joint noise reduction and missing data treatment

[lex3]

Hi, All!

While looking for a tool to restore a couple of old film-based cartoons I've found a publications page (http://brava.ina.fr/papers.en.html) of the Brava Project (Broadcast Restoration of Archives by Video Analysis).

Some titles sound very interesting:
- MCMC for joint noise reduction and missing data treatment in degraded video
- Practical MCMC for Missing Data Treatment in Degraded Video
- Image Sequence Restoration in the Presence of Pathological Motion and Severe Artifacts
(MCMC stands for Markov Chain Monte Carlo methods)

But without solid math background (:() I cannot evaluate chances to incorporate any of these modern scientific methods in a "noise killer / perfect sequence restorer" filter :)

Is here anybody who can?

[High Speed Dubb]

Ooh — nice link. Those should be interesting to look at. I do (non Bayesian) MCMC stuff for my day job in population genetics, so this isn’t too far outside my field.

Deinterlacing is a special case of missing data in the presence of noise. So that first article could well be relevant.

The nutshell explanation of MCMC is that it lets you search (and sample from) big, poorly defined spaces in a useful way. Plain Monte Carlo involves picking random examples from the space over and over. But if the space is big enough, that’s not efficient enough. MCMC instead biases toward the examples which best explain the data.

It’s a very powerful method, since it lets you use extremely complicated criteria for your search but still get an answer in a reasonable amount of time. But beware that a “reasonable amount” could well be weeks, depending on the problem.

[lex3]

Do you mean that adding even basics of these methods to a filter will mean slowing down to 1 frame per hour or day?

[High Speed Dubb]

It depends too much on the problem and the implementation for me to give a general answer. If you’re just sampling from one or two parameters in a smooth space with a single peak, MCMC is plenty fast, and might in some cases even run in real time. Then again, if you just have a couple parameters and a single peak, you could probably use something simpler, instead.

On the other hand, there are some problems which will take weeks — and even then you may not be sure of the answer, since this is a stochastic method.

Just in case anyone is up for learning about this — The good news is that the math of MCMC is easy stuff. If you can manage matrices and a little algebra, then you can handle the math involved.