Beyond CLIP: Calibrated (Few-Shot) Defect Detection With System One Decision Models
What it is
This paper proposes a framework for few-shot defect detection in manufacturing using structured decision models (Jev and Clef-Flash) instead of standard vision transformers like CLIP. It introduces an agentic approach to refine text descriptions and a calibration method to improve the reliability of defect probability estimates.
The problem it solves
Traditional defect detection models require thousands of labeled images, and even few-shot CLIP models need 50-100 images per class, which is often impossible to collect for rare defects in high-yield manufacturing.
What you could use it for
- Inspecting high-yield or low-volume manufacturing products where collecting defective samples is difficult
- Automating visual quality control with as few as two reference images per class
- Calibrating AI model outputs for reliable pass/fail inspection decisions
Why it matters
Reduces the data burden for training defect detection models from thousands of images to just two per class, making AI-based quality control feasible for niche or high-yield manufacturing processes.
Read the source ↗Back to the library →
Summary written by an AI model from the source; check the source before you rely on it.