Re: OCR/barcode reliability

On Sunday 11 July 2004 12:51 pm, Charlie Strauss wrote:
> This is a repeat of something I said before but its bears on
> the present conversation. One way to do image recognition of
> a finite set of possibilities is not to even bother with
> character by character recognition. Just recognize the whole
> item. TO give a strawman for you to whack at, just convolve
> the text to be recognized with the possible texts present.
> the signal is the peak of the convolution. Now if some letters
> were smudged, well you would not even know this from the
> convolution, The signal wold be reduced but you would not be
> attributing it to a particular character, just a deviation in
> the image.

Many years ago I read about optical computing systems that used
lenses for Fourier transforms and convolutions. They could pick
out words or phrases on a page, in the manner you describe.
Nobody could accuse us of fiddling in software with one of
those.

*<%-{]}}}

> Mathematically then, recognizing on a character by character
> basis then doing some other magic to recover lost characters
> is really nothing more than a very sophisticated (non
> euclidean) convolution. Its not "cheating" or guessing.
>
> However in all case you need in addition to the classification
> of the name, some error bars that give your certainty this is
> true.
>
> On Jul 10, 2004, at 12:49 PM, Jan Karrman wrote:
> > On Sat, 10 Jul 2004, Edward Cherlin wrote:
> >> On Saturday 10 July 2004 08:57 am, Jan Karrman wrote:
> >>> I see one big reason for not doing this: I can already
> >>> hear the enemies/critics of the OVC system claiming that
> >>> we are guessing what is written on the ballots. A misread
> >>> should always imply a hand count IMO.
> >>
> >> This is a fundamental misunderstanding of the OCR process.
> >> David was describing first and second passes within the OCR
> >> software, so that misreading one letter still gives the
> >> correct match on the whole name. Of course, we have to make
> >> sure that the candidates' names as presented on the ballot
> >> differ in more than one letter, if necessary by adding
> >> party affiliation.
> >
> > Well, I wasn't stating my understanding of the process. I
> > *do* believe it is something that could work. I was pointing
> > out a possibility for those who want to discredit the OVC
> > system. After all, it is a kind of (intelligent) guesswork,
> > where the vote recorded differs from what the scanner read.
> > Maybe it would be a strawman argument, easily refuted, but
> > I'm not so sure about that.
> >
> > Also, I did not see any mention about limiting the
> > correction to one letter. There was an example recognizing
> > names like "Krish***urthi Ram**issipan".
> >
> >> Certainly if that whole process fails, a manual inspection
> >> (not a hand count, since we are talking about individual
> >> ballots) is needed.
> >
> > Manual inspection was what I meant.
> >
> > /Jan

-- 
Edward Cherlin
Generalist & activist--Linux, languages, literacy and more
"A knot! Oh, do let me help to undo it!"
--Alice in Wonderland
http://cherlin.blogspot.com
==================================================================
= The content of this message, with the exception of any external 
= quotations under fair use, are released to the Public Domain    
==================================================================