Colorizing Balck & White Images Using Generative Adverserial Networks

Implemeted a Generational Adverserial Network for colorizing black & white images.
Dataset used : COCO-2017 training
Original paper : Image-to-Image Translation with Conditional Adversarial Networks

Some examples

Things I Tried:

• Pre-trained the Generator separately:
  • Used U-Net with a pretrained (on ImageNet) Resnet model as its backbone and then pretrained the whole generator model on the dataset.
  • Tried with Resnet18 and Resnet34 models.
• Tried different Image Augmentations: both train-time (from Albumentation library) and test-time (from TTA library)
• Tried different optimisers for Generator and Discriminator.

Things to try in Future:

• Training for more epochs.
• Hyper-parameter optimisations
• Create a web application and deploy the model over it.