KR100301216B1 - Online text recognition device - Google Patents

Abstract

An on-line character recognition apparatus is provided that prevents a decrease in recognition rate due to unexpected components in an input character pattern, and further improves the recognition rate for subsequent characters and the like.

A feature extracting unit 2 for extracting a direction and a length and a feature point for each segment obtained by cutting a coordinate point on the input character pattern from the input unit 1 with a cut line, a segment of a character in a dictionary and a segment obtained from an input character pattern; Correspondence between the feature point mapping unit 3 for performing correspondence and calculating the segment correspondence distance, and the feature information of the section determined by the feature emphasis on the input character pattern corresponding to the feature emphasis of the characters in the dictionary. A feature section extracting section 4 for extracting as a stroke feature and a feature combination section 5 for calculating candidate characters by using the calculated distance of the corresponding stroke feature and the segment correspondence distance are provided.

Description

Online character recognition device

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an on-line character recognition apparatus for handwriting characters into a pen computer and the like, and more particularly, to an on-line character recognition apparatus for improving character recognition rates for subsequent characters.

In online character recognition, which is an element technology for inputting a character code into a pen computer using a pen and a tablet as a means of input, the written character is written in a known basic stroke method (previous form of strokes). Is defined as the basic stroke, and characters are represented by a combination of the basic strokes) and other various recognition methods. However, the performance of continuous letters is not enough compared to the sea. For this reason, the research of the on-line character recognition system corresponding to the character continued conventionally has been performed. For example, the Journal of Electronics and Communication Engineers J66-D No. 5, pages 593 to 600, entitled "Online Character Recognition Not Relying on Stroke or Mandatory Order by Selective Stroke Combination". Hereinafter, this is called conventional example 1. According to the prior art example 1, the small stroke of the number of strokes (a unit of coordinate string from pen down to pen up) in the input pattern and the dictionary corresponds to one to one stroke with a large number of strokes. On the other hand, the unmatched strokes are selectively combined with the already matched strokes, the distance between the coordinate points of the dictionary and the input pattern after combining is calculated using DP (Dynamic Programming) matching, and the candidate characters are output. It allows for continuous character recognition. Since DP matching is described, for example, from page 62 of "Pattern Recognition" (Funagubo Novoru, Kyoritsu Publishing), detailed description is not made here. In Conventional Example 1, the coordinate point is used as a feature of DP matching, but there is also a method of using a direction component (direction code) between coordinate points that is equally divided according to the stroke shown in FIG.

In Fig. 18, all of the strokes are connected to each other in the same manner as the strokes are written in the stroke direction, and the strokes are approximated by appropriate widths. The direction component between each of the dividing points a1, a2, a3, a4, ..., a21 in FIG. 18 is approximated by an eight-way code shown in FIG. 19, for example, and this direction component is used as a feature of DP matching. You can also continue the character recognition.

Further, in the method using the basic stroke, there is a method of creating and recognizing a dictionary corresponding to a subsequent letter, or a method of decomposing the stroke so that the stroke number of the input pattern and the dictionary is the same with the separation information of the stroke in advance.

For example, Japanese Patent Laid-Open No. 2-10473 is a conventional example using a basic stroke. Hereinafter, this is called conventional example 2. Here, the structure and operation in this conventional example 2 are demonstrated.

20 is a block diagram showing the basic configuration of the on-line character recognition apparatus of the conventional example 2. As shown in FIG. 20 shows a line segment code for inputting the coordinate input device 21, the base line segment identification circuit 22 for inputting the output of the coordinate input unit 21, and the output of the base line segment identification circuit 22, and sequentially outputting them. The circuit 23, the line segment code buffer 24, the line segment decomposition circuit 25 which sequentially decomposes the line segment of the stroke when the determination circuit 29 is re-recognized, the control circuit 26, and the line segment code sending circuit A comparison circuit 27 for comparing the output of 23 with the output of the dictionary storage 28, a determination circuit 29 for inputting the output of the comparison circuit 27 to determine characters, and a stored dictionary The preliminary storage section 28 for sequentially sending data to the comparison circuit 27 is shown.

In the basic line segment identification circuit 22 supplied with time series information of coordinate points output from the coordinate input device 21, the stroke is approximated by a line, and the eight-directions of the line segments of the respective cut lines (segments) are shown in FIG. Represented by code. Next, using the correspondence table between the direction code string and the basic stroke shown in FIG. 21, it is determined which basic stroke the input stroke belongs to.

Next, operation | movement of the prior art example 2 is demonstrated using the pattern of FIG. In FIG. 22, strokes are written in the order of lines 101, 102, 103, 104, and 105. FIG. When each stroke is approximated by a cut line and shown in the order of order by the eight-way code, it becomes {(1), (6), (7), (1, 7), (1)}. The basic stroke trains {(1), (3), (4), (7), and (1)} are obtained using the basic stroke table shown in FIG.

Since the character pattern of FIG. 22 is written in five strokes, the combination process is performed with the five stroke dictionaries of the comparison circuit 27, and the determination circuit 29 makes a candidate character determination. As a result, the character code in the dictionary is matched and the character code is output.

Next, the operation will be described in the continuing character pattern of FIG. In FIG. 23, lines 106, 107, 108, and 109 are written in order. As for the subsequent character pattern of FIG. 23, the basic stroke sequences {(1), (3), (4), (21)} are similarly obtained by the basic line segment identification circuit 22. As shown in FIG. Since the stroke count is 4, the comparison circuit 27 compares with the 4-stroke dictionary. In this case, the judging circuit 29 cannot output characters unless there are four stroke dictionaries of "stone" in the dictionary storage unit 28. From there, the control circuit 26 is returned, and the stroke is decomposed sequentially using the line segment decomposition circuit 25. The decomposition stroke and the decomposition rule are registered in advance in the pre-memory section 28. Here, the basic stroke 21 is divided into (7) and (1) using the rules shown in FIG. Is 5. As a result, the basic stroke column of the input pattern is corrected to {(1), (3), (4), (7), (1)}, and the comparison circuit 27 has five strokes in the pre-storage unit 28. Work with dictionaries. As a result, in accordance with the character "石" in the dictionary, the determination circuit 29 outputs the result.

According to the conventional example 1, the following letters can be recognized, but for example, as shown in FIGS. 25 (a) and 25 (b), the dictionary pattern (figure 25 (a)) and the input pattern ( In the case of combining FIG. 25 (b), as shown in FIG. 25 (c), the distance between the corresponding coordinate points is increased due to position shift or deformation, and as a result, the distance from the dictionary becomes larger and thus is incorrectly read. There was a problem that was easy to do.

In addition, when the direction code is used for the feature used for DP matching as shown in FIG. 18, the character recognition can be continued, and the shift in position becomes stronger as shown in FIG. 25 (c). These similar characters, for example, "伎" and " , "," And "劫", or "村" and "杖" have a problem that is easy to misunderstand.

In addition, when a character is disturbed and a character having a large direction difference between the corresponding part with the dictionary is recognized, the cost value obtained by DP matching between the input pattern and the dictionary becomes larger than the pattern in which the character is arranged, resulting in an error in other characters. There was a problem that was easy to read.

Again, as shown, for example, in Fig. 26 (a) and 26 (b), the characters (Fig. 26 (a)) and the "folding part" having components of the "folding part" and the "pressed part" are shown. There was a problem in that the distance of matching with the dictionary (Fig. 26 (b)) which does not have a component of the "pressed part" (i.e., does not have a corresponding component) becomes large, and it is easy to read it incorrectly. On the other hand, for example, disregarding the direction code of the folding component near the end point of the stroke, the end point (for example, the starting point or the segment of the end point where the angular difference of the continuous straight line is 90 degrees or less), or performing a weighting process. The method is considered, but it is not determined if the character is not recognized whether it is the noise of the folding component near the starting point, the end point, or a feature required for the character. For this reason, there existed a problem that the component of the starting point of stroke and vicinity of an end point cannot simply be ignored.

On the other hand, a method of recognizing using stroke features such as a basic stroke has a problem in that the distance cannot be calculated unless the number of strokes of the dictionary and the input pattern match. It was necessary to describe for each character a portion that would be easy to register or lead to a character pattern of a dictionary. That is, in the conventional example 2, if the decomposition rule of the continuous stroke of the input pattern does not exist before decomposition, there is a problem that the stroke cannot be disassembled and recognized incorrectly. On the other hand, there is a problem that requires enormous dictionary capacity to cope with all the following character strokes of all characters.

For example, even if a solution having a plurality of direction codes is decomposed for each direction code without using an decomposition dictionary, the solution is to be solved by using a method other than the actual stroke that is originally expressed as a stroke. It may contain virtual strokes that do not. This example is shown in FIG. 27 (a) and 27 (b). The input pattern of FIG. 27 (a) has extra components 30 and 31, which become virtual strokes at the correct stroke number, compared to the dictionary of FIG. 27 (b). For this reason, straight code strings decomposed into straight line components do not necessarily coincide with dictionaries which have the characteristics of actual strokes, so they may be misidentified, and there is a problem that they cannot be simply divided into direction codes.

SUMMARY OF THE INVENTION The present invention has been made to solve the above problems, and an object thereof is to prevent a decrease in recognition rate due to an unexpected component in an input character pattern, and to prevent the occurrence of subsequent characters (letters written in succession). An object of the present invention is to provide an on-line character recognition apparatus for improving the recognition rate.

1 is a block diagram showing the first embodiment of the on-line character recognition apparatus according to the present invention.

FIG. 2 is a diagram showing an example of the contents of information about the character "家" recorded in a dictionary used in Example 1. FIG.

FIG. 3 is a diagram showing an example of the contents of information about the character "琢" recorded in the dictionary used in Example 1. FIG.

4 is a flowchart showing character recognition processing in the first embodiment.

5 is a flowchart showing processing performed by the feature extraction unit during the character recognition processing in the first embodiment.

6 (a) is a diagram showing an example of the 16-way code used in the first embodiment.

6 (b) shows a table in which values used for DP matching are set.

FIG. 7 is a diagram showing a pattern after the feature extraction processing is performed on the input pattern in Example 1. FIG.

8 is a table showing a form extracted by performing feature extraction processing on the input pattern in Example 1 in a tabular form.

9 is a diagram showing coordinate points of the input pattern corresponding to the start point and the end point of the character "家" in the dictionary.

FIG. 10 is a diagram showing corresponding stroke characteristics of an input pattern with respect to the character "家" in a dictionary. FIG.

FIG. 11 is a diagram showing coordinate points of an input pattern corresponding to a start point and an end point of a character "琢" in a dictionary; FIG.

Fig. 12 is a diagram showing the corresponding stroke characteristics of the input pattern with respect to the character "琢" in the dictionary.

13 (a) to 13 (d) are diagrams showing examples of portions with large variation due to character patterns.

14 is a diagram showing an example of the contents of a segment dictionary of the character "木" in the dictionary.

FIG. 15 is a diagram showing segment characteristics of the character "木" shown in FIG. 13 (d).

FIG. 16 is a diagram showing a segment dictionary of the character "木" using a direction independent code. FIG.

17 (a) and 17 (b) are diagrams used for explaining the limitation of the number of correspondences with the direction independent code in the second embodiment.

Fig. 18 is a diagram showing the features of the recognition method using the direction code in the conventional example 1.

19 shows an example of an eight-way code.

20 is a block diagram showing the basic configuration of the on-line character recognition device of Conventional Example 1. FIG.

21 is a diagram showing a correspondence table between a direction code sequence and a stroke code.

FIG. 22 is a diagram showing an input pattern of the character "石" used for explaining the operation of the conventional example 2. FIG.

FIG. 23 is a diagram showing an input pattern of the character "石" used for explaining the operation of the conventional example 2. FIG.

24 is a diagram illustrating a decomposition rule of a subsequent character stroke.

25 (a) to 25 (c) are diagrams showing the effect of deviation of the position at the time of distance calculation in the conventional example 1. FIG.

26 (a) and 26 (b) are diagrams showing examples of dictionaries that do not include an input pattern including a "bent part" and a "pressed part" and a "bent part" and a "pressed part".

27 (a) and 27 (b) are diagrams showing an example of the case where the virtual stroke is expressed by the letters that follow.

* Explanation of symbols for main parts of the drawings

1 input unit 2 feature extraction unit

3: feature point corresponding building unit 4: feature section feature extraction unit

5: feature combination unit 6: pre-memory unit

7: output unit

In order to achieve the above object, in the on-line character recognition apparatus according to the first invention, the coordinate character string data of the character pattern is inputted, and the on-line character recognition apparatus outputs a character code corresponding to the input character pattern. And inputting means for inputting coordinate point data on a stroke when describing the input character pattern, and each straight line portion obtained by performing a cut line approximation of coordinate points arranged in time series included in the coordinate point data input by the input means. Characteristic extraction means for extracting feature information about each segment and the feature point that is the end point of each segment, and dictionary storage means for storing in advance a dictionary containing feature information and feature points about the segments constituting the character for each character And feature information of each character previously described and features extracted by said feature extraction means. On the basis of the gong information, the segment constituting means for correlating the segments constituting the characters in the dictionary with the segments obtained from the input character pattern, and calculating the segment correspondence distance, and the feature point on the stroke designated by the dictionary. Designation section feature extraction means for extracting feature information of a section determined by feature stiffness on the stroke of the corresponding input pattern as a corresponding stroke feature, and corresponding stroke feature extracted by the designation section feature extraction means in the dictionary. And combination means for calculating the distance of the corresponding stroke feature, and output means for outputting a candidate character code obtained by the feature combining means, wherein the feature combining means includes the calculated distance of the corresponding stroke feature and the feature combination means. The three calculated by the feature point mapping means Characters corresponding to the input character pattern in the dictionary are specified based on the segment correspondence distance.

In the first aspect of the invention, in the on-line character recognition device according to the second aspect of the invention, the feature point correspondence means includes one coordinate point of the input character pattern for each of a start point and an end point of each stroke of the character in the dictionary. Correspondingly, the designated section feature extraction means sets the feature points of the input character pattern corresponding to the viewpoint and the feature points of the input character pattern corresponding to the endpoint.

In the on-line character recognition device according to the third invention, in the first invention, the information on each segment includes a direction and a length of each segment, and the feature point correspondence means is an angle calculated by the feature extraction means. Based on the direction and length of the segment, the cost of each associated segment is calculated, and the segment mapping distance is calculated based on the cost.

Next, the most suitable embodiment of this invention is described based on drawing.

Example 1

1 is a block diagram showing the first embodiment of the on-line character recognition apparatus according to the present invention. In the present embodiment, the online character recognition apparatus includes an input unit 1, a feature extracting unit 2, a feature point correspondence unit 3, a designation section feature extracting unit 4, a feature combining unit 5, and a pre-memory unit ( 6) and an output unit 7. The input unit 1 is prepared as an input means, and inputs coordinate point data on a stroke when the user describes character data (input character pattern) input by a pen with a tablet or the like. The feature extracting section 2 is prepared as a feature extracting means, and each of the linear portions obtained by approximating the coordinate points arranged in time series included in the coordinate sequence data input to the input section 1 to the cut lines is used as a segment. Feature information about a segment and a feature point which is an end point of each segment are extracted. The feature point correspondence unit 3 is prepared as a feature point correspondence means and inputs segments and inputs constituting each character in the dictionary based on feature information of each character previously described and feature information extracted from the feature extractor 2. The correspondence with the segment obtained from the character pattern is performed, and the segment association distance is calculated. The designated section feature extracting section 4 is prepared as a feature section extracting section, and extracts, as a corresponding stroke feature, feature information of the section determined by the feature dots on the stroke of the input character pattern corresponding to the feature points on the stroke designated by the dictionary. . The feature combining section 5 is prepared as feature combining means and calculates the distance of the corresponding stroke feature extracted by the designated section feature extracting section 4. The dictionary storage unit 6 is prepared as a dictionary storage means, and stores the dictionary mentioned above in advance. In the present embodiment, the dictionary stores feature information and feature points for segments constituting characters for each character. The output unit 7 is prepared as an output means and outputs the candidate character code obtained by the feature combination unit 5.

FIG. 2 is a diagram showing information about the character "家" in the dictionary in a tabular form, and FIG. 3 is a diagram showing information about the character "琢" in the dictionary in a tabular format. The contents and features included in the dictionary stored in the dictionary storage section 6 are the character code and the segment code as the characteristic information of the segment, the circumscribed rectangle (rectangular) width of the stroke, and the circumscribed rectangle height of the stroke. The direction code and segment length of the segment are maintained for the virtual stroke in addition to the stroke. The stroke is a unit of coordinate string from pen down to pen up, but in this case, the stroke connecting the actual stroke or the end point (pen up position) of the next stroke (pen down position) is called a virtual stroke. do. The actual stroke can hold a plurality of segments for one stroke, but the virtual stroke is one segment for one stroke. The direction of the end point from the start of the stroke having a plurality of segments is indicated in parentheses in FIGS. 2 and 3. Although not shown in Figs. 2 and 3, a stroke identification code for identifying whether each segment is a real stroke or a virtual stroke is maintained.

FIG. 4 is a flowchart of character recognition processing in the present embodiment, and FIG. 5 is a flowchart showing the processing of the feature extraction unit 2. FIG. FIG. 6 (a) shows an example of a 16-way code, and FIG. 6 (b) shows a table in which values used for DP matching are set. In this embodiment, the correspondence between the segments constituting each character in the dictionary and the segments obtained from the input character pattern is performed by DP matching.

Next, the flow of the recognition process in the present embodiment will be described based on the flowchart of FIG.

First, the input unit 1 obtains a coordinate string arranged in chronological order of handwritten character data written by the user with a pen on the tablet (step 100). Next, the feature extraction unit 2 performs preprocessing and feature extraction (step 101). Details of this processing will be described using the flowchart shown in FIG.

The feature extraction unit 2 compares the distance between the successive coordinate points with respect to the input coordinate string with the reference width and performs extraction processing of the point where the distance does not exceed the reference width (step 201). In the present embodiment, each linear portion obtained by performing a line approximation to the next coordinate, that is, the coordinate points arranged in time series, from the coordinate point after the extraction will be referred to as a segment. Next, the direction code of the segment is extracted (step 202). The segment direction code is extracted for the virtual stroke in addition to the actual stroke. Here, the direction code string is extracted using the 16-way code shown in FIG. 6 (a). Thus, the segment joining process is performed (step 203). Here, when the directions of the adjacent segments are approximate, specifically, when the direction difference between adjacent segments is ± 1, the segments are combined and the segment direction codes combined are recalculated. For example, if a segment of the direction code 9 appears after the segment of the direction code 8, these segments are combined to form a single segment of the direction code 8. However, the joining process is not executed for the virtual stroke. Thus, the segment length after joining is calculated (step 204). The segment length is indicated by the number of times the reference width used for the extraction process. In this embodiment, the direction code and the length indicating the direction of the segment are extracted as the feature information. 7 shows the input pattern after the feature extraction processing described above. FIG. 8 shows the extracted features in tabular form.

Next, returning to Fig. 4, the feature point mapping unit 3 draws out one character data from the dictionary (step 102). In this example, the dictionary of "house" shown in FIG. 2 is taken out. Next, the feature point correspondence unit 3 performs correspondence of segments by DP matching between the input pattern and the character "家" in the dictionary (step 103). DP matching is performed as follows.

Si = {si (1), si (2), ... si (i), ... si (I)} as the segments of the input pattern, and Sd = {sd (1), sd (2), ... (j), ... sd (J)}

[Equation 1]

Run Next, let this expression be an expression (1). Also, the function min is a function for finding the minimum value.

Here, the following equation (2) is used.

[Equation 2]

D [si (i + 1)] [sd (j + 1)] = a [si (i + 1), sd (j + 1)] * (| si (i + 1) | + | sd (j + 1) ｜

This expression is referred to as Equation 2 below. In Equation 1, d [i + 1] [j + 1] represents an accumulation of corresponding costs from si (i + 1) to sd (j + 1) at the time point. In Equation 2, D [si (i + 1)] [sd (j + 1)] denotes a corresponding cost of the segment si (i + 1) and the segment si (i + 1) and the segment sd (j + 1). Indicates. a [si (i + 1), sd (j + 1)] is a value determined by the direction difference between the segment si (i + 1) and the segment sd (j + 1), and is shown in FIG. Use the values in the table shown. Si (i + 1) and sd (j + 1) are the segment lengths of the segments si (i + 1) and sd (j + 1). Although not shown here, we also maintain a mapping table that gives the minimum values.

Calculate Equation 1 gradually and finally

[Equation 3]

dist dp = d [I] [J] / (I + J)

Calculate This expression is referred to as Equation 3 below. This expression (3) is taken as the cost (segment correspondence distance) of DP matching with a dictionary. Dist in Equation 3 means a function for obtaining a cost of DP matching. In addition, this DP matching uses the method described in the above-mentioned "pattern recognition" (Funagubo noboru by Kyoritsu Publishing).

Here, since the number of strokes decreases as compared to the case where the following character is written with the correct stroke number, a plurality of components of the dictionary correspond to the input pattern by matching the character dictionary of the exact stroke number with the stroke and the segment of the character input pattern. There is. However, the actual actual stroke or actual segment of the dictionary is not usually the virtual stroke of the input pattern. Therefore, it is necessary to prevent the segment corresponding to the stroke which is expressed as a part which comprises the character in a dictionary from the segment corresponding to the unexpressed part of the stroke which comprises an input character pattern. Therein, when the DP matching of the segments is performed, the penality distance large to the value of D [si (i)] [sd (j)] so that these segments do not correspond to each other when calculating the virtual stroke of the input pattern and the actual actual stroke components. , Thus preventing the correspondence that is not really possible. As a result, even a continuation character can be prevented more erroneously from the character. Of course, other methods may be used to prevent the segments from being associated.

Correlation between the features of the dictionary of "house" shown in FIG. 2 and the features of the input pattern shown in FIG. 8 is shown in Equations 1 to 3 and 6 (a) and 6 (b). And use it to get dist dp = 682.

Next, in FIG. 4, the feature point correspondence unit 3 obtains the coordinate points of the input pattern corresponding to the start point and the end point of the prestroke by using the parcel table (not shown) obtained in step 103. Next, as shown in FIG.

Next, in Fig. 4, the feature point correspondence unit 3 obtains the coordinate points of the input pattern corresponding to the start point and the end point of the prestroke by using the parcel table (not shown) obtained in step 103 (step 104). 9 shows the coordinate points of the input pattern corresponding to the start point and the end point of the stroke of the character "家" in the dictionary. Next, the designated section feature extraction unit 4 performs feature extraction between corresponding points of the input pattern (step 105). Here, the characteristic point (start) of the input character pattern corresponding to the point in time at which the point corresponding to the start point and the end point of each stroke of the "house" shown in FIG. The feature point (end point) of the input character pattern corresponding to this end point is a feature point in the input character pattern. Thus, in this input character pattern, the coordinate sequence inserted in the feature point is obtained from the start point of the character input pattern, the circumscribed rectangle width between the end points, the circumscribed rectangle height, and the direction from the start point to the end point, and the virtual stroke (corresponding to the end point). The direction and distance of the vector) from the point to the point corresponding to the starting point of the next stroke are obtained. These features will hereinafter be referred to as corresponding stroke features. The result is shown in FIG.

Next, the feature combination section 5 combines a prior stroke feature and the corresponding stroke feature of the input pattern (step 106). The combination of the corresponding stroke features is, for example, (difference of circumscribed spherical width) + (difference of circumscribed spherical height) + (difference in direction from start to end point) + (difference in direction of virtual stroke) + (length difference of virtual stroke) Calculate using If a virtual stroke corresponding to the dictionary does not exist in the input pattern, the part is not calculated. The above calculation is made between the features of FIG. 2 and FIG. 10, and the distance dist st = 93 of the corresponding stroke feature with the dictionary "house" is obtained.

Next, the feature combination section 5 determines whether a dictionary to combine exists (step 107). If there are other characters in the dictionary, the process returns to step 102 and combines with the next character. In this case, other characters exist and are combined with dictionary "琢" in FIG. The feature point mapping unit 3 performs the above processing in steps 102 and 103 and obtains a cost disp dp = 674 of DP matching of "琢". Similarly, the feature point mapping unit 3 executes step 104 to find a coordinate point corresponding to the start point and the end point of the stroke of the character “” of the input pattern. This result is shown in FIG.

Next, the designated section feature extracting section 4 executes step 105 and extracts the corresponding stroke feature as before. The result is shown in FIG. Thus, the feature combination section 5 calculates with reference to the stroke feature shown in FIG. 12 and the dictionary of FIG. 3 and obtains the distance dist st = 223 of the stroke feature.

The character referring to the above flow is continued until the character disappears in the dictionary, and when the character disappears, the output unit 7 sorts the recognition result (step 108). The sorting of the result,

[Equation 4]

dist all = α x dist dp + β x dist st (α, β are integers of weight)

Find for each dictionary. Next, the transplantation is represented by equation (4). Now let α = 1 and β = 1

dist all 「家」 = 682 + 93 = 775

dist all 「琢」 = 674 + 223 = 897

It becomes By sorting dist all in ascending order, "家" is the first candidate and "琢" is the second candidate. Finally, the output unit 7 outputs the candidate characters "家" and "琢" and ends (step 109).

As a result of the above processing, "house" is finally a recognition result. As a result of the DP matching only, "Y" is the candidate first, but the correct answer can be obtained by calculating the candidate character using the corresponding stroke in combination.

As described above, according to the present embodiment, the recognition process is performed by using the DP matching and the corresponding stroke feature together, so that it is possible to recognize the subsequent letter pattern that has been distorted in the shape of a child without maintaining the dictionary data corresponding to the subsequent letter.

Further, in the first embodiment, the input character pattern is matched with all the characters in the dictionary. However, it is also possible to perform large classification using a few features, calculate DP matching for the large classification result, and calculate corresponding stroke characteristics. In the above example, the weighting process of the DP matching and the corresponding stroke feature is made equal to α = 1 and β = 1, but this value is not limited to this. Finally, the formula for calculating the distance (Equation 4) is a sum of weighted values of the result of the DP matching and the result of the corresponding stroke feature. If the distance is larger than a certain value, use the result of DP matching only to sort the result and use the first character as the candidate character, not just as a coefficient, but to add a different calculation method or condition to get the correct solution. It may be.

Again, DP matching is used to match the segments, but the present invention is not limited to DP matching, and may be a relaxation method or another method. Incidentally, the corresponding stroke feature has been described in terms of the width, height of the corresponding part, the direction from the starting point to the end point, the width of the virtual stroke, and the direction. Alternatively, other features, for example, the basic stroke, may be extracted.

Example 2

Next, sixth (a) and sixth (b), thirteenth (a) to sixth (a) and sixth (b) methods used in the first embodiment to suppress the increase in cost due to the variation of the character pattern in the DP matching of the first embodiment. 13 (d), 14, 15 and 16 will be described. Figures 13 (a) to 13 (d) show examples of characters including a large variation, Figure 14 shows examples of the contents of the segment dictionary of the letter "木", Figure 15 shows the characteristics of the segments of Figure 13 (d), FIG. 16 is a diagram showing segment dictionaries of the character "木" using direction independent codes, respectively.

When performing DP matching using the direction code and segment length of a segment as in Example 1, when attaching a "break" to a stroke with an input character, depending on the individual, or when the end of the stroke with a character pattern and the start of the next stroke The direction code of a segment or a virtual stroke may fluctuate considerably, for example, when the distance from a distance is near. In the characters shown in FIGS. 13 (a) to 13 (c), the direction difference of the virtual stroke of the part in (circle) in the figure becomes 8 between patterns using a 16-way code, which increases the cost of DP matching. Cause. In addition, as shown in Fig. 13 (d), when there is no "breaking part" in advance for the pattern with the "breaking part", the cost is large in DP matching, and when a large number of such strokes exist among the same characters, the result is different. There is a case of misreading by character.

In this embodiment, in order to prevent this, this problem is avoided by preparing a segment which calculates using only the length information of the stroke, without calculating the direction difference, in which the direction difference greatly differs depending on individual or character pattern. I am doing it.

For example, the pattern of FIG. 13 (d) extracts the direction code string and the segment length of the segment as shown in FIG. 15, and shows the direction code string and the segment length of the segment in advance as shown in FIG. If the virtual stroke of the input pattern and the prior actual stroke are forbidden to correspond during the DP matching of the segment feature, in FIG. 14 and FIG. 15, the number of strokes is equal to four strokes, respectively. do. That is, the direction code {9, 1, 3} of the 2nd stroke of the input stroke of FIG. 15 is matched with the direction code {9} of the 2nd stroke of a dictionary, and the segment and dictionary of the "break part" of the 2nd stroke of an input pattern are referred to. The matching cost of is calculated by using Equations 2 and 6 (d), and the sum of cost = direction difference (4) * segment length = 20x (7 + 1) = 160.

On the other hand, Fig. 16 shows an example of the characteristic dictionary in the present embodiment which copes with the change in the direction code. In the present embodiment, a certain segment, i.e., the character pattern input as described above, tends to have individual differences, and it is considered that a direction such as a "folded part" is dispersed and a direction code varies considerably. Direction-independent information is added to the information. In this embodiment, the direction independent code number is used as the direction independent information. In the sixteenth direction code of FIG. 6 (a), the direction independent code number is virtually prepared as 17 and held in the second stroke of FIG. The segment calculation of the segment having the direction code of 17 is defined to perform DP matching with the direction difference set to 0, and the DP calculation of the second stroke of FIG. 16 and the stroke of the second stroke of FIG. 15 is the same as that of FIG. Calculate and calculate by using Equation 5. As in the case of FIG. 14, as a result of the second match, {9}, {9}, {17}, and {13} correspond to the dictionary {9, 17} and the input pattern {9, 13}, and The cost calculation of {17} and {13} at 5 results in a direction difference of 0x (1 + 1) = 0 and the cost calculation of {9} and {9} with 0x (7x7) = 0 gives the cost in stroke units. It must be a constant value of zero. By using the direction independent code in this way, the cost is smaller in DP matching compared to the cost 160 with the dictionary shown in FIG. 14, and as a result, the distance from the dictionary is small, and misrecognition can be prevented.

However, in the case of DP matching with the direction independent code, there is a case where the correspondence is not as expected. This will be described with reference to FIGS. 17 (a) and 17 (b). Now, it is assumed that DP matching is performed between the cutting line of Fig. 17 (a) and the dictionary of Fig. 17 (b). Fig. 17 (a) has five segments 11 to 15, and Fig. 17 (b) has two segments 16 and 17. Figs. Using Fig. 6 (a), the direction code of each segment is 9 for segment 11, 5 for segment 12, 9 for segment 13, 5 for segment 14, 9 for segment 15, The segment 16 is 9 and the segment 17 is a direction independent code 17. In addition, the segment length shall be all 1. Here, the seventeenth (a) and the seventeenth (b) are matched by using the equations (1) and (2). First, segment 11 and segment 16 are associated and its cost is direction difference 0x (1 + 1) = 0. Next, since Segment 12 and Segment 16 have a direction difference of 4 in Equation 1, the cost is 20x (1 + 1) = 40, Segment 11 and Segment 17 correspond to direction independent codes in Equation 6 (b) and Equation 2. Therefore, cost 0, segment 12, and segment 17 are also associated with the direction independent code, and thus cost 0, and segment 11 and segment 17 or segment 12 and segment 17 having the smallest cost are associated with each other. Similarly, the remaining segments 13, 14, and 15 correspond to all segments 17, and the cost is zero, respectively. As a result, the segment correspondence cost with the 17th (a) and 17 (b) degrees is 0 / (5 + 2) = 0 using equation (3).

This result implies that the seventeenth (a) and the seventeenth (b) degrees are the same, and are mismatched. If you allow such a match, it may be recognized with characters that contain direction-independent codes but are not similar in shape at all. In order to prevent this, in this embodiment, the upper limit of the number of segments obtained from the input character pattern corresponding to the segment to which the direction independent code is added is prepared.

For example, if the upper limit number corresponding to the direction independent code is 1, the costs of the 17th (a) and 17 (b) degrees correspond to the segments 11 to 15 and the segment 16, and the segment 15 and the segment 17 correspond. Build. Its cost is (segment 11 and segment 16 cost =) 0+ (cost = segment 12 and segment 16) 40+ (cost = segment 13 and segment 16) 0+ (cost of segment 14 and segment 16) 40 + (Cost = segment 15 and segment 16)) + (cost = segment 15 and segment 17) 0 = 80 and the cost as expected compared to the previous cost 0 is obtained.

Also, in the embodiment, the direction difference is set to 0 so that the cost calculated based on the segment having the direction code of 17 is set to 0, and the segments such as the "bent portion" are ignored. It can also be applied such as setting the direction difference so as not to.

In the above example, equations 1 to 3 are used for the calculation of the corresponding points of the DP matching, but the present invention is not limited thereto, and other equations may be used.

According to the present invention, a character recognition process is performed by using a segment correspondence and a corresponding stroke feature for a letter that is easy to be misrecognized (letters written in succession without writing) or letters having similar writing directions. Character verification becomes possible, and as a result, character recognition with high precision can be performed. In addition, in the same mandatory order, it is possible to recognize as one kind of dictionary even if any part of a character is continued, and it is not necessary to prepare the dictionary corresponding to each subsequent letter. Therefore, for example, the pattern of the distorted continuation letters can be recognized without maintaining the dictionary data corresponding to the continuation letters, and the various effects such as the effort to make the dictionary and the reduction of the dictionary capacity can be achieved. .

Further, when the final character recognition is performed by calculation, the weight correspondence between the segment correspondence and the corresponding stroke feature can be set, so that more accurate character recognition processing can be provided.

Further, by preparing an upper limit of the number of segments obtained from the input character pattern corresponding to the segment to which the direction independent information is added, more accurate character recognition can be achieved.

Claims (3)

An online character recognition apparatus for inputting coordinate string data of a character pattern and outputting a character code corresponding to the input character pattern, wherein the coordinate string data on a stroke when the input character pattern is described is inputted. The input means and each linear portion obtained by segmented approximation of the coordinate points arranged in time series included in the coordinate point sequence data input by the input means are segmented, and the characteristic information about each segment and each segment Feature extraction means for extracting a feature point, which is an end point of the word, dictionary storage means for storing in advance a feature containing information on the segments constituting the character and a feature point for each character, feature information of each character described in the dictionary and the Segments constituting each character in the dictionary based on feature information extracted by feature extraction means And feature point correspondence means for performing a correspondence with the segment obtained from the input character pattern, and calculating a segment correspondence distance, and by the feature acceleration on the stroke of the input character pattern corresponding to the feature tone on the stroke designated by the dictionary. Designation section feature extraction means for extracting the feature information of the section to be used as the corresponding stroke feature, and corresponding stroke feature extracted by the designation section feature extraction means are combined with the feature information in the dictionary, and thus the distance of the corresponding stroke feature. And means for outputting a candidate character code obtained by the feature combining means, wherein the feature combining means is calculated by the distance of the calculated corresponding stroke feature and the feature point correspondence means. In the dictionary based on the segment correspondence distance Online character recognition device, characterized in that for identifying the character corresponding to the input character pattern. The feature point correspondence means according to claim 1, wherein the feature point correspondence means associates one coordinate point of the input character pattern with respect to the start point and the end point of each stroke of the character in the dictionary, and the designation section feature extraction means includes: the view point; And characterizing the feature point of the input character pattern corresponding to and the feature point of the input character pattern corresponding to the endpoint. The information about each segment includes a direction and a length of each segment, and the feature point correspondence means corresponds to the segment based on the direction and length of each segment calculated by the feature extraction means. And calculating a cost of each constructed segment, and calculating a segment correspondence distance based on the cost.