3 ms·
Very nice. Does this work with screen fonts / pixel perfect fonts too? That'd be my biggest gripe with all the OCR tools: what looks like the simplest fonts of
by TacticalCoder 5y ago
Very nice. Does this work with screen fonts / pixel perfect fonts too? That'd be my biggest gripe with all the OCR tools: what looks like the simplest fonts of them all to OCR are usually the one the less well detected.
For example using tesseract on Linux and trying to OCR the "terminus" font I get better result by first resizing the screenshot to something bigger (and blurry) and even then it's far from perfect OCR'ing. When in the first place it's a pixel perfect font...
(and, yes, there are cases where OCR'ing screen fonts make sense)
- schappim 5y agoI actually use another version with Google Vision (it is in a whole other class to on device OCR), and it is quite challenging to make Google Vision not work. For example it even works for hand written text.
- TacticalCoder 5y agoOh interesting, I should check that out.
- derefr 5y agoI’ve always wondered whether you couldn’t build a not-fully-ML app for screen-font recognition, that just takes a bunch of fonts, renders out every individual glyph at every size, trims them, converts them to an alpha mask, and then generates an indexible image fingerprint of said alpha mask. The OCR software would then just need to be smart enough to recognize “things that look like glyphs”, and put bounding boxes around them; and everything from there could be implemented in logic, rather than a model. (Just apply the same transforms to the thing in the bounding box, and then search the fingerprint DB.)