CODEWITHSUNDEEP
pythonpython-imaging-library
kkgarg
kkgarg

Reputation: 1376

Partial Inversion of pixels in Pytorch transforms

I want to randomly invert some of the pixels of an image, that's what I mean by "Partial Inversion". This is definitely feasible using two for loops but I am looking for a faster operation since this operation is to be applied to a lot of images in the dataset. I didn't find any operation in the PIL documentation. I am attaching a sample image for the reference, assume it has all channels (RGB). Any help would be appreciated.

sample_image

Upvotes: 0

Views: 337

Answers (1)

Nicolas Gervais
Nicolas Gervais

Reputation: 36684

Will you consider solutions other than PIL? It's very easy to manipulate pixels with numpy. To invert, just subtract the array values from 255.

import tensorflow as tf
import torch
import matplotlib.pyplot as plt

(x, _), (_, _) = tf.keras.datasets.mnist.load_data()

x = x[0]

x = torch.cat([x[:14, :], 255 - x[14:, :]])

plt.imshow(x.numpy(), cmap='Greys')

enter image description here

You can call it in your transforms like this:

torchvision.transforms.Lambda(lambda x: torch.cat([x[:14, :], 255 - x[14:, :]]))

Upvotes: 1

Related Questions