Keras: 50,000 classes, really?

Question

Text recognising

There is a database of pictures with 50 thousand words and with different fonts. Recognizing through OCR is not realistic, at least in standard ways.
Previously, when using picture learning, the number of classes did not exceed 200 (because the set of pictures consisted of no more than 200 words). I generate pictures with words with different fonts, noises, etc. 1 word = 1 folder of pictures.
But now the task is to train 50 thousand. Do I understand correctly that there are the same number of classes for training? Or I do not understand something. Guru tell me.

Answer 1

words - dictionary or random character sets?
Usually recognized by letters, combining the result into a sequence. Text recognition is not a task of classification into N classes.
I would generally recommend not to take a steam bath, but to try the ready-made Google Cloud Vision API - the first 1000 recognitions are free, enough to try and make sure it works.