Isola, Phillip, et al. Image-to-Image Translation with Conditional Adversarial Networks. 2016. DOI.org (Datacite), https://doi.org/10.48550/ARXIV.1611.07004.

Image Coloration Using GANs

The Pix2Pix Generative Adversarial Network, or GAN, is an approach to training a deep convolutional neural network for image-to-image translation tasks.

Description

• The generator G is trained to produce outputs that cannot be distinguished from “real” coloured images by an adversarially trained discriminator, D, which is trained to do as well as possible at detecting the generator’s “fake colour images”
• Black and white image resized to (256,256) is feed into the generator trained against a discriminator which output the A channel(colour balance between Green and Magenta) and B channel(colour balance between Blue and Yellow) of LAB colour space where L channel (Lightness) is the original Black and White Image itself.

Generator Network

Generator model described in Pix2Pix research paper is the encoder - decoder architecture or U Net architecture. I used Encoder - Decoder Architecture

Discriminator Network

Below discriminator network is used to train against generator

Used BCEWithLogitsLoss and trained it with Cifar 10 datasets. Black & White images are made for input using OpenCV

Results

Model has trained pretty well and was enough to fool the discriminator

Examples: