I'm trying to build a system which can recognise simple networks of connected nodes from hand drawings
eg. 
I'm particularly interested in getting the edges between the nodes. But I'm finding it surprisingly difficult to get a good analysis. My processing pipeline is currently like this :
- blur the image
- invert
- adaptive gaussian threshold
- look for contours
- find bounding ellipses for contours
- filter to get just contours whose major axis is more than 3 or 4 times the minor axis, and whose overall size is larger than average (I'm guessing these match the edges rather than the nodes)
- find the foci of these ellipses (to get the end points of the edges)
The results really aren't good. Some edges still aren't picked up. The contours tend to pick up much too much detail eg. I might get a couple of parallel contours for the same edge. The bounding ellipse is often bigger than the contours and the foci aren't at the ends of the lines.
So ... am I approaching this all wrong? Is there a better family of algorithms for extracting this kind of information? Does anyone know of good research on this problem?
cheers
Phil