KLT based character recognition?

[SirDavidGuy]

Anyone given any thought to this?

Training some matrices with common Bold, Italicized, multiple fonts, etc., then making an OCR algorithm for determinig the character?

[MfA]

You sure you intended to ask that question on this board? :)

Dont have a lot to contribute to your question, but from my very limited knowledge I think directional transforms are more usefull than KLT for character recognition (gabor/radon/etc).

[SirDavidGuy]

I think directional transforms are more usefull than KLT for character recognition (gabor/radon/etc).

Gabor?

My memory is a little hazy, but isn't that a time/frquency transform, useful for seperating sub-bands, but not so much for similarity detection?

In radon, doesn't it handle simple on/off switches poorly?

I need to look this up, I'm getting forgetful of it.

You sure you intended to ask that question on this board?

eh? Where else?

[SirDavidGuy]

Another Radon thing:

Isn't it intended more for straight lines? seems like curves would introduce lots of high power energy into it.

[MfA]

Well somewhere with a slightly higher probability of finding people who do this kind of research :) news://sci.image.processing for instance ... maybe there is a mailing list on the topic around too?

I just know Gabor is very popular for feature extraction in pattern recognition problems ... the radon transform was just the only other directional transform which quickly came to my mind (also intuitively I would say recognition using feature vector from PCA would be very sensitive to deformation&translation). I dont claim to have much knowledge on the subject, but I had a hunch that you would not be getting a lot of answers so I said my piece anyway :)

If you want state of the art recognition you will probably have to combine several different methods anyway (which in itself is not a trivial task).