Data Augmentation for Images
Labeled images are scarce and expensive. Data augmentation creates new training examples by applying transformations that change pixels but keep the label correct.
Common transforms
- Flips and rotations present the object from new orientations.
- Crops and scaling teach the model that position and size vary.
- Color jitter alters brightness and contrast for lighting robustness.
- Noise and blur simulate imperfect cameras.
Why it helps
Augmentation acts as a regularizer. By seeing many variants of the same object, the model learns features that are invariant to those changes rather than memorizing exact pixels. This reduces overfitting and improves generalization to real world variation.
Label preservation
The golden rule is that a transform must not break the label. Flipping a cat horizontally is still a cat, so it is safe. But flipping the digit two might create something that no longer reads as a two, and mirroring text destroys it. Choosing augmentations that respect the task is essential, and applying them randomly each epoch maximizes variety.
Key idea
Image augmentation generates label preserving variants so the model learns invariant features and overfits less.