U-Net-bin: DIBCO 2017 Submission

This page is concerned to our submission that was ranked first in the Document Image Binarization Contest 2017 (DIBCO). It is a convolutional neural network (CNN) based method which uses U-Net architecture. A docker container with the network along with all the supplementary data we used in the training process has been released at ftp://smartengines.com/unetbin. Folders printed and handwritten contain grouped images from the previous DIBCO contests. Ground truth prepared by organizers is located in folders with _gt suffix. In the aux folder there are images we have additionally selected and used during the learning procedure.

Prerequisites

Install the latest version of Docker.

Usage

  • Download the docker image
wget ftp://smartengines.com/unetbin/dibco_06-30.tar
  • Unpack the docker image
docker load <dibco_06-30.tar
  • Start the running image
docker run -dt --name dibco -v <local-folder-name>:/mnt/volume dibco:06-30 /bin/bash
  • Execute
docker exec dibco /root/environment/bin/python /root/evaluate_answer.py -i <input-image-name> -o <output-image-name>