Active Learning for Classifying Template Matches in Historical Maps
B. Budig, and T. Dijk. Discovery Science, volume 9356 of Lecture Notes in Computer Science, page 33-47. Springer International Publishing, (2015)Best Applied Paper Award.
Abstract
Historical maps are important sources of information for scholars of various disciplines. Many libraries are digitising their map collections as bitmap images, but for these collections to be most useful, there is a need for searchable metadata. Due to the heterogeneity of the images, metadata are mostly extracted by hand - if at all: many collections are so large that it is important to minimise the effort spent on individual maps. We propose an active-learning approach to one of the practical problems in automatic metadata extraction from historical maps: locating occurrences of image elements such as text or place markers. For that, we combine template matching (to locate possible occurrences) with active learning (to efficiently determine a classification). Using this approach, we design a human computer interaction in which large numbers of elements on a map can be located reliably using little user effort. We experimentally demonstrate the effectiveness of this approach on real-world data.
Description
Best Applied Paper Award at Discovery Science 2015.
%0 Conference Paper
%1 budig2015active
%A Budig, Benedikt
%A Dijk, Thomas C. van
%B Discovery Science
%D 2015
%E Japkowicz, Nathalie
%E Matwin, Stan
%I Springer International Publishing
%K active-learning historical-maps myown template-matching
%P 33-47
%T Active Learning for Classifying Template Matches in Historical Maps
%U http://dx.doi.org/10.1007/978-3-319-24282-8_5
%V 9356
%X Historical maps are important sources of information for scholars of various disciplines. Many libraries are digitising their map collections as bitmap images, but for these collections to be most useful, there is a need for searchable metadata. Due to the heterogeneity of the images, metadata are mostly extracted by hand - if at all: many collections are so large that it is important to minimise the effort spent on individual maps. We propose an active-learning approach to one of the practical problems in automatic metadata extraction from historical maps: locating occurrences of image elements such as text or place markers. For that, we combine template matching (to locate possible occurrences) with active learning (to efficiently determine a classification). Using this approach, we design a human computer interaction in which large numbers of elements on a map can be located reliably using little user effort. We experimentally demonstrate the effectiveness of this approach on real-world data.
@inproceedings{budig2015active,
abstract = {Historical maps are important sources of information for scholars of various disciplines. Many libraries are digitising their map collections as bitmap images, but for these collections to be most useful, there is a need for searchable metadata. Due to the heterogeneity of the images, metadata are mostly extracted by hand - if at all: many collections are so large that it is important to minimise the effort spent on individual maps. We propose an active-learning approach to one of the practical problems in automatic metadata extraction from historical maps: locating occurrences of image elements such as text or place markers. For that, we combine template matching (to locate possible occurrences) with active learning (to efficiently determine a classification). Using this approach, we design a human computer interaction in which large numbers of elements on a map can be located reliably using little user effort. We experimentally demonstrate the effectiveness of this approach on real-world data.},
added-at = {2015-07-06T13:59:58.000+0200},
author = {Budig, Benedikt and Dijk, Thomas C. van},
biburl = {https://www.bibsonomy.org/bibtex/273395a9a8ded3c0d3c5d5efa60ae8447/thomasd},
booktitle = {Discovery Science},
description = {Best Applied Paper Award at Discovery Science 2015.},
editor = {Japkowicz, Nathalie and Matwin, Stan},
interhash = {d409e6fc1b92f8925b3fcc573008fa41},
intrahash = {73395a9a8ded3c0d3c5d5efa60ae8447},
keywords = {active-learning historical-maps myown template-matching},
note = {Best Applied Paper Award},
pages = {33-47},
publisher = {Springer International Publishing},
series = {Lecture Notes in Computer Science},
timestamp = {2015-10-21T11:56:57.000+0200},
title = {Active Learning for Classifying Template Matches in Historical Maps},
url = {http://dx.doi.org/10.1007/978-3-319-24282-8_5},
volume = 9356,
year = 2015
}