A selectional auto-encoder approach for document image binarization

document-image-binarization

The code of this repository was used for the following publication. If
you find this code useful please cite our paper:

@article{Gallego2019,
title = "A selectional auto-encoder approach for document image binarization",
author = "Jorge Calvo-Zaragoza and Antonio-Javier Gallego",
journal = "Pattern Recognition",
volume = "86",
pages = "37 - 47",
year = "2019",
issn = "0031-3203",
doi = "https://doi.org/10.1016/j.patcog.2018.08.011"
}

Below we include instructions to binarize and reproduce the experiments.

Binarize

The binarize.py script performs the binarization of an input image using a trained model. The parameters of this script are the following:

(*) By default, the model trained with all datasets will be used.

The only mandatory parameter is -imgpath, the rest are optional. You also have to choose if you want to see a demo (--demo) or to save (-save) the binarized image.

For example, to binarize the image img01.png you can run the following command:

$ python binarize.py -imgpath img01.png -save out.png

If you want to binarize this image using the model trained for Dibco 2016 and the parameters specified in the paper, you have to run the following command:

$ python binarize.py -imgpath img01.png -modelpath MODELS/model_weights_dibco_6_256x256_s96_aug_m205_f64_k5_s2_se3_e200_b32_esp.h5 -w 256 -s 96 -f 64 -k 5 -stride 2 -th 0.5 --demo

The MODELS folder includes the following trained models for the datasets evaluated:

• MODELS/model_weights_all_None_256x256_s96_aug_m205_f64_k5_s2_se3_e200_b32_esp.h5
• MODELS/model_weights_dibco_5_256x256_s96_aug_m205_f64_k5_s2_se2_e200_b32_esp.h5
• MODELS/model_weights_dibco_6_256x256_s96_aug_m205_f64_k5_s2_se3_e200_b32_esp.h5
• MODELS/model_weights_ein_-1_256x256_s96_aug_m205_f64_k5_s2_se2_e200_b32_esp.h5
• MODELS/model_weights_palm_0_256x256_s96_aug_m205_f64_k5_s2_se2_e200_b32_esp.h5
• MODELS/model_weights_palm_1_256x256_s96_aug_m205_f64_k5_s2_se2_e200_b32_esp.h5
• MODELS/model_weights_phi_-1_256x256_s96_aug_m205_f64_k5_s2_se2_e200_b32_esp.h5
• MODELS/model_weights_sal_-1_256x256_s96_aug_m205_f64_k5_s2_se2_e200_b32_esp.h5

Train

The train.py script performs the training of the proposed network from a dataset of images and a series of input parameters that allow to configure both the training process and the network topology. This executable, in addition to training, performs the validation and saves the resulting images.

The parameters of this script are the following:

The only mandatory parameters are -path and -db. To run the training you must first indicate the path where the datasets are located (with the -path parameter) and the name of the dataset to evaluate (with the -db parameter). Depending on the name of the dataset the system will create the partitions for training and validation.

The folders of each dataset must have a specific name (see table in section "Datasets" or consult the source code). Within each folder there must be two subfolders, one with the suffix \_GR with the input images in grayscale, and another with the suffix \_GT for the ground truth.

Option --aug activates the data augmentation, in this case the system will load the images stored in the folder with prefix aug_.

For example, to train a network for the dataset Dibco 2016, using data augmentation and the parameters specified in the paper, you have to run the following command:

$ python -u train.py -path datasets -db dibco -dbp 6 --aug -w 256 -s 128 -f 64 -k 5 -e 200 -b 10 -th -1 -stride 2 -page 64

Once the script finishes, it will save the learned weights and the binarized images resulting from the validation partition in a folder with the prefix _PR- followed by the name of the model. It will also show the mean squared reconstruction error and the obtained F-m. We have also used the evaluation tool provided by DIBCO (http://vc.ee.duth.gr/h-dibco2016/benchmark/), which allowed us to validate the obtained F-m and also calculate the pseudo F-m.

Datasets

Below is a summary table of the datasets used in this work along with a link from which they can be downloaded:

The ground truth of Ensiedeln and Salzinnes is available at the following link:

https://forms.gle/nSi3BZz7iammgRXWA

GitHub

GitHub - ajgallego/document-image-binarization: A selectional auto-encoder approach for document image binarization

A selectional auto-encoder approach for document image binarization - GitHub - ajgallego/document-image-binarization: A selectional auto-encoder approach for document image binarization

GitHubajgallego