Color Jitter

If you have ever worked on a Computer Vision project, you might know that using augmentations to diversify the dataset is the best practice. On this page, we will:

  • Сover the Color Jitter augmentation;

  • Check out its parameters;

  • See how Color Jitter affects an image;

  • And check out how to work with Color Jitter using Python through the Albumentations library.

Let's jump in.

To define the term, Color Jitter is a data augmentation technique that allows researchers to vary the brightness, contrast, hue, and saturation of the sample images. To understand how Color Jitter works, let's first observe the structure of digital images.

  • In general, images are stored in computers as matrixes of numbers known as pixel values.

  • Each matrix has a particular dimension, which is the height (x) multiplied by the width (y) of the image. For example, if there are 30 pixels across the height and 40 pixels across the width, the dimension of the image is 30 x 40.

  • The pixel values can vary from 0 to 255 and represent the intensity of each pixel. The range from 0 to 255 is chosen since exactly this number of values (256) can be stored in one byte (8 bits). 0 stands for black color, and 255 stands for white.

  • The matrix of numbers is also referred to as the channel.

Grayscale images have only one channel. Changing the pixel values of such images will produce different shades of gray.

Source

Typical colored images are composed of 3 matrixes or channels - Red, Green, and Blue, where each matrix contains information about the intensity of each color.
In the end, all the matrixes are superimposed so that each pixel contains:

  • A value for the red color (R);

  • A value for the green color (G).

  • A value for the blue color (B).

An example of a matrix for the Red channel

The combination of these channels (R, G, B) makes it possible to create any color the human eye can perceive.

Source

The value for each parameter below must fall between [0.0, 1.0].

  • Brightness - refers to the relative lightness or darkness of the pixel. It can vary from black (no brightness) to white (maximal brightness). Increasing the brightness makes every pixel in the image look lighter, and vice versa.


Source

Consider an image of a square made of 5 rows and 5 columns. The pixel values are all equal to zero, making the square black.

If we add +100 to each pixel value, we will receive a slightly brighter square:

Subtracting pixel values would again make the square appear darker.

  • Contrast - refers to the difference in brightness between the image's lightest and darkest areas. The square from the example above has a contrast of zero:

    Maximum pixel value (lightest) - Minimum pixel value (darkest) = 100 - 100 = 0
  • Hue - refers to the tone of a color. Hue is different from color since colors can have saturation or brightness as well as a hue.
A color wheel representation of a hue. Source
  • Saturation - refers to the intensity of a color. It can vary from a gray tone (no saturation) to an intense tone (high saturation).
Source

The picture below represents the difference between hue, saturation, and brightness.

Source
The brightness, contrast, hue, and saturation parameters will always lead to random transformations. For instance, the brightness parameter can make the image both lighter or darker to a varying extent. This is done to diversify the dataset and help your model generalize better.

The value for these parameters serves more as an amplitude:
- The closer the value is to 1, the more intense applied transformations might be;
- The closer the value is to 0, the slighter the changes in the image.

  • Probability of applying transform - defines the likelihood of applying Color Jitter to an image.

If a large fraction of training images needs to be jittered, set a high probability.
Jittered brightness
Jittered contrast
Jittered hue
Jittered saturation
python 
      import albumentations as albu
from PIL import Image
import numpy as np

transform = albu.ColorJitter(brightness=0.2, contrast=0.2, saturation=0.2, hue=0.2, p=0.5)

image = np.array(Image.open('/some/image/file/path'))
image = transform(image=image)['image']

# Now the image is transformed and ready to be accepted by the model

Boost model performance quickly with AI-powered labeling and 100% QA.

Learn more
Last modified