Abstract: Aiming at the problems of low accuracy and slow speed of surface defect identification of steel wire, a surface defect recognition method for steel wire based on improved YOLOv10 was proposed. A spatial and channel reconstruction convolution module, a multi-scale separation enhanced attention module and an interactive multi-head self-attention module were introduced to improve the YOLOv10 network, so that it had a more accurate modelling capability for complex surface defect data, and an obvious improvement in the recognition capability for the defects in a small area that were close to each other or mutually occluded, and at the same time, it further reduced the spatial and channel redundancy of the features, and lightened the structure of the network to maintain a balance between recognition performance and computational overhead. Experimental results showed that the improved YOLOv10 defect identification network achieved competitive recognition accuracy and recognition speed.