After learning how to create basic synthetic images using simple shapes and patterns, let's explore another straightforward approach to generating image data: applying noise and simple transformations, often called augmentations. While data augmentation is frequently used to modify existing real images, the same techniques can be applied to the basic synthetic images we've just discussed, or even used in combination, to create more varied datasets for training machine learning models.

Think of augmentation as slightly tweaking an image to create a new, but related, version. Why do this? Machine learning models, especially in computer vision, need to see many variations of an object or scene to learn effectively and recognize it under different conditions. If your dataset only contains perfectly centered, well-lit images, the model might struggle when faced with images that are slightly rotated, zoomed in, or have minor imperfections. Augmentation helps bridge this gap.

Adding Noise

Real-world images often contain noise due to camera sensor limitations, poor lighting conditions, or transmission errors. Intentionally adding noise to synthetic or real images can make your model more robust to these imperfections.

A common type of noise is Gaussian noise, which adds small, random values drawn from a bell-shaped (normal) distribution to each pixel's intensity. Imagine adding tiny, random specks of varying brightness across the image.

Another type is Salt-and-Pepper noise, which randomly replaces some pixels with pure white (salt) or pure black (pepper) pixels.

Adding noise simulates imperfect conditions and helps the model learn to focus on the important features rather than being distracted by minor pixel variations. The amount and type of noise can usually be controlled, allowing you to simulate different levels of imperfection.

Simple Geometric Transformations

These involve changing the position, orientation, or scale of the image content.

Rotation: This involves turning the image around its center by a specific angle. For example, rotating an image of a cat by 5 or 10 degrees creates a new training sample where the cat is slightly tilted. This helps the model recognize cats even if they aren't perfectly upright in the photo.
Scaling (Zooming): This changes the size of the image content, making it appear closer or farther away. You might scale an image down to 90% or up to 110% of its original size. This teaches the model to identify objects regardless of their apparent size in the frame.
Translation (Shifting): This involves moving the image content horizontally or vertically. Shifting an object slightly off-center helps the model learn that the object doesn't always have to be in the exact middle of the image.
Flipping: Images can be flipped horizontally (like looking in a mirror) or vertically. Horizontal flipping is a very common augmentation, especially for objects where left-right orientation doesn't change the object's identity (like cars or animals). Vertical flipping is used less often, as it's uncommon to see objects naturally upside down (e.g., an upside-down car usually signifies something different than a right-side-up car).

Color and Brightness Adjustments

Minor changes to brightness, contrast, or saturation can also simulate different lighting conditions. Making an image slightly brighter or darker, or increasing its contrast, produces variations that a model might encounter in real scenarios.

Augmentation as a Synthesis Technique

Applying these transformations creates new data points. If you start with one basic synthetic image (like a programmatically generated square) and apply five different augmentations (e.g., add noise, rotate 5 degrees, rotate -5 degrees, scale 1.1x, flip horizontally), you now have six images in your dataset instead of one. This is a simple yet effective way to expand your dataset, especially when combined with generating basic shapes or patterns.

Applying various transformations like adding noise, rotating, scaling, or flipping to an original image generates multiple new, augmented images for training.

Considerations for Augmentation

While powerful, augmentation should be applied thoughtfully:

Maintain Realism: The applied transformations should generally result in images that could plausibly occur in the real world. For example, rotating a pedestrian detection image by 180 degrees might not be useful unless you expect to see upside-down pedestrians.
Preserve Labels: The augmentation should not change the fundamental identity or label of the object in the image. A slightly rotated cat is still a cat. Adding too much noise or scaling excessively might obscure the object entirely.
Task Dependence: The best augmentations depend on the specific task and the types of variations expected in the real data.

Software libraries, which we'll touch upon in Chapter 6, provide convenient functions to apply these augmentations easily. By combining the generation of simple shapes and patterns with noise and basic augmentations, you can start building more diverse and robust datasets for your computer vision projects without needing vast amounts of real-world data initially.

Was this section helpful?