Libraries are digitizing their collections of maps from all eras, generating increasingly large online collections of historical cartographic resources. Aligning such maps to a modern geographic coordinate system greatly increases their utility. This work presents a method for such automatic georeferencing, matching raster image content to GIS vector coordinate data. Given an approximate initial alignment that has already been projected from a spherical geographic coordinate system to a Cartesian map coordinate system, a probabilistic shape-matching scheme determines an optimized match between the GIS contours and ink in the binarized map image. Using an evaluation set of 20 historical maps from states and regions of the U.S., the method reduces average alignment RMSE by 12%.

We introduce deep convolutional and recurrent neural networks for end-to-end, open-vocabulary text reading on historical maps. A text detection network predicts word bounding boxes at arbitrary orientations and scales. The detected word images are then normalized for a robust recognition network. Because accurate recognition requires large volumes of training data but manually labeled data is relatively scarce, we introduce a dynamic map text synthesizer providing a practically infinite stream of training data. Results are evaluated on a labeled data set of 30 maps featuring over 30,000 text labels.

Recognizing the place names within textual labels on historical maps is complicated by many factors, such as curvilinear baselines and dense overlap with other textual or graphical elements. However, maps' alignment with known geography and inter-label typographic style consistencies provide strong cues for resolving uncertainty and reducing text recognition errors. We present a unified probabilistic model to leverage the mutual information between text labels and styles and their geographical locations and categories. This work also introduces likelihood functions to model label placement for polyline and polygon geographical features, such as rivers or provinces. We evaluate the methods on 30 maps from 1866-1927. By interleaving automated map georeferencing with text recognition, we reduce word recognition error by 36% over OCR alone. Incorporating category-style links reduces toponym matching error by 32%.

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 search ability 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.

Using a lexicon can often improve character recognition under challenging conditions, such as poor image quality or unusual fonts. We propose a flexible probabilistic model for character recognition that integrates local language properties, such as bigrams, with lexical decision, having open and closed vocabulary modes that operate simultaneously. Lexical processing is accelerated by performing inference with sparse belief propagation, a bottom-up method for hypothesis pruning. We give experimental results on recognizing text from images of signs in outdoor scenes. Incorporating the lexicon reduces word recognition error by 42% and sparse belief propagation reduces the number of lexicon words considered by 97%.