Toponym Recognition in Historical Maps by Gazetteer Alignment

Abstract

Abstract-Historical map documents are increasingly digitized for widespread access, but most are only coarsely indexed with meta-data while the contents are largely unsearchable. We propose to increase searchability by automatically recognizing the place names in these digitized artifacts. Using a word recognition system that produces a noisy ranked list of initial hypotheses from a lexicon of viable toponyms, we form a joint probabilistic model for inferring the most likely latent alignment between image toponyms and a gazetteer of known place locations. After a robust generalized RANSAC algorithm identifies the global alignment, we rerank the toponym hypotheses by their posterior probability. Experiments demonstrate a significant boost in word recognition accuracy on a manually annotated set of 19th century U.S. state and regional maps.

Details of Deformable Part Models for Automatically Georeferencing Historical Map Images

“…5 The method included in this table was chosen for strong overall performance, but is not necessarily best on each individual map. Overall results show Weinman's toponym-based adjustments [27] reduce RMSE (in pixels) by 41% over the prior technique [26], and the approach in this work nets a further 12% reduction.…”

“…Li's method [18] included rotation, but it may not be particularly robust to noise [5, p. 1:36]. Weinman's toponym-based methods [26,27] require no such assumptions, robustly aligning with a full affine transform using only map region metadata and a gazetteer (cf. Section 3.2.1).…”

“…Georeferencing. Previous work by Weinman [26] approximately georeferences map images solely for the purpose of improving toponym recognition, assuming that knowledge of a pixel's location on the globe tightly constrains the space of possible toponyms. To that end, each word image goes through a text recognition system, which produces a list of candidate words and their corresponding probabilities.…”

“…GT is best (least-squares) affine fit to the ground truth correspondences after projection. SAC is the initial toponym-based MLESAC affine fit [26]. EM is the adjusted toponym-based affine fit [27].…”

“…This work offers a method for improving automated alignment of historical map images to the Earth's geography, a process known as georeferencing [13]. Specifically, we build on prior work that provides an approximate alignment based purely on the arrangement, textual content, and typography of toponyms (place names) on the map [26,27]. This work adapts a shape-matching algorithm (related to those previously used for word spotting [15,16]) to the task of matching contours from GIS data-geographical and political boundaries, roadways, etc.-to historical map image contents.…”