STN-OCR: A single Neural Network for Text Detection and Text Recognition
This repository contains the code for the paper: STN-OCR: A single Neural Network for Text Detection and Text Recognition
Please note that we refined our approach and released new source code. You can find the code here
Please use the new code, if you want to experiment with FSNS like data and our approach. It should also be easy to redo the text recognition experiments with the new code, although we did not release any code for that.
Structure of the repository
The folder datasets contains code related to datasets used in the paper. datasets/svhn contains several scripts that can be used to create svhn based ground truth files as used in our experiments reported in section 4.2., please see the readme in this folder on how to use the scripts. datasets/fsns contains scripts that can be used to first download the fsns dataset, second extract the images from the downloaded files and third restructure the contained gt files.
The folder mxnet contains all code used for training our networks.
Installation
In order to use the code you will need the following software environment:
- Install
python3(the code might work with python2, too, but this is untested) - it might be a good idea to use a
virtualenv - install all requirements with
pip install -r requirements.txt - clone and install
warp-ctcfrom here - go into the folder
mxnet/metrics/ctcand runpython setup.py build_ext --inplace - clone the mxnet repository
- checkout the tag
v0.9.3 - add the
warpctcplugin to the project by enabling it in the fileconfig.mk - compile mxnet
- install the python bindings of mxnet
- You should be ready to go!
Training
You can use this code to train models for three different tasks.
SVHN House Number Recognition
The file train_svhn.py is the entry point for training a network using our purpose build svhn datasets. The file as such is ready to train a network capable of finding a single house number placed randomly on an image.
Example: 
In order to do this, you need to follow these steps:
-
Download the datasets
-
Locate the folder
generated/centered -
open
train.csvand adapt the paths of all images to the path on your machine (do the same withvalid.csv) -
make sure to prepare your environment as described in installation
-
start the training by issuing the following command:
python train_svhn.py <path to train.csv> <path to valid.csv> --gpus <gpu id you want to use> --log-dir <where to save the logs> -b <batch size you want ot use> --lr 1e-5 --zoom 0.5 --char-map datasets/svhn/svhn_char_map.json -
Wait and enjoy.
If you want to do experiments on more challenging images you might need to update some parts of the code in train_svhn.py. The parts you might want to update are located around line 40 in this file. Here you can change the max. number of house numbers in the image (num_timesteps), the maximum number of characters per house number (labels_per_timestep), the number of rnn layers to use for predicting the localization num_rnn_layers and whether to use a blstm for predicting the localization or not use_blstm.
A quite more challenging dataset is contained in the folder medium_two_digits, or medium in the datasets folder. Example: 
If you want to follow our experiments with svhn numbers placed in a regular grid you'll need to do the following:
- Download the datasets
- Locate the folder
generated/easy - open
train.csvand adapt the paths of all images to the path on your machine (do the same withvalid.csv) - set
num_timestepsandlabels_per_timestepto 4 intrain_svhn.py - start the training using the following command:
python train_svhn.py <path to train.csv> <path to valid.csv> --gpus <gpu id you want to use> --log-dir <where to save the logs> -b <batch size you want ot use> --lr 1e-5 - If you are lucky it will work ;)
Text Recognition
Following our text recognition experiments might be a little difficult, because we can not offer the entire dataset used by us. But it is possible to perform the experiments based on the Synth-90k dataset provided by Jaderberg et al. here. After downloading and extracting this file you'll need to adapt the groundtruth file provided with this dataset to fit to the format used by our code. Our format is quite easy. You need to create a csv file with tabular separated values. The first column is the absolute path to the image and the rest of the line are the labels corresponding to this image.
To train the network you can use the train_text_recognition.py script. You can start this script in a similar manner to the train_svhn.py script.
FSNS
In order to redo our experiments on the FSNS dataset you need to perform the following steps:
-
Download the fsns dataset using the
download_fsns.pyscript located indatasets/fsns -
Extract the individual images using the
tfrecord_to_image.pyscript located indatasets/fsns/tfrecord_utils(you will need to install tensorflow for doing that) -
Use the
transform_gt.pyscript to transform the original fsns groundtruth, which is based on a single line to a groundtruth containing labels for each word individually. A possible usage of thetransform_gt.pyscript could look like this:python transform_gt.py <path to original gt> datasets/fsns/fsns_char_map.json <path to gt that shall be generated> -
Because MXNet expects the blank label to be
0for the training with CTC Loss, you have to use theswap_classes.pyscript indatasets/fsnsand swap the class forspaceandblankin the gt, by issuing:python swap_classes.py <original gt> <swapped gt> 0 133 -
After performing these steps you should be able to run the training by issuing:
python train_fsns.py <path to generated train gt> <path to generated validation gt> --char-map datases/fsns/fsns_char_map.json --blank-label 0
Observing the Training Progress
We've added a nice script that makes it possible to see how well the network performs at every step of the training. This progress is normally plotted to disk for each iteration and can later on be used to create animations of the train progress (you can use the create_gif.py and create_video.py scripts located in mxnet/utils for this purpose). Besides this normal plotting to disk it is also possible to directly see this progress while the training is running. In order to see this you have to do the following:
-
start the
show_progress.pyscript inmxnet/utils -
start the training with the following additional command line params:
--send-bboxes --ip <localhost, or remote ip if you are working on a remote machine> --port <the port the show_progress.py script is running on (default is 1337) -
enjoy!
This tool is especially helpful in determining whether the network is learning anything or not. We recommend that you always use this tool while training.
Evaluation
If you want to evaluate already trained models you can use the evaluation scripts provided in the mxnet folder. For evaluating a model you need to do the following:
-
train or download a model
-
choose the correct evaluation script an adapt it, if necessary (take care in case you are fiddling around with the amount of timesteps and number of RNN layers)
-
Get the dataset you want to evaluate the model on and adapt the groundtruth file to fit the format expected by our software. The format expected by our software is defined as a
csv(tab separated) file that looks like that:<absolute path to image> \t <numerical labels each label separated from the other by \t> -
run the chosen evaluation script like so
python eval_<type>_model.py <path to model dir>/<prefix of model file> <number of epoch to test> <path to evaluation gt> <path to char map>
You can use eval_svhn_model.py for evaluating a model trained with CTC on the original svhn dataset, the eval_text_recognition_model.py script for evaluating a model trained for text recognition, and the eval_fsns_model.py for evaluating a model trained on the FSNS dataset.
License
This Code is licensed under the GPLv3 license. Please see further details in LICENSE.md.
Citation
If you are using this Code please cite the following publication:
@article{bartz2017stn,
title={STN-OCR: A single Neural Network for Text Detection and Text Recognition},
author={Bartz, Christian and Yang, Haojin and Meinel, Christoph},
journal={arXiv preprint arXiv:1707.08831},
year={2017}
}
A short note on code quality
The code contains a huge amount of workarounds around MXNet, as we were not able to find any easier way to do what we wanted to do. If you know a better way, pease let us know, as we would like to have code that is better understandable, as now.
I am getting the above warning while running "make"
Can you please help me with this
444 Dec 30, 2022
99 Nov 1, 2022
27.5k Jan 8, 2023
684 Jan 6, 2023
4 Jul 11, 2022
5 Dec 6, 2021
354 Dec 12, 2022
90 Dec 22, 2022
18 Jan 27, 2022
562 Jan 3, 2023
27 Jan 8, 2023
2.4k Jan 8, 2023
24 Apr 28, 2022
21 Dec 8, 2021
363 Jan 3, 2023
1.5k Dec 28, 2022
174 Dec 31, 2022
1.9k Dec 28, 2022
144 Jan 5, 2023