Defects on steel surfaces are critical to product quality and safety in industrial production. However, existing inspection models for industrial production lines often face challenges such as high computational resource consumption and suboptimal detection performance. To address these issues, this paper proposes an enhanced model FAP-YOLO, based on YOLOv8, to improve the recognition of complex defects. FAP-YOLO introduces a redesigned lightweight C2f module with the StarBlock (via Star Operation), reducing latency while maintaining efficiency. Building on this, we propose the lightweight fusion attention mechanism (LFAM). Feature extraction is significantly improved by integrating local, channel, and spatial attention mechanisms with optimized weighting strategies. Additionally, the PIoU v2 loss function is employed, integrating target-size adaptation with anchor-box quality optimization to accelerate model convergence. On the NEU-DET and GC10-DET datasets, FAP-YOLO achieves mAP scores of 80.6% and 68.3%, surpassing the baseline by 4.2% and 1.8%, respectively. These results validate FAP-YOLO’s effectiveness in steel surface defect detection.