HSC-YOLO: steel surface defect detection model based on improved YOLOv10n

Abstract During the production process of steel, the control of surface quality is crucial to the performance of the final molded product, so it is necessary to detect defects on its surface during the production process. Aiming at the problems of low detection accuracy and insufficient feature extraction and fusion capability in steel surface defect detection, a lightweight and multi-scale feature fusion model HSC-YOLO based on the improved YOLOv10n is proposed. Firstly, the backbone feature extraction network is reset using an improved lightweight network structure based on high performance GPU network (HGNetv2) to reduce the model size. Secondly, the multilevel feature fusion module semantics and detail infusion (SDI) is used instead of the two Concat modules in neck to enhance the semantic and detail information in the image. Finally, an iterative attentional feature fusion (iAFF) mechanism is introduced and combined with cross stage partial bottleneck with 2 convolutions (C2f) to solve the problems that occur when features are fused at different scales, especially the feature fusion problem with inconsistent semantics and scale. Test results on the datasets NEU-DET and GC10-DET show that the mean average precision (mAP) of HSC-YOLO improves by 3.9% and 2.1% over the mAP of YOLOv10n, and the detection speed improves by 46.2% and 43.1%, which provides the best detection results compared to other models.