Abstract: Historical maps depict important landmarks, such as river networks or dwellings, however, in much more artistic fashion and with less precision than their contemporary counterparts. Study of such maps is of great interest to cartographers, since both the map style, accuracy, level of detail and the countryside itself has been evolving over the centuries. However, with vast map colections of cartographic institutions, this field calls for automatization of object recognition. This brings challenges of its own, with extraction of river network, recognition of old, hand-written signs, lack of annotation, etc. In your work, you will aim to improve accuracy of segmentation of river network in the map scans using deep learning methods, possibly with weak annotation.

Supervisor: Jan Schier

Collaborating institution: Department of Applied Geoinformatics and Cartography,
Faculty of Science, Charles University