JPS59231682A - Character discriminating device - Google Patents

Abstract

PURPOSE:To discriminate a candidate character category with high precision by extracting a difference between categories on a standard pattern as to the category of a candidate for an input character pattern, and detecting whether the difference is in a stroke extracted from the input character pattern. CONSTITUTION:The character pattern from an input part 1 is passed through a preprocessing circuit 2, and a rough classifying device 3 narrows down a candidate character category and passes to a structure difference extracting circuit 5, which extracts the differences from the standard patterns of two optional categories read out of a standard pattern storage circuit 4. A stroke extracting circuit 6 receives the character pattern from the circuit 2 and the structure difference extracted by the circuit 5 to calculate similarity on the basis of feature point information, and sends the value, etc., to a structure deciding circuit 7 when the similarity is larger than a threshold value. The circuit 7 decides whether the structure difference pattern received from the circuit 5 is present in a character line extracted by the circuit 6 or not. Thus, the structural difference between the candidate character categories is detected to improve character discriminating performance greatly.

Description

[Detailed Description of the Invention] (1) Description of the field to which the invention pertains A character reading device, particularly a character reading device that reads handwritten kanji, processes “major classification → identification”.

It is a common practice to divide into multiple stages such as Raru-Ii [Major classification → Subclassification → Identification]. At the stage of major classification or detailed classification, features that can be extracted easily are extracted, and candidate character categories are narrowed down by simple click matching. After that, the category to which the input character (Tan) belongs is determined and identified by structural analysis from among the limited candidate character categories.This invention is capable of identifying similar candidate character categories with high accuracy. The present invention relates to a character recognition device.

(2) Description of conventional technology Conventionally, to determine similar candidate character categories.

It was done by the following method. A set of similar candidate character categories is determined in advance through a simulation experiment or the like. In this case, 2 processes are performed to detect whether or not there are features for distinguishing between the similar candidate character categories.
It is necessary to determine it in advance depending on each individual case. There is a problem in that a great deal of effort is required to implement this method as described above. On the other hand, when 9 handwritten kanji characters are to be read, the 2 handwriting deformation is large, so there may be a case where the candidate character category of the input character burn does not exist in the set of similar candidate categories prepared in advance. In such case,
Identification processing is not performed for candidate character categories that are not prepared in advance, which often causes misrecognition, which has been a weakness of this method.

(3) Purpose of the Invention The present invention has been made in view of the problems of the conventional methods described above, and its purpose is to provide a character identification device that is simple and can deal with any combination of candidate character categories. It is about providing. Examples will be described in detail below.

(4) Description of structure and operation of the invention FIG. 1 is a block diagram of the structure of an embodiment of the character identification device of the invention. In the figure, numeral 1 is an input section that receives an input character button. Reference numeral 2 denotes a known preprocessing circuit, which processes input character patterns inputted from the input section 1.

Binarization 9 Perform preprocessing such as position and size normalization and smoothing. The preprocessed input character pattern is sent to the next device 3,
Pass it to 6. 3 is a major classification device.

For example, the number of candidate character categories of the input character pattern is narrowed down to about several dozen using a known major classification method such as line density feature. The candidate character category is passed to the next stage circuit 5.

4 is a standard button memory circuit. Standard strokes are the starting points, bending points, and intersection points of the strokes obtained by appropriately dividing the strokes that make up the character strokes.

and a series of position coordinates of the end point. Starting point above.

The refraction point, intersection point, and end point are hereinafter referred to as scotoma. FIG. 2 is an explanatory diagram of the notation 4 & format of the standard pattern of category ω. Category ωFi, consisting of M-tree strokes.

Each step) 0-ri is N1 e..., N,,...
, No number of writing points. The n-th node y□Hiro of the m-th striker, its position coordinates (
X□, y□) and the writing point are the starting point, M-fold point, intersection point,
A geometrical attribute g□ indicating which of the end points it is, and a straight line connecting the writing point Ymm and the writing point 11i÷1 (hereinafter referred to as standard segment), as in the case of the feature point? +
The number of character lines on the top, bottom, left and right of 7□) is φ□,
Represented by φ□, φ□, φ-1 set ◇ 5 in Figure 1 is a structural difference extraction circuit, which stores standard patterns of any two categories among the candidate character categories obtained by the above-mentioned major classification. The main function is to read out the standard patterns from the circuit 4, extract the difference between the two read standard patterns, and pass it to the ninth stage 6 and 7, and this is realized by the ninth stage means.

Of the two standard patterns read out from the memory circuit 4,
The standard pattern with the larger number of standard segments is designated as old, and the other is designated as R'. At this time, the standard inter-segment distance d is calculated using the 9th equation for the standard segment tree 9 of rR, for example, 1171, 1wI, and all standard segments of IR' (117m a). Jmr+ here
=Ymn+s Vmn * tqrr:n'tsuYm'n
'+s Vmj.

d (Vmn * Cape mti') -11) m
m' Yma ll + ll Ym'n'+ t
-11'mn+I 11+α・(117mn, '
llFm'n') ...Song...
・Inner use (11Here, 11111 is the norm of vector i, (V, to') Inner product of ti vector rent and l', α
is a constant. When θms is a predetermined threshold.

select. The Fqlvln*him- is passed to the next stage circuits 6 and 7 as a C-segment. w that satisfies formula (2)
:a'' does not exist in goods', it is determined that there is no standard segment of take' corresponding to da
11mm as standard pattern IR P-segment, also tR
' is assumed to be the N-segment of ', and is passed to the ninth stage circuit 6.7. After performing the above processing on all the standard segments constituting the IR, if there is a standard segment of R' that does not correspond to a standard segment of P-segment, @N
- Pass it to the next stage circuits 6 and 7 as a segment. The batan made up of the P-segment, C-segment, and N-segment is called a structural difference batan.

In the explanatory diagram illustrating an example of structural difference extraction between standard patterns in the third figure, 6 (a) is the standard pattern of "Ten J t (bl is the standard pattern of "husband"). 11.12...・,1B,21,
22. ....

29 is a solid line arrow enlarged standard segment connecting two chain points. The standard number of segments constituting the standard batan ``ten'' and ``husband'' is 7 trees and 8 trees, respectively. Therefore, by the means described above, as IR-rhuJ, fft'='heaven'.

When determining the correspondence between each standard segment, the same figure (cl
The result is obtained. The standard segment of 叡 = ``天'', which corresponds to the first standard segment of 侭 - ``husband'' (the standard segment marked by the vector from difference point 27 to 22 in the same figure) does not exist.

The standard segment is the P-segment of the beast, N of tR'
-becomes a segment. On the other hand, all standard segments of 1 (R'-rten) are associated with standard segments of 9th generation-'husband', so there are no P-segments of [R' and N-segments of IR]. .In Figure 4,
Examples of P-segments, C-segments, and N-segments, that is, structural difference patterns, obtained from the standard batons "Ten" and "Husband" by the structural difference extraction circuit 5 are shown.

Reference numeral 6 in FIG. 1 is a stroke extraction device, and in order to realize this, 9 for example, JP-A-57-182878 or literature (Takahashi, Masuda: Stroke extraction using the order of occurrence of one stroke point. ``Basic study of handwritten kanji recognition using methods'', Transactions of the Institute of Electronics and Communication Engineers (DJ, vol. J-65D)
, no, 10. pp, 1294-1301 (1982-
10)) etc. can be used.

FIG. 5 is a block diagram of the configuration of one embodiment. In the figure, numeral 33 is a known feature extraction circuit that extracts feature points representing the geometric shape of nine character lines from the thinning button using a known method. Serial numbers are assigned to the extracted feature points in order from 1. There are two types of position control for the extracted feature points (end points and 9 branch points).

(Intersection points, isolated points, sampling points) and numbers assigned to feature points connected to the feature point.

Hereinafter, this will be simply referred to as feature point information. The feature point information obtained by the circuit 35 is passed to the next stage circuit 35.

35 is a feature point storage circuit, which includes the feature point extraction circuit 3.
The feature point information extracted in step 5 is stored.

Reference numeral 37 denotes a corresponding candidate point selection circuit, which selects feature points within a predetermined number that satisfy a predetermined condition from among the feature points extracted by the feature point extraction circuit 35 and selects the structural difference read out from the structural difference extraction circuit 5. It is selected as a candidate corresponding to each writing point in the standard segment of the pattern. Means for realizing the circuit 37 include the means disclosed in the above-mentioned Japanese Patent Laid-Open No. 57-182878 or the above-mentioned documents. The selected feature points are called corresponding candidate points. The corresponding candidate points are passed to the next stage circuit 58.

3B is a connectivity determination circuit having a ninth-order function. L feature points W' + as corresponding candidate points for the writing point y□
Suppose that ft2+ ``·+tfL selects L' feature points tr'', 叡-..., 'Ll as corresponding candidate points for the writing point 1ryy1n+1 following the writing points 1rn, ll.

Select one corresponding candidate point 1, for example, n', 1'' from among the corresponding candidate points of the foil point Vyum*W+nn+1, and create a pair of corresponding candidate points such as ip=(da1.kan15(i= 1.2...., L:j'-1', 2',
..., r,).

Since L and L' corresponding candidate points are selected for the r point 1rl, In, rffl□, respectively, Lx
L' pairs of corresponding candidate points are qqed. For all the pairs of corresponding candidate points in this LxL'' set, it is first determined whether the corresponding candidate points forming the pair of corresponding candidate points are connected on the character line of the input character button.Next, the corresponding candidate points are connected to each other. Find a character line that connects the corresponding candidate points by the shortest path only for the pair of corresponding candidate points that are the same. The main function of the circuit 58 is to calculate the similarity between the first □-1rma+i lrmn, and is realized by the following method. FIG. 6 is a block diagram of the configuration of the circuit 58.

61 is a corresponding candidate point storage circuit, which stores the number of corresponding candidate points for each writing point and the previous sL selected as a corresponding candidate point! Memorizes feature point numbers.

62 is a character line tracing circuit, and the feature point storage circuit 6
Based on the connection relationship between i feature points stored in 5, it is determined whether the corresponding candidate points read from the circuit 61 are connected on the character line of the input character pattern, and if they are connected, it is determined that they correspond. Find the shortest path connecting candidate points. This can be realized using a technique known as the shortest path problem in graph theory. Only for pairs of corresponding candidate points in which corresponding candidate points are connected, the length of the shortest path connecting the corresponding candidate points is determined and passed to the ninth stage circuit 63.

Reference numeral 63 denotes a similarity calculation circuit 1. For example, corresponding candidate points 91゜1゛ forming the pair of corresponding candidate points 1p"' = (11' + ff") are connected on the character line of the input character button. .

And the length of the shortest path connecting W' t tI' is hII'
If the result is obtained by the circuit 62, the degree of similarity between the standard segment) J'17mn is calculated. Here, the input segment l −ffj'−f
If it is il.

The similarity r is given by the following equation.

Here 111711, press II, -II habe/)
hl(j *rQrnnco / le A te;6su, W
is a weight function and δ is a threshold value. FIG. 7 is an explanatory diagram of formula (3), and in FIG. 2, a and b surrounded by double circles are the writing points of a standard button, and c, d, . . . are surrounded by -type lights.

h is a feature point extracted from the input character pattern.

A solid line arrow 71 is a standard segment pointing from a to b. Fruit connecting C and d, d and e,...9g and h $972
, 75, . . . , 76 are character lines. Here a.

b as above @ avma s kmn + i s also c, d
If 1.1' is the corresponding candidate point, then 71 is a standard segment)
Illusion...

This is a measure to prevent the vector 77 of the dashed-dotted line from C to h from becoming unduly high in similarity to the character line standard segment which is greatly curved as shown in FIG. 7.

The similarity value r obtained above is r≧r'rii <r'
rit is a threshold value), the value of the similarity, the corresponding candidate points forming the corresponding candidate point pair, and the number of the feature point on the shortest path connecting the corresponding candidate points are passed to the next stage M7.

Is it the packing difference slam received from the structure difference extraction circuit 5 in the 7-loan structure determination circuit shown in Figure 1r5J?

The main function is to determine whether the shortest path extracted by the stroke extraction device 6 exists within the character line, and this is realized by the following means. FIG. 8 is a block diagram of the circuit 7. As shown in FIG.

In the figure, the horizontal axis is the X axis and the vertical axis is the y axis.

Each character storage circuit has a resolution of n (for example, 12B) pixels, and stores the input character pattern preprocessed by the preprocessing circuit 2.

Reference numerals 54 and 55 are character line number counting circuits that scan the characters stored in the character storage circuit 53 horizontally or parallel to the vertical axis and count the number of lines in one character. A certain coordinate value X'
* When Y' is specified, the counting circuit 54
The counting circuit 55 counts the number of character lines between 1 and x'-1 and between x'+1 and n, respectively. Count each number and make 9
It is passed to the next stage circuit 56.

56 is a character line number storage circuit IS') which stores the counting results of the character line number counting circuits 54 and 55 and writes them into the ninth stage circuit 57.

57 is a pair determination circuit. The extracted structural difference patterns are successively output from the structural difference extraction circuit 5. The following process is repeated for all standard segments forming the structural difference pattern. Let n be the standard segment of interest. T
For II□7, the following embedding is performed in order from the corresponding candidate point pair that gives the highest similarity value obtained by the above-mentioned straight extraction device 6. 1 by gt, 6
For example, the value of the position coordinates of all the feature points on the shortest path connecting the corresponding candidate points forming the corresponding candidate point pair -of is
The number of character lines for each feature point in each of the upper, lower, left, and right scanning directions is read from 56 above. The average value of the number of character lines is determined for each scanning direction (up, down, left, right, left, and right), and φυ is calculated for each.

ψ0. ψ1. Let ψfu. The standard segment r11.
．． ．． ．． The number of character lines on the top, bottom, left and right of φ 7. φλi, φ
瓢□φ翫ii. A certain scanning direction z (z-U, D, L
, R) Nitsuite.

Calculate the degree of dissimilarity d (zl) given by the following formula.

d(ri-1φ-1?+ Hit...song...hit...(4) ζjin d (z)≦dti(z) (dtm
tz) is the threshold value), it is determined that a character line corresponding to the standard segment ty+□, which is a structural difference pattern, exists on the input character button. On the other hand * d(z) > dt
In the case of i(z), the shortest path (character line) is rejected as not corresponding to the tQ□, and corresponding candidate point pairs having a quadratic degree of similarity are similarly examined. When all the corresponding candidate point pairs are finally rejected, it is determined that the character line corresponding to the standard segment does not exist on the input character button.

Through the above-mentioned processing, the two standard bang beasts.

It is determined whether all the standard segments of the beast' structural difference button were present on the input character button.

NP: Number of P-segments of R that existed on the input character button Nc: # C- of R that existed NN: tr N of the beast that did not exist
−〃, the match between the standard button l< and the input character button f is defined by the following equation.

fll? l=g(Np, Nc, NN)
・・・・・・・・・・・・・・・・・・・・・・・・(5) Here, g is 2 For example, g(Np+Nc+NN)=wp・Np+w6aN(+w
HaNH=-161 etc. can be used (however, +
wp, Wc, ww h are weight functions). flR
)>f, the input character BATA is determined to be structurally more similar to the category E than to the category BUT, and the result is passed to the output unit 8.

(5) Explanation of effects As explained above, according to the apparatus of the present invention.

For any combination of candidate character categories.

S) Identification processing using o-ri extraction becomes possible.

In addition, it detects character lines that are structurally different between the candidate character categories, and then determines whether or not the character line exists on the human-powered character button. By referring to the number of character lines, the positional relationship between nine character lines can be determined in advance, so that the ability to identify two characters can be greatly improved.

[Brief explanation of drawings]

FIG. 1 is a block diagram of the structure of the present character identification device, and FIG. 2 is an explanatory diagram illustrating the recording format of the standard pattern. FIG. 6 is an explanatory diagram for explaining an example of structural difference extraction, FIG. 4 is an explanatory diagram for explaining structural difference slam, FIG. FIG. 7 is an explanatory diagram for explaining similarity, and FIG. 8 is a block diagram of the structure determination circuit. 1... Input unit, 2... Preprocessing circuit, 3... Major classification device. 4... Standard button memory circuit, 5... Structural difference extraction circuit. 6... Stroke extraction device, 7... Structure determination circuit,
B...output section, 11, 12. ..., 1B, 21, 2
2. ..., 29... Vanishing point, 33... Feature point extraction circuit, 35... Feature point storage circuit, 37... Corresponding candidate point selection circuit, 3B... Connectivity determination circuit, 61. ...Corresponding candidate point storage circuit, 62...Character line tracing circuit, 63...
・Similarity calculation circuit, 71...standard segment), 72,
73. ..., 76... Character line, 77... Input segment, a, b... Line point, c, d,..., h...
Feature points, 53...Character storage circuit, 57...Pair determination circuit. 54.55...Character line number counting circuit, 56...Character line number storage circuit. Patent applicant Nippon Telegraph and Telephone Public Corporation Patent attorney Mori 1) Hiroshi 1 (21 2(2) (Q) (b) 31
21 Figure 6 Figure 7 Procedural amendment (voluntary) 1. Display of case: 1982 Patent Application No. 105490
No. 2, Title of the invention Character identification device 3, Relationship with the case of the person making the amendment Patent applicant address 1-1-6 Uchisaiwai-cho, Chiyoda-ku, Tokyo Name (4)
22) Nippon Telegraph and Telephone Public Corporation Representative Tsune Shinfuji 4, Agent address 4-17-1 M-ri, Nishi E, Arakawa-ku, Tokyo 3FC Sawara Mann Name (7484) Patent attorney Mori 1) Kan 5 , The number of inventions will not increase due to the amendment 6. Figures 1 and 5 of the drawings to be amended. 7. Contents of the amendment. Contents of the amendment (1) Figures 1 and 5 of the drawings are amended as shown in the attached sheet. do. that's all

Claims (1)

[Claims] In a character recognition device equipped with a function of extracting strokes, a standard stroke memory circuit stores standard strokes given by the position coordinates of end points, and a difference between multiple categories selected as candidates for input character strokes is detected. A structural difference extraction circuit that extracts on a standard button, detects whether the extracted difference is in the input character button or an extracted stroke, and determines whether the input character button belongs to any candidate character category. A character identification device characterized by comprising a structure determination circuit that determines the structure of a character.