Aplicação da Arquitetura U-Net para Segmentação de Mapas Históricos Holandeses do Século XVII

Abstract

Preservation and analysis of historical maps are fundamental tasks for decoding geographic, political, and cultural information from different eras of history. To ensure the conservation of these valuable documents, manual transcriptions of their content are made, thus allowing their persistence over time. With the aim to assist this process and improve efficiency in relevant data extraction, this study proposes a tool capable of segmenting historical maps to identify notable characteristics using advanced Deep Learning techniques. For this purpose, a dataset was created from 12 historical maps of the 17th century, accompanied by their respective segmentation masks. Out of these 12 maps, 11 pertain to cities in Brazil and 1 to a city in Ghana, West Africa. This dataset was used to train a convolutional neural network based on the U-Net architecture. The performance of the proposed model was evaluated using specific metrics: Intersection over Union, which indicated results above 0.75 for the segmentations performed, and Pixel Error, which value was below 0.05. These results attest to the efficiency and precision of the developed tool in the segmentation of historical maps.