Omnifont OCR with a structural model

From a model which can generate all Roman body-text faces, we derive a new approach to omnifont optical reading. Central to this approach is the definition of few basic blocks called primitives. Characters are built from primitives according to a grammar independent of the font. When known, primitives have simple enough shapes to be efficiently isolated and recognized. Their variation from font to font are small enough to enable automatic learning. Finally, segmentation of words into characters is made optimal when building characters back from consecutive primitives.