What is the best method for filter text from complex background with other text?

[Ivan1923stop]

Dear friends!

I have scanned documents (text printed in a large symbols (regular) on template with a some small size symbols (italic)).

My aim is pre-OCR filtering
(or OCR itself as the task is narrow enough - I have two well divide sets of symbols. Size ratio and italic\regular properties permanent enough).

I have got two files:

and made simpe comand:

magick.exe in.png mask.png -fx "(u|1-v)" out.png

and have got results:

There is a hole in symbol T, but I could fill it as I know shapes of all symbols I suppose.

My question is: What kind of Image Processing (pre OCR) I should use for extract large symbols only?
Would it be a simple IM substract mask filter (with auto scaling and rotating of course) or I need in Deep Learning neuro methods?