This repository is meant to be a sandboxed environment from the main Story Squad data science back-end in order to explore various processes that take place with Tesseract-OCR. Tesseract is a very complex entity and we have attempted to break down what this repository is for, what it contains and how to understand this technology and what it's doing. Once an ideal model is reached, the final <model>.traineddata file and the Tesseract.py Python script can be transferred to story-squad-ds-a to replace Google Vision endpoints.
Tesseract Quick Start Guide (more general approach)
Tesseract Onboarding Guide for Labs33
Actual Tesseract Documentation (if you're feeling adventurous)
fonts_n :: Folder containing font files (.ttf) used in synthetic data generation for training Tesseract.
langdata_lstm :: Folder containing all components of every langauge that has already been trained using Tesseract. Not used much during training but necessary for Tesseract to run.
update :: Folder used for all preprocessing steps required before training. Add a data folder that can be filled with any obtained (jpg) training images. Run the conversion.sh script first to convert all images to png file format. Next, run processing.py with Python to create and save each processed image. Finally, run the segmentation.sh script to segment each image, have it saved as a tif and generate it's ground truth text file (that will then need to be corrected by hand). Be sure that any preprocessing techniques that are added to the processing.py are added to the preprocessing function in Tesseract.py as well.
eng.traineddata :: This is the fully trained Tesseract model for the English language and needs to be placed into the Tessie_True/tesseract/tessdata directory.
storysquadset.traineddata & storysquadnew.traineddata :: These are our trained model versions so far storysquadnew being the newest model that you will pick up from. These also need to be stored in the Tessie_True/tesseract/tessdata directory.
train.sh :: This shell script runs through the entire Tesseract training process using image data rather than synthetic data.
Tesseract.py :: Python program containing a transcribe function that will eventually be used to replace the Google Vision transcribe function from story-squad-ds-a/app/api/submission.py at ~ line 55. This transcribe function calls on the preprocessing function as well so the only function that will need to be imported into submission.py is Tesseract.transcribe().
The following shell scripts are used to train Tesseract with sythetic data (fonts from above) and must be executed in order if training from sythentic data: 1gen_txt_data, 2extract_lstm, 3eval, 4finetune, 5combine