Note : Les descriptions sont présentées dans la langue officielle dans laquelle elles ont été soumises.
CA 02497586 1993-09-03
MFmHOD AND APPARATUS FOR RECOGNIZING CURSIVE WRITING
FROM SEQUENTIAL INPUT INFORMATION
15
BACKGROUND OF THE INVENTION
This invention relates to cursive script recognition
and in particular to cursive handwriting recognition methods
and apparatus, that is, recognition of characters and words
formed by a continuous stroke, wherein information on stroke
order is provided in connection with input of stroke position
information.
The field of computer-aided cursive script
recognition is of interest since many individuals do not have
the skills to communicate through a keyboard and since many
tasks would be greatly expedited by the use of direct input to
a computer through a familiar writing instrument, such as a
pen, pencil or stylus. Cursive handwriting recognition is
particularly challenging because the ciphers are formed by
continuous strokes and because handwriting differs widely among
individuals.
The work of Dr. Shelja A. Guberman of the former
Soviet Union, one of the coinventors, forms the basis of the
present invention. In a paper published in the Russian
language journal Avtomatika i Telemekhanika, by Shelya A.
Guberman and V. V. Rozentsveig under the title "Algorithm for
the Recognition of Handwritten Text," (No. 5, May, 1976, pp.
122-129, UDC 681.39.06) the developers describe the state of
the known art of cursive handwriting recognition and propose
that the dynamic parameters of pen trajectory be used in
connection with the various recognition algorithms. Among the
CA 02497586 1993-09-03
2
recognition algorithms were feature matching to identify
strokes, recognition of the start and finish of a trajectory,
and the subdivision of a trajectory into elements, or as termed
hereinafter, metastrokes. The developers, however, limited
their alphabet of metastrokes to just seven self-intersecting
elements and three arc elements. Moreover, the work was silent
about the possibility of confusion among elements in the
recognition process and did not consider the use of dynamic
programming techniques. As a consequence, further development
has been warranted in order to improve reliability and utility
of recognition.
Elements of a number of techniques similar to those
used in connection with the present invention have been
disclosed in the past. These references serve to illustrate
the state of the art. Details of specific embodiments of the
present invention which might use these prior art techniques
are therefore not described in depth. The following references
are nevertheless of interest in the field of cursive
handwriting recognition.
Ehrich and Koehler, ~~Experiments in the Contextual
Recognition of Cursive Script,~~ IEEE Transactions on Computers,
Vol. C-24, No. 2, February 1975, pp. 182-194. This paper
describes the use of segregation techniques between main bodies
of letters and ascenders and descenders as part of a
prerecognition scheme.
U.S. Patent No. 3,996,557 to Donahey describes a
similar technique to that of Ehrich et al.
U.S. Patent No. 3,133,266 to Frischkopf describes
normalization in the Y direction, use of dictionary matches for
recognition and rating of likelihood of accuracy of recognition
of individual words.
U.S. Patent No. 3,969,698 to Bollinger et al.
describes an apparatus for post processing of words which have
been misrecognized by a character recognition machine, a speech
analyzer, or a standard keyboard.
U.S. Patent No. 4,610,025 to Blum et al. describes
the isolation of ascenders and descenders as part of an early
analytical step and the isolation of words for identification.
CA 02497586 1993-09-03
3
U.S. Patent No. 4,731,857 to Tappert and U.S. Patent
No. 4,764,972 to Yoshida et al. both describe word isolation as
part of the analysis process.
U.S. Patent Nos. 4,933,977 and 4,987,603 to Ohnishi
et al. describes the elimination of extraneous marks in an
input pattern, as well as the concept of recognition of
elements that are less than complete characters, including
straight lines, arcs and loops.
U.S. Patent Nos. 3,111,646 and 3,127,588 to Harmon
describe systems using features extraction recognition
techniques as well as stroke sequence information.
U.S. Patent No. 4,754,489 to Bosker describes a
system for recognizing letter groupings called digrams and
trigrams.
U.S. Patent No. 5,034,989 to Loh describes a method
for identifying individual handwritten letters.
These prior art references provide a background
context for developing an understanding of the present
invention. The present invention builds on prior techniques,
combines many techniques not heretofore combined, and
introduces new techniques heretofore unknown in order to attain
an enhanced level of recognition.
SUMMARY OF THE INVENTION
According to the invention, a method and apparatus is
provided for cursive script recognition employing an
intermediate cursive words descriptive language constructed of
metastrokes. Metastrokes are elements or fractions of a stroke
shape which are used to represent a stroke. A stroke is
defined as the continuous segment which begins when the pen
touches the digitizer tablet's surface and ends when the pen is
lifted from the surface, comprising one or more written
letters. A cursive word may comprise one or more strokes
depending upon whether or not the pen is lifted from the
surface during the writing of the word. According to the
invention, a minimum of twenty metastrokes is typical to form
the necessary vocabulary or set of allowable stroke elemental
shapes, although as many as seventy metastrokes may be used.
CA 02497586 1993-09-03
4
Additional metastrokes increase the invention's recognition
accuracy at a cost of increased computing time. Techniques for
interpreting the input metastroke segments as words include
comparing the input metastroke segment with metastroke segments
in a dictionary of words "spelled" as metastrokes and assigning
a value for the likelihood of a match value to each word so
identified. The assigned value for the likelihood of match is
referred to as the "word metric." In further detail, this
technique involves comparing the input to each word in the
dictionary, metastroke by metastroke, assigning a metastroke
score to each position where a metastroke match is found, then
adding the scores for the matched metastrokes for each segment
so tested to obtain the word metric, ranking the words by word
metric by comparing the word metrics against a preselected
threshold as well as against each other for closeness to one
another, then, either simultaneously or consecutively,
substituting, adding and deleting metastrokes in the sequence
and then comparing the input metastroke segment so modified
with the dictionary to obtain further word metrics until a word
metric is found which satisfies the threshold criteria. In the
dictionary, there are typically many predefined metastroke
descriptions ("spellings") for each possible word because of
variations in handwriting styles. The process of the invention
involves recognition of whole words and not individual letters
as has been suggested by others. Instead of seeking to segment
a handwritten word into individual letters and seeking the best
first letter then the best second letter and so on, the
invention uses a process of computing scores of matching. The
number of possible matches is very large for typical sequences
of metastrokes and hence, according to the invention, the
process involves adding, deleting and substituting to both
decrease the size of the search and to speed the matching
process.
In a specific embodiment of the invention,
recognition equipment may include a digitizing pad for
inputting signals having values representing a sequence of
points in a coordinate system with indicia of a beginning point
and an end point and processing means for performing various
CA 02497586 1993-09-03
functions on the signals, which may be incorporated in
dedicated computer equipment or into data preprocessing means
of a computer operating system. The preprocessing means may
perform the following processes: segmenting the sequence of
5 points bounded by the beginning point and the ending point into
candidate words, strings or segments, differentiating between
the ending point of a word segment and that of an individual
word, combining word segments to form complete words,
establishing a bottom baseline and a top baseline for the
candidate word to determine height and scale of the candidate
word, reconstructing the candidate word as a reconstructed word
with a substitute sequence of points, wherein there are
interpolated points inserted and spurious points deleted,
locating and tabulating, in sequential order, critical points
in the reconstructed word, including maxima, minima,
intersections, dots and crossovers, replacing the reconstructed
word with a sequence of metastrokes according to the vocabulary
of metastrokes according to the invention, wherein each
metastroke is representative of an element of a handwritten
cipher. In particular, this replacing step may comprise
comparing the critical points against indicia of known critical
points to obtain a preliminary metastroke sequence for each
said reconstructed word. Thereafter, the technique
contemplates selecting, by means of a matrix of likelihood of
matching between each of said metastrokes, from entries in a
dictionary of entries compiled from variants of known sequences
of said metastrokes, many of the most likely matches in
likelihood order. This selecting step may comprise choosing
the preferred metastroke at selected positions according to a
maximum score calculation. There may be an analysis performed
on i) a forward sequence of the metastrokes, on ii) a reverse
sequence of the metastrokes, and on iii) most likely beginnings
of words and simultaneously on most likely endings of words.
The maximum score calculation may comprise an analysis of the
weighting of the results of additions, deletions, and
substitutions of metastrokes relative to adjacent metastrokes.
The results are provided to an output device in the form of
preferably only one, but possibly several, candidate words
CA 02497586 1993-09-03
6
based on a listing of most likely matches from the dictionary.
There is typically an indication of positive recognition of a
single candidate word if the likelihood of match figure of
merit exceeds a preselected threshold value.
In a further detail of a specific embodiment, the
methods may include measuring the average slant of sequences of
points of each candidate word segment and then dividing the
candidate word segment along horizontal boundaries into a
middle zone between the lower baseline and the upper baseline,
where bodies of letters are predicted to reside, into an upper
zone wherein ascenders of letters are predicted to reside, and
into a lower zone wherein descenders of letters are predicted
to reside. This height information, along with the shape
information provided by the identification of metastrokes,
assists the dictionary in determining the merit of any
metastroke at a position in a sequence of acceptable
metastrokes in its dictionary. The height information is also
weighted as part of the correlation.
The invention will be better understood, and greater
details on aspects of the invention will be apparent upon
reference to the following detailed description taken in
conjunction with the accompanying drawings.
BRIEF DESCRIPTION OF THE DRAWINGS
Figure 1 is a block diagram of a general recognition
system according to the invention.
Figure 2 is a block diagram of functional elements of
the invention.
Figure 3 is a table illustrating the metastrokes in
accordance with the preferred embodiment of the present
invention.
Figure 4 is a first f low chart of a method according
to the invention.
Figure 5 is a second flow chart of a method according
to the invention.
Figure 6A is an illustration of the handwritten word
"may "
CA 02497586 1993-09-03
7
Figure 68 is a possible metastroke sequence based on
the word "may" in Figure 6A.
Figure 7 is a possible dictionary spelling of the
word "may."
Figure 8 is a possible correlation matrix for the
word "may."
Figure 9 is an illustration of a sample metastroke
feature correlation table of data for substitutions used to
force an input sequence to look like a vocabulary sequence.
Figure l0A illustrates sample penalties for the
addition of selected metastrokes used to force an input
sequence to look like a vocabulary sequence.
Figure lOB illustrates sample penalties for the
deletion of selected metastrokes used to force an input
sequence to look like a vocabulary sequence.
Figure 11 is an illustration of a height correlation
table corresponding to the table of Fig. 9 used in
substitutions to force an input sequence to look like a
vocabulary sequence.
Figure 12A indicates sample height penalties
corresponding to the table of Fig. l0A used in additions to
force an input sequence to look like a vocabulary sequence.
Figure 128 indicates sample height penalties
corresponding to the table of Fig. lOB used in deletions to
force an input sequence to look like a vocabulary sequence.
Figure 13 is a complete metastroke feature
correlation table of data for weights for substitutions used to
force an input sequence to look like a vocabulary sequence, for
penalties for additions (column 1), and for penalties for
deletions (row 1) defining the similarity measure for
metastrokes for one embodiment of the invention.
Figure 14 is a complete height correlation table for
the embodiment corresponding to the feature correlation table
of Figure 13 illustrating the weights and penalties for
substitution, additions (column 1) and deletions (row 1) used
to force an input sequence to look like a vocabulary sequence.
CA 02497586 1993-09-03
8
DETAILED DESCRIPTION OF SPECIFIC EMBODIMENTS
Referring to Figure 1, there is shown a recognition
apparatus 10 for cursive handwriting in accordance with the
invention. The apparatus 10 comprises first of all as its
input device a digitizer pad 12 or equivalent mechanism, which
produces as.output in response to input from a stylus 14,
"input" signals having values representative of a sequence of
points symbolizing handwriting in a coordinate system defined
by the digitizer together with a symbol or indication of the
beginning point and ending point of each sequence of points.
The signal for the beginning point and the ending point may be
simply a negative value generated automatically whenever the
stylus 14 is lifted from the digitizer 12. The x-y coordinate
sequence of points is fed to an input interface 16, which in
turn provides the values of these coordinate points to an input
buffer means 18. Input buffer means 18 provides for temporary
or off-line storage while the input sequence is processed. The
apparatus 10 further comprises a processor 20 with associated
program storage means 22, such as read only memory (ROM), and
in-process storage means 24, such as random access memory
(RAM). Control and data lines are provided between the
processor 20 and the in-process storage means 24 and the
program storage means 22. The program storage means 22 is for
storing, permanently or semipermanently, executable computer
programs, a features table for identifying metastrokes as
hereinafter explained, correlation matrixes or tables related
to the metastrokes, and dictionaries of metastrokes for
identifying words or the like.
The in-process storage means 24 is for the temporary
storage of candidate words derived from the input data,
baselines computed from the words, reconstructed words based on
interpretation and normalization, as well as other related
processes, critical points of the reconstructed words for
comparison with the features, metastroke sequences derived from
the features tables, substitute metastroke sequences generated
as the result of maximum score calculations and figures of
merit for various maximum score calculations developed from use
of the correlation tables. The processor 20 executes the
CA 02497586 1993-09-03
9
programs of the program memory means 22 to eventually generate
a word identification or a word to be supplied to an output
device or output interface 26.
Referring to Figure 2, there is shown in greater
detail the functional elements of an apparatus 10 in accordance
with the invention. The digitizer 12, through the input
interface 16, provides the start mark, the stop mark and the
position data to buffer means 18. The input buffer means 18
may identify the boundaries of a candidate word by means of the
start mark and the stop mark, indicative of a break in the
input sequence of data. The input buffer means 18 may also
identify the boundaries of a candidate word segment by
comparing the length of different data sequence breaks;
determining which breaks are indicative of a word break and
which breaks are indicative of a break within a word. Means 28
are provided for establishing orientation and specifically for
determining the direction of writing to in such a way that the
direction of writing becomes parallel to a horizontal
ref erence .
Means 30 are thereafter provided for reconstructing
the candidate word (as a "reconstructed word") with a
substitute sequence of points. The substitute sequence of
points may have points inserted in the candidate word at
interpolated locations and spurious points deleted from the
candidate word. Spurious points are those points which seem to
have no relation to any sequence of strokes, either as a
continuation of a segment or as a point, such as a period or an
accent mark.
Thereafter, means 32 are provided for establishing a
bottom baseline and a top baseline for each candidate word to
determine height and scale. More specifically, the baseline
establishing means 32 incorporates therein means for dividing
the candidate word horizontally into a middle zone between a
lower baseline and an upper baseline calculated by examining,
for example, the density of points above and below each
baseline, and designates the space above the baseline as
ascenders of letters and designates the zone below the baseline
as descenders of letters. The element 32 also includes means
CA 02497586 2005-11-21
for measuring average slant of the selected sequences of
points, for example, as determined by average trajectory of
each of the selected ascenders and descenders in the ascender
zone and the descender zone, respectively. Means 32 also
5 includes means for normalizing the data to a desired scale.
The means 32 then conveys its data to a critical
point locator 34. The purpose of the critical point locator 34
is to locate and tabulate in sequential order all relevant
critical points in the reconstructed word segment produced by
to the point interpolator 30. Examples of critical points are
maxima, minima, intersections of line segments, dots and
crossovers. The critical points thus tabulated can then be
compared with a table containing known critical points
identifying metastrokes:
According to the invention, metastroke recognizes 36
is provided which processes the features evidenced as critical
points and substitutes for the reconstructed word segment a
string of metastroke identification codes. A minimum of twenty
metastrokes is typical to form the necessary vocabulary or set
of allowable stroke elemental shapes, although as many as
seventy metastrokes may be used. Additional metastrokes
increase the invention's recognition accuracy. In the
preferred embodiment there are thirty-three different defined
metastrokes or cursive writing features.
Referring to Figure 3, there is shown a table of
metastrokes with the definitions thereof for the preferred
embodiment. These metastrokes have been determined, according
to the invention, to provide for identification of English-
language words in a computer-stored dictionary.
The metastrokes identified in the table of Figure 3
are as follows:
a stroke with a wide break,
a "wild card" stroke which could be anything,
a generalized horizontal stroke,
a feature maximum,
a feature minimum,
an angle without a loop,
a stroke with a small break,
CA 02497586 1993-09-03
11
a dot representing a period or the like,
a crossover,
a backward up arc with a free end at the beginning of
the arc,
a backward up arc with a free end at the termination
of the arc,
a backward up arc with no free end,
an upside down gamma drawn with counter-clockwise
strokes,
a circle drawn with counter clockwise strokes,
a gamma drawn with a counter clockwise stroke, a
forward down arc without a free end,
a forward down end with a free end at the start,
a forward down arc with a free end at its
termination,
a forward up arc with a free end at its termination,
a forward up arc with a free end at its start,
a forward up arc without a free end,
an upside down gamma drawn clockwise,
a circle drawn clockwise,
a gamma drawn clockwise,
a backward down arc without a free end,
a backward down arc with a free end at its
termination,
a backward down arc with its free end at the start,
a left directed or horizontal of any arc,
a right directed or horizontal of any arc,
a generalized vertical component,
a generalized subarc at the maximum left end of a
segment, and
a down arc of any direction and a subarc at the
maximum right end of any word segment.
Referring to Figure 2, the metastroke code string
(hereinafter the metastroke string) generated by the metastroke
recognizer 36 is provided to a word recognition subsystem 38
according to the invention for recognizing words in a specific
language, such as English. The processing function of one
element thereof will be explained in greater detail
CA 02497586 1993-09-03
12
hereinafter. Elements of the word recognition subsystem 38
include a "whole-word"-based analyzer means 44 for performing
maximum score analysis by dynamic programming procedures as
hereinafter explained to attain optimal matching between a
string of input metastrokes and a "vocabulary" metastroke
string forming a known whole word. Subsystem 38 may optionally
also include a "letter"-based analyzer means 42 for analyzing
metastrokes by letter (alphanumeric character). Some of these
types of "letter"-based analyzer means 42 are known in
connection with character-based or segmented string
recognition. Such analysis is not to be confused with whole-
word-based analysis techniques which are hereinafter explained.
The letter-based analyzer means 42 may work in
combination with or complement the whole-word analyzer means
44. The word recognition means 38 may include for example a
stored dictionary 40 of whole words comprising metastrokes in
direct order and in reverse order. The letter-based analyzer
means 42 may use some of the same dictionary entries as the
whole-word analyzer means 44 in procedures operating
independently and in parallel to the whole-word analyzer means,
in an attempt to recognize characters as strings of letters
forming meaningful words.
The output of the word recognizes 38, which may be
some form of best selection chosen from a variety of analysis
procedures, is provided to a word output device 26 as
previously noted.
Referring to Figure 4, there is shown a flow chart of
processes in accordance with the invention from input to output
of metastroke recognizes means 36. According to the invention,
the input sequence is provided as a string of x and y
coordinate values for points, along with a start and stop
indicator (step A), then a bounded string is selected as a
candidate whole word from the input string received from the
input device (step H); thereafter direction of writing is
established and the candidate word is "rotated" (oriented in
its frame of reference) to make the orientation of the writing
parallel to a horizontal reference (step C). The candidate
word is then reconstructed as a reconstructed word to remove
CA 02497586 1993-09-03
13
spurious points and add missing points (step D). Thereafter, a
top baseline and a bottom baseline are established (step E),
and the candidate word is normalized (step F).
Thereafter, critical points are located in the
sequence of points (step G). Critical points, such as maxima,
minima and intersections, as well as the order of critical
points, are useful for identifying metastrokes and the order of
metastrokes. Critical points are then matched against
allowable vocabulary metastrokes to identify the string of
input metastrokes used to "spell" a candidate word (step H).
Step H is repeated for each critical point of the input word
until all of the critical points have been matched against
metastrokes (step I). The result is the string of input
metastrokes used to "spell" the candidate word. The metastroke
string is then tabulated for further processing (step J).
A flow chart for a portion of the whole-word-based
analyzer system 44 according to the invention is shown in
Figure 5. In this process, words in the dictionary are
excluded so they need not be analyzed further. The process
first tests using the direct vocabulary (the forward direction)
and then the reverse vocabulary (in the reverse direction).
First, an "empty" word is created (step K) and placed in a
buffer (step L). An "empty" word is a "null" set, that is, a
word having all of the characteristics that words should have
but having no meaningful value. It is a place-holder, much as
zero is a place-holder in an initialized storage register.
The content of the word buffer is then tested to see
if the word buffer is empty (step M). If the word buffer is
not empty, then the next entry (the string representing the
candidate word) is retrieved from the word buffer for analysis
against the dictionary(step N). This will occur whenever an
input string from the input device is placed in the word
buffer. Using the dictionary of whole words as a source, a
list is then compiled of all ("n") possible words which are
candidates to match the latest entry retrieved from word buffer
(step O). The compiled list is thereafter tested to be sure it
is not empty (step P). (The list could be empty at the outset
if the critical point test, above, has been performed and no
CA 02497586 1993-09-03
14
matches were found. The list will also be empty at the end of
a whole-word analysis when the list has been exhausted.) After
testing to determine that this list is not empty, the "next"
vocabulary word in the list is retrieved for use in analysis
against the retrieved entry, namely, against the input
candidate word retrieved from the word buffer (step Q). The
retrieved vocabulary word is compared as a whole word to the
retrieved input entry according to the invention by dynamic
programming techniques as hereinafter explained, from which an
"optimal cost value" or maximum score is derived to determine
the degree of match between the input candidate word and the
vocabulary word (step R). Steps P through R are repeated until
the list is exhausted (Step P) or until the maximum score meets
certain criteria for acceptability (step S). If the list is
emptied, a "no match" signal is generated, and the process is
continued by looking for the next word in the input word
buffer. If the acceptance criteria are satisfied by whatever
acceptance criteria are then applicable, the word is
communicated to a word output buffer (step T) for further use,
such as display or processing. The process then continues to
the next input metastroke sequence in the entry buffer (steps K
and L) .
The step of optimal match calculation employs two
inputs: the input sequence of metastrokes forming a string for
a word and the pattern or sequence of "vocabulary" metastrokes
from the dictionary representing a known word.. Each vocabulary
sequence of metastrokes describes one of the accepted ways of
rendering a word. Data for analyzing the optimal match is
extracted from a feature correlation table and a height
correlation table, as hereinafter explained.
A simplified example of the recognition procedure for
whole words follows for illustration purposes only. If the
word "run" was handwritten on a digitizer tablet, the first
step would be to construct an input metastroke string, or
sequence of preselected stroke-like ciphers based on accepted
stroke forms obtained from an analysis of curs ive handwriting.
The input metastroke string is a series of metastrokes, rather
than points or letters. The vocabulary of metastrokes for that
CA 02497586 1993-09-03
string is limited to an established number of choices, such as
20, 30, 40 or 50 different stroke shapes. Figure 3 above is an
example of one functional embodiment. The input string can be
represented compactly in a computer by a string of symbols, for
5 example the string of arbitrary symbols or their ASCII
equivalents: (~JV"V~~IIVAV. The sequence forming the input
metastroke string is then placed in the word buffer in its
symbolic form and, according to the whole word analysis
procedure of the invention, compared as a whole segment with
10 all relevant entries in a "dictionary" of whole words, spelled
out as metastrokes using as its vocabulary the same symbolic
forms used to form the input metastroke string. A sample
dictionary may for example comprise the entries:
Vocabulary Segments
15 Metastrokes W_ ord
(J JV,.V~.JI I 1 A - run
(J JV"~~~ ~( VA - run
(~)V"-'~~1 A1 A - run
- run
('JV"~~~.rJ J J - ran
A - ran
(~ J V.,N~..- ('~~ ~ - ran
(This is a very simple, two-word, seven-string dictionary.)
While comparing can be done by testing for a match
between the input metastroke string and each metastroke-spelled
word in the dictionary, metastroke by metastroke, noting by a
score or a value each position where a match between input
metastrokes and vocabulary metastrokes is found, and then
adding the scores of the matched metastrokes for each input
metastroke segment so tested to obtain a "word" metric (i.e.,
matching score) for each entry, that process is not what is
undertaken in the word analyzer means 44 in accordance with the
preferred embodiment of the invention. However, for
illustration purposes, such a matching scheme is described so
that the more complex example hereinafter given will be
understood more easily.
In the table above, the simplified comparison process
would produce results as follows, assuming that the maximum
CA 02497586 1993-09-03
16
score calculation allowed simple summation of the weights and a
maximum score for a single metastroke match is 6:
Vocabulary Input meta- "Word"
Metastr okesWord stroke segmen t Match Metric
* (~ ~V.,V~~ ~ = run (N~V"V~--1 / 6-66666666-6- 60
/ VAV
1
A
(~ ~Vr~l~'r ~ - run (N~V.,V~-~1 6-666-66-666- 54
VA / V11V
H1 = run (N~V.,V~-~1 66664-666--6- 54
A / VAV
(.~~V.,~~~ (~, - run (~)V.,V~~l / 6-666-66----- 36
) VAV
- ran (~~V"V~~1 / 6-666-66-6--- 42
VAV
* (~~V.,N~. .,__I1- ran (N~V"V~--1 / 66666-66---6- 48
VAV
- ran (N~V.,V~r~ / 6-666-66----- 36
VAV
The dash "-" appears wherever there is no match with
the metastroke at that position. The metastroke-spelled words
are typically grouped together by English-language definition,
so the "hit" is manifest (listed) as the English-language word
with a score which is the highest score of any of the
metastroke-spelled words in that grouping. This is indicated
by the asterisk "*" in the left margin, above.
The English-language words which are hits are ranked
by their highest word metric, this metric being first
normalized. Above, it would be:
"Word" Normalized
Word Metric a 'c
run 60 0.77
ran 48 0. 62
The normalized word metrics in this simplified
example can then be compared against a preselected threshold to
determine acceptability, as well as against each other for
closeness to one another to determine "confusibility." If the
word is determined to be acceptable, then the word is sent to a
word output buffer. If the word metric does not surpass the
threshold, then the input metastroke segment is changed by
substituting, adding and deleting metastrokes from the list of.
candidate English-language words.
CA 02497586 1993-09-03
17
In the above example, if a normalized metric of 0.77
did not surpass the threshold, various metastrokes might be
added, deleted or substituted in the input metastroke string to
try to obtain a better match with a known string in the
dictionary and a further comparison to the dictionary word
would be made. Additions and deletions are assigned negative
weights, whereas substitutions are assigned positive weights
depending upon the "direction" and position of the
substitution. If the last metastroke of the input segment was
deleted, the normalized metric of the first "definition" of
"run" would increase from 0.77 to 0.83, possibly meeting a
preselected threshold.
Figures 6 - 12 show a second, more accurate example
of the whole word recognition process according to the
invention more exactly illustrating how the dynamic programming
process of the present invention is carried out to attain
recognition of a whole word. Figure 6A shows the handwritten
word "may." In one embodiment of the invention this word can
be expressed as the sequence of metastrokes shown in Figure 8B.
The numbers next to particular features of the word indicate
the corresponding metastrokes (see the horizontal axis of
Figure 6B).
As indicated hereinabove, the dictionary can have
numerous metastroke "spellings" for an individual word, due to
the variations in possible writing style. In this example it
will be assumed that there is only one spelling of "may" in the
dictionary, as shown in Figure 7. Comparing the metastroke
sequence of Figure 6H with that of Figure 7 shows that the
match is not exact. (Note the letter "a".) Due to the many
variations in writing style, this is commonly the case.
To determine if the dictionary spelling of "may" is
considered a match with the input sequence, a word correlation
table is constructed comparing, as a whole, the input
metastroke string with each vocabulary metastroke string which
is a candidate for a match. As a consequence of the inventive
process, and a word metric (i.e., a value defined by the
likelihood of a match existing) and an optimal "path" through
the sequence of metastrokes results, but only after the
CA 02497586 1993-09-03
18
processing of the whole input metastroke string is completed.
The larger the word metric, the closer the match.
Figure 10 is an example of a resultant word
correlation table for the word "may' wherein the input
metastroke string of Figure 6B is matched against the
vocabulary metastroke string of Figure 7. The values in each
cell is derived from computations of added value based on
transitions between a preceding diagonal position, a preceding
row position and a preceding column position, wherein the
highest valued transition of these three choices is selected
and then the cell is assigned the sum of the weigh of the
transition and the value of the preceding cell. This is a
process of dynamic programming. The source of weights is
obtained by reference to a feature correlation table and a
height correlation table, the values of which are based on the
size and nature of the vocabulary of metastrokes and an
analysis of allowed and disallowed substitutions, insertions
and deletions between all metastrokes in the vocabulary.
Hy way of example, there is a feature correlation
table, Figure 9, which indicates the likelihood of a match
existing between individual metastrokes in a vocabulary of a
mere eight metastrokes. Each column entry represents a
metastroke found in the input string; each row entry represents
a metastroke found in the dictionary built from this (limited)
vocabulary of metastrokes. The highest value or metric is
assigned to a substitution of a vocabulary metastroke for an
identical input metastroke, i.e., along the diagonal. Lower
metrics are assigned to substitution of vocabulary metastrokes
for input metastrokes which are close in shape. Disallowed
substitutions are shown as blanks in the table, and a large
negative value is assigned to those positions in the table:
(-1000), so that computations based on those substitutions
always produce a result which is outside the range of
consideration for an overall match.
Input metastroke additions and deletions are allowed
by the invention although there is a penalty associated with
each. Figure l0A shows sample penalties for the addition of
selected metastrokes while Figure lOB shows sample penalties
CA 02497586 1993-09-03
19
for the deletion of selected metastrokes. Besides looking at
the correlation between metastrokes (positive values) and the
penalties associated with adding or deleting metastrokes
(negative or penalty values), a height comparison (relative to
a word's baseline) is also made. Figure 11 is a height
correlation table while Figures 12A-B indicate height penalties
associated with the tables of Figures l0A-B, relating
respectively to additions and deletions.
The word correlation table of Figure 8 for each
vocabulary word can be derived from data from the tables of
Figures 9, 10A, 10B, il, 12A and 12B, which is a simplified
example, or from data of Figures 13 or 14, which is from an
actual working system. (In Figures 13 and 14, the addition
vectors corresponding respectively to Figures l0A and 12A are
shown as the top row of the tables, and the deletion vectors
corresponding respectively to Figures lOB and 12B are shown as
the left-most column of the tables.) The word correlation
table of Fig. 8 is computed based upon a dynamic programming
technique for computing "optimal cost" as it is called in the
art and "path" for a transformation of one sequence into
another. In this invention the transformation or mapping is
made upon an entire input sequence to force it into the length
and form of any one of a number of known vocabulary metastroke
strings. It is therefore necessary to execute the
transformation upon the entire sequence before conclusions can
be reached about the results.
Figure 13 is a data table for showing the
transformations of input metastrokes into vocabulary
metastrokes. The left margin lists the vocabulary metastrokes,
i.e., the metastrokes found in the vocabulary. The top margin
may be labeled with the same metastrokes in the same order.
The diagonal of the table is the value assigned to a direct
transformation of an input metastroke to the identical
metastroke in the vocabulary. Off-diagonal values represent
values assigned for substitution of other vocabulary
metastrokes for the input metastroke in a process of building a
metastroke string which matches a metastroke string found in
the dictionary. Figure 14 is a height correlation table for
CA 02497586 1993-09-03
nine levels of height. Penalties are assigned for deletions
and additions, as illustrated along the zeroeth row and the
zeroeth column, respectively, of Figure 13 and as shown
separately in the example using Fig. l0A (corresponding to
5 column 0 of Figure 13) and Figure 10B (corresponding to row 0
of Figure 13.) In a deletion, a metastroke is deleted from the
input string. A deletion is indicated by a negative value of
preselected magnitude (Row Zero, any column). Similarly for
addition, a metastroke is added to a position in the input
10 metastroke string, and the action is indicated by a negative
value of a preselected magnitude (Column Zero, any row). A
"substitution" (where a vocabulary metastroke is substituted
for an input metastroke) is indicated by a positive value, as
indicated by the position in the matrix. The special case of
15 "substitution" of a metastroke for itself is the diagonal, and
is therefore assigned the highest value. There is absolutely
no requirement that the feature correlation table be symmetric,
as substitution is not a symmetric process.
The technique for interpreting the input metastroke
20 string as a word is straightforward but involves substantial
processing. In the dynamic programming approach, comparison,
weighting and modification processes are carried out
simultaneously. The process is analogous to "wiggling" a
"worm" to see if it can be made to line up with the shape of
some sample "worms." Additions, deletions and substitutions
cause the worm to wiggle. Each action results in a weighting
or a penalty.
The steps of the dynamic programming technique
include a process of comparing an entire input metastroke
string with all relevant metastroke strings in a dictionary of
words "spelled" as metastrokes to build a path for the
transformation of the input string into one of the dictionary
metastroke strings and determining the one maximum value or so-
called "optimal cost" for processing the entire input
metastroke string against each vocabulary metastroke string.
In theory, the entire input metastroke string could be
processed against all strings in the dictionary. However,
certain streamlining steps are typically taken to speed up the
CA 02497586 1993-09-03
21
processing. (To streamline the processing, not all dictionary
entries need to be examined, although the techniques to so
limit the processing are not a part of this aspect of the
invention. Such optimization techniques may take advantage of
features of the dictionary. The dictionary may for example be
arranged by number of strokes, or be "alphabetized" by
metastroke. Analysis may be performed on a forward sequence of
metastrokes, on a reverse sequence of metastrokes, on most
likely beginnings of words and on most likely endings of
words.) As previously described, included in the dictionary is
a "definition", an English-language (or any other language)
word, spelled out in Roman characters, i.e., the ASCII strings
corresponding to Roman characters, which is the single word to
be recognized. Thus there may be several dictionary entries
with the same "definition", to take into account handwriting
variations.
The metastroke feature correlation matrix (i.e., the
data for transformation of all metastrokes into each other), as
illustrated in Figure 13, is used as an integral part of the
dynamic programming process, namely, in the path building and
the optimal cost calculation. The height correlation matrix is
also used in the same processing steps. The optimal cost
calculation is a calculation of the following three formulas,
followed by a selection of a maximum based on a straight-
forward comparison. The maximum among three values is inserted
in the cell corresponding to the input metastroke-to-vocabulary
metastroke transformation in the word correlation table (e. g.,
Figure 8) for each vocabulary word so processed. Figure 8 is
an example built from a dynamic programming process using the
data from Figures 9-12.
Referring to Figure 8 for illustration purposes, the
calculation is performed by selecting the maximum value (a) for
the cell at position (a, b) in the word correlation table from
among a values calculated for addition, deletion and
substitution at those cell positions. The maxima in each
instance in Figure 8 is determined as follows:
CA 02497586 1993-09-03
22
For substitution:
ali - aji i + p(ai .bj) + q(aj.bj) [1]
For insertions or additions:
a2i - ai_1 + p(ai) + q(ai) L2]
For deletions:
a3~ - ajil + P(bj) + q(bj) [3]
wherein:
alji is the "cost value" (as used in the vocabulary
of dynamic programming), or cumulative score at the cell (i,j)
for passing from the origin via cell (i-l,j-1) in substituting
the vocabulary metastroke "a" at row position (i) for the input
metastroke "b" at column position (j) (Fig. 8);
p(ai,bj) is a similarity weighting value (extracted
from Fig. 13 or from Fig. 9) of the substitution of an input
metastroke "ai" by a vocabulary metastroke "bj";
q (a j , b j ) is the height weighting value ( from the
height correlation table, Fig. 14 or Fig. 12) for height
substitution occurring in the foregoing substitution of input
metastroke "ai" by vocabulary metastroke "bj";
a2 is the "cost value" or cumulative score for
passing from the origin via cell (i-l,j) to cell (i,j) in
inserting the vocabulary metastroke "a" at row position (i)
after the input metastroke "b" at column position (j);
a3 is the "cost value" for cumulative score for
passing from the origin via cell (i,j-1) to cell (i,j) in
deleting the input metastroke "b" at column (j) along a
sequence of metastrokes;
p(ai) is a penalty value (from Fig. 13, column 0, row
i; or from position i of Fig. l0A) for inserting the vocabulary
metastroke "a" at row position (i) after the input metastroke
"b" at column position (j);
q(ai) is the penalty value for height associated with
the foregoing insertion (Fig. 14 or Fig. 12A);
p(bj) is a penalty value (from Fig. 13, row 0, column
j; or from position j of Fig. lOB) for deleting a metastroke
nbjn.
CA 02497586 1993-09-03
23
q(bj) is the penalty value for height associated with
the foregoing deletion (Fig. 14 or Fig. 12B).
The preferred metastroke string at any cell (i,j) is
the maximum cumulative score value alpha (a) at cell (i,j)
selected from among the first cumulative score value (ai), the
second cumulative score value (a2) and the third cumulative
score value (a3). This maximum score at any cell is the
"optimal cost" at that cell along any path from the origin in
the word correlation table of Fig. 8.
A specific example may be helpful. Consider the cell
at row i=4, column j=4 in Figure 8. This cell contains the
maximum value among the computed alpha terms al, a2 or a3 for
that position.
For computing al, Equation [1] is applied. The value
aji i is 7 from examination of the adjacent diagonal cell Figure
10. The substitution shown in Figure 8 for cell (4,4) is from
the upward arc (along the top row) to a downward point (along
the side column). From Figure 9, the substitution of the
upward arc to a downward arrow, the value p(ai ,bj) is -1000
indicating a disallowed transition. The corresponding height
substitution from the fourth position of Fig. 6B to the fourth
position of Fig. 7 is from height 3 to height 5. A transition
from height 3 to height 5 has a value of -1000 from Figure 11.
The sum of aj-li-1, p and q is thus -1993.
For computing a2, Equation [2] is applied for
insertion. The value a~_1 is 11, from the adjacent row above
cell (4,4) in Figure 8. The penalty value p(ai) from Figure
10A, for insertion of an downward arrow is -2. The
corresponding height penalty value q(ai) for the downward arrow
at height 5 (Figure 7) from Figure 12A is 0. Hence the sum a2
and thus a candidate value for cell (4,4) is 9.
For computing a3, Equation [3] is applied for
deletion. The value a~-i is 12, from the adjacent column left
of cell (4,4) in Figure 8. The penalty value p(bj) from Figure
lOB, for deletion of a upward arc is -2. The corresponding
height penalty value q(ai) for the upward arc at height 3
(Figure 7) from Figure 12B is 0. Hence the sum a3 and
candidate value for cell (4,4) is 10.
CA 02497586 1993-09-03
24
Comparing al, a2 and a3, as through a simple sort,
the maximum value is 10. Hence the value 10 is inserted in
cell (4,4).
This process is continued for every cell of the word
correlation table. For every cell, the maximum a of the
substitution, addition and deletion is used. As the word
correlation table is computed, the path to each cell from the
adjacent cell which resulted in the maximum score is tabulated.
At the completion of the process, paths from any cell can be
traced back (e. g., from the right boundary or lower boundary)
to the origin (0,0). Each such path represents the optimal
path between the origin and the selected cell. In the present
invention, the right-most and lowest-most cell position from
the origin represents the "optimal cost" value for the optimal
path through the word correlation matrix for the specific input
string tested and terminated with its end mark. The specific
input metastroke string is matched to each vocabulary
metastroke string from the dictionary (built from the
vocabulary of allowable metastroke strings) to obtain numerous
word correlation tables. The optimal cost value for the input
metastroke string (which value is found at the cell in the last
column and last row and which corresponds to the termination of
the input metastroke string and vocabulary metastroke string)
of each word correlation table is then compared against all
corresponding optimal coat values from the other word
correlation tables. The maximum of these various optimal cost
values is used to identify the vocabulary metastroke string
which most closely correlates the input metastroke string with
a word ("definition") in the dictionary, assuming minimum
recognition criteria have been met. (In other words, a maximum
lower than an acceptable minimum is a basis for an indication
of a failure to recognize the input string as a word.
The system according to the invention provides a high
likelihood of recognizing words produced by the process of
cursive handwriting. There is enough redundancy built into
this system that even words written in poor handwriting with
missing letters and misspellings are given a reasonable
likelihood of recognition.
CA 02497586 1993-09-03
The invention has now been explained with reference
to specific embodiments. Other embodiments will be apparent to
those of ordinary skill in the art. It is therefore not
intended that this invention be limited, except as indicated by
5 the appended claims.
