JPS58207183A - Character discriminating system - Google Patents

Abstract

PURPOSE:To discriminate the similar characters from the crossing frequencies between the character strokes and a fixed angle, by giving vertical and horizontal scans to each point of a background part after reducing the number of candidate character types. CONSTITUTION:The number of candidate character types is reduced by an existing method. For these character types, both vertical and horizontal scans are carried out from the points showing a background. When the vertical and horizontal scans are given from a point P, the crossing frequencies between the strokes and scanning lines with a fixed angle are referred to as (a) and (b) as well as (c) and (d). Therefore a=1, b=4, c=0 and d=1 are satisfied for the point P in the figure (i). While a=4, b=1, c=3 and d=1 are satisfied for a point Q. As a result, the feature information obtained to perform discrimination among similar characters.

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a character identification method, and more particularly to a method for identifying similar characters in printed or commonly handwritten Chinese characters.

Conventionally, problems with printing or commonly used handwritten kanji recognition include:
It is known for (1) complexity of character shapes, (2) large number of character types, and (3) large number of similar character sets. Among these problems, (
1) and (2) have been largely solved using the kanji recognition methods known so far, but (3) is the final problem.

In general, character recognition methods, and handwritten character recognition methods in particular, have two methods: one method focuses on the character part and extracts stroke information to analyze the structure of the input character, and the other focuses on the background part instead of the character part and extracts stroke information. Two typical methods have been used: a method of characterizing and recognizing each point or a predetermined point.

The former aims to improve recognition performance by capturing even the detailed information of input characters, while the latter aims to extract global features in a relatively stable manner.

The former method has already been commercialized when the number of character types to be read is relatively small, such as alphabets, numbers, katakana, and special symbols, but if kanji are also included in the reading target, the above-mentioned problem will occur. The reality is that it is extremely difficult to solve problems (1) and (2).

Therefore, in kanji recognition, the latter method is used.
No. 53 Specification ``Character J Weaving Method'' etc.

Now, in order to solve problem (3), we perform the invitation process using a known method such as the one-character recognition method described in Application No. 169153/1983 above, and multiple similar character types are output. If the remaining character candidates cannot be determined as a negative character type, it may be possible to identify similar character types by extracting the stroke information mentioned above and analyzing the detailed structure. The scale may sometimes become too large, and even if the scale of the equipment is ignored, it is currently unclear whether it can be easily realized.

The purpose of the present invention is to scan each point in the background vertically and horizontally after candidate character types are narrowed down using the previously published method.
It provides a method for identifying similar characters using features characterized by the number of times they intersect with character strokes at a predetermined angle, and in particular, the above-mentioned Japanese Patent Application No. 56-169
When narrowing down the candidates using the ``Character Recognition Method'' in the specification of No. 153, most of the information to be handled is the same, making it easier to implement the actual device. It has a red advantage.

1'11・1 The present invention ζ will be explained in detail below with reference to the drawings.

FIG. 1 is a diagram for explaining the description of special information used in the present invention. When scanning upward, downward, leftward, and rightward from a point P in the background, the stroke and scanning This indicates that the number of times the lines intersect with each other at a predetermined angle is 0 times, b times, C times, and d times.

WJ2 diagram (+) (ij) (jjD is a diagram for concretely explaining the features used in the present invention.
If the line intersects with the character stroke at the same angle within the determined range, it is counted as the number of times it crosses, and if it does not, it is not counted as the number of times it crosses even if it intersects with the character area. In Qa, α=4, b=l, c=
3. d=l, but at point P, the angle between the left horizontal scanning line and the intersecting stroke is small, so it is C20, and for the rest α=1, b-4, d=l It becomes. If this is shown in correspondence with Figure 1, Figure 2 (fil, θ11)
It will be like this.

In general, in handwritten kanji, there are differences in the degree of inclination of the characters depending on the scribe, but as mentioned above, the angle between the scanning line and the stroke is used to determine the presence or absence of intersection, so the inclination of the stroke It is possible to determine the characteristics of background points that are stable even for

According to the above-mentioned specification of % Application No. 56-169153, using the above-mentioned 09, input character slams and standard character patterns can be described by multidimensional vectors, and the elements of the vectors are 0 or positive integers. be.

Next, the present invention will be explained in detail. Let ξ and SB be standard slam vectors for categories IJ-A and B, which are described by vectors based on the above characteristics. The similarity between categories IJ-A and IJ-B can be described using an m-dimensional vector as follows.

SA: c + 1) 8B=(1+ε) [-+1 iclI>IIIDI+ 1 (1+t) CI>It IL Ifl)
ε However, 8A=('I t ”'tαv ) p SB
``' (bl t ''・g b, n) gC= (C7,
・−,C,,), 1)=:(d;,...,df
f, ).

Let i = (e s , ..., e i), and l CI
= , ΣCi (Ci≧0, i = 1, −,
The same applies to mtea6)-1, II I' II, 111113 II, etc.

The satay input character slam vector is X = (x 1,・
..., X□input vector X and vector Y = (y l,
..., y, , ) as D(K, Y) = II
The above-mentioned specification of Japanese Patent Application No. 169153/1983 does not provide an example where X-Y II (=II Y-X 11 ).

When X is a member of SA or Sn1m”=
(1+'z)c+ε21)+ε3B, 11)(X,SA)=ll'IC-1-(4
z-151)+a:*l 111) (X, SB
)=II (ε1−ε)C+ε2]D+(ε3−1)
It becomes Ill.

In this case, IIc II > 1lII It ll (1+SCII)II FJIt, so 1>(X
, S, ) is greatly influenced by εIC, and p(χ, $0) is (
ε, -ε) The influence of C is strong. However, in order to identify whether the input vector

Therefore, consider the following vector operation.

"i-¥, θ) = ("1 v"'t"m) However, u1
= Knee r (xl y i-〇) (1=L;'
It is t.

Now, SAB = tt(SA ``a eθ) SBA =''
When (SB SA e) is calculated, SAB is a vector that characterizes the difference between SA and SB, and SBA is a vector that characterizes the difference between SB and SA.

Therefore, □ ξ = SA people + SAB (However, S 5 people = SA 5A
B) SR = 5I3B + SBh Ct, = sauce 5
BB=SB 5BA).

Therefore, Xa = 71 (X: -SAA, θ
)x, = M (knee SBB, o), then yA is a vector that characterizes $λ-likeness with respect to SB, while KB is a vector that characterizes SB-likeness with respect to SA. be.

Therefore, D(WA・SAB) and D(Kn*5BA
) by finding and comparing each, SA or SB
can be identified.

That is, D (XA t 5AB) < D (XB
e 5BA), then X is close to SA, and D (Xi, 5AB) > D (XB t 5BA
), then we conclude that X is close to SB.

Embodiments of the present invention will be described based on the above principle with reference to the drawings.

FIG. 3 is a block diagram showing the configuration of an embodiment of the present invention. Reference numeral 10 denotes a character pattern signal converted into a quantized signal, which becomes an input signal to the character recognition unit 9. In the character recognition unit 9, each point l of the background part of the input character pattern is quantized, and the character pattern signal is converted into a quantized signal. Details of how to rank the complementary characters and output candidate character types as signal marks will be described later. 6 is a candidate character type storage section, and 70 is a signal representing a pair of character types obtained from the candidate character type storage section, which corresponds to categories IJ A and B in the above explanation of the principle. Reference numeral 7 denotes a differential character information extraction unit, which reads the standard character pattern information corresponding to SA and SB as a signal (material) in the explanation of the principle described above.
In the above explanation of the principle $. e SAB t Outputs the differential character information corresponding to tSBB and SBA as a signal 80, and 8 is a character type pair identification section that reads the differential character information as a signal 80, and inputs the input character corresponding to X in the above explanation of the principle. Read the information as a signal (material) and use D(XA, 5An) and D(yBpsnA) in the explanation of the principle above.
When the character types to be excluded from the iron bucket character types are determined and the character types are deleted from the candidate character type storage unit 6 at signal 90, the candidate character types have been narrowed down to one character type. The character type is outputted as a signal 100.

FIG. 4 is a block diagram showing the configuration of an embodiment of the character recognition section 9, in which 1 is a feature encoding section, 2 is a feature selection section, 3 is an evaluation value calculation section, 4 is an identification section, and 5 is a standard character Memory section, 20
is a signal indicating the characteristics of each coded background point, 30 is an information signal of the input character pattern, 40 is a signal indicating the evaluation value of the input character pattern and each standard pattern, and (x) is an information signal of the standard character pattern. For details, refer to the above-mentioned % application 1986-169153.
It is omitted here as it is shown in the Meiji-sho ``Character Recognition Method''.

As explained above, according to the present invention, the above-mentioned patent application
For a character recognition method that characterizes the intersection information of strokes in the background, such as in the document No. 169153, i4U [Character Recognition Method], similar character identification processing can be realized using the same feature information as the character recognition method.

[Brief explanation of drawings]

FIG. 1 is a diagram showing an example of displaying the deviation of each point in the background part of a character pattern.
) is a diagram for explaining the features used in the recognition method that is the basis of the present invention. FIG. 3 is a program diagram showing the configuration of an embodiment of the present invention;
FIG. 4 is a block diagram showing the configuration of an embodiment of the character recognition method used in the present invention. In the figure, 1...feature encoding unit, 2...feature selection unit, 3...-evaluation value calculation unit, 4...identification unit, -15
... Standard character storage section, 6... Candidate character type storage section, 7.
・・Difference character information extraction section, ・・Character type pair identification section, 9...
Each character recognition unit is shown. ':::...

Claims (1)

[Claims] A character pattern formed by converting a character part and a background part into a quantized signal consisting of binary values is scanned in multiple directions from each point of the background part, and a predetermined value is determined for each scanning direction. By counting the number of strokes that intersect at the angle of the range, a plurality of % images are encoded and each characteristic sound is extracted, and the plurality of coded the and their characteristic sounds are used as input character information, and For each character type to be read, each standard character information stored in the same format as the features and characteristic sounds is compared with the input character information, and a recognition result is output. A character recognition unit that outputs candidate characters, a candidate character type storage unit that stores the candidate character a, and a candidate character type storage unit that selects a pair of two character types from the candidate character type storage unit, and selects a pair of two character types from the candidate character type storage unit, a differential character information extraction unit that extracts differential character information consisting of common character information and non-common character information from the standard character information; It has a character type pair identification unit that removes either one of the pairs from the candidate character types, and by sequentially performing pair determination among the candidate character types, removes one character type from the candidate character types one by one, and removes one of the candidate character types from the candidate character types. A character recognition method that is characterized by outputting the last character type as the recognition result when the number of characters is narrowed down to one type.