JPH02217981A - Device for recognizing on-line hand-written character - Google Patents

Abstract

PURPOSE:To obtain a high recognition rate by extracting a feature point from the order of successively written strokes and using an identification condition on the feature point to decide a character. CONSTITUTION:When input pattern information is applied, a feature point extracting means 1 extracts a feature point and its positional information from the input character pattern information including the order of strokes of the character. An identification condition applying means 5 retrieves an identification condition storing means 4 from a category obtained by a feature corresponding means 3 to find out an identification condition and whether the identification condition for the category is satisfied or not is decided based upon the positional information of the feature point found out by the means 1 and a correspondence recognizing result obtained from feature point correspondence information. Since whether the identification condition is satisfied or not is decided by the relation of feature point positions independently of the number of strokes, continued input characters can also be decided accurately. Thus, on-line character recognition with a high recognition rate can be attained.

Description

[Detailed Description of the Invention] [Summary] Regarding an online character recognition device that captures and recognizes handwritten character data information online in real time, the present invention aims at online character recognition with a high recognition rate, and uses input character pattern information including stroke order. A feature point extraction means for obtaining feature points and position information of the feature points from the input character pattern information, a feature point dictionary means for storing feature points of dictionary characters, and feature points of the input characters according to the stroke order. , feature point matching means for associating the feature points of the dictionary characters stored in the feature point dictionary means to obtain correspondence information, and for obtaining candidate categories as a result of the association recognition, and storing identification conditions for each target category. The identification condition storage means searches the identification condition storage means from the categories obtained by the feature correlation means to obtain identification conditions, and also searches the identification condition storage means and the feature point correspondence between the position information of the feature points obtained by the feature point extraction means and the feature point correspondence. The identification condition applying means determines whether the identification condition for the category is satisfied based on the association recognition result from the association means.

(Industrial Application Field) The present invention relates to a character recognition device, and more particularly to an online character recognition device that captures and recognizes handwritten character data information in real time online.

[Conventional technology]

A character recognition device reads prewritten characters using, for example, an image scanner, calculates the distance between the feature points of each character and the feature points of pre-stored dictionary data, and selects the character with the closest distance as a candidate character. It has been decided. The above-mentioned method of recognizing characters read with an image scanner cannot obtain the stroke order, and its recognition rate is low.

On the other hand, there is a so-called online character recognition device that does not input character data using an image scanner, but reads and recognizes information written on a tablet or the like.

This online character recognition device has a wide range of applications and is in high demand because it requires no skill compared to input devices such as keyboards.

For example, it is increasingly being used as a means of manually inputting handwritten characters in counter operations directly into a computer. Such an online character recognition device has the feature of obtaining stroke order information compared to the case of reading with a conventional image scanner, and has a higher recognition rate than the above-mentioned recognition device.

The performance of online character recognition devices has improved in recent years. In particular, technology that can also recognize characters has been developed, and the use of input at the counter described above has moved closer to the practical stage.

However, the performance in identifying characters with similar shapes is not sufficient, and high performance is also required in identifying characters with similar shapes.

Generally, an online character recognition device is composed of a major classification unit that outputs recognition candidates, and a detailed similar character identification unit that receives the recognition candidates as input and obtains correct recognition results.

FIG. 5 is a block diagram of a conventional online character recognition device. The pattern input IO is a device that converts handwritten characters into digital information, such as a digitizer. This digitizer usually consists of a board and a stylus pen, and when an inputter writes on the board with the stylus pen, the coordinates written with the stylus pen are detected. This detection result is added to the input cover pattern storage section 11 and stored. Inlet cover turn storage section 1
The digital information stored in 1 is applied to a feature point extraction section 12, which extracts feature points representing character features from the stored input cover patterns and stores them in an input feature point storage section 13.

The feature point dictionary section 14 stores feature points of dictionary characters that are standard character patterns for each character category.

The feature point matching unit 15 connects the results of the feature point extraction unit 12 stored in the input feature point storage unit 13 and the feature points of the input pattern stored in the feature point dictionary unit 14 to the feature point dictionary unit 14.
The feature points of dictionary characters stored in are associated according to the order of stroke order. That is, from the feature points obtained by the feature point extraction unit 12, the distance between the input cover pattern and the dictionary character is determined in correspondence with the brush 1 stored in the feature point dictionary unit 14. Furthermore, the feature point matching unit 15 stores the character code of the dictionary character that provides the smallest distance from the determined dictionary character in the recognition code storage unit 16 as a recognition candidate code.

The identification condition storage unit 17 stores conditions for identifying each character category as belonging to that category. The identification condition application section 18 searches for a defining condition from the top recognition candidate codes stored in the recognition code storage section 16 and applies it to the input cover turn stored in the human cover turn storage section 11 . Then, as a result of the application, the code of the dictionary character that passes the identification condition is stored in the recognition result storage section 19.

The above-described feature point matching unit 15 performs matching from images, but depending on the matching, there may be cases where incorrect matching is made. FIG. 6 is a conventional identification condition diagram of matched input characters and dictionary characters.

The input character is "2", and the character "Z" in the feature point dictionary section 14 is associated with it by the feature point matching section 15.

In such a case, the identification condition storage unit 17 contains, for example, "rZ" [The first part of the first pair of letters does not have an upwardly convex curve.]
The identification condition application unit 18 determines that the dictionary character "Z" is not an input character candidate. Similarly, if the input character is "3" and the dictionary character "5" is used as a candidate character, the dictionary character "5" is selected based on the condition that "5" is "more than two strokes". Remove from candidate characters.

Also, if you mistake "o" in katakana for "sai", it means "sai".
It was determined that the candidate was incorrect because of the identification condition, ``The second and third cars intersect more than a certain amount.''

Moreover, when "Kan" is associated with "Baku", it is detected that the association is erroneous based on the identification condition that "the 9th and 12th digits of "Baku" do not intersect.

In the case of online handwritten character recognition, it is possible to determine whether or not the candidate character is correct based on the above-mentioned strokes, stroke order, conditions for intersecting strokes, and the like. For this reason,
The recognition rate is much higher than that obtained using an image scanner or the like.

[Problem to be solved by the invention] In the online handwritten character recognition device as described above, handwritten information can be captured in real time, and by this real time capture, it is also possible to recognize the stroke order in which the text was written. Because it can be done, the recognition rate is high.

However, in the case of online handwritten character input, the inputter is not always the same, and consecutive characters, that is, continuation characters, may be input.

FIG. 7 is an input character chart that cannot be identified under the above-mentioned identification conditions. When inputting "3", if you raise the brush once and write between the top and bottom of "3", the input character will be 2.
This results in a picture that also passes the identification condition of "5". Also,
If "o" is written consecutively, that is, if the first and second strokes are written consecutively, there is no third eye, and the identification condition for "sai" cannot be applied. Similarly, if the kanji ``kan'' is followed by the ``sanzui'' part and then a part of it, the 12th ryo will be missing, and the identification condition for ``kan'' cannot be applied.

Due to the above-mentioned conditions, even if a person writes a kanji one after another, it is still considered the same kanji, but the number of strokes of the recognition device changes as a result of this, and recognition becomes incorrect.

That is, in similar character identification in conventional recognition devices, identification conditions are directly applied to the stroke order of input characters. Therefore, in order to extract character features such as stroke connection features, intersecting features, and curved features as conditions, it is necessary to find reference points from the input cover pattern. However, in order to identify characters that include zokuha characters, it is not possible to specify the number of cars to specify the identification conditions. Furthermore, if a procedure for determining a reference point from the beginning from input data is added to the identification condition, there is a problem that the storage capacity of a memory for storing the condition increases and the time required for the condition to be valid is long.

Identification for "2" in Figure 6 requires processing to find the starting part of the first car, and furthermore, "3" and "
In cases such as ``Kan'', it is not possible to deal with characters that do not have a standard number of strokes. It is also conceivable to store identification conditions one by one for how to write characters that do not have the standard number of strokes.

There are many ways to write characters that do not have the standard number of strokes, and cannot be implemented in reality.

The present invention aims at online character recognition with a high recognition rate.

[Means to solve the problem]

FIG. 1 is a functional block diagram of the present invention.

The feature point extracting means 1 receives input character pattern information including stroke order, and obtains feature points and phase information of the feature points from the input character pattern information. This input character pattern information is character information having a stroke order and a character pattern.

The feature point dictionary means 2 stores feature points of dictionary characters.

The feature point associating means 3 associates the feature points of the input character pattern information with the feature points of the dictionary characters stored in the feature point dictionary means 2 according to the stroke order to obtain correspondence information, and also determines the candidate categories of the association recognition result. seek.

The identification condition storage means 4 stores identification conditions for each target category.

The identification condition application means 5 searches the identification condition storage means 4 from the category obtained by the feature association means 3 to obtain an identification condition, and also uses the position information of the feature point obtained by the feature point extraction means 1 and the Based on the matching recognition result from the feature point matching means 3, it is determined whether the identification conditions for the category are satisfied.

[For production]

When the input cover pattern information is added, the feature point extraction means 1 obtains the feature points and the position information of the feature points from the input character pattern information including the stroke order. Then, feature point matching means 3
are the feature points obtained by the feature point extraction means 1 and the feature point dictionary means 2
The system associates the feature points of the dictionary characters stored in , obtains correspondence information, and obtains candidate categories as a result of the correspondence recognition. Candidate characters are found by this feature point matching means 3. The identification condition applying means 5 searches the identification condition storage means 4 from the categories obtained by the feature matching means 3 to obtain identification conditions, and also calculates the position information of the feature points obtained by the feature point extraction means and the feature point correspondence. Based on the association recognition result from the association information, it is determined whether the identification condition for the category is satisfied.

Since the identification condition is determined not by the number of strokes but by the relationship between the positions of feature points, it is possible to accurately identify the input character even if the character exists.

〔Example〕

Hereinafter, the present invention will be explained in detail using the drawings.

FIG. 2 is a block diagram of the online character recognition device of the embodiment.

The pattern input section 20 is, for example, a digitizer, and when an inputter writes input characters on the tablet with a stylus pen, the coordinates of the movement of the pen are read, and the input pattern storage section 21 stores the input information. This input information is memorized with a brush Jfl written with a stylus pen. In other words, this input cover pattern information includes the stroke order.

The feature point extraction section 22 reads out the input cover turn information stored in the input cover turn storage section 21, extracts feature points representing character features from the input cover turn information, and stores them in the input feature point storage section 23. Furthermore, the position of the feature point in the input cover pattern information is stored in the feature point position storage section 24. Then, the feature point associating section 25 reads out the information on the input feature points stored in the input feature point storage section 23, and also performs the matching of the feature points of the dictionary characters stored in the feature point dictionary section 26. Candidate characters with short distances, for example, three candidate characters are extracted from the characters with the least distance. and,
The extracted result is added to the recognition candidate code storage section 27, and the correspondence data in which the feature points are associated is stored in the feature point correspondence data storage section 28.

The candidate codes stored in the recognition candidate code storage section 27 are as follows:
This is a code assigned to one character, and the code is different for each character.

The identification condition storage section 29 includes the outputs of the input cover pattern storage section 21, the feature point position storage section 24, the feature point correspondence data storage section 28, and the recognition code storage section 27. Using the information stored in .28 and the identification conditions stored in the identification condition storage unit 30, the feature points of the input character corresponding to the feature points of the dictionary characters described are read out, and it is determined whether the identification conditions are satisfied. Determine. That is, the identification condition application unit 29 reads out the identification conditions stored in the identification condition storage unit 30, and applies the feature points of the input character corresponding to the feature points of the dictionary characters used in the identification conditions to the feature point data storage. This is determined from the correspondence data between the feature points of the human characters and the feature points of the dictionary characters stored in the section 28. Furthermore, points on the input character pattern corresponding to the feature points of the input character are determined using the data stored in the feature point position storage section 24.

Then, the identification conditions are applied to the input pattern and its feature points.

That is, it is determined whether the candidate character satisfies the determined identification condition, and the candidate character that satisfies the identification condition is determined as a recognition result. Note that when a plurality of candidate characters each satisfy the identification condition, the one with the smallest distance between each feature point is taken as the recognition result.

Compared to the conventional technique shown in FIG. 5, the feature point extraction section 22 stores the input feature points in the human feature point storage section 23 and also stores the feature point position information in the feature point position storage section 24.

Further, the feature point correspondence section 25 stores the recognition candidate code in the recognition code storage section 27 and stores the feature point correspondence data in the feature point data correspondence storage section 28 .

In the conventional method described above, the identification condition application unit 18 determines the identification condition without using the feature point positions and the feature point correspondence data, but in the embodiment of the present invention, the identification condition is determined by the feature point location and the feature point correspondence data. Since the discrimination is performed using the feature point correspondence data, it is possible to protect against erroneous discrimination of characters in the sequel. That is, in the past, identification conditions were used in correspondence to the number of strokes for discrimination, but in the embodiment of the present invention, a plurality of feature points are set in stroke order, and the feature point dictionary section 14 is first searched from the feature points to find candidates. I'm looking for characters. and,
The identification condition applying unit 29 determines whether the location specified by the identification condition satisfies the identification condition based on the information stored in the feature point position storage unit 24, the input cover pattern storage unit 21, and further the feature point correspondence data storage unit 28. I'm looking for.

Note that this result is stored in the recognition result storage section 31.

FIG. 3 is a chart of correspondence between feature points and identification conditions in an example. When the input character is "2" and the candidate character is "Z", based on the identification conditions of "Z" Characteristic point A of human-powered characters.

It can be determined that the input character is not "Z" based on the identification condition that there is no upwardly convex bend in the part of the pattern between B and there is no bend in the insulator.

Conventionally, it is necessary to calculate the starting part of the first car of "Z", but in this embodiment, the feature point A,
It is only necessary to find the convex curvature regarding the correspondence of B.

Also, if the input characters "3", "ono", and "kan" in Figure 6 are input as shown in Figure 7, the candidate characters "5"
Previously, the identification conditions could be passed or not applied based on the identification conditions of
As shown in the figure, if the human character is "3",
Feature points A and B of input characters corresponding to feature points A and B of "5"
B is on another stroke], it can be determined that the dictionary character "5" is not a recognition result. In addition, in the case of the input "OJ", [feature point A of the candidate character "Sai".

Feature points A and B of input characters corresponding to B, C, and D.

With respect to C and D, line segments A, B, C, and D do not intersect more than a certain amount], so that even if the first and second strokes of "o" were written consecutively, this character would be " It is possible to clearly determine that the person is not talented J. Also, when the Sanzui part of "Kan" and the 8th to 10th strokes are written consecutively, the 9th stroke and the 12th stroke do not intersect under the conventional identification conditions. However, under the identification conditions in the example, [the feature points of the input character corresponding to the feature points A, B, C, and D of the dictionary character "baku" are incorrectly recognized. A,
Line segments A, B, C, and D do not intersect with respect to B, C, and D]
From this, it can be seen that ``baku'' is not a recognition result. A and B of ``Maku'' represent 9 strokes, and C and D represent 12 strokes. In the present invention, this line segment is not represented by a stroke, but by a line segment of a feature point obtained by feature point extraction, so that it is possible to accurately determine whether a candidate character is correct based on the identification conditions.

FIG. 4 is a configuration diagram of the identification condition application section. The identification condition application section 29 includes an identification condition search section 291 and a manual character condition creation section 2.
92, an input condition storage section 293, and a determination section 294.

The identification condition search unit 291 sequentially reads the recognition candidate codes one by one from the recognition candidate code storage unit 27.

The identification condition in which the read Wm candidate code is used as an identification character is searched and read out in the identification condition storage section 30, and output to the input character condition creation section 292.

The input character condition creation unit 292 converts the identification conditions described for dictionary characters into the identification conditions described for human characters using the data stored in the minutiae correspondence data storage unit 28 and the minutiae position storage unit 24. , input condition storage section 29
Store in 3. At this time, the determination unit 294 applies the condition stored in the input condition storage unit 293 to the input feature point storage unit 293.
The input feature point data stored in 3 and the data stored in the input cover pattern storage section 21 are applied to the input character.

If the identification condition is passed, the recognition candidate code is stored in the recognition result storage section 31. If not, the above operation is repeated sequentially from 27 for the next candidate code.

As described above, the present invention extracts feature points from the sequentially written stroke order and uses the identification conditions at the feature points for discrimination, so even if consecutive characters occur, the target can be accurately identified without error. Able to distinguish between characters.

〔Effect of the invention〕

As described above, according to the present invention, it is possible to accurately identify characters even if the continuation is a character, thereby reducing the storage capacity required for identification conditions and further reducing the execution time. .

[Brief explanation of the drawing]

Fig. 1 is a functional block diagram of the present invention, Fig. 2 is a configuration diagram of an online character recognition device according to an embodiment, Fig. 3 is a diagram of feature point correspondence and identification conditions in one embodiment, and Fig. 4 is an application of identification conditions. Fig. 5 is a block diagram of a conventional online character recognition device; Fig. 6 is a diagram of matched input characters, dictionary characters, and conventional identification conditions; Fig. 7 is an input character that cannot be identified under the identification conditions. This is a diagram. I... Feature point extraction means, 2... Feature point dictionary means, 3... Feature point matching means, 4... Identification condition storage means, 5... Identification condition application means. φ

Claims (1)

[Scope of Claims] A feature point extracting means (1) in which input character pattern information including stroke order is added to obtain feature points and position information of the feature points from the input character pattern information, and feature point extraction means (1) for storing feature points of dictionary characters. a feature point dictionary means (2) for determining the correspondence between the feature points of the input character and the feature points of the dictionary characters stored in the feature point dictionary means (2) according to the stroke order; Feature point matching means (3) for obtaining candidate categories as a result; Identification condition storage means (4) for storing identification conditions for each target category; Search the identification condition storage means (4) to obtain the identification conditions, and also search the position information of the feature points obtained by the feature point extraction means (1) and the matching recognition results from the feature point matching means (3). An online handwritten character recognition device comprising: identification condition application means (5) for determining whether an identification condition for the category is satisfied based on the category.