STE-NET: An Improved Surface Defect Detection Network for Strip Steel

This study proposes STE-NET, a lightweight network for steel defect detection that improves feature preservation and texture recognition. To address YOLOv8’s complexity, we develop three key innovations: the PMSFA-block for multi-scale fusion, Star-backbone for efficient feature extraction, and Dysample block for sharper details. On NEU-DET dataset, STE-NET achieves 83.8% mAP with 15% fewer parameters, meeting real-time industrial needs while balancing accuracy and efficiency for smart manufacturing.