Can you use linear optimization to efficiently solve the OCR problem?
One reasonable linear objective function might be a weighted sum of the entropy of the text and the number of wrong ("noise") pixels, or perhaps the total absolute pixel error. You can use the standard one-of-N mixed integer linear programming approach to select which glyph is at a given position, and a reasonable entropy function might use digraph frequencies in your language of choice. You might be able to use some design variables that indicate the X-Y position of each glyph, and the font would perhaps be a shitload of parameters.
Alternatively, simple hill-climbing with a nonlinear problem state ment might be simpler to formulate and work adequately; it might "learn" the font simultaneously with the position and angle of text on the page. Gradient descent could more efficiently provide fine glyph positioning and adjust the glyph contents. (Or just simple mean/median?)