JPH07282200A - On-line character recognizing device - Google Patents

Abstract

PURPOSE:To execute the matching with high precision without increasing dictionary capacity by taking an error occurring set the time of the matching of a deformed handwritten character like an inclined character, etc., into consideration. CONSTITUTION:In a preprocessing part 2, the linearization of the handwritten character inputted from a tablet 1 is executed. In a feature point extracting part 3, feature point extraction processing is executed. In a stroke coding part 4, stroke coding is executed. In a broadly classifying part 5, broad classification using a stroke number is executed. In an intermediately classifying part 6, an inter partial-pattern vector is calculated, and inter partial-pattern vector matching processing is executed. In a partial pattern value matching part 7, a partial pattern Q value is calculated, and in a correction value processing part 12, the partial pattern Q value is corrected by a correction judging part 11 by using the correction value stored in a memory. In a partial pattern stroke code matching part 8, the matching processing of partial pattern stroke code distribution is executed.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an on-line character recognition device for identifying written characters in real time.

[0002]

2. Description of the Related Art Conventionally, as a technique in such a field,
For example, some documents were described in the following documents. Document: Japanese Patent Application No. 62-229384. Conventionally, in an online character recognition device, there is a pattern matching method as a general character recognition method. In this pattern matching method, feature points are extracted from the coordinate data string of a stroke (writing portion from pen-on to pen-off) input by handwriting, and matched with information of a registered pattern (hereinafter referred to as a registered pattern),
Performs character recognition. In this pattern matching method, it takes a lot of time to process which stroke of each stroke of the registered pattern each written stroke should be matched with. In addition, if the balance of the entire characters is disturbed, for example, when the character "kawa" has a small "shi" and a large "OK", the matching result is not similar, and it is vulnerable to handwritten character deformation. Therefore, there is an online character recognition method described in the above-mentioned document that compensates for the above-mentioned drawbacks of the pattern matching method and requires less processing amount. In the above-mentioned literature, a large classification is made according to the number of strokes of a writing character, and a portion to be written as a series in writing is defined as a partial pattern. Then, this partial pattern is subjected to middle classification by the vector between the centers of gravity, and matching is performed using the value of Q value and the value of stroke code distribution as the characteristic parameter of the partial pattern. This makes it possible to perform character recognition that is resistant to fluctuations in the overall character balance and that requires less processing.

[0003]

However, although the conventional online character recognition device has the advantage that it is resistant to fluctuations in the overall character balance and requires a small amount of processing, it has the following problems. It was In the conventional online character recognition device, in addition to the Q value, which is a characteristic parameter extracted from a normal character as a basic dictionary, in addition to the value of a tilted character, for example, a habit character such as a character that rises or falls to the extreme right, as a basic dictionary. In order to improve the identification accuracy, a dictionary is created and placed for all the characters of the dictionary definition character with the characteristic parameter corresponding to the slanted character. For this reason, there is a problem that the dictionary capacity becomes extremely large, the amount of processing for searching increases with the dictionary capacity, and the recognition time becomes long.

[0004]

In order to solve the above-mentioned problems, the first invention removes unnecessary data of a coordinate data string written on a tablet by a writer and executes a linearization process. From the coordinate data string linearized by the pre-processing unit, each stroke is defined by the feature point extraction unit that extracts the features of the strokes that make up the written character and the positional relationship between the feature points extracted by the feature point extraction unit. A stroke coding unit to be coded and a portion to be written as a series in writing are used as a partial pattern, the characteristic parameter Q value of the partial pattern is calculated from the output of the stroke coding unit, and the registered character is registered in advance. A partial pattern Q value matching unit that compares the registered characteristic parameter Q value calculated from the partial pattern with the characteristic parameter Q value is provided. An online character recognition device for performing character recognition based on a comparison result between output data of a point extraction unit or a stroke coding unit and a registered pattern of the registered character registered in advance or a comparison result of a partial pattern Q value matching processing unit. The following processing units are provided. That is, the relative distance value between the characteristic parameter Q value of the written character handwritten by the writer and the registered characteristic parameter Q value of the registered character corresponding to the written character is stored as a correction value, and the writer writes the relative distance value. A correction unit is provided for correcting the characteristic parameter Q value of the handwritten character or the registered characteristic parameter Q value of the registered character when the handwriting is input, based on the stored correction value. The partial pattern Q value matching processing unit compares the characteristic parameter Q value corrected by the correcting unit with the registered characteristic parameter Q value, or the characteristic parameter Q value and the registered characteristic parameter corrected by the correcting unit. The Q value is compared. In the second invention, the correction unit of the first invention performs the following processing. That is, in the learning mode in which the writing character of the writer corresponding to the registered character is learned by allowing the writer to input the writing character in advance, the correction unit causes the correction unit to previously input the characteristic parameter Q value of the writing character. And the relative distance value between the registered characteristic parameter Q value of the registered character corresponding to the written character is calculated, and the relative distance value is stored as a correction value. In the third invention, the correction unit of the first invention performs the following processing. That is, when the registered character corresponding to the handwritten character of the writer is character-recognized as a candidate character, and the character-recognized candidate character is confirmed by the writer, the correction unit determines the characteristic parameter of the handwritten character. And a relative distance value between the candidate character and the registered characteristic parameter are calculated, and the relative distance value is stored as a correction value.

[0005]

According to the first aspect of the present invention, since the online character recognition device is configured as described above, the correction unit includes the characteristic parameter Q value of the handwritten character handwritten by the writer and the characteristic parameter Q value corresponding to the handwritten character. The relative distance value between the registered character and the registered characteristic parameter Q value is stored as a correction value, and when the writing is input by a writer, the characteristic parameter Q value of the written character input by the writer or the registration of the registered character is registered. Since the characteristic parameter Q value is corrected based on the stored correction value, the characteristic parameter Q value due to the habit character of the handwritten character written by the writer is corrected. Therefore,
It has the function of preventing the recognition rate from decreasing due to the habit of written characters. According to the second aspect of the invention, the correction unit determines the relative distance value between the characteristic parameter Q value of the handwritten character that the writer previously inputs by hand and the registered characteristic parameter Q value of the registered character corresponding to the handwritten character. The calculated relative distance value is stored as a correction value. Then, based on this correction value, the characteristic parameter Q value or the registered characteristic parameter Q value of the written character of the writer is corrected. According to the third invention, the registered character corresponding to the written character of the writer is character-recognized as a candidate character, and when the character-recognized candidate character is confirmed by the writer, the correction unit causes the written character to be The relative distance value between the characteristic parameter of No. 1 and the registered characteristic parameter of the candidate character is calculated, and the relative distance value is stored as a correction value. Then, based on this correction value, the characteristic parameter Q value or the registered characteristic parameter Q value of the written character of the writer is corrected. Therefore, the above problem can be solved.

[0006]

[Embodiment] In the present embodiment, the following three modes (1) to (3) are provided as modes for confirming the correspondence between registered characters registered in the dictionary in advance and the written characters of the writer,
When the registered character and the written character of the writer are confirmed, the relative distance value between the registered characteristic parameter Q value of the registered character and the characteristic parameter Q value of the written character of the writer described later is registered as a correction value. (1) Learning Mode In this mode, a predetermined character is written, and the relative distance value between the characteristic parameter Q value of the written character and the registered characteristic parameter Q value of the registered character is used as a correction value. For example, when the power is turned on or reset, it is common to enter this mode and learn the first written character. (2) Next candidate selection mode In this mode, if the recognition result matches the written character, the writer checks the recognition result and the recognition result is different from the written character. If
In the mode for selecting the next candidate, when the writer confirms the recognition result, the relative distance value between the characteristic parameter Q value of the written character and the registered parameter Q value of the registered character is used as the correction value. is there. (3) Confirmation mode This mode recognizes that the written character has been confirmed by the writer by determining whether or not the written written character string has been transferred to another mode, and the written written character string is recognized. When transferred to another mode, the relative distance value between the characteristic parameter Q value of each character of the written character string and the registered characteristic parameter Q value of the corresponding registered character is used as a correction value. Generally, when creating a sentence, first, a sentence input mode is provided to write a written character in word units, and this written character string is transferred to another mode, for example, a sentence mode to create a sentence. This mode switching is used for character determination. First, the online character recognition devices according to the embodiments of the present invention corresponding to the modes (1) to (3) will be described, and then the operations of the online character recognition devices according to these embodiments will be described.

First Embodiment FIG. 1 is a functional block diagram of an online character recognition device showing a first embodiment of the present invention. The first embodiment of the present invention is an online character recognition device for realizing the learning mode. This online character recognition system can be used for individual circuits using integrated circuits, or digital signal
It is configured by program control of a processor (DSP) or the like, and has a tablet 1 for pen position input of position coordinates of characters. A pre-processing unit 2 is connected to the tablet 1, and a feature point extracting unit 3 for extracting feature points of written characters is connected to the output side thereof. A stroke code unit 4 for extracting stroke features is connected to the feature point extraction unit 3, and a large classification unit 5 for classification according to the number of strokes is connected to the output side thereof. The middle classification unit 6 is connected to the output side of the large classification unit 5, and the partial pattern Q value matching unit 7 is further connected to the output side thereof. A partial pattern stroke code distribution matching unit 8 is connected to the output side of the partial pattern Q value matching unit 7. An output terminal 10 to a display or the like is connected to the partial pattern stroke code distribution matching unit 8. A correction processing unit 12 and a correction determination unit 11 as a correction unit are connected to the partial pattern Q value matching unit 7. A correction determination unit 11 is connected to the partial pattern stroke code distribution matching unit 8 and the correction processing unit 12. Further, a correction control unit 13 is provided which determines the learning mode, causes the writer to write a predetermined character, and confirms the correspondence with the registered character. The correction processing unit 12 is connected to the correction control unit 13. Has been done. The correction processing unit 12 calculates a relative distance value between all the Q values of the written character of the writer and all the Q values of the registered character, creates a correction value based on the relative distance value, and corrects the correction value. It is output to the unit 11. Further, the correction processing unit 12 corrects the Q value of the handwritten character of the writer based on the correction value corresponding to the registered character stored by the correction determination unit 11, and the corrected Q value is used as the partial pattern Q value matching unit. 7 is output. The correction determination unit 11 determines the correction value based on the outputs from the partial pattern Q value matching unit 7, the correction processing unit 12, and the partial pattern stroke code matching unit 8.
It is determined whether or not to store in the memory provided in. Second Embodiment FIG. 2 is a functional block diagram of an online character recognition device showing a second embodiment of the present invention, in which elements similar to those in the first embodiment in FIG. is doing. The second embodiment of the present invention is an online character recognition device for realizing the next candidate selection mode, which has a correction control unit 23 connected to the tablet 1, and the correction control unit 23 has a correction processing unit. 22 is connected. The correction control unit 23 determines whether the writer has performed the next candidate selection operation from the coordinate values input from the tablet 1 before preprocessing, and outputs the determination result to the correction processing unit 22. The correction processing unit 22 calculates the relative distance value between all the Q values of the written characters of the writer and all the Q values of the registered characters based on the determination result of the correction control unit 23, and the correction value based on the relative distance value. Is generated and the correction value is output to the correction determination unit 11. Further, the correction processing unit 22 corrects the Q value of the handwritten character of the writer based on the correction value corresponding to the registered character stored by the correction determination unit 11, and uses the corrected Q value for the partial pattern Q value matching unit. 7 is output.

Third Embodiment FIG. 3 is a functional block diagram of an online character recognition device showing a third embodiment of the present invention. The same elements as those of the second embodiment in FIG. The code is attached. The third embodiment of the present invention is an online character recognition device for realizing the next candidate selection mode similar to the second embodiment. However, the difference from the second embodiment is that the writing characters of the writer are written. The relative distance value between the Q value in the Y direction and the Q value in the Y direction of the registered character is calculated, and a correction value is created based on the relative distance value. The online character recognition device in FIG. 3 is provided with a correction processing unit 32 instead of the correction processing unit 22 in FIG. The correction processing unit 32 calculates a relative distance value between the Q value of the writing character in the Y direction and the Q value of the registered character in the Y direction, creates a correction value based on the relative distance value, and calculates the correction value. It is output to the correction determination unit 11. Further, the correction processing unit 32
Corrects the Q value in the Y direction of the written character of the writer based on the correction value corresponding to the registered character stored by the correction determination unit 11, and outputs the corrected Q value to the partial pattern Q value matching unit 7. To do.

Fourth Embodiment FIG. 4 is a functional block diagram of an online character recognizing device showing a fourth embodiment of the present invention. The same elements as those of the first embodiment in FIG. The code is attached. The fourth embodiment of the present invention is an online character recognition device for realizing the confirmation mode. The correction processing unit 42 in FIG. 4 determines whether the written character string written is transferred to another mode, and based on the determination result, all Q values of the written character of the writer and all Q values of the registered character. Calculate the relative distance value between
A correction value is created based on the relative distance value, and the correction value is output to the correction determination unit 11. Further, the correction processing unit 42 corrects the Q value of the handwritten character of the writer based on the correction value corresponding to the registered character registered by the correction determination unit 11, and the corrected Q value is used by the partial pattern Q value matching unit. 7 is output. FIG. 5 is a flow chart showing the outline of the functions of the online character recognition device of FIGS. 1 to 4, and the outline of the operation of the online character recognition device of FIGS. 1 to 4 will be described using this figure. In step 1, the process of the preprocessing unit 2 is performed. In addition, when performing the correction process when selecting the next candidate, the tablet 1
The correction control unit 23 in FIG. 2 or FIG. 3 determines whether the writer has performed the next candidate selection operation based on the coordinate values before preprocessing that have been input, and performs the correction process when selecting the next candidate, which will be described later. In step 2, the characteristic point extraction processing of the characteristic point extraction unit 3 is performed. In step 3, the stroke coding unit 4 carries out stroke coding. In step 4, the large classification unit 5 performs large classification processing according to the number of strokes. Step 5
In 8, the processing in the middle classification unit 6 is performed. Step 5 is a character dictionary end determination process, step 6 is a partial pattern vector calculation process, step 7 is a partial pattern vector matching process, and step 8 is a matching result determination process.

Steps 9 to 14 are processes in the partial pattern Q value matching unit 7. Step 9 is a partial pattern Q value calculation process for calculating the characteristic parameter Q value Q _n ^* . Step 10 is the correction processing in the correction value processing unit 12 in FIG. 1, the correction value processing unit 22 in FIG. 2, the correction processing unit 32 in FIG. 3, and the correction processing unit 42 in FIG. 11
This is a process of correcting the partial pattern Q value with the correction value stored in the internal memory. Step 11 Q value of the corrected partial pattern, a partial pattern Q value matching processing of matching the Q value of Q _n ^'* and the dictionary characters. In step 12, the Q-value matching distance d of the partial pattern forming each character is based on the result of the Q-value matching in step 11.
Value relative to Q _n ^* of ^{_{Qn {dQ n / Q n *}} } m,
(Hereinafter, this value is referred to as {ΔQ _n } _m ) is secured. Here, n is the number of Q values from 1 to 16, and m is the number of the partial pattern. Step 13 is a process of calculating the distance value d _i , and step 14 is a process of sorting the distance d _i . Steps 15 to 18 are processing in the partial pattern stroke code distribution matching unit 8, and Step 1
5 is a matching process, step 16 is a partial pattern stroke code distribution calculation process, step 17 is a partial pattern matching distance ds calculation process, and step 18 is a sorting process.

Step 19 is a process in the correction control unit 13 in FIG. 1, in which a predetermined character is written, and the matching result of the characteristic parameter Q value of the written character and the registered characteristic parameter Q value of the dictionary is calculated. This is a process of determining whether the learning mode is used as a correction value. Step 20 is a process in the correction control unit 23 in FIG. 2 or FIG. 3, and the writer looks at the recognition result of the written character and, if the recognition result is different from the written character, performs the operation of selecting the next candidate. This is a process of determining whether or not. Step 21 is the correction processing unit 42 in FIG.
This is the character confirmation determination processing for determining whether or not the written character is confirmed by the confirmation of the writer and the written character string written is transferred to another mode. Step 30-
33 is a Q value correction process in the learning mode performed by the correction processing unit 12 in FIG. Step 30 is a matching distance dQ _n (n is a matching distance between a partial pattern value forming the character and a partial pattern Q value of the same portion registered in the dictionary when a predetermined character is written in the learning mode. This is a process for obtaining (Q values 1 to 16) and normalizing (relativeizing) with the written Q _n ^* value. That is, this is a process of calculating {ΔQ _n } for the number of partial patterns and obtaining {ΔQ _n ^* } _m .

Step 31 is a correction value calculation processing for performing processing such as weighting and adding {ΔQ _n ^* } _m calculated in step 30 for each partial pattern by the number of strokes of the partial pattern. Step 32 is a process of the correction determination unit 11, and it is determined whether the correction value calculated in step 31 is smaller than a threshold value ε, and when it is larger than the threshold value ε, that is, when the distance is extremely large, it is incorrect. Since there is a high possibility that the correction will be performed, the correction value storing process in the next step 33 is skipped. Step 33 is a process of storing the calculated correction value in the memory in the correction determination unit 11. Step 4
0 to 43 are Q value correction processing in the next candidate selection mode performed by the correction processing unit 22 or 32 in FIG. 2 or 3. In step 40, when the next candidate is selected in the next candidate selection mode, dQ _n (a matching distance between the partial pattern Q forming the selected character and the partial pattern Q value of the same part registered in the dictionary). n is a process of obtaining the Q values 1 to 16) and normalizing (relativeizing) the written Q _n ^* value. That is, this is a process of calculating {ΔQ _n } for the number of partial patterns and obtaining {ΔQ _n } _m . Step 41
Is a correction value calculation processing for performing processing such as weighting and adding {ΔQ _n } _m calculated in step 40 for each partial pattern by the number of strokes of the partial pattern. In step 42, it is determined whether or not the correction value calculated in step 41 is smaller than a threshold value ε, and when it is larger than the threshold value ε, that is, when the distance is extremely large, there is a high possibility that a wrong correction is performed. For,
This is processing for skipping the correction value storage processing in the next step 43. In step 43, the calculated correction value is used as the correction determination unit 1
This is a process of storing in the memory in 1.

Steps 50 to 53 are Q-value correction processing at the time of candidate determination, which is performed by the correction processing unit 42 in FIG. Step 50 is a matching distance between the partial pattern Q value forming the first character and the partial pattern Q value of the same portion registered in the dictionary when the candidate is determined, dQ _n (n is the number of Q values 1 to 1).
16) is a process of obtaining (normalizing) (relativeizing) the written Q _n ^* value. That, {Delta] Q _n} the calculated number of partial patterns, a process for obtaining a {ΔQ _{n} m.} Step 51 is a correction value calculation process for performing processing such as weighting and adding {ΔQ _n } _m calculated in step 50 for each partial pattern by the number of strokes of the partial pattern. Step 52
Determines whether the correction value calculated in step 51 is smaller than a threshold ε, and when the correction value is larger than the threshold ε, that is, when the distance is extremely large, there is a high possibility that incorrect correction is performed. This is processing for skipping the correction value storage processing in the next step 53. Step 53 is a process of storing the calculated correction value in the memory in the correction determination unit 11. 6 is shown in FIG.
4 is a diagram showing a configuration example of a character dictionary provided in the online character recognition device Nos. 4 to 4, in which the character dictionary can select a character by the number of strokes (the number of strokes). FIG. 7 is a diagram showing a configuration example of a partial pattern dictionary provided in the online character recognition device of FIGS. Below, FIG.
The processing contents of the online character recognition device of FIGS. 1 to 4 will be described in detail below in (1) to (6) with reference to FIG.

(1) Input and preprocessing (step 1) FIGS. 8A to 8C are diagrams for explaining the preprocessing of step 1 in FIG. 2, and "." It represents a writing data string from 1 or a feature point. A tablet 1 in FIGS. 1 to 4 is an input device for writing characters, and when a character is written by the tablet 1, as shown in FIG. 8A, a writing data string {(x _i ,
y _i ), i = 1, 2, ..., N _j } are extracted and sent to the preprocessing unit 2. The preprocessing unit 2 smoothes the data as shown in FIG. 8B by performing noise removal processing, moving average processing, or smoothing processing on the transmitted writing data string, and the feature point extraction unit 3 Output to. (2) Feature Point Extraction Processing (Step 2) The feature point extraction processing unit 3 of FIGS. 1 to 4 extracts feature points from the output of the preprocessing unit 2. There are several methods for this feature point extraction processing, but here, as an example, a smoothed data string {(x _i , y _i ), i = 1, 2, ..., N _j }.
A method of calculating a sign (positive, negative, sign of 0) in the x and y directions between data of _j and extracting a change point of the sign state as a feature point will be described. Signs XS _i and YS _i in the x and y directions between the data are obtained by the following equation (1), and +, 0,
Express with −. XS _i = Sign (x _i −x _{i −1} ) YS _i = Sign (y _i −y _{i −1} ) (1) The signs in the x and y directions between the respective data thus obtained are If it is the same as the signature between the previous data, if it is the same, it is not registered as a feature point. FIG. 8 (c) shows feature points to which start points and end points are added in addition to the points thus obtained. In general, this process may be called linearization. These feature points are hereinafter referred to as segments, and the feature points are represented by {(x _i , y _i ), i = 1, 2, ..., I _j } _j . The feature point information obtained as described above is output to the stroke code unit 4 and the middle classification unit 6.

(3) Stroke Encoding Process (Step 3) The stroke encoding unit 4 of FIGS. 1 to 4 encodes each stroke based on the characteristic point information obtained by the characteristic point extracting unit 3. There are many ways to do this coding, but typically one will take, for example, the X, Y signature of each segment,
Classification is performed according to the angle of the segment and the rotation angle between the segments, and coding is performed. FIG. 9 is an explanatory diagram of the above code processing. In the figure, θ ₁ , θ ₂ , and θ ₃ indicate angles of each segment (angles of Euclidean space polar coordinates), and θ ′ ₁ and θ
′ ₂ and θ ′ ₃ indicate the rotation angle between adjacent segments. The encoded stroke data is output to the large classification unit 5 and the partial pattern stroke distribution matching unit 8. (4) Major classification processing (step 4) In the major classification unit 5 of FIGS. 1 to 4, the stroke coding unit 4
, The target character is roughly classified according to the number of strokes. Therefore, a character that can be the number of strokes is prepared in advance in the character dictionary for each number of strokes, as shown in FIG. For example, it is assumed that the stroke number of the character pattern input by handwriting is 10 strokes. In this case, among the characters stored in the character dictionary, characters that can be 10 strokes as shown in FIG. 6 are selected as candidate characters such as "nuisance", ..., "Akira", ..., "double", ... To do.

(5) Medium Classification Processing (Steps 5-8) In the medium classification processing unit 6 shown in FIGS. 1 to 4, the candidate characters obtained by large classification by the large classification unit 5 are described below. The pattern is further classified into medium (step 5). Here, the partial pattern refers to a portion of one character that is written as a series of characters in writing,
The inter-partial-pattern vector is a vector whose start point and end point are the center of gravity of one partial pattern and the center of gravity of another partial pattern, respectively. First, an example of a method of calculating a vector between partial patterns will be described. Assuming that each segment in the partial pattern is x and y components are (dx _i , dy _i ), the length dl _i of each segment is expressed by the following equation (2). dl _i = (dx _i ² + dy _i ² ) ^1/2 (2) Further, the center coordinates of each segment normalized by dividing by the character widths HX and HY are (x _i ^* , y _i ^* ) Then, the barycentric coordinates (X _w , Y _w ) of the partial pattern are obtained by the following equation (3).

[0017]

[Equation 1] The center of gravity of each partial pattern is obtained by the above method, and the vector between partial patterns is obtained with the center of gravity of one partial pattern and the center of gravity of another partial pattern as the starting point and the ending point, respectively. Note that here, the inter-partition-pattern vector is considered in the x direction and the y direction. As an explanatory diagram of the inter-partition-pattern vector, FIG.
_x (mouth, nitrite) shows an example of a B _y (mouth, nitrite). However, B represents a vector. The middle classification unit 6 narrows down the candidate characters selected by the large classification unit 5 by the partial pattern vector to perform the middle classification. Here, as an example, consider the case where "Akira" is written and input. The procedure of the middle classification process for this input character will be described below.

Since the character "Akira" written and input has 10 strokes, the 10 stroke portion of the character dictionary shown in FIG. 6 is referred to.
Then, the character "mute" is arranged in the first place here, and in that column, the partial patterns constituting the "mute" and the stroke information of each partial pattern in the writing order of the partial patterns (hereinafter referred to as "cut position"). ) And the inter-partial-pattern vector value calculated in advance from the registered pattern are shown.
Below, similar information is lined up for the letters "Akira", "double", and "auto", and the middle classification unit 6 follows this letter order.
Whether or not it should be a candidate is determined, and middle classification is performed as follows. First, assuming that the character input by handwriting is "mute", the matching distance d _{vec of the} vector between partial patterns is obtained. As shown in FIG. 6, the "mute" has a cut position of (3, 7), that is, a "mouth" is formed in the first stroke to the third stroke, and a "sub" is formed in the fourth stroke to the tenth stroke. However, in this example, the input pattern is "Akira"
Since this is the cut position, the vector between partial patterns for "Akira" at this cut position is as shown in FIG. The inter-partial vector B _x ^* (mouth,
Nitrous), B _y ^* (mouth, nitrous), the partial pattern between vectors B _x (mouth that Kakare dictionary, nitrous), B _y (mouth, nitrite) the difference between the is the matching distance d _vec, It is calculated according to the following equation (4). d _vec = | B _x ^* (mouth, nitrous) -B _x (mouth, nitrous) | + | B _y ^* (mouth, nitrous) -B _y (mouth, nitrous) | ··· (4) Writing the character of When the number of partial patterns is plural, normalization is performed by the number of partial patterns BPN, and the matching distance d _vec is calculated according to the following equation (5) (step 7).

[0019]

[Equation 2] Here, a certain threshold value vecrej is set, it is determined whether or not the calculated matching distance d _vec is larger than vecrej (step 8), and the following processing is performed according to the determination result. (A) When d _vec > vecrej Assuming that it is not the referenced character (in this case, "mute"), "mute"
Matching between the partial patterns of the next character is performed without using as a candidate. (B) When d _vec ≤vecrej, it is assumed that "a mute", and calculation and matching of the partial pattern Q value described below are performed. When the input character is "Akira",
As shown in FIG. 11, the vector becomes a downward vector, but the "mute" in the dictionary is to the right, the partial pattern vector distance is large (d _vec > vecrej), and the "mute" does not remain as a candidate.

(6) Partial Pattern Q-Value Matching Process (Steps 9 to 13) In the partial pattern value matching unit 7 of FIGS. The characteristic parameter Q value of the partial pattern (hereinafter referred to as the partial pattern Q value) is calculated, and
The matching is performed with the registered characteristic parameter Q value of the partial pattern in the partial pattern dictionary shown in (also called the partial pattern Q value). The partial pattern Q value of this partial pattern dictionary is created and stored in advance by the registered pattern. Here, the partial pattern Q value is a characteristic parameter indicating the length, direction, and position of each segment. In the online character recognition, the movement of a pen to be written is obtained as important information in the X direction, Y direction, +, or − direction, and this partial information Q value is effectively used. First, a method of calculating the partial pattern Q value will be described. In the following equations (6) to (13),
Σ is the total stroke, addition for all segments, HX,
HY indicates the character width.

X-direction position of + x-direction component

[Equation 3] -X direction position of x direction component

[Equation 4] Y-direction position of + y-direction component

[Equation 5] -Y direction position of y direction component

[Equation 6] Y-direction position of + x-direction component

[Equation 7] -Y direction position of x direction component

[Equation 8] X-direction position of + y-direction component

[Equation 9] -X direction position of y direction component

[Equation 10] In the case of the expressions (6) to (13), the values are the values in the respective directional positions when the origin is set to the lower left. At this time, the values near the origin become 0 when multiplied. To prevent a 0 Therefore, also described in the same manner for the value Q ₉ to Q ₁₆ in each direction position at the time of setting the origin at the top right, the total of 16 values of Q ₁ to Q _16, the target character It shall represent the length, direction and position of the segment of each stroke.

The partial pattern Q-value matching unit 7 calculates the partial pattern Q-values for those left by classification by the vector between partial patterns (step 9).
However, for example, it is assumed that “auto” is written in and “double” remains due to the classification by the partial pattern vectors.
In this case, the “double” cut position is (2, 5, 3) as shown in FIG. 3, and the partial pattern is a + stand + mouth, so the input pattern cut position (2,5,3). Cut and calculate each Q ₁ ^{* to} Q ₁₆ ^* . In the case of this example, the first
In the cut of

[Outer 1] Each calculated partial pattern Q value is corrected by the following equation (14) with the correction value {ΔQ _n } _m stored in the memory in the correction determination unit 11 in FIGS. (Step 10). (1) For all component correction (processing in the correction processing unit 12,32,42 in Figure ^{_{1,3,4) {Q n '*}}} m = {Q n * + (ΔQ n · Q n *) } _m where, n = 1~16, m = 1~BPN ··· (14) (2) when considering the deformation by case or inclined characters (upper right or right edge) of the partial component correction (in Fig. 3 Processing by the correction processing unit 32), since only the Y-direction component needs to be corrected, for example, n = 3
(Q _3: + Y direction component in the Y direction _{position), n = 4 (Q 4} :
Y-direction position of the -Y direction _{component), n = 7 (Q 7} : -Y direction <br/> X-direction position of the _{component), n = 8 (Q 8} : + Y direction component of X
A total of eight (direction positions) etc. are corrected according to the above equation (14).

Next, the corrected partial pattern Q value Q ₁
′ To Q ₁₆ ′ and the partial pattern Q value in the partial pattern dictionary shown in FIG. 7 are matched. In this matching, first

[Outside 2] Calculated Q ₁ ^{* to} Q ₁₆ ^* and Q of partial pattern dictionary “mouth”
Match _{1 to} Q ₁₆ . The sum of the differences in these matchings is the matching distance d _BP . At this time, the matching distance d _BP represents how close the input pattern “auto” is to “double”. Normally, the number of strokes BS _j of each partial pattern is used to perform weighting as in the following equation (15), which is used as the matching distance d _BP (step 11).

[Equation 11] At this time, in order to store the correction value, the Q value distance value of each partial pattern is calculated for each character candidate by the following equation (16) to obtain a relative value. {ΔQ _n ^* } _m = {(Q _n −Q _n ^* ) / Q _n ^* } _m (16) This {ΔQ _n ^* } _m value is corrected and determined for each character candidate to calculate a correction value. It is temporarily stored in the memory of the unit 11 (step 12). FIG. 12 is a diagram showing a storage format of the matching distance. As shown in the figure, candidate characters (JIS
Code), the number of partial patterns, and each Q-value relative distance value {ΔQ _n ^* } _m of each partial pattern are stored in the order of candidates. On the other hand, the distance d _BP obtained by the equation (15) and the d _vec obtained by matching the vector between partial patterns obtained in the previous step
The weighting parameters w _vec and w _BP of the equation (17) are previously determined and stored for each number of strokes, and the weighted values w _vec and w _BP are added to obtain the distance d _i (step 13). d _i = w _{vecd vec} + w _{BPd BP} (17) The above operation is performed for all candidate characters remaining in the classification by the partial pattern vectors, and sorting by d _i is performed (step 14).

(7) Partial pattern stroke code distribution matching processing (steps 15 to 18) Partial pattern stroke distribution matching unit 8 of FIGS.
Then, receiving the outputs of the partial pattern Q value matching unit 7 and the stroke coding unit 4, the partial pattern stroke distribution is obtained for the candidate characters narrowed down by the middle classification. Then, based on this distribution and the registered pattern, it is created in advance.
The partial pattern stroke code distribution stored in the partial pattern dictionary is matched and the upper candidates are ordered (step 15). The target range for this ordering is determined by, for example, the ratio with the distance d ₁ of the first candidate obtained by sorting d _i . That is, d _i /
Ordering is performed with the candidate character of d ₁ ≦ ZRATE as the target range. Here, the threshold value ZRATE is stored in advance for each stroke number of the written character. Next, a method of calculating the partial pattern stroke code distribution will be described. As an example, it is assumed that the input pattern is “auto” and the character selected as the first candidate is “double”. From the character dictionary of FIG. 3, it can be seen that the cut position of the candidate character “double” is (2, 5, 3) and the partial pattern is “a” + “standing” + “mouth”.
At this position, cut the input pattern "Auto". The distribution of the number of stroke codes obtained by the stroke encoding unit 4 is calculated for each partial pattern obtained by cutting (step 16).

For example, in the first cutting position

[Outside 3] Therefore, the stroke code distribution in which S03 is one and S04 is one is obtained. The partial pattern stroke code distribution calculated in this way is matched with the partial pattern stroke code distribution of the partial pattern dictionary of FIG. 8 which is calculated by averaging in advance from several Touroku patterns by the same procedure. To be done. In the case of this example, first, the calculation was performed as described above.

[Outside 4] The difference from the stroke code distribution of “a” in the partial pattern dictionary of FIG. 7 is obtained as the matching distance d _s (a).

That is, input pattern partial pattern dictionary S01 ... 0 book S01 ... 0.8 book S02 ... 0 book S02 ... 1.2 book S03 ... 1 book S03 ... 0 book S04 ... 1 S04 ... 0 S05 ... 0 S05 ... 0 difference 0.8 + 1.2 + 1.0 + 1.0 = 4.0 is d _s
It is required as (a). Similarly,

[Outside 5] Get d _s (mouth).

[0027] Then, it was normalized by each partial pattern stroke code number BS _J, the sum d _s of the matching distances of the partial patterns that are normalized, calculated from the following equation (18) (step 17).

[Equation 12] Next, the weighted parameters w _vec , w _Bp , w _{s are} calculated with respect to the distance d _vec obtained by the vector matching between the partial patterns, the distance d _{BP obtained} by the partial pattern Q-value matching, and the distance d _{s obtained} by the partial pattern stroke code distribution matching obtained above. Then, the distance D is calculated according to the following equation (19) (step 18). D = w _vec · d _vec + w _BP · d _BP + w _s · d _s (19) The distance value D is used for ordering and outputting character strings as a result from the output terminal 10.

At this time, the correction values in the respective modes (i) to (iii) described below are extracted and stored. (I) Correction value extraction in learning mode (steps 30 to 3)
3) In the learning mode in which the writer writes a predetermined character and the character characteristics of the writer are extracted, correction for each candidate character temporarily stored in the memory in the correction determination unit 11 as shown in FIG. Q value distance value for use {ΔQ _n ^* } _m value is searched for a predetermined character and the correction Q value distance value is extracted by the number of partial patterns (step 30), and the relative distance value ΔQ _n ^* is calculated by the equation (17). Calculate (step 31). It is determined whether or not the added value dQ of the absolute value of the {ΔQ _n ^* } _m value in the expression (20) is smaller than a certain threshold value ε (step 32), and the following processing is performed based on the determination result. If the correction value has not already been calculated in the learning mode, the distance is extremely large when the added value dQ of the absolute value of the {ΔQ _n ^* } _m value of the equation (20) is larger than a certain threshold ε. If there is a high possibility that there is an error or a writing error, the correction value is not stored, the buzzer, warning display, etc. are displayed to the writer, and the process proceeds to the recognition process for the next character. When the sum of absolute values of {ΔQ _n ^* } _m values is smaller than a certain threshold ε, {ΔQ _n ^* } _m
Q _n ^*} the _m value as it is stored as a correction value {Delta] Q _{n} m} (step 33).

[Equation 13] If the correction value has already been calculated and stored in the learning mode,
The addition value dQ of the absolute value of the {ΔQ _n ^* } _m value is a threshold value ε.
In the following cases, the correction values are averaged by the following equation (21) and stored in the memory in the correction determination unit 11 (step 33). Addition value dQ of _{{ΔQ n} m = [{} ΔQ n} m + {ΔQ n *} m] / 2 ··· (21) the absolute value of {Delta] Q _n ^*} value, when above a certain threshold value ε is , The process proceeds to the recognition process for the next character without changing the already stored correction value.

(Ii) Extraction of Correction Value in Next Candidate Selection Mode (Steps 40-43) When the next candidate is selected and confirmed from the characters written by the writer, the character of the writer is selected from the first ranked character. To extract the features of. It is calculated by the equation (17) (step 40), and FIG.
The correction Q-value distance value {ΔQ _n ^* } _m value for each candidate character stored in the memory of the correction determination unit 11 as described above
The character as the rank is searched and the correction Q value distance value is extracted as the number of partial patterns (step 41). {Δ in equation (20)
Q _n ^*} it is determined whether the absolute value smaller than the threshold ε there is additional value dQ of the _m value (step 42), the following processing based on the determination result. If the correction value has not been calculated in the next candidate selection mode, the formula (2
When the addition value dQ of the absolute value of {ΔQ _n ^* } _m value in 0) is larger than a certain threshold value ε, the correction value is stored because the distance is extremely large and there is a high possibility that an operation error or a writing error has occurred. Instead, the buzzer, warning display, etc. are displayed to the writer, and the process proceeds to the recognition process for the next character. When the sum of the absolute values of the {ΔQ _n ^* } _m values is smaller than a certain threshold value ε, the {ΔQ _n ^* } _m value is stored as the correction value {ΔQ _n } _m as it is (step 4
3). When the correction value has already been calculated and stored in the learning mode or the like, the addition value dQ of the absolute value of the {ΔQ _n ^* } _m value is
When the threshold value is less than or equal to a certain threshold value ε, the correction values are averaged by the equation (21) and stored in the memory in the correction determination unit 11 (step 43). When the addition value dQ of the absolute value of the {ΔQ _n ^* } _m value is equal to or larger than a certain threshold value ε, the correction value already stored is not changed and the process proceeds to the recognition process of the next character.

(Iii) Extraction of correction value when the next candidate is determined (steps 50 to 53) The operation of transferring the character string written by the writer to another mode, for example, after confirming the character string written in word units, to the sentence creation mode. When, the candidate character is determined, and the character characteristics of the writer are extracted from the determined first character. It is calculated by the equation (17) (step 50), and as shown in FIG.
The correction Q-value distance value {ΔQ _n ^* } _m value for each candidate character stored in the internal memory is used to search the character ranked first and the correction Q-value distance value is extracted as the number of partial patterns (step 5).
1). It is determined whether or not the added value dQ of the absolute values of the {ΔQ _n ^* } _m values in the equation (20) is smaller than a certain threshold value ε (step 52), and the following processing is performed based on the determination result. already,
If the correction value is not calculated when the next candidate is determined, as in the learning mode, when the addition value dQ of the absolute value of the {ΔQ _n ^* } _m value in Expression (20) is larger than a certain threshold value ε, the distance is If it is highly possible that an operation error or a writing error is extremely large, the correction value is not stored, the buzzer, a warning display, etc. are displayed to the writer, and the process proceeds to the recognition process of the next character. In addition, {ΔQ _n ^* }
_{When the} sum of the absolute values of _m values is smaller than a certain threshold ε, {Δ
Q _n ^*} _m value as it is stored (step 53). If the correction value has already been calculated and stored in the learning mode or the like, and the addition value dQ of the absolute value of the {ΔQ _n ^* } _m value is less than or equal to a certain threshold value ε, the correction value is calculated by the equation (21). Average,
The data is stored in the memory in the correction determination unit 11 (step 5
3). When the added value dQ of the absolute values of the {ΔQ _n ^* } values is equal to or more than a certain threshold value ε, the correction value already stored is not changed and the process proceeds to the recognition process of the next character.

As described above, in the present embodiment, in the learning mode, when the next candidate is selected, or when the candidate is determined, the writing data of a predetermined character or the distance value of the Q value which is the characteristic parameter of the first character is used. Since the correction processing unit 13 that extracts the characteristic of the writing character peculiar to the writer and corrects it as the correction value at the time of matching is provided, the dictionary capacity can be increased even for the habit character peculiar to the writer due to inclination or the like. However, there is an advantage that an error due to the inclination of characters can be corrected and the identification rate can be improved. The present invention is not limited to the above embodiment, and various modifications can be made.
The following are examples of such modifications. (1) In the characteristic parameter Q-value matching, an example is shown in which the matching distance is sequentially stored for each candidate character in matching. However, when the number of recognition target characters increases and the memory increases, the learning mode, the next candidate selection, and the corresponding character are selected. It is also possible to search the character dictionary and the partial pattern dictionary again and calculate the matching distance {ΔQ _n ^* } _m . (2) Rather than having a correction value for the characteristic parameter Q value for each partial pattern, 16 correction values for each character {Δ
It is also possible to _{Q n} (n = 1~16)} . At this time, for the partial pattern distance value {ΔQ _n } _m , the number of strokes BS _{m of} each partial pattern is increased according to the following equation (22) in order to increase the weighting as the amount of information increases, that is, the number of strokes increases.
Are weighted to obtain a correction value {ΔQ _n }.

[Equation 14] Then, by using this correction value {ΔQ _n }, at the time of correction, each Q _n value of the partial pattern of the handwritten input character is multiplied, and the relative correction value is converted into an absolute value by weighting the stroke number of the partial pattern. Equation (14)
Instead of, it is corrected by the following equation (23). ^{_{{Q n '*} m =}} {Q n * + (ΔQ n · Q n *) BS m / N} m ··· (23) Thus, of having a correction value for each partial pattern for characters Instead, the capacity for storing the correction value can be reduced by averaging the characters by weighting with the number of strokes of the partial pattern. (3) It is also possible to correct the registered characteristic parameter Q value of the registered character by the correction value, and match the corrected registered pattern Q value and the written character characteristic parameter Q value.

[0032]

As described in detail above, the first to third aspects
According to the invention, a correction unit is provided, and the partial pattern Q value matching processing unit compares the characteristic parameter Q value corrected by the correction unit with the registered characteristic parameter Q value, or the characteristic parameter Q value is corrected by the correction unit. The registered registered characteristic parameter Q value is compared. Therefore, even for a habit character peculiar to a writer due to inclination or the like, an error due to the inclination or the like of the character can be corrected without increasing the dictionary capacity, and the identification rate can be improved.

[Brief description of drawings]

FIG. 1 is a functional block diagram of an online character recognition device according to a first embodiment of the present invention.

FIG. 2 is a functional block diagram of an online character recognition device according to a second embodiment of the present invention.

FIG. 3 is a functional block diagram of an online character recognition device according to a third embodiment of the present invention.

FIG. 4 is a functional block diagram of an online character recognition device according to a fourth embodiment of the present invention.

5 is a flowchart for explaining the operation of the online character recognition device of FIGS. 1 to 4. FIG.

FIG. 6 is a diagram showing a configuration example of a character dictionary.

FIG. 7 is a diagram showing a configuration example of a partial pattern dictionary.

FIG. 8 is a diagram for explaining the preprocessing of FIG.

FIG. 9 is a diagram for explaining stroke encoding processing.

FIG. 10 is a diagram for explaining vectors between partial patterns.

FIG. 11 is a diagram for explaining a partial pattern vector of a handwritten character example “Akira”.

FIG. 12 is a diagram showing a storage format of a matching distance {ΔQ _n ^* } _m .

[Explanation of symbols]

1 Tablet 2 Pre-Processing Part 3 Feature Point Extracting Part 4 Stroke Encoding Part 5 Major Class Part 6 Medium Classification Part 7 Partial Pattern Q Value Matching Part 8 Partial Pattern Stroke Code Matching Part 11 Correction Judgment Part 12, 22, 32, 42 Correction Processing unit 13, 23 Correction control unit

Claims (3)

[Claims] 1. A preprocessing unit that removes unnecessary data in a coordinate data string written by a writer by a writer and performs linearization processing, and a written character from a coordinate data string linearized by the preprocessing unit. A feature point extraction unit that extracts the features of the strokes that make up the stroke, a stroke encoding unit that encodes each stroke according to the positional relationship of the feature points extracted by the feature point extraction unit, and a portion that is written as a series in writing. As a partial pattern, the characteristic parameter Q value of the partial pattern is calculated from the output of the stroke encoding unit, and the registered characteristic parameter Q value and the characteristic parameter Q value calculated from the registered partial pattern of the registered character registered in advance. And a partial pattern Q-value matching unit that compares the partial pattern Q value with the output data of the feature point extraction unit or the stroke coding unit in advance. In the online character recognition device for performing character recognition based on the comparison result of the registered character recorded with the registered pattern or the comparison result of the partial pattern Q value matching processing unit, the characteristics of the handwritten character written and input by the writer. The relative distance value between the parameter Q value and the registered characteristic parameter Q value of the registered character corresponding to the writing character is stored as a correction value, and when the writing input is performed by the writer, the writing character of the writing input character is input. A correction unit that corrects the characteristic parameter Q value or the registered characteristic parameter Q value of the registered character based on the stored correction value is provided, and the partial pattern Q value matching processing unit is corrected by the correction unit. The parameter Q value is compared with the registered characteristic parameter Q value, or the characteristic parameter Q value and the registration corrected by the correction unit are registered. Online character recognition apparatus characterized by being adapted to compare the symptom parameter Q value. 2. A learning mode in which a writing character of a writer corresponding to the registered character is learned in advance by allowing the writer to input the written character, wherein the correction unit has a characteristic of the writing character that the writer has previously input. The online according to claim 1, wherein a relative distance value between the parameter Q value and the registered characteristic parameter Q value of the registered character corresponding to the written character is calculated, and the relative distance value is stored as a correction value. Character recognizer. 3. The registered character corresponding to the written character of the writer is character-recognized as a candidate character, and when the character-recognized candidate character is confirmed by the writer, the correction unit is configured to detect the written character. 2. The relative distance value between the characteristic parameter of the above and the registered characteristic parameter of the candidate character is calculated, and the relative distance value is stored as a correction value.
Online character recognition device described.