Training-Free Few-Shot Defect Segmentation Framework for Steel Tubes Surfaces Based on DINOv3 and SAM2

The identification of surface defects in steel tubular products is of paramount importance for ensuring industrial product quality and safeguarding manufacturing process safety. However, extant deep learning models necessitate voluminous training data, a condition that is challenging to fulfill in industrial settings. In this paper, we propose a training-free few-shot defect segmentation framework that integrates DINOv3 for feature association inference and employs SAM2 for prompt-based segmentation optimization. The framework comprises a Correlation Fusion Module (CFM), Scale-Adaptive Multi-Level Fusion (SAM-F), and a Geometric Prompt Module (GPM). The integration of these modules enables precise defect localisation using limited annotated samples without the necessity of updating model parameters. Experimental results show that the proposed method achieves superior detection performance on the CGFSDS-9 benchmark, with an overall FBIoU of 86.17\%, providing a practical solution for seamless steel tube surface defect detection under limited data conditions.