I want to train my own model to detect and recognize ID card with Tesseract. I want to extract the key information like name, id from it. The data looks like: [sample of data]
The introduction of training can only input text with single line.I'm confused how to train the detection model in Tesseract and should I label single character or label the whole text line in each box. (https://github.com/tesseract-ocr/tesstrain) enter image description here
1 by One Character Replacement from image to text is based on training in groups.
so here in the first tesseract training test sample, the idea is to let tesseract understand that the ch ligature is to be output as two letters the δ is to be lower case d with f as k and that Uber is Aber etc.
However that does not correct spelling of words without a dictionary of accepted character permutations and thus you need to either train all words you could expect like 123 is allowed but not 321 or else you allow all numbers.
The problem then is should ¦ be i | l or 1 ! ? and only human intelligent context is likely to agree what is 100% correct, especially when italics so is / = i | l or 1 ! or is it italic / ?
The clearer the characters are compared in contrast to the background, is usually going to produce the best result, and well defined void space within a character will help to distinguish well between B and 8 thus resolution is also a help or hindrance.
= INT 3O 80 S~A MARIA
A dictionary entry of BO and STA would possibly help in this case.
Oh, I think I get it. Tesseract doesn't need a detection model to get the position of the text line, it recognize each blob(letter) and uses the position of each letter to locate the text line.
