Learning regions of interest in postal automation

The general layout of mail pieces depends on certain rules, the sender's country, conventions and so on. One central task in postal automation is the localization of target addresses on mail pieces. After preprocessing, a mail piece consists of a set of bounding boxes for each region. In this paper a formal graph representation for mail pieces will be shown, in which relative constellations of bounding boxes are represented in a qualitative way. From a labeled set of such graphs, a model graph of the domain represented by the training set can be derived automatically. The model graphs only hold qualitative relations between items found on mail pieces of the training set. An inference mechanism based on the A* algorithm is used to find inexact subgraph isomorphism between a learned model and a formerly unseen mail piece. The implemented system has been evaluated on real letters and the recognition results are discussed.