Retrain the Model
This page explains when and how to retrain your AI model on the OV20i system, whether you're using Classification or Segmentation, to keep inspections accurate as parts or production conditions change.
AI performance depends on the relevance of training data. Retrain when parts, conditions, or inspection requirements evolve.
When Should You Retrain?
Retraining ensures the model keeps up with production realities. Apply this guidance to both Classification and Segmentation projects.
Retrain If:
- You're inspecting a new SKU or part variant
- Your inspection requirements have changed (e.g., now detecting surface defects or grease)
- You've changed the fixture, robot, or part presentation
- Lighting has changed significantly (e.g., reflections, angle, intensity)
- Accuracy has dropped — more false positives/negatives
- You need tighter confidence thresholds or more precise results
- The model shows signs of overfitting or under-generalizing
How to Retrain (for Both Model Types)
-
Capture new sample images from your current production setup
-
For Classification: Label images or ROIs with class names
For Segmentation: Draw pixel-level masks on defects (or good/bad regions)
-
Choose the appropriate training mode:
- Classification:
- Fast – For quick testing or iteration
- Accurate – For production use
- Segmentation:
- Accurate – Only one mode, optimized for precision
- Classification:
-
Run training inside the Recipe interface
-
Review model outputs and test live inspections
-
Deploy the new model when confidence and coverage meet expectations
Model Type Quick Guide
| Model Type | Best For | Training Modes | Output |
|---|---|---|---|
| Classification | Good/Bad or discrete state decisions | Fast, Accurate | Whole image or ROI class |
| Segmentation | Pixel-level defect or region mapping | Accurate only | Labeled mask (highlighted areas) |
Sample Use Cases
| Example | Model Type |
|---|---|
| Detecting missing bolts | Classification |
| Checking for scratches or dents | Segmentation |
| Verifying grease presence | Classification or Segmentation (depends on precision needed) |
| Measuring foam coverage | Segmentation |
Best Practices for Retraining
- ✅ Always use fresh production images
- ✅ Include a mix of pass/fail cases, especially edge cases
- ✅ Use at least 30–50 images per class (Classification)
- ✅ Make sure ROI boundaries or masks match the part layout
- ✅ Use Accurate mode before deployment
- ❌ Don't train with blurry, low-light, or misaligned images