JPH0950490A - Handwritten character recognizing device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To improve the discrimination rate of characters which are similar to each other in shape and to improve the recognition rate of a right character. SOLUTION: Input holograph is normalized in a linear normalization circuit 1 and a nonlinear normalization circuit 2, supplied to a bit map plotting circuit 3, and the plotted pattern is supplied to a feature vector generation circuit 4 and the generated feature vector is stored in a storage circuit 5. Further, the stored feature vector is supplied to a comparison and calculation circuit 6, the feature vector is compared with the feature vector of the recognition object character from a reference pattern data base 7 and the evaluation value of a degree of similarity is calculated. These calculated evaluation values are stored ascending order of the evaluation value in descending order of value having the higher degree of similarity) in a candidate character string storage circuit 8. Further, the number of strokes of the stroke string by the input holograph is stored in a number of strokes storage circuit 9, this number of strokes is supplied to a number of strokes certification circuit 10 and the number of strokes is compared with the number of strokes of the recognition object character from a number of strokes information data base 11. By this comparison result, the order of the candidate character strings stored in the storage circuit 8 is rearranged and the character strings are outputted.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a handwritten character recognition apparatus, and more particularly to a verification of candidate strings of recognition results.

[0002]

2. Description of the Related Art For example, in a portable information terminal device, an input device in the form of a transparent touch panel is provided on the surface of a display device such as a liquid crystal display so that information can be input by so-called handwriting. Is being implemented.

As a method for recognizing a handwritten character or the like in such an apparatus, a character recognition by a so-called bitmap pattern has been widely used conventionally. That is, in the character recognition using this bitmap pattern, the recognition process is performed by superimposing it on the standard pattern, which is characterized by the shape of the character pattern itself.

Therefore, in this method, it is important to absorb the positional deviation of the patterns in order to perform the correct overlay. Therefore, generally, in character recognition using a bitmap pattern, pattern normalization processing, blurring processing, pattern mosaicing processing, and the like are performed.

As the above-mentioned normalization processing, for example, Japanese Patent Application No. 62-185826 or "Sho 62 National Conference on Information Systems, No. 68, 198" is used.
7. Non-linear normalization in handwritten Chinese character recognition "has been proposed.

[0006] That is, in this means, the input handwriting from when the pen is dropped on the above-mentioned input device (pen down) to when it is released (pen up) is approximated by a polygonal line, and this approximated coordinate point sequence is called a stroke. Then, the size is corrected for each stroke, the balance of the positions of the plurality of size-corrected strokes is corrected, and a bitmap pattern is drawn using the balance-corrected stroke sequence. The recognition process is performed by superimposing the standard pattern with the standard pattern.

[0007]

However, in the recognition processing using the bit map pattern as described above, if the patterns are not correctly superimposed, another character having a similar shape is overlapped, resulting in incorrect recognition. It may result.

That is, for example, when a character "Oi" is to be input, in the recognition processing using the bitmap pattern, for example, "slow, Oi, Liao, advance, multiply,
The letters "..., ..." Are extracted as a candidate string. Here, the character "Oi" is not always the first candidate, and therefore, in such a candidate string, there is a fear that the characters cannot be smoothly input.

This application has been made in view of such a point, and the problem to be solved is that in the recognition processing by the bitmap pattern, when the patterns are not superposed correctly, It is likely that an incorrect recognition result is generated.

[0010]

Therefore, in the present invention, in addition to the recognition processing by the bitmap pattern, the candidate character string is rearranged by using the data of the strokes of the input handwriting and the stroke number of the recognition target character. According to this, the recognition rate of characters having similar shapes can be increased and the recognition rate of correct characters can be improved.

[0011]

BEST MODE FOR CARRYING OUT THE INVENTION That is, according to the present invention, in a handwritten character recognizing device using a recognition process using a bit map pattern, a stroke number storing means for storing the number of strokes of an input handwriting and a stroke number data for a character to be recognized are stored. A stroke number database is provided to sort the recognition target characters stored in the candidate character string storage means according to the number of strokes of the input handwriting from the stroke number storage means and the stroke number of the recognition target character from the stroke number database. is there.

The present invention will be described below with reference to the drawings. FIG. 1 is a block diagram showing the configuration of an example of a handwritten character recognition apparatus according to the present invention.

That is, in FIG. 1, for example, the stroke sequence based on the input handwriting written in the above-mentioned input device (not shown) is supplied to the linear normalization circuit 1. The stroke is represented by a sequence of coordinate points obtained by approximating the input handwriting from pen-down to pen-up of the input device with a polygonal line as described above.

Then, the linear normalization circuit 1 corrects the size of the handwriting of the input stroke, and outputs the size-corrected stroke sequence. Further, the stroke sequence whose size is corrected by the linear normalization circuit 1 is supplied to the non-linear normalization circuit 2. Then, the non-linear normalization circuit 2 receives the stroke sequence having the corrected size, and outputs the stroke sequence in which the position balance of each stroke is corrected.

Further, the stroke sequence whose position balance has been corrected by the non-linear normalization circuit 2 is supplied to the handwriting bitmap drawing circuit 3 and drawn in a bitmap pattern using the above-mentioned normalized stroke sequence as an input. Pattern is formed. Note that the linear and non-linear normalization processing is shown in, for example, the above-mentioned document, and since it is not directly related to the essence of the invention of the present application, detailed description thereof is omitted.

The pattern drawn by the handwriting bitmap drawing circuit 3 is supplied to the feature vector generating circuit 4. In addition, the feature vector generated by the feature vector generation circuit 4 is stored in the feature vector storage circuit 5.

Further, the feature vector stored in the feature vector storage circuit 5 is supplied to the feature vector comparison calculation circuit 6. Then, the feature vector comparison / calculation circuit 6 compares the feature vector of the recognition target character from the reference pattern database 7 with the feature vector stored in the feature vector storage circuit 5, and calculates, for example, the score of the portion where these do not match. Then, the evaluation value of these similarities is calculated.

The above-mentioned reference pattern database 7
Stores the value of each feature vector of the recognition target character category. The feature vector of each character category is obtained by averaging the feature vectors of some sample handwritings belonging to that character category.

Therefore, in the calculation of the above-described feature vector comparison / calculation circuit 6, the higher the degree of similarity, the lower the evaluation value. Therefore, the recognition target characters for which the calculation has been performed are stored in the candidate character string storage circuit 8 in order from the one with the lowest evaluation value (the highest similarity). The candidate character string storage circuit 8
The character code such as the JIS code of the recognition target character and the information such as the calculated evaluation value are stored in the.

Further, in this apparatus, the stroke sequence based on the above-mentioned input handwriting is supplied to the input handwriting stroke number storage circuit 9, and the number thereof is stored in the input handwriting stroke number storage circuit 9. Further, the stroke number of the input handwriting stored in the input stroke number storage circuit 9 is supplied to the stroke number verification circuit 10.

Further, in the stroke number verification circuit 10, the stroke number of the character to be recognized from the stroke number information database 11 is compared with the stroke number stored in the input handwriting stroke number storage circuit 9. By this comparison, the ranks of the candidate character strings stored in the candidate character string storage circuit 8 are rearranged, and the rearranged candidate character strings are output as the final recognition result.

Next, FIG. 2 is a flowchart showing a processing procedure when the above-mentioned recognition device is realized by software.

In FIG. 2, when the processing is started,
First, in step [1], a linear normalization process of the input handwriting corresponding to the above-described linear normalization circuit 1 is performed. In step [2], the non-linear normalization processing of the input handwriting corresponding to the non-linear normalization circuit 2 is performed. Further, in step [3], the number of strokes of the input handwriting is calculated, and the number of strokes of the input handwriting is stored in the input handwriting stroke number storage circuit 9.

In step [4], drawing of a normalized stroke sequence corresponding to the handwriting bitmap drawing circuit 3 into a bitmap pattern is performed. Then, in step [5], a feature vector is generated from the drawn bitmap pattern corresponding to the above-mentioned feature vector generation circuit 4, and the feature vector of the input handwriting is stored in the feature vector storage circuit 5. .

Further, in step [6], the set representing the recognition result candidates is initialized to an empty set. That is, the above-mentioned candidate character string storage circuit 8 is initialized. In step [7], the read position of the reference pattern database 7 is initialized to the leading position.

In step [8], one record (feature vector) stored in the designated read position of the reference pattern database 7 is read. Further steps

In [9], the distance (evaluation value) between the feature vector read from the reference pattern database 7 and the feature vector of the input handwriting is calculated.

In step [10], addition to the candidate set is performed according to the calculated distance. That is, information such as the character code of the recognition target character and the calculated evaluation value is stored at the position determined by the rank of the evaluation value in the candidate string stored in the above-mentioned candidate character string storage circuit 8.

Further, in step [11], it is judged whether or not the feature vector read from the reference pattern database 7 is the final record. If it is not the final one, the reference pattern database 7 is selected in step [12].
The reading position of 1 is advanced by 1 and returned to step [8], and the above-mentioned processing is repeated.

When the feature vector read out from the reference pattern database 7 in step [11] is the final record, the stroke number test process described later is performed in step [13] and the process ends.

Therefore, the stroke number verification process in step [13] is performed as shown in the flowchart of FIG. 3, for example.

That is, in FIG. 3, when the process is started, first, in step [21], 1 is assigned to an arbitrary variable i. Here, i represents the number of the candidate to be tested. Next, in step [22], the difference between the evaluation values of the first-ranked candidate and the i-th candidate is calculated, and when the difference is equal to or larger than a threshold value T, the process of stopping the stroke number test is performed. Thereby, the stroke number test can be performed only on the candidate having a high degree of similarity with the first-ranked candidate.

Further, in step [23], the difference between the input stroke number and the stroke number of the i-th candidate is calculated, and this stroke number difference is stored. In step [24], it is judged whether or not the value of i exceeds the number of candidates.
In step 5], 1 is added to the variable i, the process returns to step [22], and the above process is repeated.

When the value of i exceeds the number of candidates in step [24], the candidates from the first place to the i-th place are sorted by the distance between the stroke number and the input handwriting in step [26]. In the stroke number verification circuit 10, the candidate character strings stored in the candidate character string storage circuit 8 are rearranged. As a rearrangement method, for example, the one having the smallest stroke number difference is first, and the one having the same stroke number difference is the smallest evaluation value.

Therefore, in the above-mentioned device, for example, "addition"
In the recognition process using the bitmap pattern, the following evaluation value is obtained for each recognition target character when the character is input.

Slow [18530], Oi [18558], Liao [18572], Susumu [18991] Increment [19532], Tsu [19564], Haruka [19594], [19844] Diff [19948], Hen [20079] , 逼 [20114], 逗 [20121] Evacuation [20134], Oops [20245], Atsushi [20305], Haruka [20393] Sending [20396], Retiring [20444], Sako [20502], Lotus [20536]

In other words, in this case, the candidate character string is [delaying, delaying, passing, passing, construction, and alienation, and avoiding, lotus].

On the other hand, if the threshold value T is set to T = 1000, for example, the first four characters are subject to the stroke number test, and the stroke numbers are slow [12 strokes] and additional strokes (9 strokes). , Ryo [15 strokes] and Shin [11 strokes] are stored in the stroke number information database 11.

On the other hand, when the stroke number of the input handwriting is, for example, 8 strokes, the stroke number difference from each character of the stroke number test is slow [4 strokes], additional [1 stroke], Liao [7 strokes]. ], And progress [3 strokes].

[0039] Therefore by this stroke count assay, the candidate character strings described above, [Tsuishin OsoRyo逓通Herikudazo違遍逼逗避遇渥遁Yashisa SakoRen], and the final this sorted string It is output as a recognition result.

That is, in this case, the leading four characters are rearranged as shown by underlining, and, for example, the candidate character string in which the character of the "additional" that is to be input is at the beginning is the final recognition result. Is output.

Therefore, in the present invention, there is a problem that an erroneous recognition result is apt to occur when the patterns are not correctly superposed only by the conventional recognition process using the bitmap pattern. By rearranging the candidate character strings using the data of the number of strokes of the input handwriting and the number of strokes of the recognition target character, it is possible to increase the recognition rate of characters with similar shapes and improve the recognition rate of correct characters. .

In the above apparatus, the threshold T in step [22] of the flowchart of FIG. 3 may be set to infinity and the stroke number test may be performed for all candidates.

Further, the rearrangement in step [26] is not limited to the one having the smallest difference in the number of strokes as in the embodiment, but E _i = α ×, where the difference in the number of strokes is d _i and the evaluation value is e _i. Two values such as d _i + β × e _i (α and β are constants) are combined, and this value E _i is used as a new evaluation value, and the smallest value E _i is rearranged from the beginning. May be.

According to this, for example, the number of strokes is likely to fluctuate due to the continuation of handwriting, etc. By combining this with the original evaluation value, it is possible to reduce the influence of the variation in the number of strokes.

Further, the stroke number information database 11 described above is used.
By giving the number of strokes stored in the space such as slow [10-12 strokes], additional [7-9 strokes], Liao [12-15 strokes], and progression [10-11 strokes], It is also possible to reduce the influence of the number of strokes. In this case, the stroke number verification process in step [13] is performed as shown in the flowchart of FIG. 4, for example.

That is, in FIG. 4, when the process is started, first, in step [31], 1 is substituted into an arbitrary variable i. Here, i represents the number of the candidate to be tested. Next, in step [32], the difference between the evaluation values of the first-ranked candidate and the i-th candidate is calculated. Thereby, the stroke number test can be performed only on the candidate having a high degree of similarity with the first-ranked candidate.

Further, in step [33], for example, the input stroke number and the minimum value (min) of the i-th candidate are calculated, and this value is stored as the stroke number difference. In step [34], it is judged whether or not the value of i exceeds the number of candidates, and if it does not exceed 1, the variable i is incremented by 1 in step [35] and returned to step [32]. The process is repeated.

When the value of i exceeds the number of candidates in step [34], the candidates from the 1st place to the i-th place are sorted by the distance between the stroke number and the input handwriting in step [36]. In the stroke number verification circuit 10, the candidate character strings stored in the candidate character string storage circuit 8 are rearranged. In this way, by making the number of strokes stored in the stroke number information database 11 wide, the influence of the number of strokes can be reduced.

Thus, according to the above-mentioned handwritten character recognition device, in the handwritten character recognition device using the recognition processing by the bitmap pattern, the stroke number storage means for storing the number of strokes of the input handwriting and the stroke number data of the recognition target character are stored. A stroke number database to be stored is provided, and the recognition target characters stored in the candidate character string storage means are rearranged according to the number of strokes of the handwriting input from the stroke number storage means and the stroke number of the recognition target character from the stroke number database. This makes it possible to increase the identification rate of characters having similar shapes and improve the recognition rate of correct characters.

In the above-described embodiment, the case where the input handwriting is subjected to the pattern normalization processing and the recognition processing by the bitmap pattern is performed as the pre-processing is explained. Other processing such as the mosaic processing of the above may be included.

[0051]

As described above, according to the present invention, there is a problem that an erroneous recognition result is apt to occur when pattern superimposition is not correctly performed only by the conventional recognition processing using a bitmap pattern. By rearranging the candidate character strings using the data of the number of strokes and the number of strokes of the recognition target character, the recognition rate of characters with similar shapes is increased,
It has become possible to improve the recognition rate of correct characters.

That is, according to the present invention, the classification result by the bitmap pattern is tested by using the information of the number of strokes that does not appear in the bitmap pattern, so that the ability to identify similar patterns can be improved. It is possible to improve the recognition rate.

[Brief description of drawings]

FIG. 1 is a configuration diagram of an example of a handwritten character recognition device to which the present invention is applied.

FIG. 2 is a flowchart for explaining the operation.

FIG. 3 is a flowchart for explaining the operation.

FIG. 4 is a flowchart for explaining the operation.

[Explanation of symbols]

 1 linear normalization circuit 2 non-linear normalization circuit 3 handwriting bitmap drawing circuit 4 feature vector generation circuit 5 feature vector storage circuit 6 feature vector comparison calculation circuit 7 reference pattern database 8 candidate character string storage circuit 9 input handwriting stroke count storage circuit 10 strokes Test circuit 11 Stroke information database

Claims (2)

[Claims] 1. A stroke sequence normalizing means for normalizing a stroke sequence representing an input handwriting as an input, and a stroke sequence normalized by the normalizing means is drawn in a bit map pattern and its feature vector is drawn. And a reference vector database for storing the feature vector of the recognition target character. The feature vector generation means for generating the feature vector, the feature vector storage means for storing the feature vector generated by the generation means, and the reference pattern database for storing the feature vector of the recognition target character Comparison calculation means for calculating the similarity between the feature vector of the input stroke sequence and the feature vector of the recognition target character from the reference pattern database, and the recognition target character in descending order of similarity calculated by this comparison calculation means. In the handwritten character recognition device provided with a candidate character string storage means for storing An input handwriting stroke number storage means for storing the number of strokes and a stroke number information database for storing information on the stroke number of the recognition target character are provided, and the stroke number of the input handwriting from the stroke number storage means and the stroke number information database are provided. A handwritten character recognition device for rearranging the recognition target characters stored in the candidate character string storage means according to the number of strokes of the recognition target characters from. 2. The handwritten character recognition device according to claim 1, wherein the number of strokes of the input handwriting and the number of strokes of the recognition target character are used only for a candidate having a high degree of similarity calculated by the comparison calculation means. A handwritten character recognition device, characterized in that the recognition target characters stored in the candidate character string storage means are rearranged.