AI defect detection in factories usually needs thousands of labelled photos before it works. A study involving University of Dayton statistician Ying-Ju Chen suggests that a pretrained model from OpenAI can do the job after seeing far fewer examples. According to a University of Dayton release dated 22 October 2026, the research was published in the Journal of Quality Technology. The paper itself appeared online on 3 June 2026, so the release is new but the study is several months old. This article is based on the release and the published paper.
The problem with traditional AI defect detection
Factories have used cameras and software to catch bad parts for years. The barrier, the university explains, is getting started. Before an inspection system can recognise a crack, a scratch or a bad weld, it usually needs thousands of labelled photos of those exact flaws. Collecting them can take months, and for a small or mid-size manufacturer, the cost alone can end the conversation.
“Automated visual inspection is not new, but many existing approaches require large amounts of labeled data, which can be difficult and costly to collect,” Chen says.
How CLIP handles AI defect detection
The team adapted CLIP, a model built by OpenAI that was trained on hundreds of millions of everyday images. Because it has already seen so much visual data, it does not need to learn what a defect looks like from scratch.
The paper is titled “Adapting OpenAI’s CLIP model for few-shot image inspection in manufacturing quality control: An expository case study with multiple application examples”. It uses CLIP’s image embeddings, which are numerical summaries of each picture, and classifies new images by comparing them with a small set of labelled examples using cosine similarity, a nearest-neighbour approach. The authors compare two versions of the model, ViT-L/14 and ViT-B/32, and present the method as a baseline.
Five tests across manufacturing
The study applied the method to five case studies:
- Metallic pan surfaces
- 3D printing extrusion profiles
- Stochastic textured surfaces
- Automotive assembly
- Microstructure classification
According to the paper’s abstract, the method reached high accuracy with roughly 50 to 100 examples per class for single-component and texture defects. Performance degraded in complex scenes with multiple components, such as assembly checks. The study also offers practical guidance on applying and evaluating the method, including its strengths and limits.
Who was involved
The team came from universities in the US and Europe. Authors include Fadel Megahed of Miami University, Chen, Bianca Maria Colosimo and Marco Grasso of Politecnico di Milano, L. Allison Jones-Farmer, Sven Knoth of Helmut Schmidt University, Hongyue Sun of the University of Georgia and Inez Zwetsloot of the University of Amsterdam, the corresponding author.
“I am excited about what this research could mean for manufacturing and aerospace companies in and beyond our region,” Chen says. For another example of AI reading images, see our story on Penn’s Percival CT scan AI.
The case for and against
The case for this approach is cost. If a manufacturer can start inspecting parts with dozens of labelled images instead of thousands, automated quality control becomes realistic for smaller firms. The method is also simple, which makes it easy to test.
The case for caution is that the authors themselves frame it as a baseline, not a finished system. It struggled with complex, multi-part scenes, which are common on real production lines. The release does not describe any factory deployment.
My view on AI defect detection
As a writer, I feel this is a useful reminder that not every AI advance needs a giant new model. Reusing a general-purpose model with a small number of examples could open quality control to many more manufacturers, including smaller ones in countries like India. I would like to see it tested on live lines.
FAQ: CLIP AI defect detection
What model did the researchers use?
They adapted OpenAI’s CLIP, comparing the ViT-L/14 and ViT-B/32 versions.
How many examples does it need?
The paper’s abstract reports high accuracy with about 50 to 100 examples per class for single-component and texture defects.
Where does it struggle?
Performance degrades in complex scenes with multiple components, such as automotive assembly.