CNN-BASED TEMPLATE MATCHING FOR DETECTING FEATURES FROM HISTORICAL MAPS

Abstract

Efficiently detecting features from historical maps is a challenging task due to its inconsistent manual scribbling styles and the lack of large scale labelled training data. To tackle this issue, this paper proposes an automatic feature detection pipeline utilizing CNN-based template matching (TM), which can lead to efficient feature extraction with minimal input, i.e. one single template. Three CNN-based TM models equipped with different feature extractors are investigated and compared in this research, namely pre-trained VGG19 CNNs, autoencoders, and the combination of both. Experiments conducted on six tiles of the Swiss Old National Map demonstrate that the combined architecture achieves the best result in wetlands detection, resulting in a mean intersection over union (IoU) of 69% and an average F1 measure of 82%.