In response to the initial YOLOv5 target detection algorithm's problems of inadequate extraction of defect features in complex targets, imprecise localization, low detection accuracy and high miss detection rate, an improved YOLOv5 steel surface defect detection algorithm is proposed, in which the anchor frame is calculated with the K-Means++ algorithm for anchor frame selection, so that the randomly selected clustering center tends to the global optimal solution as much as possible and the prediction frame is more accurate. The CBAM attention mechanism is added to give higher weights to the defects in complex images and enhance the attention to key information. The results after the experimental comparison show that the improved YOLOv5 algorithm has better detection performance.