JPS63308690A - Holograph recognition - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

[Detailed description of the invention] [Industrial application field] The present invention relates to a method for recognizing handwritten handwriting.

[Prior Art] Many problems have been raised through many years of research on machine recognition RK of handwritten handwriting. Below, I will briefly describe these problems and how they were resolved.

The basic elements of a typeface (not to mention special symbols) are:
These are letters and numbers, and are hereinafter referred to as "typeface elements." These elements are collected and arranged to form words, and recognition begins with individualization. In printed typefaces, the characters and other symbols are separated (letters), so their boundaries do not pose any problems. In contrast, in connected typefaces, the stroke elements that represent the typeface elements cannot be individuated (this is generally the case for letters, since numbers are usually distinct from each other), and therefore the elements To recognize,
It is necessary to mark the boundaries in advance or to separate the characters in advance.

Once the handwriting has been separated into elements, it is necessary to identify each element by determining which typeface element these elements correspond to.

Various criteria such as size, curvature, inclination, direction, etc. must be used for this identification.

The first attempt was to search for a standard that would enable the "absolute" recognition of typeface elements, that is, a standard that would be valid for any type of handwriting. This attempt is successful only if the writer agrees to follow a very strict writing style, that is, to follow well-formed letters of a certain size and direction.

Efforts have therefore been made to circumvent the above limitations using a training or learning phase. In such cases, in the pre-recognition stage, the writer traces the known text and retains the features of each letter (obtained by successively applying various criteria) for later comparison with other features. use. In fact, this process makes it possible to recognize handwriting. However, it requires a very large number of basic operating steps. This is because comparisons are made by applying all criteria to any handwriting element.

[Means for solving problems]

Therefore, the method of the present invention does not enable real-time recognition of almost any kind of tracking by using equipment of reasonable power, such as a microcomputer.

The present invention relates to a handwriting recognition method that is performed in real time using equipment that does not exceed the power of a microcomputer.

This invention was developed based on the size and shape of handwriting (non-alphape, g,
The present invention relates to a method that enables recognition of all handwriting, regardless of ideograms, specific symbols, etc.).

In detail, the present invention is designed not only to select the most appropriate criterion for the handwriting in the particular case considered, but also to set the most suitable sequence of operation steps (decision tree) for processing the handwriting. , a dR method that utilizes the results of a learning phase, or the results of a step-by-step recognition of handwriting, or both.

The handwriting recognition method according to the present invention includes the following steps.

applying predetermined criteria according to a predetermined sequence of operational steps to the handwriting or to the elements of this handwriting to determine some features of the handwriting or to the elements; and converting the thus determined features into features representative of known typeface elements. and when the comparison of features gives a predetermined result, confirming one element of the handwriting as a known typeface element.

A feature of this recognition method is that it includes the step of setting a predetermined operation step sequence according to the determined characteristics by applying a criterion to handwriting elements.

Advantageously, the step of setting up the sequence of operational steps includes determining the order of the criteria to be applied to the handwriting or handwriting elements. It may also be advantageous for this step of establishing to include selecting only some criteria from among a large number of available criteria.

Advantageously, the selection of features representing known typeface elements, the order of the sequence of operational steps and/or the criteria is updated gradually during different operational steps of handwriting recognition.

The recognition method of the present invention includes a pre-stage of training, in which elements of the handwriting corresponding to known typeface elements such as letters and numbers are applied with some predetermined criteria; Features representative of are determined and retained or stored. This preliminary step of training involves setting up the sequence of operational steps used for recognition purposes.

Preferably, the method of the invention includes encoding the salient points of the handwriting or handwriting elements prior to applying the predetermined criteria. It is also advantageous that the method of the invention includes the step of removing some salient points according to a criterion of proximity.

In this advantageous embodiment, the recognition method of the invention also includes the step of cutting the handwriting into its components. This stage of handwriting cutting includes the steps of determining particular points of possible cuts and evaluating for each particular point the possibility that this point actually constitutes a cut between two handwriting elements. Moreover, it is advantageous that the sequence of operating steps used to determine the probability that a point constitutes a cut between two handwriting elements is set according to the characteristics determined during the training phase.

Preferably, the method of the invention further comprises comparing a dictionary containing the typeface elements with the previously identified element group.

The invention, objects, features, details and advantages may be better understood by reference to the accompanying drawings, which give non-limiting examples only of preferred specific embodiments of the invention.

〔Example〕

The following description essentially includes a training phase that requires a handwritten version of the text known from the machine, and a recognition phase that utilizes the results of the training phase. Additionally, recognition is used to update the results given by the training phase.

However, this update is not necessarily necessary if the operator's handwriting changes only slightly over a long period of time. Additionally, previous training may be omitted or waived. , in this way recognition is not very effective at first. That is, the machine used must receive a lot of assistance from the operator (location of cutting characters, nature of the characters, etc.). This interactivity is very useful within the scope of the invention both during the training phase and also during the recognition phase.

In the considered embodiment, the recognition method according to the invention comprises two
a training phase relating to known typefaces and a phase of recognition of unknown typefaces, preferably with updating of the training results during a kind of "continuous training" period. uR
The stages must be considered first to understand the process and criteria used and the features that should be known, and then the training stages must be considered to understand what primitive features are obtained. No.

Although the following description relates to the specific case of a - set of criteria, other sets of criteria may be used. Accordingly, the following description is merely an example and does not limit the scope of the invention in any way.

Recognition begins with the application of predetermined criteria.

However, the handwriting to be identified is demarcated and separated into individual handwriting elements. This delimitation is performed by first encoding the handwriting into salient points by applying a first criterion.

This criterion is effective in determining the end points of handwriting, i.e., high points,
The determination of the low point, the right edge, the left edge, the rise of the writing instrument (this is relevant in the case of dynamic acquisition of handwriting).

FIG. 1 is an enlarged view of a portion of handwriting that includes four characters belonging to two words (two characters are connected and the other two characters are not connected).

The lower part of the figure shows how to interpret the salient points of the handwriting. These notable points are high points (■4);
It includes the low point (L), the left edge (G), the right edge (D), and the rising edge of the writing instrument (L). Therefore, in the first stage, this handwriting is HB HB HB HB D HL -BL-L-
It is encoded in the format BDHL-GBDHL.

First, this code is filtered to remove points that are too close together in the main reference directions X and Y by applying an I threshold. In the particular case of Figure 1, the points that are removed are points 12 (too close to point 16), points 15. ). Note that although point 7 is close to points 6 and 8, it is not removed.

Point 7 must not be removed because it corresponds to a very important feature, namely a reversal or a goal conceded. Therefore, the new code is HB [-1HHB I-T B D HL -L
-L -DHL-BDL.

It should be pointed out that the above process, unlike other criteria such as the determination of curvature, has a very short calculation time. Therefore, since it has the ability to discriminate against all types of handwriting, it is desirable to employ this process in a systematic manner.

As already mentioned, before recognition a cutting, ie cutting the handwriting into individual handwriting elements (each handwriting element corresponds to one character) must be carried out.

Advantageously, this cutting involves two stages: a first step in which all possible cut points are determined, and a second step in which the probability that the character actually constitutes such a cut point is evaluated for each cut point. It has stages.

Figure 2 shows a word with all possible cut points indicated by small black squares. These points will define "segments" between them. The evaluation of said possibilities consists of calculating the distance or spacing between the segment in question and its previous segment and determining the relative position according to this distance.
and calculating whether there is a stroke. These calculations allow it to be determined whether there are character breaks, word breaks or accents or return to lines in cases below 95. The remaining five cases are handled by extending the concept of characters to character strings during the training period. Of course, other criteria can be used to determine such a cut.

The standards thus defined or established for character recognition can vary widely, many of which will be dealt with5.

The first group of criteria concerns the size of the bounded element. Some of these criteria are shown in Figure 6. In the figure, "n" is an HB H shape of equal height (note that the salient points that actually represent the cutting points have been removed). rhJ is the reduced form of HBH, “ruL”
is the incremental form I-I B M. The loop is fully open, open,
It can be either closed, right open, left open, right closed,
It can be any degree of difference from the left opening.

The second group of criteria relates to size. The handwriting elements are extremely dog (handwritten f), large (1), and small (a), and the sentences are extremely small (1).
). These criteria can be applied to any element or only one segment of it (between two salient points).

Similarly, a handwriting element can be very wide, wide, narrow, or extremely narrow.

The sixth group of criteria is curvature and includes clockwise curvature, counterclockwise curvature, and reflection or loss points. This sixth group also includes the absence of curvature (straight segments) and the presence of cuts.

The fourth group of standards is square. The square is high (letter "1")
, either low (letter "m") or square (letter "a"). These square criteria apply to all or only some of the handwriting elements.

The fifth group of criteria is the relative position of the salient points. This group includes return, immobility, advance, rapid advance, decrease, equal, increase,
Near, far, etc. characteristics can be determined.

The sixth group of criteria is related to direction and gives the following characteristics: high, stone high, right, low right, low, current, left, and high. It also gives the general direction between two falls, between two rises, between a fall and a rise, and between a rise and a fall.

Additionally, the direction may be determined for one segment or for all first or last strokes or lines.

The seventh group of criteria is positioning, which concerns the determination of the position of the handwriting relative to a salient line (baseline, distance between two lines). These criteria allow the determination of intermediate areas, accents and commas, and the presence of upward and downward strokes.

An eighth possible group of criteria is the treatment of points removed during the first filter link of the code.

The ninth group of criteria is the end point symbol, which indicates how to determine the ends of a handwriting element (falling, rising, cutting).

Other criteria and other groups of criteria can also be used if they have discriminative capabilities.

After determining the features of the handwriting elements, these features constitute the features of the handwriting elements determined during the training period.

However, all these criteria require too many buttons to be used in real time during recognition. Attempts have therefore been made to arrange the criteria by organizing them according to a decision tree, the first rubel of which is the most discriminative criterion. Nevertheless, since this organization was to be appropriate for any type of handwriting to be identified, the decision tree must have all the criteria and must have a very large number of levels. There wasn't. Real-time recognition could not be achieved in this way without extremely powerful equipment.

However, only a few criteria are generally sufficient for recognition. To this end, consider the example of the French word ``addresseJ'' illustrated in Figure 4. In the same figure, the black squares indicate possible cutting points. or a handwriting element extending between the start of the handwriting and the first possible cutting point. The first operating step is to determine the salient points according to a purely automatic process.

Next, there are the operational steps of partial recognition, cutting into characters according to this partial recognition (i.e. collecting some segments into one character) and performing final recognition according to cutting into complete characters. This is sometimes done by making use of lexicographical rules or actual dictionaries.

In other words, the French word [adre
In the particular case sseJ, the first criterion is “H
GBDJ (encoding already explained with reference to Figure 1)
It is. The first segment with the code 1-IGB is either C or part of the letters A,Q. The second segment is the code B I-I B and the I scarecrow or letters 1-1 . Represents a part of F. The connection between the first segments is not Q (in case of a connection, it is simply D or A. The sixth segment has the same sign t as the first segment.
r G [3, and therefore C or letters, A, Q
is part of either. The fourth segment is coded BD
It has BGB and represents only L or letters, part of A (combined with the previous segment) or part of H (combined with the next segment). The segment that follows has the symbol BHB. This code can represent various characters alone or together with subsequent segments, but
Combined with subsequent segments to form H. Similar considerations apply to subsequent segments.

Therefore, one standard (limited to the beginning of a word) is CLCL, CICH, and CID.

CIA, ACL, ACH, AD, AA, DCL, DCH
.

It must be pointed out that a combination of DD and DA is available.

Therefore, the "increase-same-decrease" criterion (the first group of criteria mentioned above) is used. This criterion applied to the first two groups of segments gives an "equal or reduced" result, and the third
and when applied to the fourth group of segments gives an "increase" result. Therefore, the fourth and eighth to twelfth possible solutions given by the first criterion are eliminated. Only 6 possible combinations remain.

Another criterion then allows for additional removal, and a comparison with the dictionary removes the final uncertainty.

It should be noted that a very small number of criteria gives a very wide selectivity. Moreover, these criteria are extremely easy to obtain and do not require lengthy mechanical processing (calculations). Real-time recognition of handwriting is made possible because these criteria are selected that are allowed by the configuration of the sequence of operation steps for specific recognition of the handwriting, and are small in number.

The determination of the criteria to be used, the construction of the sequence of operating steps and the determination of the characteristics of the handwriting elements are provided by training. That is, the training phase is performed according to the invention in order to select the most discriminative criterion in the particular case of the handwriting to be processed, as well as the fastest (most effective recognition) criterion for this particular case. Judgment tree (basic operation step sequence) that gives
used to set.

The stages of training will now be considered. The purpose of this stage is the selection and organization of criteria. Therefore, a very large number (perhaps the largest number of criteria) is utilized.

The validity of each criterion is then determined and a criterion selection is made. Also, criteria that give highly unstable results are avoided.

Training consists of the acquisition of a known text, the cutting of this text into characters and the determination of features by applying criteria (measurements) (i.e. the operational steps of processing the values of the features, calculating the probability of recognition and constructing a decision tree). column settings).

The first stage (or learning) involves the user copying the text fragment by means such as a graphical table connected to a microcomputer. Preferably, the text consists of the set of symbols used (e.g. letters,
It is best to include all symbols (numbers, alphabetic punctuation marks, correction marks, etc.). Of course, the more the text samples represent recognition, the more effective the training will be.

The next stage is cutting. The cutting into segments is carried out entirely by a machine that does not require operator assistance due to its knowledge of the text.

However, the selection of cut points (character delimitation) involves the operator's responsibility for the results obtained. This is because it is the operator who identifies and ultimately determines the handwriting as one or several characters. Following cutting, one handwriting element is associated with one symbol (letter, number, etc.).

Training will be speeded up if the learning equipment (a microcomputer connected to a graphical table) itself indicates where to make the cuts. The operator merely has to enable a cut between characters or, conversely, indicate where the cut should be made.

The next step is the measurement of the features by applying the criteria mentioned above. The measurement involves the principle of performing a complete handwriting analysis according to all criteria. All characters associated with the handwriting after cutting are associated with a list of values (list of reference values). This list of values is called the measurement element that makes up the set of features.

The table illustrated in FIG. 5 shows a set of letters in overall structure with different variants of each letter, and associates its measurement elements with the respective handwriting.

This table is a simple example to help you understand the organization of the structure. According to this table, analysis of known texts reveals five different variants of letters a and 1, letters ◇ and C
It is possible to identify four variations of the letter d and three variations of the letter d. Each letter and its variants are identified by the numbers that make up the code in the first column. Columns I, II, ■ each show the values obtained by applying one criterion. In other words, column 2 shows the values forming the result of applying the reference HBGD (as defined in connection with FIG. 1). Column ■【 shows the result of applying the size standard to each character. This criterion allows for four distinctions, represented by four codes 1 to 4, meaning "extremely small,""small,""dog," and "extremely large," respectively. Column ■ represents relative size, 0 to 4 representing "uncalculable,""decrease,""equivalent," and "increase," respectively.
04 diacritics with 1 decrease-equal-increase” criterion.

Although this table only shows three columns of criteria, it must be understood that there are as many columns as criteria. After setting up the overall structure of the characters and their measurement elements, they are subjected to self-help processing. This will be explained in Figures 6 to 10, but it is possible to design a decision tree to be used during recognition of text written by a person whose handwriting is new to the device and whose handwriting allows the setting of the table shown in Figure 5. become.

Self-help processing begins by applying the criterion HB GD that has proven to be the most discriminative to all characters. Column I contains 5 encoded distinguishing elements, and application of this criterion yields 5 subsets SE1 to 5E5 (
There is one subset for each distinguishing element. (see Figure 6) are set.

In this table, the distinguishing element B I-1 appears only once,
Indicates the character 4, which is a type of a. Distinguishing element D I-I
Q is code number 10, 11, 12, 13, 17°18.
19,20.21 is included in the measurement element of the character. Therefore, the characters e and the characters representing 1 will be included in subset SE2.

Similarly, a subset SE6 containing the letters 6 and 8 which is the letter d, a subset SE4 containing the letters having the code number 7 and which is the letter ji, code numbers 1.2, 3゜5.14.15
．． 16 and representing the letters a and d can be obtained.

Since the characters SEL, SEL, SF6, and SF3 contain only one character, the recognition process for the corresponding characters is complete. However, Sapse, ToSE
2, SF3 contains different characters. That recognition requires a more complete analysis.

The latter consists of applying to these subsets of all the criteria already mentioned, excluding the criterion HB GD.

FIG. 7 shows how criterion II is applied to subset S E 2. Since this criterion contains four distinct possibilities, the structure representing the data processing of subset SE2 has four branches, each terminating in a subset SE2, denoted by 5SE1-4. code 1
The subset 5S given by the distinguishability with
E1 includes the handwriting of characters having the symbols 10 and 11, which is the letter e.

The subsera given by the sign 20 distinguishing element or distinguishability) SSF2 also extends to the sign which is the letter e, 13
has a character with . Subsets 5SE3 and 5SE4
are respectively character 1, character 17, 19 and character ! It has the character 18°20.21. Calculation of likelihood determines the validity of the criterion. Therefore, K, each subset has a majority 1ett
er) must remain. The number of such characters is calculated and divided by the total number of characters of this form in the table. The values determined for the subsets 5SE1 to 5SE4 are then added. In the case of FIG. 7, each subset only contains characters representing one character. This character is the main character.

That is, the subset s s Ei contains two characters C, and since the total number of characters C in the table is 4, it contains a coefficient with probability 2/4. Subsets 5SI32 to SSIJ4 each give a coefficient of probability 2/4.215.315. The sum of these four coefficients is 2 and constitutes the value that determines the validity of the criterion I[.

FIG. 8 represents the application of criterion 2 with four distinctness possibilities 0, 1, 2.3 to subset 5E2 (therefore resulting in a structure with four branches). Since in the table the distinguishability 0.1.2.3 is not associated with the handwriting or characters of subset SE2, these distinguishability do not result in the setting of subset SE2. On the other hand, distinguishability 2
yields a subset 5SE6 containing all the characters of SF2 (ie 4 characters e and 5 characters 1). letter e and 1
The distinguishability coefficient of the 0 distribution is 4/4°515, respectively. The sum of the distinguishability coefficients is equal to one.

This factor indicating criterion validity is smaller than the factor determined using Criterion II. Therefore, the latter is considered the most discriminative criterion and is retained for setting up the decision tree shown in FIG.

FIGS. 9 and 10 show the processing of standard B I-(subsera including application of all standards other than GD) SF3. Since this example is limited to criteria II and 1■, only these two criteria have to be retained for the processing of subset SE2. Figure 9 is the standard ■【
The result of applying to subset SE5 is given. 4 subsets) SSE1 to 5SE5 are found. Supplement) SS
El contains characters 1 and 2, all of which are the characters a. The distinguishability factor is 215. Subset l-8SE2 includes letters 6 and 5, which are the letter a, and a letter 14, which is the letter d, and by calculating their distinguishability, a coefficient of 215 is given to the letter a, and a coefficient of 1 is given to the letter d. /6 is given.

Therefore, the letter a is the main letter and the coefficient 215 must be retained for the calculation of the validity of the criterion I. Since SSE3 and 5SE4 each contain only one letter d, the coefficient of distinguishability is 1/3.1
/6.

Adding all the coefficients of the main characters, the discrimination factor is 22/1
5 is found.

FIG. 10 shows the results of using criterion 11 to identify the characters included in subset SE5. To the extent that the criterion ■ has four distinct possibilities, four subsets must be found.

The SE50 character has a distinguishability of 0 and 1 in its measurement elements, i.e. in all of these feature sets.
Since we do not show the symbols representing , only distinguishability 2 and 6 will each give one subset. Subset) SF3 is the letter a, four letters 1.2.3.5
Therefore, the distinguishability coefficient 415 is determined. Subset 5SE4 proves character 14 to be character d,
It has 15°16. Since the total number of letters d is 6, the distinguishability factor is 6/3. Adding both distinctness coefficients yields the value 915 (=27/15).

Comparing the factors showing the effectiveness of discrimination of criteria ■ and ■ applied to subset SE5, factor 27/15
Criterion ■ with a factor of 22/15 is better than criterion II with a factor of 22/15.
It turns out that it is more effective than

Therefore, in the decision tree of FIG. 6, the criterion ■ is held.

The detailed description given so far illustrates the principle of self-help processing that leads to the construction of a decision tree that retains only the most discriminating characters and determines the criteria for which the criteria are determined. Used at each node or junction. If the code HB GD is used as the first criterion, g200 branches are found for normal handwriting. Some branches end up in subsets that form leaves of the tree, where some characters are classified and separated from other characters. The training process ends with the construction of a decision tree, and the settings of a given handwriting feature file are represented in the form of a flowchart in FIG. In the first operating step, a person whose handwriting is to be identified by the device writes a known text on a support (e.g. a table to be digitized). There is.

Tables, which are themselves well known, generate from the handwriting a series of electrical signals, each representing one point of the handwriting, and send these electrical signals to a handwriting processing device, such as a microcomputer. The microcomputer is programmed to master, i.e. to cut the text into elements indicated by the letter names as described above, to set up the measuring elements for the various letters and to write them in the table according to FIG. Create an organized structure in the manner shown. The microcomputer then performs a self-help process of generating a decision tree by applying criteria.

FIG. 12 shows in a flowchart the process of recognizing text that is unknown to the device and written by someone who has already configured the device with a feature file containing a decision tree. In this example, the letter d is written on the table to be digitized. The microcomputer is programmed to learn this character in the same way as before. The microcomputer is then programmed to carry out operational steps suitable for the recognition of the letter d by reference to a decision tree created for the person whose handwriting is to be identified. Criterion 1 applied to the handwriting of the letter d to be identified (B According to the decision tree, criterion II should be applied.If we apply criterion I[I to this same handwriting and the code 3 is given as a result, then the identified The microcomputer knows that the characters should be in the subset 5SE4.This subset has the symbols 14 and 15.
It contains a letter with ゜16, which actually indicates the letter d according to the table in Figure 5. In other words, according to the present invention, a microcomputer is able to identify the letter d by applying just two of the extremely diverse criteria that have been maintained for establishing the measurement elements of each letter. It will be possible. It should be appreciated that the present invention enables real-time recognition using standard microcomputers (eg, of the type commonly employed in text processing equipment).

If the handwriting that created the composition of the decision tree undergoes a modification, the device of the invention simultaneously recognizes, updates the measurement elements of the modified character and, if necessary, updates the decision tree.

Another useful feature of the invention is the possibility of identifying words without necessarily identifying each letter by comparing groups of handwriting elements that form a word into a dictionary of word forms, or how many handwritings there are. In this way it is possible to eliminate the uncertainty arising from the fact that it is properly cut. According to the invention, these problems are solved by a structure called a distribution grid or pattern, and by an ordered dictionary of specific structures. FIG. 16 shows the French word [admise J] with the cut made by the microcomputer and the grid or pattern set by the microcomputer below the letter C). This grid or pattern comes from the French word 1” admi se
It shows various possibilities for recognizing handwriting segments of J. However, in this example, only the most likely characters of each distribution are shown.

Since the letters "d" and rhJ have one segment in common, both letters are mutually exclusive.

Uncertainty is removed by reference to a specific dictionary, the principle of which is shown in FIG. 14.

It has a dendritic structure, with one branch for each letter of the alphabet, for example. All these branches start from a common point.

Each branch has a node or junction, from which other branches arise, and each branch is a single letter that forms with the first branch a letter combination that seems to have some meaning. associated with. Each of these twigs forms the start of a new branch associated with a single letter, forming a new meaningful letter combination. This structure also allows to take into account various grammatical forms of words. For example, a single dendritic structure containing 80,000 words and considering different grammatical forms of the words, i.e. 200,000 proper letter combinations, is extremely unsatisfactory for real-time recognition of handwriting. The structure requires, for example, 180 kilo octets of storage space.

By setting up a decision tree during the training process and, if necessary, by the presence of a dictionary organized in a dendritic structure, the method of the invention can Real-time recognition of handwriting is possible using standard equipment that requires only For example, a microcomputer incorporating a central memory of 512 oct. and increasing the storage space suitable for a dictionary is extremely satisfactory.

The method of the present invention has many applications. Namely, it constitutes a learning device for a microcomputer, and together with a rotary control member now called a "mouse", a keyboard (taple, etc.) is used.
Direct writing of text, instructions, etc. on the board). It also enables direct typewriting of handwritten text by a person whose handwriting has been previously analyzed, scanned and learned during the training process to construct the decision tree.

[Brief explanation of the drawing]

FIG. 1 shows some characters and how the encoding according to the invention is performed. FIG. 2 shows a word and how to cut the word into letters. Figure 3 shows some characters and how the encoding is interpreted. FIG. 4 shows the handwritten words used to describe the operational steps performed during the recognition period. Figure 5 shows
3 is a table showing the analysis results of handwritten text according to the present invention. FIG. 6 shows a decision tree set from the table of FIG. 7, 8, 9 and 10 illustrate the operational steps for setting up a decision tree according to the invention. FIG. 11 shows a flowchart for setting up a decision tree according to the invention. FIG. 12 shows a flowchart of a single character recognition process according to the present invention. FIG. 16 illustrates the principle of a distributed grid or pattern according to the invention. FIG. 14 shows the organization of an ordered dictionary according to the present invention. HBHBHBHBBDHLBL LBDHLGBDH
L no 2 Ko r3 φ11 2 Scream+2 2 10
...+r++12
221:l
INK+
2”
G11ll 21! i 128
113 16
0ml 41f
lN1G 318
INI
G4IQ
INIG320
l decoy
4J Engineering qS S5 Ward 1 + 醋ノ♀逼41 衷41L” (10)
II! W,'/SSM:l l1
fflE4J Ex-Kai

Claims (10)

[Claims] (1) applying a predetermined criterion based on a predetermined sequence of operation steps to handwritten strokes or handwriting elements to determine features of the handwriting or handwriting elements; and converting the thus determined features into features representing known typeface elements; A method for recognizing handwriting, comprising: comparing; and identifying the handwriting element as a known writing element when the comparison between the features yields a predetermined result; 1. A recognition method comprising the step of setting a predetermined sequence of operation steps according to the characteristics determined by the application. (2) In the recognition method described in claim 1,
A recognition method, characterized in that the step of setting the predetermined sequence of operation steps includes determining the order in which criteria are to be applied to the handwriting or handwriting elements. (3) In the recognition method described in claim 2,
A recognition method characterized in that the step of setting includes selecting only a small number of criteria from among a large number of criteria. (4) In the recognition method described in claim 1,
A recognition method further comprising the step of gradually updating features representing known typeface elements, the order of the sequence of operational steps, or the selected criteria. (5) In the recognition method described in claim 1,
Known typeface elements, such as letters, numbers, etc., are subjected to the application of predetermined criteria to determine or retain the characteristics representing these known typeface elements, with a training pre-step, which is used for recognition and A recognition method comprising setting an operation step sequence including criteria with discriminative power. (6) In the recognition method described in claim 1,
A recognition method characterized by encoding salient points of handwriting and handwriting elements. (7) In the recognition method described in claim 6,
A recognition method comprising the step of removing some salient points according to a criterion of proximity. (8) In the recognition method described in claim 1,
A recognition method comprising the step of cutting handwriting into handwriting elements. (9) In the recognition method described in claim 8,
The step of cutting the handwriting involves discrimination of specific points of possible cutting;
and, for each particular point, an evaluation of the possibility that the point actually constitutes a cut between two elements of the handwriting. (10) A recognition method according to claim 1, characterized in that the recognition method includes a comparison between a group of previously recognized elements and a dictionary containing the group.