Project MANGA PSIML2017

This repo contains a project that we created during the Petnica Summer Institue: Machine Learning seminar. For intro to the project I suggest that you look at this blog post.

UNET without hint

UNET with hint

Datasets

Dataset images can be loaded or patched from a giant montage of all images.

• Manga546 - 546 images of color and bw - size 950x640

  Created out of ScottPilgrim comics and is not openly available, but it we weren't able to make much use of it.

• MangaOnline - 42215 images of color - size 512x512

  Consists of anime/manga images and can be downloaded from the safebooru.org website using the script that is created by @kvfrans and provided in the master branch.

Ideas

Every list entry is an idea implemented on the branch of the same name

• UNET - U-net autoencoder on patched Manga546
• UNET-LOAD - U-net autoencoder on loaded Manga546 (whole images are loaded)
• UNET-ONLINE - U-net autoencoder on loaded MangaOnline
• UNET-ONLINE-HINT - U-net autoencoder trained with image hints on loaded MangaOnline
• simple-segmentated - Few convolutional layers on patched Manga546
• simple-conv - Few convolutional layers on patched Manga546

Installation

pip3 install opencv-python tensorflow keras h5py scikit-image
git clone https://github.com/dulex123/manga

# This is our best model, you can use any model/branch
git checkout UNET-ONLINE-HINT

Usage

In order to run prediction on a single image you can use predict.py files while the training can be executed by running the branch-name.py file. The best results are on the UNET-ONLINE-HINT and UNET-ONLINE branches

python3 predict.py img/to/predict.ext

For training datasets should be downloaded at the following locations:

• data/MangaOnline/train and /data/MangaOnline/test

• data/Manga546/

Credits

Released under the MIT License.
Authored and maintained by Dušan Josipović & Nikola Jovičić.

Blog nikola-j.github.io  ·  Github @nikola-j

Blog dulex123.github.io  ·  GitHub @dulex123  ·  Twitter @josipovicd