Skip to main content

Dataset Augmentaion

This article will cover how you can increase the size of your original dataset with the help of data augmentation. Data augmentaion is a practice of altering samples in your dataset, making them distinct enough from the original sample to be considered a new sample, and keeping alterations small enough to keep them recognizable as a part of the dataset's original data domain.
Examples: Adding slight noise to audio samples and mirroring images.

Image augmentaion

The simplest way to add data augmentaion to your training pipeline is to use Albumentations library.

Starting from the most basic ones, here are some augmentaion tricks you can use:

  • Original image: 
    image = Image.open("testImg.jpg")
image_np = np.array(image)
  • Image flipping or mirroring: 
    transform = A.Compose([A.HorizontalFlip(p=1.0)])
transformed_image = transform(image=image_np)["image"]



    transform = A.Compose([A.VerticalFlip(p=1.0)])
transformed_image = transform(image=image_np)["image"]
  • Image rotation: 
    transform = A.Compose([A.Rotate(p=1.0, limit=45, border_mode=0)])
transformed_image = transform(image=image_np)["image"]
  • HSV Jitter: 
    transform = A.Compose([A.ColorJitter(p=1.0)])
transformed_image = transform(image=image_np)["image"]
  • Gaussian Noise: 
    transform = A.Compose([A.GaussNoise(p=1.0, var_limit=(1000.0, 5000.0))])
transformed_image = transform(image=image_np)["image"]
Augmentaions almost always combined with each other: 
  transform = A.Compose([
      A.HorizontalFlip(p=1.0),
      A.VerticalFlip(p=1.0),
      A.Rotate(p=1.0, limit=45, border_mode=0),
      A.RandomBrightnessContrast(p=1.0, brightness_limit=(0.15,0.25)),
      A.ColorJitter(p=1.0),
      A.GaussNoise(p=1.0, var_limit=(1000.0, 2000.0),),
  ])
  transformed_image = transform(image=image_np)["image"]

Above is an extremecase of image augmentaion, we still want to keep the resulting images as close to the original data distribution as possible: 

  transform = A.Compose([
      A.HorizontalFlip(p=0.5),
      A.Rotate(p=0.5, limit=15, border_mode=0),
      A.RandomBrightnessContrast(p=0.5, brightness_limit=(-0.1,0.1)),
      A.ColorJitter(p=0.5),
      A.GaussNoise(p=0.5, var_limit=(50.0, 250.0),),
  ])
  transformed_image = transform(image=image_np)["image"]
Back to top