Comicolorization

This is an implementation of the Comicolorization : Semi-automatic Manga Colorization.

With this repository, you can

• run sample codes
• train colorization task
• use as colorization library

Sample Code

We prepare two sample codes, sample_painting.py and sample_detecting_panels.py.

1. sample_painting.py is example of colorization the manga page.
2. sample_detecting_panels.py is example of detection auto panel rectangle.

First of all, install the requirements.

pip install -r requirements.txt

In addition, you must install OpenCV-Python.

Colorization

python sample/sample_painting.py

# Help
# python sample/sample_painting.py --help

Auto Panel Rectangle Detection

1. get manga-frame-extension and build

git submodule init
git submodule update
cd manga-frame-extraction/MangaFrameExtraction
cmake ./
make

Please read manga-frame-extraction's README.md for details.

2. run

cd ../../
python sample/sample_detecting_panels.py

# Help
# python sample/sample_detecting_panels.py --help

Copyright

Following images are from Manga109 dataset.

©Ishioka Shoei	©Sakurano Minene	©Sakurano Minene	©Tanaka Masato

Training

There are two training task.

1. the colorization task for generate the low resolution colorized image.
2. the super resolution task for generate the higher resolution colorized image.

First of all, install the requirements.

pip install -r requirements.txt

In addition, you must install OpenCV-Python.

Colorization Task

1. prepare dataset

For training, three type data are required.

• color images that Pillow can load
• json file written label ID list that has 'id' key
  • ex. { "id": [ "a", "b", "c" ] }
• json file written label ID list for each image
  • the key is image file name without extension
  • ex. { "imageX": ["a", "b"], "imageY": ["a"], "imageZ": ["c"], ... }

2. run

# run same as paper
python bin/train.py \
    /path/to/images/directory \
    /path/to/save \
    --path_tag_list /path/to/label_ID_list.json \
    --path_tag_list_each_image /path/to/label_ID_list_for_each_image.json \
    --network_model LTBC \
    --num_dataset_test 1000 \
    --batchsize 30 \
    --size_image 224 \
    --augmentation True \
    --size_image_augmentation 256 \
    --save_result_iteration 1000 \
    --random_seed_test 0 \
    --loss_type Lab \
    --alpha_ltbc_classification 0.00333333333 \
    --ltbc_classification_loss_function multi_label \
    --line_drawing_mode otsu_threshold \
    --max_pixel_drawing 15 \
    --use_adversarial_network \
    --blend_adversarial_generator 1.0 \
    --discriminator_first_pooling_size 2 \
    --optimizer_adam_alpha 0.0001 \
    --blend_mse_color 1.0 \
    --mse_loss_mode color_space \
    --weight_decay 0.0001 \
    --log_interval 200 \
    --gpu -1 \
    --ltbc_classification_num_output_list 512 428 \  # last number should be same as number of labels
    {color_feature}  # other params depending on task, please see below

# Help
# python bin/train.py --help

When palette mode, change color_feature to

    --threshold_histogram_palette 0.0 \
    --use_histogram_network \
    --num_bins_histogram 6 \
    --use_multidimensional_histogram

When histogram mode, change color_feature to

    --use_histogram_network \
    --num_bins_histogram 6 \
    --use_multidimensional_histogram

When without any color feature, color_feature is empty.

Super Resolution Task

1. modify config file

Modify the information with these keys in bin/config_super_resolution.json.

• dataset/images_glob: set the image file names
• model/other/path_result_directory: set the colorization model directory
• project/name: set the name of this task (this will be the directory name)
• project/result_path: set the path for the result directory

2. run

python bin/train_super_resolution.py bin/config_super_resolution.json

Use as colorization library

Install

pip install git+http://github.com/DwangoMediaVillage/Comicolorization

Download trained model

Please download model files at model directory in this repository.

How to use

1. import comicolorization and comicolorization_sr
2. initialize Drawer
3. load model
4. initialize PagePipeline
5. call process method

Please see the sample code for more details.

License

MIT License, see LICENSE.