Abstract: Industrial defect detection (IDD) is critical for ensuring product quality and production efficiency, yet traditional methods are hindered by the high cost of manual annotation and the challenge of identifying complex defect patterns. Active learning (AL) greatly reduces the annotation budget by selecting the most valuable samples for annotation, but it still faces the issue of cold start. A novel cold start active learning framework, namely SAM-based uncertainty-guided feature weighting (SUGFW) is proposed in this paper to keep low annotation cost for defect detection. Using the zero-shot segmentation capabilities of the segment anything model (SAM), SUGFW employs regional uncertainty estimation and a patch-based global distinct representation (PGDR) strategy to generate highly representative and diverse sample subsets for annotation. Based on a sample selection strategy that combines cluster and uncertainty, SUGFW ensures uniform sample distribution in both feature space and uncertainty intervals. Experiments on the NEU-Seg dataset for hot rolled steel strip surface defect detection demonstrate that SUGFW outperforms the advanced methods at present by 4 percentage points at least in mean intersection over union (IoU) at 3%, 10% and 20% annotation ratios. It has achieved a mean IoU of 0.842 3 under an annotation ratio of 20%, closely approaching the fully annotated performance of 0.859 9. This framework significantly reduces annotation costs and provides an efficient and robust solution for industrial defect detection.