JPH0981690A - Handwritten character on-line recognition device and character style registering and learning method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To highly efficiently recognize characters with a small dictionary registration number even for the peculiar form and abbreviated forms, etc., of individuals. SOLUTION: When dictionary registration is performed for an input character pattern 'GO' including the peculiar form or the like in a part (left-hand side of character, for instance) the input character pattern is compared with a standard pattern corresponding to it, a partial pattern (left-hand side of character) unsmilar to a radical pattern for constituting the standard pattern is detected, the partial pattern is automatically divided from the input character pattern and the partial pattern is registered as the pattern of a partinent radical. When 'HANASHI' whose left-hand side part is turned to the similar peculiar form is written and inputted, recognition rate is improved when the standard pattern matched with it is constituted by combining the 'left-hand side' of the peculiar form and the standard pattern 'SHITA'.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a handwritten character online recognition device for recognizing characters handwritten on a tablet online, and particularly to recognize handwritten characters different from standard typefaces with high efficiency from a small user registration dictionary. The present invention relates to a suitable font registration / learning method.

[0002]

2. Description of the Related Art In a device for recognizing a character pattern handwritten from a tablet, a matching process is performed on an input handwritten character using a dictionary of standard fonts registered in advance in the device. If the handwritten input characters are written in the right stroke order and carefully in a regular style like a standard typeface, the recognition rate will be high. However, if the stroke order of the handwritten input characters is incorrect, the order of the feature point sequences of the input characters will be significantly different from the order of the feature point sequences of the standard font, and the recognition rate will be low.

Therefore, for example, Proceedings of the 50th National Convention of the Information Processing Society of Japan (the first half of 1995), page 2-61, 4D-
4 "High-speed online handwritten character recognition method that does not depend on the number of strokes" As described in Lei Van Too et al. (Tokyo University of Agriculture and Technology), a technology that uses structured font expression in combination has been proposed. This structured font representation is shown in FIG. 12 and FIG.
This will be described with reference to FIG.

In the example shown in FIG. 12, the standard font pattern "light" is obtained from the pattern "day" and the pattern "month" by the following steps (1) to (4), and this is input. I try to compare it with the letters. (1) Convert the standard pattern of the component "day" into the position (in this case, the biased position) and size of the first part of the character "bright". (2) Convert the standard pattern of the component "month" into the position (in this case, the position of the straw) and the size of the second part of the character "bright". (3) Set the end point E of the standard pattern converted in (1) above to (2)
It connects to the starting point S of the standard pattern converted by. (4) Convert the created pattern to the size of one character.

As described above, the character A (“bright” in the above example) is composed of the component B (“day” in the above example) and the component C (“month” in the above example). When the number of patterns registered as standard patterns (the number of patterns with different stroke order or handwriting) is k, the number registered as component B is m, and the number registered as component C is n,
The total number of standard patterns to be compared with the input characters is m × n + k
Becomes

FIG. 13 is a diagram showing an example of standard pattern formation of "cat". If there are 3 standard patterns of beast bias, 3 standard patterns of canopy, 2 standard patterns of "rice field", and 1 standard pattern is registered as "cat", The total number of standard patterns of “cat” is 3 × 3 × 2 + 1 = 19.

[0007]

By adopting the above-described structured font expression, all characters (Kanji characters have the first character
It is possible to improve the recognition rate of handwritten characters without user registration of a total of 6300 surplus characters of the level and the second level. However, personal habits and abbreviations are written in single-stroke or double-stroke, etc., so even if the above structured font expression is adopted, the underlying dictionary pattern will be written in regular writing. Then you will not be able to recognize it at all. In addition, different characters cannot be recognized. In order to be able to recognize such abbreviations and variant characters with high efficiency,
There is a problem in that it is necessary to register individual habit characters for all the characters, and a great deal of time is required for registration work.

It is an object of the present invention to provide a font registration / learning method and a handwritten character online recognition device which enable highly efficient recognition with a small number of dictionary registrations for individual habit characters and abbreviations.

[0009]

[Means for Solving the Problems] The above object is to provide a standard pattern composed of a combination of various radical patterns registered and an input character pattern acquired by using a tablet which outputs handwriting as a coordinate data string. In a handwritten character online recognition device that performs matching processing to recognize the input character pattern, when a dictionary registration is performed for the input character pattern, the input character pattern is compared with the standard pattern corresponding thereto, By automatically dividing a partial pattern that is not similar to the radical pattern forming the standard pattern from the input character pattern and registering the partial pattern as the pattern of the corresponding radical,
Achieved.

When a person makes a peculiar writing on a part of a character such as a radical, in many cases, a person systematically makes the peculiar writing on all characters including the radical as a part of the character shape. . If the part where the peculiar handwriting is performed is specified by a certain character and the specific part can be registered, the labor of character registration will be greatly reduced by adopting the structured font expression described above.

Therefore, when recognizing a written character based on the character pattern data acquired using a tablet that outputs a handwriting as a coordinate data string, when a character pattern different from the standard font is registered, regarding this registered character pattern, Further, the following processing is performed.

The registered character pattern (for example, a character pattern in which the word bias is abbreviated with the word "word") is collated with the dictionary pattern, and the matching partial pattern ("word""go"). ) Is determined. If there is a matching partial pattern, register the remaining parts as corresponding (if the "go" part matches in the above example,
Register the remaining abbreviation pattern and the word as matching.) Then, this abbreviation part is used at the time of recognition processing of other characters including the same abbreviation part (characters such as "detailed" and "theory") by the structured font representation.

[0013]

BEST MODE FOR CARRYING OUT THE INVENTION An embodiment of the present invention will be described below with reference to the drawings. FIG. 3 is a configuration diagram showing a font registration function for character recognition in the handwritten character online recognition device according to an embodiment of the present invention. This character online recognition device registers (learns) without dividing the input character pattern.
Pattern whole learning unit 1, an automatic division learning unit 2 that automatically divides an input character pattern and registers (learns) a portion to be registered (learned), a structured font expression dictionary 3, and an input character pattern It is determined whether or not there is only radicals. If only basic radicals (= cannot be divided), the input character pattern is not passed to the overall pattern learning unit 1. If not, the input character pattern is automatically divided into learning units 2. Determination unit 4
Then, the automatic division learning unit 2 fails to automatically divide the input character, and a division request is output to the application (the application requests the user to divide and input the automatic division failure character by this division request). And outputs a guidance for prompting the user to input the divided character. Then, the input character learning unit 5 learns the input divided characters and registers them in the structured font expression dictionary 3.

FIG. 4 is a flowchart showing an automatic division learning processing procedure performed by the automatic division learning unit 2. The definitions of variables and symbols used in this Figure 4 are: In_ptt: Normalized input pattern B, C: Character (registered character) that you want to output as a recognition result (registered character) A component Cat (A): Character or Category pattern set of radical A (set of standard patterns that represent A) Des_Score: Similarity between input pattern and misrecognition result when misrecognized MaxScore (In_ptt, Cat (A)): Input pattern and Cat (A) Is the maximum value of the similarity with each of.

In the processing procedure of FIG. 4, first, whether or not a portion similar to the standard pattern of one constituent element B of the registered character A (consisting of constituent elements B and C) can be detected from the input pattern In_Ptt. It is determined (step 1). If this determination result is negative (not similar), the process proceeds to step 5 described later, and if a part similar to the component B can be detected, the remaining part is cut out and registered as the component C (step 2). And by this registration,
The maximum similarity between the input pattern and the registered character category pattern set is MaxScore (In_Ptt, Cat (A)).
Des_Score (If a new standard pattern is registered for an input pattern, if the similarity between the input pattern and the new standard pattern does not exceed the similarity with the standard pattern that was misrecognized before registration, However, the one with the highest similarity still becomes the standard pattern that was erroneously recognized, so it is necessary to exceed it,
Des_Score is the target similarity. ) It is determined whether it becomes larger (step 3). This judgment result is affirmative (y
If it is es), the learning is terminated, and if negative (No), the registration in step 2 is canceled (step 4), step 5
Proceed to.

In step 5 and subsequent steps, the same processing as in steps 1 to 4 described above is performed on the other constituent element C, and division is not possible (if the judgment result in step 5 is negative or if step 8 is reached). ), A division request is output to the application because the automatic division learning has failed.

FIG. 5 is a flow chart showing a detailed procedure of step 1 of FIG. 4 (processing for detecting a part similar to the component B). In the process of FIG. 5, the process is compared with the category pattern set of the component B, the most similar partial pattern of the input patterns is detected, and a process of returning a delimiter point that separates this partial pattern from the remaining part is performed. This Figure 5
The definitions of variables and symbols described in each step of the flowchart of are as follows.

When the j-th writing point (same as the feature point obtained by the known technique) of the stroke (one stroke drawn from the time when the pen is lowered) is the starting point Is_Start (j) = 1 When the ending point Is_End (j) = 1 Part (X: s to e): Part of pattern X from the s-th point to the e-th point Pattern Points (X): Number of writing points of pattern X Points1 (X): First part of pattern X Number of writing points Points2 (X): Number of writing points in the second part of pattern X n: Number of standard patterns in the category pattern set of component B i: Index PB (i): B indicating the standard pattern in the category pattern set of B -Th standard pattern in the category pattern set of Cut_start = 2: end point when the number of strokes in the first part is the minimum Cut_end = Points (In_Ptt) -3: end point when the number of strokes in the first part is the maximum Cut_run : Variable for searching the end point of the first part Cut_candidate: End point candidate of the first part Part_score: Similarity between the first part of the input pattern and the standard pattern of B category Part_maxscore: Maximum value of Part_score OK_score: Similarity threshold for determining whether two patterns are similar or more is the definition .

In the processing of FIG. 5, the input pattern is divided into two parts, and the first part is the component B.
Returns its breakpoint, assuming that it looks like some standard pattern of. For all the standard patterns PB (i) of the category pattern set of the component B, the end point (break point) of the first part of the input pattern is the second writing point (feature point) (in this embodiment, Since the starting point is 0th,
2nd is the 3rd writing point counting from the starting point) to Points
(In_ptt) -The partial pattern is cut out as being between the 3rd point (similarly from the last point of the writing point to the 2nd point), matched with PB (i), and the standard pattern with the highest similarity is obtained. Determine the breakpoints in the input pattern and return the breakpoints.

The first part is the part from the start point of the input pattern to the break point, and the part from the next point of the break point (if it is the end point of the stroke, then the next point) to the end point of the input pattern. Return as the second part. However,
This similarity is “0” or too small (Threshold OK_sc
If it is smaller than ore), a similar standard pattern is returned as not found.

The detailed procedure of FIG. 5 will be described below. First, in step 10, initialization is performed by setting the values of the variable i and the variable Part_maxscore to “0”, and in the next step 11, it is determined whether or not the value of the variable i is smaller than the standard pattern number n of the component B. To do. The value of the variable i is set in step 18 described later.
Is incremented by 1, and the following process is repeated by the number n of standard patterns. Since i = 0 at first, the determination result of step 11 is Yes and the process proceeds to step 12. In step 12, a variable Cut_ for searching for a break point
The value of the variable Cut_start as the value of run (in this case, "2")
And enter step 13. The value of the variable Cut_run is incremented by "1" in step 17 described later.

In step 13, it is judged whether or not the value of the variable Cut_run is less than or equal to the value of the variable Cut_end. In the present embodiment, in order to check which of the writing points (feature points) of the input pattern will be the break point, one point is assumed to be the breaking point from the starting point side to the end point side on the writing point of the input pattern. It shifts and compares the partial pattern before the break point with the standard pattern to obtain the degree of similarity. Therefore, in step 13, the assumed break point is the end point of the first part (the writing point of the second part of the input pattern is at least 2).
Since there are dots, the maximum end point of the first portion is two points before the end point of the input pattern. ) Is reached.

If the result of the determination in step 13 is Yes, the process proceeds to the next step 14 and the partial pattern from the start point of the input pattern to the writing point of the current value of Cut_run,
The degree of similarity of the component B with the i-th standard pattern is calculated. Then, in the next step 15, this similarity is a variable Part_maxscore indicating the maximum value of the similarity scores to date.
Is larger than the value of. If the degree of similarity becomes large as a result of this determination, the process proceeds to step 16 and the current value of Cut_run is set as a variable Cut_can indicating the end point candidate of the break point.
The value of didate and the value of the similarity Part_score obtained in step 14 are input as the value of the variable Part_maxscore. Then, the process proceeds to step 17, the value of Cut_run is incremented by 1, and the process returns to step 13. If the degree of similarity is low in the determination in step 15, the process skips step 16 and proceeds to step 17.

When the value of Cut_run reaches the end point of the first part by repeating the processing from step 13 to step 17, the process proceeds from step 13 to step 18 to increment the value of the variable i and make it the standard to be compared. After changing the pattern, the process proceeds to steps 11 and 12 and step 13
The processes from to 17 are repeated.

When the processing is completed for all the numbers of standard patterns to be compared, the determination result in step 11 is negative (No), and the process proceeds to step 19. This step 19
Then, it is determined whether the value of the variable Part_maxscore is larger than the value of OK_score (threshold value), and if it is larger than the threshold value, it is assumed that a partial pattern similar to the standard pattern of the component B is found, and the threshold value is set. When it is smaller than the value, it is determined that a partial pattern similar to the standard pattern was not found.

The above is the detailed procedure of step 1 in FIG. Since the detailed procedure of step 5 of FIG. 4 is basically the same as the processing procedure of FIG. 5, detailed description thereof will be omitted. In this step 5, the second part (straw part) similar to the standard pattern (component C) is detected. This step 5
Now, with the part from the assumed break point to the end point as the second part, the matching with the standard pattern of category C is repeated to search for the break point. And if successful,
The part from the start point of the input pattern to the previous point of the break point (if it is the start point of the stroke, further before it) is returned as the first part.

In the automatic division learning method shown in FIG. 4, if a standard pattern of constituent elements similar to the first part or the second part of the input pattern is in the dictionary, the input pattern in which both parts are connected by a stroke is formed. On the other hand, the automatic division succeeds.
FIG. 1 shows a specific example. In FIG. 1, the box on the left is an input pattern of "words" written in abbreviated "words".
Suppose that you want to automatically divide and learn this input pattern.

Since the kanji "word" consists of "word" and "go", those standard patterns are converted into a predetermined size ("word" is a biased size, and "go" is a straw size). . In this case, it is assumed that the category pattern set of "go" has only one standard pattern.

To detect a portion similar to the standard pattern of "go" from the input pattern, the following process is performed. Cut_run
In the range from Cut_start to Cut_end, the input pattern is divided into two parts by the assumed break point, and the second
Match the part with the standard pattern of "go". The point with the highest similarity and greater than OK_Score is the final breakpoint. In the example of FIG. 1, the point C is determined to be a break point, and the portion from the start point of the input pattern to the point C ′ before C is returned as the pattern of the first portion.

Although the above-mentioned embodiment is an example in which a kanji is used as an input pattern, the present invention can be applied to the recognition of English words. When recognizing English words, by considering units such as prefixes, word stems, and inflection endings as word constituents, it can be applied in the same way as in the case of the above-mentioned kanji. FIG. 2 shows a specific example thereof. The word <software> and the element <soft>
It is assumed that the standard pattern shown in the right side of FIG. 2 is registered in each of the <ware>. In the case of English words, the standard pattern of words is generated by concatenating the standard patterns of its constituent elements.

As the input pattern, the cursive pattern of the individual habit character shown on the left side is entered, which is automatically divided and registered. In this case, from the input pattern <soft>
Or it judges whether the part similar to the standard pattern of <ware> can be detected. Like the example of Kanji mentioned above, Cu
The input pattern is divided into two parts by using the writing point of t_run as an assumed break point in the range from Cut_start to Cut_end, and the first part is matched with the standard pattern of <soft>, or the second part is <ware>. Matches the standard pattern of. And the similarity is highest, and OK_Sco
For example, the second part <ware> is cut out with the point that is larger than re as the final delimiter, and the remaining part (the part where the peculiar character is not very similar to the dictionary pattern) is <soft> and the structured font expression dictionary. Register with.

FIG. 6 is a flowchart showing a detailed processing procedure of the divided instruction pattern learning unit 5 shown in FIG. The learning processing of the divided instruction pattern is started when the application outputs guidance to the user when the automatic division learning is not successful, and when the user inputs the divided pattern based on the guidance. The learning unit 5 is provided with an input pattern of the entire character and a total of three patterns including the two partial patterns. The definitions of variables and symbols used in each step of the flowchart shown in FIG. 6 are as follows.

In_Ptt: normalized input pattern In_Ptt1, In_Ptt2: two normalized input partial patterns Score1: In_Ptt1 after registration of the maximum similarity between In_Ptt and Cat (A) Score1: In_Ptt2 after registration Maximum similarity score between In_Ptt and Cat (A) of: In_Ptt after registering In_Ptt1 and In_Ptt2
The maximum similarity between and Cat (A).

First, in step 21, Score1 and Score2 are initialized to "0". And the next step 22
Then register In_Ptt1 as component B. Next, in step 23, matching with the category pattern set of the registered character is performed, and the maximum similarity score MaxScore is set to S.
After entering core1, registration as component B is canceled (step 24).

Next, in step 25, In_Ptt2 is registered as the constituent element C. Then, the next step 26
Then, after matching with the category pattern set of the registered character and putting the value of the maximum similarity MaxScore into Score2, the registration as the constituent element C is canceled (step 27).

In the next step 28, Score1> Score2
If Yes, then in step 29, Sc
It is determined whether ore1> Des_Score or not, and if No, it is determined in step 30 whether or not Score2> Des_Score is satisfied (for Des_Score, refer to the definition of the flowchart of FIG. 4). If the determination result in step 29 is Yes, In_Ptt1 is registered as the constituent element B in the next step 31, and the process ends. The judgment result in step 30 is Ye
If s, In_Ptt2 is the constituent element C in the next step 32.
Then, the process ends.

When the judgment results in Steps 29 and 30 are No, the process proceeds to Step 33, In_Ptt1 is registered as the constituent element B and In_Ptt2 is registered as the constituent element C. In Step 34, the registered characters are standardized. Match the pattern category set to the maximum similarity in Score. And step 35
Then, it is determined whether or not Score> Des_Score. If Yes, the process ends. If No, the registration in step 33 is canceled (step 36), and the entire pattern is registered ( (Step 37) The process ends.

As described above, in the automatic division learning according to the embodiment of the present invention, the optimum automatic division, that is, the break point is determined while comparing the input pattern with the standard pattern of the dictionary. In the case of a basic kanji character whose input pattern is not divided into a plurality of constituent elements, the pattern whole learning unit shown in FIG. 1 often works, but the divisible character shape (kanji)
Automatic division learning is effective for this pattern. This is because even if the input pattern includes abbreviated characters or non-standard stroke order, in most cases, at least one part is similar to the standard pattern in the dictionary.

As shown in FIG. 7, when a character including the abbreviations of "Koto" (upper row) and a character including the different characters of "Ki" (lower row) are handwritten with a pen, the input patterns are the dictionary. It is not recognized because it is not similar to the standard pattern of. Therefore, for a handwritten input pattern of appropriately selected characters, for example, "word" and "pause", the abbreviation "word" bias and the heterograph "tree" are separated by the automatic division learning according to the embodiment of the present invention described above. If registered with, the recognition rate becomes extremely high as shown in FIG.

The dictionary pattern to be compared with the input pattern is
The larger the number, the higher the recognition rate, but on the other hand, there is a problem that the time required for the recognition processing becomes long. Therefore,
A process of sieving out a dictionary pattern candidate having a high probability of being similar to the input pattern from a large number of dictionary patterns (hereinafter, this process is referred to as "major classification process").
Is preferably performed before the character recognition processing. Hereinafter, this large classification process will be described.

In the large classification process, a character pattern formed by combining constituent elements from a structured font expression dictionary or a learning character pattern is obtained without dividing it into constituent parts such as bias, straw, and crown. A standard pattern is created by extracting the feature (unstructured feature).

First, as shown in FIG. 9A, the character pattern is divided into a 4 × 4 mesh. Next, the area of FIG.
Consider a small area with adjacent neighbors, such as ACMK, BDNL, CEOM, and FHRP. And each one is P00, P01, P
I will call it Pyx like 02 and P10. Pyx x
For y and y, the small area in the upper left is P00, and x is incremented by 1 each time it is shifted to the right, and y is incremented by 1 each time it is shifted down by 1. This creates nine subregions from P00 to P22. Fig. 9 (b) shows P03 from Fig. 9 (a).
Is an example of taking out.

Then, for each small area, each component of the stroke of the input pattern (components in four directions (|, \,-, /))
Find out how much each is included. In FIG.
An example of the value is shown. A series of these values is called a direction feature vector.

In the large classification process, similar characteristics are extracted from the input pattern, the characteristics are matched with the characteristics of the standard pattern of each character, and a plurality of character categories of the standard pattern having a high degree of similarity are selected and returned as candidates. .

In the character recognition processing described above, it is known which partial pattern of the input pattern to which constituent element needs to be registered. In the structured font representation dictionary in the character recognition process, it is effective to register only one new partial pattern in this component and register it in all the character categories included in a part of the character shape. However, such a method cannot be realized with an unstructured font representation dictionary that does not share constituent elements. Therefore, for all the character categories that include the component to be registered in the glyph (you can see from the structured font expression dictionary which characters include that component), the character that includes the partial pattern you want to additionally register. The pattern is formed from the structured font representation dictionary, the unstructured features described above are extracted, and the unstructured font representation dictionary is updated to allow the features.

FIG. 11 shows an example. It is assumed that FIG. 11A is written for the “word” and that the registration is performed.
In this case, the word deviation portion of the abbreviation is cut out and registered. Kanji including this bias (for example, "ron" shown in Fig. 11 (c))
Is obtained from the structured font representation dictionary. Therefore, the direction feature vector is extracted for each of the cut-out abbreviations “words” (FIG. 11B) and other kanji of the word bias, and the unstructured font representation dictionary is updated.

[0047]

According to the present invention, when a character shape including a part of an abbreviated character or an atypical character is registered, a partial pattern of the abbreviated character or an atypical character is cut out and registered. Even if all the other character shapes are not registered, the recognition rate can be improved by using this registered partial pattern.

[Brief description of drawings]

FIG. 1 is a diagram specifically illustrating automatic division of Chinese characters in a handwritten character online recognition device according to an embodiment of the present invention.

FIG. 2 is a diagram specifically illustrating automatic division of English words in the handwritten character online recognition device according to an embodiment of the present invention.

FIG. 3 is a configuration diagram showing a font registration function for character recognition in the handwritten character online recognition device according to an embodiment of the present invention.

FIG. 4 is a flowchart showing an automatic division learning processing procedure performed by the automatic division learning unit shown in FIG.

5 is a flowchart showing a detailed procedure of step 1 (processing for detecting a part similar to the component B) in FIG. 4;

6 is a flowchart showing a detailed processing procedure of a divided instruction pattern learning unit shown in FIG.

FIG. 7 is a diagram showing a recognition result for a handwritten input pattern before learning by automatic division.

FIG. 8 is a diagram showing a recognition result for a handwritten input pattern after learning by automatic division.

FIG. 9 is a diagram illustrating a character pattern small area in a large classification process.

FIG. 10 is a diagram showing an example of a direction feature vector in a large classification process.

FIG. 11 is an explanatory diagram of an unstructured font expression dictionary for character categories including word bias.

FIG. 12 is a diagram showing an example of forming a standard pattern of “bright”.

FIG. 13 is a diagram showing a pattern configuration of “cat”.

[Explanation of symbols]

1 ... Pattern whole learning unit, 2 ... Automatic division learning unit, 3 ... Structured character expression dictionary, 4 ... Judgment unit, 5 ... Divided instruction pattern learning unit.

[Procedure amendment]

[Submission date] December 28, 1995

[Procedure amendment 1]

[Document name to be amended] Drawing

[Correction target item name] Fig. 4

[Correction method] Change

[Correction contents]

FIG. 4

Claims (11)

[Claims] 1. A standard pattern composed of a combination of various radical patterns registered, and an input character pattern acquired by using a tablet which outputs a handwriting as a coordinate data string are subjected to matching processing to obtain the input character. In a handwritten character online recognition device for recognizing a pattern, when an input character pattern is registered in a dictionary, the input character pattern is compared with the corresponding standard pattern, and a radical pattern forming the standard pattern is compared. A character registration / learning method for a handwritten character online recognition device, characterized in that a partial pattern not similar to is automatically divided from the input character pattern and the partial pattern is registered as a corresponding radical pattern. 2. The input character pattern by performing a matching process between a standard pattern composed of a combination of registered partial glyph patterns and an input character pattern acquired using a tablet that outputs handwriting as a coordinate data string. In the handwritten character online recognition device for recognizing, when a dictionary is registered for an input character pattern, the input character pattern is compared with the standard pattern corresponding to the input character pattern, and a partial glyph pattern forming the standard pattern is obtained. A font registration / learning method for a handwritten character online recognition device, wherein a dissimilar partial pattern is automatically divided from the input character pattern and the partial pattern is registered as a corresponding partial character shape pattern. 3. The input character by performing a matching process between a standard pattern composed of a combination of various radical patterns registered and an input character pattern acquired using a tablet which outputs handwriting as a coordinate data string. In the handwritten character online recognition device for recognizing a pattern, when performing dictionary registration for an input character pattern, a partial pattern is input from the tablet for each radical constituting the input character pattern, and for each input partial pattern,
A standard pattern is generated by using the input partial pattern as a part of the standard pattern, and a maximum similarity is calculated by performing matching processing between each standard pattern corresponding to each input partial pattern and the input character pattern. A font registration / learning method for a handwritten character online recognition device, characterized in that an input partial pattern giving the maximum similarity is registered as a radical pattern when the value exceeds the similarity. 4. The method according to claim 3, wherein when the maximum similarity does not reach the target similarity, the similarity between the standard pattern formed by the combination of the input partial patterns and the input character pattern is calculated, and the similarity is the target similarity. A font registration / learning method for a handwritten character online recognition device, characterized in that a combination of input partial patterns is registered when the value exceeds a degree. 5. The font registration / learning method of the handwritten character online recognition device according to claim 4, wherein when the similarity does not reach the target similarity, the entire input character pattern is registered. 6. The input character is obtained by performing a matching process between a standard pattern composed of a combination of various radical patterns registered and an input character pattern acquired using a tablet which outputs handwriting as a coordinate data string. In a handwritten character online recognition device for recognizing a pattern, a means for comparing the input character pattern with the standard pattern corresponding thereto when the dictionary is registered for the input character pattern, and the standard pattern is configured as a result of the comparison. And a means for automatically dividing the partial pattern from the input character pattern when a partial pattern not similar to the radical pattern is detected, and a means for registering the partial pattern as a corresponding radical pattern. An online recognition device featuring handwritten characters. 7. The input character pattern by performing a matching process between a standard pattern composed of a combination of registered partial glyph patterns and an input character pattern acquired using a tablet that outputs handwriting as a coordinate data string. In a handwritten character on-line recognition device for recognizing, when a dictionary is registered for an input character pattern, a means for comparing the input character pattern with the corresponding standard pattern, and as a result of the comparison, constructs the standard pattern. And a means for automatically dividing the partial pattern from the input character pattern when a partial pattern not similar to the partial character pattern is detected, and a means for registering the partial pattern as a corresponding partial character pattern. Handwriting character online recognition device. 8. The input character is obtained by performing a matching process between a standard pattern composed of a combination of various radical patterns registered and an input character pattern acquired using a tablet which outputs handwriting as a coordinate data string. In a handwritten character online recognition device for recognizing a pattern, when performing dictionary registration for an input character pattern, a means for inputting a partial pattern for each radical constituting the input character pattern from the tablet to obtain a coordinate data string, A means for generating a standard pattern by using the input partial pattern as a part of the standard pattern for each input partial pattern, performing matching processing between the standard pattern corresponding to each input partial pattern and the input character pattern, and obtaining a maximum similarity. When the maximum similarity exceeds the target similarity, the maximum similarity is given. And a means for registering an input partial pattern as a corresponding radical pattern. 9. The method according to claim 8, wherein when the maximum similarity does not reach the target similarity, a means for determining the similarity between the standard pattern composed of a combination of the input partial patterns and the input character pattern, and the similarity is the target similarity. An online recognition device for handwritten characters, comprising: means for registering a combination of input partial patterns when the value exceeds a degree. 10. The handwritten character online recognition device according to claim 9, further comprising means for registering the entire input character pattern when the similarity does not reach the target similarity. 11. When the characteristic amount of a standard pattern of each character shape is stored in a dictionary for each character shape and a matching process between a large number of standard patterns and an input character pattern is performed to perform character recognition of the input character pattern. Prior to the character recognition processing, a candidate for a standard pattern having a feature amount similar to the feature amount of the input character pattern is searched in the dictionary and preselected, and a matching process between the standard pattern candidate and the input character pattern is performed. In the handwritten character online recognition device for performing the above, when a partial pattern is registered by the font registration / learning method according to any one of claims 1 to 3, the feature amount of a standard pattern having the partial pattern is obtained to obtain the dictionary. An on-line recognition device for handwritten characters, comprising: means for updating.