Color Detection Of An Image

[anthony]

I would probably do a color reduction to say 3 or so colors to find the basic colors used in an image. A histogram would then tell you which was the more common color.

Code: Select all

   convert image.png -colors 3 -depth 8 -format %c histogram:info:- 

For the built-in rose: image this showed that a darkish gray and a red were the premonident colors of this image.

Hmmm... this may be a good why of sorting images for comparision, or finding images of people with a predominate 'skin' color.

The number of colors may been adjustment. but should still work well.

NOTE that removing the outside borders of the image before looking for a predominate color may also be a very good idea.
Also you can use a -scale to get an average color!

Code: Select all

   convert image.png -gravity center -crop 70x70%+0+0 -scale 1x1\! -depth 8 txt:- 

I have not tried these except in a superficial way.
Please let me know what you discover, so that others can benifit and refine the methdos further.


I have added these notes to IM Examples Image Compare -- determining and sorting images...
http://www.cit.gu.edu.au/~anthony//grap ... #imagetype

[magick]

The -segment option might be useful here. It analyzes the color histogram to find predominant mean colors in the image.

[anthony]

yes -segment will quantize the colors but without limiting it to a specific number of colors. That is it will NOT merge two seperated sets of colors, just near-colors. that is something using -colors with a low number is prone to do.

[anthony]

Additional...
I recently found an artical on what Color Segmentation is suposed to do....
http://www.leptonica.com/color-segmentation.html

The IM segementation algorithm does NOT do what this artical says it should do.

1. The controls do not seem to make any realistic difference to output.
   That is, while you can get some differences between 1x1 and 1x2, the first
   number does not appear to have any effect, and beyond those numbers the
   second number also has a almost nill effect on the results.
2. The results does not ignore things like black edges unless image is blurred first
3. And -segment fails completely on images less than about 300x300 pixels.
   that is it will work (badly) on logo: (unless blured first), but it will just produce
   a single color for a image like rose: