Antti Kaihola wrote:
>
> I've tried some OMR (optical music recognition) programs, and although
> their performance is still poor compared to OCR, I've been quite
> impressed of what they can do.
>
> Imagine we have an OMR package. It can recognize without help from the
> user say 25 % of all musical symbols on an image of a music sheet. On
> the other hand, given any non-musical image it does not misinterpret
> more than half of that amount as music. If such programs do exist, I'd
> say that we have an AI algorithm strong enough for a visual music search
> engine.
>
> We could try and see. There are free versions and trial versions of OMR
> software available out there.
I've been thinking along the same lines myself. Although the OMR
packages I've tried are hardly useful for transcribing music, *some* of
them are at least able to determine whether a graphics file includes
notated music or not, and that's all we need for indexing purposes.
There are still a couple of obstacles, though:
1. Most of the graphics on Internet are 72 dpi. I still haven't found an
OMR that is able to cope with such low resolution.
2. There are far more fragments (music examples etc.) than complete
tunes on Internet. We need a way to filter those out to avoid the search
results to be loaded down with useless hits.
3. We need to be able to read the title - and preferably composer to be
able to index. This information is sometimes embedded in the graphics,
sometimes it's in HTML. And the formatting of these "header fields"
might differ so much that the only computers able to sort it all out is
the ones we all carry between our ears.
It would certainly be an interesting project, though.
Frank Nordberg
To subscribe/unsubscribe, point your browser to: http://www.tullochgorm.com/lists.html
