Re: (statistics)how to make date more like Laplacian distribution?

"walala"

> Why people think I want to lie upon seeing my question? Oh, it's my
> problem
> that I did not clearly present the background...

Are you quite sure you are describing your problem clearly? Reading between
the lines you have a poor non invertible integer DCT which you are trying to
compensate for.

I am going to assume that is true for now ... my apologies if it isnt true.

Statistics and deblocking are not the answer, your transform is the problem.
Either use more precision to get closer to the true DCT transform, or
create an invertible transform using lifting/ladder structures.

> Here is the story: in deblocking of block DCT coded JPEG images, it was
> known that the DCTed coefficients are Laplacian distributed... But now I
> am
> looking at low bit rate JPEG images, so there are someking of artifacts...
> in order to reconstruct the original images... many algorithms have been
> devised... one possibility is to make the image coefficients more Laplcian
> like...

This will do nothing, the idea of using the laplacian distribution with
deblocking works by using a reconstruction point for the quantized
coefficient inside its quantization interval which will minimize the
expected distortion. In very simple terms, the coefficient is more likely to
be in the low end of the quantization interval, so we assign a value in that
region to the dequantized coefficient.

If you want to use this method to do deblocking you dont have to make the
image coefficients more laplacian, but you simple have to fit the
distribution of the coefficient (across the image) to a laplacian ... then
you take the parameter for that laplacian and determine from it where in the
quantization interval to reconstruct the coefficient.

If you are trying to use this in some fashion to combat a poor transform you
are barking up the wrong tree. You need to work on the transform, the paper
I mentioned earlier has most of the relevant math :
http://citeseer.nj.nec.com/kim98fixedpoint.html

If your goal is not to get that close to the DCT but simply get a DCT like
transform which can be implemented cheaply then you should look at
invertible approximations like the binDCT (google for it, someone on the
comp.dsp newsgroup posted a link to source code in the past I think).
Without the invertible property the low precision integer math you want to
use will create an unuseable transform, and no amount of deblocking or
laplacian statistics is going to fix that.

Marco