Abstract: The research and implementation of postal numerals segmentation and recognition system for window letters are discussed in the paper here. Row information of characters is gained from envelope’s image according to projection, then the sub image of postal numerals is segmented. The numerals are recognized by two independent classifiers, and their results are combined to form final results. A new combination mode of multiple classifiers based on statistical rules is utilized in order to achieve good performance for high recognition rate and low substitution rate. It is proved that the performance of the system is good, which suffices to the requirement of the practical use.