To effectively analyze the stochastic behavior of data augmentation, an auxiliary function can be defined to repeatedly apply a specific augmentation method to an input image. By collecting and displaying the results of these multiple applications, practitioners can visually observe the full range of random variations introduced by the transformation.

Visualizing Image Augmentation Effects

In deep learning practice, image augmentation that incorporates random operations—such as random cropping or flipping—is typically applied exclusively to the training dataset. During evaluation phases, such as when using a validation set for hyperparameter tuning or a test set for final prediction, these random operations are disabled. This ensures that no randomness is introduced during evaluation, allowing the model to be tested on deterministic, unaltered input images.

Image Augmentation in Training vs. Evaluation

Image augmentation leverages random operations to generate diverse, novel images from existing training data. Due to this stochastic nature, applying the exact same augmentation procedure multiple times to a single image will yield distinct, randomized variations. This deliberate introduction of randomness dynamically expands the dataset and is fundamentally designed to improve the generalization ability of deep learning models.

Claude

For example, instead of taking more cat pictures, you can use one or a combination of the following methods:
- Mirroring (Horizontally/vertically flipping)
- Random cropping (Zooming in)
- Rotating
- Shearing
- Local wrapping
- Color shifting

Methods of Data Augmentation in Image Recognition Through Deep Learning

Dive into Deep Learning

By adding distortions to the R, G, and B values of images, we still get images in which the same objects are recognizable.

Color Shifting for Data Augmentation in Image Recognition Through Deep Learning

Randomness in Image Augmentation

Flipping an image horizontally (left-to-right) or vertically (up-and-down) is a foundational and widely used image augmentation technique. Because flipping typically does not alter the category of the object depicted, it safely expands the dataset. While horizontal flipping is extremely common, vertical flipping is employed less frequently but remains useful when the vertical orientation of objects does not hinder recognition.

Image Flipping Augmentation

In deep learning practice, a single augmentation technique is rarely applied in isolation. Instead, multiple distinct image augmentation methods—such as horizontal flipping, random resized cropping, and color jittering—are combined into a unified transformation pipeline. By applying these diverse transformations sequentially through a composition function (often denoted as `Compose`), the range of possible random variations per image increases significantly, yielding more robust and invariant models.

Learn Before

Related

Learn After