- Data Driven Approach for Compound Figure Separation Using Convolutional Neural Networks
- This repository contains an implementation of compound figure separator using a covolutional neral network (CNN).
- tensorflow 1.0 https://www.tensorflow.org
- opencv 3 http://opencv.org
If you are new, I strongly recoomend Anaconda and then install tensorflow and opencv.
pip install tensorflow
#pip instal tensorflow-gpu #in case you want to use GPU
conda uninstall -c menpo opencv #in case you have opnecv2
conda install -c menpo opencv3
If you find this tool useful, please consider to cite:
@inproceedings{figure-separate,
title={{A Data Driven Approach for Compound Figure Separation Using Convolutional Neural Networks}},
author={Satoshi Tsutsui, and David Crandall},
booktitle={{The IAPR International Conference on Document Analysis and Recognition (ICDAR)}},
year={2017}
}
Let's donwload the pretrained model at the ./data directory. I uploaded onto multiple places.
Google Drive: https://drive.google.com/open?id=0B046sNk0DhCDems2am5YV3NLeDQ
Dropbox: https://www.dropbox.com/s/xug7uw1rrq7ljy0/figure-sepration-model-submitted-544.pb?dl=0
Sure, just use this command:
python main.py --images ./imgs --annotate 1
See the results directory. You have sub-figures at ./results . That's it!
Here is other options:
--images IMAGES the directory that has figures
--model MODEL model pb file. Default is ./data/figure-sepration-model-submitted-544.pb
--thresh THRESH sub-figuere detection threshold. Default is 0.5
--output OUTPUT output directory Default is ./results
--annotate ANNOTATE save annotation to the image or not. 1 is yes, 0 is no. Default is 0.
Output json is somethine like:
[
{
"x": (x coordinate of left top point of the sub-figure),
"y": (y coordinate of left top point of the sub-figure),
"w": (width of the sub-figure),
"h": (height of the sub-figure),
"conf": (confidence value of the extaction),
} ,....
]
Sure. You can use FigureSeparator class. See simple_example.py which is:
from code.FigureSeparator import FigureSeparator
fig_separator=FigureSeparator("./data/figure-sepration-model-submitted-544.pb")
sub_figures=fig_separator.extract("./imgs/PMC4076561-Figure5-1.png")
print(sub_figures)
Giving you the bounding box of sub-figures with confidence values.
The training is done by darknet and then ported to tensorflow model using darkflow. Part of the code is re-used from darkflow. Thank you very much for the authors of these two repositories.