JPS62229383A - On-line character recognizing device - Google Patents

Abstract

PURPOSE:To decrease the influence of the variance of a character shape by providing a large group classifying part by the number of the strokes of a writing character and a middle group classifying part by the vector between part patterns, selecting a candidate character and executing the sequencing with a part pattern Q value. CONSTITUTION:A large group classifying part 5 execute the large group classification by the stroke of the writing character of the output of a stroke coding part 4. A middle group classifying part 6 calculates the center of gravity of respective part patterns from the segment information to link the section between characteristics points extracted by a characteristic point extracting part 3, obtains the vector between part patterns between the respective centers of gravity centers, executes the matching with the vector between the part patterns prepared beforehand from a registering pattern and selects a candidate character. A part pattern Q value matching part 7 executes the matching with the part pattern Q value to show the characteristic of a part pattern and determines the sequence of the candidate character. A part pattern stroke code part matching part 8 determines a final sequence by the part pattern stroke code distribution.

Description

DETAILED DESCRIPTION OF THE INVENTION (Field of Industrial Application) The present invention relates to an online character recognition device that identifies written characters in real time.

(Prior Art) The following method is generally used as a recognition method in an online character recognition device. The data string of the input stroke (the written part from pen-on to pen-off) is
The feature points are expressed in Y coordinates or r and θ polar coordinates, and feature points are extracted from these data (hereinafter referred to as input cover patterns). Character recognition was performed by matching information on the extracted feature points with information on a pattern (hereinafter referred to as registered pattern) in which feature points were extracted and registered in advance using the same method.

Hereinafter, a conventional online character recognition method will be explained with reference to FIGS. 10 to 12.

First, FIG. 10 shows an example in which feature points are expressed by the start and end points of each stroke. The j stroke data string IDj input by hand is rDj=((xl, ys)j), i=1.2. ,,n-
= (1) nj'' If the number of data for j strokes is, the starting point (Xs + ys)' is (XI r y
t ) j, the end point (Xe + ye)j is (Xnj
l ynj) Δ.

Here, the registered patterns found and registered in a similar manner are assumed to be (xs, ys) JI and (Xe, ye) JI. When a simple matching method is adopted, when the number of character strokes is N, the distance d between the input cover pattern and the registered pattern is calculated according to the following. Then, matching with each registered pattern is performed, and among them, one with a small distance d is output as a recognition result.

(Problems to be Solved by the Invention) However, the conventional online character recognition method using such a matching method has the following problems.

According to this method, a handwritten item with a shape similar to the registered pattern is recognized with a recognition rate of nearly 100%. However, in general, there are many cases where the handwritten input pattern does not have the same shape as the registered pattern. Figure 11 (
A) “River” whose shape is deformed with respect to the registered pattern
An example of the input cover pattern is shown in FIG. 11(b), and a registered pattern corresponding to this is shown in FIG. In this case, "shi" and "" are shifted up and down, and their respective sizes vary greatly. Therefore, the distance d between the input cover pattern and the registered pattern becomes large, increasing the possibility of misrecognition. As described above, according to the conventional method, there is a problem in that when a handwritten input is performed in a shape that is modified from the registered pattern, the misrecognition rate increases.

Furthermore, in a simple matching of the order of written strokes based on equation (2), the distance d becomes large when the stroke order of the input characters is different. Therefore, in reality, processing is required to optimally associate matching strokes.

That is, it is necessary to determine which stroke of the registered pattern should be matched with each handwritten stroke. This process consumes a lot of time.

For example, let's consider the optimal correspondence when "shi'°" as shown in Fig. 12(a) is input by hand. Fig. 12(a)
b) is a diagram showing the correspondence registration pattern at this time, and FIG. 12(C) is an explanatory diagram of the optimum stroke correspondence processing.

One way to make the correspondence is to first find the distance dll between the first stroke of the incoming turn (■ in Figure 12(a)) and the first stroke of the registered bat-axe turn (■ in Figure 12(b)). is calculated using formula (2), and then the registered pattern ■
distance d, 2. The distance d13 from the registered pattern ■ is calculated, and the minimum stroke of dII + d12 + d13 is associated.

In this example, d11 is the minimum, and 9■-hang 1 is associated as the optimal stroke. Similarly, the second stroke of the input cover pattern (■ in FIG. 12(a)) is associated with the remaining strokes excluding the registered pattern strokes that have already been associated (■ in this case). Then, the respective optimal strokes are associated with each other.

This processing takes less time when the number of strokes is small, but for characters with a large number of strokes (characters with a large number of strokes), the number of processing M is as follows, where the number of strokes is N, M= Σ i ... ( 3) Since it increases at the ratio of i-+, the amount of processing becomes enormous. In order to compensate for this increase in processing amount, it is necessary to use a high-speed ALU or the like, so conventional online character recognition devices cannot be miniaturized and have the disadvantage of increasing costs.

As described above, the conventional method is sensitive to changes in the shape of the input cover pattern input by handwriting, and is prone to misrecognition.Additionally, it requires processing to perform optimal stroke correspondence in order to relax the stroke order, which requires a large amount of processing. There was a drawback.

SUMMARY OF THE INVENTION An object of the present invention is to solve the problems of the prior art described above, and to provide an excellent online character recognition device that is resistant to deformation of input cover patterns input by hand, and that requires less processing. do.

(Means for Solving the Problems) The present invention includes a preprocessing unit that removes unnecessary data from a coordinate data string obtained by handwriting input on a tablet, and performs linearization processing; A feature point extraction unit that extracts feature points representing characteristics of strokes that make up a written character from a coordinate data string, and a stroke coding unit that codes each stroke based on the positional relationship of the feature points extracted by the feature point extraction unit. The present invention aims at an online character recognition device that performs character recognition by comparing the output data of the feature point extraction section or the stroke coding section with registered pattern data registered in advance, and solves the problems of the conventional technology. In order to solve the problem, there is a large classification section that receives the output of the stroke encoding section and roughly categorizes the written characters according to the number of strokes; The part is treated as a partial pattern, and the center of gravity of each partial pattern is calculated from the information on the segments connecting the feature points extracted by the feature point extraction section. Based on the calculated center of gravity data, the center of gravity of one partial pattern and that of another partial pattern are calculated. Find vectors between sub-patterns with the center of gravity as the starting point and end point, respectively, match these inter-sub-pattern vectors with vectors between sub-patterns previously created from registered patterns in the same way, and perform medium classification according to the matching results. For the candidate sentences narrowed down by the intermediate classification unit, a partial pattern Q value representing the characteristics of the partial pattern extracted from each stroke segment information of the partial pattern is determined, and this partial pattern Q value and The partial pattern Q is matched with the partial pattern Q value created in advance from the registered pattern in the same way, and the candidate characters are ranked according to this matching result and the matching result by the intermediate classification section.
For each partial pattern, for the top candidate characters ranked by the value matching section and the partial pattern Q value matching section,
The stroke code distribution extracted by the stroke coding unit is calculated, and this stroke code distribution is matched with a partial pattern stroke code distribution previously created from registered patterns in the same way, and candidates are generated according to this matching result. The system also includes a partial pattern stroke code distribution matching section that performs final ranking of characters.

(Function) In the present invention, since the online character recognition device is configured as described above, each technical means functions as follows.

The preprocessing unit performs noise removal processing, moving average processing, smoothing processing, etc. on the character coordinate data string handwritten on the tablet, and linearizes the data. For example, the feature point extraction unit may be configured to
Feature points are extracted by a method such as finding a sign in the Y direction and using a change point in the state of the sign as a feature point. The stroke encoding unit classifies strokes based on the extracted feature point information, for example, by the sign of each segment, the angle of the segments, and the rotation angle between the segments, and encodes the strokes.

This coded stroke data is used for the following classification and ranking of each means, which is a feature of the present invention.

The major classification section performs major classification based on the number of strokes of written characters. For this reason, for each number of strokes, characters corresponding to the number of strokes are stored in advance in a character dictionary, for example.

The intermediate classification section classifies candidate characters roughly classified according to the number of strokes by 1 based on the matching of vectors between partial patterns, and further narrows down the candidate characters. Specifically, a vector between partial patterns is obtained for each candidate character, and this is compared with, for example, an inter-pattern vector in a character dictionary, and the difference is obtained as the distance between both vectors. This distance is compared with a predetermined threshold value, and if it is larger than the threshold value, it is removed from the candidate characters, and if it is less than the threshold value, it is left as a candidate character.

The partial pattern Q value matching unit ranks the medium-classified candidate characters by matching the partial pattern Q values. For this reason, the partial pattern Q value obtained in advance from the registered pattern is stored in, for example, a partial pattern dictionary. Note that matching data from the intermediate classification section is also used for this ranking. The partial pattern Q value represents the length, direction, position, etc. of each segment, and by effectively using this information, effective ranking can be performed.

The partial pattern stroke code distribution matching section performs final ranking on the thus ranked top candidate characters by matching the partial pattern stroke code distribution (7).

Therefore, it is possible to deal with fluctuations due to variations in the shapes of characters input by hand, eliminate the need for optimum correspondence processing of strokes due to differences in stroke order, reduce the amount of processing, and solve the problems of the prior art.

(Embodiment) Hereinafter, an online character recognition device according to an embodiment of the present invention will be described in detail.

FIG. 1 is a block diagram showing the configuration of the online character recognition device of this embodiment. As shown in the figure, this online character recognition device includes a tablet 1, a preprocessing section 2, a feature point extraction section 3, a stroke encoding section 4, a major classification section 5, a medium classification section 6, and a partial pattern Q value matching section. 7 and a partial pattern stroke code distribution matching section 8. Note that 9 in the figure indicates an output terminal to a display device (not shown) or the like.

FIG. 2 is a flowchart showing an outline of the operation of the apparatus shown in FIG. The major classification section 5 executes step 104, and the intermediate classification section 6 executes steps 105 to 1.
08, the partial pattern Q value matching unit 7 executes steps 109 to 112, and the partial pattern stroke code distribution matching unit 8 executes steps 113 to 116.

FIG. 3 is a diagram showing the structure of a character dictionary included in the apparatus of this embodiment, and this dictionary allows characters to be selected based on the number of strokes.

Further, FIG. 4 is a diagram showing the configuration of a partial pattern dictionary included in the apparatus of this embodiment.

Hereinafter, the operation of the apparatus of this embodiment will be explained in detail step by step for each part.

First, the tablet l is used to input characters by hand, and when characters are input by hand, a string of writing data ((xi+yi), i=1.2-nj)j is generated as shown in FIG. 5(a).
is extracted and sent to the preprocessing section 2.

The preprocessing unit 2 linearizes the data as shown in FIG. 5(b) by performing noise removal processing, moving average processing, and smoothing processing on the received writing data string, and then the feature point extraction unit 3 Output to.

Next, the operation of the feature point extraction section 3 will be described. There are several methods for feature point extraction processing performed by the feature point extraction unit 3, but here we will use a linearized data string ((Xi +
yi)i=1.2. -X between data of njlj.

A method of calculating the sign (positive, negative, and 0 signs) in the X direction and extracting points of change in the sign state as feature points will be described.

Find the X, X direction sine between the data, XS・, YSi, +, 0. Expressed with -. Compare the signs in the X direction and X direction between each piece of data obtained in this way with the signs between the previous data, and if they are the same, they are not registered as feature points, and if they are different, the state is considered to have changed and the feature point is Register as. In addition to the points obtained in this way, Fig. 5 (C) shows the starting point,
Shows the feature points plus the end point. Generally, this process is sometimes called linear approximation. The space between these feature points is hereinafter referred to as a segment, and the feature points are ((X; , Y; ), j−”1
, , , tj)j.

The feature point information obtained as described above is output to the stroke coding section 4 and the medium classification section 6.

Next, the operation of the stroke coding section 4 will be described. The stroke encoding section 4 encodes each stroke based on the feature point information obtained by the feature point extraction section 3. There are many methods for this encoding, but in general, the encoding is performed by classifying the X and Y signs of each segment, the angle of the segment, and the rotation angle between the segments. FIG. 6 is an explanatory diagram of this encoding process, in which θ2. θ2. θ3 indicates the angle of the segment (the angle formed in the +xX direction), iz,
, a2 indicates the rotation angle between nine adjacent segments. The coded stroke data is output to the major classification section 5 and the partial pattern stroke code distribution matching section 8.

Next, the operation of the major classification section 5 will be described. The major classification section 5 receives the output from the stroke encoding section 4 and performs major classification of the target characters based on the number of strokes. Therefore, for each number of strokes, nine characters corresponding to the number of strokes are prepared in advance in a character dictionary as shown in FIG. For example, assume that the number of strokes of a character pattern input by hand is 10 strokes. In this case, among the characters stored in the character dictionary, characters "kuchia", "koi", "ai", etc., which can have 10 strokes as shown in FIG. 3, are selected as candidate characters.

Next, the operation of the intermediate classification section 6 will be described. The middle classification section 6 is
The candidate characters obtained by major classification according to the number of strokes in the major classification section 5 are further divided into medium classifications using vectors between partial patterns, which will be described below. Here, a partial pattern refers to a part of one character that is written as a series of characters, and a vector between partial patterns refers to the center of gravity of one partial pattern and the center of gravity of another partial pattern as the starting point and ending point, respectively. is the vector.

First, an example of a method for calculating vectors between partial patterns will be described. X for each segment in the partial pattern.

When the y component is (dxi, dyi), the length dll of each segment is expressed as dli = ('; Let the center coordinates of each segment be (Xi + yi), then the barycenter coordinates of the partial pattern (XW).

Yw) is calculated by Using the above method, the center of gravity of each partial pattern is determined, and a partial pattern vector is determined using the center of gravity of one partial pattern and the center of gravity of another partial pattern as the starting point and end point, respectively. Note that here, the inter-partial pattern vectors are considered separately for the X direction and the X direction. Figure 7 shows the partial pattern “mouth” of the character “all”, each center of gravity of “t”, and the vector between partial patterns X ((II l stop)
, y(a, sub).

The medium classification section 6 performs medium classification by narrowing down the candidate characters selected by the main classification section 5 using the inter-partial pattern vectors explained above.
Let us consider the case where the input character is input by hand, and the procedure for medium classification for this input character will be explained below.

Since the handwritten character "Ai" has 10 strokes, first, the 10 stroke part of the character dictionary shown in FIG. 3 is referred to. Then, the character "at" is placed first here, and in that column there are the partial patterns that make up "at", the handwriting of the partial patterns, and the information on the number of strokes for each partial pattern (hereinafter referred to as cut position). ), and vector values between each partial pattern calculated in advance from the registered pattern are shown.

In the following order, similar information is arranged for the characters ``゛welt'', ``逢'', etc., and the intermediate classification unit 6 determines whether or not each character should be a candidate according to the character, and divides it into intermediate classifications as follows. Do the following.

First, assuming that the characters input by hand are 0&==, the matching distance 1itltdvec of vectors between partial patterns is determined. As written in the character dictionary in Figure 3, “o
f'', the cut position is (3,7), that is, the 1st to 3rd strokes form an ``O'', and the 4th to 10th strokes form a ``late hole'', but in this example, the input cover turn is Since it is “Ai”, 1 at this cut position.
If we consider the inter-subpattern vector for ``Ai'', it becomes as shown in Fig. 8.The inter-partial vector of the input pattern obtained as shown in Fig. 8 ~ x (o, Omi), i (o, t)
and the vector between partial patterns written in the dictionary X(O,
A>.

The difference from 7(C1, I) is the matching distance avec, and dvec = l x (σ;, taste) - ma (b, taste) 1
+1 bu (σ1g-) - bu (ro, sub) 1 ... (calculated by force).

Generally, when the number of partial patterns of a written character is multiple,
The matching distance avec is calculated according to the following by normalizing with the number of partial patterns BPN.

Here, a certain threshold VECREJ is set, and the calculated dve
It is determined whether c is larger than VECIJ.

When dvec>vECREJ, it is assumed that the character is not the referenced character (in this case, "at"), and matching of vectors between partial patterns of the next character is performed.

When avec≦VECREJ, it is assumed that the partial pattern is likely to be “0 minister”, and calculation and matching of the partial pattern Q value, which will be explained next, is performed.

Next, the operation of the partial pattern Q value matching section 7 will be described. The partial pattern Q value matching unit 7 calculates the partial pattern Q value for the candidate characters remaining after the intermediate classification using the inter-partial pattern vectors in the intermediate classification unit 6, and matches it with the partial pattern Q value in the partial pattern dictionary shown in FIG. Do the following. The partial pattern Q values of this partial pattern dictionary are created and stored in advance from registered patterns. Here, the partial pattern Q value refers to a characteristic parameter representing the length, direction, and position of each segment. In online character recognition, the X, Y directions, + and - directions are obtained as important information regarding the movement of the writing pen, and this partial pattern Q value is an effective use of this information.

First, a method for calculating the partial pattern Q value will be explained.

Note that in the formula below, Σ indicates addition regarding all strokes and all segments. Further, HX and HY indicate character width.

■ X-direction position of +X-direction component■ -X-direction position of X-direction component■ +yX-direction component y-direction position■ -X-direction position of y-direction component■ +X-direction position of X-direction component■ -X-direction position of X-direction component ■ +X-direction position of the y-direction component■ -X-direction position of the y-direction component■~■ In the case of the origin, it is the value of each direction position when the origin is set to the lower left, but in this case, the value near the origin is multiplied. When the origin is set to the upper right, the values Q9 to Q18 in each direction are written in the same way, and a total of 16 values are obtained from Q1 to Q. The length, direction, and position of each stroke segment of the target character are fully expressed by the value of .

The partial pattern Q value matching unit 7 calculates the partial pattern Q value described above for those remaining after classification based on vectors between partial patterns.
Suppose that ``逢'' is input into the car, and ``farewell'' remains after classification based on vectors between partial patterns. In this case, the cut position of "dust" is tll [as shown in Figure 3, it is cut at (3, 2, 5), and the partial pattern is 1+h+height, so the incoming cover turn t cut position is (3, 2, 5). Cut with Q1~
Calculate ax QCs. At this time, in this example, as shown in FIG. 9, "person" becomes "r", "two" becomes "ha", and "tsuji".

correspond to "粱" respectively. The calculated partial pattern Q values ˜Q+a are matched with the partial pattern Q values in the partial pattern dictionary shown in FIG. 4. That is, first, Q I-Q +a calculated for "person" is matched with Q1 to Qla of the partial pattern dictionary 14", and then "
Q1~Q+e calculated in “2” and partial pattern dictionary “Ha”
Matching Q1~Qta of ``Tsuji'', Q1~Q of partial pattern dictionary ``Kyu'' with Q, ~Q, a calculated by "Tsuji"
Match 16. The sum of the differences in these matchings is defined as the matching distance dBP. At this time, the distance dBP represents how close the input cover turn "Ai" is to "1Ki".

Generally, I is determined by the number of stores Bsj of each partial pattern.
) Weighting is performed as shown in equation (17), and this is taken as the matching distance dBP.

Then, a distance di is obtained by adding the distance dBP obtained as above and dVeC obtained by matching the inter-partial pattern vectors obtained in the previous step.

di = dvec + dnp −(
The operations above 181 are performed on all candidate characters remaining after classification based on inter-subpattern vectors, and sorting based on di is performed.

Next, partial pattern stroke code distribution matching section 8
We will describe the operation of. The partial pattern stroke code distribution matching section 8 is the partial pattern Q value matching section 7.
In response to the output from the stroke coding unit 4, a partial pattern stroke code distribution is obtained for the candidate characters narrowed down by the intermediate classification, and the partial pattern stroke code distribution is created in advance from this distribution and the registered pattern and stored in the partial pattern dictionary shown in FIG. Matching with the pattern stroke code distribution is performed, and the top candidates are further ranked. The range of targets for this ranking is determined, for example, by the ratio to the distance d of the first candidate obtained by sorting di. i.e. d
Ranking is performed with candidate characters of j/d, ≦ZRATE as the target range.

Here, a method for calculating the partial pattern stroke code distribution will be explained. -For example, the first cover pattern is “Ai”.
Assume that the character selected as a candidate is "Ai". Third
From the character dictionary in the figure, the cut position of the candidate character “Ai” is (3,
4, '3) shows that the partial pattern is L+'l-+'5, so cut the incoming cover turn "Ai" at this position. In this case, the partial patterns obtained by cutting are the same as the contents of the character dictionary, but the distribution of the number of stroke codes obtained by the stroke coding section 4 is calculated for each partial pattern. For example, if you look at the phantom partial patterns, they are "〆", "Z", and "-", so
The stroke chord distribution is determined in the form of one line for l, one line for s03, and one line for s05.

The partial pattern stroke code distribution calculated in this way matches the partial pattern stroke code distribution of the partial pattern dictionary shown in Figure 4, which was created by calculating and averaging several registered patterns in advance using the same procedure. be done. In the case of this example, first, the difference between the partial pattern stroke code distribution of "mata °" calculated as above and the stroke code distribution of "ku" in the partial pattern dictionary in Fig. 4 is calculated by the matching distance ds (, L).

That is, input cover turn partial pattern sQl...1 line sol...0.9 line 302.
...0 pieces 302...0.1 pieces s03...1 piece
s03...0.4 pieces s04 =-0 pieces s04
-= 0.6 lines 305...1 line s05...1 line difference 0.1 + 0.1 + 0.6 + 0.6 =
1.4 lines are obtained as d s (4). Similarly, matching is performed for Ill I n, @l,'' to obtain ds(1), ds(L), and the total distance d8 of each partial pattern 9 matching distance is ds''ds(also)+ds()) Calculated from 10ds(L) -u9.

The partial pattern stroke code distribution matching section 8 calculates this ds for each candidate character of dj/d, ≦ZRATE, and ranks the candidate characters according to the calculation result. Then, the recognition result is outputted from the output terminal 9 to a display device (not shown) or the like.

Although not mentioned in the above embodiment, partial patterns "+" and "dust" are used for those whose handwriting shapes are very similar in the ranked recognition results, such as "dust" and "dust".
A more reliable recognition result can be obtained by providing a detailed check section that recognizes "dust" if there is an intersection of ○ marks as a difference between "A" and "dust" if not.

In addition, in the explanation of the above embodiment, the partial pattern Q value matching unit 7 avec + (lap
Although it has been described that the candidates are ranked based on the added value of dBP, it is also possible to rank the candidates only based on dBP.

-Furthermore, in addition to narrowing down the candidates by matching vectors between partial patterns in the medium classification unit 6, a process is added to exclude candidates when the distance is greater than a certain threshold by partial pattern stroke code distribution matching, and candidates are subjected to post-processing. It is also possible to narrow down the characters.

(Effects of the Invention) As described in detail above, according to the present invention, by extracting and matching feature parameters for each partial pattern, it is possible to eliminate variations in shapes input by hand, which is a drawback of pattern matching methods (especially Since fluctuations in distance values due to positional deviations between partial patterns constituting characters are absorbed, erroneous recognition caused by these fluctuations can be prevented. Also,
Narrowing down the candidates by matching vectors between partial patterns and matching using partial parameter Q values representing information on each segment of each stroke for each partial pattern eliminates the need for optimal stroke matching processing and shortens the processing time. Ru. As described above, the present invention can provide an excellent online character recognition device that is resistant to variations in writing shapes and that requires less processing.

[Brief explanation of drawings]

FIG. 1 is a block diagram showing the configuration of an online character recognition device according to an embodiment of the present invention, FIG. 2 is an operation flowchart of the device shown in FIG. 1, FIG. 3 is a diagram showing an example of the configuration of a character dictionary, and FIG. The figure shows an example of the structure of a partial pattern dictionary, FIG. 5 is an explanatory diagram of preprocessing, FIG. 6 is an explanatory diagram of stroke encoding processing, FIG. 7 is an explanatory diagram of vectors between partial patterns, and FIG. FIG. 9 is an explanatory diagram of calculation of vectors between partial patterns. FIG. 9 is an explanatory diagram of partial pattern Q value calculation. FIG. 10 is a diagram showing feature points of input cover turns and registered patterns. FIG. 12 is an explanatory diagram of stroke association processing. DESCRIPTION OF SYMBOLS 1... Tablet, 2... Preprocessing part, 3... Feature point extraction part, 4... Stroke encoding part, 5... Main classification part, 6... Middle classification part, 7. . . . Partial pattern Q value matching section, 8 . . . Partial pattern stroke code distribution matching section, 9 . . . Output terminal.

Claims (1)

[Scope of Claims] A preprocessing unit that removes unnecessary data from a coordinate data string obtained by handwritten input on a tablet and performs a linearization process, and a written character from the coordinate data string linearized by the preprocessing unit. A feature point extraction unit that extracts feature points representing the characteristics of the strokes that make up the stroke, and a stroke coding unit that codes each stroke based on the positional relationship of the feature points extracted by the feature point extraction unit. An online character recognition device that performs character recognition by comparing the output data of the stroke encoding section or stroke encoding section with registered pattern data registered in advance, receives the output of the stroke encoding section and roughly categorizes the written characters according to the number of strokes. The major classification section performs the following: The part of the candidate character selected by the major classification section that is written as a series in handwriting is defined as a partial pattern, and each character is extracted from the segment information connecting the feature points extracted by the feature point extraction section. Calculate the center of gravity of the partial patterns, and based on the calculated center of gravity data, find a vector between partial patterns whose starting and ending points are the center of gravity of one partial pattern and the center of gravity of another partial pattern. The intermediate classification section performs matching with vectors between partial patterns created from the registered patterns in the same way, and performs intermediate classification according to the matching results. A partial pattern Q value representing the characteristics of the partial pattern extracted from each segment information of each stroke is calculated, and this partial pattern Q value and a partial pattern Q previously created from registered patterns in the same manner are calculated.
A partial pattern Q value matching unit performs matching with the value and ranks candidate characters according to this matching result and the matching result by the intermediate classification unit, and the top candidates ranked by the partial pattern Q value matching unit. For each character, the stroke code distribution extracted by the stroke coding unit is calculated for each partial pattern, and this stroke code distribution is matched with the partial pattern stroke code distribution previously created from registered patterns in the same way. and a partial pattern stroke code distribution matching section that performs the following: and performs a final ranking of candidate characters according to the matching results.