DE4345583B4 - Cursive handwritten word recognition from sequential input information - using construction word correlation tables to select strings of vocabulary metastrokes corresponding to candidate word and identifies most likely match - Google Patents

Abstract

The recognition method involves receiving input signals having values representative of a sequence of points symbolizing handwriting. A beginning point and ending point of the sequence is also provided by the input signals. A sequence of points, bounded by the beginning point and the end point, is selected as a candidate word. The candidate word is replaced with a string of input metastrokes in sequential order. Each metastroke is representative of an element of a cursive handwritten cipher to obtain a preliminary metastroke string for each candidate word. Word correlation tables are constructed by using a feature correlation table of all metastrokes forming a vocabulary. The feature correlation table relates input metastrokes to vocabulary metastrokes. Strings of vocabulary metastrokes are selected corresponding to the candidate word. The strings of vocabulary metastrokes comprise a listing of entries compiled from variants of known strings of the metastrokes. A single candidate maximum score is extracted for each one of the word correlation tables only upon completion of the word correlation table. The most likely match between the candidate word and one of the strings of the vocabulary metastrokes is identified to an output device.

Description

TECHNICAL BACKGROUND

The The invention relates to handwriting recognition and relates in particular Handwriting recognition method and apparatus for handwriting recognition, i.e. the recognition of existing from a continuous polyline Signs and words, being information about the stroke sequence in connection with the input of bar position information are provided.

Of the Field of computer-aided Handwriting recognition is of interest, since many people do not the ability have, over a keyboard to communicate and since then many tasks through a direct input to a computer with a familiar writing instrument, e.g. a pen, pen or pen significantly accelerated become. Handwriting recognition is therefore a particular challenge, because the characters consist of solid lines and because people's manuscripts are very different from each other.

The Work of Dr. med. Shelja A. Guberman of the former Soviet Union, one of the co-inventors, forms the basis of the invention. In one Essay written in Russian in the journal Avtomatika i Telemekhanika, by Shelja A. Guberman and V.V. Rozentsveig the title "Algorithm for the recognition of a handwritten text "(No. 5, May, 1976, pages 122-129, UDC 681.39.06) The inventors describe the state of the art of handwriting recognition and suggest that the dynamic parameters of the pen line guide in conjunction with various Detection algorithms are used. To the recognition algorithms counting a feature comparison to identify the lines, the recognition of the beginning and end the trajectory and the subdivision of a trajectory into elements that hereinafter also referred to as metastrokes. The inventors have their metastroke alphabet on just seven self-cutting Limited elements and three arc elements. But the work is silent The possibility a confusion of elements in the recognition process and taken into account not the use of dynamic programming techniques. consequently Further development was focused on reliability and to improve the usability of recognition.

In In the past, elements of various techniques were revealed which are similar to those which are used in connection with the invention. These citations serve to illustrate the state of the art. Therefore are Details of special embodiments of the invention making use of these known techniques, not described in detail. The following pamphlets are but still important in the field of handwriting recognition.

Ehrich and Koehler, "Experiments in the Contextual Recognition of Cursive Script, "IEEE Transactions on Computers, Vol. C-24, No. 2, February 1975, pages 182-194. This Essay describes the use of separation techniques between the main parts of the letters and the ascenders and descenders of the Letters as part of a pre-recognition scheme.

US-A-3,996,557 (Donahey) describes a technique similar to that of Ehrich et al. is similar.

US-A-3,133,266 (Frischkopf) describes the standardization in the Y direction, the Use of dictionary contrasts for detection and estimation the probability of accuracy of detection of individual Words.

US-A-3,969,698 (Bollinger et al.) Describes a device for retrofitting Processing words, that of a word recognition engine, a language analyzer, or a standard keyboard incorrectly were detected.

US-A-4,610,025 (Blum et al.) Describes the isolation of upper and lower lengths of the Letters as part of a previous analytical step and the isolation of words for identification.

US-A-4,731,857 (Tappert) and US-A-4,764,972 (Yoshida et al.) Describe word isolation as part of an analysis procedure.

US-A-4,933,977 and US-A-4,987,603 (Ohnishi et al.) Describe the elimination of foreign characters in an input pattern as well as the concept of recognizing elements that are not complete characters, eg straight lines, arcs and loops.

US-A-3,111,646 and US-A-3,127,588 (Harmon) describe systems used by detection techniques for extraction characteristic features as well as bar sequence information usage do.

US-A-4,754,489 (Bosker) describes a system for recognizing letter groups, which are called digrams and trigrams.

US-A-5,034,989 (Loh) describes a method for identifying individual handwritten ones Letters.

Out GB-A-2087616 is an apparatus and method for generating an alphanumeric output based on shorthand known. The short characters entered on a tablet 10 are recognized and displayed on a screen.

Further Techniques for computer-based input of handwriting by means of Abbreviations are from US-A-5140645, US-A-4718102 and camel, K .; Imam, I. "A computerized transcription system for cursive shorthand writing "Southeastcon '88., IEEE Conference Proceedings, 11-13, Apr 1988, Pages: 336-339 known.

in the In view of the prior art, the object of the following The invention is to provide a method and a system with which handwritten words can be recognized more efficiently and reliably.

These Task is solved by the subject matter of claims 1 and 9.

preferred embodiments are the subject of the dependent claims.

The handwriting recognizing method and apparatus of the present invention makes use of an intermediate description language for handwritten words composed of metastrokes. Metastrokes are elements or sections of a stroke that are used to represent a stroke. A stroke is defined as a contiguous segment that begins when the stylus touches the surface of the digitizer's tablet and ends when the stylus is lifted off the surface, the stroke comprising one or more written characters. A cursive word may include one or more dashes depending on whether the pen is lifted off the surface during the writing of the word or not. According to the invention, at least twenty metastrokes are generally required to form the necessary vocabulary or set of allowed elementary stroke shapes, although no less than seventy metastrokes can be used. Additional metastrokes improve the recognition accuracy of the invention at the expense of increasing computation time. The techniques for interpreting the input metastroke segments as words involves comparing the input metastroke segments with metastroke segments from a dictionary of words "written" as metastrokes, and assigning a probability value of a match value for each word thus identified. The assigned value for the likelihood of the match is called a "word measure". Specifically, this technique involves comparing the input with each word in the dictionary, metastroke for metastroke, assigning a metastroke score to each metastroke match position, adding the metastroke match scores for each segment so tested to the dictionary Word order, arranging the words by the word measure by comparing the word measures both with a predetermined threshold and with each other in terms of their match, substituting, adding and deleting metastrokes in the sequence either simultaneously or consecutively and then comparing the words thus modified input metastroke segment with the dictionary to obtain further word measures until a word measure satisfying the threshold criteria has been found. Because of the different manuscripts, each dictionary typically has many predefined metastroke descriptions ("spellings") for each possible word. The method of the invention involves the recognition of whole words and not individual letters as suggested by others. Instead of searching to decompose a handwritten word into individual letters and instead of searching for the best first letter, then the best second letter, etc., the invention makes use of a method of calculating arithmetic ratings of the data moods. For typical metastroke sequences, the number of possible matches is very large, and thus the inventive method involves adding, erasing, and substituting to both reduce the scope of the search and increase the speed of the comparison procedure.

In a particular embodiment of the invention the recognition device a digitizing keyboard for input of signals with values representing a sequence of points in a coordinate system represent with indicia of a start and an end point, and a processing device for performing various functions with the signals coming in a dedicated computer device or in a data preprocessing device of a computer-aided system can be included. The preprocessing device may be the perform the following procedures: Dividing the sequence between the start and end points lying points in test words, strings or segment, distinguishing between the starting point of a Word segment and that of a single word, combining the Word segments to form more complete words, setting a lower baseline and an upper baseline for the check word, around the size and the scale of the test word determine the reconstructing of the checkword as a reconstructed one Word with a replacement sequence of points, with interpolated points added and false points deleted be, locating and tabulating critical points in the reconstructed word in sequential order, where the points maxima, minima, intersections, points and crossing points include reinstating the reconstructed word with a Sequence of metastrokes corresponding to the metastroke vocabulary according to the invention, wherein each metastroke is an element of a handwritten character represents. This re-insertion step can be described in detail Compare the critical points with indicia of known critical points include a preliminary Metastroke sequence for each to get reconstructed word. Accordingly, the invention is based on selecting of many of the best matching matches in a probability order from the entries in a compiled entries Dictionary of known metastatic sequence variants by means of a matrix of the match probability the individual metastasis. This selection step may include selecting one preferred metastroke at selected positions depending from a maximum score calculation. It can be an analysis on i) a forward sequence the metastroke, ii) a backward sequence the metastrokes and iii) the best fitting word beginnings and at the same time the best matching word ends are performed. The maximum score calculation may be an analysis of the weighting results, additions, deletions and substitutions include the metastrokes with respect to the neighboring metastases. The results are preferably in the form of, possibly also in the form of several test words based on a list of the most appropriate matches from the dictionary delivered to an output device. Normally a positive Detection of a single test word displayed when the probability of the correspondence figure of significance exceeds a predetermined limit.

In a particular embodiment The procedure may be measuring the average slope of Point sequences of each check word segment and dividing the check word segment along horizontal boundaries into a middle zone between one lower and upper baseline, assuming that yourself the body the letter in an upper area where the ascenders of the Letters be adopted and located in a lower area, in which the descenders the letters are accepted. This size information supports in conjunction with the contour information obtained by the identification of the metastrokes the dictionary in it, the value of each metastroke at a position in a sequence of acceptable metastrokes in the dictionary. The size information is also weighted as part of the correlation.

Under Reference is made to the following detailed description in conjunction with the accompanying drawings The invention will be better understood and further details of aspects of the invention will become apparent.

BRIEF DESCRIPTION

1 shows a block diagram of the general detection system according to the invention.

2 shows a block diagram of the functional elements of the invention.

3 Figure 14 shows a table illustrating the metastrokes according to a preferred embodiment of the invention.

4 shows a first flow chart of the method according to the invention.

5 shows a second flowchart of the method according to the invention.

6A shows a representation of the handwritten word "may".

6B shows a possible metastroke sequence based on the word "may" in 6A based.

7 shows a possible dictionary spelling of the word "may".

8th shows a possible correlation matrix for the word "may".

9 Figure 12 is a representation of a metastroke feature correlation table of data for substitutions to adapt an input sequence to a vocabulary sequence in appearance.

10A illustrates sample additions for the addition of selected metastrokes to match an input sequence in appearance to a vocabulary sequence.

10B illustrates sample scratches for deleting selected metastrokes to match an input sequence in appearance to a vocabulary sequence.

11 FIG. 12 is an illustration of a size correlation table corresponding to the table of FIG 9 to adapt an input sequence to a vocabulary sequence in substitutions.

12A shows sample size gains according to the table of 10A to match in additions an input sequence in appearance to a vocabulary sequence.

12B shows sample size reductions according to the table of 10B to customize an input sequence in erasures according to a vocabulary sequence.

13 Figure 15 shows a complete metastroke correlation table of weight substitution data to fit an input sequence in appearance to a vocabulary sequence, for addition (column 1) and deletion complications (row 1) aggravation measures, for the metastroke similarity measure for one embodiment to define the invention.

14 FIG. 12 shows a complete magnitude correlation table for the embodiment according to the feature correlation table of FIG 13 which illustrates the weights and aggravations for substitutions, additions (column 1) and deletions (row 1) to adapt an input sequence to a vocabulary sequence in appearance.

DESCRIPTION THE PREFERRED EMBODIMENTS

1 shows the detection device according to the invention 10 for italic handwriting. The device 10 initially has as an input device a digitizing keyboard 12 or a similar mechanism depending on the input by means of a pen 14 generates as output "input signals" having characteristic values for a sequence of dots symbolizing the handwriting in a coordinate system defined by the digitizer together with a symbol or a display of the start and end points of each sequence of dots. The signal for the starting point and the end point can simply be a negative value that is automatically generated when the pen 14 from the digitizer 12 is lifted. The XY coordinate point sequence is applied to an input interface which in turn sends the values of these coordinate points to an input buffer device 18 emits. The input buffer device 18 causes temporary or system independent storage while the input sequence is being processed. The device 10 further includes a processor 20 with an associated program storage device 22 , eg a read only memory (ROM), and an intra-process memory device 24 , eg a random access memory (RAM). Between the processor 20 and the intra-process storage device 24 and the program storage device 22 Control and data lines are provided. The program storage device 22 is used for permanent or semi-permanent storage of executable computer programs, a metastroke identification feature table, as described below, metastatic correlation matrices or tables and dictionaries of metastrokes for identifying words, etc.

The intra-process storage device 24 It is used for caching check words derived from the input data, baselines calculated from the words, reconstructed words based on interpretation and normalization, and other related processes, critical points of the reconstructed words for comparison with the features, feature tables derived metastroke sequences, as a result of the maximum scoring computations generated substitute metastroke sequences and figures of significance of various maximum scoring calculations developed from the correlation tables. The processor 20 executes the programs of the program storage device 22 Finally, to generate a word identification or word that can be sent to an output device or an output interface 26 should be delivered.

2 shows in a more detailed representation of the functional elements of the device according to the invention 10 , Via the input interface 16 gives the digitizer 12 the start mark, the stop mark and the position data to the buffer device 18 out. The input buffer device 18 can identify the boundaries of a checkword by means of the start and stop flags indicating an interrupt in the data entry sequence. The input buffer device 18 can also identify the boundaries of a check word segment by comparing the length of different data sequence breaks, determining which interrupts indicate a word break and which breaks indicate a break within a word. The device 28 is intended to provide orientation and, in particular, to determine the direction of the writing so that the writing is aligned parallel to a horizontal reference.

The device 30 is suitably provided to reconstruct the checkword (as a "reconstructed word") with a replacement sequence of points. The replacement sequence of points may include points inserted in the checkword at interpolated locations and false points deleted from the checkword. False points are such points which have no relation to any sequence of strokes, neither as a continuation of a segment nor as a point, eg as a period or an accentuating feature.

The device 32 is similarly provided to establish a lower baseline and an upper baseline for each checkword to determine magnitude and scale. The baseline detection device 32 comprises means for horizontally dividing the checkword into a middle region between a lower baseline and an upper baseline calculated by examining, for example, the density of points above and below each baseline, and characterizing the space above the baseline as ascenders of the letters and indicates the area below the baseline as descenders of the letters. The element 32 further comprises means for measuring the average slope of the selected point sequences as determined, for example, by the average trajectory of each of the selected descenders in the upper and lower length zones, respectively. The device 32 also includes means for normalizing the data to a desired scale.

The device 32 then delivers its data to a localization facility 34 for critical points. The task of the localization facility 34 for critical points lies in getting all relevant critical points in the reconstructed word segment through the point interpolator 32 is generated, in a sequential order to locate and tabulate. Critical points are, for example, maxima, minima, intersection points of line segments, points and crossing points. The critical points thus tabulated can then be compared to a table containing known metastrophic critical points.

A metastroke recognizer 36 is provided according to the invention, which processes the features identified as critical points and replaces a reconstructed word segment with a string of metastroke identification codes. In general, at least twenty metastrokes make up the required vocabulary or set of allowed primitive shapes, although not less than seventy metastrokes can be used. Additional metastrokes improve the recognition accuracy of the invention. In a preferred embodiment, there are thirty-three differently defined metastrokes or italic features.

3 shows a table of metastrokes with their definitions for the preferred embodiment. These metastrokes have been established in accordance with the invention to identify English-language words in a computer-stored dictionary.

In the table of 3 the following metastases are identified:
a dash with a wide interruption,
a "wild card" stroke that can mean anything
a general horizontal line,
a maximum feature,
a minimal feature,
an angle without a loop,
a dash with a small interruption,
a point that represents a period, for example,
a crossing point,
a backward upward bow with a free end at the beginning of the bow,
a backward upward bow with a free end at the bow end,
a backward upward bow without free end,
an inverted gamma drawn in counterclockwise strokes,
a circle drawn in counterclockwise strokes,
a gamma drawn with a clockwise stroke,
a forward bow, with no free end,
a forward bow with a free end at the beginning,
a forward bow with a free end at its end,
a forward bow with a free end at its end,
a forward bow with a free end at its beginning,
a forward bow without a free end,
a clockwise inverse gamma,
a circle drawn in a clockwise direction,
a clockwise drawn gamma,
a backward down bow without a free end,
a backward bow with a free end at its beginning,
a backward-looking bow with its free end at the beginning,
any left or horizontal arc,
any right-handed or horizontal arc,
a general vertical component,
a general arc portion at the leftmost end of a segment, and
a downward arc of any direction and an arc portion of the rightmost end of any word segment.

The through the metastroke recognizer 36 generated metastroke code string (hereinafter referred to as metastroke string) is according to the invention a word recognition subsystem 38 supplied to recognize the words in a specific language, eg English. The processing of one of these elements will be explained in detail below. The elements of the word recognition subsystem 38 comprise a "whole word" based analyzer device 44 for performing maximum evaluation analysis in dynamic programming procedures, as explained below, to obtain an optimal match between a string of input metastrokes and a known whole word "vocabulary" metastroke string. The subsystem 38 Thus, optionally, a based on "letter" analyzer 42 for analyzing metastrokes by letters (alphanumeric characters). Some of these letter based analyzer facility types 42 are known in conjunction with a letter or split string based recognition. Such analysis should not be confused with the word-based analysis techniques described below.

The letter based analyzer device 42 may be combined or supplemented with the whole word based analyzer device 44 work. The word recognition device 38 may include, for example, a stored dictionary of whole words that have metastrokes in direct and reverse order.

The letter based analyzer device 42 For example, in procedures that operate independently and in parallel with the whole word based analyzer, some of the same dictionary entries may be used as the whole word based analyzer 44 to try to recognize the characters as a string of letters that make meaningful words.

The output of the word recognition device 38 that can have a shape that's the best choice from different evaluation procedures is sent to a word output device 26 delivered as previously mentioned.

4 shows a flowchart of program flows according to the invention between the input and output of the metastroke detection device 36 , According to the invention, the input sequence is formed as a string of X and Y coordinate values for dots together with a start and stop indicator (step A), then a delimited string is selected as a whole check word from the input string obtained from the input device (step B) ; then the direction of the font is set and the check word is "rotated" (oriented in its frame of reference) to align the font parallel to a horizontal reference (step C). Then the check word is reconstructed as a reconstructed word to remove the wrong points and add missing points (step D). Thereafter, a lower base line and an upper base line are set (step E), and the check word is normalized (step F).

After that the critical points are located in the sequence of points (Step G). Both the critical points, e.g. Maxima, minima and Intersections, as well as the order of critical points are for identification the metastrokes and the order of the metastases. The critical points then become permissible vocabulary metastrokes faced, to identify the string of input metastrokes that for "writing" a test word are used (step H). Step H will work for everyone critical point of the input word repeated until all critical points were confronted with the metastases (step I). The result is a string of input metastrokes that for "writing" the test word is used. The metastroke string will then be for the further processing is tabulated (step J).

5 FIG. 10 is a flowchart for a portion of the whole word based analyzer system of the present invention. FIG 44 , In this method, words in the dictionary are excluded so that they need not be further analyzed. The tests are done in the process first using the direct vocabulary (the forward direction) and then using the opposite vocabulary (in the backward direction).

First an empty word is generated (step K) and stored in a buffer (step L). An "empty" word is a "null" sentence, i. a word, all the features of words has, but has no meaning. It's a placeholder, like zero is a wildcard in an initialized memory register.

Of 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 (containing the check word representing string) from the word buffer for evaluation against the dictionary found out (step N). This always happens when one Input string stored by the input device in the word buffer is. Using the dictionary of whole words as a source, then becomes a list of all ("n") compiled possible words, the test words are, which match the last entry found from the word buffer (Step O). The compiled list is then tested to be safe to be that is not empty (step P). (The list may be empty at the beginning, if the above mentioned Test of critical points has been done and no matches were found. The list can also be at the end of the whole words Analysis will be empty if the list has been used up.) After the Testing to determine that this List is not empty, the "next" vocabulary word in the list figured out in the analysis against the found entry, i. opposite to the word buffer found input check word, to be used (step Q). The retrieved vocabulary word is reflected as a whole word with the input entry found out of the invention by the dynamic programming techniques explained below compared, from which an "optimal Cost value "or a maximum rating is derived to match the degree of agreement between the input test word and the vocabulary word (step R). The steps P to R are repeated until the list is exhausted (step P) or until the maximum rating meets certain acceptance criteria (step S). If the list is cleared, a signal "no match" is generated and the process becomes continued by following the next Word in the input word buffer. if the Acceptance criteria met are, whatever for Acceptance criteria are then applicable, the word is sent to the word output buffer (Step T) for further use, e.g. for display or processing issued. The process then goes to the next Input metastroke in the input buffer via (Steps K and L).

The optimal fit calculation step includes two inputs: the input sequence of metastrokes forming a string for a word and the pattern or string of a known word "vocabulary" metastrokes from the dictionary. Every vocabulary sequence from Me Tastrichen describes one of the permitted ways to render a word. The data for analyzing the optimal fit is then extracted from a feature correlation table and a size correlation table, as described below.

Die Sequenz, die die Eingabe-Metastrichzeichenfolge bildet, wird dann in ihrer symbolischen Form in einem Wortpuffer abgelegt und entsprechend der auf ganzen Wörtern basierenden Auswertungsprozedur der Erfindung als ein ganzes Wortsegment mit allen relevanten Einträgen in einem "Wörterverzeichnis" von ganzen Wörtern verglichen, die als Metastriche geschrieben sind, wobei als Vokabular dieselben symbolischen Formen benutzt werden, welche die Eingabe-Metastrichzeichenfolge bilden. Ein Beispielwörterverzeichnis kann beispielsweise die folgenden Einträge umfassen:By way of illustration only, a simplified example of the whole word recognition procedure follows. If the word "run" is written by hand on a digitizing panel, the first step is to construct an input metastroke string or a sequence of selected string-like numerals based on approved stroke shapes obtained by analyzing the italic handwriting , The input metastroke string is more a series of metastrokes than dots or letters. The metastroke vocabulary for this string is limited to a fixed number of choices, such as 20, 30, 40 or 50 different stroke styles. The above 3 is an example of a functional embodiment. The input string can be represented contiguously in a computer by a string of symbols, eg the string of random symbols or their ASCII equivalents:

The sequence forming the input metastroke string is then stored in its symbolic form in a word buffer and compared as a whole word segment with all relevant entries in a "dictionary" of whole words corresponding to the whole word based evaluation procedure of the invention Metastasis are written using the same symbolic forms that form the input metastroke string as a vocabulary. For example, a sample dictionary may include the following entries:

vocabulary segments

(This is a very simple dictionary with two words and seven strings.)

While the comparison can be made by checking the input metastroke string and any metastatic spelled word in the dictionary for correspondence, Metastrich for metastroke records that every item with a score or value that matches the input Metastrokes and the vocabulary metastrokes and that the metastroke scores for each input metastroke segment tested in this way are then added together to obtain a "metric" (ie, a match score) for each entry, this process is not what is according to the preferred embodiment of the invention in the word analyzer 44 is carried out. This matching scheme will be described for the purpose of illustration only to facilitate understanding of the more complex embodiment below.

In The above table results from the simplified comparison method the following results assuming that the maximum score calculation a simple summation of the weights allowed and a maximum rating for one single metastroke match 6 is the same:

Of the Hyphen "-" occurs at the point on, at which a match not present with the metastroke. The spelled as metastases words are usually summarized by an English-language definition, so that the "hit" as an English-language Word is presented (listed) with a rating which the highest Evaluation of all words written as metastrophic in this group is. This is indicated by the asterix "*" am left margin, see above.

The English words, the hits are classified by their highest word size, being that measure first is normalized. It follows:

The normalized word measures can in this simplified example then both with a predetermined Limit value to determine acceptability, as well as among themselves in terms of their compliance compared to determine the "confusion". If it is stated that the Word is acceptable, the word is then sent to a word output buffer output. When the word measure the Does not exceed the limit, is the input metastroke segment by substituting, adding and Clear changed the metastroke from the list of English-language test words.

If in the above example, a normalized measure of 0.77 did not exceed the limit has different metastrokes in the input metastroke string added, deleted or substituted to try to get a better match with a known string in the dictionary and it will be another comparison with the word from the dictionary made. Additions and deletions are characterized by negative weightings, while substitutions by positive weightings depending on the "direction" and position of the Substitutions are marked. If the last metastroke of the input segment deleted becomes, the normalized measure of the first "definition" of "run" from 0.77 to 0.83 to, possibly it meets a predetermined limit.

The 6 to 12 show a second more specific example of a whole word recognition method according to the invention, which illustrates in more detail how the dynamic programming process according to the invention is executed in order to recognize a whole word. 6A shows the handwritten word "may". In one embodiment of the invention, this word may be used as a string of the type defined in 8B expressed metastrokes. The numbers near special features of the word indicate the corresponding metastrokes (see the horizontal axis of 6B ).

As noted above, due to differences in possible handwriting styles, the dictionary may have numerous metastroke "spellings" for a single word. In this example, it is assumed that there is only one notation of "may" in the dictionary, as in 7 is shown. A comparison of the metastatic sequence of 6B with the of 7 shows that the match is not exact. (Note the letter "a.") This is generally the case due to the various variations in the handwriting styles.

To determine if the dictionaries of "may" are the same as the input sequence In the meantime, a word correlation table is created that as a whole compares the input metastroke string with each vocabulary metastroke string that is eligible for matching. As a consequence of the method according to the invention, and a word measure (ie a value defined by the probability of an existing match) and an optimal "path" through the sequence of metastroke results, but only after the processing of the whole metastroke string has ended. The larger the word size, the greater the match.

10 shows an example of a word correlation table resulting from the word "may", wherein the input metastroke string of 6B the vocabulary metastroke string of 7 is faced. The values in each cell are derived from the calculation of an addition value based on the transitions between a previous diagonal position, a previous row position and a preceding column position, from which three selected values the highest value transition is selected and then the cell the sum of the weight of the transition and the value of the previous cell is assigned. This is a process of dynamic programming. The source of weights is obtained by reference to a feature correlation table and a size correlation table whose values are based on the size and nature of the metastrogram vocabulary and an analysis of allowable and invalid substitutions, insertions and deletions between all metastrokes in the dictionary.

9 For example, Figure 10 shows a feature correlation table indicating the likelihood of a match occurring between individual metastrokes in a dictionary of only eight metastrokes. Each column entry represents a metastroke found in the input string. Each row entry represents a metastroke found in the dictionary made up of this (limited) metastroke vocabulary. The highest value or measure is assigned to a substitution of an identical input metastroke by a metastroke from the dictionary, ie along the diagonal. Low dimensions are assigned substitutions of input metastrokes by metastatic ones from the dictionary that come close in shape. In the table, invalid substitutions are represented by blanks, and a large negative value is assigned to those positions in the table: (-1000), so that the calculations based on these substitutions always result in a result that is out of the range of view for a complete match ,

Input metastroke additions and deletions are allowed in accordance with the invention, although there is some difficulty associated with each. 10A shows sample drawbacks for the addition of selected metastrokes while 10B Sample drawbacks for deleting the selected metastrokes. Apart. from observing the correlation between metastases (positive values) and aggravations (negative values or aggravated values) associated with added and deleted metastrokes, a size comparison (with respect to the base line of the word) is also performed. 11 shows a size correlation table while the 12A -B Show size scores that match the tables of the 10A -B and refer to additions or deletions.

The word correlation table of 8th for each word of the dictionary can be derived from the data of tables 9 . 10A . 10B . 11 . 12A and 12B derived, which is a simplified example, or from data from the 13 or 14 which corresponds to the actual work system. (In the 13 and 14 are each the 10A and 12A corresponding addition vectors are shown in the headers of the table and the respective 10B and 12B corresponding deletion vectors are shown in the left column of the tables.) The word correlation table of 8th is calculated on the basis of a dynamic programming technique to calculate "optimal cost", as referred to in the art, and on the basis of "a path" for transforming one sequence into another. In the invention, the transformation or mapping is made on the basis of an entire input sequence to bring it out of the dictionary in the length and form of any string of a number of known metastroke strings. Therefore, it is necessary to perform the transformation on the whole sequence before conclusions can be drawn from the results.

13 Figure 12 shows a data table illustrating the transformations of input metastrokes into vocabulary metastrokes. The left margin lists the vocabulary metastrokes, ie the metastrokes found in the dictionary. The top border can be labeled with the same metastrokes in the same order. The diagonal of the table is a value associated with a direct transformation of an input metastroke to the identical metastroke in the dictionary. Non-diagonal values represent values associated with the substitution of an input metastroke for other vocabulary metastrokes in a process in which a metastroke string is constructed that matches a metastroke string found in the dictionary , 14 shows a size correlation table for nine size levels. Complications are assigned for deletions or additions, such as along the zero row or the zero column of 13 is illustrated and in the example according to 10A (this corresponds to zero column of 13 ) and 10B (this corresponds to row zero of 13 ) is shown separately. In a deletion, a metastroke is deleted from the input string. A deletion is indicated by a negative value of a selected size (row zero, any column). With addition, a metastroke is added to a position in the input metastroke string, and the action is indicated by a negative value of a selected size (zero column, any row). A "substitution" (where an input metastroke is substituted by a vocabulary metastroke) is indicated by a positive value, as indicated by the position in the matrix. The special case of a "substitution" of a metastroke by itself is the diagonal and this is therefore the largest value assigned. It is absolutely not necessary for the feature correlation table to be symmetrical since the substitution is not a symmetric process.

The Technique for interpreting the input metastroke string as A word is straightforward, but requires extensive programming. In the dynamic programming attempt, comparison, weighting and modification procedures performed simultaneously. The process corresponds the "snaking" of a "worm" to see if these are aligned to the shape of "pattern worms" can. The additions, deletions and substitutions to that the worm meanders. Each action results in a weighting or an aggravation.

The Steps of the dynamic programming technique include a process in which an entire metastroke string with all relevant metastroke strings in a dictionary of words which are "written" as metastrokes to a path for the transformation the input string into one of the dictionary metastroke strings in which a maximum value or so-called "optimal cost" is determined, around the whole input metastroke string towards everyone To process vocabulary metastroke string. Theoretically you can the entire input metastroke string against all strings in the dictionary are processed. But there are usually certain accelerating Steps executed to the process accelerate. (To the process too speed up, not all dictionary entries need to be checked, although the techniques to limit the process in this way, are not part of this aspect of the invention. These optimization techniques use the features of the dictionary out. The dictionary can e.g. be formed by a number of strokes, or it can be "alphabetized" by metastases. The evaluation can on a forward directed metastroke sequence, on one opposite directed metastrogram sequence, on the best fitting word beginnings and the most appropriate word endings.) As previously described is included the dictionary a "definition", an English-language one Word (or a word in another language) that is in Romance Characters, i. corresponding to the Romanesque characters ASCII strings are written, which is the only one to recognize Word is. Consequently, there are different dictionary entries with the same "definition" to the variations to take into account in the manuscript.

The metastroke feature correlation matrix (ie the data for the transformation of all metastrokes into others) is considered an integral part, such as 13 shows, used in the dynamic programming process, namely in the path building and in the optimal cost calculation. The size correlation matrix is also used in the same process steps. The optimal cost calculation is a calculation of the following three formulas followed by a selection of a maximum based on a simple comparison. The maximum of the three values is inserted into the cell, which is the transformation of input metastrokes into vocabulary metastrokes in the word correlation table (eg 8th ) for each word thus processed from the dictionary. 8th shows an example resulting from the dynamic programming process that extracts the data from the 9 to 12 used.

For illustration is on 8th Referenced. The calculation is performed by selecting the maximum value (α) for the cell in position (a, b) in the word correlation table from α values calculated for addition, deletion and substitution at these cell positions. The maxima are in each example of 8th determined as follows: For the substitutions: α1 j i = α j-1 i-1 + p (a i , b j ) + q (a j , b j ) (1)

For the inserts or additions: α2 j i = α j i-1 + p (a i ) + q (a i ) (2)

For the deletions: α3 j i = α j-1 i + p (b j ) + q (b j ) (3) wherein:
α1 ^j _{i is} the "cost value" (used in the vocabulary of dynamic programming) or the cumulative score in element (i, j) for passing from the origin through element (i-1, j-1) by the input Metastroke "b" in the column position (j) is replaced by the vocabulary metastroke "a" in the row position (i) ( 8th );
p (a _i , b _j ) on (off 13 or 9 obtained) similarity weighting value of the substitution of an input metastroke "a _j " with a vocabulary metastroke "b _j ";
q (a _j , b _j ) of (obtained from the magnitude correlation table, 14 or 12 ) Size weighting for the size substitution that occurs in the prior substitution of the input metastroke "a _i " by the vocabulary metastroke "b _j ";
α2 is the "cost value" or the cumulative score for passing from the origin via element (i-1, j) to element (i, j), by placing the vocabulary metastroke "a" in the line position (i) after the input Metastrich "b" is inserted in column position (j);
α3 is the "cost" for the cumulative score for passing from origin over item (i, j-1) to item (i, j) by deleting the input metastroke "b" in column (j) along a sequence of metastrokes becomes;
p (a _j ) an aggravated value (off 13 , Column 0, line i; or from position i of 10A ) for inserting the vocabulary metastroke "a" in row position (i) after the input metastroke "b" in column position (j);
q (a _i ) is the complication value for the quantity assigned to the previous insertion ( 14 or 12A );
p _(bj) is a penalty value (from 13 , Line 0, column j; or from position j of 10B ) for deleting a metastroke "b _j ";
q (b _j ) is an aggravated value for the quantity assigned to the previous deletion ( 14 or 12B ).

The preferred metastroke string in each cell (i, j) is the maximum cumulative score alpha (α) in cell (i, j) that consists of the first cumulative score (α1), the second cumulative score (α2), and the third cumulative score (α3) is selected. This maximum score in each cell represents the "optimal cost" in that cell along any path from the origin in the word correlation table 8th represents.

A special example can be helpful. Let's look at the element in line i = 4, column j = 4 in 8th pointed. This element contains the maximum value from the calculated alpha thermals α1, α2 or α3 for this position.

To calculate α1, equation [1] is used. From checking the adjacent diagonal cell ( 10 ) results in a value of 7 for α ^j-1 _i-1 8th The substitution shown for cell (4, 4) is from the upward-facing arc (along the header) to a downward-pointing point (along the page column). Out 9 , the substitution from the upward bow to a downward arrow, the value p (a _i , b _j ) equals -1000, which indicates an illegal transfer. The corresponding size substitution from the fourth position of 6B up to the fourth position of 7 is from size 3 to size 5. A transfer from size 3 to size 5 has a value of -1000 according to 11 , The sum of α ^j-1 _i-1 , p and q is thus equal to -1993.

To calculate α2, equation [2] is used for insertion. From the adjacent line above the cell (4, 4) in 8th the result is a value α ^j _i-1 of 11. For the insertion of a downward arrow, the complication value p (a _i ) of 10A equals -2. The corresponding magnitude weighting value q (a _i ) for the downward arrow at level 5 ( 7 ) is after 12A equals 0. Thus the sum α2 and a test value for cell (4, 4) equals 9.

To calculate α3, equation [3] is used for insertion. From the adjacent left of cell (4, 4) lying in column 8th the result for α ^j-1 _{i is} a value of 12. For deleting one after top-directed arc is the complication value p (b _j ) of 10B equals -2. The corresponding magnitude weighting value q (a _i ) for an upward bow in the amount 3 ( 7 ) is after 12B equals 0. Thus, the sum α3 and the test value for cell (4, 4) equals 10. From a simple comparison of α1, α2 and α3, a maximum value of 10 results. The value of 10 is thus stored in cell (4, 4 ) used.

This Procedure is for every cell of the word correlation table is performed. For each cell, the maximum α of the Substitution, addition and deletion used. When the word correlation table is calculated, the Path to each cell from the neighboring cell, giving the maximum rating results, tabulated. At the end of the procedure, the paths of each cell can (e.g., from the right border or lower border) to the origin (0, 0) are returned. Each of these paths represents the optimal path between the origin and the selected one Cell. According to the invention, the cell position is rightmost and at the very bottom of the origin the "optimal Cost value "for the optimal Path through the word correlation matrix for the specific with their Endmark tested and specific input string. The specific input metastroke string becomes any vocabulary metastroke string from the (from the vocabulary of permissible metastroke strings composing) dictionary faced, to create numerous word correlation tables. The optimal Cost value for the input metastroke string (whose value is derived from the cell in the last column and last row yields and the end the input metastroke string and the vocabulary metastroke string corresponds) of each word correlation table is then with all corresponding optimal cost values from the other word correlation tables compared. The maximum of these different optimal cost values is then used to identify the vocabulary metastroke string which best fits the input metastroke string in one word ("Definition") from the dictionary correlates, assuming that the minimum recognition criteria enough becomes. (In other words, a maximum that is below one acceptable minimum, a basis for the display of an error, recognize the input string as a word.)

With the system according to the invention can words, which are written in italics by hand, most likely be recognized. In the system a sufficient redundancy is provided, that with appropriate probability even words recognized can be in a bad handwriting with missing letters and Spelling errors are written.

Claims (14)

Method of recognizing italic handwritten Words with the following procedural steps: Receiving digital point-indicating information that has a starting point and a Include endpoint; Choose the points delimited by the starting point and the end point as a checkword; represent of the test word by a distinctive variety of metastases; Create a list of possible true meta-catchwords using a dictionary of metastroke words and the distinctive multitude of metastases; to compare potential true meta-catchwords from the list of possible true meta-catchwords with the distinctive multitude of metastases; To calculate a word measure based on the comparison; and Issuing a first possible true word if the word size exceeds a threshold. The method of claim 1, wherein calculating the first word measure substituting a second metastroke with a first one Metastrich includes, with the second metastroke being part of the distinctive variety of metastrokes, and calculating the word measure is one Substitution weighting considered. The method of claim 2, wherein calculating the word measure an insertion of a third metastroke into the distinctive multitude of metastases and calculating the word measure is an add-emphasis weight considered. The method of claim 3, wherein calculating the word measure a delete a fourth metastroke of the distinctive variety of Metastrokes, and calculating the word measure includes an erasure weighting weighting considered. The method of claim 1, wherein calculating the first word measure comprises substituting a second metastroke with a first metastroke, the second metastroke being part of the distinctive plurality of metastrokes, and calculating the word measure a substitution penalty weight considered. The method of claim 1, wherein calculating the first word measure an insertion a first metastroke for includes a second metastroke, the second metastroke part is the distinctive variety of metastrophes, and calculating of the word measure considers an addition weighting weighting. The method of claim 1, wherein calculating the first word measure a delete a first metastroke for includes a second metastroke, the second metastroke part is the distinctive variety of metastrophes, and calculating of the word measure an erasure weighting considered. At least one computer-readable medium of one Computer executable Instructions which, when executed, follow the procedure Perform claim 1. System in a computer environment comprising: medium for receiving stroke data with a handwritten input correspond; Metastroke recognizer means for generating Input metastrokes from the stroke data; Means of interpretation the input metastrokes as words, comprising means for comparison the input metastrokes with dictionary metastrokes and assign a value for the probability of a match by comparing the input metastrokes with at least some from the dictionary metastases and Assign a score if a metastroke match is found and using the scores to calculate a value for each Word; and Means to list the words according to their values. The system of claim 9, further comprising means to replace a metastroke with another metastroke and Recalculating the value for a word based on substitution aggravation. The system of claim 9, further comprising means to paste of a metastroke and recalculating the value of a word based on an insertion difficulty. The system of claim 9, further comprising means to delete of a metastroke and recalculating the value of a word based on a deletion aggravation. The system of claim 9, further comprising means to compare each value with a threshold. The system of claim 9, further comprising means to compare a value for a word with a value for another word.