Revision history [back]

Berak's suggestions are good, but one other idea is kmeans clustering. The large cluster is the background, and everything else is foreground. Then morphological operations to clean up the map.

I don't know how effective it would be, but it has the advantage of being easy to try.