Rare defects
Important classes have few or no useful images.
A real problem. The right specialists.
Too fewdefect images for AI?
Good parts are common, real defects are rare. We assess whether targeted data collection, anomaly detection or synthetic defect images can support your inspection project.
Assess your data situation Initially free
and non-binding
and non-binding
No account requiredThe method can remain open
Aaron ScallaManaging Director, CorDev GmbH
You describe the problem. We clarify the next technical step.Component → objective → suitable specialist
Your starting point
Is your dataset holding the project back?
Unrepresentative images
Variants, illumination or actual production conditions are missing.
Tests look too good
Development and evaluation use the same or very similar images.
The inspection principle
Add data. Evaluate on new real images.
Real source images
Synthetic defect variants
Development & training
Match the approach to the data and defect task.
Independent real test
Schematic principle, not to scale. The drawings explain the inspection task; performance must be validated on real components.
A practical technical start
Match the approach to the data.
The defect task and honest evaluation on held-out real images are decisive.
- Collect targeted data
- Which defects or variants are missing, and can they be captured on real samples or agreed trials?
- Detect anomalies
- Are good images representative? Can relevant defects be separated from acceptable variation?
- Add synthetic data
- Are generated defects physically plausible, and do they improve detection on independent real images?
What happens next
A clear next step for your inspection task.
- 01
Describe the problem
A short description is enough to begin. Add photos or requirements if available.
- 02
Clarify the task
CorDev clarifies open questions with you and identifies the support your task needs.
- 03
Agree the next step
We agree the involvement of suitable specialists with you. Testing or a pilot requires a separate agreement.
Your inspection task
What do you need to establish on your component?
Describe what you observe and what you want to achieve. You do not need to know the right technology yet.
Aaron ScallaYour contact at CorDev
TopicToo few defect images for AI inspection
Initially free and non-binding. No account required.
Before you ask
Frequently asked questions.
How many defect images do we need?
There is no universal minimum. Defects, variation, image quality and the method determine the need.
Can synthetic images replace real defects?
They can support development. Their benefit must be demonstrated on independently captured real good and defective images.
Can we start with good images?
Some anomaly approaches support this. Representative good images and real defect examples remain important for evaluating relevant failures.
Does submitting create a paid order?
No. Paid work requires a separate, express agreement.