Few‑Shot for quality inspection

In brief: In industry, collecting hundreds of “defective” examples per reference is costly and often impossible at product launch. Few-shot approaches with vision-language models (LVLM) like IADGPT revolutionize quality inspection with only a few reference images. Dans l’industrie, collecter des centaines d’exemples « défectueux » par référence est coûteux et souvent impossible au lancement d’un produit. Les approches few‑shot avec modèles vision‑langage (LVLM) comme IADGPT révolutionnent l’inspection qualité avec seulement quelques images de référence.

1) Technology & challenges

In industry, collecting hundreds of “defective” examples per reference is costly and often impossible at product launch. Few-shot approaches with vision-language models (LVLM) fill this gap: the IADGPT pre-publication shows that a single model can detect, localize and explain anomalies from a few reference images, thanks to progressive training and in-context learning.

In parallel, the new MVTec AD 2 dataset raises the bar (transparency, backlight/dark-field, micro-defects) and reminds us that robustness in real conditions is the real judge (SOTA performance < 60% AU‑PRO on average).

2) Problem & solution approach

Typical problem: starting visual quality control on a new SKU with few images.

Pragmatic approach: build 10–20 “good” images (+ some defects if available), evaluate a classic baseline and a few-shot LVLM approach like IADGPT on readable metrics (AUROC image-level, AU‑PRO pixel-level) and stress the system under lighting/pose variations inspired by AD 2.

This protocol, backed by historical MVTec AD benchmarks, allows to quickly obtain a “go/no‑go” signal and objectify gains before pilot deployment.

Bottles on production line: anomaly detection with few-shot approaches and LVLM models

3) Conclusion & ARCY

At ARCY, we integrate these few-shot and LVLM approaches into our quality inspection and industrial vision solutions (defects, counting, PPE), with end-to-end support: KPI scoping, supervised POC, on‑prem/edge deployment if needed., nous intégrons ces approches few‑shot et LVLM dans nos solutions d’inspection qualité et de vision industrielle (défauts, comptage, EPI), avec un accompagnement de bout en bout : cadrage KPI, POC supervisé, déploiement on‑prem/edge si besoin.

Are you starting with little data but strong traceability requirements? Contact us to design a short, measurable, and transferable protocol for production.

4) Sources (selection)