Support output of word-level bounding boxes

[timwedde]

First of all: Great project! I've developed a very similar thing for a project at work some years ago and ran into very similar issues as I, too, was using Python at that time.

One question/request though: As far as I can tell you only output paragraph/line level bounding boxes right now. This is fine, but in many cases a higher 'resolution' is preferable. As such it would be nice if this tool could output word-level bounding boxes as well, either by default or via an additional switch.

The current implementation stops at the line level.

I've implemented this exact thing (using Apple OCR) in Python in an old package of mine here for reference. Unfortunately Apple doesn't provide this functionality out of the box, but with a tiny bit of regex it's possible to approximate this with reasonable accuracy. There is a boundingBox(for: range) method in the SDK that helps with disassembling each line.

This project is basically what I dreamt of doing but never found the time to accomplish, so excellent job on that front :)

[AmineDiro]

Thank you for your interest for this project and love to see the appreciation for it !
Regarding the word lvl bbox, I think you're right. In ferrules I am basically merging OCR and nonOCR extracted text at line level. Then merging lines into blocks using the layout information and additional heuristics.

I think I'll open up an issues to keep line information in the parsed blocks first then keep char/span level information about each char. This could be useful for benchmarks for example.