Automatic Image Colorization

Overview

This is a TensorFlow implementation of the Residual Encoder Network based on Automatic Colorization and the pre-trained VGG16 model from https://github.com/machrisaa/tensorflow-vgg

For latest TensorFlow with estimator support, check tf-1.12 branch. (still under development, the training code is working now)

Structure

• config.py: config variables like batch size, training_iters and so on
• image_helper.py: all functions related to image manipulation
• read_input.py: all functions related to input
• residual_encoder.py: the residual encoder model
• common.py: the common part for training and testing, which is mainly the workflow for this model
• train.py: train the residual encoder model using TensorFlow built-in AdamOptimizer
• test.py: test your own images and save the output images

TensorFlow graph

How to use

• Download pre-trained VGG16 model vgg16.npy to vgg folder

• Option 1: Use pre-trained residual encoder model

  • Download model here
  • Unzip all files to summary_path (you can change this path in config.py)

• Option 2: Train your own model!

  1. Change the batch_size and training_iters if you want.
  2. Change training_dir to your directory that has all your training jpg images
  3. Run python train.py

• Test

  1. Change testing_dir to your directory that has all your testing jpg images
  2. Run python test.py

Examples

• 

• 

• 

• 

• 

• 

• 

• 

• 

• 

• 

• 

• More example output images can be found in sample_output_images folder.

References

• Automatic Colorization
• pavelgonchar/colornet
• raghavgupta0296/ColourNet
• pretrained VGG16 npy file

Contributing

See CONTRIBUTING.md

License

GNU GPL 3.0 for personal or research use. COMMERCIAL USE PROHIBITED.