A probabilistic approach to automatic handwritten address reading

Abstract

To sort handwritten mail pieces, the French Postal Office Research Centre has developed a reading device. After recalling the different processing stages of our system, we focus on the problem of merging the decisions made by its different sub-modules (postal code and city name hypotheses) to produce a unique and reliable final decision. To achieve this goal we use a probabilistic modelling of the system behaviour taking into account the characteristics of all the processing stages involved in handwritten address recognition. This approach has been rested on large reference sets consisting of live mail handwritten envelopes and already produces very promising results.