Learning character recognition by localized interpretation of character-images

Recognition algorithms encompass segmentation, feature extraction and classification, but these components might be difficult to isolate because of strong interactions between them, and the lack of crisp criteria telling where one stops and where the other begins. Extreme variability of text images, and of hand written texts in particular, makes it difficult to tune any of those three parts of a recognition algorithm to real data. Automatic parameter tuning (training or learning) requires parametrization of at least a part of the algorithm. As it is more convenient to parametrize classification than the rest of the recognition algorithm, machine learned recognition usually means that the recognition classifier has been trained or tuned automatically. We show that our box connectivity approach to feature extraction, and localized interpretation within the classifier, provide solutions to the outlined problems, and allow efficient implementation of direct learning.