That said, ML is really most practical on a cloud system where it is learning from everyone in the world instead of just the corrections *I* make, which is essenially why Google OCR is so darned good. These days to improve recognition of those situations I'd use a little bit of ML and user correction feedback (ie, app scans as "Torn" and I correct it to "Tom" then the pattern it scanned as "rn" gets added as a "less likely rn, more likely m" lesson and over time it improves on the fonts people actually use). Places where it will often fail on standard screen fonts are 'm' and 'rn' distinctions (when I clip the Participants list in a Zoom window, my name is often "Torn" instead of "Tom"), adding/removing '.' and spaces, and smaller text. It isn't as accurate as a Google-based OCR, but the main advantage is that *nothing leaves your computer*, which makes it a viable product to use when dealing with text that isn't compatible with a cloud networking solution, or when you don't have a network connection at all. ![]() PicaText works "pretty well", which requires some extra explanation.
0 Comments
Leave a Reply. |
AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |