Aiming at the issues of outdated existing models for steel surface defect detection, limited accuracy, and the existence of misjudgment and omission, an improved YOLOv10 method for steel surface defect detection, named FAA-YOLO, is proposed. This method introduces the lightweight network FasterNet and multi-scale attention mechanism EMA, designing the C2f_Faster_EMA module to balance the lightweight of the backbone network with enhanced feature extraction capabilities. An adaptive fine-grained channel attention mechanism AFGC is added at the end of the backbone network to enhance the preliminary feature extraction ability of the model′s backbone network, thereby improving the model′s detection accuracy. The Neck part is replaced with an attention scale sequence fusion framework ASF to enhance the model′s ability to integrate multi-scale feature information. Comparative experiments and ablation experiments on the NEU-DET steel surface defect dataset show that the proposed FAA-YOLO model reduces the number of parameters by 11.01%, the computational load by 7.69%, and increases the detection accuracy by 2.9 percentage points, achieving a detection accuracy of 83.6%. This method reduces the complexity of the model while achieving high detection accuracy, demonstrating high usability and real-time performance in industrial systems.