Image augmentation is a technique used to increase the diversity of training images by applying random transformations while preserving labels. Common augmentations include rotations, translations (shifting image), flips (horizontal/vertical), scaling/zooming, cropping, brightness/contrast adjustments, adding noise, blur, color jitter, etc. For example, an image of a cat might be randomly rotated by 15 degrees and still be an image of a cat – using it in training helps the model become invariant to rotation. Augmentation is especially helpful when the training set is limited; it acts as a form of regularization, reducing overfitting by making the model see varied versions of each image. Some augmentations are application-specific (e.g., random erasing, Mixup or Cutout in images, or time-warping in audio). Key is that the semantics of the image don’t change. Most deep learning frameworks support on-the-fly augmentation during training.