JPH1049631A - Online handwritten character recognition method and apparatus - Google Patents

Abstract

(57) [Summary] [Problem] It is necessary to be able to recognize even if there is a stroke order error or a difference in the number of strokes, and to recognize characters at high speed with a small amount of calculation. SOLUTION: In a dictionary, only a minimum feature amount necessary for constituting a character to be recognized and which can be clearly distinguished from other characters is registered. Then, the feature amounts constituting a certain character registered in the dictionary are read one by one, and the read feature amount is compared with each feature amount constituting the input character to obtain a similarity. Based on the obtained similarity, the feature amount of the dictionary is associated with the feature amount of the input character (steps S2 to S6). An error value corresponding to the feature amount is obtained (steps S7 to S9), and in consideration of the error value, a total similarity between the character read from the dictionary and the input character is obtained. A recognition result is output based on the similarity (steps S10 to S12).

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an online handwritten character recognition for recognizing a character by associating a characteristic amount of an input character with a characteristic amount registered in a dictionary. The present invention relates to an on-line handwritten character recognition method and apparatus that can be recognized even if omitted.

[0002]

2. Description of the Related Art In general, online handwritten character recognition is performed by matching input handwriting information with handwriting information registered in a dictionary, and using a dictionary pattern having a high degree of similarity as a recognition result. In such online handwritten character recognition, strokes (referred to as input strokes) that constitute an input character are generally used as handwriting information, using strokes corresponding to the number of strokes of the character and features of partial strokes obtained by dividing the stroke. ) And a stroke registered in a dictionary (referred to as a dictionary stroke), the recognition is usually performed.

In such a recognition method, there is a method of performing recognition using stroke order information and stroke number information indicating the number of strokes of the input stroke in the character.

[0004] In character recognition using such information on the stroke order and the number of strokes, when calculating the similarity, the input stroke and the dictionary stroke are associated with each other to perform matching and narrowing down to the same image. There is an advantage that the recognition process can be performed at a low speed, but it is not possible to recognize that the stroke order and the number of strokes of the input characters are incorrect.
According to the rules, it is necessary to rearrange the input strokes or to collate the input strokes with the dictionary strokes one by one to make correspondence. Furthermore, in order to cope with an incorrect number of strokes, simply rearranging the input strokes in accordance with the rules will cause the correspondence between the dictionary strokes to shift due to excessive or insufficient input strokes, making accurate correspondence impossible. It is a general practice to collate and associate with each other.

[0005] Here, the difference in the number of strokes may be the case where the input stroke is continued and the input stroke is omitted, or the case where a noise stroke is input. If there is such a difference in the number of strokes, excess or deficiency occurs when the input stroke and the dictionary stroke are collated one by one. As a conventional method for coping with the excess or deficiency, (a) a method of leaving the larger of an input stroke and a dictionary stroke (Japanese Patent Publication No. 63-62784: hereinafter, referred to as a first conventional technique).

(B) A method of combining an extra stroke with a stroke in the immediately preceding stroke order (Japanese Patent Publication No. 2-1950: hereinafter referred to as a second prior art).

(C) A method of combining a surplus stroke with a stroke in either the front or rear stroke order (Japanese Patent Publication No.
-66703: hereinafter referred to as a third conventional technique).

(D) A method in which surplus strokes are collated with strokes described in a dictionary, and the strokes are combined in accordance with the ease with which the strokes can be combined (Japanese Patent Publication No. 4-4633: hereinafter referred to as a fourth prior art).

[0009] There is.

[0010]

According to the first prior art, when there is an inclusion relationship between an input pattern and a dictionary pattern, one of them is embedded in the other and causes erroneous recognition. There is a problem in that the overlapping of strokes of only the number of strokes cannot cope with a unique handwritten deformation that occurs when writing with the wrong number of strokes. These problems have been pointed out in the second prior art.

For example, as shown in FIG.
Considering the dictionary pattern of the character "day" and the input pattern of the character "eye", the dictionary pattern of the character "day" is included in the input pattern of the character "eye". When trying to associate strokes of characters having such an inclusion relationship with each other using the first conventional technique,
As shown in FIG. 6B, a remainder occurs in the input stroke (in this example, the third stroke is left), but in this case, depending on the state of association between other strokes. , The input character may be erroneously recognized as “day”.

The second to fourth prior arts are techniques for combining the remaining strokes by respective methods. However, all of them involve a large amount of calculation and have a problem in processing speed.

Accordingly, the present invention is capable of recognizing even if the stroke order and the number of strokes are wrong, enabling recognition between characters having an inclusive relationship, and being able to cope with both square and continuous characters. An object of the present invention is to realize an on-line handwritten character recognition method and apparatus capable of reducing the amount of calculation and increasing the recognition speed.

[0014]

According to an on-line handwritten character recognition method of the present invention, when a certain character is inputted, a characteristic amount constituting the character is extracted, and the extracted characteristic amount is registered in a dictionary. In the online handwritten character recognition method for recognizing an input character by collating and associating a feature amount with each other, each feature amount constituting one character of a recognition target character registered in the dictionary Is a feature amount that is minimum necessary to compose the character and can be clearly distinguished from other characters, and when associating with a feature amount of an input character, the character amount constituting a certain character registered in the dictionary is The feature values are read one by one, and the read feature values are compared with the respective feature values constituting the input character to obtain a similarity. Based on the obtained similarity, a dictionary feature value and an input value are input. Characters And associating the feature quantity, the feature quantity dictionary surplus without being correspondence (Dictionary remaining features) or feature amount of correlation is not in excess input character (input remaining features)
, An error value corresponding to the feature value is obtained from the data, and the total similarity between the character read from the dictionary and the input character is obtained in consideration of the error value.

The weight for each error value is changed between the error value set for the dictionary remaining feature amount and the error value set for the input remaining feature amount. Weighting is performed so that the set error value is larger than the error value set for the remaining input feature amount.

The error value set for the dictionary remaining feature amount is weighted so as to increase the error value for a feature amount having a high appearance frequency of the feature amount constituting each character when the dictionary is created. To do.

The error value set with respect to the remaining input feature may be obtained by dividing the feature of the input character into several partial features and collating them. The weight is changed between the case of the isolated feature and the case of the partial feature in which a part of a certain feature remains, and weighting is performed so that the error value of the isolated feature is larger than the error value of the partial feature. To do.

The feature amount is a stroke.

Further, the on-line handwritten character recognition device of the present invention extracts a characteristic amount constituting a character when a certain character is input, and extracts the extracted characteristic amount and the characteristic amount registered in the dictionary. In an online handwritten character recognition device for recognizing input characters, a feature amount extracting means for extracting each feature amount constituting the character from the input character, A dictionary in which, for each character, a feature amount that is minimum necessary to compose the character and that can be clearly distinguished from other characters is registered, and a feature amount that is registered in this dictionary. A similarity calculating unit for calculating a similarity between the characteristic amount and the characteristic amount of the input character extracted by the characteristic amount extracting unit; and a similarity calculating unit that calculates the similarity between the characteristic amounts based on the similarity calculated by the similarity calculating unit. The matching is performed, and the feature amount of the dictionary that has not been matched by the matching process (dictionary remaining feature amount) or the feature amount of the input character that has not been matched (the remaining input feature amount) is An association processing unit that obtains an error value corresponding to the feature amount, and calculates a total similarity between a character read from the dictionary and an input character in consideration of the error value. And

The weight for each error value is changed between the error value set for the dictionary remaining feature amount and the error value set for the input remaining feature amount. Weighting is performed so that the set error value is larger than the error value set for the remaining input feature amount.

The error value set for the dictionary remaining feature amount is weighted so as to increase the error value for a feature amount having a high frequency of appearance of the feature amount constituting each character when the dictionary is created. To do.

The error value set for the remaining input feature may be obtained by dividing the feature of the input character into several partial features and collating them. The weight is changed between the case of the isolated feature and the case of the partial feature in which a part of a certain feature remains, and weighting is performed so that the error value of the isolated feature is larger than the error value of the partial feature. To do.

The feature quantity is a stroke.

By using such means, the present invention gives an error value to the surplus feature amount that is not associated, and calculates the total similarity in consideration of the error value, thereby obtaining the stroke order and the number of strokes. Can be recognized even if a character is erroneously input, and even if there is an inclusive relationship between the input character pattern and the dictionary pattern, recognition can be performed at a high recognition rate without erroneous recognition. In addition, since there is no need to perform a process of combining the remaining features, the processing speed can be increased.

In the present invention, the error value is weighted according to various conditions. First, weighting is performed so that the error value set for the dictionary remaining feature amount is larger than the error value set for the input remaining feature amount. The error value set for the dictionary remaining feature value is weighted so that the error value is increased for a feature value having a high frequency of appearance of the feature value constituting each character when the dictionary is created. This means that it is difficult to associate with the input character when there is a surplus of important features in forming the character. In such a case, the error value is increased to prevent erroneous recognition. Like that. Further, the error value set for the input remaining feature amount is, in the case where the feature amount of the input character is divided and divided into several partial feature amounts for collation, regarding the input remaining feature amount,
The error value of the isolated feature is made larger than the error value of the partial feature. This is because the isolated feature is originally an input feature to be written, and when there are remaining input features to be written, a larger error value allows more accurate association.

As described above, by weighting the error value according to each condition, accurate recognition corresponding to various conditions can be performed.

Further, by making the feature amount a stroke,
The feature amount can be easily extracted, the similarity can be easily calculated, and highly accurate recognition can be performed.

[0028]

Embodiments of the present invention will be described below.

FIG. 1 is a block diagram for explaining an embodiment of the present invention. A character input unit 1 for outputting coordinate data for a character input with a pen or the like on a tablet as time-series data, and this character input unit 1 A stroke extraction unit 2 for extracting a stroke as a feature amount constituting the character based on the time-series coordinate data sent from the A dictionary 3 in which stroke data which is at least necessary to constitute the character and which can be clearly distinguished from other characters is registered; one of strokes constituting a certain character read out from the dictionary 3; A similarity calculation unit 4 that collates the dictionary strokes with the respective input strokes constituting the input characters and calculates the similarity, and calculates the similarity by the similarity calculation unit 4. Based on the obtained similarity, the dictionary strokes associated with the dictionary strokes read from the dictionary are associated with each other. From the stroke, an error value corresponding to the characteristic amount of the stroke, for example, a length is obtained from the position data of the start point / end point, an error value corresponding to the length is obtained, and the error value is taken into consideration from the dictionary. An association processing unit 5 for calculating the total similarity between the read character and the input character, and a recognition result output for outputting a recognition result of the input character based on the result obtained by the association processing unit 5 It is composed of a unit 6 and the like.

As described above, the dictionary 3 includes, for each character to be recognized, a stroke which is at least necessary for composing the character and which can be clearly distinguished from other characters. Data is registered, and an example of the stroke is shown in FIG.

FIG. 2 shows strokes L1, L2,..., L7 which are the minimum necessary to construct the character "sea" and can be clearly distinguished from other characters. Such strokes L1, L2,.
, L7 data is registered as a dictionary stroke.
The minimum necessary stroke for constituting the character refers to a stroke that is rarely omitted when a large number of writers write a certain character. Further, strokes necessary for distinguishing from other characters are registered for each character.

When collating with each input stroke constituting an input character, the collation order of several dictionary strokes constituting a certain character can be variously considered. You may read one by one according to the correct stroke order in the dictionary stroke,
The order in which the matching is easy may be determined in advance.

For example, when collation with an input stroke is performed, the collation order may be set such that the collation is performed in an order in which there is little correspondence error with each input stroke. The order in which the number of errors in the association is small includes, specifically, (1) an order in which matching is performed in order from a characteristic stroke (for example, a stroke having a long length) among strokes constituting one character. Write it down.

(2) The order is described so that the collation is performed in order from the stroke located at the outer edge of the character among the strokes constituting one character.

(3) The order is described so that the matching is performed in order from the stroke located in a specific direction among the strokes constituting one character.

(4) An order in which the above (1) to (3) are arbitrarily combined, for example, a combination of the above (1) and (2), which is located at the outer edge and has a characteristic stroke (for example, long The order is described so that the matching is performed in order from the (stroke).

And the like.

However, the order of collation in this embodiment is described in a dictionary such that the registered strokes are read out one by one according to the correct stroke order.

Now, an input stroke i as shown in FIG.
It is assumed that 1, i2,..., I7 have entered (the stroke input at that time is indicated by a thick line).

In the dictionary 3, dictionary strokes for each character are registered for all the characters to be recognized. FIG. 3 shows the dictionary stroke L1 of the character "Umi" as one of the characters to be recognized. , L2, ...
., L7 are read out in the order described in advance (in this case, the order according to the correct stroke order) (the read dictionary strokes are indicated by bold lines). The registered dictionary strokes L1, L2,..., L7 are
As described above, for each character, a stroke that is at least necessary to compose the character and that can be clearly distinguished from other characters is registered. In the case of “sea,” a stroke as shown in FIG. 2 is registered. Are the dictionary strokes L1, L2,.
., L7.

First, the dictionary strokes L1 set in the first order are read from the dictionary 3, and the similarity calculation unit 4 reads the read first dictionary strokes L1 and the input strokes i1, i2,. .., I7 are collated in order from the input stroke i1, and respective similarities are calculated. By this, the similarity of the input strokes i1, i2,..., I7 to the dictionary stroke L1 is obtained. Here, in order to simplify the drawing, the dictionary stroke L1 and the input strokes i1, i2,.・, I
Among the similarities between 7, only the maximum similarity is shown. That is, in FIG. 3, the similarity d1 obtained between the dictionary stroke L1 and the input stroke i1 is obtained.
1 is the maximum similarity, and indicates that the dictionary stroke L1 and the input stroke i1 are associated with each other.

Next, the dictionary stroke L2 set in the second order from the dictionary 3 is read out, and the similarity calculation section 4 reads out the read out second dictionary stroke L2 and the input strokes i1, i2, .., I7 are collated in order from the input stroke i1, and the similarity is calculated. Thereby, the similarity of the input strokes i1, i2,..., I7 to the dictionary stroke L2 is obtained. In this case, the similarity d22 obtained from the dictionary stroke L2 to the input stroke i2 is obtained. Is the maximum similarity, and the input stroke i2 is associated with the dictionary stroke L2.

Thus, the dictionary strokes are sequentially read from the dictionary 3, the read dictionary strokes are collated with the input strokes i1, i2,..., I7, and the similarities are calculated. Based on the obtained similarity, the association processing unit 5 associates the dictionary stroke with the input stroke. Thereby, the first dictionary stroke L1 is associated with the first input stroke i1, the second dictionary stroke L2 is associated with the second input stroke i2, and the third dictionary stroke L2 is associated with the first dictionary stroke L2. L3 is associated with the third input stroke i3, and the input stroke is associated with each of the dictionary strokes L1, L2,..., L7.

This example is an example in which the dictionary strokes and the input strokes are associated with each other without excess or deficiency. The case where either the dictionary strokes or the input strokes remain will be described below.

First, the case where there is an extra dictionary stroke will be described with reference to FIG. The example of FIG.
An example will be described in which, when the character “every” is input, the input stroke of the character “every” is compared with the dictionary stroke of “sea” among the recognizable characters registered in the dictionary 3. Is what you do. In this case, "sea"
The dictionary stroke of the character
Dictionary strokes L1 and L2 are minimum strokes required to compose "sea" and can be clearly distinguished from other characters.
L2,..., L7 are registered.

First, the dictionary stroke L1 set in the first order is read from the dictionary 3, and the similarity calculation unit 4 reads the first dictionary stroke L1 and
Input strokes i11, i12,.
, I16 are collated in order from the input stroke i11,
The similarity calculator 4 calculates each similarity. Thus, the similarities of the input strokes i11, i12,..., I16 with respect to the dictionary stroke L1 are obtained. In the same manner as described above, the maximum similarity that can be associated with the dictionary stroke L1 to simplify the drawing. Only shows. That is, in FIG. 4, the dictionary stroke L1
, The similarity d11 obtained with the input stroke i11 is the maximum similarity, and the dictionary stroke L1
And the input stroke i11 are associated with each other.

Next, the dictionary stroke L2 set in the second order from the dictionary 3 is read out, and the similarity calculator 4 reads out the read out second dictionary stroke L2 and the input strokes i11, i12,. .., I16 are sequentially collated from the input stroke i11, and their similarities are calculated. Thus, the similarity of the input strokes i11, i12,..., I16 to the dictionary stroke L2 is obtained. In this case, a large value is obtained between any of the input strokes i11, i12,. No similarity is obtained, and it is not associated with the dictionary stroke L2.

Next, the dictionary stroke L3 set in the third order from the dictionary 3 is read out, and the similarity calculation unit 4 reads out the read third dictionary stroke L3 and the input strokes i11, i12,. ,..., I16 are collated in order from the input stroke i11, and the similarity calculating section 4 calculates respective similarities. Thereby, the input strokes i11, i12,.
.., I16 are obtained respectively, and in this case,
For the dictionary stroke L3, the input stroke i12
And the similarity d32 obtained between is the maximum similarity,
The input stroke i12 is associated with the dictionary stroke L3.

In this way, the dictionary strokes are sequentially read from the dictionary 3, and the read dictionary strokes are compared with the input strokes i11, i12,..., I16, and the similarities are calculated. Thereby, the input stroke i11 is associated with the dictionary stroke L1 with the similarity d11, the input stroke i12 is associated with the dictionary stroke L3 with the similarity d32, and the dictionary stroke L4 is associated with the dictionary stroke L4. , The input stroke i13 is associated with the similarity d43, and for the dictionary stroke L5, the input stroke i14 is associated with the similarity d5.
4, the input stroke i15 is associated with the dictionary stroke L6 with the similarity d65, and the input stroke i16 is associated with the dictionary stroke L7 with the similarity d76.

As can be seen from FIG. 4, in this example, the dictionary stroke L2 remains without being associated with any input stroke. As described above, when there is a surplus stroke, a total similarity D is determined in consideration of the surplus stroke as an error value. The total similarity D is D = d11 + d32 + d43 + d54 + d65 + d76− (a × Ld2) when the dictionary stroke is left.
It is determined by (1). In this equation (1), Ld
2 is an error value obtained for the extra dictionary stroke, and a is a weight coefficient. As described above, the error value (a × Ld2 in the above-described example) obtained from the sum of the similarities between the associated dictionary strokes and the input strokes (d11 + d32 + d43 + d54 + d65 + d76 in the above-described example). Minus the total similarity between the input character and the character read from the dictionary.

In the example of FIG. 4, the total similarity D is obtained by multiplying the error value set for the remaining dictionary stroke L2 by a certain weight from the sum of the similarities between the other strokes. It will be required after deduction. At this time, if the error value for the extra dictionary stroke is large, the total similarity becomes a small value, and "every" is not erroneously recognized as "sea". The setting of the error value and the setting of the weight will be described later.

Next, a case where the input stroke is excessive will be described with reference to FIG. The example in FIG.
When the character "Umi" is input in nine strokes (the number of strokes is 9), among the nine input strokes of the character "Umi" and the recognizable characters registered in the dictionary 3,
An example will be described in which matching is performed with a dictionary stroke of “sea”. Also in this case, as described above, the dictionary strokes of the characters “sea” are the minimum necessary for constituting the “sea” and can be clearly distinguished from other characters, as the dictionary strokes L1, L2, L2. ...
It is assumed that L7 has been registered.

First, the dictionary stroke L1 set in the first order is read from the dictionary 3, and the similarity calculation unit 4 reads out the read first dictionary stroke L1.
Input strokes i21, i22,.
, I29 are collated in order from the input stroke i21,
The similarity calculator 4 calculates each similarity. Thus, the similarities of the input strokes i21, i22,..., I29 with respect to the dictionary stroke L1 are obtained. As described above, in order to simplify the figure, the maximum similarity that can be associated with the dictionary stroke L1. Only shows. That is, in FIG. 5, the dictionary stroke L1
, The similarity d11 obtained with the input stroke i21 is the maximum similarity, and the dictionary stroke L1
And the input stroke i21 are associated with each other.

Next, the dictionary stroke L2 set in the second order from the dictionary 3 is read out, and the similarity calculation unit 4 reads out the read out second dictionary stroke L2 and the input strokes i21, i22, , I29 are collated in order from the input stroke i21, and the similarity is calculated for each. Thereby, the similarity of the input strokes i21, i22,..., I29 to the dictionary stroke L2 is obtained. In this case, the similarity d23 obtained between the dictionary stroke L2 and the input stroke i23 is obtained. Is the maximum similarity, and the input stroke i23 is associated with the dictionary stroke L2.

Next, the dictionary stroke L3 set in the third order is read from the dictionary 3, and the similarity calculation unit 4 reads the read third dictionary stroke L3 and the input strokes i21, i22, , I29 are collated in order from the input stroke i21, and the similarity is calculated for each. Thereby, the similarity of the input strokes i21, i22,..., I29 to the dictionary stroke L3 is obtained. In this case, the similarity d35 obtained between the dictionary stroke L3 and the input stroke i25 is obtained. Is the maximum similarity, and the input stroke i25 is associated with the dictionary stroke L3.

In this way, the dictionary strokes are sequentially read from the dictionary 3, the read dictionary strokes are collated with the input strokes i21, i22,..., I29, and the similarities are calculated. . Thereby, the input stroke i21 is associated with the dictionary stroke L1 with the similarity d11, the input stroke i23 is associated with the dictionary stroke L2 with the similarity d23, and the dictionary stroke L3 is associated with the dictionary stroke L3. , The input stroke i25 is associated with the similarity d35, and for the dictionary stroke L4, the input stroke i26 is similar to the similarity d4.
6, for the dictionary stroke L5, the input stroke i27 is associated with the similarity d57, for the dictionary stroke L6, the input stroke i28 is associated with the similarity d68, and for the dictionary stroke L7. On the other hand, the input stroke i29 is associated with the similarity d79.

As can be seen from FIG. 5, in this example, the input strokes i22 and i24 are left. As described above, when the input stroke side is left, the total similarity D for associating is obtained as follows.

D = d11 + d23 + d35 + d46 + d57 + d68 + d79
−b × (id2 + id4) (2) In the equation (2), id2 and id4 are the remaining input strokes i22,
error values determined for i24, b
Is a weight coefficient. As described above, the error values (b × (in the above-described example, b × ( id2 + id4)) minus
It is the similarity between the input character and the character read from the dictionary.

By the way, the error value corresponding to the above-mentioned surplus stroke and its weight coefficient are set as follows.

(A) Each error value corresponding to an extra dictionary stroke (referred to as a remaining dictionary stroke) or an input stroke (referred to as a remaining input stroke) is determined based on the length of the remaining dictionary stroke and the remaining input stroke. Determine the value. Specifically, the length of the stroke is obtained from the position data of the start and end points of the remaining stroke, and an error value is determined based on the length. Therefore, the longer the stroke length, the larger the error value.

(B) The weights of the remaining dictionary strokes and the remaining input strokes with respect to the error values are changed, and the weight of the remaining dictionary strokes with respect to the error value is set to a value larger than the weight of the remaining input strokes with respect to the error value. That is,
The weighting coefficient a in the above equation (1) and the weighting coefficient b in the equation (2) are set to a> b.

(C) The weight of dictionary remaining strokes that are easily omitted is reduced.

(D) The weight of the dictionary remaining stroke is determined based on the appearance frequency of the stroke when the dictionary is created. In other words, when a dictionary is created, when a large number of writers write a certain character, the determination is made based on how much each stroke constituting the character is written. Specifically, when 100 writers write a certain character, strokes written by all 100 writers have a large weight (for example, a weight of 1.0) and a stroke of 10
Of the zero strokes, strokes for which only about 80 strokes are written are determined based on the appearance frequency of the strokes such that the weight is made small (for example, the weight is 0.8).

(E) In the case where the input stroke is divided and collation is performed (one stroke is divided into several partial strokes), the remaining input stroke is the case of the isolated stroke and the case where a part of the stroke remains. The weight is changed in the case of a stroke (partial stroke).

In (e), changing the weight between the case of an isolated stroke and the case where a part of the stroke remains, for example, when taking a small size as an example, the stroke f1 shown in FIG. , F2, and f3 are isolated strokes, respectively. A case where such an isolated stroke is left as an input remaining stroke is shown in FIG.
Is divided into several partial strokes f11, f12,..., F14, and these partial strokes f11, f12,.
In other words, when the pattern stroke is compared with the dictionary stroke, the weight is changed depending on whether the partial stroke is left as the remaining input stroke.

That is, regarding the remaining input stroke, the weight of the isolated stroke is made larger than that of the remaining input stroke which is not the isolated stroke. The remaining input strokes other than the isolated strokes indicate, for example, if there is a partial stroke obtained by dividing one stroke, the remaining partial stroke.

Specifically, in the input strokes shown in FIGS. 6A and 6B, as a result of the collation, for example, a case where a stroke (isolated stroke) f2 is left, and a case where a stroke f13 (a part which is not an isolated stroke) stroke)
When there is a surplus, since the stroke 13 is not an isolated stroke, the weight is increased for the stroke f2 which is an isolated stroke. This is because an isolated stroke is originally an input stroke to be written, and if there are extra input strokes to be written, a more accurate association can be made by increasing the error value.

As described above, while setting an error value corresponding to the stroke feature value,
By performing weighting under various conditions and calculating the similarity in consideration of the weighted error value, it is possible to easily perform processing when there is an extra dictionary stroke or input stroke.

For example, FIG. 4 shows an example in which dictionary strokes are left, and the condition is that the dictionary strokes are left, and that the surplus dictionary strokes L2 are strokes with a high appearance frequency (frequently omitted). From weight a
Is a large value, a × Ld2 in the equation (1) is obtained.
Is a large value, so that the total similarity D is a small value, so that "every" is not erroneously recognized as "sea".

FIG. 5 shows an example in which an input stroke is left. In this case, two input strokes i22 and i24 are used.
Although two input strokes are left, the weight b is small because the input strokes are left, and the lengths of the input strokes i22 and i24 are short. Therefore, b × (id2 + id4) in the equation (2) is a very large value. Does not. Further, even if b × (id2 + id4) is subtracted, the similarity (d11 + d) obtained between other strokes is obtained.
23 + d35 + d46 + d57 + d68 + d79) is a large value, so that the input character can be recognized as "sea" without much affecting the total similarity D.

As described above, by setting the length of the surplus stroke as an error value and changing the weight of the error value under the above-described conditions, the surplus of the dictionary stroke and the input stroke can be obtained. Accurate recognition becomes possible, and erroneous recognition of characters having an inclusive relation is prevented. Further, by changing the weight of the error value with respect to the remaining input stroke, it is possible to cope with both the square writing and the broken character. That is, if the weight of the error value with respect to the remaining input stroke is increased, the recognition engine conforms to the square writing, and if the weight of the remaining input stroke is reduced, the recognition engine conforms to the broken character. This is because when input strokes are divided and collated by dividing them into several partial strokes, there is a high possibility that partial strokes will be left in the case of broken characters, and there is little surplus in the case of regular writing. is there.

FIG. 7 is a flowchart for explaining the overall processing of the embodiment described above. In FIG.
When the stroke of the input character is extracted (step S1), a certain character is read from the registered recognition target characters (step S2), and the stroke of the read character is read. One dictionary stroke is read (step S3).

Next, one read dictionary stroke is compared with one of the strokes constituting the input character (step S4), and the similarity is calculated (step S5). An association process is performed based on the similarity (step S6).

Then, it is determined whether or not the processing has been completed for all the strokes constituting the input character (step S7). If the processing has not been completed, the process returns to step S4 and repeats steps S4 to S7. When the matching process of one read dictionary stroke with all input strokes constituting the input character is completed, it is determined whether or not the process has been completed for all dictionary strokes constituting the read character. It is determined (step S8), and if not completed, the process returns to step S3 to read the next (in this case, the second) dictionary stroke,
Perform steps 3 to S8.

When the collation processing is completed for all the dictionary strokes constituting the read character (the state shown in FIGS. 3, 4 and 5), the remaining dictionary strokes or input stroke errors due to the processing up to that point are completed. The value is calculated in consideration of the above-mentioned weight (step S9), and the similarity of the total in which the error value is considered as in the above equation (1) or (2) is calculated using the obtained error value ( Step S10).

Next, it is determined whether or not the processing has been completed for all the characters to be recognized (step S11). If the processing has not been completed, the process returns to step S2, and among the dictionary strokes for the next character, the dictionary stroke Are read out one by one, and the same processing as described above is performed. By repeating such processing, the similarity between the input character and all the recognition target characters is obtained, and the recognition result is output based on the obtained similarity (step S12).

In the above-described embodiment, the length of the surplus dictionary stroke or input stroke is obtained from the start point / end point position data of the surplus stroke by the associating processing unit 5, and the length is obtained as an error. Although it was made to be a value, instead of calculating by calculation each time, an error value for each stroke is set in advance,
You may make it memorize | store in the memory | storage part. The point is that an appropriate error value can be obtained for the extra stroke.

In the above-described embodiment, the total similarity is obtained by subtracting the error value from the sum of the similarities obtained between the strokes. However, the present invention is not limited to this. To the sum of the error values found in
The error value set for the surplus stroke may be added to obtain a total error value, and the similarity may be calculated from the value.

Further, when reading out a dictionary stroke of a certain character from the dictionary, in this embodiment, the sequence is read out in accordance with the correct stroke order among the registered dictionary strokes, but is limited to this. Instead, for example, a convenient order for performing the above-described collation may be determined in advance, and the data may be read out according to the order.

In addition, a stroke corresponding to the number of strokes of a character is used as a feature amount for associating with a dictionary. The feature amount may be a partial stroke obtained by dividing one stroke into several strokes. Other features may be used. The point is that any feature amount can be matched with the contents registered in the dictionary and a similarity can be obtained.

Further, the program for performing the processing of the present invention can be stored in a storage medium such as a floppy disk, and the present invention includes the storage medium.

[0082]

As described above, according to the present invention,
In the dictionary, only the minimum amount of features that are required to compose each character for each recognition target character and that can be clearly distinguished from other characters are registered. For each one of the target characters, the characteristic amount registered for the character and the characteristic amount of the input character are set to 1
The error values obtained corresponding to the surplus features (features registered in the dictionary or the features of the input characters) are weighted according to various conditions. Since the similarity is calculated in consideration of the weighted error value, it is possible to recognize even a character input with an incorrect stroke order or stroke number, and even if there is an inclusive relationship between the input character pattern and the dictionary pattern. It is possible to perform recognition at a high recognition rate without erroneous recognition. In addition, since it is not necessary to perform a process of combining the remaining strokes, the processing speed can be increased. Furthermore, if the weight is increased with respect to the error value of the remaining input stroke, the recognition process can be performed in accordance with the standard style. Recognition processing can be performed, and both square writing and broken characters can be handled.

[Brief description of the drawings]

FIG. 1 is a block diagram illustrating a configuration of an embodiment of the present invention.

FIG. 2 is an exemplary view showing an example of a stroke of a character registered in a dictionary in the embodiment.

FIG. 3 is a specific example of processing according to the embodiment (part 1);
FIG.

FIG. 4 is a specific example (part 2) of a process according to the embodiment;
FIG.

FIG. 5 is a specific example of a process according to the embodiment (part 3);
FIG.

FIG. 6 is an exemplary view for explaining an isolated stroke and a stroke in which a part of the stroke remains in the embodiment;

FIG. 7 is an exemplary flowchart illustrating the flow of a process in the embodiment.

FIG. 8 is a view for explaining processing of the related art.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 Character input part 2 Stroke extraction part 3 Dictionary 4 Similarity calculation part 5 Association processing part 6 Recognition result output part d11, d12, ... Associated similarity L1, L2, ..., L7 Dictionary stroke i1 , I2, ..., i7 Input stroke i11, i12, ..., i16 Input stroke i21, i22, ..., i29 Input stroke f1, f2, ..., f14 stroke

Claims (10)

[Claims] When a certain character is input, a characteristic amount constituting the character is extracted, and the extracted characteristic amount is collated with a characteristic amount registered in a dictionary. In the online handwritten character recognition method for recognizing input characters, each of the feature amounts constituting each of the characters to be recognized registered in the dictionary is at least necessary to constitute the character, and When making a feature amount that can be clearly distinguished from the character of the input character and associating it with the feature amount of the input character, the feature amounts constituting a certain character registered in the dictionary are read out one by one, and the read out feature is read out. The amount is compared with each feature amount constituting the input character to obtain a similarity, and based on the obtained similarity, the dictionary feature amount and the input character feature amount are associated with each other. Unexpected dictionary features An error value corresponding to the feature amount is obtained from the amount (remaining dictionary feature amount) or the feature amount (input remaining feature amount) of the input character that has not been associated, and the error value is taken into consideration. An online handwritten character recognition method, wherein a total similarity between a read character and an input character is obtained. 2. An error value set for the dictionary remaining feature amount and an error value set for the input remaining feature amount, weights for the respective error values are changed, and 2. The online handwritten character recognition method according to claim 1, wherein weighting is performed so that the set error value is larger than the error value set for the remaining input feature amount. 3. An error value set for the dictionary remaining feature amount is weighted such that an error value is increased for a feature amount having a high appearance frequency of a feature amount constituting each character at the time of creating a dictionary. 2. The online handwritten character recognition method according to claim 1, wherein 4. An error value set for the input remaining feature amount is obtained by dividing the input character feature amount into several partial feature amounts and collating them. The weight is changed between the case of an isolated feature and the case of a partial feature in which a part of a certain feature remains, and weighting is performed so that the error value of the isolated feature is larger than the error value of the partial feature. The online handwritten character recognition method according to claim 1, wherein: 5. The online handwritten character recognition method according to claim 1, wherein the feature amount is a stroke. 6. When a certain character is input, a feature amount constituting the character is extracted, and the extracted feature amount is matched with a feature amount registered in a dictionary by matching. An on-line handwritten character recognition device for recognizing an input character, a feature amount extracting means for extracting each feature amount constituting the character from the input character, and a feature amount extracting unit for forming the character for each character to be recognized. A dictionary in which, at least, a feature amount necessary for constituting the character and clearly distinguishable from other characters is registered, and the feature amount registered in this dictionary and extracted by the feature amount extracting means. A similarity calculating unit for calculating a similarity between the input character and the obtained characteristic, and a method for associating the characteristic with each other based on the similarity obtained by the similarity calculating. An error value corresponding to the feature amount is obtained from the feature amount of the dictionary remaining without being matched (dictionary remaining feature amount) or the feature amount of the input character remaining without being matched (input remaining feature amount). An association processing unit for calculating a total similarity between a character read from the dictionary and an input character in consideration of the error value. 7. An error value set for the dictionary remaining feature amount and an error value set for the input remaining feature amount, the weight for each error value is changed, and 7. The online handwritten character recognition device according to claim 6, wherein weighting is performed such that the set error value is larger than the error value set for the remaining input feature amount. 8. An error value set for the dictionary remaining feature amount is weighted such that an error value is increased for a feature amount having a high appearance frequency of a feature amount constituting each character when a dictionary is created. 7. The online handwritten character recognition device according to claim 6, wherein: 9. The error value set for the remaining input feature is determined by dividing the feature of the input character into several partial features and collating them. The weight is changed between the case of an isolated feature and the case of a partial feature in which a part of a certain feature remains, and weighting is performed so that the error value of the isolated feature is larger than the error value of the partial feature. 7. The online handwritten character recognition device according to claim 6, wherein: 10. The online handwritten character recognition device according to claim 6, wherein the feature amount is a stroke.