Abstract: The single-stage target detection network YOLOv5 has certain deficiencies when dealing with feature extraction and sensory feature fusion for surface defects on hot-rolled steel. This paper proposes an optimized YOLOv5 algorithm for surface defect detection of hot rolled strip steel, which adjusts the anchor clustering anchor frame setting by the IOU-K-means++ algorithm, increases the dynamic head target detection head, introduces the channel attention mechanism (C3_CA), and combines the hard swish activation function with the WioU_loss bounding box regression function, effectively improving the comprehensive accuracy of hot rolled strip steel surface defect detection. Test results from the NEU-DET dataset show that compared to the single-stage YOLOv5 algorithm fusion results, the mean average precision (mAP) of the optimized YOLOv5 network model can be improved up to 75.7%, and the network constraint rate can be effectively improved by 6.1%. The above optimized YOLOv5 algorithm is a useful reference for hot rolled strip steel surface defect location surveys, classification points, and impact assessments and provides important support for high-precision screening of metal surfaces.