This work deals with processing historical cadastral maps. It is part of a complex system for seamless map creation. The main goal is to detect and recognize so-called nomenclatures, text information composed of several components that specify the position of a map sheet in the coordinate system. This information will then be used to create a large seamless map, allowing for better online presentation. The main contribution of this work is the utilization of a modern end- to-end approach and its comparison with the currently used two-step approach combining text detection and OCR techniques. We chose a visual document understanding model, concretely Donut, for our experiments. The results prove that such models can be successfully trained for our task and outperform the traditional methods. |
*** Title, author list and abstract as submitted during Camera-Ready version delivery. Small changes that may have occurred during processing by Springer may not appear in this window.