JPH08249424A - Handwriting recognition device - Google Patents

Abstract

PURPOSE: To recognize a handwritten character including a (running hand character) and a (simplified form character) stably at a high recognition rate. CONSTITUTION: A dictionary storage part 10 stores a reference pattern in unit of sub-pattern easy to write with one stroke. A stroke processing part 5 extracts a feature point representing the feature of a curvature part in advance before normalization. A feature resampling part 8 performs the normalization, interpolation and noise elimination at every chain of stroke, and interpolates an off- stroke section with the feature point. A character recognition part 9 performs DP matching using the reference pattern in unit of sub-pattern. In such a way, the feature of the curvature part is prevented from being lost due to the normalization by sampling the feature point of the curvature part before the normalization. Also, the accuracy of the DP matching is improved and a time is reduced by the feature point extracted by using both an equal division method and an angle method and a standard pattern for every sub-pattern. Also, it is possible to deal with the (running hand character) and the (simplified form character) by performing write with one stroke by interpolating the off-stroke section.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a handwriting recognizing device which is used in a word processor, a personal computer or the like to recognize a handwritten character or the like input from a tablet by a pen.

[0002]

2. Description of the Related Art At present, a method of using a keyboard for inputting characters to an electronic notebook, a word processor, a personal computer or the like is mainly used. However,
Character input using this keyboard requires a certain amount of skill, and there is a problem that the area for the keyboard cannot be secured because the target device for inputting characters is becoming smaller like an electronic notebook. Is happening. Therefore, online handwritten character input using a tablet or the like has begun to attract attention.

On-line handwritten characters When recognizing an input character, the handwritten character written on the tablet is usually represented by coordinates (x, y) on the tablet. Then, the stroke of the handwritten character is used as a basic unit for recognition, and feature extraction is performed for each stroke. Then, assuming that the character is a set of strokes, the feature sequence of strokes associated with the category number is registered in the stroke dictionary as a standard pattern, and the feature sequence extracted from the recognition-target handwritten character is collated with the standard pattern. Then, the category of strokes forming the recognition target handwritten character is recognized. Then, the input character is recognized by recognizing the combination of strokes using the character dictionary. When registering a standard pattern in the dictionary, a character pattern is divided into one or more sub patterns such as "kihen", "gonben", or "furutori" and registered as subpatterns. There is also a method.

As the feature extraction method for each stroke,
There are the following feature extraction methods. (1) The stroke start point to end point is divided into a predetermined number, and the features of each division point are extracted. (2) The stroke start point to end point are divided at equal intervals, and the features of each division point are extracted. (3) A stroke is traced, a bending point having a predetermined angle or more is detected as a feature point, and the feature of each feature point is extracted.

The above classification is a classification in which conventional handwriting recognition is viewed from the feature extraction method. On the other hand, from the perspective of the handwriting recognition method as a whole, a handwriting recognition device has been proposed that recognizes from "regular characters" to "some continuous characters". Is expected. However,
Despite such research being carried out for a long time, none that has reached a practical level has yet appeared. In addition, in the case of online handwriting recognition, in addition to the above-mentioned problem of stroke deformation, there is a big problem to be solved, that is, “stroke order variation”.

[0006] The following studies have been made to deal with the "continuation character" and the "collapsed character" that have been proposed so far. [i] Online handwritten character recognition using deformation vector prediction based on deformation vector field (IEICE Transactions '86 /
12 Vol.J69-D No.12) After determining the stroke correspondence between the input pattern and the standard pattern in the dictionary, the corresponding feature points are determined for each stroke pair by DP matching. Then, the predicted deformation vector obtained from the deformation vector field between the corresponding feature points is superimposed for each attribute to deform the standard pattern, and DP matching is performed between the deformed standard pattern and the input pattern to obtain the distance. Of.

[Ii] Real-time recognition of comb characters by fuzzy inference (The Institute of Electronics and Communication Engineers PRU 89-29 (1989)) Fourier transform is performed on the x, y coordinate sequence for each stroke,
The fuzzy data is obtained based on the Mebership function of each coefficient (3 in x and y in the literature). Then, similarly to the fuzzy data, the dictionary data represented by the fuzzy data is recognized for each stroke. As the dictionary data, all possible continuous character patterns are prepared for each character.

[Iii] Attributed String Matching by S
plit-and-Merge for On-Line ChineseCharacter Recogn
ition (IEEE TRANSACTIONS ON PATTERN ANALYSIS AND M
ACTINEINTELLIGENSE VOL 15, NO.2 FEBRUARY 1993) For example, under a fixed stroke order condition that only the same person uses, the offstroke part of the input is associated with the offstroke part of the dictionary (after all, the strokes are associated with each other). Then, detailed DP matching is performed for each corresponding stroke to obtain the distance.

As described above, the conventional "continuous character" and "collapsed character"
The online recognition method of can be roughly divided into the following three basic approaches. (4) The stroke order is a free method, and the stroke order is absorbed by the stroke correspondence, and is recognized by using detailed features and methods. (5) It is a fixed stroke order method, and is recognized under the condition that the stroke order can be ignored, such as the same user, or limited to a certain number of variations.

A method has also been proposed in which the captured character data is treated as image data and a pattern recognition of a handwritten character is performed by using a character recognition method in an OCR (optical character recognition device). For example, if the input character image is 12 ×
The number of display pixels in each cell is divided into 12 cells, the distribution of the display pixels is used as a characteristic pattern, and the input character is recognized by collating with the standard pattern registered in the dictionary.

[0011]

However, the above conventional handwritten character recognition device has the following problems. (A) In the feature extraction method (1) for each stroke, since the entire stroke is divided into a predetermined number, when the target stroke is very long, the feature points become too long, In some cases, the shape of the original stroke cannot be reproduced due to the sequence of. In that case, there is a problem in that the degree of coincidence decreases during collation with the stroke dictionary and a high recognition rate cannot be obtained.

(B) Feature extraction method for each stroke
In (2), since the entire stroke is divided at equal intervals, the number of feature points becomes very large when the length of the target stroke is very long. By setting the division interval to be short within an appropriate range, the original stroke shape can be faithfully reproduced from a series of feature points even if a complicated Chinese character or the like is input by handwriting in one stroke. However, on the other hand, since there are many characteristic points, it takes time to collate with the stroke dictionary.

(C) Feature extraction method for each stroke (3)
In the above, since the bending point having a predetermined angle or more is detected as the characteristic point, the number of detected characteristic points depends on the threshold value of the angle. As a result, when the angle threshold value is increased, the number of feature points is decreased and the matching time is shortened, but the recognition rate is reduced. On the other hand, when the angle threshold is reduced, the recognition rate can be increased, but there is a problem that the number of feature points increases and the matching time increases. In addition, even if the threshold value of the angle is appropriately set, the position of the characteristic point is not stable due to individual differences in curvature in the case of a stroke having many curved lines such as a kana character. Therefore, it is difficult to obtain a high recognition rate when the coordinates of the feature points are used as the feature amount.

(D) In the above feature extraction methods (1) and (2), when feature extraction is performed from an input coordinate sequence consisting of a plurality of normalized strokes, the extracted feature amount indicates the feature of the input character. It may disappear. Therefore, it is necessary to extract a large number of feature points in order to surely extract the feature points that accurately represent the features, and it takes time for dictionary matching.
On the other hand, if the number of extracted feature points is reduced, there is a problem that the probability of accurately expressing the feature decreases, and the recognition rate decreases. For example, in FIG. 28A, consider a case where a feature is extracted from a normalized input coordinate sequence (displayed by a solid line).
In that case, feature points (“points”) extracted from the input coordinate sequence
28), there are characteristic points as if they are arranged on a straight line (one row above) as shown in FIG. 28 (b), even though there is a bent portion (α) in the input coordinate sequence. To be extracted. In that case, the recognition result of the handwritten character radical "Gyoninben" becomes "Ninben".

(E) In each of the above-mentioned feature extraction methods (1) to (3), since the stroke is recognized as a unit, the strokes that the writer does not intend, "entry" and "bounce", are not included in the character configuration. There is a problem that the recognition rate is lowered because the strokes that cannot be distinguished cannot be distinguished from each other. Further, the characteristic of the bent portion of the stroke may disappear by normalization, and in that case, it is impossible to accurately extract the characteristic of the bent portion from the normalized stroke.

Further, the above-mentioned conventional "continuous character" and "collapsed character" online recognition methods have the following problems. (F) In the stroke order free method of the above online recognition method (4), a. It takes too much time to process the stroke association itself, and the processing speed is slow as a whole. b. If there is excessive deformation or omission, the performance of stroke correspondence deteriorates, and the recognition rate decreases. The problem arises. Here, the final purpose is "character recognition".
However, there is a problem that the amount of stroke matching processing before character recognition becomes large, for example, for online recognition of "continuation characters" and "collapsed characters" in a poor hardware configuration such as an electronic notebook. It will not be possible.
Further, the problem that the stroke matching performance deteriorates due to excessive deformation or omission makes it difficult to expect the development of online recognition of "continuous characters" from online recognition of "collapsed characters".

(G) In the fixed stroke order method of the above-mentioned online recognition method (5), although the recognition performance of "successive characters" and "collapsed characters" is relatively good, there is a problem that it is extremely weak against fluctuations in the stroke order. The simplest and only solution to this problem is to have all stroke order variations in the dictionary, but since the dictionary capacity monotonically increases by the number of variations, it is virtually impossible to build a general system. I will end up.

(H) When the character recognition method using OCR is used, since the pattern of the entire character is the recognition target, many characters are required to accurately recognize the "continuation character" or the "collapsed character". It is necessary to register the modified pattern as a standard pattern in the dictionary, and it takes a long time for matching. Therefore, there is a problem in that, when trying to recognize the stroke unit, the stroke break is not clear and the recognition rate is lowered.

(I) In addition, the online recognition method (4),
In common with (5) and the character recognition method using OCR, there is a problem that the recognition process cannot be started unless the user's input is completed because it is a character-by-character recognition process. Normally, human writing takes about 2 seconds. Therefore, performing the pre-recognition process using this writing time is particularly effective in a poor device such as an electronic notebook. However,
It is difficult for the conventional recognition method to realize the pre-recognition process using this writing time.

SUMMARY OF THE INVENTION An object of the present invention is to provide a handwriting recognition device capable of stably recognizing handwritten characters including "continuation characters" and "collapsed characters" at a high recognition rate at a high speed.

[0021]

In order to achieve the above object, the invention according to claim 1 is a tablet for generating a coordinate data string formed by arranging the coordinates of handwriting input by handwriting in time series, and the coordinates. A feature extraction unit that extracts the feature points that represent the features of the handwriting pattern from the data sequence and the feature amount of the feature points, and the feature pattern that is a sequence of feature points that has the extracted feature amount and the dictionary are matched to make the handwriting. In a handwriting recognition device having a recognition unit for recognizing a pattern and a result display unit for displaying a recognition result by the recognition unit, the dictionary has subpatterns formed by separating a representative pattern into patterns that are easy to write in a statistical manner and each pattern. This is a sub-pattern dictionary in which a sequence of feature points having feature quantities extracted from a modified pattern of a sub-pattern is used as a standard pattern. When the feature points and the feature amount are extracted from the data sequence, the recognition unit includes off-stroke interpolation means for interpolating the off-stroke between strokes with the feature points, and the recognition unit is characterized by the strokes and the off-strokes. It is characterized in that a feature point sequence having a quantity is regarded as one feature pattern and collated with the sub-pattern dictionary.

According to a second aspect of the present invention, in the handwriting recognition device according to the first aspect of the present invention, the feature extraction unit determines the feature points of the bent portion in the handwriting pattern from the coordinate data string generated by the tablet. The first feature extraction unit that extracts the feature amount and the feature point sequence extracted by the first feature extraction unit are normalized, and the bent portion is calculated based on the normalized feature point sequence and the normalized value of the coordinate data. A feature is that a feature point that approximates the feature points at equal intervals and a second feature extraction unit that extracts a feature amount are provided.

According to a third aspect of the invention, in the handwriting recognition device of the first aspect of the invention, the feature amount of the feature points on the feature pattern obtained by the feature extraction section and the feature point on the standard pattern are obtained. A table storing unit that stores a table in which the feature amount and the value of the difference between the two feature amounts are associated with each other, and the recognizing unit includes each feature point on the feature pattern and each feature point on the standard pattern. Based on the feature amount of the feature point pair, the table is drawn to find the difference in the feature amount for each feature point pair, and the distance between the feature pattern and the standard pattern is calculated based on the value of each difference. It is characterized in that it is provided with a distance calculating means.

The invention according to claim 4 is the handwriting recognition apparatus according to claim 1, wherein the recognition section is
It is characterized by including a DP matching means for performing a DP matching process in which a variation weight is given to each feature point on a pattern to be matched.

According to a fifth aspect of the present invention, in the handwriting recognition apparatus of the fourth aspect of the invention, the feature extracting unit and the DP matching unit are controlled, and the feature extracting unit extracts a learning pattern prepared in advance by the feature extracting unit. Generate a learning feature pattern consisting of a sequence of feature points with features,
The DP matching processing of the learning feature pattern and the standard pattern is performed by the DP matching means to obtain the matching distance for the point on the matching path, and the feature distance for all learning is associated with each feature point in the standard pattern. The present invention is characterized in that a variation weight calculation unit that calculates the degree of variation in matching distance and uses the variation weight as the variation weight is provided.

According to a sixth aspect of the present invention, in the handwriting recognition apparatus according to the fourth aspect of the invention, the distance from the feature point on the feature pattern extracted by the feature extraction unit to the next feature point and the feature A feature is that a variation weight calculation unit that calculates a value of a ratio to a distance connecting all feature points in a pattern in an input order and sets the value as the variation weight of a section related to the feature point is provided.

According to a seventh aspect of the present invention, in the handwriting recognition apparatus according to the first aspect of the present invention, a representative pattern configuration table in which a representative pattern and sub-patterns forming the representative pattern are associated with each other is stored. With the pattern configuration table storage unit, the recognition unit,
Sub-pattern recognition means for recognizing the handwriting pattern for each sub-pattern by comparing the characteristic pattern with a sub-pattern dictionary, a combination of sub-pattern candidates obtained as a result of recognition by the sub-pattern recognition means, and the representative pattern It is characterized in that a handwriting pattern recognition means for recognizing the handwriting pattern by collating with the configuration table is provided.

According to an eighth aspect of the present invention, in the handwriting recognition apparatus according to the seventh aspect of the present invention, the feature extraction unit has a straight line direction connecting the coordinate points in the coordinate data sequence in a predetermined value. Bent portion feature extraction means for extracting the changing points as the feature points, straight line feature extraction means for extracting the feature points that approximate the feature points extracted by the bent portion feature extraction means at equal intervals, and The present invention is characterized by including a direction value calculation unit that calculates, as one of the feature amounts, a direction value that represents the direction to the next feature point in each feature point extracted by the feature extraction unit.

According to a ninth aspect of the invention, in the handwriting recognition device of the eighth aspect of the invention, the approximate straight line number of the handwriting pattern is calculated from the number of the characteristic points of the bent portion extracted by the bent portion feature extracting means. The number of strokes of the feature pattern is defined from the approximate number of straight lines and the number of strokes of the handwriting pattern, and the sub-pattern dictionary is provided for each standard pattern with the original pattern of the standard pattern. And has header information consisting of the number of strokes of the representative pattern, and the recognizing unit selects from the sub-pattern dictionary only standard patterns whose header information is the number of strokes within the range defined by the stroke number limiting unit. It is characterized in that a feature pattern selecting means is provided.

The invention according to claim 10 is the invention according to claim 8.
In the handwriting recognition device of the invention according to, the sub-pattern dictionary, for each standard pattern, has header information consisting of the effective range of the direction value from the final feature point of the standard pattern to the leading feature point in the next stroke, The recognition unit determines that only a standard pattern having header information as an effective range in which the direction value from the final feature point in the feature pattern calculated by the direction value calculation means to the first feature point in the next stroke is used as the valid standard pattern. It is characterized in that it is provided with an effective standard pattern determining means.

The invention according to claim 11 is the invention according to claim 7.
In the handwriting recognition device of the invention according to, the recognition unit,
A standard pattern selection means for obtaining a sub pattern that can follow the sub pattern candidate chain of the recognized handwriting pattern by referring to the representative pattern configuration table and selecting only the standard pattern of the obtained sub pattern from the sub pattern dictionary. It is characterized by having.

The invention according to claim 12 is the invention according to claim 1.
In the handwriting recognition device of the invention according to, while the coordinate data string generated by the tablet is written,
The input coordinate buffer consisting of a dual-port memory from which the written coordinate data sequence is read by the feature extraction unit and the feature pattern extracted by the feature extraction unit are written, while the written feature pattern is A feature buffer composed of a dual port memory read by the recognition unit, a feature extraction process by the feature extraction unit and a recognition process by the recognition unit, and a coordinate data string by the tablet by controlling the feature extraction unit and the recognition unit. It is characterized by having a control unit that executes the generation process in parallel.

[0033]

In the invention according to claim 1, when the tablet is handwritten, a coordinate data string is generated. Then, when the generated coordinate data string is a coordinate data string of a plurality of strokes, the off-slot talk is interpolated by the feature points by the off-slot talk interpolation means of the feature extraction unit. In this way, the feature extraction unit extracts the feature points and the feature amount of each stroke and the off-stroke from the coordinate data sequence of a plurality of strokes. Then, the recognition unit considers the sequence of feature points having the feature amounts related to the plurality of strokes and the off-slot talk as one feature pattern, performs matching with the sub-pattern dictionary, and recognizes the handwriting pattern. It

In this way, as a result of off-slot talk being interpolated at the time of feature extraction, a standard pattern of a modified pattern having "entry", "bounce", "continuation", etc. is registered in the sub-pattern dictionary. , It is easy to recognize the handwriting pattern of "continuation character" or "collapsed character".

Further, in the invention according to claim 2, when the coordinate data string from the tablet is input to the feature extraction unit, first, the first feature extraction unit extracts the bent portion in the handwriting pattern from the coordinate data string. Feature points and feature quantities are extracted. Next, the second feature extraction unit normalizes the feature point sequence of the bent portion, and extracts feature points and feature amounts that approximate the normalized feature points at equal intervals. Thus
By extracting the feature points and the feature amount of the bent portion before the normalization, even if the feature of the bent portion disappears due to the normalization, the feature points and the feature amounts representing the bent portion can be reliably extracted.

Further, in the invention according to claim 3, the distance calculating means of the recognizing unit calculates the table based on the characteristic amount of the characteristic point pair consisting of the characteristic points on the characteristic pattern and the characteristic points on the standard pattern. Is subtracted to obtain the difference between the feature quantities associated with the feature point pair. Then, the distance between the characteristic pattern and the standard pattern is calculated based on the value of each difference. In this way, the matching between the characteristic pattern and the standard pattern by the recognition unit is quickly performed by a simple operation of pulling a table.

Further, in the invention according to claim 4, the DP matching means of the recognition section performs the DP matching processing in which the variation weight is given to each feature point on the pattern to be matched. In this way, the time extension characteristic during DP matching is changed according to the appearance of the feature along the stroke of the handwriting pattern.

Further, in the invention according to claim 5, the feature extracting section and the DP matching means are controlled by the variation weight calculating section to generate a learning characteristic pattern from a learning pattern prepared in advance, and the learning characteristic pattern is generated. By the DP matching process between the pattern and the standard pattern, the matching distance related to the points on the matching path is obtained, and the degree of variation in the matching distance related to all the learning characteristic patterns is calculated as the variation weight for each characteristic point in the standard pattern. It is calculated. In this way, DP matching is performed in which the section in which the characteristic points are stably obtained is emphasized.

Further, in the invention according to claim 6, the variation weight calculation unit calculates the distance from the feature point on the feature pattern extracted by the feature extraction unit to the next feature point and all the feature points in the feature pattern. The value of the ratio to the distance connecting in the input order is calculated as the variation weight of the section related to the feature point. In this way, DP matching in which the characteristic of the bent portion in the handwriting pattern is emphasized is performed.

In the invention according to claim 7, the sub-pattern recognizing means of the recognizing unit collates the characteristic pattern with the sub-pattern dictionary to recognize the handwriting pattern for each sub-pattern. Then, the handwriting pattern recognition means collates the combination of the subpattern candidates obtained as a result of the recognition by the subpattern recognition means with the representative pattern configuration table to recognize the handwriting pattern. In this way, the recognition process is performed in units of sub-patterns that are statistically easy to draw with one stroke, and the handwriting pattern is quickly recognized with a small number of recognition candidates.

Further, in the invention according to claim 8, the bending point feature extraction means of the feature extraction unit detects a change point at which the direction of the straight line connecting the coordinate points in the coordinate data sequence changes in a predetermined value or more. It is extracted as a characteristic point of the bent portion. Further, the straight line feature extraction means extracts feature points that approximate the feature points of the bent portion at equal intervals. Then, the direction value calculation means calculates the direction value representing the direction to the next feature point among the feature points extracted by the feature extraction means as one of the feature amounts. Thus, the characteristic points are obtained by the angle method and the equal division method,
The direction value of each feature point is used as one of the feature amounts used at the time of dictionary matching.

Further, in the invention according to claim 9, the stroke number limiting unit obtains the approximate straight line number of the handwriting pattern from the number of characteristic points of the bent portion, and the approximate straight line number and the stroke number of the handwriting pattern. From the range of the number of strokes of the characteristic pattern is determined. Then, the characteristic pattern selecting means of the recognizing unit selects a standard pattern having the number of strokes within the defined range as header information from the sub-pattern dictionary, and uses only the selected standard pattern to perform the recognition. The above-mentioned dictionary collation is quickly performed by the department.

According to the tenth aspect of the invention, the direction value calculating means of the feature amount extracting section calculates the direction value from the final feature point in the feature pattern to the leading feature point in the next stroke. Then, the valid standard pattern determination means of the recognition unit determines that the standard pattern having the valid range in which the calculated direction value is included as header information is the valid standard pattern. Then, the recognition unit uses only the standard pattern determined to be the effective standard pattern to quickly perform the dictionary matching.

According to the eleventh aspect of the invention, the standard pattern selection means of the recognition section refers to the representative pattern configuration table to identify subpatterns that can follow the subpattern candidate chain of the recognized handwriting pattern. A standard pattern of the obtained sub-patterns is selected from the sub-pattern dictionary. Then, only the selected standard pattern is used, and the recognition unit quickly performs the dictionary matching.

According to the twelfth aspect of the invention, the feature extraction section and the recognition section are operated in parallel with the tablet under the control of the control section. Then, the coordinate data string generated by the tablet will be
It is written from the input port to the input coordinate buffer consisting of memory. Then, the written coordinate data string is read from the output port by the feature extraction unit in parallel with the writing, and the feature extraction is performed. further,
The extracted feature pattern is written from the input port to the feature buffer composed of the dual port memory. Then, the written feature pattern is read from the output port by the recognition unit in parallel with the writing,
Collation with the dictionary is performed. In this way, the feature extraction process and the recognition process are performed in parallel with the coordinate data string generation process.

[0046]

The present invention will be described in detail below with reference to the embodiments shown in the drawings. It should be noted that the first to fourth examples mainly relate to the extraction method of the characteristic points and the characteristic amounts, and the fifth example mainly relates to the recognition method. <First Embodiment> FIG. 1 is a block diagram of a handwriting recognition apparatus in this embodiment. This handwriting recognition device is roughly configured by a recognition unit 1, a tablet 2, and a result display unit 3.

The tablet 2 is formed as a display-integrated type, for example, and converts handwriting data pen-input by a writer into a coordinate data string represented by an electric signal. Recognition unit 1
Recognizes a handwritten character or the like by performing a handwriting recognition process, which will be described in detail later, based on the obtained coordinate data string.
The result display unit 3 is composed of a liquid crystal display device or the like, and includes a recognition unit 1.
Display the recognition result by.

The recognition unit 1 is composed of a stroke end determination unit 4, a stroke processing unit 5, a feature storage unit 6, a character end determination unit 7, a feature re-extraction unit 8 and a character recognition unit 9. Hereinafter, the function and operation of each unit will be sequentially described.

The stroke end judging section 4 fetches the coordinate data sequence sequentially sent from the tablet 2,
Pen-up is detected on the basis of the coordinate data string taken in and a stroke end code is added. The stroke processing unit 5 extracts the coordinate points of the bent portion of the stroke as the characteristic points based on the coordinate data string forming the stroke. The coordinate data string of the feature point to which the stroke end code is added thus obtained is stored in the feature storage unit 6. The character end determination unit 7 detects the end of one character based on the coordinate data string from the tablet 2 and outputs the character end information.

The feature re-extracting section 8 includes a character end determining section 7
Based on the character end information from, the feature point sequence for one character is cut out from the feature point sequence stored in the feature storage unit 6, a stroke chain is generated and normalized, and the distance and direction between each feature point are calculated. Add / delete feature points to align. Further, the feature points are inserted on the assumption that there is a stroke even in a section where the pen is not actually input (hereinafter referred to as an off stroke). The character recognition unit 9 includes a feature re-extraction unit 8
In order to recognize the sub-pattern by performing the weighted DP matching between the feature point sequence finally obtained in step 3 and the standard pattern registered in the dictionary storage unit 10, the character dictionary registered in the dictionary storage unit 10 Recognize handwritten input characters using. Then, special classification processing is performed on characters that cannot be distinguished by their shapes.

Here, the standard patterns registered in the dictionary storage unit 10 are as follows so as to be able to deal with "entry", "bounce", "continuation character" and "collapse character" of the pen. Will be generated. Generally, when characters are handwritten, there are the following rules of thumb.・ It is easy to write with a single stroke in a pattern that can be separated into left and right, such as Kanji characters "bias" and "straw".・ Low stroke number characters are easy to write with one stroke.・ Some characters are written empirically, experimentally, or instructively. Therefore, in the present embodiment, 1
The definition of the sub-patterns that make up the character pattern is defined as follows. -Basically, a pattern that can be separated into left and right is a sub pattern.・ Patterns with a low stroke count and patterns that are statistically easy to write with a single stroke are sub-patterns.

Sub-patterns are set according to such a definition, and further, the stroke order differences and deformation variations of each sub-pattern are prepared (hereinafter, individual stroke-order patterns and deformation variation patterns are also referred to as sub-patterns).
To do. Then, the off-stroke in each sub-pattern is interpolated, the same processing as the feature extraction processing by the stroke processing unit 5 and the feature re-extraction unit 8 described later is performed to extract the feature point and the feature amount, and the feature amount is added. The pattern of the extracted feature points is registered in the sub-pattern dictionary of the dictionary storage unit 10 as a standard pattern. FIG. 2 shows an example of the sub pattern. 2 (a) and 2 (b) are sub-patterns "words", and FIGS. 2 (c) and 2 (d) are sub-patterns "day".
2E is the sub-pattern “Haya”. The numbers in the figure are stroke numbers.

Here, as the sub-pattern in this embodiment, a pattern that can be basically divided into right and left is made one sub-pattern for the following reason. That is, as shown in FIG. 3, the patterns “rice” and “tree” in the pattern “fruit” that can be separated into upper and lower parts are generally easy to be “continued”.
Therefore, when "Tar" and "Thu" are registered as the sub-patterns, they will not be recognized when the handwritten pattern as shown in FIGS. 3B and 3C is input. However, in the present embodiment, the “result” is registered as one sub-pattern, and the difference in the stroke order is also used as the sub-pattern, so that a high recognition rate can be obtained even if “continuation” is performed. You can do it.

As described above, since a pattern that is statistically easy to write with one stroke is registered as one sub-pattern, its deformation variation is limited to some extent. Therefore, the number of deformation variations may be small even when one character is used as a unit for DP matching. That is, according to the sub-pattern of the present embodiment, the required storage capacity of the dictionary can be reduced as compared with the conventional case where a stroke or one character is used as a unit for DP matching. Conversely, since more patterns can be registered with the same storage capacity than before, the deformation variation can be set in more detail and a high recognition rate can be obtained.

As shown in FIG. 4, the character dictionary registered in the dictionary storage unit 10 is made by associating a character with a sub-pattern constituting this character, and
It is referred to when generating the character candidate by combining the sub-pattern candidates recognized by the P matching. In this embodiment, since the character is formed by the combination of the sub-patterns, which is a set of a plurality of strokes as described above, the number of constituent elements of the character is larger than that in the case of forming the character by the combination of strokes as in the conventional case. Can be greatly reduced, and the capacity of the character dictionary can be reduced.

FIG. 5 is a flow chart of the handwriting recognition processing operation executed by the handwritten character recognition device having the above configuration. Hereinafter, the handwriting recognition processing operation will be described in detail with reference to FIG.

In step S1, the tablet 2 determines whether or not there is a pen input. If there is a pen input, the process proceeds to step S2. In step S2, the tablet 2 converts the input handwriting pattern of the character into a coordinate data string. Then, the stroke end determination unit 4 determines the end of the stroke, and the coordinate data string to which the stroke end code is added is sent to the stroke processing unit 5. Here, the stroke end determination by the stroke end determination unit 4 is performed as follows. That is, the coordinate data sequence sequentially sent out from the tablet 2 is fetched, and the pen-up is detected when the fetching interval exceeds the scanning timing interval in the tablet 2, and it is determined that one stroke is completed. . Then, the stroke end code is added to the end of the coordinate data sequence that has already been fetched.

In step S3, the stroke processing unit 5
Thus, the stroke end code added by the stroke end determination unit 4 is referred to, the feature extraction processing is performed, and the feature point sequence extracted in stroke units is stored in the feature storage unit 6. Here, the stroke processing unit 5 performs stroke feature extraction as follows. That is, first,
A stroke end code is detected from the fetched coordinate data string, and the coordinate data fetched next to this stroke end code is used as the coordinates of the start point of the stroke. Next, paying attention to the starting point, the distance from this starting point to the next coordinate point is calculated. If this distance is less than or equal to the predetermined distance Nlmin (for example, 0.25 mm), the distance to the next coordinate point is calculated. Thus, the distance is the above-mentioned predetermined distance Nlmin.
The point of interest is moved to the coordinate point when it becomes larger. After that, similar processing is performed until the next stroke end code is fetched, the obtained target point is set as a feature point, and its coordinates are set as one of the feature amounts.

Next, as shown in FIG. 6A, the characteristic point P1 next to the starting point P0 in the characteristic point sequence forming the stroke
Is set as the point of interest, the difference between the direction from the starting point P0 to the point of interest (feature point P1) and the direction from the point of interest P1 to the feature point P2 is calculated, and the value of this difference is calculated from the predetermined angle Ndist (for example, 30 degrees). If it is smaller, the point of interest P1 is removed. Hereinafter, the above-mentioned point of interest is sequentially shifted to each of the feature points P2, P3, ..., Pn, ..., The direction from the feature point Pn-1 to the point of interest Pn and the direction from the point of interest Pn to the feature point Pn + 1. Is calculated, and the point of interest Pn is calculated based on the value of this difference.
Is determined. As a result, a feature point sequence as shown in FIG. 7B is created from the input coordinate sequence as shown in FIG. 7A. In this way, by extracting the characteristic points of the bent portion of the stroke before the characteristic is lost by the normalization, the characteristic of the bent portion is retained even after the normalization.

In step S4, the feature re-extracting section 8 determines whether or not one character has been input based on the character end information from the character end determining section 7. as a result,
If it is completed, the process proceeds to step S5. On the other hand, if not completed, the process returns to step S1 to continue the character input and the stroke feature extraction. The character end determination by the character end determination unit 7 is performed as follows.
That is, when the coordinate data sequence sequentially sent from the tablet 2 is fetched, one handwritten character is generated because a predetermined time Ntime (for example, 1 second) has elapsed from the last coordinate data fetch to the next coordinate data fetch. The end of the coordinate data sequence related to is detected. Then, the character end information indicating the end of the character is sent to the feature re-extracting section 8.

In step S5, the feature re-extracting unit 8 adds the feature point coordinate data stored in the feature storing unit 6 to the feature point of the stroke in one character section based on the character end information. The coordinate data string being read is read. Then, the feature re-extraction processing is performed as follows.

A plurality of feature points read from the feature storage unit 6 are arranged in the order of input strokes. Then, the strokes are sequentially connected from the first stroke to the final stroke, and normalized for each sequentially obtained stroke chain. This normalization is performed as follows. That is, the maximum coordinate value xmax of the x axis and the maximum coordinate value ymax of the y axis in the target stroke chain are obtained. Further, the minimum coordinate value xmin of the x-axis and the minimum coordinate value ymin of the y-axis are obtained. Then, according to the equations (1) and (2), the coordinates of each feature point are 0 to Nnorm on the x-axis and the y-axis, respectively.
It is normalized so as to be distributed in the range (for example, 128 dots). xn ′ = {(xn−xmin) / (xmax−xmin)} · Nnorm (1) yn ′ = {(yn−ymin) / (ymax−ymin)} · Nnorm (2) where xn: feature X coordinate of point n xn ': x coordinate of feature point n after normalization yn: y coordinate of feature point n yn': y coordinate after normalization of feature point n

The distance between the feature point of interest and the feature point immediately after it in each feature point normalized as described above is obtained. If this distance is equal to or greater than the predetermined distance Nfmax (for example, 4 mm), a feature point selected from the coordinate data sequence corresponding to the feature points and normalized is newly added to the feature point in the section. The distance between them is set to a predetermined distance Nflen (for example, 2 mm) or less. Thus, by interpolating between the characteristic points of the bent portion and performing equal-interval approximation, the characteristic points can be extracted more appropriately, and the subsequent DP matching time can be shortened and a high recognition rate can be obtained. Further, among the directional differences between the direction from the target feature point to the immediately preceding target feature point and the direction from the target feature point to the immediately subsequent feature point, when the distance is equal to or less than a predetermined distance Nfmin (for example, 1 mm). When the smaller value is smaller than the predetermined direction difference Nddst (for example, 90 degrees), the target feature point is regarded as noise and is deleted from the feature point sequence.

Here, when the above-mentioned stroke chain is formed, if the above-mentioned off-stroke exists between the strokes, it is assumed that the stroke actually exists in the off-stroke section, and the off-stroke is concerned. Feature points that divide the section into equal parts or less Nflen or less are added. In this way, by interpolating the off-stroke section, it is possible to accurately extract the "entry" or "bounce" portion or the "continuation character" or "collapse character" that the writer does not intend.

The above-mentioned feature point addition / deletion operations are
The process is repeated until no new feature point is added. As a result, a feature point sequence having three elements of x-coordinate, y-coordinate and absolute direction as shown in FIGS. 7 (c) and 7 (d) is created from the feature point sequence as shown in FIG. 7 (b). Is done. still,
FIG. 7C shows an example of a series of characteristic points in the chain of the first stroke a1 and the second stroke b1, and the broken line e shows the interpolated off-stroke section. The absolute direction is the direction of each feature point to the next feature point, and is represented by the degree of deviation from the reference direction.

In step S6, the character recognition unit 9 performs a weighted DP matching process between the feature point sequence obtained from the stroke chain and the standard pattern registered in the sub-pattern dictionary of the dictionary storage unit 10. The matching distance (hereinafter referred to as partial matching distance) Δd between the corresponding feature points of the characteristic pattern and the standard pattern used in the DP matching in the present embodiment is the Euclidean between the corresponding feature points of the characteristic pattern and the standard pattern. The distance ΔL and the direction difference ΔH are used to obtain the value by the equation (3). Δd = Wp · ΔL + Wd · ΔH (3) However, Wp: Weight related to distance (for example, Nwl between all feature points
en (= 1)) However, Wd: weight related to the direction (for example, Nwd between all feature points)
st (= 8))

The Euclidean distance ΔL is given by the equation (4). Further, the direction difference ΔH is a difference between the directions from the feature point immediately before the feature point of interest to the feature point immediately after,
It is given in (5). Here, the value of the direction difference ΔH between the start point and the end point is “0”. ΔL = {(xm−xn) ² + (ym−yn) ² } ^1/2 (4) ΔH = | dn−dm | (5) where xn: x coordinate yn of the characteristic point n on the characteristic pattern : Y coordinate of the feature point n xm: x coordinate of the feature point m on the standard pattern ym: y coordinate of the feature point m dn: feature point (n-1) immediately before the feature point n to the feature point (n + 1) immediately after the feature point n Direction dm: direction from the feature point (m-1) immediately before the feature point m to the feature point (m + 1) immediately after

Then, the sum of the distances along the matching path by the weighted DP matching using the partial matching distance Δd (hereinafter, simply referred to as "matching distance") D
Sub-pattern candidate lattices are generated based on

Here, the feature re-extracting process by the feature re-extracting unit 8 and the weighted DP matching by the character recognizing unit 9 are continuously performed while sequentially growing the stroke chain. Note that the growth of the stroke chain at that time is up to a predetermined stroke number (for example, 16 strokes) set in consideration of the stroke numbers forming the sub patterns registered in the sub pattern dictionary. Hereinafter, the feature re-extraction processing and the weighted DP matching will be described more specifically with reference to FIGS. 7B, 7C, and 7D.

First, the feature re-extracting section 8 selects one of the feature point sequences (see FIG. 7B) created by the stroke processing section 5 based on the input character pattern having the sub-pattern "month".
The characteristic point sequence of the first stroke a1 is read out, and the above-described normalization and characteristic re-extraction processing is performed to obtain the characteristic pattern a2 of the first stroke a1 (see FIG. 7C). Next, the character recognition unit 9 performs weighted DP matching between the characteristic pattern a2 of the first stroke a1 and the standard pattern registered in the sub-pattern dictionary to obtain a sub-pattern candidate whose matching distance D is a predetermined value or less. The obtained sub-pattern candidate is added to the matching distance D and stored in the internal buffer.

Next, the feature re-extracting section 8 normalizes the feature point sequence of the first stroke a1 and the second stroke b1 and then, between the first stroke a1 and the second stroke b1. Since there is an off-stroke section, the off-stroke section is interpolated. Then, characteristic re-extraction processing is performed to obtain a characteristic pattern of a stroke chain composed of the first stroke a2, the off-stroke section e, and the second stroke b2. Then, the character recognition unit 9 performs weighted DP matching to obtain a sub-pattern candidate whose matching distance D is a predetermined value or less, and adds the matching distance D to the obtained sub-pattern candidate and stores it in an internal buffer.

Thereafter, similarly, strokes c1 and d1 are sequentially added to the stroke chain of the first stroke a1 and the second stroke b1, and the strokes are interpolated in the off-stroke sections f and g, and the feature re-extraction processing and weighting DP are performed. Match. Such processing is continued until the number of stroke chains reaches the predetermined number of strokes "16" or the feature point sequence ends, and a sub pattern candidate lattice is generated in the internal buffer of the character recognition unit 9. It is. FIG. 7D is the characteristic pattern of the sub-pattern “month” finally extracted by the characteristic re-extracting unit 8 in this way.

In step S7, based on the sub-pattern candidate lattice obtained as a result of the weighted DP matching in step S6 by the character recognition unit 9,
The character dictionary registered in the dictionary storage unit 10 is referenced to generate character candidates. Then, from the character candidate having the smallest total value of the matching distances D associated with the sub-patterns forming each character candidate, for example, the character candidates up to the tenth place are output as the recognition result. In step S8, special classification processing is performed on characters such as the characters "ya" and "ya" that cannot be distinguished by the shape of the stroke. For example, in the case of the classification of the characters “ya” and the characters “ya”, since the shapes are completely the same, it is impossible to identify them by DP matching using the characteristic pattern extracted after normalization.
Therefore, in step S3, the stroke processing unit 5
The maximum distance in the y-axis direction (character height) of the character to be recognized is calculated based on the coordinates of the feature points extracted by, and the height of this character must be less than 1/2 of the character recognized immediately before. For example, the small letter “ya” is determined to be the first recognition result candidate.

In step S9, the result display unit 3 displays the first recognition result candidate on the display panel integrally laminated on the lower side of the tablet 2. In step S10, it is determined whether or not the end of the handwriting recognition processing operation has been instructed. As a result, if not instructed, the process returns to step S1 and waits for the next pen input,
If instructed, the handwriting recognition processing operation ends.

As described above, in the present embodiment, the stroke processing unit 5 extracts the coordinate points whose direction difference between the coordinate points is larger than the predetermined value Ndist for each input stroke as feature points. There is. That is, in the present embodiment, the characteristic points and the characteristic amounts representing the characteristic of the bent portion are extracted in advance before being normalized. Therefore, even if the characteristic of the bent portion is lost during the later normalization, the feature amount and the characteristic point representing the bent portion can be reliably extracted.

Further, the feature re-extracting section 8 sets 1
Noise is removed by interpolating between feature points in which the distance between normalized feature points is greater than a predetermined value Nfmin for each stroke chain forming a character pattern. Further, the off-stroke section is interpolated at the characteristic points so as to be equally divided. In this way, the feature points are extracted again after the above-mentioned normalization, and the equal interval approximation is performed. That is, the extraction of the feature points in this embodiment is a combination of the equal division method of the conventional feature extraction method (2) and the angle method of the feature extraction method (3) described above. Therefore, the feature points can be optimally extracted by compensating for the drawbacks of the conventional feature extraction methods (2) and (3), and the accuracy of the subsequent DP matching can be improved and the time can be shortened. Also, by interpolating the off stroke,
It becomes possible to deal with "entry" and "bounce" or "continuation character" and "collapse character" of the above pen.

Furthermore, since the character recognition unit 9 performs DP matching using a standard pattern in units of sub-patterns that are easy to write with one stroke, the number of lattices generated is small, and expansion into character candidates is easy and fast. Can be done. Therefore, according to this embodiment, the recognition time can be shortened and the recognition rate can be increased. Also, the storage capacity of the sub-pattern dictionary and the character dictionary can be reduced.

<Second Embodiment> This embodiment relates to generation of the Euclidean distance ΔL and the direction difference ΔH used in the weighted DP matching in the character recognition unit 9 in the first embodiment.

FIG. 8 is a block diagram of the handwriting recognition device in this embodiment. Tablet 12, result display unit 13,
The stroke end determination unit 14, the stroke processing unit 15, the feature storage unit 16, the character end determination unit 17, the feature re-extraction unit 18, and the dictionary storage unit 20 are the tablet 2, the result display unit 3, and the stroke end determination in the first embodiment. The unit 4, the stroke processing unit 5, the feature storage unit 6, the character end determination unit 7, the feature re-extraction unit 8 and the dictionary storage unit 10 have the same functions and operate in the same manner.

The storage unit 21 stores an absolute difference table, a distance table and a direction difference table used when the character recognition unit 19 generates the Euclidean distance ΔL and the direction difference ΔH. The absolute difference table and the distance table are used when generating the Euclidean distance ΔL. Here, since the characteristic points when the Euclidean distance ΔL is obtained are the normalized characteristic points, the x and y coordinates of the characteristic points fall within the range of 0 to Nnorm. Therefore, for example, when Nnorm = 128 dots, as shown in FIGS. 9A and 9B, the difference between the coordinates of the feature point n on the feature pattern and the coordinate m on the standard pattern is calculated. A 128 × 128 absolute difference table having absolute values as elements is created for each x coordinate and y coordinate. The absolute value of the coordinate difference is within the range of 0 to 127. Therefore, a 128 × 128 distance table having the Euclidean distance ΔL as an element as shown in FIG. 9C is created.

The direction difference table is used when generating the direction difference ΔH. Here, the absolute direction of the characteristic point n on the characteristic pattern or the characteristic point m on the standard pattern is defined so that one rotation is 360 degrees with a certain direction being "0", and an absolute value of "0-359" is set. Express with a numerical value.
Then, as shown in FIG. 10, the feature point (n-1) on the feature pattern and the feature point on the standard pattern
A 360 × 360 direction difference table having the direction difference ΔH from (m−1) as an element is created.

In the above structure, the character recognition unit 19 generates the Euclidean distance ΔL and the direction difference ΔH as follows. That is, first, the absolute value of the difference between the x coordinate "xn" of the characteristic point n on the characteristic pattern and the x coordinate "xm" of the characteristic point m on the standard pattern is obtained by looking up the absolute difference table. Similarly, the absolute value of the difference between the y-coordinate “yn” of the characteristic point n and the y-coordinate “ym” of the characteristic point m is obtained by drawing the absolute difference table. Then, the Euclidean distance ΔL is obtained by drawing the distance table based on the obtained absolute values “| xn−xm |” and “| yn−ym |”. Next, the difference between the direction dn toward the feature point (n + 1) at the feature point (n-1) on the feature pattern and the direction dm toward the feature point (m + 1) at the feature point (m-1) on the standard pattern. ΔH is obtained by drawing the direction difference table.

After that, the Euclidean distance ΔL and the direction difference ΔH obtained as described above are used to calculate the partial matching distance Δd according to the equation (3), and DP matching is performed using this partial matching distance Δd. Of.

As described above, in this embodiment, the storage unit 21 in which the absolute difference table, the distance table and the direction difference table are registered is provided. Then, the character recognition unit 19
Using the above absolute difference table and distance table,
Feature point n on the feature pattern and feature point m on the standard pattern
And the Euclidean distance ΔL is calculated. Further, using the direction difference table, the feature point (n-1) on the feature pattern
To the feature point (n + 1) and the feature point on the standard pattern
The direction difference ΔH from the direction from (m−1) to the feature point (m + 1) is calculated. Therefore, according to the present embodiment, the Euclidean distance Δ can be obtained by a simple process of accessing each table stored in the storage unit 21.
It is possible to obtain L and the direction difference ΔH, and it is possible to increase the generation processing speed of the partial matching distance Δd, and thus the DP matching processing speed.

<Third Embodiment> This embodiment relates to a change in weight used when performing the weighted DP matching in the character recognition unit 9 in the first embodiment.

FIG. 11 is a block diagram of the handwriting recognition apparatus in this embodiment. Tablet 32, result display unit 3
3, stroke end determination unit 34, stroke processing unit 35,
The feature storage unit 36, the character end determination unit 37, and the feature re-extraction unit 38 are the tablet 2, the result display unit 3, the stroke end determination unit 4, the stroke processing unit 5, the feature storage unit 6, and the character end determination in the first embodiment. It has the same functions as the unit 7 and the feature re-extracting unit 8 and operates in the same manner.

In the present embodiment, the weight is calculated for each of the feature points (m-1) to (m + 1) (hereinafter referred to as section m) in the standard pattern, and the obtained weight is the feature of the standard pattern. It is added to the point m and registered in the dictionary storage unit 40.

The method of calculating the weight for each section will be described in detail below. In the present embodiment, the character recognition unit 3
The weight Wp related to the distance and the weight Wd related to the direction used in the calculation of the partial matching distance Δd in 9 are represented by Expression (6) and Expression (7). Wp = Nwlen · Wlwm (6) Wd = Nwdst · Wdwm (7) where Nwlen, Nwdst: uniform weights among all feature points as in the first embodiment Wlwm, Wdwm: in section m in the standard pattern Fluctuation weight

The above-mentioned fluctuation weights Wlwm and Wdwm are obtained as described below under the control of the weight calculator 41. That is, first, a learning character set including coordinate data strings of various handwriting patterns for all recognition target characters is prepared. Then, a coordinate data string of one sub-pattern relating to each recognition target character is arbitrarily selected and input to the stroke end determination unit 34, and feature extraction / re-extraction is performed in the same manner as in the first embodiment, The obtained feature string is registered in the weight calculation buffer 42 as an initial standard pattern. Next, the stroke end determination unit 3 determines a coordinate data string having a sub-pattern of a learning character in the learning character set.
4 and the feature extraction / similar to the case of the first embodiment.
Re-extraction is performed to obtain a learning feature pattern. The character recognition unit 39 reads the initial standard pattern related to the sub-pattern from the weight calculation buffer 42 and performs DP matching between the initial standard pattern of the sub-pattern and the learning feature pattern. DP matching at that time is the first
This is DP matching using the partial matching distance Δd of Expression (3) in the embodiment.

As a result of the DP matching, if the sum of the distances along the matching path is larger than a predetermined value, it is determined that the learning feature pattern is not in the category of the sub-pattern, and the variation weight is calculated. remove.

When the learning feature pattern is to be the object of variable weight calculation, the Euclidean distance ΔL for each point on the matching path calculated during the DP matching is calculated.
And the direction difference ΔH are stored in the weight calculation buffer 42 in association with the section related to the feature point on the initial standard pattern.

The above-described processing is performed for all the coordinate data strings in the recognition target characters of all the learning character sets to obtain the distribution of the Euclidean distance ΔL and the direction difference ΔH for each section. Then, the weight calculation unit 41 calculates the variance ρL of the Euclidean distance ΔL and the variance ρH of the directional difference ΔH for each section based on the Euclidean distance ΔL and the direction difference ΔH stored in the weight calculation buffer 42. The variance ρLm of the Euclidean distance ΔL related to the section m is calculated and used as the fluctuation weight Wlwm. On the other hand, the direction difference Δ in the section m
The variance ρHm of H is used as the fluctuation weight Wdwm. The variation weights Wlwm and Wdwm thus obtained are added to the feature points in the initial standard pattern and stored in the dictionary storage unit 40 as the standard pattern and its variation weights.

In the present embodiment, the variance weights Wlwm and Wdwm are calculated by calculating the variance ρL of the Euclidean distance ΔL and the variance ρH of the direction difference ΔH for each section.
The average μL of the Euclidean distance ΔL and the average μH of the direction difference ΔH may be used as the fluctuation weights Wlwm and Wdwm. The point is that it is only necessary to be able to represent the degree of variation in the learning character set with respect to the initial standard pattern.

Thus, in this embodiment, in advance,
A learning character set of all recognition target characters is prepared, one coordinate data string relating to each recognition target character is arbitrarily selected, and a feature string is obtained to be an initial standard pattern. Then, the character recognition unit 39 performs DP matching between the learning feature pattern from each coordinate data string in the learning character set and the initial standard pattern related to the learning character, and the Euclidean character corresponding to each point along the matching path. The distance ΔL and the direction difference ΔH are calculated. Then, the variance ρL of the Euclidean distance ΔL and the variance ρH of the direction difference ΔH related to the feature points in the section m in the initial standard pattern are set as the variation weights Wlwm, W.
Get as dwm. After that, when the handwriting is recognized by the character recognition unit 39, the variation weight Wlw thus obtained is obtained.
Weighted DP matching using m and Wdwm is performed.

Therefore, according to the present embodiment, when the characteristic points of the curved portion of the handwritten character cannot be stably extracted, the DP matching is performed by adding the weight according to the instability. High recognition performance can be secured.

<Fourth Embodiment> In this embodiment, the variation weight Wlw
The present invention relates to another embodiment in which weighted DP matching is performed using n and Wdwn. FIG. 12 is a block diagram of the handwriting recognition device in this embodiment. Tablet 52, result display unit 53, stroke end determination unit 54, stroke processing unit 5
5, the feature storage unit 56, the character end determination unit 57, the feature re-extraction unit 58, and the dictionary storage unit 60 are the tablet 2, the result display unit 3, the stroke end determination unit 4, the stroke processing unit 5, and the feature in the first embodiment. The storage unit 6, the character end determination unit 7, the feature re-extraction unit 8 and the dictionary storage unit 10 have the same functions and operate in the same manner.

In this embodiment, the weight Wp related to the distance and the weight Wd related to the direction used in the calculation of the partial matching distance Δd in the character recognition unit 59 are expressed by the equations (8) and (9).
And Wp = Nwlen · Wlwn (8) Wd = Nwdst · Wdwn (9) where Wlwn, Wdwn: characteristic point n in the characteristic pattern
Variation weight for

The weight calculation section 61 has the feature re-extraction section 58.
Variation weights Wlwn, Wdw based on the feature pattern obtained in
n is calculated, and the obtained variation weights Wlwn and Wdwn are added to the feature pattern and sent to the character recognition unit 59. Then, the character recognition unit 59 performs the same weighting as in the third embodiment based on the characteristic pattern and the variation weights Wlwn, Wdwn from the weight calculation unit 61 and the standard pattern of the sub-pattern dictionary stored in the dictionary storage unit 60. DP matching is performed. At this time, the calculation of the variation weights Wlwn and Wdwn by the weight calculation unit 61 is performed as follows.

For each feature point n in the feature pattern from the feature re-extracting unit 58, the feature point (n-1) and the feature point
The distance from (n + 1) is calculated. In addition, the distance (hereinafter, simply referred to as "total length") connecting all the feature points of the stroke in the input order is calculated based on the coordinates of all the feature points forming the stroke to which the feature point n belongs. Then, the variation weights Wlwn and Wdwn are calculated according to the equation (8). Wlwn = Wdwn = (distance between feature point (n-1) and feature point (n + 1)) / (total length of stroke) (10)

As described above, in the present embodiment, the values of the fluctuation weights Wlwn and Wdwn in the above weighted DP matching are calculated by the ratio of the distance between the characteristic points on both sides of the characteristic point n to the stroke length. There is. Therefore, the partial matching distance Δd between the characteristic points on the straight line portion is increased, and the deviation of the coordinates of the characteristic points of the bent portion between the characteristic pattern and the standard pattern can be easily absorbed. That is, according to this embodiment, it is possible to realize a handwriting recognition device that is resistant to local deformation of the handwriting input pattern.

<Fifth Embodiment> This embodiment relates to an input pattern recognition method. In the present embodiment, in order to simplify the description, a handwritten character recognition device that recognizes a handwritten character pattern will be described as an example, but the same applies to a pattern other than a character such as a gesture. Can be done by

After briefly describing the outline of the invention, the data structure, the recognition method and the like will be described in detail. FIG. 13 is a block diagram of the handwriting recognition device in this embodiment.
This handwriting recognition device is roughly configured by a recognition unit 71, a tablet 72, a feature extraction storage unit 73, a recognition storage unit 74, a sub-pattern dictionary 75, and a result display unit 76. Here, the definition of the sub-pattern in this embodiment is as follows:
It is exactly the same as the definition in the embodiment.

The tablet 72 is the first to fourth embodiments.
Similar to the embodiment, for example, it is formed as a display-integrated type and converts the handwriting data pen-input by the writer into a coordinate data string represented by an electric signal. The recognition unit 71 is a tablet 72.
The feature of the coordinate data string input from is extracted, the handwritten character is recognized using the sub-pattern dictionary 75, and displayed on the result display unit 76. The feature extraction storage unit 73 stores various buffers used when extracting features of the input coordinate data sequence. Further, in the recognition storage unit 74,
Various buffers used during the handwriting recognition process are stored. The character structure table in the recognition storage unit 74 is the same as the character dictionary in the first embodiment and has the structure shown in FIG.

The coordinate data string input from the tablet 72 is input to the normalization unit 78 via the input unit 77 and is normalized for each one or more stroke chains. Then, the feature extraction unit 79 extracts feature points and their feature amounts from the normalized stroke chain as described later. The direction difference determination unit 80 obtains the direction difference related to the final feature point of the stroke among the feature points extracted by the feature extraction unit 79 as described in detail later. Then, the validity of the standard pattern used for DP matching is determined based on this direction difference. Stroke limited part 8
1 limits the number of regular strokes of the input pattern from the tablet 72 to a certain range. By limiting the number of strokes of the input pattern in this way, the speed of the DP matching performed later can be increased.

The DP matching unit 82 has a pattern of a feature point sequence having the feature quantity extracted by the feature extracting unit 79.
The sub-pattern is recognized by performing DP matching between the (feature pattern) and the standard pattern in sub-pattern units registered in the sub-pattern dictionary 75. The character generator 84
The sub pattern candidate sorting unit 83 combines the sub pattern candidates to generate a character candidate. The output unit 85 sends the generated character candidates to the result display unit 76 in order of high similarity and displays them. Control unit 8
6 is the input unit 77, the normalization unit 78, the feature extraction unit 79,
The handwriting character recognition process is executed by controlling the direction difference determination unit 80, the stroke number limiting unit 81, the DP matching unit 82, the sub pattern candidate sorting unit 83, the character generation unit 84, and the output unit 85.

FIG. 14 is a flowchart showing the general flow of the handwritten character recognition processing operation performed under the control of the control unit 86. The outline of the handwritten character recognition processing operation will be described below with reference to FIG. Step S
At 11, the coordinate data string from the tablet 72 is fetched by the input unit 77 and stored in the input coordinate buffer of the feature extraction storage unit 73. In step S12, the normalization unit 78, the feature extraction unit 79, the direction difference determination unit 80, the stroke number limiting unit 81, the DP matching unit 82, and the sub-pattern candidate sorting unit 83 sequentially generate a Slotalk chain for normalization, Feature extraction, stroke number limitation, off direction difference calculation to the next stroke, sub pattern recognition and first sub pattern candidate sorting are performed,
The first sub-pattern candidate is determined. The first sub-pattern is the first sub-pattern that constitutes the input character pattern, and usually corresponds to the deviation of characters. In step S13, the normalization unit 78 and the feature extraction unit 7
9, direction difference determination unit 80, stroke limit unit 81, DP matching unit 82, sub-pattern candidate sorting unit 83
And the character generation unit 84 performs sub-pattern recognition while limiting subsequent sub-patterns based on the determined first sub-pattern candidate, and combines the obtained sub-pattern candidates to generate a character candidate. In step S14, the output unit 85 sorts the character candidates sorted in descending order of similarity by the character candidate sorting unit 88 of the character generation unit 84 in descending order of similarity.
It is output to 6 and displayed.

The first sub-pattern recognition processing and character generation processing in the above general flow will be described in detail below. Here, in order to facilitate the description of the recognition process and the character generation process of the first sub-pattern, the normalization process by the normalization unit 78, the feature extraction process by the feature extraction unit 79, the direction difference by the direction difference determination unit 80. Judgment process, stroke number limiting process by stroke number limiting unit 81, DP
The DP matching process by the matching unit 82 and the structure of the sub-pattern dictionary 75 will be described in advance.

(I) Normalization processing The stroke data for one or more strokes (variable length coordinate sequence data) is normalized so that the circumscribing frame of the entire stroke data becomes Nnorm dots vertically and horizontally. The number of dots "Nnorm" is a value that can be arbitrarily determined according to the ability of the recognition device. In this embodiment, Nnorm =
It is 200 dots.

(II) Feature Extraction Processing With respect to the stroke data after normalization, the off-stroke section is interpolated with the normalization value of appropriate coordinate data to make a single stroke of the input pattern. After that, all coordinate data are scanned, and a point whose direction has changed by a threshold value or more is obtained as a bending point. When the distance between the bending points is equal to or greater than a predetermined value, interpolation points are obtained at w (for example, "70") dot intervals, and the bending points and the interpolation points (equal spacing points) are combined and extracted as feature points. . Then, the direction value d of each feature point to the next feature point is obtained and combined with the coordinate value (x, y) to obtain the feature amount Fn. Therefore, the feature amount Fn is a variable-length coordinate sequence with a direction represented by the equation (11). _{Fn = (x 1, y 1} , d 1), (x 2, y 2, d 2), ..., (x n, y n, d n) ... (11)

The direction value d is defined so that one rotation is 360 degrees with a certain direction being "1", and is expressed by an integer value of 1 to D (for example, "64"). While the effective range of the coordinates (x, y) depends on the number of dots Nnorm in the normalized frame,
The effective range of the direction value d depends on the quantization number D of the direction. However, the dot number Nnorm, the quantization number D, and the interpolation interval w
Is a value that can be arbitrarily determined according to the ability of the recognition device.

(III) Directional Difference Determination Processing This processing is processing for determining whether or not skip processing for speeding up the DP matching processing is to be performed. Specifically, two directional differences from the final feature point of one stroke to the leading feature point of the next stroke are obtained, and the value of this directional difference and the directional evaluation stored in the sub-pattern dictionary 75 for each standard pattern are evaluated. This is a process of determining an effective standard pattern with a value. If there is a standard pattern which is determined not to be an effective standard pattern in this process, the DP matching process using the standard pattern is not performed and the next standard pattern is skipped.

Here, the direction difference is defined as follows. That is, since the direction value d is a cyclic value, its difference
(Ddiff (d1, d2)) is defined by equation (12). Ddiff (d1, d2) = min (| d1-d2 |, D- | d1-d2 |) (12) where d1 and d2 are direction values in two directions at a certain feature point.
(1≤d1≤D, 1≤d2≤D), and min (a, b) is a,
This means taking the smaller value of b.

As described above, the direction value d is a cyclic value. Therefore, the average and standard deviation cannot be directly calculated. Therefore, in this embodiment, two orthogonal axes are used.
(Direction value) is prepared, the direction difference between these two axes and the direction value d is calculated, and the average value and standard deviation value of the two calculated direction differences are substituted. That is, d _i is the i (1 ≦ i ≦ I) th direction value, the direction values of the first and second axes are R ^j (j = 1, 2),
If the direction difference between the first and second axes is dd ^j , the direction difference dd ^j is
(13) is determined, the mean m ^j direction difference dd ^j given by Equation (14), the standard deviation v ^j direction difference dd ^j is determined by Equation (15).

[Equation 1] The first and second axes are two orthogonal axes, and can be arbitrarily determined according to the ability of the recognition device. In this embodiment, R ¹ = 1 and R ²
= 16.

The skip process using the direction difference determination process result is performed as follows. For example, in the sub-pattern “shellfish” in the character “bet”, the direction of the off-stroke to the next sub-pattern “person” (hereinafter, referred to as the off direction) is limited to a specific range (upper right direction). Therefore, by using this off-direction information, it is checked whether the off-direction to the next stroke related to the current feature point is within the allowable range of the off-direction added to the current standard pattern. In some cases, the current standard pattern is not targeted for DP matching processing, and the next standard pattern is skipped.

Specifically, the skip processing is performed as follows. First, the off direction from the final feature point of the final stroke in the standard pattern to the next stroke is obtained, and the direction difference dd ¹ , between the off direction and the first axis (direction value d1) and the second axis (direction value d2). Find dd ² . Next, from the header part of the current standard pattern in the sub-pattern dictionary 75, the direction difference average value m ¹ with the first axis and the direction difference average value m ² with the second axis, and the direction difference with the first axis. The standard deviation value v ¹ and the directional difference standard deviation value v ² between the second axis are read out to obtain the equation (16),
It is determined whether Expression (17) is satisfied. If not satisfied, the process skips to the next standard pattern. | dd ¹ −m ¹ | <v ¹ · Cs… (16) | dd ² −m ² | <v ² · Cs… (17) where Cs: a constant (eg, “3”)

The average value m ¹ , m ² of the direction difference between the first and second axes and the standard deviation value v ¹ , v of the direction difference between the first and second axes written in the sub-pattern dictionary 75. ² is the direction difference average value and the standard deviation value with respect to the off-direction value from the target sub-pattern to the next sub-pattern, and is the standard deviation value, based on the coordinate data string of a large amount of handwritten characters for each sub-pattern. Is what I asked for.

(IV) Stroke Number Limiting Process The stroke number limiting process is a process of limiting (estimating) the regular stroke number of the input pattern to a certain range.
This is a process for speeding up the P matching process. As shown in Expression (18), the limit range of the stroke number sl is such that the lower limit is the input stroke number rs of the input pattern, and the upper limit is an expected value that can be calculated from the number of bending points pc obtained during the feature extraction processing. To do. rs ≦ sl ≦ pc + 1 (18)

DP according to the above-mentioned stroke number limiting processing result
The speed of the matching process is increased as follows. That is, the sub-pattern dictionary 75 is sorted in advance in the order of regular stroke numbers. Then, in the DP matching processing, all the standard patterns are not used, but only the standard pattern of the sub-patterns having a regular stroke number within the limited range of the stroke sl related to the feature pattern of interest is used. To do.

(V) DP matching process This is a process for calculating the distance between the characteristic pattern from the coordinate data string of the input pattern and the standard pattern registered in the sub pattern dictionary 75. This DP matching processing is to obtain the distance between two data having different lengths.
The definition of the matching path, the width of the matching window, and the definition of the matching distance can be arbitrarily determined according to the ability of the recognition device.

In this embodiment, the partial matching distance d (m, n) is defined by the equation (19), and the partial sum D (m, n) of the distances along the matching path is defined by the equation (20). To do.

[Equation 2]

The width ww of the matching window of DP matching is variable, and is calculated by the equation (21) based on the number N of characteristic points of the characteristic pattern and the number M of characteristic points of the standard pattern. ww = max (Cw, | MN |) (21) The above max (a, b) means to take the larger value of a and b. Further, Cw is a constant that can be arbitrarily set according to the ability of the recognition device, and is “8” in this embodiment.
Is. That is, in the DP matching processing in the present embodiment, the partial sum D (m, n) of the distances along the matching path within the range of 1 ≦ m ≦ M, 1 ≦ n ≦ N and the matching window ww is obtained,
Finally, the value obtained by dividing the value of D (M, N) by N becomes the matching distance dist. dist = D (M, N) / N (22)

(VI) Structure of Sub-Pattern Dictionary 75 Standard patterns in each sub-pattern unit are registered in the sub-pattern dictionary 75. The data of the standard pattern is a variable-length coordinated coordinate sequence similar to the feature amount Fn of the above-described feature pattern, and the normalization processing and feature extraction by the normalization unit 78 are performed on the coordinate data sequence of the sub-pattern prepared in advance. It is created by performing the feature extraction processing by the unit 79. Further, various information of each sub-pattern is added to the sub-pattern dictionary 75 as header information. Some header information is as shown in FIG. In FIG. 15, the "stroke number part" stores the regular stroke number, and the "first direction difference average part" and the "second direction difference average part" have the direction difference average value m with the first axis. ¹ , the direction difference average value m ² with the second axis is stored, and the direction difference standard deviation value v ¹ with the first axis is stored in the “first direction difference deviation section” and the “second direction difference deviation value”. The directional difference standard deviation value v ^{2 from} the second axis is stored. The dictionary structure of the sub-pattern dictionary 75 is composed of the header section and variable-length coordinate data string per sub-pattern dictionary.

The recognition process of the first sub-pattern will be described below. This process is performed in the order of the matching distance dist, which is obtained by the DP matching process with respect to the characteristic pattern extracted from the input coordinate data sequence, in the order of increasing
This is a process of creating first sub-pattern candidates up to the rank.
Normally, when performing online handwritten character recognition, this first sub-pattern recognition processing is realized as pre-recognition processing in parallel with the user's input operation. However, in the present embodiment, in order to simplify the explanation. The operation after the input processing will be described (the method of parallel processing will be described later).

FIG. 16 shows the normalizing unit 78, the feature extracting unit 79, the direction difference determining unit 80, the stroke number limiting unit 81, and the DP.
9 is a flowchart of a first sub-pattern recognition processing operation by the matching unit 82 and the sub-pattern candidate sorting unit 83. Hereinafter, the first sub-pattern recognition processing operation will be described in detail with reference to FIG. Step S21
, The start stroke counter ststrk and the end stroke counter en set in the feature extraction storage unit 73.
The contents of strk are initialized to "1" respectively. At step S22, the contents of the sub-pattern candidate buffer in the recognition storage section 74 are initialized. The sub-pattern candidate buffer has a structure as shown in FIG. 17 for each candidate. In addition, the number of sub-pattern candidates obtained in this embodiment is “50”. Therefore, the sub-pattern candidate buffer in this embodiment has a memory capacity / structure in which 50 structures shown in FIG. 17 are arranged. here,
In the initialization of the sub-pattern candidate buffer, the maximum distance value maxdist (“255” in this embodiment) is set as the value of all 50 DP matching distances.

In step S23, the normalization unit 78 performs the above-described normalization processing on the stroke data from the "ststrk" th stroke to the "enstrk" th stroke. In step S24, the feature extraction unit 7
In step 9, the above-mentioned feature extraction process is performed on the stroke data after the normalization process in step S23 to obtain a feature pattern. The feature pattern thus obtained is stored in the feature buffer of the feature extraction storage unit 73. In step S25, the stroke number limiting unit 81
The above-described stroke number limiting process is performed. That is, the bending point number pc is obtained based on the direction value d, which is one of the feature quantities extracted in step S24,
The limited range of the stroke number sl is calculated according to (18). The limited range of the stroke number sl thus obtained is stored in the input data buffer of the feature extraction storage unit 73. In step S26, the off-direction difference to the next stroke is calculated by the direction difference determination unit 80. That is,
"Enstrk" from the last feature point of the "enstrk" th stroke
The off direction to the leading feature point of the +1 "th stroke is obtained, and the direction differences dd ¹ and dd ² between this off direction and the first and second axes are obtained. The direction differences dd ¹ and dd ²
The value of is stored in the input data buffer of the feature extraction storage unit 73.

In step S27, the coordinate data string of each feature point and the stroke number s obtained in the above-mentioned steps and stored in the feature buffer, the input data buffer, etc.
The limited range of l, the difference of two directions dd ¹ and dd ² , the start stroke number and the end stroke number are input data, and the above D
The P matching unit 82 executes a sub-pattern recognition processing subroutine, which will be described in detail later. In step S28, the content of the end stroke counter enstrk is incremented. In step S29, it is determined whether or not the content of the end stroke counter enstrk is less than or equal to the number of strokes in which stroke data is stored in the input coordinate buffer of the feature extraction storage unit 73. As a result, if it is larger than the input stroke number rs, the process proceeds to step S30. On the other hand, if the number of input strokes is less than or equal to rs, the process returns to step S23, and shifts to the normalization process for the stroke data from the "ststrk" th stroke to the incremented "enstrk" th stroke. In this way, the normalization process, the feature extraction process, the stroke number limitation process, the off-direction difference calculation process, and the sub-pattern recognition process are repeated while the stroke chain is sequentially grown until the stroke number reaches the input stroke number rs.

In step S30, the first sub-pattern candidates generated in the sub-pattern candidate buffer as a result of the sub-pattern recognition processing in step S27 are sorted by the sub-pattern candidate sorting section 83 in ascending order of matching distance dist. It After that, the first sub-pattern recognition processing operation is ended and the operation proceeds to the character generation processing operation.

FIG. 18 is a flowchart of the sub pattern recognition processing subroutine. Hereinafter, the sub-pattern recognition processing operation performed by the DP matching unit 82 will be described in detail with reference to FIG. In step S31, the limited range of the number of strokes sl stored in the input data buffer is read out, and based on this limited range, the reading start when reading the standard pattern from the sub-pattern dictionary is started as follows. The address and the read end address are set. Here, the sub-pattern dictionary 75 is sorted in ascending order of the regular stroke number. Then, in the sub-pattern dictionary 75, a table in which the start address and the end address of the registered standard pattern having the same number of strokes are associated with the number of strokes is registered as additional information. Therefore, the table is referred to based on the lower limit value of the limited range of the stroke number sl to obtain the read start address when reading the standard pattern. The read start address thus set is set in the start pointer dicp set in the internal memory. In step S32, the table is referred to based on the upper limit value of the limited range of the number of strokes sl, and the read end address for reading the standard pattern is set. And the end pointer set in the internal memory
Set to endp. In step S33, the maximum distance is searched from within the "distance part" in the sub-pattern candidate buffer. Then, the obtained maximum distance is stored in the maximum distance register maxdist.

In step S34, the sub-pattern dictionary 7
The header part is read from the address designated by the start pointer dicp in 5 and stored in the internal memory. In step S35, based on the direction difference information dd ¹ , dd ^{2 of the} standard pattern and the direction difference information in the internal memory obtained in step S26 in the flowchart of the first sub-pattern recognition processing operation (see FIG. 16). The skip processing described above is performed. Then, it is determined whether or not the standard pattern related to the header part is skipped to the next standard pattern without being the target of the DP matching process. As a result, if skipped, the process proceeds to step S40, and if not skipped, the process proceeds to step S36. Step S36
Then, the DP of the characteristic pattern and the standard pattern
The matching process is executed, and the obtained matching distance dist is stored in the distance register dist. Step S37
Then, the contents of the above distance register dist is the maximum distance register ma
It is determined whether or not it is less than or equal to the content of xdist. As a result, if the following is true, the process proceeds to step S38, and if not, the process proceeds to step S40.

In step S38, the information of the sub-pattern candidate having the maximum distance maxdist in the sub-pattern candidate buffer is rewritten with the information of the standard pattern. This rewriting of information is performed as follows. That is, the "distance part" stores the contents of the distance register dist, and the "start stroke part" contains the start stroke counter.
The contents of ststrk are stored, and the "stroke part" is (contents of end stroke counter enstrk-start stroke
The calculation result of the counter ststrk + 1) is stored, and the "sub-pattern code part" stores the sub-pattern code of the header part. Here, when there are a plurality of sub-pattern candidates having the maximum distance maxdist, the information of the sub-pattern candidate having the minimum address is rewritten. In step S39, the maximum distance is searched from within the "distance portion" in the sub-pattern candidate buffer. Then, the obtained maximum distance is stored in the maximum distance register maxdist.

At step S40, the content of the start pointer dicp is incremented. In step S41, it is determined whether or not the content of the start pointer dicp is smaller than the content of the end pointer endp. As a result, if it is smaller, the process returns to the step S34 to shift to the DP matching process using the next standard pattern for the feature pattern. On the other hand, if it is equal to or more than the end pointer endp, the sub pattern recognition processing subroutine is ended and the process returns to the main routine.

Thus, when the sub-pattern recognition processing subroutine is completed, a maximum of 50 sub-pattern candidates consisting of standard patterns similar to the characteristic pattern are stored in the sub-pattern candidate buffer.

FIG. 19 shows the normalizing unit 78, the feature extracting unit 79, the stroke number limiting unit 81, the DP matching unit 82,
9 is a flowchart of a character generation processing operation by a character generation unit 84, a character candidate expansion unit 87, and a character candidate sorting unit 88. Hereinafter, the character generation processing operation will be described in detail with reference to FIG. When the first sub-pattern recognition processing operation is completed, the character generation processing operation is started.

In step S51, the character candidate buffer of the recognition storage unit 74 is determined by the contents of the first sub-pattern recognition result obtained by the first sub-pattern recognition processing operation and stored in the sub-pattern candidate buffer. It is initialized. This initialization is performed as follows. That is, in FIG. 20, “DP matching distance”, which is the content of “distance part” of the sub-pattern candidate buffer, is stored in “distance part”. Similarly, the "end stroke part" stores the "input stroke count", which is the content of the "stroke count part". In addition, "sub pattern
The "code part" stores "sub-pattern code" which is the contents of the "sub-pattern code part", and "1" is stored in the "sub-pattern number part". The symbol Nsp in the data length of the pattern code part is 1
It is the maximum number of sub-patterns that form a character and varies depending on the definition of the sub-pattern ("8" in this embodiment).

The character candidate buffer is shown in FIG.
The memory capacity / structure in which the 200 structures shown in FIG. Further, a maximum of 50 first sub-pattern candidates are stored in the sub-pattern candidate buffer. Therefore, at the stage when the character candidate buffer is initialized, a maximum of 50 pieces of character candidate information are written in the ascending order of the DP matching distance dist. FIG. 21 schematically shows the contents of the character candidate buffer immediately after the initialization when the character "behind" is input by handwriting with a "9" stroke. In addition, in FIG.
The items in FIG. 20 are rearranged for easier viewing.

In step S52, the minimum end stroke number is searched from the "end stroke number portion" in the initialized character candidate buffer. Then, the obtained minimum end stroke number is the minimum end stroke number register minens.
Stored in trk. Then, the character candidate buffer is searched again, and the first sub-pattern candidate having the minimum end stroke number is selected. In step S53, it is determined whether or not the content of the minimum end stroke number register minenstrk is equal to the input stroke number. As a result, if they are equal, the process proceeds to step S65, and if they are not equal, the process proceeds to step S65.
Proceed to 54. In the example shown in FIG. 21, the first sub-pattern candidates “ten”, “na”, and “jam” having the minimum end stroke number “2” are searched and it is determined that they are not equal to the input stroke number “9”. It

In step S54, the character configuration table in the recognition storage section 74 is referred to, and the second sub-pattern candidates following the first sub-pattern candidate selected in step S52 are limited. For example, as the second sub-pattern candidates following the first sub-pattern candidates "ten", "na" and "ja", the first sub-pattern candidate is the second sub-pattern candidate "ten" → "force (cooperative)", " My brother (Katsu) ", ..." Na "→" Mu (male) "," Engineering (left) ", ..." Weapon "→" Ta (ridge) "," Sato (back) "... etc. By doing so, the DP matching process at the time of recognizing the input patterns after the second sub-pattern can be speeded up.

In step S55, the start stroke counter ststrk is set to "minimum end stroke number register minens".
The content of trk + 1 is set. In step S56, the content of the minimum stroke count register minenstrk + 1 is set in the end stroke counter enstrk. In step S57, the first sub-pattern recognition processing operation is performed. The contents of the sub-pattern candidate buffer are initialized in the same manner as in step S22, and it is determined in step S58 whether the contents of the end stroke counter enstrk are smaller than the number of input strokes. , If it is smaller than the number of input strokes, proceed to step S59,
If it is equal to or larger than the number of input strokes, the process proceeds to step S64.

In steps S59 to S63, the first
From step S23 in the sub-pattern recognition processing operation
Similar to step S25, step S27 and step S28, the normalization process, the feature extraction process, the stroke number limiting process, the sub pattern recognition process subroutine and the end stroke counter enstrk increment are executed,
Return to step S58. At that time, the DP matching at the time of the sub-pattern recognition processing is performed using only the standard pattern of the sub-patterns limited in step S54. In this way, the time required for DP matching is minimized and the recognition time is shortened.

In this way, the above operation is performed until the content of the end stroke counter enstrk becomes equal to or more than the number of input strokes. If it is determined in step S58 that the content of the end stroke counter enstrk is equal to or greater than the number of input strokes, the process proceeds to step S64.

In step S64, the character candidate expansion section 87 is used.
As a result, a character expansion processing subroutine, which will be described in detail later, is executed and the process returns to step S52. After that, the above operation is performed until the content of the minimum end stroke number register minenstrk becomes equal to the input stroke number. If it is determined in step S53 that the content of the minimum end stroke number register minenstrk is equal to the input stroke number, the process proceeds to step S65. In step S65, the character candidate sorting unit 88 sorts the character candidates generated in the character candidate buffer in ascending order of distance to obtain a recognition result, and ends the character generation processing operation.

That is, in this character generation process, the first sub-pattern candidate having the minimum stroke number is selected from the first sub-pattern candidates obtained as a result of the first sub-pattern recognition process, and the next sub-stroke number is selected. 16 is executed as the new starting stroke number, the first stroke recognition processing operation of FIG. 16 is executed, a sub pattern candidate following the first sub pattern candidate is obtained, and a character candidate is obtained based on the obtained sub pattern candidate. To generate.

FIG. 22 is a flow chart of the character expansion processing subroutine. Hereinafter, the character expansion processing operation executed by the character candidate expansion unit 87 will be described in detail with reference to FIG. This character expansion process is a process of inputting the contents of the character candidate buffer and the sub-pattern candidate buffer and referring to the character configuration table to create a new character candidate. When it is determined in step S58 of the flowchart of the character generation processing operation that the content of the end stroke counter enstrk is equal to the number of input strokes, the character expansion processing operation starts.

In step S71, the character candidate buffer is stored.
The maximum distance is searched from the "distance part" in the
Set to Dista maxdist. In step S72, above
New character candidate counter new set in the recognition storage unit 74 c
In count, "current character candidate counter c contents of count (character
The number of character candidates generated in the complementary buffer) +1 "is set.
Be done. In step S73, it is set in the recognition storage unit 74.
The initial value "1" is set in the selected character candidate counter cc.
It

At step S74, the contents of the character candidate counter cc is changed to the current character candidate counter c. It is determined whether or not it is less than or equal to the content of count. As a result, if below, step S75
Otherwise, to step S86. In step S75, the content of the "end stroke number part" of the "cc" th character candidate in the character candidate buffer (that is, the end stroke number) is stored in the minimum end stroke number register m.
It is determined whether or not it is equal to the content of inenstrk (hereinafter referred to as the noticed end stroke number). As a result, if they are equal, the process proceeds to step S76, and if they are not equal, the step S85.
Proceed to.

In step S76, the initial value "1" is set in the sub-pattern candidate counter sc set in the recognition storage section 74.
Is set. The sub-pattern candidate counter is set to the step S62 of the character generation processing operation shown in FIG.
Sub-pattern candidates after the second sub-pattern are stored by the sub-pattern recognition in. Step S
At 77, it is determined whether or not the content of the sub-pattern candidate counter sc is less than or equal to the number of sub-pattern candidates. As a result, if it is below, the process proceeds to step S78, and if not, the process proceeds to step S85. In step S78, the "cc" th character candidate in the character candidate buffer (there may be more than one)
, And the content of the "sub-pattern code part" of the "sc" -th sub-pattern candidate in the sub-pattern candidate buffer is generated. That is, a sub-pattern code string of the first sub-pattern candidate having the number of strokes of the attention end and the subsequent sub-pattern candidates is generated.

In step S79, the above-mentioned character composition table is
By referring to the sub generated in step S78,
There is a character whose pattern code string is a constituent sub-pattern.
It is determined whether or not it exists. As a result, if it exists,
If not, the process proceeds to step S84. Step
In step S80, select the character searched in step S79.
Average distance c dist is calculated according to equation (23). c dist = [{CD_p-1・ (P-1)} + SD_p] / P (23) where, p: sub-pattern forming the target character
Number of CDs_p-1: Up to the (p-1) th which constitutes the target character
Average distance of character candidates consisting of subpattern sequence of SD_p: P-th sub-pattern that makes up the target character
DP matching distance of candidate

In step S81, the average distance c calculated in step S80 is calculated. The value of dist is the maximum distance register ma
It is determined whether or not it is larger than the content of xdist. As a result, if it is larger, the process proceeds to step S84, and if not, the process proceeds to step S82. In step S82, the above step S79
Since the character searched in is smaller than the maximum distance of the character candidates currently obtained, the above step S78
A new character candidate (new character candidate) is created on the basis of the sub-pattern code string created in (3) and added to the character candidate. Note that the new character candidate creation is specifically performed as follows.

That is, "new" in the character candidate buffer c The contents of "end stroke number part" and "sub pattern number part" of "cc" th character candidate are copied as the contents of "end stroke number part" and "sub pattern number part" of the count "th character candidate. Next, the following processing is performed based on the content of the "sc" th sub-pattern candidate in the sub-pattern candidate buffer: "new" in the character candidate buffer c The "distance part" of count "th character candidate (hereinafter referred to as new character candidate)" is
Average distance c calculated in step S80 dist is stored.・ "Sc" is added to the "end stroke part" of the above new character candidates.
The contents of the "stroke part" in the th sub-pattern candidate are added. -"1" in the content of "number of sub-patterns" of the above new character candidate
Is added. -The contents of the "subpattern code part" in the "sc" th subpattern candidate are added to the end of the contents of the "subpattern code part" of the new character candidate. The content of the "subpattern code part" of the above new character candidate is "subpattern" in the "sc" th subpattern candidate.
When the sub-pattern code string obtained by adding the contents of the "code part" is the only sub-pattern code string in the character configuration table, it is composed of the sub-pattern code string in the character configuration table. The code of the character to be stored is stored in the "character code part" of the new character candidate.

In step S83, the new character candidate counter
new c The contents of count are incremented. After the content of the sub-pattern candidate counter sc is incremented in step S84, the process returns to step S77. In this way, the above operation is performed until the content of the sub pattern candidate counter sc becomes larger than the number of sub pattern candidates.
Then, in the above step S77, the number of sub-pattern candidates
If it is determined that the content of sc is larger than the number of sub-pattern candidates, the process proceeds to step S85. After the content of the character candidate counter cc is incremented in step S85, the process returns to step S74. Thus, the above-mentioned operation is performed such that the contents of the character candidate counter cc is the same as the current character candidate counter c. co
It is performed until it becomes larger than the contents of unt. Then, in step S74, the current character candidate counter c count
If it is determined that it is larger than the content of, the process proceeds to step S86.

In step S86, character candidates are packed. Here, the filling of the character candidates is performed as follows. That is, when the character expansion process is executed, the character candidates having the number of end strokes of interest are added to the second and subsequent sub-patterns and reborn as the new character candidates and newly registered. . Therefore, the character candidates having the number of end strokes of interest are no longer required. Therefore, from the first character candidate in the character candidate buffer above, "c Of the character candidates up to the "count" th character candidate, the number of character candidates having an end stroke number less than or equal to the noted end stroke number is counted to erase the character candidate counter e. Set to count. After that, the character candidates having the end stroke number equal to or less than the noted end stroke number are deleted, and the storage area of the non-erased character candidates is sequentially moved to the empty area on the low address side. Step S87
Then, the above character candidate erase count counter e Based on the contents of count, the current character candidate counter c The contents of count are updated according to equation (24). C COUNT ＝ (NEW COUNT) − (E COUNT) (24) where C COUNT: Current character candidate counter c contents of count NEW COUNT: New character candidate counter new c contents of count E COUNT: Character candidate erase count counter e Contents of count After that, the character expansion processing subroutine is terminated, and the processing returns to the character generation processing operation routine.

By executing the character expansion processing sub-routine in this manner, the sub-pattern candidates after the second sub-pattern candidate are added to the sub-pattern candidates less than the number of input strokes among the first sub-pattern candidates, Character candidates are obtained.

Hereinafter, the above-mentioned character expansion processing will be described more concretely, taking as an example the case where the character “gaku” is input by handwriting with the stroke number “5”. Here, as a result of the first sub-pattern recognition processing, it is assumed that the contents of the character candidate buffer are initialized as shown in FIG.
Also, the above-mentioned character generation processing ends at the minimum number of minks.
It is assumed that the contents of the sub-pattern candidate buffer are as shown in FIG. 24 after completion of k = 3.

First, the maximum distance register maxdist = 8
0 is set, new character candidate counter new c count = current
Character candidate counter c count + 1 = 4 is set. …
Steps S71, S72 Next, the character candidate counter cc = 1 is set, and
The set value is the current character candidate counter c count = 4 or less
Therefore, the "cc = 1" th sub-pattern from the character candidate buffer
The candidate "engine" is read. And the sub putter
The current end stroke number “3” of the candidate “Engine” ends
Since it matches the number of strokes (= minenstrk = 3),
Set the pattern candidate counter sc from "1" to "sub pattern
Sub-pattern candidates while incrementing to complement = 3 "
Sub pattern following the first sub-pattern candidate "Kou" from the buffer
The candidate candidates are searched. And the searched sub putter
The candidate "power", "one", "parallel" is attached to the first sub-pattern candidate "engine"
In addition, the sub-pattern code string "work", "force" and "work",
“Person”, “work”, and “parallel” are generated. However, in this case
Is a sub-pattern code string "Kou", "kata" and "Kou", "parallel"
Since it does not exist in the character composition table, the subpattern
Average distance c only for code string "Engineer" and "Force" dist becomes the formula (23)
Therefore, it is calculated. Thus, the average distance c calculated dis
t = 36 is less than maximum distance register maxdist = 80
Then, a new character candidate "Kou, Riki" is created and shown in Fig. 25.
Sea urchin "new c registered as count = 4 "th character candidate
Be done. In addition, the character candidate "Kou" is included in the "stroke part".
End stroke number = 3 and the sub-pattern candidate "power" strikes
"5" that is the sum of the number of roques = 2 is stored, and "sub pattern" is stored.
The number of sub-patterns of the character candidate "Kou" is set to 1
The value "2" obtained by adding "1" is stored, and the new character candidate count
New c count is incremented to "5". So
After that, the above character candidate counter cc is incremented
It ... Steps S73 to S84, Step S77, Step S8
5

"Cc = 2" th sub-pattern candidate "Sat"
Since the current end stroke number “3” of “3” matches the target end stroke number “3”, a sub pattern candidate following the first sub pattern candidate “Sat” is searched. Then, the sub-pattern / code sequence “Soil”, “Force” and “Sat”, “One” and “Sat”, and “Parallel” are generated. However, in this case, the subpattern
Since the code strings "earth", "force" and "earth", "parallel" do not exist in the character composition table, the average distance c of only the sub-pattern code series "earth", "one" dist is calculated. And the obtained average distance c Since dist = 61 is smaller than the maximum distance register maxdist = 80, a new character candidate “Sat, kata” is created, as shown in FIG.
As shown in “new c count = 5 Registered as the 5th character candidate, the contents of the "end stroke number part" and the "subpattern number part" are stored, and the new character candidate counter n
ew c count is incremented to "6". After that, the character candidate counter cc is incremented.
... Steps S74 to S84, Step S77, Step S85

The "cc = 3" th sub-pattern candidate "gas" is read. However, since the current end stroke number “4” in the sub-pattern candidate “gas” does not match the target end stroke number = 3, the character candidate counter cc is incremented. Steps S74, S75, S85 and thereafter, similar processing is performed, and character candidates as shown in FIG. 26 are developed in the character candidate buffer.

Next, in the character candidates in the character candidate buffer shown in FIG. 26, the "end stroke number" is equal to the noticed end stroke number (= minenstrk = 3).
Is deleted and the remaining character candidates are packed. Step S86 After that, after returning to the above character generation processing operation to sort the character candidates, the character candidate "gas" whose number of end strokes is less than the number "5" of input strokes is deleted. As a result, character candidates as shown in FIG. 27 are obtained. … Step S65

As described above, the first sub-pattern recognition process, the character generation process, and the character expansion process are performed, and
When a character candidate as shown in 7 is formed in the character candidate buffer, the output unit 85 displays the character code of the recognition result “gong” generated from the first character candidate “work, strength”. It is sent to the unit 76 and displayed.

As described above, in the present embodiment, the standard pattern registered in the sub-pattern dictionary 75 is provided with the header portion, and the header portion has the average value m ¹ , m of the directional difference between the two axes. ² and standard deviation values v ¹ and v ² are stored. Then, the direction difference determination unit 80 calculates the difference in the off direction to the next stroke with dd ¹ and dd ² by the equation (13), and the calculated results dd ¹ and dd ² and the standard pattern read from the sub-pattern dictionary 75. the average value of the direction difference is added to m ^1, m ²
It is determined that the standard pattern in which the standard deviation values v ¹ and v ² do not satisfy the equations (16) and (17) is not effective for DP matching. Therefore, later DP
The DP matching process performed by the matching unit 82 can be quickly performed using only valid standard patterns.

Further, the stroke number limiting unit 81 limits the regular stroke number of the characteristic pattern which is the target of the DP matching processing to a certain range according to the equation (18). Then, the DP by the DP matching unit 82
In the matching process, only standard patterns within the above-mentioned limited range are selected and used. Therefore, the DP matching process can be performed more quickly using only the standard pattern similar to the characteristic pattern.

Further, a character structure table in which characters and structure sub-patterns are associated with each other is registered in the recognition storage unit 74, and in the character generation processing by the character generation unit 84, the character structure table is referred to, Sub-pattern candidates following the recognition result of the first sub-pattern are limited. Therefore, the standard pattern used for DP matching in the character generation processing can be narrowed down according to the limited result, and the character recognition processing can be performed quickly.

That is, according to the present embodiment, it is possible to perform the recognition processing of the characters input by handwriting at high speed. This not only makes it possible to quickly obtain the recognition result, but also increases the number of variations of the standard pattern registered in the sub-pattern dictionary 25 to increase the recognition rate for "continuation characters" and "collapsed characters". Is possible.

In each of the above-mentioned embodiments, the feature re-extracting units 8, 18, 3 in the first to fourth embodiments are described for the sake of simplicity.
The feature re-extraction processing by 8, 58, the feature extraction processing and the first sub-pattern recognition processing in the fifth embodiment are
After the coordinate data string of the feature point extracted from the input pattern for one character is stored in the feature storage unit 6, 16, 36, 56, or the coordinate data string or feature pattern for one character is input to the input coordinate buffer. It is described as being executed after being stored in the feature buffer. However, usually
Since it takes about 2 seconds to write one character by a human, the writing time may be used to perform the feature re-extracting process or the feature extracting process and the first sub-pattern recognition process in parallel with the above-mentioned input process. For example, online handwriting recognition is possible.

In this case, the feature storage units 6, 16, 36, 56 in the first to fourth embodiments, or the input coordinate buffer, feature buffer, input data buffer, etc. in the fifth embodiment are dual-ported. -A memory may be used to write and read information in parallel. still,
Such online handwritten character recognition processing can be realized only when the handwritten character recognition processing can be executed at high speed as described above.

Distance between coordinate points Nim in the first embodiment.
in, direction difference Ndist, stroke end determination time Ntime, normalized range Nnorm, interpolation determination distance Nfmax, inter-feature point limit distance N
flen, noise judgment distance Nfmin, noise judgment direction difference Nddst,
Alternatively, the normalization range Nnorm, the interpolation interval w, the direction value quantization number D, the direction value R ^{j of} two orthogonal axes, and the constant C in the fifth embodiment.
Numerical values such as s, direction difference weight Wd, constant Cw, matching window range ww, etc. are merely examples, and may be appropriately set according to the recognition accuracy and the usage situation. Further, each of the above-mentioned embodiments may be carried out independently, but may be appropriately combined with each other. For example, the preprocessing in the first embodiment for extracting the features before and after the normalization is applied to the preprocessing in the fifth embodiment, or the DP in the third and fourth embodiments is applied.
It is conceivable to apply the matching weight change to the DP matching in the fifth embodiment.

When the preprocessing in the first embodiment is applied to the preprocessing in the fifth embodiment, it is possible to obtain a high recognition result for continuous characters as follows. Ie 4
The standard patterns of about 14,000 sub-patterns created from handwriting data of ten thousand characters are registered in the sub-pattern dictionary. Then, handwriting data was recognized by inputting handwriting data (about 400 characters × 67 names) of about 27,000 kanji characters in which a plurality of strokes were continuously written. As a result, as shown in Table 1, the correct answer rate for the first-ranked recognition candidate is as high as 96.0%, and it is even higher if the second-ranked and lower-ranked recognition candidates are included.
We were able to obtain a correct answer rate of%. This correct answer rate is significantly higher than that of a conventional handwriting recognition device (announced at D578, National Convention of the Institute of Electronics, Information and Communication Engineers, 1993).

[Table 1]

As described above, the capacity of the dictionary can be reduced by using the sub-pattern, which is a set of a plurality of strokes, as a recognition unit. This will be described with a specific example. Here, the recognition target pattern is “fast”. The number of types of strokes that can occur when handwriting the pattern "Haya" is "18". -The number of strokes that can be combined when forming the pattern "Haya" and the number of combinations in that case are as follows. Number of strokes Number of combinations Number of strokes Number of combinations 6 1 3 10 5 5 2 5 4 4 10 1 1 is assumed.

(A) Conventional Dictionary (Slotok Dictionary + Pattern Dictionary) In the conventional dictionary, all strokes that can occur when handwriting the pattern “Haya” are registered in the stroke dictionary to form the pattern “Haya”. All stroke combinations are registered in the pattern dictionary. Slotalk dictionary The storage capacity per stroke is stroke code = 2 bytes Directional sequence of 16 sections expressing stroke = 16 bytes Total = 18 bytes, the number of possible strokes is "18", so the stroke dictionary The minimum required capacity is 18 bytes x 18 strokes = 324 bytes.

Pattern dictionary: Storage capacity per pattern candidate: pattern code = 2 bytes Stroke number = 1 byte Direction information = 4 bytes Stroke code string = 2 bytes x number of strokes Position information between strokes = 1 byte x stroke Assuming that the total number = 7 bytes + 3 bytes × stroke number, the number of strokes that can be combined and the number of combinations generated in that case are as described above. Therefore, the minimum required capacity of the character dictionary is as follows. 1 x (7 bytes + 3 bytes x 6 strokes) + 5 x (7 bytes + 3 bytes x 5 strokes) + 10 x (7 bytes + 3 bytes x 4 strokes) + 10 x (7 bytes + 3 bytes x 3 strokes) + 5 x (7 bytes + 3 Byte x 2 strokes) + 1 x (7 bytes + 3 bytes x 1 stroke) = 560 bytes.

Required minimum capacity of conventional dictionary Required minimum capacity of stroke dictionary + Minimum required capacity of pattern dictionary = 324 bytes +560 bytes = 884 bytes

(B) Dictionary of this Embodiment (Sub-Pattern Dictionary) In the dictionary of this embodiment, the sub-pattern “Haya” shown in FIG.
Is registered in the sub-pattern dictionary.・ Sub-pattern dictionary Storage capacity per sub-pattern Sub-pattern code = 2 bytes Stroke number = 1 byte Coordinate number = 1 byte Direction information = 4 bytes Feature data = 4 bytes x number of feature points Total = 8 bytes + 4 bytes x Features 2 (e), the minimum required capacity of the sub-pattern dictionary is 8 bytes + 4 bytes × 20 points = 88 bytes.

As described above, the dictionary capacity can be greatly reduced by using subpatterns that are patterns that are statistically easy to write with one stroke. Therefore, it becomes possible to register a large number of modified patterns of sub patterns, and the recognition rate of handwriting patterns can be greatly improved.

[0173]

As is apparent from the above, the handwriting recognition apparatus according to the first aspect of the present invention interpolates the off-stroke and the sub-pattern dictionary that stores the standard pattern obtained from the sub-pattern and the modified pattern of each sub-pattern. Since the recognizing unit has the off-stroke interpolating means for performing the collation with the sub-pattern dictionary, the sequence of the feature points having the feature amounts related to the plurality of slot strokes and the off-strokes is regarded as one feature pattern. If you register a standard pattern of transformation patterns that include, "enter", "bounce", "continue", etc. in the above sub-pattern dictionary in advance, you can consistently recognize "continuation characters" and "collapsed characters" with a high recognition rate. It is possible to recognize handwritten characters and the like. Further, since the sub-pattern that is easily written in one stroke statistically is used as the unit for dictionary matching, the capacity of the matching dictionary can be reduced and the matching speed can be increased as compared with the case where strokes or one character are used as the unit.

That is, according to the present invention, the "continuation character" or the "collapsed character" in a poor hardware structure such as an electronic notebook.
It makes it easy to recognize online.

Further, the feature extraction unit in the handwriting recognition apparatus of the invention according to claim 2 connects between the first feature extraction unit for extracting the feature point and the feature amount of the bent portion in the handwriting pattern and the normalized feature point. Since the second feature extraction unit that extracts the feature points and the feature amount that are approximated at equal intervals is included, the feature point and the feature amount of the bent portion can be extracted from the coordinate data sequence before normalization. Therefore, even if the feature of the bent portion disappears due to the later normalization, the feature point and the feature amount representing the bent portion can be reliably extracted. Further, since the feature of the bent portion is not eliminated by the above normalization, feature extraction using both the equal division method and the angle method can be performed, and a feature amount sequence that accurately represents the feature of the handwriting stroke can be obtained. it can.

Further, the recognizing unit in the handwriting recognizing device of the invention according to claim 3 draws a table by the distance calculating means to calculate the feature amount of each feature point on the feature pattern and each feature point on the standard pattern. Since the difference is calculated and the distance between both patterns is calculated based on the value of each difference, the variable value for distance calculation of both patterns can be calculated only by a simple operation of drawing a table. Therefore, according to the present invention, high-speed dictionary matching can be performed.

Further, the recognizing unit in the handwriting recognizing apparatus of the invention according to claim 4 performs the DP matching processing in which the DP matching means carries out the DP weighting with the variation weight for each feature point on the pattern to be collated. Depending on the appearance of the feature along the stroke of the handwriting pattern, D
The time extension characteristic during P matching can be changed. Therefore, according to the present invention, it is possible to perform DP matching that emphasizes the characteristics of the handwriting pattern, and it is possible to further improve the recognition rate.

Further, the handwriting recognition apparatus of the invention according to claim 5 relates to the point on the matching path when the DP matching processing of the plurality of learning feature patterns and the standard pattern is performed by the variation weight calculation section. Since the matching distance is obtained and the degree of variation in the matching distances related to all learning feature patterns is used as the variation weight, DP matching is performed by emphasizing a section in which stable feature points are obtained. Therefore, according to the present invention, a high recognition rate can be stably obtained.

Further, in the handwriting recognition apparatus of the invention according to claim 6, the variation weight calculation unit calculates the ratio of the distance between the characteristic points on the characteristic pattern and the distance connecting all the characteristic points in the characteristic pattern in the input order. Since the value is used as the variation weight, DP matching that emphasizes the characteristic of the bent portion in the handwriting pattern is performed. Therefore, according to the present invention, it is possible to absorb the deviation of the bent portion and obtain a high recognition rate.

Further, the recognizing unit in the handwriting recognizing device of the invention according to claim 7 recognizes the handwriting pattern for each subpattern by the subpattern recognizing means, and combines the subpattern candidates obtained by the handwriting pattern recognizing means. Since the above-mentioned handwriting pattern is recognized by comparing with the representative pattern configuration table, it is possible to perform the recognition process in units of sub-patterns that are easy to write with one stroke. Therefore,
According to the present invention, it is possible to quickly recognize the writing pattern with a small number of recognition candidates. Further, it is possible to provide a handwriting recognition processing device that is strong against "continuation characters" and "collapsed characters".

In the feature extraction unit of the handwriting recognition apparatus according to the eighth aspect of the present invention, the bending feature extraction unit changes the direction of the straight line connecting the coordinate points in the coordinate data string by a predetermined value or more. Since the points are extracted as feature points, the straight-line feature extraction means extracts feature points that approximate the feature points at equal intervals, and the direction value calculation means calculates the direction value of each feature point as one of the feature amounts. The feature points can be extracted by using the equal division method and the angle method together, and the feature pattern that accurately represents the features of the handwriting pattern can be extracted.

The recognizing unit in the handwriting recognizing device of the invention according to claim 9 is characterized by the feature pattern selecting means.
Since only the standard pattern having the number of strokes within the range defined by the stroke number limiting unit as the header information is selected from the sub-pattern dictionary, dictionary matching can be performed quickly using only the highly similar standard pattern.

Further, the recognizing unit in the handwriting recognizing device of the invention according to claim 10 calculates the direction value from the final characteristic point of the characteristic pattern to the leading characteristic point in the next stroke by the direction value calculating means, and the effective standard Since the pattern determining unit determines only the standard pattern having the effective range in which the calculated direction value is included as the header information as the effective standard pattern, the dictionary matching can be quickly performed using only the effective standard pattern.

Further, the recognizing unit in the handwriting recognizing device of the invention according to claim 11 obtains, by the standard pattern selecting means, a sub pattern which can follow the recognized sub pattern candidate chain by referring to the representative pattern configuration table. Since only the standard pattern of the obtained sub-patterns is selected from the sub-pattern dictionary, it is possible to quickly perform dictionary matching using only the valid standard pattern.

Further, the handwriting recognition apparatus of the invention according to claim 12 has an input coordinate buffer and a feature buffer which are dual port memories, and is controlled by the control section.
Since the feature extracting unit and the recognizing unit are operated in parallel with the tablet, writing of the coordinate data string by the tablet to the input coordinate buffer, reading by the feature extracting unit, and feature pattern by the feature extracting unit for the feature buffer. And the reading by the recognition unit can be performed in parallel. Therefore, according to the present invention, the feature extraction processing and the recognition processing and the coordinate data string generation processing can be executed in parallel, and online handwriting pattern recognition can be performed at high speed. In other words, the present invention easily realizes online recognition of "continuation characters" or "collapsed characters" with a poor hardware configuration such as an electronic notebook.

[Brief description of drawings]

FIG. 1 is a block diagram of a handwriting recognition device according to a first embodiment of the present invention.

FIG. 2 is a diagram showing an example of a sub pattern.

FIG. 3 is a diagram showing an example of a recognition result of sub patterns that are vertically separable.

FIG. 4 is a diagram showing a configuration example of a character dictionary.

FIG. 5 is a flowchart of a handwriting recognition processing operation in the first embodiment.

FIG. 6 is an explanatory diagram of feature point extraction by a stroke processing unit in FIG. 1.

FIG. 7 is a correspondence diagram of an input coordinate sequence and a feature point sequence.

FIG. 8 is a block diagram of a handwriting recognition device according to a second embodiment of the present invention.

FIG. 9 is an explanatory diagram of an absolute difference table and a distance table.

FIG. 10 is an explanatory diagram of a direction difference table.

FIG. 11 is a block diagram of a handwriting recognition device according to a third embodiment of the present invention.

FIG. 12 is a block diagram of a handwriting recognition device according to a fourth embodiment of the present invention.

FIG. 13 is a block diagram of a handwriting recognition device according to a fifth embodiment of the present invention.

FIG. 14 is a flowchart showing a general flow of a handwritten character recognition processing operation performed under the control of the control unit in FIG.

15 is an explanatory diagram of a structure of a header added to the sub pattern dictionary in FIG.

FIG. 16 is a flowchart of a first sub-pattern recognition processing operation.

FIG. 17 is an explanatory diagram of a structure of a sub pattern candidate buffer.

FIG. 18 is a flowchart of a sub pattern recognition processing subroutine.

FIG. 19 is a flowchart of a character generation processing operation.

FIG. 20 is an explanatory diagram of a structure of a character candidate buffer.

FIG. 21 is a schematic diagram showing an example of the contents of a character candidate buffer after initialization.

FIG. 22 is a flowchart of a character expansion processing subroutine.

FIG. 23 is a schematic diagram showing the contents of a character candidate buffer after initialization different from that of FIG. 21.

24 is a schematic diagram showing the contents of the sub-pattern candidate buffer corresponding to the contents of the character candidate buffer shown in FIG. 23.

25 is a diagram showing changes in the contents of the character candidate buffer shown in FIG. 23 during the character expansion processing operation.

FIG. 26 is a diagram showing a transition of the contents of the character candidate buffer subsequent to FIG. 25.

27 is a diagram showing the contents of the character candidate buffer when the character candidates are packed and sorted with respect to the contents of the character candidate buffer shown in FIG. 26.

FIG. 28 is an explanatory diagram of the influence of normalization processing on a conventional feature point.

[Explanation of symbols]

1, 11, 31, 51, 71 ... Recognition unit, 2, 12, 32, 52,
72 ... Tablet, 3,13,33,53,76 ... Result display section 5,15,35,55 ... Stroke processing section 6,16,
36, 56 ... Feature storage unit, 8, 18, 38, 58 ... Feature re-extraction unit, 9, 19, 39, 59 ... Character recognition unit, 10, 20, 4
0,60 ... dictionary storage unit, 41, 61 ... weight calculation unit,
73 ... Feature extraction storage unit, 74 ... Recognition storage unit,
75 ... Sub-pattern dictionary, 78 ... Normalization unit, 79 ... Feature extraction unit, 80 ... Direction difference determination unit, 81 ... Stroke number limiting unit,
82 ... DP matching unit, 84 ... Character generation unit, 86 ... Control unit, 87 ... Character candidate expansion unit.

Front page continued (72) Inventor Yoshihiro Kitamura 22-22 Nagaike-cho, Abeno-ku, Osaka-shi, Osaka

Claims (12)

[Claims] 1. A tablet that generates a coordinate data string formed by arranging the coordinates of handwriting input by handwriting in time series, a feature point representing a feature of a handwriting pattern from the coordinate data string, and a feature amount of the feature point. A feature extraction unit to extract, a recognition unit that recognizes a handwriting pattern by matching a feature pattern consisting of a sequence of feature points having the extracted feature amount with a dictionary, and a result display that displays the recognition result by this recognition unit In the handwriting recognition device having a section, the dictionary standardizes a sequence of feature points having a feature amount extracted from a subpattern formed by separating a representative pattern into a pattern that is statistically easy to write with one stroke and a modified pattern of each subpattern. It is a sub-pattern dictionary for patterns, wherein the feature extraction unit extracts the feature points and the feature amount from the coordinate data sequence of a plurality of strokes, and The substroke dictionary includes off-stroke interpolation means for interpolating off-strokes with feature points, and the recognition unit considers a feature point sequence having feature amounts related to the plurality of strokes and off-strokes as one feature pattern. A handwriting recognition device characterized by performing collation with. 2. The handwriting recognition device according to claim 1, wherein the feature extraction unit extracts a feature point and a feature amount of a bent portion in a handwriting pattern from the coordinate data sequence generated by the tablet. And a feature point sequence extracted by the first feature extraction unit, and a feature that approximates the feature points of the bent portion at equal intervals from the normalized feature point sequence and the normalized value of the coordinate data. A handwriting recognition device comprising a second feature extraction unit for extracting points and feature amounts. 3. The handwriting recognition apparatus according to claim 1, wherein the feature amount of the feature points on the feature pattern obtained by the feature extraction unit, the feature amount of the feature points on the standard pattern, and the difference between the two feature amounts. And a table storage unit that stores a table in which the values are associated with each other, and the recognition unit determines a feature amount of a feature point pair including each feature point on the feature pattern and each feature point on the standard pattern. Based on the above, the table is drawn to obtain the difference between the feature amounts for each of the feature point pairs, and a distance calculating means for calculating the distance between the feature pattern and the standard pattern based on the value of each difference is provided. Characterized handwriting recognition device. 4. The handwriting recognition device according to claim 1, wherein the recognition unit includes a DP matching unit that performs a DP matching process in which a variation weight is given to each feature point on a pattern to be matched. A handwriting recognition device characterized by the above. 5. The handwriting recognition apparatus according to claim 4, wherein the feature extraction unit and the DP matching unit are controlled, and the feature extraction unit comprises a feature point sequence having a feature amount from the learning pattern prepared in advance. A learning characteristic pattern is generated, and a DP matching process between the learning characteristic pattern and the standard pattern is performed by the DP matching means to obtain a matching distance associated with a point on a matching path. 1. A handwriting recognition apparatus, further comprising: a variation weight calculation unit that calculates a degree of variation in matching distances related to all learning feature patterns and uses the variation weight as the variation weight. 6. The handwriting recognition device according to claim 4, wherein the distance from a feature point on the feature pattern extracted by the feature extraction unit to the next feature point and all feature points in the feature pattern are connected in the input order. A handwriting recognition device comprising: a variation weight calculation unit that calculates a value of a ratio to a distance and uses the variation weight for the section related to the feature point. 7. The handwriting recognition apparatus according to claim 1, further comprising: a representative pattern configuration table storage unit that stores a representative pattern configuration table that associates a representative pattern with sub-patterns that form the representative pattern, The recognition unit collates the feature pattern with a sub-pattern dictionary to recognize the handwriting pattern for each sub-pattern, and a sub-pattern candidate obtained as a result of recognition by the sub-pattern recognition unit. A handwriting recognition device comprising a handwriting pattern recognition means for recognizing the handwriting pattern by collating a combination with the representative pattern configuration table. 8. The handwriting recognition apparatus according to claim 7, wherein the feature extraction unit uses a change point at which a direction of a straight line connecting the coordinate points in the coordinate data string changes in a predetermined value or more as a feature point. Bent portion feature extracting means for extracting, straight line feature extracting means for extracting feature points that approximate the feature points extracted by the bend portion feature extracting means at equal intervals, and each feature extracted by both the feature extracting means A handwriting recognition device comprising a direction value calculation means for calculating a direction value representing a direction to a next characteristic point at one point as one of the characteristic amounts. 9. The handwriting recognition apparatus according to claim 8, wherein an approximate straight line number of the handwriting pattern is obtained from the number of characteristic points of the bent portion extracted by the bent portion feature extracting means, and the approximate straight line number and the approximate straight line number are calculated. The stroke number limiting unit that determines the range of the stroke number of the characteristic pattern from the stroke number of the handwriting pattern is provided, and the sub-pattern dictionary includes, for each standard pattern, the stroke number of the representative pattern that is the source of the standard pattern. It has header information, and the recognition unit includes a characteristic pattern selection unit that selects only a standard pattern having the number of strokes within the range defined by the stroke number limiting unit as header information from the sub-pattern dictionary. Characterized handwriting recognition device. 10. The handwriting recognition apparatus according to claim 8, wherein the sub-pattern dictionary includes, for each standard pattern, an effective range of a direction value from a final feature point of the standard pattern to a leading feature point in a next stroke. The standard pattern having the header information is a valid range in which the direction value from the final feature point in the feature pattern calculated by the direction value calculating means to the head feature point in the next stroke is included in the recognition unit. A handwriting recognition device comprising an effective standard pattern determining means for determining only an effective standard pattern. 11. The handwriting recognition apparatus according to claim 7, wherein the recognition unit obtains a sub-pattern that can follow a sub-pattern candidate chain of recognized handwriting patterns by referring to the representative pattern configuration table, A handwriting recognition apparatus comprising a standard pattern selection means for selecting only the standard pattern of the obtained sub patterns from the sub pattern dictionary. 12. The dual-port memory according to claim 1, wherein the coordinate data sequence generated by the tablet is written, and the written coordinate data sequence is read by the feature extracting unit. An input coordinate buffer, a feature buffer extracted from the feature extraction unit, and a feature buffer including a dual port memory from which the written feature pattern is read by the recognition unit; and the feature extraction unit. And a control unit that controls the recognition unit to execute the feature extraction process by the feature extraction unit and the recognition process by the recognition unit and the coordinate data string generation process by the tablet in parallel. Recognition device.