Image augmentation that preserves the label meaning

Image augmentation encodes an invariance: after a transformation, the label should still be correct. That assumption depends on the task. A horizontal flip may preserve “vertical stripe,” while a 90-degree rotation turns a vertical stripe into a horizontal one. Applying it without changing the label trains a contradiction.

A transformation test

The image lab applies two transforms to 20 original teaching images and checks predictions from its fitted convolutional model.

from deep_learning_cases import augmentation_case

result = augmentation_case()
assert result["horizontal_flip_label_preserving_by_definition"] is True
assert result["ninety_degree_rotation_requires_label_swap"] is True
assert result["rotation_swapped_label_accuracy"] == 1.0
print(result["flip_prediction_agreement"], result["rotation_original_label_accuracy"])

Horizontal flips preserve every prediction in this case. After a 90-degree rotation, accuracy against the original label is 0%, while accuracy against swapped labels is 100%. The transform is valid only with a label mapping for this task.

Build a label-semantics table

For each transform, ask a domain reviewer what changes:

• Horizontal flip may be invalid for text, laterality or traffic signs.
• Vertical flip may be unrealistic for natural scenes or medical orientation.
• Rotation may change digit or directional classes.
• Cropping may remove the labeled object.
• Color jitter may erase a diagnostic color signal.

Record permitted ranges and transformation probability. Visualize augmented examples with labels before training.

Keep augmentation inside training

Create validation and test images from untouched source groups. Applying stochastic training augmentation to test makes the metric depend on random views and changes the target population. Test-time augmentation is a separate inference method that needs its own fixed protocol and evaluation.

Seed randomness for reproduction, but also repeat training seeds to learn whether an improvement is stable. Compare no augmentation, one transform family and the final composition under the same budget. Measure class and slice behavior, not only aggregate accuracy.

Augmentation cannot manufacture missing populations. Flipping images from one device does not establish performance on another device or environment. Collect representative validation data for the intended use.

The Data Science course links augmentation choices to label definitions and split provenance.

Exercise

Write a transformation-label table for a real project. Implement tests for output shape, value range and label mapping. Have a reviewer inspect a seeded grid before running the model comparison.

Continue learning

This article is part of the Deep learning and computer vision sequence. Use the neighbouring tasks when you need the prerequisite or the next application.

• Review the prerequisite or neighbouring task in CNN image classification with a clean data split.
• Continue with Transfer learning: freeze, fine-tune and compare baselines.

Reference: PyTorch transforms documentation.