HoughCircles() requires too much fine tuning

[chenhe95]

It's very hard to tune HoughCircles() to work well for all hand drawn graphs because it is extremely sensitive to the parameters such as radius and minimum distance.
If we were to replace HoughCircles with some sort of closed boundary detection by pixel iteration, it might work better

[chenhe95]

A possibly new circle detection algorithm would be to look at every black pixel and search for closed loops. Then we could take the average point of the loop as the (x, y) and we could take the average distance from the average point (x, y) to every point in the loop as the radius. Then we can get the normalized squared error from the circle model to the actual circle to see how good of a fit it is. I say normalized here because larger circles will have more error, so we normalize it by the radius of the circle (or circumference), or something along those lines.