Introduction
In the recent years most libraries around the world have initiated
projects for the complete digitization of their map collections so as
to enable easy access to information. When it comes to paper
documents, optical character recognition (OCR) techniques are widely
used for text digitization. Although maps contain a wealth of
alphanumeric information, recent literature has indicated very limited
examples of OCR implementation on maps. These examples are mainly
focused on the extraction of textual information, whereas the
extraction of numerical information has not been examined
separately. Purpose of this research is to evaluate how an OCR
algorithm can be used to extract numerical information from maps in
form of bathymetric points and house numbers, which can ultimately
lead in the automatic generation of geographic information
layers. Additionally, the possibility of an automated georeference
method is examined based on the recognition of coordinates on the map
borders and their assignment to the corresponding grid intersections.
Methodology
To evaluate the proposed method an OCR algorithm was developed in
Python language, using the OpenCV library. The overall process can be
separated into 3 main stages (Fig.1). The first is the pre-processing of the
image, which includes the conversion to greyscale
and the binarization of the image. Additional filters of Gaussian blur
and Erosion/Dilation are used in cases of complicated background or
noise (Fig.2). The feature extraction stage involves the digit/graphics
separation, the segmentation of individual digits and the extraction of
a feature set that can describe each of them (Fig.3). In this research the
intensity values of the pixels of each segmented images were used as
descriptors. The last stage of the OCR algorithm is the
classification, where each segmented image is classified as one of the
ten possible digits. The adjacent digits that form one multi-digit
number are grouped together and then assigned in the correct position
on the map (Fig.4). Regarding the proposed method of georeference, Hough
transform is used for the detection of the grid lines and then the
intersections, which will serve as control points, are identified (Fig.6). OCR
is applied then on the close area of each intersection and the
recognised coordinates are being assigned to the intersection (Fig.5).
Results and further improvements
The produced algorithm achieves a very good recognition rate and most digits - in form of
bathymetric points - were succesfully recognised (Fig.4). However,
the overall accuracy of the algorithm is highly dependent on the characteristics of each map
and the complexity of the graphics. Thus, customisation is necessary for application on different map
series. Regarding the georeference, the complete automation of the process is not an easy task, as each individual map has to be handled differently, requiring the user’s intervention in multiple steps throughout the process.
(Fig.5). Undoubtedly, the algorithm could be improved to tackle some
of the problems that
are encountered throughout the whole process. For example, the multi-oriented
digits can be succesfully reognised using rotation inveriant
descriptors. Additionally, the developed OCR algorithm could be
enriched to support the recognition of latin alphabet characters. This
would enable the recognition of toponyms, which would in turn facilitate the
georeference process by locating roughly the position and scale of the
map with an alignment to a gazeteer of known place names.
Recommented Literature
1. Adam, S., Ogier, J., Cariou, C., Mullot, R., Labiche, J.,
Gardes, J., 2000, Symbol and character recognition: application to
engineering drawings, International Journal on Document Analysis and
Recognition, 3(2), pp. 89–101.
2. Cheriet, M., Kharma, N., Liu, C.L., Suen, C.Y., 2007,
Character Recognition Systems: A guide for Students and Practitioners,
New Jersey: John Wiley & Sons Inc.
3. Chiang, Y., Knoblock, C.A., 2010, An Approach for Recognizing
Text Labels in Raster Maps, Proceedings of the 20th International
Conference on Pattern Recognition, pp. 3199-3202.
4. Chiang, Y., Knoblock, C.A., 2011, Recognition of
multi-oriented, multi-sized, and curved text, Proceedings of the
11th International Conference on Document Analysis and Recognition,
pp. 1399-1403.
5. Chiang, Y., Knoblock, C.A., 2012, Generating Named Road Vector
Data from Raster Maps, Geographic Information Science, Lecture Notes
in Computer Science, 7478, pp. 57-71.
6. Deseilligny, M. P., Mena, H. L., Stamonb, G., 1995, Character
string recognition on maps, a rotation- invariant recognition
method, Pattern Recognition Letters, 16(12), pp. 1297–1310.
7. Jatnieks, J., 2010, Open Source Solution for Massive Map Sheet
Georeferencing Tasks for Digital Archiving, Lecture Notes in Computer
Science, 6102, pp. 258-259.
8. Pouderoux, J., Gonzato, J.C., Guitton, P., Pereira, A., 2007,
Toponym Recognition in Scanned Color Topographic Maps, Document
Analysis and Recognition, 1, pp. 531-535.
9. Titova, O.A., Chernov, A.V., 2009, Method for the Automatic
Georeferencing and Calibration of Cartographic Images, Applied
Problems - Pattern Recognition and Image Analysis, 19(1), pp. 193-196.