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

"walala"

> I just wanted to say that binDCT itself does not give good quality, it is
> the EZW-like packetization scheme make it good PSNR... If binDCT is then
> quantized with normal JPEG quantization matrix, that will give lower
PSNR...
> I guess...

You shouldnt compare it against the DCT, but against your _approximation_ of
the DCT.

If you want a transform which can be implemented in hardware at a far lower
cost than an IEEE compliant DCT (I forgot the exact number of the standard)
then you should consider the binDCT ... if you can afford the cost of a good
enough DCT and dont need the ability to losslessly invert the transform then
sure, DCT is better.

> No I am using the exact formula in that paper to compute variance... it
says
> that variance = (F_reconstructed)^2

Actually it uses square brackets ... that notation means you need the second
moment of a set of numbers, ie. variance, not the square of a single number.

You need to calculate the variance for each given coefficient, using the set
composed by that coefficient for all the blocks in the image. From that you
determine a single bias, which you subtract from each and every one.

So bias should be a 8x8 matrix too.

Marco