Abstract

The large volume of mail pieces that must be handled every day, the intensive process of mail handling, and the increase of the manual processing costs have made automatic sorting systems very important and economically attractive. One of the challenges to make the postal automation technology more accurate is to improve the accuracy of the address-block locating task. This task can be decomposed into: address-block segmentation, which involves segmenting the mail piece into different regions and address-block selection, which involves selecting the segmented region that satisfies the optimal destination address-block among all the segmented regions. In this thesis we focused on the segmentation part of the address-block locating task. We investigated whether some clustering techniques and ensemble clustering techniques will help to improve the accuracy of address-block locating. For this a workbench with several clustering methods as well as ensemble methods was developed. Very little prior knowledge of the images is required. The implemented system has been evaluated on mail piece images captured live from real postal pieces at the postal sorting centers. The results of this approach will be described and illustrated with tests carried out on different images (parcels, magazines, postcard, etc.) where there are no fixed position for the address-block, postmarks and stamps. A ground-truth strategy is employed to evaluate the accuracy of segmentation. The agglomerative based clustering and their ensemble version achieved the best clustering results. However they suffer from high run-time. Therefore further development will be needed to achieve better computation time.