Application of the error-correcting grammatical inference algorithm (ECGI) to planar shape recognition

ECGI is an error-correcting-based learning technique that aims at obtaining structural finite-state models of (unidimensional) objects from samples of these objects. The learning procedure captures certain useful regularities of the training data in the object-models, while also obtaining appropriate models of the `irregularities´ (errors and distortions) that these data tend to exhibit with respect to the learnt object-models. In the test phase, both the object-models and the corresponding error-models are cooperatively used to recognize new objects through stochastic error-correcting parsing. The application of ECGI to planar shape recognition is discussed and an example is given which consists of the recognition of arabic numerals from 0 to 9 that were handwritten by several writers. The results are compared with those of another more conventional (non-structural) recognition technique showing that not only ECGI clearly outperforms this technique, but it also seems capable of providing greater recognition accuracy than many other approaches reported in the literature