Training Tesseract 4 on real images

[Sim Tov]

Hello,

I would like to train Tesseract 4 to recognize certain scripts/languages based on real images rather than synthetic ones. Here are my questions:

1. Is there a tool, preferably cross-platform (Windows/Linux) GUI, that assists in creating .box file based on scanned images? How to get coordinates of textlines? etc...

2. Is there a youtube/video tutorial describing .tiff/.box files preparation based on real scans?

3. What provides better recognition - training on real images or training on synthetic images?

4. How many textlines of real scans do I need to get proper recognition?

Thank you very much!

ST

[Murtuza Dahodwala]

I also want to know that how we can train on real images which are not single lines?