Tesseract font matching

[vroscigno]

I am using Tesseract 3.04 on Windows to analyze scanned paper forms which often contain non-contiguous text labels of various size, position, and font style. I am attempting to deduce simple typeface characteristics such as serif vs sans-serif, fixed vs variable pitch, italics, bold, etc, in an effort to loosely classify identified text labels.

I started out by using LTRResultIterator::WordFontAttributes for recognized words, but then learned that the returned font properties are from the *best* matching font, not from an accumulation of actual character attributes for the recognized word.

As an example of this, I have observed cases where sequences of ARIAL (sans-serif, variable-pitch) characters are measured and determined to be fixed-pitch (for example: "BOOK"), and the best matching font is a COURIER variant (fixed-pitch, serif). In this case, none of the characters have serifs, but the determined pitch (fixed) seems to carry significant weight when matching fonts.

I intend to study the font classification logic a bit to be sure I understand it.

I also suspect that the Adaptive Classifier may propagate this effect for 'downstream' results. (True/False? Opinions?)

I thought about exploring the following:

1. disabling fixed-pitch character classification and handling

2. disabling the adaptive classifier or limiting it's influence

Does anyone have any suggestions or opinions?

Thanks,

Vince Roscigno