Version 1(CNN AutoEncode-Decoder)

ImageColorization

• Implemented the model based on the paper Colorful Image Colorization
• Implemented in Keras
• Custom Dataset built by taking an youtube video of mario gameplay and extacting frames

Training

• Run the train_script.py for training any custom dataset
• Run colorization.py to convert images from grayscale to rgb

CNN Model Architecture:

Comparision of output of both the models

1.Ground Truth      2.Cnn Output(Version 1)      2.GAN Output(Version 2)

• 
• 
• 

Version 2 (PIX2PIX GAN)

• This is the updated Image Colorization model built using the PIX2PIX Gan architecture

• Due to great results on the first dataset, I used a different dataset, collected from Pixabay and containing more vibrant colors to train this network

• Using Gan for Image colorization provided better results as compared to Version 1 of the model

• The input images for the network should be 256*256 in height and width for the training

• ./Gans/trainGan: Run this script for training the model on the dataset

• ./Gans/imgColor: Use this script to convert GrayScale to RGB images

Generator Model:

Discriminator Model:

Version 2 Converted Image Samples:

1.Ground Truth      2.GrayScale Version       3.Gan Output