Paper:
*****************************************************************************
DeepOtsu: Document Enhancement and Binarization using Iterative Deep Learning
Sheng He, Lambert Schomaker
Pattern Recognition, vol 91, July 2019.

Email: heshengxgd@gmail.com

*****************************************************************************

Please cite this paper if you use this code.


The code is on: http://www.ai.rug.nl/~sheng/DeepOtsu_tf.tar.gz


python binTestImprovedFusion.py --dataset directory_or_image/ --imgtype png --overlap 0.1 --multiscale True --num_block 6

--dataset: the directory of the input images
--imgtype: image type, such as png,jpg,tiff, etc.
--overlap: the overlap of crop small patches (default: 0.1) The larger of this value (maximum: 1.0), the smaller of the number of patches and the faster for running.
-- multiscale: whether use multiscale patches (default: True) Set it to False will fast the running.
--num_block: (number of UNets used in the program, default: 6), but 3 might also good. The smaller of this number, the faster of running.

The program actually runs on gray-scale images. If you input a color image, it will read as gray-scale image by the code.