[mythtv] A thought on Commercial Flagging -- Long (And rambling)

[Robert Johnston]

On 07/08/05, Jason W. Thompson <jason at jasonandmary.com> wrote:
> I'm not so sure how one would be able to implement all this, but it
> seems like a really good idea in theory.  I was thinking that perhaps
> when one edits out the commercials it is sent to a database so that next
> time someone watches that show (that is a repeat), the machine could
> just pull up data and not have to think about it.  But then I remembered
> that the length of commercials per showing of a program (or the length
> of the program) is often altered.  So that doesn't work either.

Well, the way I saw it was to use either a Neural Network, or some
kind of Bayesian filter, to "Learn" what is, and what is not, a
commercial. In thoery, as it's only data the filter would be dealing
with, we don't need to write a system to downscale the video, or
resample the audio, as part of the "Learning" will be for the NN to
spot patterns in the data, and use that to make it's own rules,
learning and refining as it goes.

This is my understanding of how a Neural Net works. You feed it
stimuli and give it a response that it's supposed to be able to
determine, and as the NN grows and learns it gets more and more
accurate at predicting what the response should be.

Now, initially the NN will be slower than molasses in January as it
attempts to find, track, and utilise all of the rules it has built to
try and determine the "Commercial" status, but as it uses Natural
Selection to weed out the cruft from the solid rules, the NN will
become faster.

I was thinking the NN could be very simple in what it has to achieve.
Just to say "Commercial" or "Not Commercial" for a given set of frames
(say, a second's worth).

I think I'll be doing more research into Neural Nets. :)
-- 
Robert "Anaerin" Johnston