05-31-2022, 12:30 PM
I'm still looking for good FLOSS HWR, possibly https://github.com/githubharald/SimpleHTR , and so far tinkered with OCR e.g Tesseract. It's terrible https://twitter.com/utopiah/status/1321152581717078017 but feasible even on less powerful hardware.
On "boxier" boxes I had similar discussion on Inkscape and there is some work e.g https://inkscape.org/~qpad/%E2%98%85shape-recognition which ironically enough does not demand anything fancy, namely that if you have a certain type of curve, basically something close enough according to your threshold, to a line then you remove the number of points. If you do it properly it will flatten it. You can also go the ML way with shape recognition from a catalogue of shape though or mix the two.
On "boxier" boxes I had similar discussion on Inkscape and there is some work e.g https://inkscape.org/~qpad/%E2%98%85shape-recognition which ironically enough does not demand anything fancy, namely that if you have a certain type of curve, basically something close enough according to your threshold, to a line then you remove the number of points. If you do it properly it will flatten it. You can also go the ML way with shape recognition from a catalogue of shape though or mix the two.
