Re: interpolation for a color image?

On May 27, 11:35 am, aruzinsky <aruzin...@xxxxxxxxxxxxxxxxxxxx> wrote:
On May 26, 4:28 pm, da...@xxxxxxxxx (Dave Martindale) wrote:





Harris <xgeorg...@xxxxxxxxxxxxx> writes:
...

The problem I see with this is that you're effectively calculating the
higher-resolution image by recombining a luminance/intensity image
that was obtained by linear interpolation and colour information that
was upsampled by nearest-neighbour.  The latter is a terrible upsampling
method that doesn't actually interpolate anything.

But if the two pixels have the same intensity but different colour,
there will be an abrupt colour change half-way along a line between the
two original pixel locations.  So the colour component of the image will
have some nasty artifacts in it.

Now, you can argue that under certain viewing conditions, your eye's
colour resolution is bad enough that you can't see the colour artifacts,
so they don't matter.  True.  But you will get an image that looks
significantly worse up-close than interpolating in all 3 components
(whether that's in RGB or YCbCr).

In Harris's defense, it won't look "terrible" or "significantly worse"
but just bad enough to negate the advantage of the computation savings
which are much smaller than he imagines.    The eye is relatively
insensitive to inaccuracies in the Cb and Cr channels and that is why
they are compressed more than Y in JPEG.  I have used schemes similar
to Harris's for Xin Li (NEDI) enlargement.  In NEDI, counterparts to
bilinear weights are calculated Wiener style which is very time
consuming.  I calculate the weights based only on the Y channel and
apply them to all three channels, Y, Cb, Cr.   It is good enough
considering the ~2/3 time savings.  However, my default weights for
homogeneous Y are bilinear and not nearest neighbor as in Harris's
case.- Hide quoted text -

- Show quoted text -

Here is why one should not put theory ahead of empiricism:

1. Input Image: http://www.general-cathexis.com/interpolation/clown.png
2. 4X Bilinear on RGB channels: http://www.general-cathexis.com/images/clown4Xbilinear.jpg
3. 4X Bilinear on Y, Nearest Neighbor on R/Y etc. :
http://www.general-cathexis.com/images/clown4XbilinearYRatio.jpg
4. 4X Bilinear on Y, Nearest Neighbor on Cb and Cr:
http://www.general-cathexis.com/images/clown4XbilinearYboxCbCr.jpg

The Y here is Y = (R+G+B)/3. Conventional YCbCr gave worse results
for 3.

This was done with SAR Image Processor with no additional
programming. You can verify these results with free SAR save disabled
demo. No. 3 involved circuitous calculations involving log RGB color
space where division is subtraction and multiplication is addition. I
mention this because the No. 3 result should be checked. 3 is the
worse result. If you look below the triangular white area of the
right car, you will see bright red streaks. No. 4 is almost
indistinguishable from 2.

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

Relevant Pages

  • Re: Basic Texture Mapping Question/Problem
    ... I'm trying to get the basics of tecture mapping down and I ... and the exterior edges of the square with the background. ... of the nearest voxels with the border value. ... You can set the interpolation to NEAREST to get sharp edges. ...
    (comp.graphics.api.opengl)
  • Re: Yeah, I have desktop effects:
    ... Nearest, Bilinear or Trilinear? ... Direct Rendering? ...
    (Fedora)
  • Re: Avoiding for-loop
    ... I have 2 mesh matrices ... that contain a grid of u,v coordinates. ... Specify 'nearest' for interpolation method to keep ...
    (comp.soft-sys.matlab)
  • Re: Interpolation
    ... >Is there any kind of interpolation like spline, nearest ... >neighborthat can be used for interpolating a nonuniform data ...
    (sci.math.research)
  • Interpolation
    ... Is there any kind of interpolation like spline, nearest ... neighborthat can be used for interpolating a nonuniform data ...
    (sci.math.research)
Loading