Train a Classifier

This guide shows you how to configure and train a classification model on the OV20i camera system. Use this procedure when you need to automatically categorize objects into different classes based on visual features.

When to Use Classification: Sorting parts by type, size, color, or condition; identifying different product variants; quality control with multiple acceptable categories.

Prerequisites

• Active recipe with imaging settings configured
• Template image and alignment completed (or skipped)
• Inspection ROI(s) defined
• Sample objects representing each class you want to detect

Step 1: Access Classification Block

1.1 Navigate to Classification

1. Click "Classification Block" in the breadcrumb menu, OR
2. Select from dropdown in navigation bar

1.2 Verify Prerequisites

Ensure the following blocks show green status:

• ✅ Imaging Setup
• ✅ Template and Alignment (or skipped)
• ✅ Inspection Setup

Step 2: Create Classification Classes

2.1 Define Your Classes

1. Click Edit under "Inspection Types"
2. Add classes for each category you want to detect

2.2 Configure Each Class

For each class:

1. Enter Class Name: Use descriptive names (e.g., "Small", "Medium", "Large")
2. Select Class Color: Choose distinct colors for visual identification
3. Add Description: Optional details about the class
4. Click Save

2.3 Class Naming Best Practices

Good Names	Poor Names
Small_Bolt, Medium_Bolt, Large_Bolt	Type1, Type2, Type3
Red_Cap, Blue_Cap, Green_Cap	Color1, Color2, Color3
Good_Part, Defective_Part	Pass, Fail
Screw_PhillipsHead, Screw_Flathead	A, B

Step 3: Capture Training Images

3.1 Image Capture Process

For each class, capture minimum 5 images (10+ recommended):

1. Place object representing the class in the inspection area
2. Verify object is within ROI boundaries
3. Click Capture to take training image
4. Select appropriate class from dropdown
5. Click Save to store labeled image
6. Repeat with different examples of the same class

3.2 Training Data Requirements

Class	Minimum Images	Recommended Images	Notes
Each class	5	10-15	More images = better accuracy
Total dataset	15+	30-50+	Balanced across all classes
Edge cases	2-3 per class	5+ per class	Borderline examples

3.3 Training Image Best Practices

Do:

• Use different examples within each class
• Vary object orientations and positions
• Include good lighting conditions
• Capture edge cases and borderline examples
• Maintain consistent ROI framing

Don't:

• Use identical objects repeatedly
• Include multiple objects in one ROI
• Mix classes in single images
• Use blurry or poorly lit images
• Change ROI size between captures

3.4 Quality Control

After capturing each image:

1. Review image quality in preview
2. Verify correct class label assignment
3. Delete poor quality images using Delete button
4. Retake if necessary

Step 4: Configure Training Parameters

4.1 Access Training Settings

1. Click Train Classification Model button

4.2 Select Training Mode

Choose based on your needs:

Training Mode	Duration	Accuracy	Use Case
Fast	2-5 minutes	Good for testing	Initial model validation
Balanced	5-15 minutes	Production ready	Most applications
Accurate	15-30 minutes	Highest precision	Critical applications

4.3 Set Iteration Count

Manual iteration setting:

• Low (50-100): Quick testing, basic accuracy
• Medium (200-500): Production quality
• High (500+): Maximum accuracy, slower training

4.4 Advanced Settings (Optional)

Batch Size:

• Smaller batches: More stable training, slower
• Larger batches: Faster training, may be less stable

Learning Rate:

• Lower values: More stable, slower learning
• Higher values: Faster learning, risk of instability

Recommendation: Use default settings unless you have specific performance requirements.

Step 5: Start Training Process

5.1 Initialize Training

1. Review training configuration
2. Click Start Training
3. Monitor progress in training modal

5.2 Training Progress Indicators

Monitor these metrics:

• Current Iteration: Progress through training cycles
• Training Accuracy: Model performance on training data
• Estimated Time: Remaining training duration
• Loss Value: Model error (should decrease over time)

5.3 Training Controls

Available actions during training:

• Abort Training: Stop training immediately
• Finish Early: Stop when current accuracy is sufficient
• Extend Training: Add more iterations if needed

5.4 Training Completion

Training stops automatically when:

• Target accuracy reached (typically 95%+)
• Maximum iterations completed
• User manually stops training

Step 6: Evaluate Model Performance

6.1 Review Training Results

Check final metrics:

• Final Accuracy: Should be >85% for production use
• Training Time: Note duration for future reference
• Convergence: Verify accuracy stabilized

6.2 Model Quality Indicators

Accuracy Range	Quality Level	Recommendation
95%+	Excellent	Ready for production
85-94%	Good	Suitable for most applications
75-84%	Fair	Consider more training data
<75%	Poor	Retrain with more/better images

6.3 Troubleshooting Poor Performance

Problem	Likely Cause	Solution
Low accuracy (<75%)	Insufficient training data	Add more labeled images
Training doesn't improve	Poor image quality	Improve lighting/focus
Classes confused	Similar-looking objects	Add more distinguishing examples
Overfitting	Too few images per class	Balance dataset across classes

Step 7: Test Classification Performance

7.1 Live Testing

1. Click Live Preview to access real-time testing
2. Place test objects in inspection area
3. Observe classification results:
   • Predicted class name
   • Confidence percentage
   • Processing time

7.2 Validation Testing

Systematic validation process:

Test Object	Expected Class	Actual Result	Confidence	Pass/Fail
Known Class A object	Class A	_____	____%	☐
Known Class B object	Class B	_____	____%	☐
Borderline example	Class A or B	_____	____%	☐
Unknown object	Low confidence	_____	____%	☐

7.3 Performance Validation

Verify these aspects:

• Accuracy: Correct classifications for known objects
• Confidence: High confidence (>80%) for clear examples
• Consistency: Repeatable results for same object
• Speed: Acceptable processing time for your application

Step 8: Model Optimization

8.1 If Performance is Unsatisfactory

Iterative improvement process:

1. Identify problem areas:
   • Which classes are confused?
   • What objects are misclassified?
   • Are confidence levels appropriate?
2. Add targeted training data:
   • More examples of confused classes
   • Edge cases and borderline examples
   • Different lighting/positioning conditions
3. Retrain model:
   • Use "Accurate" mode for better performance
   • Increase iteration count
   • Monitor improvement in accuracy

8.2 Advanced Optimization

For critical applications:

• Data augmentation: Use varied lighting and positions
• Transfer learning: Start from similar trained models
• Ensemble methods: Combine multiple models
• Regular retraining: Update with new production data

Step 9: Finalize Configuration

9.1 Save Model

1. Verify satisfactory performance
2. Model automatically saves when training completes
3. Note model version for documentation

9.2 Documentation

Record these details:

• Training date and version
• Number of images per class
• Training mode and iterations used
• Final accuracy achieved
• Any special considerations

9.3 Backup Configuration

1. Export recipe for backup
2. Save training images separately if needed
3. Document model parameters

Success! Your Classifier is Ready

Your trained classification model can now:

• Automatically categorize objects into defined classes
• Provide confidence scores for each prediction
• Process images in real-time for production use
• Integrate with I/O logic for automated decision-making

Ongoing Maintenance

Regular Model Updates

• Monitor performance over time
• Add new training data as needed
• Retrain periodically to maintain accuracy
• Update classes for new product variants

Performance Monitoring

• Track accuracy metrics in production
• Identify drift in model performance
• Schedule retraining based on performance degradation

Next Steps

After training your classifier:

1. Configure I/O logic for pass/fail decisions
2. Set up production workflows in IO Block
3. Test complete inspection system end-to-end
4. Deploy to production environment

Common Pitfalls

Pitfall	Impact	Prevention
Insufficient training data	Poor accuracy	Use 10+ images per class
Imbalanced classes	Biased predictions	Equal images across classes
Poor image quality	Inconsistent results	Optimize lighting and focus
Overly similar classes	Confused classifications	Choose distinct class definitions
No validation testing	Production failures	Always test with unseen objects