Many different approaches have been tried to build accurate, efficient symbol recognizers. Under the umbrella of prototype-batted symbol recognizers, there is a spectrum ranging from rigid template matching to matching via deformable templates. However, the former is fast but not so accurate and the latter are very accurate but also very slow. We consider merging the best of both worlds into a new prototype-based classifier, one that is fast and robust.; We present an efficient, adaptable representation of prototypes as vector templates, and an image metric, the Inkwell Hausdorff distance, that is fast yet tolerant to small misalignments. This technique is shown to be faster that existing techniques, with only a slightly lower accuracy.
展开▼