Classification
What it does: Assigns a single category label to an entire image.
Input
A whole image (e.g., a photo of a product on a conveyor belt)
Labels
This model expects image-level labels, such as:
- "defective" vs "good"
- "product_type_A" vs "product_type_B"
- "empty" vs "product_present"
- Use clear and consistent labels: e.g., always "good" instead of mixing "ok", "okay", etc.
- Avoid ambiguous labels; uncertain images can confuse the model
- Review labels before training to prevent errors
Output
A class label (e.g., "defective", "good", "product_type_A")
How it looks in the platform
- The model returns a single label for each uploaded image
- Confidence scores are provided for each class (optional display)
When to use classification
Use classification when you need to:
✅ Know what’s in the image, without locating it
✅ Make a single decision per image – One dominant object or condition
✅ Detect presence or absence – "Is there a product?" "Is the cap on?"
✅ Perform binary quality control – Pass/fail, good/defective
Example use cases
| Application | What It Classifies |
|---|---|
| Product presence | "Empty" vs "Product present" |
| Overall quality | "Good" vs "Defective" |
| Product type | "Type A" vs "Type B" vs "Type C" |
| Orientation | "Upright" vs "Tilted" vs "Fallen" |
Limitations:
- ❌ Cannot locate defects or objects
- ❌ Cannot handle multiple objects in the same image
- ❌ Does not provide spatial information
- ❌ Cannot distinguish between multiple instances
- ❌ Less robust – relies only on global image features
- ❌ Needs enough images per class (minimum ~50-100 recommended)
- ❌ Sensitive to class imbalance – rare classes may be misclassified
- ❌ Image resolution affects accuracy; very small images may reduce performance
Better alternatives:
Prefer Segmentation or Instance Segmentation because they:
- ✅ Provide geometric information (shape, size, position)
- ✅ Leverage color and spatial features
- ✅ Are more accurate and robust
- ✅ Can be simplified to classification if only detection is needed
Exception: Use classification only for very simple cases where a single image-level decision is enough and localization is not required.
Image classification models
| Model | Use for | Example Applications | Input | Output |
|---|---|---|---|---|
| Generic classification | Product categorization, quality grading, scene recognition | Sorting products by type, identifying defective vs normal items | Whole images | Class labels with confidence scores |
| Dynamic classification | Applications where categories may evolve over time | Detecting new product variants, adapting to seasonal changes | Whole images | Adaptable class labels with confidence scores |
| Generic classification embedding | Similarity search, clustering, anomaly detection | Finding visually similar products, spotting unusual items | Whole images | Feature vectors that capture image characteristics |
Considerations for training
- Input size: Typically 224x224 or 256x256 px; larger sizes can improve accuracy but increase computation.
- Data augmentation: Rotations, flips, and color changes improve model robustness.
- Class balance: Ensure all classes have enough examples to avoid bias.
- Transfer learning: Using pre-trained models speeds up training and improves performance.
- Batch size and learning rate: Experiment to optimize training efficiency.
- Regularization: Dropout or weight decay to prevent overfitting.
Additional configuration
- Dynamic classification: Ensure proper label mapping and update logic for evolving classes.