Automated surface defect detection often faces a “cold-start” problem due to limited annotated data for new anomalies. Traditional object detectors struggle to converge in such few-shot settings. To address this, we adapt Vision–Language Models (VLMs), specifically YOLO-World. We use semantic pre-training to mitigate data scarcity. We evaluate this approach on the MVTec AD dataset in bounding-box format. We use a strict 1:9 train-validation split, resulting in an average of 11.8 defect instances per category. YOLO-World surpasses traditional baselines, like YOLOv11s and YOLOv26s, in 12 of 15 categories. The optimized VLM pipeline achieves up to 64.9% mAP@50 on texture-heavy categories, such as Tile, with only nine training instances. Ablation studies show standard optimization techniques are limited under 10-shot constraints. We find a critical augmentation divide. Disabling spatial distortions (Mosaic) is vital to preserving rigid-object geometry. The Normalized Wasserstein Distance (NWD) improves the localization of microscopic anomalies. Varifocal Loss (VFL) often causes model collapse. Ultimately, VLMs offer a superior foundation for cold-start inspection but require carefully tailored pipelines for robustness.
Sub-r-pa et al. (2026) studied this question.