Important TorchVision Random Transforms used in Image Augmentation - Python PyTorch

 

Wondered what Torch Vision transforms actually do? These transforms are most common in deep learning to perform image data augmentation. These are common image transformations used different application of computer vision and image processing. 

In this post we will try to understand seven random transforms very important for image data augmentation. So what is a random transform. A random transformation will produce different result each time it is applied to an image.

What is Image Augmentation? 

Through image augmentation we can create more image date by applying different transformations on the existing data. Applying these transformation we basically alter the original image so that we can create a bigger dataset. This dataset is used to train deep learning models. 

Why random transforms are important for image augmentation?

The random transforms are more important to create the a bigger dataset with different types of images which are not available in the original dataset. So training process makes more robust.

What is importance of image data augmentation?

To create an image dataset is more time consuming and costly. So with sufficient images we can create a bigger dataset using image data augmentation. 

Table of Contents:

• RandomCrop
• RandomSizedCrop
• RandomRotation
• RandomAffine
• GaussianBlur
• RandomInvert
• ColorJitter

Import Required Libraries

import torch
import torchvision.transforms as T
from PIL import Image

The above imports are required to perform the transforms.

We  use PIL to read and display the image. You can use torchvision.io.read_image(path)  to read the image as a PyTorch tensor.

Read Input Image

The next step to read the input image. We use the following image "Koala.jpg" to demonstrate all transformations.

img = Image.open('Koala.jpg')
img.show()

Output:

Koala.jpg

Note: All transforms accepts both PIL image and Tensor. If the input image is a torch Tensor, it must be in [..., H, W] shape. 

RandomCrop

The RandomCrop transform crops the input image at random location. 

Syntax 

torchvision.transforms.RandomCrop(size)

Parameter: size- it is the expected size of cropped output image.

Example

cropper = T.RandomCrop(size=(256, 256))
cropped_img = cropper(img)
cropped_img.show()

Output:

Output- Random cropped image

Reference: torchvision.transforms.RandomCrop 

RandomResizedCrop

It crops a random portion of image and resize it to a given size. A crop of the original image is made and this crop is finally resized to the given size. It's used to train the Inception networks.

Syntax 

torchvision.transforms.RandomResizedCrop(size)

Parameter: size- it is the expected size of cropped output image.

Example:

resized_cropper = T.RandomResizedCrop(size=(256, 256))
resized_crop = resized_cropper(img)
resized_crop.show()

Output:

Output: Random resize cropped image

Reference: torchvision.transforms.RandomResizedCrop

RandomRotation

It rotate the image by an angle. 

Syntax 

torchvision.transforms.RandomRotation(degrees)

Parameter: degrees - it is range of degree (min, max). The angle by which the image is rotated is selected from this range.

Example

rotater = T.RandomRotation(degrees=(0, 180))
rotated_img = rotater(img)
rotated_img.show()

Output:

Output: random rotated image

Reference: torchvision.transforms.RandomRotation

RandomAffine

It make affine transformation of the image keeping center invariant.

Syntax 

torchvision.transforms.RandomAffine(degrees)

Parameter: degrees -  it is range of degree (min, max). The angle by which the image is rotated is selected from this range.

Example:

affine_transfomer = T.RandomAffine(degrees=(30, 70), translate=(0.1, 0.3), scale=(0.5, 0.75))
affine_img = affine_transfomer(img)
affine_img.show()

Output: Random affined image

Reference: torchvision.transforms.RandomAffine

GaussianBlur

It blurs image with randomly chosen Gaussian blur.

Syntax 

torchvision.transforms.GaussianBlur(kernel_size, sigma=(0.1, 2.0))

Parameter: kernel_size- it is size of Gaussian kernel.

Example

blurrer = T.GaussianBlur(kernel_size=(5, 9), sigma=(0.1, 5))
blurred_img = blurrer(img)
blurred_img.show()

Output:

Output: Gaussian Blur Image

Reference: torchvision.transforms.GaussianBlur

RandomInvert

It inverts the colors of the given image randomly with a given probability.

Syntax 

torchvision.transforms.RandomInvert(p=0.5)

Example

inverter = T.RandomInvert()
invertered_img = inverter(img)
invertered_img.show()

Output:

Output: Random Inverted Image

Reference: torchvision.transforms.RandomInvert

ColorJitter

It randomly change the brightness, contrast, saturation and hue of an image.

Syntax 

torchvision.transforms.ColorJitter(brightness=0, contrast=0, saturation=0, hue=0)

Example:

jitter = T.ColorJitter(brightness=.5, hue=.3)
jitted_img = jitter(img)
jitted_img.show()

Output:

Output: Random Color Jitter Image

Reference: torchvision.transforms.ColorJitter.

We will discuss the following random transforms in the next tutorial-

• RandomPerspective
• RandomPosterize
• RandomSolarize
• RandomAdjustSharpness
• RandomAutocontrast
• RandomEqualize
• AutoAugment
• RandAugment

Randomly Applied Transforms

• RandomHorizontalFlip
• RandomVerticalFlip
• RandomApply

Useful Resources:

• Computer Vision
• How To
• PyTorch
• Image Processing

Next Post: 3D Shape Analysis with Geometric Deep Learning- Research Plan

Previous Post: Write a program in python to read string and print longest word and its position