III: Small: Collaborative Research: Adaptive Integration of Textual and Geospatial Information for Mining Massive Map Collections

点击翻译按钮获取中文摘要

Libraries and archives are digitizing historical maps for widespread online access. Without technology for searching them, large map collections relevant to a given problem or question may remain obscure even in online archives. If all of the text in a map can be read automatically by computer, a wealth of information becomes quickly available -- location names, geographic features, and often statistics. This project will increase capacity for search and analysis of historical maps by automatically recognizing place names and other text in these digitized artifacts while simultaneously aligning them with modern geography. The improvements this project will make to current text recognition methods will facilitate more powerful uses of humanity's trove of old maps -- for example, by allowing scientists and policymakers to establish changes in land usage, waterways, or borders over time. By creating free, open-source tools for studying historical maps, this project will increase public engagement with science and technology and empower any Internet user to explore the intersection of technology and history. This research will train a diverse group of graduate and undergraduate students in constructing, learning, and making predictions with adaptive models featuring heterogeneous yet highly interdependent entities.Although many institutions are digitizing hundreds of thousands of historical maps, most digitized map images are poorly annotated, limiting their usefulness. Manual annotation and metadata association is highly laborious. This project's primary objectives are (1) to fully automate text and shape-based georeferencing (aligning map images to the known global geography) while (2) indexing words and place names (for search) by enhancing text detection and recognition methods in these complex artifacts. These innovations will address the shortcomings of manual georeferencing and current automated text recognition algorithms. The researchers will employ an iterative interpretation process for solving problems including text/graphics separation, text recognition, and georeferencing. For example, the fact that all members of a given class of text entities on a map (e.g., county names) are typically rendered in the same text style can be used to inform predictions about difficult members of the category with information derived from more easily-recognized members. The researchers will use a dataset of annotated maps containing over 12,000 words in 9,000 place names as benchmark data for testing the algorithms developed in the project. Software, data, and benchmarks will be broadly distributed on the project website (http://www.cs.grinnell.edu/~weinman/research/maps.shtml). Findings will be shared with the research community through journals and conferences in the computer vision, artificial intelligence, and GIS communities.