StackMix and Blot Augmentations for Handwritten Recognition using CTCLoss

This paper presents a new text generation method StackMix. StackMix can be applied to the standalone task of generating handwritten text based on printed text.

Config file

Create a new config file need in configs/__init__.py. An individual config file is required for each dataset.

Datasets

There are two ways to get a dataset:

The first way:

Download a dataset and annotations (for example Bentham: http://www.transcriptorium.eu/~tsdata/BenthamR0/). Prepare dataset and create marking.csv file

The second way:

Downlad prepared dataset using the script download_dataset.py (for example Bentham: python scripts/download_dataset.py --dataset_name=bentham) And now you can use train script.

You can change out folder by key --datadir='your path', by default --datadir=../StackMix-OCR-DATA. All dataset names: bentham, peter, hkr, iam.

Dataset format

The dataset should contain a directory with images and a csv file marking.csv with annotations. The csv file must contain the "stage" field with information about which sample the image belongs to (train / valid / test). An example of the structure and content of a csv file is given below sample_id,path,stage,text 270-01,washington/images/270-01.png,train,"270. Letters, Orders and Instructions. October 1755." 270-03,washington/images/270-03.png,train,"only for the publick use, unless by particu-" 270-04,washington/images/270-04.png,train,lar Orders from me. You are to send 270-05,washington/images/270-05.png,train,"down a Barrel of Flints with the Arms, to"

Setup

install requirements: pip install -r requirements.txt -f https://download.pytorch.org/whl/torch_stable.html

Example run train

python scripts/run_train.py \ --checkpoint_path "" \ --experiment_name "HKR_exp" \ --dataset_name "hkr" \ --data_dir "data/HKR/" \ --output_dir "exp/" \ --experiment_description \ "[Base] Training Base OCR on HKR dataset" \ --image_w 512 \ --image_h 64 \ --num_epochs 200 \ --bs 256 \ --num_workers 8 \ --use_blot 0 \ --use_augs 1 \ --use_progress_bar 0 \ --seed 6955

Example run evaluation

python scripts/run_evaluation.py \ --experiment_folder "exp/" \ --dataset_name "hkr" \ --data_dir "data/HKR/" \ --image_w 512 \ --image_h 64 \ --bs 128

Generating allcharmasks.json for stackmix

python scripts/prepare_char_masks.py \ --checkpoint_path "exp/HKR_exp/best_cer.pt" \ --data_dir "data/HKR/" \ --dataset_name "hkr" \ --image_w 512 \ --image_h 64 \ --bs 128 \ --num_workers 8

dataset_name must be taken from the config

Image generation

The code for generation can be found here

Example of generating images with stackmix

Supported by:

• Sber
• OCRV
• Sirius University
• RZHD

Citation

Please cite the related works in your publications if it helps your research: @inproceedings{10.1145/3476887.3476892, author = {Mark, Potanin and Denis, Dimitrov and Alex, Shonenkov and Vladimir, Bataev and Denis, Karachev and Maxim, Novopoltsev and Andrey, Chertok}, title = {Digital Peter: New Dataset, Competition and Handwriting Recognition Methods}, year = {2021}, isbn = {9781450386906}, publisher = {Association for Computing Machinery}, address = {New York, NY, USA}, url = {https://doi.org/10.1145/3476887.3476892}, doi = {10.1145/3476887.3476892}, abstract = { This paper presents a new dataset of Peter the Great’s manuscripts and describes a segmentation procedure that converts initial images of documents into lines. This new dataset may be useful for researchers to train handwriting text recognition models as a benchmark when comparing different models. It consists of 9694 images and text files corresponding to different lines in historical documents. The open machine learning competition ”Digital Peter” was held based on the considered dataset. The baseline solution for this competition and advanced methods on handwritten text recognition are described in the article. The full dataset and all codes are publicly available.}, booktitle = {The 6th International Workshop on Historical Document Imaging and Processing}, pages = {43–48}, numpages = {6}, keywords = {handwritten text recognition, Digital Peter, historical dataset, Russian}, location = {Lausanne, Switzerland}, series = {HIP '21} }

Contacts

• A. Shonenkov shonenkov@phystech.edu
• D. Karachev
• M. Novopoltsev
• D. Dimitrov
• M. Potanin