle with real-time performance and excessive parameters on embedded devices, this study proposes a lightweight MSL-YOLO detection model. The model designs a MiniStar (Miniature Star) backbone network to reduce structural complexity, introduces a Lightweight Adaptive-Weighted DownSampling (LADS) method to balance computational efficiency and feature retention during resolution compression, and constructs a Lightweight Shared-Detail-Enhanced Convolutional Detection Head (LSDECD_Detect) to minimize parameter size and computational burden. Experimental results on the GC10-DET and NEU-DET datasets demonstrate that the model achieves a 25.4% improvement in detection speed compared to the baseline YOLOv8, with mean average precision (mAP) reaching 69.0% and 77.1%, respectively. The model has been successfully deployed on low-power devices such as Raspberry Pi, enabling accurate identification and localization of typical metal surface defect.