Indexing of Handwritten Historical Documents - Recent Progress

R. Manmatha
Toni M. Rath
Center for Intelligent Information Retrieval
Computer Science Department
University of Massachusetts Amherst

Abstract
Indexing and searching collections of handwritten
archival documents and manuscripts has always been
a challenge because handwriting recognizers do not
perform well on such noisy documents. Given a collection of documents written by a single author (or a
few authors), one can apply a technique called word
spotting. The approach is to cluster word images
based on their visual appearance, after segmenting
them from the documents. Annotation can then be
performed for clusters rather than documents.
Given segmented pages, matching handwritten
word images in historical documents is a great challenge due to the variations in handwriting and the
noise in the images. We describe investigations
into a number of different matching techniques for
word images. These include shape context matching,
SSD correlation, Euclidean Distance Mapping and
dynamic time warping. Experimental results show
that dynamic time warping works best and gives an
average precision of around 70% on a test set of
2000 word images (from ten pages) from the George
Washington corpus.
Dynamic time warping is relatively expensive and
we will describe approaches to speeding up the computation so that the approach scales. Our immediate goal is to process a set of 100 page images with
a longer term goal of processing all 6000 available
pages.

Introduction

Libraries contain an enormous amount of handwritten historical documents. Such collections are interesting to a great range of people, be it for historians,
students or just curious readers. Efficient access to
such collections (e.g. on digital media or on the Internet) requires an index, for example like in the
back of a book. Such indexes are usually created by
manual transcription and automatic index generation from a digitized version. While this approach
may be feasible for small numbers of documents, the

cost of this approach is prohibitive for large collections, such as the manuscripts of George Washington
with well over 6000 pages.
Using Optical Character Recognition (OCR) as
an automatic approach may seem like an obvious
choice, since this technology has advanced enough
to make commercial applications (e.g. tablet PCs)
possible. However, OCR techniques have only been
successful in the online domain, where the pen position and possibly other features are recorded during writing, and in offline applications (recognition
from images) with very limited lexicons, such as
automatic check processing (26 words allowed for
legal amount field). However, for general historical documents with large lexicons, and the usually
greatly degraded image quality (faded ink, ink bleedthrough, smudges, etc.), traditional OCR techniques
are not adequate. Figure 1 shows part of a page
from the George Washington collection (this page is
of relatively good quality).
Previous work [23] has shown the difficulties that
even high-quality handwriting recognizers have with
historical documents: the authors aligned a page
from the Thomas Jefferson collection with a perfect
transcription of the document. The transcription
was used to generate a limited lexicon for each word
hypothesis in the document. With a limited lexicon, the recognizer only has to decide which of a
few possibilities the word to be recognized is. However, even for very small lexicon sizes of at most 11
(ASCII) words per word hypothesis in the image,
only 83% of the words on a page could be correctly
aligned with the transcription.
The wordspotting idea [12] has been proposed as
an alternative to OCR solutions for building indexes
of handwritten historical documents, which were
produced by a single author: this ensures that identical words, which were written at different times,
will have very similar visual appearances. This fact
can be exploited by clustering words into groups
with image matching techniques. Ideally, each clus-

Figure 1: Part of a scanned document from the George Washington collection.

ter of word images consists of all occurrences of a
word in the analyzed document collection. Clusters
that contain the most words can then be annotated
in order to allow an index generation for the respective words1 .
Apart from the problem of segmenting images
from documents (see [14] for a scale-space approach),
the crucial step is to determine the similarity between images of words. In this work, we present
our recent investigations into robust matching techniques for words in historical manuscripts, and evaluate their performance on a subset of the George
Washington collection.

Matching Techniques

The matching techniques we have investigated fall

roughly into two categories: image matching approaches that compare images pixel-by-pixel, and
feature-oriented techniques, that extract image features and compare them after determining correspondences between them.

2.1

Pixel-by-Pixel Matching

The matching techniques in this section compare two

images pixel-by-pixel after aligning them initially.
The alignments compensate part of the variations
which are inherent to handwriting (e.g. shear and
scale changes). Some of the matching techniques
that have been investigated include:
1. XOR[8, 13]: The images are aligned and then
a difference image is computed. The difference
pixel count determines the cost.
1 Clusters

of stop words, such as the are not annotated.

2. SSD[8]: translates template and candidate image relative to each other to find the minimum
cost (= matching cost) based on the Sum of
Squared Differences.
3. EDM[8, 13]: Euclidean Distance Mapping.
This technique is similar to XOR, but difference pixels in larger groups are penalized more
heavily, because they are likely to result from
structural differences between the template and
the candidate image, not from noise.
Early versions of the above algorithms were proposed by [12, 13]. Kane et al. [8] improved them by
using extensive normalization techniques that align
images and also conducted more systematic experiments . The above algorithms (including the normalization techniques) are detailed in [8].

2.2

Feature-Oriented Matching

A number of feature based techniques have also been

investigated for matching word:
1. SLH[8, 13]: recovers an affine warping transform (using the Scott and Longuet-Higgins algorithm [22]) between sample points taken from
the edge of the template and candidate image. The residual between template points and
warped candidate points is used as the matching cost.
2. SC [1, 15]: Shape Context matching. This algorithm is currently the best classifier for handwritten digits. Two shapes are matched by establishing correspondences between their outlines. The outlines are sampled and shape context histograms are generated for each sample

point: each histogram describes the distribution

of sample points in the shape with respect to the
sample point at which it is generated. Points
with similar histograms are deemed correspondences and a warping transform between the
two shapes is calculated and performed. The
matching cost is determined from the cost associated with the chosen correspondences. [15]
tested this algorithm for word matching in word
spotting.
3. DTW[15]: A fixed number of features is extracted per image column, resulting in sets of
times series (one per extracted feature) with
the horizontal axis representing time. These
time series can then be jointly aligned and compared with the Dynamic Time Warping algorithm. Examples of the time series features include projection profiles and upper/lower word
profiles of the words.
4. CORR[18]: Using correlation, this technique recovers the correspondences between points of
interest in two images. These correspondences
are then used to construct a similarity measure.
Unlike the other techniques mentioned above,
both the dynamic time warping and the point correspondence techniques share the property that they
do not assume a global transformation between
words2 . These two techniques turn out to be the
best performing techniques with DTW being somewhat better than the other technique. In the following sections we will discuss the DTW and CORR
techniques in some more detail. For more details of
the use of the SLH and SC algorithm in word spotting, see [8] and [15].

2.3

Matching Words with DTW

A person writing a word usually moves the pen from

left to right in English (from right to left in some languages like Arabic and Hebrew). While a word is inherently two dimensional, there is some association
between image columns and the time they were written. By carefully pre-processing word images, one
can minimize variations in the vertical dimension
and then recast word matching as a 1-dimensional
problem along the horizontal axis.
The slant and skew angles at which a person
writes, are usually constant for single words, and
can be normalized using a global transform. On the
other hand, the inter-character and intra-character
spacing is subject to larger variations. DTW [20] offers a way to compensate for these variations, which
2 The warping transform used in the shape context algorithm is rather rigid.

is more flexible than linear scaling: in the matching algorithm that we describe here, image columns
are aligned and compared using DTW. In our framework, each image column is represented by a time series sample point. Figure 2 shows an example alignment of two time series using dynamic time warping.
value

value

samples

samples

samples

Figure 2: Alignment of two similar time series using

dynamic time warping.
A single feature vector consists of one feature
value per column of the image it is calculated for.
For example, if image A = (a(i, j)) is wA pixels wide,
a feature f (A) would be a vector of length wA 3 :
f (A) = (f (a(1, )), f (a(2, )), . . . , f (a(wA , ))). (1)
The dynamic time warping matching algorithm
simultaneously aligns two sets of feature vectors FA
and FB which are extracted from the images A and
B (FB similarly):
FA = (FA (1, ), FA (2, ), . . . , FA (wA , )),

(2)

where every entry FA (x, ) is a d-dimensional vector

consisting of all extracted feature values for image
column x. That is, FA and FB consist of d individually calculated features that will be aligned together
by the dynamic time warping algorithm. The matching error4 for matching images A and B is defined
as
1
merr(A, B) = merr(FA , FB ) = DT W (wA , wB ),
l
(3)
where l is the length of the warping path recovered
by the dynamic time warping algorithm DT W (, )
which uses the recurrence equation

DT W (i 1, j)
DT W (i, j) = min DT W (i, j)
+ d(i, j),(4)

DT W (i, j 1)
d(i, j) =

d
X

(FA (i, k) FB (j, k))2 .

(5)

k=1

To prevent pathological warpings, global path

constraints are used to force the paths to stay close
to the diagonal of the DTW matrix [19]. More details of the dynamic time warping algorithm are presented in [15].
3 The notation implies that every feature value is calculated
strictly from the pixels in the corresponding image column.
This constraint can be relaxed.
4 Matching scores can be obtained from errors by negation.

Another approach, which uses dynamic time

warping to compare features from a template word
image to feature representations of whole lines of
handwritten historical text, was described by Kolcz
et al. in [10]. The main differences to our work
are the application (retrieval by example in Kolczs
case), the matching framework (independent alignment of time series vs. constrained alignment in our
approach) and the limited evaluation (4 query examples vs. thousands in our case).
In the following section we present a number of
features that we used for matching words as images,
using the above dynamic time warping algorithm.

2.3.1

Features for Dynamic Time

Warping

The images we operate on are all grayscale with

256 levels of intensity [0..255]. Before column features can be extracted from an image, a number of
processing steps have to be performed: first, parts
from other words that reach into this words bounding box have to be removed and the background
is cleaned; then inter-word variations such as skew
and slant angle have to be detected and normalized;
next, the bounding box is cropped so that it tightly
encloses the word; finally, the image is padded with
extra rows either on top or on the bottom, to move
the baseline5 to a predefined location. Figure 3
shows an original image and the result of the above
processing steps.

(a) original image,

(b) cleaned and normalized version.

Figure 3: Original image and result after cleaning

and normalization.
All of the column features we describe in the following are normalized to a maximum range of [0..1],
so they are comparable across words. Our goal was
to choose a variety of features presented in the handwriting recognition literature (e.g. [3] or [24]), such
that an approximate reconstruction of a word from
its features would be possible.
We use the following four features to represent
word images (all of the example Figures were obtained from the image in Figure 3(b)):
5 The

baseline is the imaginary line people write on.

Figure 4:
Projection profile feature (rangenormalized and inverted).

Figure 5: Upper and lower word profile features

(normalized and inverted).
1. Projection Profile: these time series result from
recording the sum of the pixel intensity values
in every image column. The result is a profile
that captures the distribution of ink along the
horizontal axis of a word. Figure 4 shows a
typical result.

2/3. Word Profiles: for every image, we can extract an upper and lower profile of the contained word, by going along the top (bottom)
of the enclosing bounding box, and recording
the distance to the nearest ink pixel in the
current image column. Identifying ink pixels
is currently realized by a thresholding technique, which we have found to be sufficient
for our purposes. For more sophisticated foreground/background separation, see [11]. Together, these features capture the shape of the
word outline (see Figure 5).

4. Background/Ink Transitions: for every image

column, we record the number of transitions
from a background- to an ink-pixel. This feature captures the inner structure of a word (see
Figure 6 for an example).

Figure 6: Background/ink transition-count feature

(normalized).
We also tried a number of other features, including
Gaussian derivatives and projection profiles that are
calculated for parts of words. A discussion of the
results can be found in [16].

2.3.2

Speeding up Dynamic Time

Warping

While the word matching based on dynamic time

warping works very well (see results in section 3),
the computational load is quite high: for two images of width n, the algorithms complexity is O(n2 ).
Here we present preliminary investigations into possibilites for speeding up DTW.
Using a global path constraint like the SakoeChiba band [19] speeds up the computation, because
the DTW matrix only has to be evaluated in a region around the diagonal. The complexity of the
resulting algorithm is still O(n2 ), but with a lower
constant.
Another approach is to use the lower-bounding
paradigm (e.g. see [5]): the idea is to use a lowerbounding function lb(A, B), which always underestimates the real matching distance merr(A, B):
AB :

lb(A, B) merr(A, B).

(6)

Such a lower-bounding function can be used for finding a time series in a collection, that has the lowest
distance merr to a given query series.
The approach is still to sequentially scan the data
base for the best matching series, but lb is used to
compare the query Q to a candidate C: if lb(Q, C)
is greater than the distance merr(Q, M) to the currently best matching series M, it is not necessary to
evaluate merr(Q, C), since
merr(Q, M) < lb(Q, C) merr(Q, C).

(7)

We only need to evaluate merr(Q, C), if lb(Q, C) is

less than merr(Q, M).
Of course, this strategy is only useful if lb has
a lower complexity than merr. Several researchers
have proposed lower-bounding functions for time series comparisons with DTW, with [9] being the tightest. The tightness is an important aspect of lower
bounds, since it determines how often merr has to
be evaluated for time series that do not yield a distance which is lower than the current minimum. The
lower bound in [9] was proposed for univariate time
series. We have extended the approach to multivariate time series [17].

2.4

Word Matching using Point

Correspondences

The image matching approach based on point correspondences identifies image corners in the input
images using the Harris detector. Then, the similarity between these points is determined by correlating their intensity neigborhoods using the sum
of squared differences measure. Then the recovered
correspondences are used to calculate a measure of
similarity between the input word images.

Recovering correspondences for all pixels in one

word image would be an expensive operation, considering the search space, which is of quadratic size
in the number of sample points in the two input
images. Using the Harris detector allows us to select a limited number of points of interest, which
are repeatable under a range of transformations and
invariant to illumination variations [6].
The Harris detector operates on the matrix
!
I
I
I 2
)
( x
)( y
)
( x
,
M=
I
I
I 2
( x
)( y
)
( y
)
where I is the gray level intensity image. A corner
is determined when the 2 eigenvalues of M are large
since this indicates grayscale variations in both the
x and y direction.

2.4.1

Recovering Corner
Correspondences

Here we determine pairs of corresponding corner

points in the two input images. Most correspondence methods compare the characteristics of the
local regions around feature points, and then select
the most similar pairs as correspondences. The characteristics of local regions can be represented by either a feature vector (e.g. see [21]), or by windows
of gray-level intensities.
We use the sum of squared differences (SSD) error
measure to compare gray-level intensity windows,
which are centered around detected corner locations.
The reasons for selecting the SSD measure are its
simplicity and the small number of operations required to calculate it - an important consideration
when comparing a large number of image pairs.
In this simple approach, false point matches can
be caused by a number of factors. We try to alleviate
them with constraints:
the size of the query word may be different
from that of the candidate word image, that is,
they have different resolutions. Assuming tight
bounding boxes for all words, we resize all candidate images to the size of the query image.
For a given feature point, there might be several feature points in a candidate image, which
result in small SSD errors. to reduce this possibility, we constrain corresponding feature points
in the candidate image to lie in the neighborhood of the corner point in the template image.
This constraint also has the desirable effect of
speeding up the algorithm.
In order to further decrease the computational
load, we reduce each image to half-size. In essence,
this can be regarded as doubling the size of the SSD

correlation windows without slowing down the implementation. Using larger SSD windows can help
prevent false matches, because of the added context
that is taken into account when comparing image
regions.

quality (see Figure 8(a)). The second set is very

degraded (see Figure 8(b)) - even humans have difficulties reading these pages. We prepared four test
sets for our evaluation:
A: 15 images in test set 1.
B: 2372 images of good quality (test set 1).
C: 32 images in test set 2.
D: 3262 images of bad quality (test set 2).

Figure 7: Recovered correspondences in two word

images.
Experiments showed that adding the above constraints greatly improved the matching accuracy (see
Figure 7 for an example of recovered correspondences).

2.4.2

Distance Measure Calculation

The correspondence between pairs of feature

points captures the similarity between local regions
of two images. In order to judge the similarity of two
word images, this local information is now combined
into a global measure of similarity.
After investigating various approaches for distance measurements, we used the following distance
measurement:
P p
(xbi xai )2 + (ybi yai )2
i
D(A, B) =
#correspondences
#feature points in A

,
(8)
#correspondences
where A is the query image, and B a candidate image; (xai , yai ) and (xbi , ybi ) are the coordinates of a
pair of corresponding feature points, in A and B respectively. Essentially, we are calculating the mean
Euclidean distance of corresponding feature points.
Additionally, we penalize for every point in image
A, that does not have a correspondence in image B6
by multipliying the average distance with a weight.
Thus, the fewer corresponding feature points are
found in the candidate image B, the larger the distance between B and A.
More details on the point correspondence technique can be found in [18].

Results

We conducted experiments on two labeled data sets,

both 10 pages in size. Data set 1 is of acceptable
6 This can happen if the search area in B for a corner point
in A does not contain any points of interest.

Test sets A and C are mainly for the purpose of

comparison to previously reported techniques (in [8])
and for quick performance tests of new techniques.

3.1

Experimental Methodology

For a given test set/matching algorithm pair, the

following evaluation was performed: each image in
the test set was regarded as a query, which was
used to rank the rest of the images in the collection according to their similarity to the query.
Some query/candidate pairs are not compared by
the matching function (pruned), because they are
dissimilar according to a set of simple heuristics (image length, aspect ratio of bounding box, etc.).
The ranked lists of retrieved word images were
evaluated using the mean average precision measure [25], which is commonly used in the information retrieval field. For the purpose of evaluation,
a candidate image was considered relevant to the
query image, if the image labels (ASCII annotation)
matched.
Previous work by [8] used a somewhat misleading evaluation, that considered each query image as
a candidate. With this evaluation, the results are
biased, since most of the techniques always retrieve
the query image at rank 1, if it is part of the candidate set. In table 1, we have provided both the old
evaluation and a new version, which removes query
images from the candidate set. The new evaluation
values for some of the discussed matching techniques
appear in the 4 right-most columns.

3.2

Result Discussion

As can be seen from table 1, DTW and CORR

work best, with similar performance on all data sets.
While CORR performs better on the smaller sets A
and C, DTW seems to have a slightly better overall
performance (data sets B and D). The EDM matching technique seems to perform well on data set A,
but its performance is significantly lower on set C.
Additionally, it is unclear what its overall performance is, since we had no raw results available for
sets B and D. The rest of the matching approaches
(XOR, SSD, SLH and SC) does not perform nearly

(a) example from test set 1 (good quality),

(b) example from test set 2 (bad quality).

Figure 8: Examples from the two test sets used in the evaluation.
Run
A
B
C
D

XOR
SSD
SLH
SC
EDM
DTW
54.14% 52.66% 42.43% 48.67% 72.61% 73.71%
n/a
n/a
n/a
n/a
n/a
65.34%
n/a
n/a
n/a
n/a
15.05% 58.81 %
n/a
n/a
n/a
n/a
n/a
51.81%

CORR
73.95%
62.57%
59.96%
51.08%

SC
EDM
DTW CORR
40.58% 67.67% 67.92% 69.69%
n/a
n/a
40.98% 36.23%
n/a
n/a
13.04% 14.84%
n/a
n/a
16.50% 15.49%

Table 1: Average precision scores for all test runs (XOR: matching using difference images, SSD: sum of
squared differences technique, SLH: technique by Scott & Longuet-Higgins [22], SC: shape context matching
[1], EDM: euclidean distance mapping, DTW: dynamic time warping matching, CORR: recovered correspondences). Four right-most columns show corrected evaluation results.
as well as DTW and CORR, with mean average precision values in the 40-50% range.
The general performance difference of DTW and
CORR on data sets A and B can be explained by
the pruning heuristics, which work much better on
set A than on set B: in set A, only 10% of the valid
matches are discarded in the pruning, while in set
B, almost 30% are discarded. A similar observation
can be made for the test sets C and D: on both
sets, the pruning discards around 45% of the valid
matches. This can be seen in the smaller differences
in mean average precision for both DTW and CORR
on these data sets. The reason for the high rejection
rate of valid matches by the pruning lies in the word
segmentation, which is heavily affected by the bad
quality of the document images.

Conclusions and Outlook

Given the challenges in recognizing words from the

large vocabularies of handwritten manuscript collections, word spotting involving word matching is a
reasonable approach for solving the problem of in-

dexing such manuscript collections. We have discussed a number of different approaches to matching
with the best performing ones being dynamic time
warping and a point correspondence based technique. Challenges remain, including the creation of
word clusters and the necessity of speeding up these
algorithms sufficiently, so that large collections can
be handled in a reasonable amount of time.

Building a system involves creating a user interface. While it is straightforward to imagine a visual
index with pictures and links to pages, it is not clear
whether users would be able to use such an index effectively. An ASCII user interface can be created
by annotating the matched word clusters manually
- permitting a more traditional index. Recent advances in automatic picture annotation [2, 4, 7], using machine learning and information retrieval techniques, may permit a completely different approach
to this problem through automatic annotation of
word image clusters.

Acknowledgments
We would like to thank Jamie Rothfeder and Shaolei
Feng for contributing to the work on point correspondences for word spotting. We also thank the
Library of Congress for providing the images of the
George Washington collection.
This work was supported in part by the Center
for Intelligent Information Retrieval and in part by
the National Science Foundation under grant number IIS-9909073. Any opinions, findings and conclusions or recommendations expressed in this material
are the author(s) and do not necessarily reflect those
of the sponsor.

References
[1] S. Belongie, J. Malik and J. Puzicha: Shape
Matching and Object Recognition Using Shape
Contexts. IEEE Trans. on Pattern Analysis and
Machine Intelligence 24:24 (2002) 509-522.
[2] D. M. Blei and M. I. Jordan: Modeling Annotated Data. Technical Report UCB//CSD-021202, 2002.
[3] C.-H. Chen: Lexicon-Driven Word Recognition.
In: Proc. of the Third Intl Conf. on Document Analysis and Recognition 1995, Montreal,
Canada, August 14-16, 1995, pp. 919-922.
[4] P. Duygulu, K. Barnard, N. de Freitas and
D. Forsyth: Object Recognition as Machine
Translation: Learning a Lexicon for a Fixed Image Vocabulary. In: Proc. 7th European Conference on Computer Vision, Copenhagen, Denmark, May 27-June 2, 2002, vol. 4, pp. 97-112.
[5] C. Faloutsos: Multimedia IR: Indexing and
Searching. In: Modern Information Retrieval,
R. Baeza-Yates and B. Ribeiro-Neto; AddisonWesley, Reading, MA, 1999.
[6] C. Harris and M. Stephens: A Combined Corner
and Edge Detector. In: Proc. of the 4th Alvey
Vision Conf., 1988, pp. 147-151.
[7] J. Jeon, V. Lavrenko and R. Manmatha: Automatic Image Annotation and Retrieval Using
Cross-Media Relevance Models. CIIR Technical
Report MM-41, 2003.
[8] S. Kane, A. Lehman and E. Partridge: Indexing
George Washingtons Handwritten Manuscripts.
Technical Report MM-34, Center for Intelligent Information Retrieval, University of Massachusetts Amherst, 2001.
[9] E. Keogh: Exact Indexing of Dynamic Time
Warping. In: Proc. of the 28th Very Large

Databases Conf. (VLDB), Hong Kong, China,

August 20-23, 2002, pp. 406-417.
[10] A. Kolcz, J. Alspector, M. Augusteijn, R. Carlson and G. V. Popescu: A Line-Oriented Approach to Word Spotting in Handwritten Documents. Pattern Analysis & Applications 3 (2000)
153-168.
[11] G. Leedham, S. Varma, A. Patankar and
V. Govindaraju: Separating Text and Background in Degraded Documents Images - A
Comparison of Global Thresholding Techniques
for Multi-Stage Thresholding. In: Proc. of the
8th Intl Workshop on Frontiers in Handwriting Recognition 2002, Niagara-on-the-Lake, ON,
August 6-8, 2002, pp. 244-249.
[12] R. Manmatha, C. Han, E. M. Riseman and
W. B. Croft: Indexing Handwriting Using Word
Matching. In: Digital Libraries 96: 1st ACM
Intl Conf. on Digital Libraries, Bethesda, MD,
March 20-23, 1996, pp. 151-159.
[13] R. Manmatha and W. B. Croft: Word Spotting: Indexing Handwritten Archives. In: Intelligent Multi-media Information Retrieval Collection, M. Maybury (ed.), AAAI/MIT Press 1997.
[14] R. Manmatha and N. Srimal: Scale Space
Technique for Word Segmentation in Handwritten Manuscripts. In: Proc. 2nd Intl Conf. on
Scale-Space Theories in Computer Vision, Corfu,
Greece, September 26-27, 1999, pp. 22-33.
[15] T. M. Rath and R. Manmatha: Word Image
Matching Using Dynamic Time Warping. to appear in Proc. of the Computer Vision and Pattern Recognition Conf. 2003.
[16] T. M. Rath and R. Manmatha: Features for
Word Spotting in Historical Manuscripts. to appear in Proc. of the 7th Intl Conf. on Document
Analysis and Recognition 2003.
[17] T. M. Rath and R. Manmatha: LowerBounding of Dynamic Time Warping Distances
for Multivariate Time Series. CIIR Technical
Report MM-40, 2003.
[18] J. L. Rothfeder, S. Feng and T. M. Rath: Using
Corner Feature Correspondences to Rank Word
Images by Similarity. CIIR Technical Report
MM-44, 2003.
[19] H. Sakoe and S. Chiba: Dynamic Programming Optimization for Spoken Word Recognition.
IEEE Trans. on Acoustics, Speech and Signal
Processing 26 (1980) 623-625.

[20] D. Sankoff and J. B. Kruskal: Time Warps,

String Edits, and Macromolecules: The Theory
and Practice of Sequence Comparison. AddisonWesley, Reading, MA, 1983.
[21] C. Schmid and R. Mohr: Local Grayvalue Invariants for Image Retrieval. IEEE Trans. on
Pattern Analysis and Machine Intelligence 19:5
(1997) 530-535.
[22] G. L. Scott and H. C. Longuet-Higgins: An Algorithm for Associating the Features of Two Patterns. Proc. of the Royal Society of London B224
(1991) 21-26.
[23] C. I. Tomai, B. Zhang and V. Govindaraju:
Transcript Mapping for Historic Handwritten
Document Images. In: Proc. of the 8th Intl
Workshop on Frontiers in Handwriting Recognition 2002, Niagara-on-the-Lake, ON, August 6-8,
2002, pp. 413-418.
[24] . D. Trier, A. K. Jain and T. Taxt: Feature
Extraction Methods for Character Recognition A Survey. Pattern Recognition 29:4 (1996) 641662.
[25] C. J. van Rijsbergen: Information Retrieval.
Butterworth, London, England, 1979.