I am currently applying some pre-processing to this dataset. Some of the digits are extremely rotated and I would like to know how others managed to get them 'upright'.
I have implemented something on matlab which finds the difference between (most bottom pixel) - (most top pixel) and (most right pixel) - (most left pixel). And a statement to check that the vertical distance is greater the horizontal. However a few numbers like 2 are written more wider than taller. This gives me more dirty than clean data.
Anyone else got any ideas on some logical statement to tell if this digit needs rotating or not?