Abstract: The traditional YOLOv5 algorithm faces challenges such as significant variations in defect scales and high inter-class similarity during steel surface defect detection. These limitations greatly increase the difficulty of the detection task and significantly impair the performance of YOLOv5 in the field of defect detection on steel surfaces. To address the series of challenges mentioned above, a steel surface defect detection model by integrating MobileViTv3 with YOLOv5s is proposed to enhance the global information processing capability of YOLOv5s.Furthermore, to achieve the efficient fusion of multi-scale defect features, a novel PGCSP block is proposed to replace the C3 block, The PGCSP block, combined with GSConv, forms a new neck network LWT-Neck. Additionally, WIoU, a novel loss function, has been introduced to improve small object detection performance and mitigate scale-variation issues. Experimental validation on the NEU-DET dataset demonstrates that the proposed model achieves real-time detection with an average model accuracy (mAP) of 77.83%. This represents an improvement of 7.69% over the baseline YOLOv5s model and 4.27% over Faster R-CNN.