Tesseract Train (intended for train with Ground Truth)

Introduction

This project is configured tesstrain with all the necessary params to train your ground truth image lines to Tesseract OCR

Setup

First of create your tessdata dir using: mkdir -p tesstrain/usr/share/tessdata Into data folder you need to create your ground truth directory, this directory have to follow the rules below:

• 1: Let's suppose that you want to call your model as spa_001
• 2: You need to create a folder using your model name + -ground-truth and the result is the follow -> spa_001-ground-truth
• 3: This folder has to be created inside data folder: Eg: mkdir -p tesstrain/data/spa_001-ground-truth
• 4: Ground Truth files has to be sequential numbers like 001.jpeg, 001.gt.txt (This pair is required)
• 5: Inside data folder you have to create .numbers, .punc, .wordlist with your model name as prefix: Eg: spa_001.punc
• 6: If you want to continue from some start model, you have to move the start model traineddata file into tesstrain/usr/share/tessdata folder.

After all set, you have to start your training using

make training MODEL_NAME=spa_001 PSM=7 START_MODEL=spa

Where:

• MODEL_NAME is the name of your new model (we used it as spa_001)
• PSM is the Page Segmentation Mode (7 is for line data)
• START_MODEL is the model you want to continue training from (we used standard spa.traineddata from tesseract repositories)
  • You can found initial models in: https://github.com/tesseract-ocr/tessdata_best