A probabilistic model for postcode recognition: a first step towards probabilistic address interpretation

In order to solve the problem of address interpretation using a probabilistic framework, we need to calculate the probabilities of the individual address elements. An important address element is the postcode. In this paper some basic research is presented concerning the calculation of the probability of the postcode, given the character recognition scores of a handwritten boxed postcode and the postcode dictionary. The resulting probability model is presented. The accuracy of the probabilities obtained by this model was investigated using a test set of 105,761 images. The probabilities were found to be very accurate: in the higher range of the probability values, an accuracy of 0.2% was measured. Furthermore, the error-reject characteristic of the probabilistic method was determined. Compared with the method that is currently operational, the new method has a yield that is 4.3% higher at an error level of 1.0%. The results of this paper provide a solid foundation to continue the development of a probabilistic framework for address interpretation