Identifying the place without text annotations: an assembled neural network framework for content-based raster map retrieval with cartographical morphological pattern

Abstract:

Currently, a majority of maps originate from volunteered sources. These volunteered maps have been created with the raster data structure and failed to follow the professional mapping principles. As the primary map languages, text and symbol annotations might be incorrect, or even missing. This poses a challenge for state-of-the-art map content recognition approaches that mainly focus on map text information. Under this occasion, graphical information could be an alternative solution for retrieving these maps. However, map graphs might significantly vary and overlap on complex backgrounds in massive volunteered maps. To address this challenge, we propose a concept called a cartographical morphological pattern, and an assembled neural network framework for retracing raster maps based on labeled and unlabeled datasets. The experiments prove that the feature maps generated from deep learning models can represent the shape characteristics of the target place, and our proposed integrated framework enables effective volunteered map retrieval based on cartographical morphological patterns. We hope our work can provide a novel strategy for content-based map retrieval.

Citation:

Xiran Zhou, Yi Wen, Zhenfeng Shao, Wenwen Li, Guochao Hu, Xiao, Xie, Qunshan Zhao. (2025). “Identifying the place without text annotations: an assembled neural network framework for content-based raster map retrieval with cartographical morphological pattern.“ Geo-Spatial Information Science, 1–17. https://doi.org/10.1080/10095020.2025.2522146.