IJDAR (2000) 2: 177–185
Integrated text and line-art extraction from a topographic map
Luyang Li1 , George Nagy2 , Ashok Samal3 , Sharad Seth3 , Yihong Xu4
1
2
3
4
Panasonic Information and Networking Laboratories, Princeton, NY, USA
Rensselaer Polytechnic Institute, Troy, NY, USA
University of Nebraska – Lincoln, Department of Computer Science, Lincoln, NE 68588, USA
Hewlett-Packard Laboratories, Palo Alto, CA, USA
Received January 1, 2000 / Revised January 21, 2000
Abstract. Our proposed approach to text and line-art
extraction requires accurately locating a text-string box
and identifying external line vectors incident on the box.
The results of extrapolating these vectors inside the box
are passed to an experimental single-font optical character reader (OCR) program, specifically trained for the
font used for street labels. In the first evaluation experiment, automated techniques are used to identify the
boxes and the line vectors. In the second, more comprehensive, experiment an operator marks these using a
graphical user interface. OCR results on 544 instances
of overlapped street-name boxes show the following improvements due to the integrated processing: the error
rate is reduced from 4.1% to 2.0% for characters and
from 11.8% to 6.4% for words.
Key words: Graphics recognition – Map processing –
Text-graphics segmentation – Geographic information
systems – Character recognition – Cooperative processing – Interactive processing – Adaptation – Form processing
1 Introduction
Both government and business organizations must frequently convert existing paper maps of cities and larger
regions into a computer-readable form that can be interfaced with existing geographical information systems
(GIS). Often, the task involves a batch of maps of the
same kind, but each covering a different geographic area.
Examples might be a set of topographic maps or road
maps of a foreign country issued by the same agency.
Until recently, coordinate digitizing tables were used
for such conversion. Currently, the most efficient method
is to raster-scan the maps, display them in segments on
a computer screen, and digitize them using interactive
CAD (e.g., Autocad) or GIS (e.g., ARC/INFO) tools.
Parts of the process, such as tracing a single curve, might
Correspondence to: S. Seth
be automated. Completely automatic conversion, while
still the objective of many academic research projects,
remains elusive. However, when a large amount of similar
data needs to be converted at the same time, it is highly
desirable to reduce the amount of manual intervention
required as the task progresses.
In 1997, the United States National Imagery and
Mapping Agency (NIMA) sponsored a series of research
activities as part of the Intelligent Map Understanding
Project. The project was designed to fill both in-house
and commercial needs. The University of NebraskaLincoln and Rensselaer Polytechnic Institute jointly participated in these activities. Our primary contribution in
this work was to identify the bottlenecks in large-volume
conversions using a general-purpose system and develop
a framework under which the operator time for these
tasks could be progressively reduced.
Our specific task was to develop methods to extract
the street network and street labels from the bitmap representation of a United States Geological Survey (USGS)
topographic map, and convert them to a GIS format similar to that of the TIGER database developed by the Bureau of Census [USD97]. Here the streets are polylines,
and the associated street names are in ASCII. (We did
not attempt to duplicate the TIGER city-block structure.) The main research objective was to demonstrate
that iterated processing, feedback from downstream processes, and judicious use of operator interaction improve
the results.
In the USGS map, street lines and names are all
printed in black. Thus separating the black layer with
a color filter provides a starting point for processing the
image. However, major challenges remain even with this
simple and effective step of information reduction. The
text and line layers frequently overlap; text strings appear in many orientations; the line layer contains not
only streets but also grid lines and political boundaries.
Similarly, the black text layer contains not just the street
names but also elevation values and names of neighborhoods, political districts, schools, hospitals, other builtup areas, physiographic and topographic features. These
challenges are well illustrated in the black layer of a 3-
178
L. Li et al.: Integrated text and line-art extraction from a topographic map
Fig. 1. The black layer of a section of USGS map
inch square chip from the Washington D.C. East map in
Fig. 1.
Because map conversion is a niche market, with government agencies serving as the primary customer base,
it has not received as much attention in the commercial
world as the conversion of forms and legal documents.
However, with the spread of the Web and GPS technologies, navigational databases are finding use in invehicle navigation, intelligent traffic and transportation
systems, and such online applications as routing and city
guides. Street layer extraction is the first step towards
building detailed navigational databases. As such, any
significant improvement in this process will help alleviate the sparse coverage of the existing databases. For
example, in the Navtech database [Nav], detailed area
maps are currently available for only four cities in the
entire states of Iowa, Kansas, Missouri, Nebraska, North
Dakota, and South Dakota.
Preliminary results were presented at a session organized by NIMA at the Conference on Geographic and
Land Information Systems (GIS/LIS) [NSS+ 97], at the
Workshop on Graphics Recognition of the International
Association of Pattern Recognition [NSS+ 98], and at
the International Conference on Document Analysis and
Recognition (ICDAR) [LNS+ 99]. The results also appear in four theses [Kal97,Li98,Siv98,Xu98]. This paper presents new results obtained since the conclusion of
the NIMA project but it draws on methods discussed in
detail in these references.
Briefly, in our past work we accomplished the following steps on a single map. We isolated the black layer
of the map, which contains both the street lines and the
street labels, by appropriate thresholding of the RGB
color components viewed as an HSV (hue-saturationvalue) representation. We separated the 18, 848 × 23, 631
pixel black layer into 32,000 connected components (this
took about 4 min ‘system time’ on a Sun 4M SPARCstation). Then we carried out a preliminary classification
of each component, according to geometric features and
black pixel density, into line art, characters, and icons.
The line art was vectorized using ArcScan [ESRI97],
and refined (‘beautified’ [PW85]) using line-width, linelength, and line-pairing criteria, and finally the constraint
was applied that the street casings must form city blocks.
The double-line street casings were then reduced to a
street centerline representation. The characters were
grouped into ‘word boxes’, rotated to horizontal, and
submitted to an experimental OCR program that did not
require character-level separation and which was specifically trained for the font used for street labels. (We also
tried a commercial omnifont system but found that it
produced extremely poor results on the map text.)
The labels were analyzed according to the direction
and spacing of nearby street lines. The specific and
generic components of the labels (like ‘George Washington’ and ‘Ave’) were assembled, and the street lines
were traced as far as possible and associated with the
complete (and occasionally repeated) labels.
Other important parts of the project were the implementation on top of ARC/INFO of a display of partial
or complete results to allow interactive correction of errors, of a logging system that tracks the time spent by
the operator on various tasks, and of a cost model of
the complete process for estimating the conversion time
required for new tasks in terms of map and operatorrelated parameters.
The novel aspect documented here is the following.
We demonstrate the effect of cooperative processing of
text and line art. This is necessary because the text and
line art often overlap and recognition results on one can
help recognition of the other. We use information from
the street lines to locate and orient label boxes and to
remove overlaying line segments, and use the output of
the character recognition system to refine the street-line
sublayer. We show that with cooperative processing the
error rate of our word-based OCR algorithm can be reduced by a factor of two on noisy samples.
We briefly review relevant sources of information.
The definitive authority on text-graphics separation is
Kasturi. His work with his students stretches over a
decade [FK88, KA88, TK98]. It usually gives much more
emphasis to image processing than to OCR. A good
source of other techniques devised in the last 20 years
for isolating text from illustrations in printed documents,
and for line-text separation in engineering drawings, is
[OK95]. Brief reviews appear in [NSS+ 97] and
[Li98].
There are many, many papers on map conversion in
general, with a particularly valuable early paper by a
practitioner, Rhind [Rhi74]. Good examples of recent
projects are [StK95, Har95], while surveys of the required
tools appear in [EAK95, JV97]. Since we use commercial
software, we don’t review the even larger field of thinning
and vectorization, but a recent survey is [Doe98].
There is not a mass of literature on associating streets
with street labels on scanned maps. The most relevant
paper we found is by G. Myers and his colleagues at SRI
L. Li et al.: Integrated text and line-art extraction from a topographic map
Fig. 2. A street label and street casings from a USGS map
(enlarged)
[MMC+ 96]. They propose a verification-based approach,
but compared to ours the line art and the text are processed essentially separately.
Finally, we note that the problem of separating text
from lines arises in forms processing, which is a major
industry. Although most of the successful techniques are
proprietary, some recent results appear in [NSY96].
2 Data
The source map for all of the following experiments was
WASHINGTON EAST (D.C.) QUADRANGLE, based
on aerial photography taken in 1955. It was revised in
1965, photorevised (as shown in purple) and edited, but
not field checked, in 1979. The map has a scale of
1:24,000 and elevation contour intervals of 10 feet. It is a
polyconic projection based on the 1927 North American
Datum. It shows the Maryland and Virginia state coordinate systems as well as the 1000-m Universal Transverse
Mercator grid.
The size of the map (46 cm × 58 cm) dictates special
processing considerations. The map was color-digitized
by NIMA at 1000 dpi ( 40 lpm) with 8 bits per pixel
onto a CD-ROM. The uncompressed image is 455 MB,
which reduces to 71 MB after lossless compression. Since
such a large image can neither fit in typical workstation
memory nor can be displayed at full resolution, we implemented the retrieval of rectangular ‘chips’ of arbitrary
size and location using either pixel or latitude-longitude
coordinates. The map can thus be processed as a mosaic
of adjacent chips.
The USGS 7.5-minute series of topographic maps is
typical of the best in traditional cartography and packs
an enormous amount of information for diverse uses. The
extraction of data from such a map is correspondingly
more difficult than from specialized maps like cadastral and road maps, or from map separates (overlays).
High-resolution digitization is desirable because at lower
sample-spacing fine lines break, character shapes are distorted, it is even more difficult to segment glyphs from
street lines, and the regularity of the halftone pattern is
destroyed by aliasing.
179
The focus here is on street lines and labels in the
black layer (Fig. 2). In addition to street casings, 0.1 mm
wide and separated by 0.5 mm, the black layer graphics include solid blocks showing buildings in non-builtup areas like parks - some with flag (school) or cross
(church) icons attached; black outlines for running tracks
and other surface features; thick railroad lines; dashed
political boundary lines; and straight grid lines slightly
thinner than the street casings.
There are half-a-dozen fonts and sizes of text. The
street labels are slanted sans-serif caps 1.25 mm high or,
for principal streets, 1.4 mm high. Text in other fonts
denotes districts, monuments, schools, elevation benchmarks, etc. The street labels are typically oriented from
left to right for E-W streets, and from bottom to top
for N-S streets for viewing from the right. These conventions conflict for streets that run NW to SE: there
is a 10-degree region where the orientation is inconsistent. Street labels frequently overlay intersecting street
casings, and occasionally touch icons.
3 Methodology
The interaction between the line art and text processing at several points in the data flow is shown in Fig. 3.
The initial processing examines every pixel only once to
recover the black layer and extract its connected components. Then the candidate constituents of street names
and street casings are identified. The initial classification
is, however, error prone and the resulting errors affect
subsequent OCR and vectorization. The results can be
improved by detailed examination of the vicinity of identified character candidates (Fig. 4). The local processing
consists of two interlaced steps: character grouping and
line structure recovery.
Character grouping
Character grouping is based on the character neighborhood graph, which represents the spatial relations between connected components in a manner similar to that
of O’Gorman’s DocStrum [O’G93]. Characters are assembled into word boxes according to the constraint that
they lie on a straight line and their centroids are separated by about 1.4 times their average width.
In order to extract only street name boxes, the neighboring black pixels are searched for parallel line configurations with the appropriate spacing to establish the
dominant direction of the street-name string. This is particularly useful when nearly all the characters in a string
touch street lines. Combining the information from the
neighboring street casings and the character neighborhood graph, characters are grouped into word boxes to
form the longest possible aligned string. Strings are recursively merged as long as they don’t violate orientation
and word separation constraints. The resulting strings
are subjected to an additional uniformity-of-height constraint to distinguish street words from other text labels.
An example of the resolution of groups of adjacent characters into street-label word boxes is shown in Fig. 5.
180
L. Li et al.: Integrated text and line-art extraction from a topographic map
Color
Map Image
Color Separation
Black Layer
Initial Layer Separation
Text Layer
Line Layer
Character
Grouping
Vectorization
Line Vectors
Text Bounding Boxes
Line Structure
Recovery
Marked Text Image
Character
Template
Matching
Character Templates
Cooperative Processing
Line
Image
Fig. 5. Street-label character grouping. In the label ‘UPSHUR’, three of the six characters are touched by overlapping
street lines, yet the word is found
Text
Image
Fig. 3. Sublayer separation
Fig. 6. Line structure recovery based on graphics extrapolation. From top to bottom: original label box; segmented text;
and segmented street lines
Fig. 4. Local analysis. Above, vectors in the vicinity of a
label box. Below, paired vertical vectors identified and extrapolated as street-lines
The above character-grouping scheme is not yet robust enough to fully recover street labels with a curved
baseline that follow curving streets. Such labels are rare
in Washington East (D.C.) but might be more common
in some suburbs. The scheme also has other shortcomings. It cannot reliably distinguish street-name font from
other text or graphics.
Line structure recovery
The objective of this step is to recover street-line casings
that are interrupted because they overlap text. A black
pixel in a word-box can belong to a character, to a casing,
or to both.
All the vectors in the vicinity of each word box are analyzed. Short vectors which are not aligned with longer
vectors are typically parts of characters. Long vectors
that intersect the word bounding box are extended and
connected, and assigned to the street-line layer. The output of the character recognition routine is used to mark
the location of text pixels. The resulting sublayers, segmented text and segmented street lines, are both subsets
of the original black layer. An example is shown in Fig. 6.
Character template matching
Our OCR is based on template matching which, in addition to converting the word-box bitmaps to symbolic
alphanumeric ASCII labels, allows further improvement
of the line art and text segmentation.
Prototypes for each character class are first extracted
from bitmaps of operator-labeled training data. We note
that labeling a hundred or so word boxes is quite rapid,
because the operator need not segment the characters.
L. Li et al.: Integrated text and line-art extraction from a topographic map
181
Fig. 9. Templates constructed from the training data (first
experiment)
Fig. 7. Label box, matching templates, and residue
Fig. 10. Examples of test data for street labels, rotated to
horizontal
Fig. 8. Examples of training data for street labels, rotated
to horizontal
The segmentation-free prototype extraction and recognition algorithms are described in [NX97,XN99].
The extracted prototypes are matched against the
word boxes. The classification is based on the best fit
for the whole word box, which is found using a levelbuilding algorithm. The matching algorithm can ignore
pixels that are flagged as belonging to line art on grounds
that we cannot know the color of the underlying character pixel. Thus the recognition result itself can discriminate between these two possibilities, as the best-fitting
template should overlay the character pixels.
The residual line art overlaying some character can
introduce a mistake only if it distorts the overlaid character into another recognizable class, which is relatively
rare. (This type of error could be corrected by reference
to a street-name gazetteer.)
The important point is that the best-fitting templates
usually form a clean street label, and the residue that is
not covered by them (Fig. 7) can be analyzed, as mentioned above, for segmentation. The recognized characters could in turn be used to improve the initial templates which are based on too few operator-identified
character samples. All the word boxes could then be reclassified with the improved templates.
4 Experiments
The first experiment shows the performance of our custom character-recognition system on unsegmented label
boxes. The 32 templates extracted from the 92 streetname boxes used as the training set (examples in Fig. 8)
are shown in Fig. 9. Darker shades indicate higher probability of black. A few examples of the 733 automaticallyextracted street name boxes used for testing are displayed in Fig. 10.
After determining the best-fitting combination of
templates that fits a label, the character recognition system may still reject some characters if they are not
Table 1. Results of street label recognition
Street Labels
Other Text
Graphics
Correct
Wrong
Rejected
1285
151
0
57
57
46
94
275
573
matched with high enough confidence. Characters that
are part of street labels are recognized with an accuracy of 95% at a reject rate of 8%. (The error rate was
determined by a string-comparison algorithm.) Most of
the graphics fragments are rejected, but some lines perpendicular to the street box are recognized as ‘I’ or ‘L’
(here the line pixels were not marked). Most alphabetic
characters that do not originate from street labels are
rejected, but when they are not, they are often misrecognized. The complete character-level results are presented
in Table 1.
In our earlier work, the street-line layer was evaluated
on an 8” × 8” section of the map [NSS98]. Comparison
against the digital line graph (DLG) indicated that the
vectorization was 97.5% accurate before operator correction, but only 36% of the street lines were extracted.
The positional accuracy of the intersections was 12 m,
which is just within the National Map Accuracy Standards. An interactive session raised the completeness of
the vectorization to 97%, with an accuracy of 8 m. Of
the 3% residual error 1% is due to some missed streets,
and 2% to some of the corrected streets that ran outside
the chip boundary, where there was no DLG data.
The DLG does not contain street names, so these
could not be verified independently. The geometrical accuracy of the TIGER database, which does contain the
street names, was too low to be used for comparison.
Because of the aforementioned limitations of the current box-finding algorithm (see Fig. 10) and the vectorizer, a comprehensive evaluation of our integrated approach to text and line-art extraction was not possible
with automatically extracted boxes and vectors. Therefore, a custom graphic user interface was designed to
182
L. Li et al.: Integrated text and line-art extraction from a topographic map
Fig. 11. Left: an example street-word box (rectified) with
incident vectors. Right: the result of extrapolation of line vectors
allow an operator to draw the word boxes and the incident vectors. Additionally, the operator could transcribe
the word in ASCII to serve as ground truth for OCR.
The interface was instrumented to log the time spent on
individual steps. The black layer for the whole map was
divided into 12 overlapping chips to simplify data entry. Pause and resume buttons allowed the operator to
break the task of capturing all the street-name boxes on
the map into several sessions.
The left part of Fig. 11 shows an example box with
incident line vectors. The time log revealed that it took
the operator 7.6 h to enter 966 boxes. Of this total time,
60% was spent in drawing boxes, 24% in drawing vectors,
and only 16% in entering text. To simplify the dataentry task, the boxes were initially marked with color
markers on hard copies of chip images. This took about
4 h. Note that these operations were necessary only for
the evaluation of the new OCR algorithm and will not
be required in an automated system.
The operator was asked to ignore the 64 street labels with curved baselines (0.2% of the total number of
boxes). The data files containing the box and the incident vector information were post-processed to extrapolate the vectors within the box and mark the corresponding pixels for OCR with the street-line information (see
the right part of Fig. 11).
Of the 966 boxes, 422 were clean (had no line overlap). Ignoring the many duplicate generic names (‘ST’,
‘AVE’, etc.) from among these, a total of 185 were used
as the training set for the OCR program. The 544 boxes
with line overlap were used for evaluation. Figure 12
shows the set of templates constructed from the clean
glyphs. Note that because these are based on a clean and
much larger sample compared to those shown in Fig. 9,
better OCR results can be expected.
The results of OCR with and without vector information are shown in Table 2. The chips were 7-inch square
(with one inch overlap horizontally and vertically with
the adjoining chips) except for the right-most chips (second index = 2) and the bottom-most chips (first index
= 3) which were smaller. The second and fifth columns
show, respectively, the total number of characters and
words in each chip. The third and fourth columns show
the number of character errors for OCR without and
with vector information, respectively. The corresponding word-error data is shown in the sixth and seventh
Fig. 12. Templates constructed from the training data (second experiment)
(a)
(d)
(b)
(e)
(c)
(f)
Fig. 13. A sample of OCR results with and without consideration of lines. In each case, from top to bottom: original
box image, lines, OCR with lines, and OCR without lines
columns. The cumulative results in the last row show
that with the vector information the character error was
reduced from 4.1% to 2.0% and the word error was reduced from 11.6% to 6.4%. The error count does not include ‘1-I’ confusions, as those are identical glyphs on the
map and cannot be discriminated without context. The
count does include nine character errors due to the lack
of a ‘Q’ template because of which all the Q’s were recognized as O’s. We emphasize that only the overlapped
boxes were considered in this test. The error rate would
be lower if a representative sample of clean boxes were
included in the test.
The six sample results in Fig. 13 illustrate text and
line-art extraction with and without vector information.
In each case, the first row shows the original box image with crossing lines, the second row the extrapolated
lines, the third row the OCR result without vector information, and the fourth row the OCR result with vector
information.
In case a, the letter ‘U’ is recognized as the pair
‘6I’ because of the line crossing this letter in the center. When the line pixels are ignored, the ‘U’ is correctly
recognized. In case b, the line next to ‘L’ causes the best
match to occur for the template ‘E’ instead of ‘L’; when
the line pixels are ignored, this error disappears. Case c
illustrates the rare situation when ignoring the line pixels causes an error. Here OCR makes different errors in
L. Li et al.: Integrated text and line-art extraction from a topographic map
183
Table 2. OCR performance with and without vector information
Chip ID
# chars
Char Errors
w/o vector
with vector
information information
# Words
Word Errors
w/o vector
with vector
information information
s00
s01
s02
s10
s11
s12
s20
s21
s22
s30
s31
s32
250
251
98
303
368
216
167
146
171
108
194
71
10
9
10
6
21
7
11
4
1
4
12
1
3
5
4
8
11
5
2
2
1
2
3
1
60
49
19
63
75
49
43
38
39
31
53
25
5
6
5
4
14
7
8
3
2
2
8
1
3
3
4
4
7
5
2
2
1
2
1
1
Total
2343
96(4.1%)
47(2.0%)
544
64(11.8%)
35(6.4%)
the two cases. With the line pixels considered, the bestmatching template is ‘8’. Without the ggggg the bottom
stroke of ‘E’ disappears and because of the residual line
pixels at the top the best match is for ‘T’. In cases d
and e, the crossing lines add enough black pixels to cause
erroneous recognition of the letters ‘E’ and ‘P’, respectively; these errors are corrected when the line pixels are
ignored.
Finally, case f shows one of the three words that were
correctly recognized without the line information but incorrectly recognized with this information. The line pixels deleted enough of the stem pixels of ‘I’ to cause a
better match for ‘T’. What is also clear from these examples (and from Table 2) is that the OCR algorithm
is able to tolerate a fair amount of noise but its performance is further improved with the line information.
5 Conclusion
Beginning with the connected components of the black
layer of the map, we have shown that almost all streetlabel boxes can be located and extracted using character alignment and size, and the graphic context of the
location of labels adjacent to street lines. Most boxes
are correctly oriented, but sizing these boxes precisely
is more difficult. Even after extensive processing, many
boxes contain some graphics fragments, and some text
characters are missing. This is an area we are currently
pursuing for further improvement.
We have presented a template-based character recognition method that does not require character-level segmentation because it recognizes an entire string of characters. This recognizer tolerates overlapping graphics
segments much better than commercial omnifont OCR
and is able to take advantage of the overlapping line
information, thereby further reducing its error rate by
half. The resulting accuracy appears sufficient to invoke
context-based processing (i.e., a gazetteer) to advantage.
We have shown that the integrated approach also
helps eliminate character fragments (within the boxes)
from the vector layer. In future research we intend to
exploit this to improve vectorization.
The percentage of correctly extracted street vectors,
and of correctly extracted and recognized labels, is evidently far too low to be useable without extensive operator interaction. In previous work, we estimated that the
automated processing that we had developed reduced the
time required to digitize the street network from about
24 h to 16 h. With some of the additional improvements
mentioned here, especially close interaction between text
and graphics processing, it appears feasible to reduce the
overall time to about 10 h.
According to the operator log resulting from the correction of the 8” × 8” chip, the most time-consuming
aspects of operator intervention are vectorization, identification of the street label locations, and association
of street labels with street lines. Our methods are quite
effective for the last two of these three tasks, label box
positioning and association. As shown earlier, few labels
are missed entirely, and given correct street vectors and
street labels, the association is practically foolproof.
It therefore appears desirable to concentrate further
effort on more complete vectorization, perhaps even by
relaxing the constraints that now guarantee high accuracy. This will have the greatest direct impact on reducing operator time, as well as a beneficial secondary
effect on label box location. Further improving the accuracy of character recognition will have only a minor
effect, because typing-in the correct labels of the rotated
label boxes, without the need to point-click the location
or association, is very fast.
Wholly automatic conversion of documents of the
complexity of topographic maps is still well over the horizon, but tools can be developed for reducing the cost of
conversion. We are continuing our research with this objective in mind.
Acknowledgements. We gratefully acknowledge the hours of
careful work by Chakravarthy Terlapu in capturing the
street-label box information. This work directly followed the
184
L. Li et al.: Integrated text and line-art extraction from a topographic map
recent project supported by the National Imagery and Mapping Agency as a part of the Intelligent Map Understanding
Project. Support was also provided by the Central Research
Laboratory, Hitachi, Inc., and the NRI Geospatial Decision
Support Grant of the University of Nebraska. Part of this
work was carried out in the New York State Center for Automation Technologies (CAT) at Rensselaer Polytechnic Institute. The CAT is partially funded by a block grant from
the New York State Science and Technology Foundation.
References
[Doe98]
D. Doermann. An introduction to vectorization
and segmentation. Graphics Recognition: Algorithms and Systems, LNCS 1389, Berlin Heidelberg New York: Springer, pp. 1–8, 1998
[EAK95] L. Eikvil, K. Aas, H. Koren. Tools for interactive
map conversion and vectorization. IEEE Conf.
on Comput. Vision and Pattern Recognition, pp.
937–930, 1995
[ESRI97] Environmental Systems Research Institute Inc.
ARC/INFO Software. Redlands, CA, 1997
[FK88]
L.A. Fletcher, R. Kasturi. A robust algorithm for
text string separation from mixed text-graphics
images. IEEE Trans. on Pattern Anal. and Mach.
Intell., 10(6): 910–918 (1988)
[Har95]
J. Hartog. A framework for knowledge-based map
interpretation. Doctoral Dissertation, Technical
University of Delft, 1995
[JV97]
R. Jansen, A. Vossepoel. Adaptive vectorization
of line drawing images. Comput. Vision and Image Understanding, 65(1): 38–56 (1997)
[KA88]
R. Kasturi, L. Alemany. Information extraction
from images of paper-based maps. IEEE. Trans.
SE, 14(5): 671–675 (1988)
[Kal97]
K. Kalafala. Identification of street lines and
intersections in a scanned urban topographic
map. MS Thesis, Rensselaer Polytechnic Institute, Troy, 1997
[Li98]
L. Li. Adaptive text/line separation in document
images based on vectorization and OCR. MS Thesis, University of Nebraska - Lincoln, 1998
[LNS+ 99] L. Li, G. Nagy, A. Samal, S. Seth, Y. Xu. Cooperative text and line-art extraction from a topographic map. Proc. 5th Int. Conf. on Document
Anal. and Recognition (ICDAR ’99), pp. 467–470,
1999
[MMC+ 96] G.K. Myers, P. Mulgaonkar, C. Chen, J. DeCurtins, E. Chen. Verification based approach
for automated text and feature extraction
from raster-scanned maps. Graphics Recognition:
Methods and Applications, LNCS 1072, Berlin
Heidelberg New York: Springer, pp. 190–203,
1996
[Nav]
Navtech. Navigational Technologies, home page:
http://www.navtech.com/
[NSS+ 97] G. Nagy, A. Samal, S. Seth, T. Fisher, E. Guthmann, K. Kalafala, L. Li, S. Sivasubraniam,
Y. Xu. Reading street names from maps - technical challenges. Proc. GIS/LIS Conf., Cincinnati,
OH, pp. 89–97, 1997
[NSS+ 98] G. Nagy, A. Samal, S. Seth, T. Fisher, E. Guthmann, K. Kalafala, L. Li, S. Sivasubraniam,
Y. Xu. A prototype for adaptive association of
street names on maps. Graphics Recognition: Algorithms and Systems, K. Tombre,A. Chhabra,
editors, LNCS 1389, Berlin Heidelberg New York:
Springer, pp. 302–313, 1998
[NSY96] S. Naoi, M. Suwa, M. Yabuki. Quantitative evaluation of improved global interpolation in the segmentation of handwritten numbers overlapping a
border. IEICE Trans. Inf. Syst., E79-D(5), May
1996
[NX97]
G. Nagy, Y. Xu. Automatic prototype extraction
for OCR. Proc. Int. Conf. on Document Anal. and
Recognition, ICDAR’97, pp. 278–282, 1997
[O’G93]
L. O’Gorman. The document spectrum for page
layout analysis. IEEE Trans. PAMI, 15(11):
1162–1173 (1993)
[OK95]
L. O’Gorman, R. Kasturi. Document Image Analysis. IEEE Computer Society Press, New York,
1995
[PW85]
T. Pavlidis, C.J. Van Wyk. An automatic beautifier for drawings and illustrations. Procs. ACM
SIGGRAPH, pp. 225–234, 1985
[Rhi74]
D. Rhind. An introduction to the digitizing and
editing of mapped data. Automation in Cartography. In: P. Dale (ed.), British Cartographic Society, 1974
[Siv98]
S. Sivasubramaniam. Street name association in
digitized maps. MS Thesis, Rensselaer Polytechnic Institute, Troy, 1998
[StK95]
A. Smeulders, T. ten Kate. Software systems design for paper map conversion. Graphics Recognition Methods and Appl., pp. 204–211, 1995
[TK98]
P. Tofani, R. Kasturi. Segmentation of text from
color map images. Technical Report CSE-98-007,
Department of CSE, Penn State University, 1998
[USD97] USDC. Tiger tales: a presentation of the automated geographic support system. US Department of Commerce, Bureau of Census, 1997
[XN99]
Y. Xu, G. Nagy. Prototype extraction and adaptive OCR. IEEE Trans. PAMI, 21(12): 1280–1296
(1999)
[Xu98]
Y. Xu. Prototype extraction and OCR. PhD Thesis, Rensselaer Polytechnic Institute, Troy, 1998
L. Li et al.: Integrated text and line-art extraction from a topographic map
185
Luyang Li
Sharad Seth
George Nagy
Yihong Xu
Ashok Samal