JPS58186882A - Input device of handwritten character - Google Patents

Abstract

PURPOSE:To perform accurate approximation with less information by approximating each stroke of an inputted character pattern arcuately and extracting features. CONSTITUTION:When a handwritten character pattern is inputted on a coordinate input device 11, a sequence of position coordinates of its strokes is inputted to a buffer successively and a preprocessing part 12 performs smoothing, etc., and sends the result to a feature extraction part 13 to perform feature extraction by arcuate approximation. Features of respective strokes of the input character pattern are found by this feature extraction and stored in a feature buffer 14 while classified by the strokes. A collation arithmetic device 15 collates them with features of respective strokes of standard characters registered in a dictionary 16 and a recognition decision part 17 outputs a category name which has the highest similarly and sufficient difference from succeeding similarity as the solution to the input character pattern.

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a highly accurate handwritten character input device that identifies and inputs handwritten characters based on information on handwritten characters.

[Technical backbone of the invention and its problems]

In Japan, where character strings containing kanji and kana are used as a means of writing everyday language, as computer use spreads and becomes more sophisticated, the development of devices that can manually write a wide variety of Japanese characters, including kanji, has become an important issue. I'm waiting. And recently,
With the development and practical use of Japanese word processors, it has become relatively easy to input and edit Japanese characters.

However, Japanese society has valued handwritten characters for many years, and it is thought that this trend will not change in the future.

From this, OC can read handwritten characters as they are.
There is a strong desire for the development of an R device or a simple device that can identify and input the above-mentioned characters from information obtained in the process of handwriting the above-mentioned characters.

Now, various studies have been conducted on devices for inputting handwritten characters online.
Input character types are expanding from English letters, numbers, and symbols to katakana and hiragana. Typical examples include Odaka, Buyou, and Masuda, “Online recognition of handwritten characters using point approximation of strokes,” Journal of the Institute of Electronics and Communication Engineers 82/2, Vol J63-D.
There is a point approximation of the stroke described in A2.

This point approximation method calculates the individual strokes of the input character from 3 to 6
This approximation is performed using five representative points and compared with a dictionary of characters registered in advance, making it possible to recognize approximately 1,000 types of characters. According to this method, the strokes that constitute characters are approximated by straight lines or straight line segments, so that it is possible to effectively recognize and input kanji and katakana, which basically consist of straight lines. However, even in this case, problems remain regarding the stroke order and number of strokes of the input characters. However, it appears most frequently in Japanese character strings,
Moreover, it was difficult to recognize hiragana characters, which are often composed of curved lines. Furthermore, even if katakana or kanji are deformed by handwriting and become curved, there is a problem in that it is not possible to adequately deal with this.

Therefore, highly accurate recognition of input characters could not be expected. .

[Purpose of the invention]

The present invention was made in consideration of the above circumstances, and its purpose is to provide highly practical handwritten characters that can easily and accurately identify and input handwritten characters that include many curves. Its purpose is to provide an input device.

[Summary of the invention]

The present invention approximates the stroke information obtained in the above-mentioned input process of handwritten characters input through a coordinate input device by circular arcs for each stroke, and identifies the information series approximated by these circular arcs to obtain the handwritten characters. It is designed to be input.

〔Effect of the invention〕

Therefore, according to the present invention, it is possible to effectively approximate the curved strokes that are often found in handwritten characters to express their characteristics well, and it is possible to perform highly accurate identification with less information and input the handwritten characters with high precision. I can do it. Moreover, since each stroke is approximated by a circular arc, it is possible to accurately find information about the straight line components of kanji that are curved by handwriting, improving identification ability and making it possible to effectively write many types of characters manually. This has great practical effects, such as making it possible to

5- [Embodiments of the invention] Embodiments of the present invention will be described below with reference to the drawings.

FIG. 1 shows the configuration of a conventional online handwritten character recognition processing device. Handwritten character input/main turn is input from a two-dimensional coordinate input device (not shown) such as a tablet using a time-series signal Pi(x , y) to the preprocessing unit 1. In this preprocessing section 1, for example, the above-mentioned time series signal Pi
Smoothing processing is performed by calculating a moving average between three points before and after (x, y). The feature extraction unit 2 extracts stroke original information of the input character/mother turn from this smoothed signal, that is, the time series signal Pi (x,
The feature that best approximates the sequence of y) is extracted.

For example, according to the method shown in the above-mentioned literature, this feature is obtained by linearly approximating the stroke of the input character/mother turn, and when the stroke is approximated by a single straight line, the starting point and ending point of the stroke are approximated by a straight line. Stroke features are indicated by position coordinates. In addition, when approximation is performed using two straight lines, the stroke characteristics are indicated by the positional coordinates of the starting point, midpoint, and ending point indicated by the two broken lines.

However, the information on the stroke feature series of the input characters and patterns extracted in this way is sequentially stored in the stroke feature series buffer 3, and when all the information constituting the input characters and patterns is obtained, it is verified Transferred to Section 4. The matching section 4 sequentially matches the standard stroke features of each character registered in advance in the stroke feature series dictionary 5 with the stroke features of the input character pattern, and calculates, for example, their similarity. Based on this matching result, the category of the standard pattern with the highest degree of similarity is recognized and inputted as the input result of the input character i4 turn.

In such basic recognition processing, the stroke feature extraction and dictionary matching described above are the most important techniques in this type of recognition processing and influence the recognition accuracy.

The present invention is characterized in that circular arc approximation is used in place of the conventional straight line approximation to extract stroke features in the character recognition input process. The details of this circular arc approximation will be described below, but first an overview thereof will be explained.

Figures 2 (a) to (C) and 3 (a) to (C) show the difference between linear approximation and circular arc approximation for the input character/mother turn. These are examples for each of the following. In the figure, (-) indicates an input character/mother turn, (b) a straight line approximation mother turn, and (e) an arc approximation pattern.

As shown in these nos and turns, kanji characters have a larger number of strokes than flat names, but each stroke is almost entirely straight. On the other hand, although flat name names have fewer strokes, most of them are often represented by curved lines.

Therefore, in this case, it is necessary to linearly approximate the curved stroke by a combination of a plurality of partial linear strokes. In comparison, the arc approximation allows each stroke to be accurately approximated with a smaller number of divisions than the linear approximation described above.

The fact that the stroke approximation accuracy in feature extraction is high means that there is a high possibility that highly accurate recognition processing can be performed. Further, as mentioned above, the stroke can be effectively approximated by reducing the number of divisions of the stroke. In other words, the fact that the amount of information required for approximation may be small means that high-speed recognition processing can be performed with simpler hardware.

Conceptually, it can be said that the effect of feature extraction by arc approximation in the apparatus of the present invention is very high.

Now, arc approximation of the stroke of an input character pattern according to the present invention is performed as follows.

Figure 4 shows an example of this, where i is indicated by a thick solid line A.
The f-turn indicates the stroke to be approximated. Both ends of this stroke A are RQ, and the straight line connecting them is B
Then, the approximate midpoint P of the stroke A is defined as the (9) midpoint on the straight line B (PQ). And the stroke A is
It is approximated by a circular arc C (indicated by a broken line in the figure) whose chord is the straight line B. Also, the length of this arc C is S1 the straight line B
The slope of is defined as θ (or θ'). There are various methods for determining the arc C of the stroke A approximated in this way in one section, but here, for example, the following four pieces of information are used.

(1) Coordinates of midpoint P (Px, Py) (ii) Inclination of straight line B (θ) < ij) Length of arc C (S) (1v) Difference between the length of arc C and the length of straight line B ( D) The linearity U of the stroke A is defined as U=(length of straight line B)/(length of arc C). Further, the slope θ of the straight line B as the information (11) is generally defined as the slope θ' of the straight line B itself with respect to the reference line as shown in FIG. − is defined as the angle θ measured in the direction in which the arc C exists with respect to the straight line B by adding an angle π force to the above-mentioned inclination θ′, that is, the normal direction of the straight line B. Thereby, the direction of the arc C is uniquely determined. As described above, in the arc approximation according to the present invention, the stroke A has the above-mentioned four pieces of information (PX.

py), θ, S, and D.

By the way, according to the linear approximation shown in the above-mentioned literature,
Stroke A is (Qx, Qy, Tx, Ty, RX.

Ry) is approximated by a broken line. In comparison, by employing circular arc approximation, it is possible to compress the amount of information by approximately 17f6.

However, when inputting a character or a turn through a coordinate input device such as a tablet, such an amount of information is generated by a position coordinate series ("i") indicating a stroke A in the input process.
, 3'i) i=1.2-t It can be obtained as follows. That is, the coordinate information at the beginning of stroke A (Q-work, Qa) and the coordinate information at the end of stroke A (Rx, R
y), the position coordinates of the midpoint P are determined as Px = (QX + RX)/2 py = (Qy+Ry)/2. Then, the slope of straight line B can be found as. However, regarding this slope θ, the above Rx, Ry
, Qx, Qy, the angle information calculated in advance is RO
Store it as a table in M etc., and convert this table to (
Ry-Qy), (RX-QX) may be searched for as an address.

Furthermore, regarding the length of the arc C, if this arc C closely approximates the stroke A, it can be said that it is approximately equal to the length of the stroke A, so the length of the arc C can be calculated from the position coordinate series (xl, yl). I'm asking for an example. Historically, the difference in length between arc C and straight line B is
It is determined by subtracting the length of straight line B from the value of S.

These operations (calculations) are executed by microprocessor firmware or dedicated digital hardware, and are characteristic sounds that can be extracted relatively easily.

Further, the linearity U of the stroke A calculated from the characteristic sound obtained in this manner indicates whether the stroke A is linear or curved. Therefore, using this information U, one step) o-
It is also possible to divide the stroke into a plurality of partial strokes, perform the above-described arc approximation for each partial stroke, and express the stroke using N consecutive circular arcs.

In this way, it is possible to further improve the accuracy of the circular arc approximation.

FIGS. 5(-) to (f) show "combined" and "full" strokes in which the number of strokes approximated by the inner arc is the same as described above.

13- Characteristics of each stroke of the input character ``gutaan'',
That is, the angle θ' and the linearity U are shown in the radiation diagram as a basic meter. In this radiation diagram, the radial direction is the axis of the linearity U, and when the stroke is a straight line (U=1), information on the stroke is plotted at the outermost position. Further, the stronger the stroke curve and the stronger the condition (U<1), the more the f lot position is determined inside the circle. In this way,
Even if only the features related to the angle θ' and the linearity U are taken into account, it can be seen that the features of the input characters/mother turns of the above-mentioned three input characters with the same number of strokes and the same number of strokes are significantly different. This means that the above-mentioned similar input character patterns can be clearly identified. According to the stroke arc approximation according to the present invention, in addition to these features, the length Sl of the stroke A and its position it (Px, Py) are also used as identification information, so that the features appearing in the radiographic diagram are The above differences can be obtained for each input character pattern. J14- So, for example, “me” and “and”, “mi” and “su”, and “to”.
It is also possible to effectively identify character patterns that have strong similarities, such as `` and ``.

FIG. 6 is a schematic configuration diagram of an embodiment of a device for recognizing and inputting input handwritten characters by performing the above-mentioned stroke arc near (JJ process); 1) is a position coordinate input device such as a tablet, 12 13 is a preprocessing unit, 13 is a feature extraction unit that extracts features of an input character pattern by circular arc approximation of strokes, 14 is a feature buffer, and 15 is a matching calculation device. Reference numeral 16 denotes a dictionary in which features of each category used in the matching calculation are registered, 17 a recognition determination section, and 18 a control section responsible for a series of these controls.

However, now, using the coordinate input device 11, character No. 4? When a turn is input by hand, the position coordinate series ("i, 3'1) i=1.201.t indicating the stroke
Store in next buffer. And this position coordinate series (xl
. Through this feature extraction process, the features (PX, Py, θ) are extracted for each stroke of the input character/gutaan.

S, D) are required. These features are then stored in the feature buffer 14 for each stroke. The above processing is performed until the feature extraction of all the strokes constituting the input character pattern and the storage of the features in the buffer 14 are completed. Upon completion of this processing, the number of strokes of the input character pattern is also detected. be remembered. Verification calculation device 1
5 is a dictionary 16 for the characteristics of each stroke of input characters and turns registered in the buffer 14 in this way.
The characteristics of each stroke of a plurality of standard characters and patterns recorded in advance are compared and verified. For example, as shown in FIG. 7, the dictionary 16 stores the category name and the number of strokes for each category of a plurality of standard patterns as a header 21, and stores each stroke of the number of strokes indicated by the header 2. Information 22 is stored respectively. This stroke information 22 is the feature information (px, py, θ, S, D) obtained by the circular arc approximation described above.
It is registered for each stroke in a format as shown in FIG. 8, for example. Note that it is desirable that the stroke information 22 be registered in accordance with the stroke order of the character pattern. At this time, in addition to the above-described information, coefficients kl, 2, 3, etc. indicating the importance of the stroke are also registered in the stroke number N and stroke division serial number nXtlj.

However, the above-mentioned input character and
The number of strokes of each turn is compared with the number of strokes of each standard character registered in the dictionary 16, and matching categories are extracted. Information on each of the strokes with the same number of strokes is compared with the corresponding stroke of the input character string by performing a similarity calculation. This matching process is performed for all strokes forming the - character, and the results are stored.

This matching process can also be performed directly between the input character/mother turn and standard characters/mother turns of other categories that have the same number of strokes, and the matching result obtained thereby and its category name are respectively memorized.Then, in the recognition determination section 17, the matching results obtained for each category 8 are compared with each other,
For example, the category name having the highest similarity to j1 and having a sufficient difference from the next similarity is output as the answer to the input character/gutern. By inputting the categorical name that is the answer, the input of the handwritten input character pattern is completed.

By the way, in the above explanation, all the strokes that make up the character/P turn are approximated by one circular arc to find their characteristics, but as mentioned earlier in 1411, strokes with strong curves (for multiple parts)
It is also possible to divide the stroke into o-ways and extract features by approximating each of these partial strokes.
It is 1 set. That is, in this case, in the above sentence 416,
'Find the division sequence number n of the stroke that was made by the child ball, and calculate 18
-n) If the condition is 1, the corresponding stroke may be divided into n parts according to the value of n, and arc approximation may be performed anew for each of them.

Then, a comparison may be performed for each feature of these divided partial strokes.

Note that the above-described similarity calculation between strokes is performed, for example, as follows.

Here, the symbol with a cross (-) indicates the stroke information of a turn in a standard character registered in the dictionary 16 in advance. Furthermore, 1 + k2 + k31'j is a coefficient given to each feature category, and is determined according to the importance of the feature in handling.

That is, in the above equation, the first term indicates a difference in stroke position, the second term indicates a difference in stroke inclination, and the third term indicates a difference in stroke shape. Therefore, for example, in distinguishing between "n" and "su", the difference in the slope of the first stroke is the determining factor, so in this case, the coefficient is given a heavier weight. By weighting such features, it becomes possible to reliably identify characters having similar strokes.

As explained above, according to the present invention, each stroke of an input character pattern is approximated by a circular arc to extract features, so each stroke can be accurately approximated with a small amount of information. In addition, it is possible to effectively extract the characteristics of hiragana, which contain many curved strokes and appear frequently, making it possible to recognize and input these types of characters, which have traditionally been difficult to recognize, with a high level of accuracy. becomes. In addition, by adaptively changing the number of stroke divisions and the degree of preferential treatment of each stroke feature according to the complexity of the stroke shape and the severity of the stroke, the discrimination between similar characters can be made more advanced. It becomes possible to do so. Furthermore, by variably adjusting the number of stroke divisions and the importance meter while repeatedly learning and training, character recognition input performance can be easily improved successively.

Furthermore, the characteristics of a stroke can be obtained instantaneously during the input process or at the end of writing, and can be processed at high speed using simple hardware. Therefore, the practical advantages of this device in terms of recognition accuracy and other aspects are extremely high. Moreover, it produces various special effects that cannot be expected from conventional devices.

Note that the present invention is not limited only to the above embodiments. For example, it is useful to normalize the size of input characters and the size of characters stored in a dictionary. In this case, the area where the four turns of the input character exist is examined, the vertical and horizontal dimensions of the character/arm turn are determined, and the position (Px) is determined within this area.

Py) may be used as the relative position information. Also, regarding the difference between the arc length S and the length in this normalization, the upper = t s and 21-D of each stroke may be divided by the total length of all strokes that make up the character. .

Further, the restrictions on the order of each stroke can be removed as follows. That is, the position of the input stroke consisting of (
pX, pA) and the stroke position registered in the dictionary (p
x, py), and select the one with the smallest difference as the matching target. Furthermore, it is also useful to define a virtual line segment connecting one stroke to the next stroke using a straight line connecting the end point of the above stroke and the start point of the next stroke, and add this as one of the identification information. It is. In this way, the relative positional relationship between two strokes that occur one after the other becomes clear, which facilitates the recognition process.

Also, the geometric relationship between strokes,
For example, it is useful to add information such as protrusion, intersection, parallelism, and branching as one of the identification information. As described above, the present invention can be variously modified and applied without departing from the gist thereof.

[Brief explanation of drawings]

The figures are for explaining one embodiment of the present invention. Figure 1 is a schematic configuration diagram of an online handwritten character recognition device, and Figure 2 (
a) to (c) and Figure 3 (a) to (c) are letters and
Figure 4 is a diagram showing the concept of circular arc approximation; Figures 5 (a) to (f) are diagrams showing the character/arm turn and the features approximated by circular arc; Figure 4 is a diagram showing the concept of circular arc approximation; FIG. 6 is a schematic block diagram of the embodiment apparatus, FIG. 7 is a memory block diagram of a dictionary, and FIG. 8 is a diagram showing a data format of stroke characteristics. 11... Coordinate input device, 12... Preprocessing unit, 13.
...Feature extraction device, 14...Feature buffer, 15...
- Verification calculation device, 16... Dictionary, 17... Recognition determination unit, 18... Control unit. Applicant's agent Patent attorney Takehiko Suzue 23-

Claims (4)

[Claims] (1) Stroke information obtained in the above input process of handwritten characters input via a coordinate input device is approximated by circular arcs for each stroke, and these arc approximated information series are identified to obtain information on the handwritten characters. A handwritten character input device that is capable of inputting . (2) The circular arc approximation of a stroke is based on information on the coordinates of the midpoint of a straight line connecting both ends of the stroke, information on the slope of the straight line, information on the total length of the above-mentioned stroke, and the relationship between the total length of this stroke and the length of the straight line. The handwritten character input device according to claim 1, wherein approximation processing is performed by obtaining difference information. (3) A patent claim in which the means for identifying the o-ri by approximating a circular arc divides the stroke into a plurality of partial strokes, and then performs the identification process by approximating each partial stroke to a circular arc one by one. The handwritten character input device according to item 1. (4) The handwritten character input device according to claim 3, wherein the number of divisions of a stroke into a plurality of partial strokes is adaptively determined according to the degree of complexity of the stroke.