JPH02300986A - Hand-written character recognizing method - Google Patents

Abstract

PURPOSE:To deal with the various kinds of the modifications of a hand-written character by utilizing the individuality of a writer. CONSTITUTION:A discriminated result storage part 3, a type quantifying part 4, an average quantity calculating part 5, and an error detecting part 6 are provided. Then, as for plural character patterns which are entered by the same writer and is discriminated into the same category by an existing recognizing method, a quantity vector to represent their types is calculated, and the character pattern having the vector different widely from others is rejected. Thus, an error unavoidable in a recognition system whose object is unspecified writer can be detected and rejected.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention relates to a handwritten character recognition method, and more particularly to a handwritten character recognition method that utilizes the individuality of a scribe to detect errors in recognition results.

[Conventional technology]

Mechanical character recognition is usually achieved through a series of processes: pre-processing, feature extraction processing, and identification processing (for example, "Introduction to Character Recognition", edited by Shinichi Hashimoto, Telecommunications Association, 1982),
(See pages 11-40).

The outline of the conventional character recognition method is as follows.

In preprocessing, the size of the input character pattern is normalized.

Perform noise removal, etc. In the feature extraction process, a feature amount representing the shape of the character is extracted from the input character stamp, and a feature vector is created. In the identification process, the feature vector extracted from the input character bang is compared with a set of feature values that define the glyphs of each category stored in the character identification dictionary, and the 0 character identification process determines the category of the input character bang. The dictionary is created statistically using character pattern data entered in advance by a large number of scribes. Further, the recognition target categories stored in this character identification dictionary are often composed of subcategories subdivided by character shape in order to deal with various character transformations.

[Problem to be solved by the invention]

In the above-mentioned conventional method, it is difficult to completely identify characters with various deformations depending on the scribe. One reason for this is that the writing habits of scribes often result in
For example, a font in one category of one scribe may be extremely similar to a font in another category of another scribe. This is particularly likely to occur with simple character types such as numbers; in other words, in conventional methods, there is fundamentally an overlap of feature distributions between each category, which improves recognition accuracy. This was a major hindrance.

SUMMARY OF THE INVENTION An object of the present invention is to provide a handwritten character recognition method that utilizes the individuality of the scribe to solve the above-mentioned problems and that can handle various handwritten character transformations.

[Means to solve the problem]

In order to achieve the above object, the present invention calculates quantity vectors representing the character shapes of multiple character patterns written by the same scribe and identified in the same category by existing recognition methods, and It is designed to reject character patterns that have vectors that are significantly different from the .

[For production]

In general, if the scribe is the same, the variation in character shape within the same category is small, and characters in different categories are always written with distinction in character shape. Therefore, for example, when calculating the average of the quantitative vectors expressing the character shapes of multiple character patterns made by the same scribe and identified in the same category, if the distance from the average is more than a certain value, The character pattern is unnatural for characters of the same category written by the same scribe and should be rejected.

By using the present invention, errors that could not be avoided in recognition systems targeted at unspecified scribes can be detected,
It can be rejected.

〔Example〕

An embodiment of the present invention will be described below with reference to the drawings.

FIG. 1 is a configuration diagram of an embodiment of the present invention, in which 1 is the character L & Oribe, 2 is a character identification dictionary, 3 is an identification result storage unit, and 4 is a block diagram of an embodiment of the present invention.
5 is a shape quantification unit, 5 is an average quantity calculation unit, and 6 is an error detection unit. Here, the part surrounded by the broken line is the part added in the present invention.

A character recognition unit 1 recognizes one handwritten input character pattern using a character recognition dictionary 2. The character identification dictionary 2 stores a set of feature values that define the character shapes of each recognition target category for unspecified scribes.In order to cope with various character deformations, as shown in FIG. Each of the recognition target categories is subdivided into subcategories by character shape.The recognition process in the character recognition unit 1 is the same as before, and as a result of the comparison with the character recognition dictionary 2, only one candidate is found for the input character pattern. When a subcategory is obtained, or when multiple subcategories of the same category are obtained as candidates, the candidate subcategory number and category number are output. Furthermore, if a plurality of subcategories of different categories are obtained as candidates, or if no candidate subcategories are obtained at all, a reject code indicating unrecognizability is output.

The output of the character recognition section 1 is stored in the identification result storage section 3 after being given a Chinese writer's name or a scribe number given in advance by reading field information or the like.

The glyph quantification unit 4 searches the identification result storage unit 3, extracts the candidate subcategory number of each character pattern for a set of character patterns identified by the same scribe and in the same category, and extracts the candidate subcategory number of each character pattern. , calculate the vector of quantities that represent those glyph shapes as follows. First, for a character pattern n, a sample data vector Qn Qn= (qnl, qn2. ---,
qnS, -)'(qns=loro) First, use the conversion matrix We corresponding to category C prepared in advance, and multiply Qn by Wc to create one character bang n.

Quantity vector vn expressing the glyph vn= (yn(1', vn"v -* vn(k)
, −・−door=WcQn ゛, n(10=Σs wk
Convert to +rqns.

The output of the glyph quantification unit 4, that is, the quantity vector νn of each character stamp identified by the same scribe and in the same category
The set is stored again in the identification result storage unit 3.

Here, the conversion matrix WC may be determined by any method, but if Qn of a large amount of character patterns for learning is It is appropriate to use the category weight vector obtained by applying the Japanese Society of Science and Technology, 1981). Using this theory, νn for the WC and learning character pattern is calculated by X = ( ,1(k), ,2(10,·,·, ,
n(k), ,,・)t and Y= (wkl, wk2. , − wks, −)
”Correlation coefficient ρXY ρXY:σχY/ (σχ° σY)crt=ΣnΣ
s qns1vn′′wks−(ΣnΣ5qns”v
n(k Lee(ΣnΣsqnsIIwks)/Nσx2=
ΣnXs qns happyvn(k)”−(ΣnΣs qns
@yn(k"/Nσ, +=ΣnΣs qnrwks"
−(ΣnΣs qns 0wks)”/NN=ΣnΣ
It is determined so that s qns is maximized. This quantity vector νn
is called the sample score vector.

In addition, the recognition system is about one character slam.

In the case of a configuration in which recognition processing is repeated several times by changing recognition conditions such as a binarization threshold value, the above method can be applied as is by setting qs to be the number of times an occurrence occurs during all recognition processing.

The average quantity calculation unit 5 searches the identification result storage unit 3 and calculates the average quantity vector νn of character stamps identified by the same scribe and in the same category, νν=(ν(1), (
1),...,ν(k),...)t and variance σ8 σ3=(σ(1)!, (!12....,σ(10
! ,...) tσ is calculated.

The error detection unit 6 calculates the average vector ν and the quantity vector νn of each character stamp used to calculate this ν.
Euclid distance Dn normalized by the variance of
A character slam whose Dn is equal to or greater than a certain threshold value e is treated as an error, and the identification result is changed to a reject code.

FIG. 3 is an operational flowchart of the character shape quantification section 4, average quantity calculation section 5, and error detection section 6 in FIG. In FIG. 3, p is a scribe number counter, C is a category number counter, and n is a character slam counter.

In step 11.13, p, c are cleared to O in advance,
In step 12.14, each is counted up by one to its maximum value. In step 15, for a certain combination of scribe number and category number, the glyph quantification unit 4 sequentially reads out the identification result categories and candidate subcategories of each character stamp stored in the identification result storage unit 3, and is converted into a quantity vector wage n and stored in the identification result storage unit 3. At this time, according to the determination in step 16,
When the number of registered character bangs is less than 3, the identification result is accepted as the correct answer and the process proceeds to the first category number. When the number of registered character bangs is 3 or more, in step 17, the average quantity calculation unit 5 The quantity vector νn of each character stamp stored in the separate result storage section 3 is read out, and their average ν and variance 62 are calculated. Next, in step 18, n is cleared to O, and in step 19, it is counted up by 1, and in step 20, the error detection unit 6 again calculates the quantity of each character button stored in the identification result storage unit 3. The vector νn is read out one by one, and the quantity vector v
A distance Dn normalized by n and the variance 62 of the average quantity vector wage is calculated. In step 21, it is determined whether or not there is a character slam whose distance Dn is larger than a certain threshold value, and if it exists.

In step 22, the character slam is considered to have a high possibility of being an error, and the identification result category is changed to a reject code. This is repeated for the number of registered character buttons (step 3), and the above process is executed for all p and c (step 4.25).

FIG. 4 is a diagram showing the relationship between error rate and jet rate. In Figure 4, the basic recognition method is
Uses a recognition method that combines the topological structuring method and the contour feature extraction method (Miyamoto et al.: "High-precision character reading device for alphanumeric and kana characters", NTT Research Practical Application Report Vol. 37, No. 2, p. 175, 1987). did. From FIG. 4, it can be seen that the error rate can be reduced to about 173 by allowing about three times as much glue jet as the conventional method.

FIG. 5 shows a specific example of the present invention. In FIG. 5, the numbers at the bottom left of each pattern indicate the correct answer, and the patterns within the broken line are, from the left, 7 becomes 1, 2 becomes 3, and 0. was mistakenly recognized as 6. By applying the present invention, the pattern within the broken line was detected as an error pattern. As described above, the present invention can easily detect error patterns with extremely confusing character shapes, which were difficult to detect using conventional recognition methods targeted at unspecified scribes.

In the example shown in FIG. 1, an average quantity vector is calculated for each character pattern written by the same scribe and identified in the same category, and characters whose distance from this average quantity vector is a certain value or more are determined. Although it is assumed that a pattern is to be rejected, the point is that it is only necessary to reject a character pattern having a quantity vector significantly different from other patterns, and it goes without saying that the method is not limited to the method of the embodiment.

〔Effect of the invention〕

As described above, according to the present invention, errors that could not be avoided in conventional recognition systems targeted at unspecified scribes can be easily detected and rejected.

Furthermore, in general, it is extremely difficult to control the ratio of error-correcting jets by modifying the identification dictionary, but according to this method, the ratio of error-correcting jets can be easily controlled by changing the distance threshold. be able to. This control can be performed finely for each character string to be recognized within the same document.

Furthermore, the present invention has the advantage that it can be easily added to existing recognition systems.

[Brief explanation of the drawing]

FIG. 1 is a block diagram of an embodiment of the present invention, FIG. 2 is a conceptual diagram of a character identification dictionary, and FIG. 3 is a character shape quantification section in FIG.
FIG. 4 is a diagram showing the relationship between the error rate and the jet rate, and FIG. 5 is a diagram showing a specific example of the present invention. 1...Character recognition unit, 2...Character result storage unit, 3.
... Identification result storage unit, 4... Character shape quantification unit. 5... Average quantity calculation unit, 6... Error detection unit. Figure 1

Claims (1)

[Claims] (1) In a character recognition method that identifies handwritten character patterns using a character recognition dictionary that subdivides each recognition target category into subcategories by character shape, handwritten character patterns are identified by the same scribe and in the same category. For each of a plurality of character patterns, a quantity vector expressing the character shape of the character pattern is calculated based on the appearance tendency of candidate subkadekori obtained as a result of identification processing, and a character pattern having a quantity vector significantly different from the others is rejected. A handwritten character recognition method featuring: