The angle corresponding
to the projection with the greatest positive derivative is used to detect
the least amount of overlap between vertical strokes and, therefore,
the dominant slant angle. In [18], slant detection is performed by dividing
the image into vertical and horizontal windows. The slant is
estimated based on the center of gravity of the upper and lower half
of each window averaged over all the windows. Finally, in [90], a
variant of the Hough transform is used by scanning left to right across
the image and calculating projections in the direction of 21 different
slants. The top three projections for any slant are added and the slant