Katja Worm

Image Based Mail Piece Identification Using Unsupervised Learning

Next generation postal sorting machines reuse once extracted mail piece addresses in different sorting steps by means of the mail piece image. Based on the mail piece uniqueness, characteristics derived from the image guarantee the assignment of stored addresses. During the first sorting step mail piece characteristics are extracted and stored together with the target address in a database. In subsequent sorting steps the address is accessed by determining the corresponding mail piece characteristics in the database. Appropriate mail piece image characteristics and procedures for their distance measurement were presented in a previous work.

Handling of Surface Modifications for Robust Image Based Mail Piece Comparison

Image based mail piece identification is a new technology for optimizing the postal sorting process. Exploiting the mail piece surface uniqueness, characteristic features are used to identify each mail piece within a large amount of mail pieces. This process facilitates the storage of mail piece relevant data and its assignment in different sorting steps without manipulating the mail piece surface. Mail piece transportation and the mechanical mail piece sorting process may cause address and label movements, image rotations and further application-specific modifications. In contrast to other document identification systems image based mail piece identification requires high robustness towards above mentioned surface modifications for stable identification. This paper introduces four identification models for these special scenarios. Based on the proposed models text based feature extraction methods, feature representations and in particular appropriate distance metrics are presented which guarantee a robust mail piece identification. The applicability of the proposed procedure is shown in different experiments.