It's more an addition to Stevens answer:
- Histogram based features are e.g. HOG features, or Histograms of LBPs.
- Profile features are computed from the outer boundary of your letter / word in a specific direction, e.g. left profile: go from the left bounding box border to the right until you hit your first pixel and do this for each row (thus a proper binarization is important) and you'll get a left profile. You can do the same for the other directions too (top to bottom, bottom to top, left to right, right to left). From these profiles you can then compute several statistics (mean, std-dev., etc.).
