Aiming at the problems of many types of defects on the surface of steel plate, large defect differences, high leakage detection rate, etc., a defect detection algorithm to improve YOLOv9 is proposed. Firstly, the algorithm improves the RepNCSPELAN4 module in the feature extraction network through the FasterBlock in FasterNet, and the RepNCSPELAN4-FB module is designed to realize the multi-scale feature fusion, so as to reduce the number of parameters of the model, and secondly, using the inverse residual structure of iRMB and a kind of highly efficient multi-scale attention module, EMAttention, to combine to form a new iEMA module that improve the accuracy of the network, and finally, using the Inner-WIOU loss function to improve the bounding box regression loss, which improves the model′s detection performance for inhomogeneous distributions and target defects at different scales. Through experiments on the GC10-DET dataset, the improved algorithm improves the precision, recall and map@0.5 by 3.5%、 3% and 2.1% compared with the original algorithm.The model shows good performance in steel surface defect detection.