Generating Historical Maps from Online Maps
Zekun Li · 2019
This paper proposes an automatic system to generate a large amount of data for the training of text detection systems for historical maps. The system takes online maps as input and learns a conditional GAN model, to generate realistic historical map images from existing geographic datasets. Then the system uses the generated images as the base map and inserts synthetic text. Since the system has the control of text content, font style, and location, the system can obtain ground truth information (minimum bounding boxes) of the synthetic text. To overcome the challenge of content mismatch, the proposed system uses a novel loss function to encourage the generation of historical cartographic symbols in the foreground areas and discourage the generation in the background. The final output is a set of images resembling historical maps and the minimum bounding boxes around text regions on the images as annotations.