Simple Tesseract 5.0 Training Docker

This image contains the bare minimum code to train the tesseract 5.0 lstm English model. The training requires:

• Train_data: A folder with the train dataset composed of .tiff images and their corresponding .box files.

• Validation_data: A folder with the validation dataset composed of .tiff images and their corresponding .box files.

Build and train

To build the image just do:

• docker build --build-arg UBUNTU_VERSION=20.04 -t tesseract_trainer:5.0 .

To fine tune your data execute:

•       docker run -u $UID:$UID \ 
                -v <host path to "Train_data">:/home/tesstrain/train_samples \   
                -v <host path to "Validation_data">:/home/tesstrain/validation_samples \
                -v <host path to output directory>:/home/tesstrain/output   \ 
                -it tesseract_trainer:5.0 \ 
                                        <"0" to process all .tiff and .box into .lstm or "1" to skip this if already done> \ 
                                        <The psm mode (default is 7) to detect text>\ 
                                        <Number of train iterations>
  

For example in linux:

docker run -u $UID:$UID -v /foo/bar/train_samples:/home/tesstrain/train_samples -v /foo/bar/validation_samples:/home/tesstrain/validation_samples -v /foo/bar/output:/home/tesstrain/output -it tesseract_trainer:5.0 0 7 10000

This will process the .tiff and .box files to generate the .lstm files using psm=7 and then train the best English model for 10000 iterations.

TODO

• Add example data with VIA.
• Add language selector.
• Add script to convert VIA box coordinates to .box files.
• Add more training parameters configuration.