image-colorization-using-gans

Gray-scale Image Colorization using Conditional Generative Adversarial Network. This is a PyTorch implementation of the DCGAN as described in the paper Image Colorization using Generative Adversarial Networks

Method

In a traditional GAN, the input of the generator is randomly generated noise data z. However, this approach is not applicable to the automatic colorization problem due to the nature of its inputs. The generator must be modified to accept grayscale images as inputs rather than noise. This problem was addressed by using a variant of GAN called conditional generative adversarial networks. Since no noise is introduced, the input of the generator is treated as zero noise with the grayscale input as a prior:

Network Architecture

The architecture of generator is inspired by U-Net:



The architecture of the model is symmetric, with n encoding units and n decoding units. For discriminator, we use similar architecture as the baselines contractile path.

Datasets

We used CIFAR-10 dataset to train model. Download dataset here.