Here is a small approach I will use in your case:
- Find octagon by template matching (doc). Use an empty octagon for template.
- subtract the empty octagon for each founded => you have ``only'' the lines made by user (and some noise remove with a threshold on white value, may be convert the image in HSV before doing the subtraction with cvtColor.
- In the binary mask coming from subtraction, count the pixels along all potential lines (1-2,1-3,...) using an acceptable width, ie: count not only for a line of one pixel, but for 5 or 6 thick. If the value is high (to determine by samples...) => you have a connection between numbers.
To avoid problem when you count a pixel for a line, remove it from the mask if the line is accepted.
If the line are too thin, try morphological maths to emphasized it with this.
Let us know if it working.
