Why is the 'Median Filter' limited to a radius of 100 pixel maximum?

[Matthias Schade]

Hello

it is not possible to execute the 'Median Filter' nor the 'Remove Outliers..' function via user-interface for radii > 100 pixels. Historically, considering processing power or RAM this might have been a reasonable limit. But nowadays?

I was following recommendations from the following paper on the inference of an illumination function where the authors apply a median filter (r=500 pixels) on an image of 1280x1024 pixels:
Singh, S., Bray, M., Jones, T. R., & Carpenter, a E. (2014). Pipeline for illumination correction of images for high-throughput microscopy. Journal of Microscopy, 256(3), 231–6. doi:10.1111/jmi.12178

The ImageJ java code which I think is responsible for catching any radius>100 pixel is found in "RankFilters.Java" (latest version for 1.49m) and states: "int maxRadius = (filterType==MEDIAN || filterType==OUTLIERS || filterType==REMOVE_NAN) ? 100 : 1000;"

This limitation is void/ignored for ImageJ macros, i.e. using a line such as: run("Median...", "radius=200 slice");

Cheers
Matthias

--
Matthias Schade

Humboldt-Universität zu Berlin
Molekulare Biophysik  -  Institut für Biologie
Invalidenstraße 42  -  10115 Berlin, Germany

fon      [phone]
fax:     [phone]
office: 216

email: [hidden email]
www: http://www2.hu-berlin.de/biologie/molbp/

--
ImageJ mailing list: http://imagej.nih.gov/ij/list.html

[Michael Schmid]

Hi Matthias,

the reason for limiting the radius for the median is computing time:
For the median, ImageJ uses quickselect (Hoare's selection algorithm), with some additional tricks for optimization. This is a good choice for the typical small kernel sizes, but slow for large sizes of the kernel.
On my computer (2.4 GHz Core2Duo), the median for a noisy 2048x2048-pixel image with r=100 takes needs almost 7 minutes!
With these computing times, I think that it makes no sense to have larger radii; the user might think it got stuck if the progress bar does not change. A large radius is useful for large images only, and there it takes excessively long.

For 8-bit and 16-bit images and a large radius, you are much better off with a histogram-based median. If you are happy with a square (instead of circular) kernel area, you could try this one:
  http://ij-plugins.sourceforge.net/plugins/filters/

Maybe there is a reasonably fast 'true' median algorithm also for floating-point images; if someone knows one, let me know, please!

An approximation for the median, also with a square kernel, can be found in the Fast Filters plugin. It needs 5 seconds for the 2048x2048 image.
  http://imagejdocu.tudor.lu/doku.php?id=plugin:filter:fast_filters:start


What is the application of a median with such a large size?


Michael
________________________________________________________________
On Dec 15, 2014, at 20:29, Matthias Schade wrote:

> Hello
>
> it is not possible to execute the 'Median Filter' nor the 'Remove Outliers..' function via user-interface for radii > 100 pixels. Historically, considering processing power or RAM this might have been a reasonable limit. But nowadays?
>
> I was following recommendations from the following paper on the inference of an illumination function where the authors apply a median filter (r=500 pixels) on an image of 1280x1024 pixels:
> Singh, S., Bray, M., Jones, T. R., & Carpenter, a E. (2014). Pipeline for illumination correction of images for high-throughput microscopy. Journal of Microscopy, 256(3), 231–6. doi:10.1111/jmi.12178
>
> The ImageJ java code which I think is responsible for catching any radius>100 pixel is found in "RankFilters.Java" (latest version for 1.49m) and states: "int maxRadius = (filterType==MEDIAN || filterType==OUTLIERS || filterType==REMOVE_NAN) ? 100 : 1000;"
>
> This limitation is void/ignored for ImageJ macros, i.e. using a line such as: run("Median...", "radius=200 slice");
>
> Cheers
> Matthias
>
> --
> Matthias Schade
>
> Humboldt-Universität zu Berlin
> Molekulare Biophysik  -  Institut für Biologie
> Invalidenstraße 42  -  10115 Berlin, Germany
>
> fon      [phone]
> fax:     [phone]
> office: 216
>
> email: [hidden email]
> www: http://www2.hu-berlin.de/biologie/molbp/
>
> --
> ImageJ mailing list: http://imagej.nih.gov/ij/list.html

--
ImageJ mailing list: http://imagej.nih.gov/ij/list.html