> An ideal OCR would use the "lower level" results from the letter
> recognition to influenece the "high level" word recognition and
> vice versa.
>
> I'm not familiar with this Neural Net speak, but I think this
> sort of thing is possible and should help to achieve much higher
> accuracy levels.
Good news: You are describing one of the uses for backward error
propagation in neural nets.
Bad [news?]: I am not aware of any serious study of this method's
consistent application in the system with large number of levels involved
(here it's shapes -> letters -> words [-> sentences]). OCR systems mostly
use "simple hacks", and probably neural-net-like things inside one level.
-- Alex
-- Source code, list archive, and docs: http://www.mostang.com/sane/ To unsubscribe: echo unsubscribe sane-devel | mail majordomo@mostang.com