Re: interpolation for a color image?

On May 28, 9:58 am, Harris <xgeorg...@xxxxxxxxxxxxx> wrote:
aruzinsky <aruzin...@xxxxxxxxxxxxxxxxxxxx> wrote innews:5f90d431-efc6-4ce2-a629-fbfbd9c6304f@xxxxxxxxxxxxxxxxxxxxxxxxxxxx:

If you have any doubt,
see:http://en.wikipedia.org/wiki/Bilinear_interpola
tion.

That doesn't apply to enlargement because the problem is separable and
performed as 2 1D interpolations.  Assume input is m x n and output is............
............

I hope you understand the fault in your thinking. Resize is non-separable oprtation in 2D because you
have to take into account a "box" instead of a "line" - this why the linear interpolation is called "bi-
linear" and no just linear. What you are describing will not work if the contents of the image are
correlated statistically in both directions, since artificially separable operations usually cause
discontinuities in the resulting image (similar to the jpeg "block" artifacts). Still, you may choose to use
that, if you feel there is some other application-specific advantage (e.g. speed?)


The fault in your thinking is that you ignore the reality of what I
do. I indeed do bilinear interpolation as a separable operation and
the results are exactly the same as for doing it as a non-separable
operation. It does not depend on the image. You are denying reality
and preferring your thoughts in the manner of a schizophrenic.



In hindsight, if Harris's ratios were a good idea, they would probably
be used in lossy image compression instead of Cb and Cr.

I usually don't write follow-ups in my Usenet posts because sometimes it seems like flaming someone
(which I never do). I would never suggest nearest neighbor or bilinear for serious image resizing,
especially for upscaling. Furthermore, lossy image compression has nothing to do with image resize.


Seeing the relation requires intelligence and creativity.


My "optimality" criterion regarding my posts here were entirely towards speed and ease of code
optimization (less divisions, more integer operations).


Then just use nearest neighbor.

Everyone who has spent years working in this field (professionally or academically) knows the
advantages and disadvantages of each method, there is no point discussing things that are easily
available in any good textbook in signal/image processing.


And, that would be who?
.

Loading