battletoad
Expert Member
- Joined
- Mar 10, 2009
- Messages
- 1,461
- Reaction score
- 52
Have a look at TROCR and docTr, both can run locally.
The former requires you to do a bit of training for the weights, the latter works right off the bat (resnet50 with rotation will handle angled forms).
These simply perform ocr and it’d be on you to handle making sense of the forms.
On the other hand - although I’m not 100% sure - docvqa, donut and Google’s pixt2struct may just be able to extract the data automatically with some training (pixt2struct’s GitHub outlines its training method).
Also on Levenshtein: chuck the records into Postgres and have Levenshtein run there. Enrich records with additional data so that you can score results not only on the edit distance, but other reasonable signals too.
The former requires you to do a bit of training for the weights, the latter works right off the bat (resnet50 with rotation will handle angled forms).
These simply perform ocr and it’d be on you to handle making sense of the forms.
On the other hand - although I’m not 100% sure - docvqa, donut and Google’s pixt2struct may just be able to extract the data automatically with some training (pixt2struct’s GitHub outlines its training method).
Also on Levenshtein: chuck the records into Postgres and have Levenshtein run there. Enrich records with additional data so that you can score results not only on the edit distance, but other reasonable signals too.
