Matsakis MEng 99

Recognition of Handwritten Mathematical Expressions
by
Nicholas E. Matsakis
Submitted to the Department of Electrical Engineering and
Computer Science
in partial fulfillment of the requirements for the degrees of
Bachelor of Science in Electrical Engineering and Computer Science
and
Master of Engineering in Electrical Engineering and Computer
Science
at the
MASSACHUSETTS INSTITUTE OF TECHNOLOGY
May 1999

c Nicholas E. Matsakis, MCMXCIX. All rights reserved.
The author hereby grants to MIT permission to reproduce and
distribute publicly paper and electronic copies of this thesis
document in whole or in part.
Author . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Department of Electrical Engineering and Computer Science
May 21, 1999
Certified by . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Paul A. Viola
Associate Professor
Thesis Supervisor
Accepted by . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Arthur C. Smith
Chairman, Department Committee on Graduate Students
Recognition of Handwritten Mathematical Expressions
by
Nicholas E. Matsakis
Submitted to the Department of Electrical Engineering and Computer Science

on May 21, 1999, in partial fulfillment of the
requirements for the degrees of
Bachelor of Science in Electrical Engineering and Computer Science
and
Master of Engineering in Electrical Engineering and Computer Science
Abstract
In recent years, the recognition of handwritten mathematical expressions has re-
cieved an increasing amount of attention in pattern recognition research. The di-
versity of approaches to the problem and the lack of a commercially viable system,
however, indicate that there is still much research to be done in this area. In this
thesis, I will describe an on-line approach for converting a handwritten mathemat-
ical expression into an equivalent expression in a typesetting command language
such as TEX or MathML, as well as a feedback-oriented user interface which can
make errors more tolerable to the end user since they can be quickly corrected.
The three primary components of this system are a method for classifying iso-
lated handwritten symbols, an algorithm for partitioning an expression into sym-
bols, and an algorithm for converting a two-dimensional arrangements of symbols
into a typeset expression. For symbol classification, a Gaussian classifier is used to
rank order the interpretations of a set of strokes as a single symbol. To partition
an expression, the values generated by the symbol classifier are used to perform a
constrained search of possible partitions for the one with the minimum summed
cost. Finally, the expression is parsed using a simple geometric grammar.
Thesis Supervisor: Paul A. Viola

Title: Associate Professor
2
Acknowledgments
I would like to thank my advisor Paul Viola for providing the motivation and the
resources for this research to be done. In particular, I would like to thank him for
his continual suggestions for improving system performance and for suggesting
the minimum spanning tree constraint. I would also like to thank Erik Miller for
his willingness to help me work through a stubborn problem, for his insight into
the nature of mathematical expression recognition, and for quickly providing me
with counter examples for every deep truth I ever thought I had divined.
In addition, I would like to thank John Winn for his help working through the
stroke reversing heuristics, Victor Luchangco and Kevin Simmons for their helpful
suggestions for improvements to the manuscript of this thesis, and Jeff Norris for
being a supportive officemate.
I would like also to thank my parents, Elias and Joanne Matsakis, siblings
Chris and Antigone Matsakis, grandparents Nicholas and Theodora Matsakis and
Christopher and Antigone Dafnides, Godparents Michael and Penny Pilafas, and
the rest of my family for their unwavering support both during this work and in all
my endeavors. Finally, I would like to express my deepest gratitude to Terri Iuz-
zolino for all the encouragement and support she has provided throughout this
project, and in particular during the final stages of writing.
3
Contents
1 Introduction 9
1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.2 System Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.2.1 Isolated Symbol Classification . . . . . . . . . . . . . . . . . . 11
1.2.2 Expression Partitioning . . . . . . . . . . . . . . . . . . . . . . 11
1.2.3 Parsing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.2.4 User Interface . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2 Problem Overview 15
2.1 Terminology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.2 Qualities of the Problem . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.3 Previous Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
3 Isolated Symbol Classification 20

3.1 Problem Description and Assumptions . . . . . . . . . . . . . . . . . 20
3.2 Data Collection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.3 Preprocessing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.3.1 Stroke Direction and Order Invariance . . . . . . . . . . . . . . 22
3.3.2 Scaling, Shifting, and Resampling . . . . . . . . . . . . . . . . 25
3.3.3 Derivative Terms . . . . . . . . . . . . . . . . . . . . . . . . . . 26
3.3.4 Principal Component Analysis . . . . . . . . . . . . . . . . . . 26
3.4 Gaussian Density Estimation . . . . . . . . . . . . . . . . . . . . . . . 27
3.5 Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
4
3.5.1 Confusion Matrix . . . . . . . . . . . . . . . . . . . . . . . . . . 28
3.5.2 Generalization Ability . . . . . . . . . . . . . . . . . . . . . . . 30
4 Partitioning an Expression 32
4.1.1 The Existence of a Partition . . . . . . . . . . . . . . . . . . . . 32
4.1.2 Optimal Substructure . . . . . . . . . . . . . . . . . . . . . . . 33
4.1.3 Additive Cost Function . . . . . . . . . . . . . . . . . . . . . . 34
4.2 Orders of Growth . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
4.3 Constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
4.3.1 Time Ordering . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
4.3.2 The Minimum Spanning Tree Constraint . . . . . . . . . . . . 37
5 Parsing 41
5.2 The Structure of Mathematical Expressions . . . . . . . . . . . . . . . 42
5.2.1 Basic Mathematical Typesetting . . . . . . . . . . . . . . . . . . 42
5.3 Box Relationships and the Geometric Grammar . . . . . . . . . . . . . 43
5.4 The Parsing Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
5.5 Superscripts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
6 Demonstration Program Engineering 48

6.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
6.2 User Interface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
6.2.1 Pen Interface . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
6.2.2 Error Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
6.2.3 Error Correction . . . . . . . . . . . . . . . . . . . . . . . . . . 51
6.2.4 Preferences . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
7 Conclusions 54
A Glossary 56
5
References 58
6
List of Figures
1-1 Screen image of the recognition system . . . . . . . . . . . . . . . . . 11

1-2 The effect of partitioning on interpretation . . . . . . . . . . . . . . . . 12
1-3 An expression correctly partitioned into symbols . . . . . . . . . . . . 13
2-1 Two characters represented by the same symbol . . . . . . . . . . . . 16

2-2 Two symbols representing the same character . . . . . . . . . . . . . . 16
3-1 The Data Collection Program . . . . . . . . . . . . . . . . . . . . . . . 21

3-2 The Angles Used to Find a Canonical Stroke Ordering . . . . . . . . . 24
3-3 Eigenvalues Associated with the Principal Components . . . . . . . . 27
3-4 Examples of Reconstructed Characters . . . . . . . . . . . . . . . . . . 28
3-5 Gaussian Classifier Confusion Matrix . . . . . . . . . . . . . . . . . . 29
3-6 Symbols Commonly Confused by a Gaussian Classifier . . . . . . . . 29
3-7 Two styles of writing the digit 5 . . . . . . . . . . . . . . . . . . . . 30
3-8 Classifier Generalization Performance . . . . . . . . . . . . . . . . . . 31
4-1 An Expression with Script Characters . . . . . . . . . . . . . . . . . . 33

4-2 An Expression Correctly Partitioned into Symbols . . . . . . . . . . . 33
4-3 A typical minimum spanning tree . . . . . . . . . . . . . . . . . . . . 38
4-4 Using a Minimum Spanning Tree to Partition . . . . . . . . . . . . . . 39
4-5 The chain formation of a minimum spanning tree . . . . . . . . . . . 40
4-6 The star formation of a minimum spanning tree . . . . . . . . . . . . 40
5-1 Some bounding box relationships . . . . . . . . . . . . . . . . . . . . . 44

5-2 A Set of Partitioned Symbols . . . . . . . . . . . . . . . . . . . . . . . 45
7
5-3 Characters which share the same baseline . . . . . . . . . . . . . . . . 47
6-1 Screen Image of the Client Application . . . . . . . . . . . . . . . . . . 49

6-2 An Example of Error Detection . . . . . . . . . . . . . . . . . . . . . . 52
8
Chapter 1
Introduction
1.1 Motivation
The problem of machine recognition of handwritten expressions has long been a

focus of study in the field of pattern recognition. Research in this area has been
driven by a desire to combine the natural advantages of handwritten input, in-
cluding a simple interface and a well-established stylistic vocabulary, with the data
processing capabilities of computers. Recently, the problem has been approached
with increased vigor with the advent of palmtop computers with pen interfaces,
which possess enough processing resources to handle the demands of machine
recognition. As a result, a number of commercially successful products are avail-
able which recognize a users natural handwriting and use this ability to perform
simple tasks such as scheduling appointments and writing memos.
Most scientists and engineers, however, are unable to take advantage of these
computers for their technical work due to the lack of effective algorithms for in-
terpreting more complex handwritten expressions, particularly diagrams, graphs,
equations, and other mathematical forms. While computers can store these forms
as digital ink, the inability to work with the expressions in a meaningful way
after they have been entered has prevented these systems from replacing pencil
and paper. Compared to the effort put into the recognition of printed and cursive
prose the recognition of more complex forms has received only minor attention in
9
pattern recognition research. In addition, the diversity of approaches to the prob-
lem and the lack of a commercially viable system indicate that there is still much
research to be done in this area.
As more powerful computers with better displays and input devices become
available, demand will increase substantially for software systems which can work
with the type of handwritten data that one would find in a research notebook or
technical document. Mathematical expressions are a natural place to begin such
research as they are critical to virtually all technical writing and there already ex-
ists a wide body of literature on recognizing handwritten letters and words, major
subcomponents of these expressions. Combining mathematical expression recog-
nition capabilities with existing algebra solving software, graphing programs, and
simulation systems would be a first step towards a superior user interface for do-
ing technical work with a computer.
1.2 System Overview
In this thesis, I describe an on-line approach for converting a handwritten math-

ematical expression into an equivalent expression in a typesetting command lan-
guage such as TEX or MathML. In addition, I describe a feedback-oriented user
interface which renders errors in a recognition system more tolerable to the end
user since they can be quickly detected and corrected. Figure 1-1 provides a screen
image of the system.
The problem of interpreting an expression can be divided into three modular
subproblems called isolated symbol classification, expression partitioning and parsing.
This division has the advantage that each subproblem can be solved and its per-
formance evaluated essentially independent of the others, so that improvements
can be made in each area while still maintaining the integrity of the entire system.
In processing an expression, it is first partitioned into symbols in a process
called expression partitioning. The symbol classifier is used in this process to eval-
uate the likelihood that particular strokes should be combined into symbols. The
10
Figure 1-1: Screen image of the recognition system
resulting set of symbols is then parsed by assigning characters to the symbols and
determining a structure for the expression, resulting in an interpretation of the ex-
pression as a typesetting command.
1.2.1 Isolated Symbol Classification
One of the basic problems in handwriting recognition is determining, out of con-

text, which symbol is best represented by a set of strokes. For this task I created
Gaussian models of a set of common symbols from examples of my handwriting.
At run time these models were used to rank possible interpretations of a new set
of strokes.
1.2.2 Expression Partitioning
Handwritten expressions typically contain more than a single symbol, so the capa-
bilities of the symbol classifier need to be used within a larger framework to deter-
mine the quantity, location, and identity of the symbols in an expression. There-
fore, the expression partitioning algorithm attempts to find an optimal partitioning
11
of an expressions strokes into a set of symbols.
The correct partitioning of an expression is not always obvious, even to a hu-
man reader. Consider the example given in figure 1-2. This diagram shows how
the strokes in an ambigious expression may be partitioned into two different sets
of symbols, which are indicated by the placement of a grey bounding box around
the symbols. Depending on the partitioning of strokes, this particular expression
could be interpreted as either 1 < x or kx. A good partitioning algorithm needs to
be able to consider both of these possibilities and determine which one is prefer-
able according to some cost function.
Ambiguous Expression
Possible Partitions
Multiple Interpretations
1< x kx
Figure 1-2: The effect of partitioning on interpretation
An additive cost model is used to partition an expression, where the cost for
considering that a particular set of strokes belongs to the same symbol comes from
the values assigned those strokes by the symbol classifier. The cost of a partition
of a set of strokes, then, is the sum of the costs of the symbols in the partition.
12
Figure 1-3: An expression correctly partitioned into symbols
In addition, a minimum spanning tree based in interstroke distances is used to

constrain the set of partitions searched to a reasonable size.
1.2.3 Parsing
Once the expression has been correctly partitioned into symbols, as in figure 1-3,
there still remains the problem of determining which characters the symbols rep-
resent and deciding on the structure of the resulting typeset expression. This can
be done using a geometric grammar whose elements are inspired by the atomic
elements of TEXs typesetting engine. In this grammar, each character belongs to
a single grammar type and combines with other characters in well defined ways
based on simple relationships between their bounding boxes. In addition, sim-
ple characters also have a baseline which aids in determining when a character is
superscripted.
1.2.4 User Interface
An accurate recognition algorithm still needs a good user interface if it is to be a

viable alternative to pen and paper. An interface needs to be simple, since it may
be used on a computer with only a stylus and touch screen for input. In addition, it
should allow the user to quickly correct errors, because some expressions are am-
biguous and people often make mistakes while writing long expressions. Finally,
it is important for the interface to be able to provide immediate feedback to the
user so that errors can be quickly detected.
13
The programs user interface provides a number of features which allow the
user to immediately detect and correct errors. One way of detecting potential
errors is to set a recognition threshold for symbols, and change the color of any
strokes that can not be assigned to a symbol with a cost below the threshold as
they are written. Another way of detecting errors is to draw faint boxes around
the partitioned symbols, so that the user immediately knows that a two or three
stroke symbol has been recognized.
The system also provides two ways of correcting errors. The simplest is the
ability to erase a set of strokes simply by scribbling over them with the stylus.
A small part of an expression can then be changed without effecting the rest of
the expression. Another feature for error correction is a pop-up correction menu,
which allow the user to change the character assigned to a symbol by specifying
an alternative from a list of options. Using these features, expressions with minor
interpretation errors can be quickly corrected.
14
Chapter 2
Problem Overview
2.1 Terminology
Research in handwriting recognition has produced a rich glossary of terms pertain-

ing to every aspect of handwriting, with different approaches often necessitating
the use of different terminology. Generally, recognition systems can be classified
as either on-line or off-line systems. In on-line systems [4], the writing is captured
by the computer as the user writes on a pen tablet or other similar input device.
Pen data is typically collected as a series of strokes, data structures which store the
trajectory of a pen between the time it touches down on the writing surface and
the time it lifts off, expressed as a series of location and time coordinates (x, y, t).
A handwritten expression, then, consists of a time-ordered set of strokes in the same
coordinate frame. In contrast, the input data to an off-line system is an expression
that has already been written on paper, typically acquired as an image using an
optical scanner. This data contains no timing data so that recognition systems of
this type work by analyzing the pixels of the image.
It is common in handwriting recognition literature to refer to characters as the
items being recognized. However in this thesis the word character is used in a
slightly more precise manner than usual to denote the characters in a typesetting
language. In this sense, a character is not only specified by its appearance but
also by a well defined set of allowable geometric relations with other characters.
15
For example, consider the capital Greek sigma character and the summation sign
character. Though they may be drawn in precisely the same manner with a pen,
as in figure 2-1, they are actually two distinct characters. As a summation sign
the symbol is used as a group operator which may have other symbols appearing
above and below it as bounds on the summation. As the letter sigma such bounds
are not permitted while exponential relations are. Characters in this thesis will be
written in quotation marks, as in the character .
Figure 2-1: Two characters represented by the same symbol
Since distinct characters may be drawn in the same way, the term symbol will
be used to describe the items which are actually drawn. It is important to note
that not only may a single symbol represent distinct characters, but distinct sym-
bols may represent the same character, as in figure 2-2. Thus, the mapping from
symbols to characters is generally many to many, though often a single character is
represented by a single symbol. Symbols in this thesis will be written in brackets,
as in the symbol [].
Figure 2-2: Two symbols representing the same character
2.2 Qualities of the Problem
Any approach to machine recognition of handwriting must address certain prop-

erties inherent to the problem. Perhaps the most daunting is the size of the space
16
of possible interpretations; for a given set of strokes, there are a super-exponential
number of possible expressions that they might represent 1 . Therefore, to correctly
interpret a set of strokes a system needs not only to be able to rank the likelihood of
particular interpretations, but it also needs to effectively prune its search space to
a reasonable size, perhaps before even computing the likelihood of any particular
interpretation fully.
Fortunately, the search space is not without structure. Handwritten expressions
are, in some sense, decomposable. That is, small sets of strokes can be interpreted
in isolation and then combined with other stroke sets to form larger expressions.
It is this property which makes a dynamic programming approach feasible, which
then allows for the dynamic reinterpretation of expressions as the user continues
to write.
The method suggested in this thesis takes advantage of this decomposability
in two ways. First, symbol recognition is done in isolation; the likelihood that a
particular set of strokes form a symbol is estimated only from the appearance of
those strokes and not from that of the surrounding strokes. Though it is clear that
the surrounding strokes might provide information about what symbol a partic-
ular set of strokes represents, recognizing the symbols in isolation considerably
reduces the complexity of the symbol models. Furthermore, it allows the results
of symbol recognition to be reused as more strokes are added to the expression,
speeding up recognition considerably.
Decomposablility is also utilized in expression partitioning, where the cost of
a partition of a subset of the strokes does not depend on the surrounding strokes.
While this may not always be a valid assumption, it is sufficient to find a good
partition in most cases and allows the use of dynamic programming to produce a
solution with relatively little computation.
It is important to note that expressions are not completely decomposable. The
1
Given a set of symbols on a page, the number of possible assignments of characters is expo-
nential in the number of symbols. However, for mathematical expressions there is the additional
4
factor of structure, so that 23 , 234 , and 234 are distinct from one another, adding to the size of the
search space.
17
best interpretation of the whole expression cannot simply be constructed from the
best interpretation of its subexpressions, and it is certainly possible that an inter-
pretation of a particular set of strokes may change dramatically as other strokes are
added. For this reason, when dynamically reinterpreting an expression the parti-
tioning and parsing steps are redone every time new strokes are added or deleted
to allow the user the flexibility to change the meaning of any particular stroke.
Still, the results of previous computation are reused when possible.
In addition to these more general qualities of handwriting recognition, the
problem of recognizing mathematical expressions has a number of properties that
distinguish it from the problem of recognizing printed text. The most obvious of
these is its two dimensional structure; whereas prose is written linearly, mathe-
matical symbols can be written above, below, and inside one another, and these
spatial relationships are crucial to the interpretation of the expression. This means
that even if all the characters are correctly recognized, there is still the possibility
that the structure may be interpreted incorrectly.
Another aspect of the problem is that the symbols in handwritten mathemat-
ical expressions are almost always printed rather than written in a cursive script.
That is, each symbol is written separately from every other so that each stroke in
an expression belongs to one and only one symbol. This is advantageous from
the standpoint of machine recognition, since segmentation of strokes into symbols
is much easier when each stroke only belongs to a single symbol. Therefore, the
partitioning algorithm suggested in this thesis assumes this to be the case. Thus, it
requires that words that appear in mathematical expressions, such as the abbrevi-
ations for the trigonometric functions (e.g. sin, tan), also be printed.
2.3 Previous Work
Many previous attempts at equation recognition [9, 7, 14] aimed at only interpret-
ing expressions that were valid mathematical expressions according to some gram-
mar. These grammars enforced a number of reasonable constraints, requiring that
18
parentheses match, binary operators have valid operands, and integrals have dif-
ferentials, among other things. Initially this seems like a strong set of constraints
on the space of possible interpretations, since of the set of expressions that can be
4=s
typeset, such as +r
, only a very small number actually mean anything, and of
those an even smaller fraction are likely to ever occur in a handwritten expression.
The problem with this type of approach is that it greatly reduces the decom-
posability of the problem. In a valid mathematical expression, there can be a large
number of long distance dependencies at work. For example, in an integral ex-
pression, the integral and its differential may be separated by great distances, so
that the likelihood of a symbol on the left hand side of an expression being an inte-
gral sign depends on whether there are symbols on the right hand side which can
represent good differentials. These sorts of dependencies are extremely difficult to
model accurately and evaluate quickly. Since the proposed system does not model
these types of dependencies, it has the potential to produce syntactically invalid
interpretations. One solution to this problem is to find a preliminary interpreta-
tion first, and then to correct the expression syntactically using a narrow range of
options. This approach has been tried before [16], and syntax-checking would be
a straightforward extension to the system proposed in this thesis.
19
Chapter 3
Isolated Symbol Classification
3.1 Problem Description and Assumptions
Isolated symbol classification is the problem of providing a ranked list of inter-

pretations of a set of strokes as a single symbol. It is the problem with the most
previous work among those discussed in this thesis, since it lies at the core of al-
most all handwriting recognition applications. See [4] for a survey of common
techniques.
To solve this problem I used a statistical approach, creating a single Gaussian
model for each symbol class based on examples of my own handwriting. The ex-
ample symbols were preprocessed and then used to estimate the model parameters
in a training phase. At run time potential symbols were similarly processed and
compared to the models for classification.
Since the models are estimated from a single users handwriting, performance
with other users writing is sporadic, performing well only when there is a great
deal of similarity between that users handwriting and the training examples. Ad-
justing this technique to the problems of a multiuser system would introduce many
issues which are peripheral to the problem of mathematical expression recognition
and so it was decided that initially the system would be tuned to a particular user.
In principle, however, only the symbol classifier would need to be modified to
make the system work with multiple users.
20
3.2 Data Collection
The data for a symbol model were collected in a set of 500 or more examples,
called a training set, in a single sitting using a Wacom digitizing tablet and a custom
collection program written in Java. An image of the collection program is shown in
figure 3-1. Precise timing data was disregarded, so that the order in which the pen
touched various points was maintained, but the actual time difference between
samples was lost. Therefore, the resulting data for a particular stroke consisted of
an ordered set of (x, y) pairs, with the first pair corresponding to the touchdown
point of the pen, and successive pairs added as the pen moved from one sampling
point to the next, ending with the lift-off point. Symbols were treated as time
ordered sets of strokes, and a training set consisted of a set of symbols each drawn
using the same number of strokes in the same direction and order.
Figure 3-1: The Data Collection Program
3.3 Preprocessing
The purpose of preprocessing is to convert raw stroke data into a uniform format
so that training data can be used to create symbol models. In addition, it is often
21
desireable to simplify the models by removing information from the stroke data
during preprocessing which will not be useful for classification. One example of
extraneous information is the scale of the symbols, since the identity of a symbol
is largely independent of its scale.
Preprocessing must necessarily be done without regard to the symbol class that
is being modelled, so that both training examples and new symbols are prepro-
cessed in the exact same manner. The preprocessing used in this method converts
a set of strokes in the form of ordered (x, y) coordinates into a single vector in a 74
dimensional space. In this way, any set of strokes can be compared against every
model to determine which symbol they best represent.
3.3.1 Stroke Direction and Order Invariance
One of the pitfalls of on-line approaches to handwriting recognition is that there

is a much greater variation in the manner in which a writer might have drawn a
particular symbol than with its actual appearance on the page. For example, one
may choose to pick up the pen at an arbitrary point and then continue the previous
motion with a new stroke, or one might retrace a portion of an already written
stroke. Both actions are nearly impossible to detect by looking at the resulting
image, but have the potential to affect on-line recognition performance adversely.
Some of the most common variations occur in the direction that simple strokes
are drawn and in the order that strokes in multi-stroke symbols are drawn. For
example, the symbol [+] is typically drawn with two strokes. The horizontal stroke
can be drawn in two directions, with the pen moving either from left to right or
right to left. The vertical stroke can likewise be drawn both from top to bottom or
from bottom to top. Furthermore, either the horizontal stroke or the vertical stroke
may be drawn first. Unfortunately, allowing the user to enter this symbol in any
of these ways without preprocessing would result in a multi-modal distribution
which would be poorly modeled by a Gaussian distribution.
One way of combating this problem is to treat each way of drawing a symbol
22
to be an entirely different symbol and create a separate model for each one. While
this is the most reliable method, it is computationally expensive, since the num-
ber of models for single stroke symbols would double, multiplied by a factor of
eight for two stroke symbols, and every three stroke symbol, such as [], would be
represented by forty-eight separate models.
An alternative is to take each example symbol in the data set and try to assign a
canonical direction for each of its strokes and a canonical ordering of strokes before
building symbol models. Then, redirect and reorder strokes in the same manner
before comparing them to the models. Finding qualities of a stroke that are con-
sistent for the same stroke in a symbol can be difficult. Still, there are some reg-
ularities in handwritten strokes that may be taken advantage of through heuristic
techniques that can make symbol recognition fairly robust to such variations.
Stroke Reversal
To attempt to put a stroke in a canonical direction, the first point (xf , yf ) and the
last point (xl , yl ) of the stroke were used to determine two ratios,
|xf xl | |yf yl |
Rx = , Ry =
D D
where D is the length of the diagonal of the bounding box of the stroke. These
ratios are a weighted measure of the change in the x and y coordinates between the
beginning and the end of these strokes. Then, using a threshold value, [0, 1],
these ratios are used to classify a stroke as being one of four general types: closed,
horizontal, vertical, or diagonal. A stroke is closed if Rx < and Ry < , horizontal
if Rx and Ry < , vertical if Rx < and Ry , and diagonal if Rx and
Ry . Horizontal strokes were then reversed if xf < xl , vertical and diagonal
strokes were reversed if yf < yl , and closed strokes were never reversed. Because
closed strokes are not reversed, this resulting models will still be sensitive to the
initial starting place on a closed stroke, such as that of the symbol [].
23
Figure 3-2: The Angles Used to Find a Canonical Stroke Ordering
Stroke Ordering
After being properly directed, the strokes in a multi-stroke symbol are put into a
canonical ordering. Strokes are first assigned a value according to the magnitude
of the interior angle between the upper edge of the symbols bounding box and the
line segment between the the upper left corner of the bounding box and the last
point in the stroke. They are then sorted in ascending order based on this value.
Figure 3-2 illustrates this angle on the strokes in the symbol [] and indicates the
resulting ordering.
Types of Errors
Used heedlessly, these heuristics are liable to produce just the types of effects that
they are intended to eliminate variation in the directions and order of drawn
strokes. However, their effectiveness can be tested using a collection of data with
the strokes drawn in the same order and direction.
There are two types of errors that can occur when trying to assign a canonical
direction to a stroke. The first is simply to not assign the same direction to a par-
ticular stroke and its reverse. This only happens when a stroke is determined to
24
be closed, since neither a closed stroke nor its reverse will be reversed. Obviously,
this problem can be avoided simply by lowering the threshold value so that no
stroke is determined to be closed. Lowering this value, however, causes errors of
a more significant type.
The second type of error is to reverse some of the strokes in a set of example
symbols, but not their corresponding strokes in other examples. This is the more
severe of the two types of errors because it can take a set of consistently directed
examples and reverse some of them, resulting in precisely the type of multi-modal
distribution this heuristic seeks to avoid. Errors of this type can be prevented sim-
ply by raising the threshold value so that all strokes are determined to be closed
and not flipped. If this value is raised too high errors of the first type will result.
The proper setting for the threshold value is the one which most accurately
identifies closed strokes as being closed. I found that = 0.37 performed well at
this task. In a trial of 20,500 strokes belonging to 41 types of symbol, I found errors
of type 1 occurred on 15% of the strokes. Errors of this type on all but 1% of the
strokes belonged to symbols which were either closed, such as [8], [0], [o], [],
or symbols which are very unlikely to be drawn backwards such as [d], [p], [].
With this same threshold value, I found that less than 1% of the strokes had errors
of type two, and most of those were concentrated in the symbol [].
3.3.2 Scaling, Shifting, and Resampling
After the strokes are rearranged, they are further preprocessed to ensure scale and
translation invariance and then resampled to ensure that all symbols can be ana-
lyzed in the same vector space. First, every point in each stroke is shifted so that
the origin is the upper left corner of the symbols bounding box. Then, all of the
points are scaled so that the symbol fits inside a unit square, preserving aspect
ratio. Next, each stroke was resampled at equidistant points along its arc length
using linear interpolation between points. The number of the sample points for a
36
stroke was N
, where N is the number of strokes in a character. Finally, the (x, y)
25
coordinates for each character were assembled into a single 72 dimensional col-
umn vector with the x coordinates occupying rows 1 36, and the y coordinates
occupying rows 37 72. The points of multiple strokes characters were added in
their canonical order.
3.3.3 Derivative Terms
In addition to the 72 stroke dimensions, the symbol vector has two more dimen-
sions containing the summed squared second derivative of the x and y coordinates
with respect to the arc length. These components measure the amount of curva-
ture in a stroke in both the x and y directions. For multi-stroke symbols, these
dimensions contain the sum of these values over all strokes. These features were
added to aid the classifier in detecting the difference between consistent curves
in strokes, important for classification, from variable curves due to writer noise.
These dimensions were scaled to make their variance comparable to that of the
stroke dimensions.
3.3.4 Principal Component Analysis
The final stage of preprocessing a set of strokes is projecting the symbol vector
into the vector space spanned by the first 15 principal components of the entire
training data set. These components are the eigenvectors of maximal variance of
the covariance matrix of the entire pool of training data. This subspace contains
the components of symbol vectors which are the most variable across all symbols.
The technique used to compute the projection matrix is a modification of standard
principal component analysis where each class of symbol contributes equally to
the choice of components, regardless of the number of examples in that class. This
multi-class PCA algorithm is useful in that it allows the projection matrix for a
large data set to be recomputed quickly when a new class is added.
This dimensionality reduction is justified by the rapid fall-off of the eigenvalues
associated with the principal components, plotted in figure 3-3, since presumably
26
1.6
1.4
1.2
Eigenvalues
1
0.8
0.6
0.4
0.2
0
0 20 40 60 80
Number of Eigenvectors
Figure 3-3: Eigenvalues Associated with the Principal Components
components which are essentially constant across all symbols will not be useful
in distinguishing between them. Further justification comes from the reasonable
appearance of reconstructed characters, produced by projecting the data into the
subspace and then back to the full space, shown in figure 3-4.
3.4 Gaussian Density Estimation
In the training phase, the projected data points of each class are used to estimate a
mean () and covariance matrix () for a Gaussian density in the PCA space. At
run time, an unknown symbol vector is compared to the models by computing the
density of the vector under each model producing a real number for each of the
possible symbols.
(x j )T 1
!
1 j (x j )
dj (z) = D 1 exp (3.1)
(2) 2 |j | 2 2
The density of unknown symbol z under model j
These values are then used by both the expression partitioning and expression
parsing algorithms to determine a final interpretation for the stroke set.
27
Original Characters
Reconst. Characters
Figure 3-4: Examples of Reconstructed Characters
3.5 Performance
3.5.1 Confusion Matrix
One standard measure of classifier performance is the rate at which it misinter-

prets one symbol as another. This can be estimated by first training a classifier on a
body of data, and then testing it using a different data set to determine the misclas-
sification rate for each type of symbol. These values can be depicted graphically
in a confusion matrix of the classifier, as shown in figure 3-5. The rows in this ma-
trix represent the actual classes of the test symbols, and the columns indicate the
classes which those symbols were assigned by the classifier. A particular element
(i, j) has a magnitude proportional to the number of times a character of type i
was classified as type j. Thus, the confusion matrix of the perfect classifier has
nonzero entries only along the diagonal. In figure 3-5, larger values are denoted
by darker cells, indicating that the classifier generally performed well, classifying
most classes with over 99% accuracy.
Notice that there are two areas where there is significant confusion, however.
Class one is often confused with class forty-three, and class two is often confused
with class forty. In these experiments, classes one and forty-three were the classes
for symbols [0] and [o] respectively, and classes two and forty were the classes for
28
0
10
20
Actual Class
30
40
50
60
0 10 20 30 40 50 60
Assigned Class
Figure 3-5: Gaussian Classifier Confusion Matrix
symbols [1] and [l].

This example illustrates one of the finer points of making a distinction between
symbols and characters. As seen in figure 3-6, it is debatable whether these are
actually separate symbols at all. Generally, it is reasonable to have the same symbol
represent multiple characters only when a human observer would always require
context to determine the intended character. However, if the distinction between
two symbols is subtle enough not to be modeled by the classifier, it may be more
simple to treat them as if they were the same. Since the symbol classifier was able
to perform significantly better than chance at classifying the types of symbols in
3-6, it was decided that they were distinct symbols, though ones which would be
often confused.
Digit 0 Letter o Digit 1 Letter l
Figure 3-6: Symbols Commonly Confused by a Gaussian Classifier
Determining which symbols to use would be even more problematic in a multi-
29
user system, where users may not always write the same characters with the same
symbols. The character 5, for example, is sometimes written with a single stroke
symbol, and sometimes with a two stroke symbol, as in figure 3-7. Though they
look very similar, the points in single and double stroke symbols are treated dif-
ferently by this classifier, and so to accurately recognize both styles of 5, models
would need to be made of each symbol.
Figure 3-7: Two styles of writing the digit 5
3.5.2 Generalization Ability
A second measure of performance seeks to answer a more subtle question, that

of the classifiers ability to generalize from the training data. It is undesirable to
request that a user enter hundreds of examples of each of the symbols that they
wanted to classify. Therefore, a classifier should be able to perform well when
trained with only a few training examples. Figure 3-8 shows performance statis-
tics on a test set for classifiers trained on 50, 100, 150, and 200 training examples
from each class. Error rates are plotted as a function of training set size for the
best and worst performance on the test classes as well as for the average across
all classes. Even with only 50 training examples, performance was fairly good
for most classes, and even 100% correct for some. More training examples did
not affect the performance of these models. However, it did help improve the
performance obtained on the commonly confused classes previously mentioned,
therefore raising the average performance slightly.
30
1.2
Percent Correct Across All Classes
Maximum
1
Average
0.8
0.6
0.4 Minimum
0.2
0
50 100 150 200
Number of Positive Training Examples
Figure 3-8: Classifier Generalization Performance
31
Chapter 4
Partitioning an Expression
The symbol classifier described in the previous chapter classifies symbols in isola-
tion; given a set of strokes, it can determine the symbol they best represent. How-
ever, there is no way of knowing a priori which strokes in an expression should be
combined together into symbols or even how many symbols are present. There-
fore, the capabilities of the symbol classifier need to be used within a larger frame-
work to determine the quantity and locations of the symbols present in an expres-
sion. Solving this problem for printed text is equivalent to finding the best group-
ing of a set of strokes into a set of symbols, called a partition of an expression.
4.1.1 The Existence of a Partition
The most fundamental assumption of the proposed solution is that such a parti-
tion even exists. As discussed in section 2.2, handwritten mathematics is typically
printed, so that each stroke belongs to a single symbol. Without this property, it
would still be possible to assign strokes to symbols, but the nature of the solution
would be very different, since a single stroke may then belong to multiple sym-
bols. The consequence of this assumption is that there is no straightforward way
to extend this solution to the types of script forms that sometimes appear in hand-
32
written mathematics, such as those in figure 4-1. In this example, the characters
c, o, and s are all written with a single stroke. To recognize such a form in
this framework they would need to be considered as a one stroke symbol, [cos], as
would every other script combination.
Figure 4-1: An Expression with Script Characters
4.1.2 Optimal Substructure
A second crucial assumption of this solution is that optimal partitions have optimal
substructure or obey what Bellman called the principle of optimality. For this prob-
lem, this means that an optimal partition of a large expression can be found by
combining optimal partitions of smaller expressions.
Figure 4-2: An Expression Correctly Partitioned into Symbols
This property is best illustrated by an example. Figure 4-2 shows a set of strokes
correctly partitioned into symbols, where the strokes in each identified symbol
have been boxed. Some symbols in the expression, such as the [3] and [2], are one
stroke symbols, while others, such as the [4] and [=], are two stroke symbols. If this
partition exhibits optimal substructure, the strokes in any subset of the symbols
are also optimally partitioned. If they were not, then the principle of optimality
dictates that the partition of the complete expression would be suboptimal, since it
33
could be improved by improving the partition of that subset. For a more complete
discussion of why this might be so and other properties of problems with optimal
substructure, see [6] and [3].
4.1.3 Additive Cost Function
Any stroke set partitioning solution must have a means of measuring the quality
of a particular partition. For a solution to assume optimal substructure, then, is
equivalent to its measure of quality exhibiting this structure. In this case, the cost
of a partition is the sum of costs of the symbols identified by that partition. The
cost of a single symbol is the log likelihood of the best interpretation of that stroke
set given by the symbol classifier. This measure has an additive cost structure typ-
ical of such optimization problems and therefore can take advantage of dynamic
programming techniques.
In order to make this measure viable, though, an additional factor is necessary.
If the summed cost of the symbols in a partition were the only measure of optimal-
ity, there would be a natural tendency to prefer partitions with fewer symbols to
those with more symbols, even if the average cost of the symbols in the latter ex-
pression was lower. For example, if a three stroke expression was partitioned into a
single symbol, the cost of the partition would only be the cost of interpreting those
strokes as one symbol. On the other hand, if the same expression was partitioned
into three separate symbols, the cost would be the sum of the cost of all of them.
Thus, such a measure would tend to combine strokes into multi-stroke symbols
when it was unwarranted. To counteract this tendency, a term called the combi-
nation weighting is added to the cost of multiple stroke symbols to equalize their
weight with those of single stroke symbols. The combination weighting term is
multiplied by the number of strokes in the potential symbol, less one, so that single
stroke symbols are not weighted, double stroke symbols have a single additional
weight, and triple stroke symbols are incremented by two times the combination
weighting.
34
4.2 Orders of Growth
Within this framework, the only way to guarantee that the best partition is found is
to evaluate the cost of every possible partition and choose that with the lowest cost.
A quick analysis of the size of this search space reveals that, even using dynamic
programming, an exhaustive search is not practical for this problem and so search
constraints are necessary to find a good partition in a reasonable amount of time.
The order of growth of the search space can be analyzed with respect to two in-
teresting quantities. The first quantity, F (N ), is the number of stroke sets that are
examined as possible symbols, as a function of the number of strokes, N . Since this
operation involves processing the strokes and then making a comparison with ev-
ery symbol model, it is a fairly time intensive operation that should be performed
rarely, if possible. If a symbol is allowed to have an arbitrary number of strokes,
F (N ) grows with the size of the power set of the set of strokes in an expression,
since each possible subset of that set will need to be separately compared to the
symbol models. The power set is exponential in size.
F (N ) = 2N = (2N )
In practice, however, the number of strokes that can be considered as a symbol

is limited to the number of strokes of the largest symbol model, some small value k,
typically less than 5. In this case, we find that the the number of possible symbols
examined in an exhaustive search grows polynomially with the number of strokes
in an expression.
k
!
N
= (N k )
X
F (N ) =
i=1 k
There is another important quantity which affects the order of growth of this
problem, though, which is the number of actual sub-partitions considered in a
search, G(N ). Using dynamic programming, the optimal partition is found by first
finding the optimal sub-partitions and then combining these together to produce
35
a solution for the entire expression. Again, this is possible because of the optimal
substructure of an additive cost function.
In a completely general search, every possible sub-partition needs to be exam-
ined at least once. As with F (N ), G(N ) grows with the size of the power set of
N,
G(N ) = 2N = (2N )
Even if the number of strokes allowed in a single symbol is bounded as before,

this quantity remains the same, since there is no reason not to consider a partition
based only on its size. Therefore, additional constraints will be necessary to keep
the search space to a manageable size.
4.3 Constraints
4.3.1 Time Ordering
One simple constraint is to consider only strokes written consecutively in time

as potentially belonging to the same symbol. With this restriction, a user would
need to write an expression one symbol at a time, completing the strokes in each
symbol before moving on to the next. This constraint, used by [10], creates a chain
structure, where each stroke can only be combined with the strokes that occurred
before or after it in time. In this case, the sub-partitions that need to be explored
are simply sets of strokes who form a complete sub-chain, significantly reduces the
search space of the problem such that,
F (N ) = (kN )
Since for each additional stroke, st , added to the expression at time t, only k
36
possible stroke sets need be considered as symbols:
{st }, {st , st1 }, . . . , {st . . . stk+1 }

N +1
Additionally, G(N ) = 2
= (N 2 ), since choosing a sub-chain is equivalent
to choosing an beginning an end to the chain from N + 1 possible choices. Taking
advantage of the chain structure, G(N ) can be reduced further to (kN ), since the
optimal partitioning after stroke st can be determined using only the k possible
symbols st may belong to and the results of previous partitions.
4.3.2 The Minimum Spanning Tree Constraint
While the time ordering constraint is a good first step to reducing this problem
to a manageable size, it is overly restrictive. Writers often want to change their
previously written expressions by augmenting a symbol, e.g. converting a [<] to
a []. It is especially common for writers to cross a [t] or dot an [i] after other
symbols are written. To deal appropriately with these issues, a constraint based on
a common structure known as a minimum spanning tree (MST) may be used.
A spanning tree of a connected, undirected graph is a set of edges which con-
nect all of the vertices of the graph with no cycles so that a path exists between any
two vertices. To define a minimum spanning tree, the graph edges also need to be
weighted, in which case an MST is a spanning tree of a graph whose summed edge
weight is less than or equal to that of all other spanning trees of that graph. Mini-
mum spanning trees can be efficiently computed using the algorithms of Kruskal
and Prim [6].
A minimum spanning tree can also be defined for a set of points in a metric
space. Here, the points represent the vertices of a fully connected graph and the
weight of the edge between two vertices is the distance between their associated
points. The MST for this graph can then be computed as before. Similarly, one
can define an MST for a set of strokes by choosing a distance metric between the
strokes. For simplicity, I chose to define the distance between any two strokes to
37
be the Euclidean distance between the centroids of their bounding boxes1 . Figure
4-3 shows a minimum spanning tree of the strokes in a typical expression.
Figure 4-3: A typical minimum spanning tree
Once a minimum spanning tree is found for a set of strokes, the search for
an optimal partition can be constrained by considering only partitions that form
connected subtrees in the MST. Consider the example in figure 4-4. The stroke
set represents eight symbols, including three double stroke symbols: [x], [+], and
[y]. If the correct partition is to be found, then the system needs to consider the
possibility that the strokes in these symbols belong to the same symbol. If the
tree did not contain an edge between these strokes, then the possibility that they
belong to the same symbol is not explored and they will either be partitioned into
their own symbol or combined with other strokes. For example, this MST prevents
the strokes in the [2] and [] from ever being considered as a single symbol since
there is no direct path between them. This is an effective use of the constraint
since these two strokes do not belong to the same symbol. There is an edge in the
tree between the strokes in every two stroke symbol, however, meaning the correct
partition was covered by the MST search.
Practical Running Time Analysis
As with any constrained search, the partition resulting from an MST search may
be suboptimal; This is the price paid for not examining all possible partitions. In
practice, though, the MST method is very good at covering the correct partition in
a wide range of expressions. This is because multiple stroke handwritten symbols
1
The intersection of the diagonals of the smallest box containing all the points of the stroke.
38
Raw Strokes
Minimum Spanning Tree
Optimal Partition
Figure 4-4: Using a Minimum Spanning Tree to Partition
almost always have strokes which are written closer together than strokes of dif-
ferent symbols. The MST constraint is also a fast way of solving a difficult problem
both because computing the tree can be done quickly and because the tree struc-
ture is easy to search recursively, much like the chain structure discussed earlier.
To repeat the growth analysis for searches constrained by minimum spanning
trees, it is necessary to consider both a best and worst case formation for the tree.
In the best case, the tree forms a chain, as shown in figure 4-5. Here, each stroke
is only connected to two others, resulting in a structure identical to that produced
by the time ordering constraint. Therefore, F (N ) = (kN ) and G(N ) = (kN ) as
before.
39
In the worst case, the tree forms a star, as shown in figure 4-6. In this case, the
number of symbols explored, F (N ), grows polynomially with N , since any two of
the outer strokes have a path connecting them through the center stroke.
k
!
N 1
= (N k1 )
X
F (N ) =
i=1 k1
The number of partitions necessary to consider in a star formation is equal to

the power set of N-1, so that,
F (N ) = 2N 1 = (2N )
While in the worst case the problem is still exponential, minimum spanning
trees for mathematical expressions are typically long chains punctuated with small
star formations. Quite often making the actual running time close to the theoretical
best case, resulting in a very efficient algorithm in practice.
Figure 4-5: The chain formation of a minimum spanning tree
Figure 4-6: The star formation of a minimum spanning tree
40
Chapter 5
Parsing
After an expression has been partitioned into symbols, the most difficult part of
the work is done. However, there still remains the problem of parsing the expres-
sion. That is, determining which characters those symbols represent, deciding on
the structure of the expression, and finally generating an interpretation of the ex-
pression as a command in a typesetting language.
Under the umbrella of mathematical notation there exists a wide range of dis-
tinct notations, including those for simple algebra, calculus, matrices, theorems
and proofs, and others. As a result, any mathematical expression recognition sys-
tem will most likely be unable to deal with every notational form ever devised.
For this reason, the parsing framework described in this chapter aims not only to
recognize a number of common notations, but also to provide a general foundation
for expression parsing which is expandable to other notational forms.
The current system is able to recognize common forms including fractions, rad-
icals, summations, accents, and superscripts. It is still very limited in the input it
can accept, however. In particular, it assumes that the expression to be parsed is a
single mathematical statement. So, if the user decides to draw two lines of math-
ematics, let alone a diagram or graph, the system will attempt to find a single line
of mathematics that best fits the input, however meaningless it may be.
41
5.2 The Structure of Mathematical Expressions
Up until this point, the fact that the expression being recognized is mathemati-
cal has not been used in any significant capacity to aid in its recognition. Indeed,
the symbol recognition and expression partitioning algorithms would probably
work well for any handwritten notation with printed symbols, whether it is text,
mathematics, or even musical notation. However, if the symbols are to be given
syntactic or semantic meaning, then some knowledge of the structure of mathe-
matics will need to be incorporated into the recognition algorithm. Inspiration for
this task may be drawn from typesetting languages, since they were designed with
the specific intent of formalizing the relationships that appear on the printed page.
Additionally, having a structure based on a typesetting language makes it easier to
generate output for existing typesetters.
5.2.1 Basic Mathematical Typesetting
The structure of mathematics is sufficiently different from that of handwritten

prose that typesetting languages often have a set of commands devoted entirely
to mathematics. To handle the tree-like structure of mathematics, languages such
as TEX use a recursive command structure to describe the layout of characters on a
page.
The most basic a mathematical expressions in TEX are merely lists of characters
to appear on a page from left to right. For example the command,
\alpha+\beta=\gamma
tells TEX to create the expression,
+ =
42
More complex characters require special constructs, however. For example,
a fraction bar requires a numerator and denominator as arguments, as in the
following command,
\frac{1+3}{x}
which tells TEX to create the expression,
1+3
x
For most characters in mathematics, the appearance of the symbol which rep-
resents them on the page is defined entirely by their size and the location relative
to the baseline. Digits and letters, such as the 1, +, 3, and x in the ex-
pression above fall into this category. The appearance of other types of characters,
however, depends on the characters which form their arguments. In the example
above, the fraction bar has a numerator and denominator which determine its
length. Another type of non-simple character is an accent, whose location depends
on the character it is accenting, as with the character hat in t vs. a
.
Some characters also have the ability to have nested arguments, whereby an
arbitrary expression can appear as an argument for a character. In the following

example, the expression 22 appears as arguments for a fraction bar, radical,
and summation sign. Note that an expression does not need to be valid mathe-
matics to be typeset.
q
22
22 X
22
22
22
5.3 Box Relationships and the Geometric Grammar
TEX typesets an expression by using the command codes to arrange boxes on a page.
Boxes are items which represent a region of the page and its associated image. One
of the most common types of boxes is that which contains a single symbol from a
43
typeface. Boxes can also contain other boxes, though, and TEX uses a sophisti-
cated set of rules to convert the structure inherent in the command language into
an arrangement of boxes. The problem of parsing a set of symbols can be par-
tially solved by reversing this process. Unfortunately, the process is not entirely
reversible since multiple characters may map to the same symbol.
Much of the ambiguity can be avoided by ensuring that the parse is valid under
a geometric grammar. In this grammar, every character belongs to a grammar type
according to its usage in mathematics. Most characters are categorized as simple,
meaning they take no arguments, but accents, delimiters, grouping operators, fractions
and radicals are each represented by different grammar types.
These grammar types can be combined with one another using a well defined
set of geometric relationships. The relationship between any two boxes is classi-
fied as one of 6 fundamental types, in/out, up/down, left/right, upper left/lower
right, upper right/lower left, or identical. Examples of some of these relationships
are shown in figure 5-1.
Up/Down Relationship In/Out Relationship
Left/Right Relationship Upper Right/Lower Left Relationship
Figure 5-1: Some bounding box relationships
The location of a symbol can then be used to determine which character it rep-
resents. For example, the characters and tilde accent both use the same
symbol. However, within the framework of this grammar a tilde accent needs
44
to have a simple character below it while a may not. These two characters can
then be easily distinguished in an expression such as
. Some characters can
be easily confused with others of their same grammar type, though, such as the
characters b and 6. In this case, the grammar is unable to disambiguate the
interpretation since they are both simple characters.
5.4 The Parsing Algorithm
Given a set of symbols, the parsing algorithm begins by selecting one of the sym-
bols to be the key. Which symbol becomes the key is generally unimportant, as
long as the key lies at the root level of the expression. That is, the key must not be
a symbol in an expression which is an argument to another character. For example,
in the expression in figure 5-2, the symbols [4] and [3] should not be chosen as the
key because they are arguments to the fraction bar, nor can the symbol [2] be
chosen because it is an argument of the radical. There appears to be a dilemma
in choosing a key, however. In order to be absolutely sure to choose an appropri-
ate key, the structure of the expression needs to be known. But to determine the
structure, a key must first be chosen. This dilemma is avoided by simply picking
the widest symbol as the key, since it is almost invariably lies at the root level of
the expression. Special cases can then be checked for and the key reassigned if
necessary.
Figure 5-2: A Set of Partitioned Symbols
Once a key has been chosen, all the other symbols in the expression are grouped
according to their bounding box relationship to the key symbol. If there are sym-
45
bols above, below, or inside the key, those symbols, along with the key, are as-
signed to the best characters whose grammar type compatible with the observed
relationship. For example, if one symbol is observed above the key and another
below, then the key may assigned to either a fraction bar or a grouping opera-
tor, since those types of characters are allowed to have other characters above and
below them. This assignment is accomplished by using the known mapping be-
tween symbols and characters and the rankings provided by the symbol classifier
to create an ordered list of possible characters for a set of symbols. This list is then
examined to find the best character with a specific grammar type.
If the key is determined to be a character which takes arguments, the arguments
are recursively parsed. Finally, all the characters to the right and left of the key
character are recursively parsed, and the subexpressions are linked together to
form a tree structure, which is traversed to generate the final typesetting command.
5.5 Superscripts
This parsing algorithm can also be extended to contain rudimentary support for
superscripted expressions, such as ex . Detecting a superscripted expression is a
more difficult task than detecting a grouping operator because whether or not one
symbol is a superscript of another depends on the baseline of those symbols.
An additional property of boxes not mentioned previously is that a baseline
can be defined for each box. The baseline of a character is not always the lowest
point in its bounding box because some characters descend below the baseline.
Likewise, the bounding box of some characters does not ascend to the full height
of the symbol since some characters such as [a], [e], and [o] are short. Determining
the baseline and size of the characters in an expression is necessary for detect-
ing superscripts since superscripted characters are typically written smaller than
characters on the baseline. Figure 5-3 gives some examples of characters which
all share the same baseline, even though their bounding boxes differ in size and
vertical position.
46
Figure 5-3: Characters which share the same baseline
For this reason, the simple relationships defined in the geometric grammar,
such as up/down or right/left, are not fine enough to detect superscripted charac-
ters. So, each simple character is also assigned a baseline type such as descender(e.g.
g, q), ascender (e.g. f , t), or short (e.g. a, c). After an expression is
parsed it is searched for superscript relations by keeping track of the number of
c
levels that have been superscripted in a stack (ab is a two level deep superscripted
expression). A seperate stack is maintained for each subexpression in nested ex-
pressions, as well.
While this scheme is of some utility, it will eventually need to be replaced with
a more robust method of determining sub and superscripted expression. In par-
3
ticular, it is unable to handle complex exponentiations such as 2 4 or symbols with
both sub and superscripts, such as x2i . Robust detection of baselines would also be
necessary to deal appropriately with multiple line expressions and complex matrix
forms such as the one in the following expression.

k

a b c
A matrix form unable to be handled appropriately by the current system
47
Chapter 6
Demonstration Program Engineering
6.1 Overview
The demonstration program for this thesis was written in Java as an example of
how these techniques could be incorporated into a general purpose mathemati-
cal expression recognition engine. Utilizing object oriented design, the engine is a
software object with a well-defined application programming interface (API) that
client software can use to send basic stroke data to the engine and request ob-
jects representing interpreted expressions. In addition, the recognition engine has
methods for changing recognition parameters, such as the combination weight-
ing and rejection threshold, as well as for graphically displaying research-related
information, such as the minimum spanning tree and the optimal partition.
In the system, the client software is a simple Java application which first collects
the stroke data as mouse events, then sends the data to the recognition engine, and
finally displays the interpretation in three forms, as a image of the final typeset
expression, as a TEX string, and as a MathML expression. A screen image of the
client application can be seen in figure 6-1.
The demonstration software was largely written by myself, utilizing standard
Java utilities and the Java abstract windowing toolkit (AWT) for the graphical user
interface elements. As an object oriented program, most the elements in the recog-
nition process have associated Java objects. In addition to the class representing the
48
Figure 6-1: Screen Image of the Client Application
recognition engine itself, there are classes representing bounding boxes, strokes,
symbol models, symbols, characters, expressions, as well as a number of utility
classes.
In addition to these classes, third party software is used for both numerical
analysis and displaying typeset expressions. Numerical analysis, including all
matrix operations, is done using the Java Matrix Package, a joint effort from The
MathWorks and the National Institute for Standards and Technologies. Typeset
expressions are displayed using WebEQ from Geometry Technologies.
6.2 User Interface
The user interface of the demonstration program was designed to explore some of
the concepts which would make a mathematical recognition system practical a vi-
able alternative to pen and paper. It is important for such an interface to be simple,
since it may be used on a computer with only a pen and touch screen for input.
In addition, it is also important that the interface be able to provide immediate
feedback to the user, so that errors can be quickly detected and corrected.
49
6.2.1 Pen Interface
Modern graphical interfaces are geared towards providing the user with a multi-
tude of options at the click of a mouse button. Many systems have separate menus
of options for each of three buttons on the mouse, which can be further augmented
by combining keyboard strokes with button presses. Palmtop computers do not
have the space for a full-sized keyboard and mouse, relying instead on a stylus
and touchscreen for input.
Since complex sequences of menus are difficult to navigate with a pen, the sys-
tem must be more attuned to what the user is trying to do and present them with
a limited set of options at any time. Furthermore, they must be able to effectively
distinguish writing strokes from command strokes. Often, this is accomplished by
setting aside a special area of the screen for written input and using the rest of the
screen for command strokes. Instead, the demonstration interface uses the same
region for both written expression input and command strokes, such as deleting
and correction command. Additional functionality is provided through a row of
buttons along the top of the writing surface, visible in figure 6-1.
6.2.2 Error Detection
Since the system is not actually capable of knowing for certain that it has made an
error, the term potential error detection might be more appropriate. The user inter-
face uses two very different styles of error detection, stroke rejection and symbol
boxes. In stroke rejection it is the system which is detecting potential errors, while
symbol boxes are a feedback mechanism used by the user to detect errors.
Stroke Rejection
During the partitioning stage, the strokes of an expression are partitioned into
symbols. Each of those symbols has a cost which is the lowest cost assigned to
the symbol by any of the symbol models. Any particular stroke or set of strokes
may be far from every model, however. In this case, it is likely that the user entered
50
an erroneous stroke, or that they drew a particularly poor example of the symbol
they were attempting to write. When this occurs it is often best to simply ignore
those strokes rather than pass them on to the parsing algorithm.
This can be accomplished by setting a rejection threshold, such that strokes that
cannot be assigned to symbols with costs less than that threshold are rejected.
These strokes can then be drawn in a different color on the screen to indicate to
the user that the stroke was determined to be unrecognizable. In practice such a
threshold works well, though many types of erroneous strokes are also good ex-
amples of simple symbols like [], [1], or [], and so are not rejected.
Symbol Boxes
Another type of error rejection can be achieved by drawing faint boxes around the
partitioned symbols. The user can then immediately see when a multiple stroke
symbol has been partitioned improperly. While this form of error detection is done
by the user rather the system, it is a relatively nonintrusive way of giving the user
access to some of what the system has decided about the expression they are draw-
ing. If the boxes are too much of an annoyance, they may always be turned off us-
ing the system preferences. Figure 6-2 shows an example of both of these forms of
error detection working in parallel. In this example, the erroneous stroke through
the [y] has been faded and left unboxed, to indicate that it is unrecognizable.
6.2.3 Error Correction
The complement to error detection is error correction. If the user was required
to rewrite an expression from scratch every time an error was made, they would
quickly tire of using the system. For this reason, good error correction techniques
can greatly improve the usability of the system. Since error correction is another
form of input, it is important that it neither require a keyboard or nor involve user
interface elements more suited to a mouse than a pen, such as nested menus.
51
Figure 6-2: An Example of Error Detection
Erasing
One of the simplest forms of error correction is to allow the user to erase a set of
erroneous strokes. Some digitizing tablets are able to distinguish the point from
the eraser end of the pen, and so allowing the user to erase strokes with the eraser
would be a guaranteed method of determining when the user wanted to erase
strokes. However, most handheld computers do not support this feature so an
alternative way to erase strokes is necessary.
One simple way of erasing strokes is giving the user the option to delete the
last stroke written. In the demostration program, this option is provided through
the use of a button above the writing region. Often, however, the user wants to
erase a stroke that was written long before the last stroke. To do this the system
introduces erasing gestures, whereby a user can erase a stroke or set of strokes by
moving the pen rapidly back and forth over the strokes. This is already a common
feature on many pen-based computers.
52
Correction Menus
Another way of correcting errors is through the use of pop-up menus. After an
interpretation of a symbol has been made the user can tap the pen on the symbol
and be presented with a short list of possible characters the symbol may represent.
This list is generated by using the ranked list of interpretations generated by the
symbol classifier along with the character to symbol mappings. If the user specifies
a particular character for a symbol, then this interpretation is fixed and cannot
be changed unless the user recalls the menu and selects a different character. In
addition, the menus for multiple stroke symbols give the option of breaking the
symbols apart so that the strokes will not be considered as being part of the same
symbol in future interpretations.
6.2.4 Preferences
Finally, the demonstration program allows the user to set preferences for the re-
jection threshold and combination weighting, as well provides them with display
options for displaying the minimum spanning tree and symbol boxes as in figure
6-2. Furthermore, when these options are changed the current expression is reeval-
uated in light of the new parameters, so that the effect of different values on the
interpretation of the same expression can be readily observed.
53
Chapter 7
Conclusions
The system described in this thesis has already performed well in a practical set-
ting, as it was used to efficiently typeset many of the mathematical expressions in
this thesis. Furthermore, recognition accuracy rates within its domain have been
very promising. However, it should be viewed as only a first step towards truly
robust recognition of handwritten mathematics. If these techniques are to be ex-
tended to a truly practical setting, a number of limitations will need to be over-
come. Some of these limitations constitute simple improvements to various parts
of the system, while other limitations arise at a more fundamental level.
The most immediate limitation of the current program is that it is still a single
user system, since the symbol recognition algorithm is naturally sensitive to vari-
ations in writing that were not present in the training data, such as that between
examples written by different writers. Modelling this variation is difficult, and still
an active area of research in handwriting recognition. One possible solution is a
user adaptive system, which uses erroneously classified symbols to adjust its symbol
models to a particular user. Another is to use a mixture model for each individual
symbol. No matter how it is approached, the problem of creating a truly user in-
dependent system lies largely in improving isolated symbol recognition, since the
other elements of the system are not nearly as sensitive to the variations in other
users writing.
Of the three main subproblems of this thesis, the stroke partitioning algorithm
54
is perhaps the most robust. The system is currently able to correctly partition the
symbols in expressions which it is entirely unable to parse, indicating that this
portion of the research should scale very well to more complex expressions. There
still remains some work to be done on normalizing the costs of multiple stroke
symbols with those of single stroke symbols. I view this problem to actually be a
problem in the symbol recognition algorithm though, rather than a limitation of
the stroke partitioning algorithm.
In addition, the area where the most research still needs to be done is in cor-
rectly parsing complex expressions. In particular, for the system to iterpret more
than simple expressions it will be necessary to incorporate a more complete use of
symbol baselines into the parsing algorithm. The expressions in this thesis which
could not be typeset using the system typically had either complex superscript or
subscript forms or matrix notations. A better understanding of how the baseline
of symbols effects the structure of the expression will be necessary if this problem
is to be solved.
Finally, there is still much work to be done on the user interface. As more
complex functionality is added to the system, the limitations of a pen interface
will quickly become appearent. At the very least, a full functioned equation editor
will need to be incorporated within the system.
55
Appendix A
Glossary
Bounding Box The smallest box which contains all of the points in a stroke or
symbol.
Centroid The intersection of the diagonals of a box.
Character A character is an element in a typesetting language.
Combination Weighting A parameter for determining how easily the system will
combine strokes into multiple stroke symbols. The lower the weighting, the
easier it is for strokes to be combined.
Expression A time-ordered set of strokes which are all written in the same coor-
dinate frame.
Parse A parse of a partitioned expression is both an assignment of characters to

symbols, and an interpretation of the structure of the expression.
Partition A partition of an expression, or set of strokes, is an assignment of strokes

to symbol such that no stroke belongs to more than one glyph.
Rejection Threshold A value for determining when a particular symbol is unrec-

ognizable. If a symbol is not able to be classified by one of the symbol models
to within that cost, then it is rejected.
56
Stroke The trajectory of a pen, expressed as a series of (x, y, t) coordinates, be-
tween the time it touches down on the writing surface and the time it lifts
off.
Symbol A handwritten symbol that represents a character. This term is used for
both for a particular instance of a symbol and for an entire class of symbols.
57
References
[1] R.H. Anderson. Syntax-directed recognition of handprinted two-dimensional

mathematics. Proc. 5th Int. Conf. Pattern Recognition, pages 436459, 1968.
[2] A. Belaid and J.P. Haton. A syntactic approach for handwritten mathematical
formula recognition. IEEE Trans. Pattern Anal. Machine Intell., PAMI-6:105
111, 1984.
[3] Dimitri Berksekas. Dynamic Programming and Optimal Control. Athena Scien-
tific, Belmont, Mass., 1995.
[4] Toru Wakahara Charles C. Tappart, Ching Y. Suen. The state of the art in on-
line handwriting recognition. IEEE Transactions on Pattern Analysis and Ma-
chine Intelligence, 12(8):789808, August 1990.
[5] Philip A. Chou. Recognition of equations using a two-dimensional stochas-

tic context-free grammar. In SPIE Vol. 1199 Visual Communications and Image
Processing, SPIE, pages 852863, Murray Hill, NJ, 1989. SPIE, SPIE.
[6] Thomas H. Cormen, Charles E. Leiserson, and Ronald L. Rivest. Introduction

to Algorithms. MIT Press, Cambridge, Mass., 1991.
[7] Yannis A. Dimitriadis and Juan Lopez Coronado. Towards and art based
mathematical editor, that uses on-line handwritten signal recognition. Pat-
tern Recognition, 28(6):807822, 1995.
[8] Richard J. Fateman and Taku Tokuyasu. Progress in recognizing typeset math-
58
ematics. In Document Recognition III, SPIE Volume 2660, pages 3750, Murray
Hill, NJ, 1996. SPIE, SPIE.
[9] Jesse F. Hull. Recognition of mathematics using a two-dimensional trainable

context-free grammar. Masters thesis, Massachusetts Institute of Technology,
Department of Electrical Engineering and Computer Science, June 1996.
[10] M. Koschinski, H. J. Winkler, and M. Lang. Segmentation and recognition of

symbols within handwritten mathematical expressions. In ICASSP, 1995.
[11] Hsi-Jian Lee and Min-Chou Lee. Understanding mathematical expressions

using procedure-oriented transformation. Pattern Recognition, 27(3):447457,
1994.
[12] Hsi-Jian Lee and Jiumn-Shine Wang. Design of a mathematical expression

understanding system. Pattern Recognition Letters, 18:289298, 1997.
[13] William A. Martin. A fast parsing scheme for hand-printed mathematical ex-
pressions. MIT AI Project Memo 145, Massachusetts Institute of Technology,
Computer Science Department, MIT, Cambridge, MA, October 1967.
[14] Erik G. Miller and Paul A. Viola. Ambiguity and constraint in mathemati-
cal expression recognition. In AAAI-98/IAAI-98 Proceedings, pages 784791.
AAAI, 1998.
[15] Z. X. Wang and C. Faure. Structural analysis of handwritten mathematical

expressions. 9th International Conference on Pattern Recognition, pages 3234,
1988.
[16] H. J. Winkler, H. Fahrner, and M. Lang. A soft-decision approach for struc-

tural analysis of handwritten mathematical expressions. In ICASSP, 1995.
59

Matsakis MEng 99

Uploaded by

Copyright:

Available Formats

Matsakis MEng 99

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Matsakis MEng 99

Uploaded by

Copyright:

Available Formats

Recognition of Handwritten Mathematical Expressions

Submitted to the Department of Electrical Engineering and Computer Science

Thesis Supervisor: Paul A. Viola

3 Isolated Symbol Classification 20

6 Demonstration Program Engineering 48

1-1 Screen image of the recognition system . . . . . . . . . . . . . . . . . 11

2-1 Two characters represented by the same symbol . . . . . . . . . . . . 16

3-1 The Data Collection Program . . . . . . . . . . . . . . . . . . . . . . . 21

4-1 An Expression with Script Characters . . . . . . . . . . . . . . . . . . 33

5-1 Some bounding box relationships . . . . . . . . . . . . . . . . . . . . . 44

6-1 Screen Image of the Client Application . . . . . . . . . . . . . . . . . . 49

The problem of machine recognition of handwritten expressions has long been a

1.2 System Overview

In this thesis, I describe an on-line approach for converting a handwritten math-

1.2.1 Isolated Symbol Classification

One of the basic problems in handwriting recognition is determining, out of con-

1.2.2 Expression Partitioning

In addition, a minimum spanning tree based in interstroke distances is used to

1.2.4 User Interface

An accurate recognition algorithm still needs a good user interface if it is to be a

Research in handwriting recognition has produced a rich glossary of terms pertain-

Figure 2-1: Two characters represented by the same symbol

Figure 2-2: Two symbols representing the same character

2.2 Qualities of the Problem

Any approach to machine recognition of handwriting must address certain prop-

2.3 Previous Work

Isolated Symbol Classification

3.1 Problem Description and Assumptions

Isolated symbol classification is the problem of providing a ranked list of inter-

Figure 3-1: The Data Collection Program

3.3.1 Stroke Direction and Order Invariance

One of the pitfalls of on-line approaches to handwriting recognition is that there

3.3.2 Scaling, Shifting, and Resampling

3.3.3 Derivative Terms

3.3.4 Principal Component Analysis

Figure 3-3: Eigenvalues Associated with the Principal Components

3.4 Gaussian Density Estimation

The density of unknown symbol z under model j

Figure 3-4: Examples of Reconstructed Characters

3.5.1 Confusion Matrix

One standard measure of classifier performance is the rate at which it misinter-

Figure 3-5: Gaussian Classifier Confusion Matrix

symbols [1] and [l].

Digit 0 Letter o Digit 1 Letter l

Figure 3-6: Symbols Commonly Confused by a Gaussian Classifier

Determining which symbols to use would be even more problematic in a multi-

Figure 3-7: Two styles of writing the digit 5

3.5.2 Generalization Ability

A second measure of performance seeks to answer a more subtle question, that

Figure 3-8: Classifier Generalization Performance

4.1 Problem Description and Assumptions

4.1.1 The Existence of a Partition

Figure 4-1: An Expression with Script Characters

4.1.2 Optimal Substructure

Figure 4-2: An Expression Correctly Partitioned into Symbols

4.1.3 Additive Cost Function

In practice, however, the number of strokes that can be considered as a symbol

Even if the number of strokes allowed in a single symbol is bounded as before,

4.3.1 Time Ordering