Course contentsShow
← Machine Learning and Deep Learning
Lesson 766 of 3,53818. Regularization Techniques for Deep NetworksFree lesson

Basic Image Augmentation Techniques

Geometric transformations: rotations, flips, crops, and scaling for computer vision tasks.

Basic Image Augmentation Techniques

What You'll Learn

You'll discover how simple geometric transformations of training images act as powerful regularization by teaching your model to recognize objects from different perspectives.

The Core Idea

Imagine teaching a child to recognize cats. If you only show them photos of cats sitting upright and facing forward, they might not recognize a cat that's lying down or viewed from the side. Basic image augmentation solves this problem for neural networks by artificially creating variations of your training images through geometric transformations.

These transformations don't require collecting more data—you simply modify existing images during training in ways that preserve their meaning but change their appearance.

Common Geometric Transformations

Rotations

Rotating images by small angles (e.g., ±15°) helps your model recognize objects regardless of orientation. A rotated dog is still a dog.

Flips

Horizontal flips work well for most tasks (a mirror-image car is still a car), but vertical flips are typically avoided unless your task specifically needs them (medical scans might benefit; street signs wouldn't).

Crops

Random cropping forces the network to recognize objects even when only partially visible, simulating real-world occlusion and different framing.

Scaling (Zoom In/Out)

Changing image scale teaches the model to identify objects at different distances from the camera—crucial for detection tasks where object size varies.

Why This Works as Regularization

Like dropout randomly disabling neurons, augmentation randomly transforms images, preventing the network from memorizing exact pixel patterns. Instead, it learns robust, invariant features that generalize better to new data—reducing overfitting without collecting more samples.

Key Takeaway: Basic image augmentation applies geometric transformations (rotations, flips, crops, scaling) to training images, creating diverse variations that teach your network robust, position- and orientation-invariant features while acting as powerful implicit regularization.