[IPOL discuss] [IPOL tech] Démo de la méthode K_SVD

[Miguel Colom]

Quoting Nicolas Limare <nicolas.limare at cmla.ens-cachan.fr>:
> It's not the "demo system" (ie python code) that is stopping processes
> demanding too much memory, it is the operating system kernel.
Well, I consider our configuration or policy about the memory as part  
of the demo system.

> And for the general discussion, could you elaborate on why denoising
> methods need more than 1G memory? I have no opinion on this point, but
> I would like to understand where this memory need comes from. 1Gb is
> more than 130 million double-precision variables.

Despite all the optimizations we put, denoising (and also noise  
estimation) are very costly in memory. I'm thinking on the (yet  
unpublished) Noise Clinic.

In general, you consider overlapping blocks of size w*w. A typical  
value if w=8, so in practice the amount of memory you need just to  
keep the image in memory is w^w=64 bigger. And of course, some  
algorithms need to deal with color, so they need to have all channels  
loaded, and can't process them sequentially.

And also, the denoising algorithms are multiscale, that means that  
they create a sort of pyramid structure of the images, which means  
more memory.

And they are also multifrequency, so they need information about the  
image at each frequency.

Denoising is likely to be one of the topics that demands more  
resources (memory and CPU).