This is a fork of the TextRecognitionDataGenerator.
The fork is used for generating syntehic data for training the excellent docTR OCR engine.
I would use the original lib, but some dependencies seems outdated and I could not get it to work.
I just updated the dependency arabic-reshaper to 3.0.0 and it seems to work for my purpose.
If TextRecognitionDataGenerator is updated I will switch back to the original lib.
git clone https://github.com/diversen/TextRecognitionDataGenerator.git
cd TextRecognitionDataGenerator
python3 -m venv venv
source venv/bin/activate
pip install -r requirements.txt
You may edit the settings.py file to change the font and font_size used for generating the images.
You may also alter the fonts in the fonts folder.
There is also a bunch of other settings you may alter.
This will generate at least 1000 unique words in a sqlite3 database.db placed in train-data.
The genereated words will use only characters that are in the doctr vocab danish.
The words will be generated from the da wikipedia.
Words generated will be added to the words table in the database.db file.
python generate-words.py --num-words 1000 --output-dir train-data --vocab danish --lang da
In the train-data/images generate 1000 x 2 images using da words. The generated images will also be added
to the labels table in the database.db file.
python generate-img.py --num-words 1000 --num-images-per-word 2 --output-dir train-data --lang da
Extract the labels from the database.db file and write them to the train-data/labels.json file.
python generate-labels.py --output-dir train-data --lang da
If you have followed the above steps you should have a train-data folder with images and labels,
which can be given as input to the docTR OCR training script.
In practice you will have to generate many more images and labels. Maybe 250000 words and e.g. 8 images per word.
Likewise you would generate a folder with validation-data for validating the training.
The easiest way to train is to use the docTR OCR training scripts.
In the following there is instructions for pytorch training.
Clone the docTR repo and run the training script.
git clone https://github.com/mindee/doctr
cd doctr
python3 -m venv venv
source venv/bin/activate
pip install "python-doctr[tf]"
pip install -r references/requirements.txt
python references/recognition/train_pytorch.py crnn_vgg16_bn --vocab danish --train_path train-data --val_path validation-data --epochs 5
python references/recognition/train_pytorch.py crnn_vgg16_bn --max-chars 32 --vocab danish --train_path train-data --val_path validation-data --epochs 5 --pretrained
python references/recognition/train_pytorch.py crnn_vgg16_bn --max-chars 32 --vocab danish --train_path train-data --val_path validation-data --epochs 5 --pretrained --resume crnn_vgg16_bn_20240316-233300.pt
python references/recognition/train_pytorch.py crnn_vgg16_bn --vocab danish --train_path train-data --val_path validation-data --epochs 1 --resume /home/dennis/.cache/doctr/models/crnn_vgg16_bn-9762b0b0.pt
python references/recognition/train_pytorch.py crnn_vgg16_bn --max-chars 32 --vocab danish --train_path train-data --val_path validation-data --epochs 1 --pretrained --resume crnn_vgg16_bn_20240317-095746.pt --push-to-hub --name doctr-torch-crnn_vgg16_bn-danish-v1
python convert.py