ImageColorization

Forked from github.com/praywj/Interactive-Deep-Colorization-and-Compression

This originally was the code for the paper "Interactive Deep Colorization and Its Application for Image Compression" by Xiao et al.
We expanded it with a few methods to choose the local cue points. Including a method, that worked better than the approach used by Xiao et al.

We also added a GAN Compression system by Agustsson et al., instead of the original BPG compression.

Point Picking Methods

Getting Started

Prerequisites

• A Linux of choice, for the feeling of superiority.
• NVIDIA GPU for training, or CPU for running only with at least ~8 GB of RAM
• Tensorflow 2.X + python

Pretrained models

The pretrained models are available here: models
They were trained on ~100.000 images of the ImageNet train set cropped to 256x256.

Preparing for training

The file preprocess.py will preprocess the images used for training. For both the colorization and compression.

Adjust the code in main() to your needs before running.
Call dataset_prepare(set="train", max_img=100_000) with the number of images you want to use of this set. This will symlink random images into res/set/original_img.
Note: This code was written with the ImageNet Dataset folder structure in mind. For other datasets, you might need to adjust the code in other places as well.
preprocess_color and preprocess_grayscale will generate the required data for training, but also for some recolorization and compression steps. You can choose what to generate and whether to overwrite as well. The random crop uses the filename as a seed, to produce the same crop for color and grayscale images.

Run

TODO

Training

TODO