Detection of Edge-blurred Small Targets of Steel Surface Defects Based on SAM-YOLO

To address the issue of challenging detection tasks for tiny and medium-sized objects because of backdrop confusion and inadequate feature representation of steel surface flaws, this paper proposes an efficient feature-enhanced, feature-fused SAM-YOLO detector based on the YOLOv8 model. It adds a new feature fusion module MSFF based on the original YOLOv8 model and replaces the bottleneck module in the C2f module with a feature enhancement module SACM. These two modules improve the enhancement and detection of defect edge features, local area sensing, and the network capability of multi-scale feature fusion, and the effect of SAM-YOLO is verified through several experiments on the NEU-DET steel surface defect dataset. The mAP50 of SAM-YOLO reaches 77.9%. mAP50:95 reaches 49.1%, exceeding the original YOLOv8 baseline model by 6.9% and surpassing several other benchmark models. To confirm the legitimacy of each SAM-YOLO model module, ablation tests were conducted on the model in the interim.