jacob-horton / sudokusolver Goto Github PK

Recognising digits from the MNIST and EMNIST dataset have around 80-90% success rate, which should be higher.
Recognising digits from processed sudokus has a very variable success rate, sometimes as low as 30%

The neural network digit classifier may identify a number as 0, which is not a valid number in sudoku

If one corner is inside the others, the convex hull algorithm used to get the points in a clockwise pattern will only return 3 points, which cannot be used in the quadrilateral transform. This arrangement of points, however, will not be possible as a square could never be projected to that shape, so the user must be informed that they are at fault.

The corners can be dragged into a shape such that an hour glass is made. This is corrected using the clockwise hull, but that is only computed at the end. If the user drags the points into an hour glass shape, a square should still be drawn.