PageNet is a Deep Learning system that takes in an image with a document in it and returns a quadrilateral representing the main page region. We trained PageNet using the library Caffe. For details, see our paper.
There are three scripts in this repo. One for training networks, one for predictions using pre-trained networks, and one for rendered quadrilateral regions.
We have provided two pretrained models from our paper. One model is trained on the CBAD dataset and the other is trained on a private collection of Ohio Death Records provided by Family Search.
test_pretrained.py
has the following usage
usage: test_pretrained.py [-h] [--out-dir OUT_DIR] [--gpu GPU]
[--print-count PRINT_COUNT]
image_dir manifest model out_file
Outputs binary predictions
positional arguments:
image_dir The directory where images are stored
manifest txt file listing images relative to image_dir
model [cbad|ohio]
out_file Output file
optional arguments:
-h, --help show this help message and exit
--out-dir OUT_DIR
--gpu GPU GPU to use for running the network
--print-count PRINT_COUNT
Print interval
image_dir
is the directory containing images to predict. The file paths listed in manifest
are relative to image_dir
and are listed one per line. model
should be either cbad
or ohio
to select which trained model to use. out_file
will list the coordinates of the quadrilaterals predicted by PageNet for each of the input images.
--gpu
is for passing the device ID of the GPU to use. If it is negative, CPU mode is used. Specifying --out-dir
will allow you to dump both the raw and post processed predictions as images.
train.py
has the following usage
usage: train.py [-h] [--gpu GPU] [-m MEAN] [-s SCALE] [-b BATCH_SIZE] [-c]
[--image-size IMAGE_SIZE] [--gt-interval GT_INTERVAL]
[--min-interval MIN_INTERVAL] [--debug-dir DEBUG_DIR]
[--print-count PRINT_COUNT]
solver_file dataset_dir train_manifest val_manifest
Outputs binary predictions
positional arguments:
solver_file The solver.prototxt
dataset_dir The dataset to be evaluated
train_manifest txt file listing images to train on
val_manifest txt file listing images for validation
optional arguments:
-h, --help show this help message and exit
--gpu GPU GPU to use for running the network
-m MEAN, --mean MEAN Mean value for data preprocessing
-s SCALE, --scale SCALE
Optional pixel scale factor
-b BATCH_SIZE, --batch-size BATCH_SIZE
Training batch size
-c, --color Training batch size
--image-size IMAGE_SIZE
Size of images for input to training/prediction
--gt-interval GT_INTERVAL
Interval for Debug
--min-interval MIN_INTERVAL
Miniumum iteration for Debug
--debug-dir DEBUG_DIR
Dump images for debugging
--print-count PRINT_COUNT
How often to print progress
solver_file
points to a caffe solver.prototxt file. Such a file is included in the repo. The training script expects that the network used for training to begin and end like the included train_val.prototxt
file, but the middle layers can be changed.
dataset_dir
is the directory containing the training and validation images. The file paths listed in train_manifest
and val_manifest
are relative to dataset_dir
and are listed one per line.
--gpu
is for passing the device ID of the GPU to use. If it is negative, CPU mode is used. --debug-dir
defaults to debug
and if it is not the empty string, predictions and metrics will be dumped at intervals specified by --gt-interval
and --min-interval
. This can help with selecting the best model from the snapshots.
The optional arguments have reasonable defaults. If you're curious about their exact meaning, I suggest you look at the code.
If you have trained your own model with train.py
, you can test it with test.py
. The usage is
usage: test.py [-h] [--out-dir OUT_DIR] [--gpu GPU] [-c] [-m MEAN] [-s SCALE]
[--image-size IMAGE_SIZE] [--print-count PRINT_COUNT]
net_file weight_file dataset_dir test_manifest out_file
Outputs binary predictions
positional arguments:
net_file The deploy.prototxt
weight_file The .caffemodel
dataset_dir The dataset to be evaluated
test_manifest Images to predict
out_file output file listing quad regions
optional arguments:
-h, --help show this help message and exit
--out-dir OUT_DIR Dump images
--gpu GPU GPU to use for running the network
-c, --color Training batch size
-m MEAN, --mean MEAN Mean value for data preprocessing
-s SCALE, --scale SCALE
Optional pixel scale factor
--image-size IMAGE_SIZE
Size of images for input to prediction
--print-count PRINT_COUNT
Print interval
The optional arguments for this script mirror those for train.py
and should be set to the same values. The required arguments are the same as for test_pretrained.py
, except you manually specify network file
(e.g., train_val.prototxt
) and the weight_file
.
The usage for render_quads.py
is
python render_quads.py manifest dataset_dir out_dir
manifest
lists the image file path and quadrilateral coordinates. It should be the out_file
of test_pretrained.py
. The filepaths in manifest
are relative to dataset_dir
. out_dir
is an output directory where quadrilateral region images are written
The python scripts depend on OpenCV 3.2, Matplotlib, Numpy, and Caffe.
For those who don't want to install the dependencies, I have created a docker image to run this code. You must have the nvidia-docker plugin installed to use it though you can still run our models on CPU (not recommended).
The usage for the docker container is
nvidia-docker run -v $HOST_WORK_DIRECTORY:/data tensmeyerc/icdar2017:pagenet python $SCRIPT $ARGS
$HOST_WORK_DIRECTORY
is a directory on your machine that is mounted on /data
inside of the docker container (using -v). It's the only way to expose files to the docker container.
$SCRIPT
is one of the scripts described above. $ARGS
are the normal arguments you pass to the python script. Note that any file paths passed as arguments must begin with /data
to be visible to the docker container.
There is no need to download the container ahead of time. If you have docker and nvidia-docker installed, running the above commands will pull the docker image (~2GB) if it has not been previously pulled.
For those wanting to replicate the baselines we reported in our paper, we've included the two scripts we used.
baselines/grabCutCropEval.py
evaluates using the Grab Cut approach described in the paper. As input arguments, it takes the groudtruth file, image directory, directory to store intermediate results, and the number of threads to run the script on. It produces a file which is the ground truth file name with _fullgrab.res
appended.
python grabCutCropEval.py gtFile imageDir interDir numThreads
baselines/noCutMeanCutCropEval.py
evaulates using full image (no cropping) and a mean quadrilateral. The mean quadrilateral is computed from the ground truth if not given in the arguments. It prints the computed mean so it can be reused. As input arguments, the script takes the ground truth file, image directory, and optionally the four (normalized 0.0 - 1.0) quadrilateral points (in clockwise order, starting at top left). It produces two files which are the ground truth file name with _fullno.res
and _fullmean.res
appended.
python noCutMeanCutCropEval.py gtFile imageDir [mean_x1 mean_y1 mean_x2 mean_y2 mean_x3 mean_y3 mean_x4 mean_y4]
If you find this code useful to your research, please cite our paper:
@article{tensmeyer2017_pagenet,
title={PageNet: Page Boundary Extraction in Historical Handwritten Documents},
author={Tensmeyer, Chris and Davis, Brian and Wigington, Curtis and Lee, Iain and Barrett, Bill},
journal={arXiv preprint arXiv:1709.01618},
year={2017},
}