Zooming and smoothing

[Gluender-2]

Bob,

I'm not totally sure about interpolation used by IJ.

As far as I remember the interpolation for zooming (which from my
point of view is definitely superfluous) is bilinear for speed
reasons but the interpolation used for scaling is cubic. (Just
checked it and it appears to be so.)

Then there is this great transformation plug-in "TransformJ" that
offers a great variety of different interpolation schemes (3 cubic
types and even one quintic). Furthermore, there is the plug-in
"Resize" that offers a few interpolation schemes.

In any case you have to justify the "simulation" of a finer grid...

HTH

Herbie

-------------------

>Herbie, Gabriel, Volker, and everyone,
>
>I prepare my acoustic images by computing them on a coarse grid
>(related to the phased array resolution) and then interpolating to a
>finer grid for presentation.  I found that the bilinear
>interpolation in ImageJ was not doing what I needed, so I wrote a
>bicubic interpolation code by extending the FloatProcessor class and
>overriding getInterpolatedPixel.  There are  other bicubic
>interpolation plugins available for ImageJ, including an example
>from the Burger and Burge book, but in each case the interpolation
>is part of something else, not a direct effort to extend ImageJ.  I
>posted my code on my web site last year in the hopes that someone
>would run with it by implementing the algorithm for the other image
>types and offering it to Wayne.  I was a little tentative because I
>was not sure that my somewhat invented algorithm would work out.
>I've been using it a lot, and it seems to be perfect for my needs.
>If it is time for ImageJ to go bicubic, here is a simple path for
>it.  I do not have time to do the coding right now.
>
>Bob
>
>
>Robert Dougherty, Ph.D.
>President, OptiNav, Inc.
>4176 148th Ave. NE
>Redmond, WA 98052
>[phone]
>FAX [phone]
>www.optinav.com
>[hidden email]

[Robert Dougherty]

Herbie,
>
>
> In any case you have to justify the "simulation" of a finer grid...
>
> HTH
>
> Herbie

Interesting point.  The acoustical images are generated on a range of  
scales depending on frequency and the processing algorithm, but they  
always have to end up a certain size to match the optical reference  
image and to look right in PowerPoint, Word, or YouTube.  I could  
compute them in that size (I'm giving YouTube 640 x 360 now) but it  
would take longer than it needs to in some cases (low frequency).  I  
could scale up the acoustic image without interpolation, but then it  
would look seriously wrong, and would not be what I would have gotten  
by computing every pixel.  I convinced myself that the band-limited  
nature of the imaging is consistent with the interpolation process,  
and validated this with some test cases.  In the tests, the  
interpolated results matched the detailed computation.   Application  
of a deconvolution algorithm makes the situation more interesting.  It  
that case, it is possible to generate high resolution nonsense.

Bob


Robert Dougherty, Ph.D.
President, OptiNav, Inc.
4176 148th Ave. NE
Redmond, WA 98052
[phone]
FAX [phone]
www.optinav.com
[hidden email]

[Gluender-2]

Bob,

thanks for providing more details of the problem.

>Herbie,
>>
>>In any case you have to justify the "simulation" of a finer grid...
>>HTH
>>Herbie
>
>Interesting point.  The acoustical images are generated on a range
>of scales depending on frequency and the processing algorithm, but
>they always have to end up a certain size to match the optical
>reference image and to look right in PowerPoint, Word, or YouTube.
>I could compute them in that size (I'm giving YouTube 640 x 360 now)
>but it would take longer than it needs to in some cases (low
>frequency).  I could scale up the acoustic image without
>interpolation, but then it would look seriously wrong, and would not
>be what I would have gotten by computing every pixel.  I convinced
>myself that the band-limited nature of the imaging is consistent
>with the interpolation process, and validated this with some test
>cases.  In the tests, the interpolated results matched the detailed
>computation.   Application of a deconvolution algorithm makes the
>situation more interesting.  It that case, it is possible to
>generate high resolution nonsense.
>
>Bob
>
>
>Robert Dougherty, Ph.D.
>President, OptiNav, Inc.
>4176 148th Ave. NE
>Redmond, WA 98052
>[phone]
>FAX [phone]
>www.optinav.com
>[hidden email]

1.
Of course I've assumed that your data essentially conforms with the
sampling theorem.

2.
As far as I understand the rest of the story, you want to match
images of different _true_ resolutions (defined by the acquisition
process; wavelength, type of radiation etc.) and display them
together.

3.
This apears to be a situation in which interpolation can make sense.
If these images are of scientific importance you should point out the
numerical values of the various true resolutions.

4.
In this context I still have problems with expressions such as "look
right" or "look seriously wrong". But...

5.
...there is an important aspect:
The usual visual representation of pixel values as blocks introduces
structures that have nothing to do with underlying image information.
That said, it is obvious that interpolation helps avoiding these
alien structures when inspecting and perhaps displaying spatially
discrete images at high magnifications.

6.
Originally this thread dealt with another intention, namely that to
make high resolution selections from interpolated images and this
appears to be scientifically highly problematic, not to say dishonest.
I cite:
"I draw a free hand selection and it creates a zigzag in pixels which makes it
impossible to follow the round.  Likewise for the thresholding - the
pixelation is very obvious."


Best

Herbie