FCN-Boosted Historical Map Segmentation with Little Training Data

Josef Baloun and Ladislav Lenc and Pavel Král
ICDAR (2023)



This paper deals with automatic image segmentation in poorly resourced areas. We concentrate on map content segmentation in historical maps as an example of such a domain. In such cases, conventional computer vision (CV) approaches fail in unexpected unique regions such as map content area exceeding the map frame, while deep learning methods lack boundary localization accuracy. Therefore, we propose an efficient approach that combines conventional CV techniques with deep learning and practically eliminates their drawbacks. To do so, we redefine the learning objective of a simple fully convolutional network to make the training easier and the model more robust even with few training samples. The presented method provides excellent results compared to more sophisticated but solely deep learning or traditional computer vision techniques as shown in “MapSeg” segmentation competition, where all other approaches were significantly outperformed. We further propose two additional approaches that improve the original method and set a new state-of-the-art result on the MapSeg dataset. The methods are further tested on an extended version of the Map Border dataset to show their robustness.



@inproceedings{baloun2023fcn, title={FCN-Boosted Historical Map Segmentation with Little Training Data}, author={Baloun, Josef and Lenc, Ladislav and Kr{\'a}l, Pavel}, booktitle={International Conference on Document Analysis and Recognition}, pages={520--533}, year={2023}, organization={Springer} }
Back to Top