JPH03291781A - Character recognizing device - Google Patents

Abstract

PURPOSE:To enable the recognition of characters even between similar ones without erroneous recognition by verifying candidates based on the candidates of a block arrangement type and a block attribute. CONSTITUTION:The picture data of a check is read and inputted by a picture input part 2, a preprocess part 3 performs the process of the preprocessing and thinning a line to input the data to a global variable (memory) 20. An inference mechanism 40 collates the condition part of the rule part of a knowledge base 10 and the content of a working memory 30 to control the firing of the rule. The result of execution parts 12B-12E is inputted to the working memory 30 to display it on a display part 4 through the global variable 20.

Description

[Detailed Description of the Invention] Purpose of the Invention (Industrial Application Field) The present invention is a character recognition system that recognizes characters by utilizing character arrangement types in order to efficiently and reliably recognize handwritten characters. Regarding equipment.

(Prior Art) As a conventional method for recognizing handwritten characters, there is, for example, a character recognition method described in Japanese Patent Laid-Open No. 1-233585, which was devised by the present applicant. In this method, after performing pre-processing of smoothing and thinning of character data and line structure analysis, first, a recognition target candidate block that is determined to be one character higher as a global structure recognition is extracted,
Character recognition is performed for each block. As for blots that have not been recognized, segmentation, line structure analysis, or data reconstruction is further performed until recognition is completed to advance character recognition.

(Problem to be solved by the invention) However, in the above-mentioned conventional handwritten character recognition method, for example, the example shown in FIG. It had become unrecognizable. Specifically, when recognition of the handwritten characters in the case of FIG. 4 is proceeded, the blocks B1,1 to BL8 will finally be recognized as shown in FIG. Then, when recognition is performed for each block, blocks Ill, ], B1.2BL3. Bl, 5. B10. I
Can IL8 recognize characters?Block B14 and old 6
There will be a misunderstanding about this.

In other words, blocks B] and 4 are recognized as the number 1゛°, and blocks B1 and 6 are also recognized as the number "1".
It is recognized as ゛°.

Furthermore, if we add symbol recognition information such as Sen 1 ~ bar, cent point, etc. (-1) to the above conventional method, 1 will be the old block, and 4 will be the prison's "bi" or "sen".
It is difficult to determine whether the block is a bar or not, and it is also not possible to determine whether the number '1'' tt is a bar or not. 2 types of ingredients-1-
In some cases, the problem was that they were unrecognizable.

By the way, as shown in Figure 4, in the case of a U.S. check, it is composed of letters that are written in the form of letters or truota cent ota, cent bar, sen 1 - point 1 -, and is written in the form of a handrail. Although there are individual differences, most can be categorized into one type or another.

In the case of U.S. check handles, most of them are
The actual situation is whether it belongs to any of types A to F in the figure. Incidentally, this is not only the case with American checks; if there is a certain standard of writing format, then you can see the signs of this.

This invention was made due to the above-mentioned circumstances,
An object of the present invention is to provide a character recognition device that can reliably recognize characters based on the arrangement type by using a layout type that morphs the handwritten characters with a certain fixed writing format. Our goal is to provide the following.

Structure of the Invention (Means for Solving the Problems) This invention provides a knowledge base that stores rules necessary for recognition from information about a group of unrecognized characters, and an inference that uses the knowledge base to make inferences. Regarding a character recognition device that recognizes characters using a mechanism, the above object of the present invention is to extract a single character block from the recognized character block T, and to create a block candidate for each character block. stomach·
a block arrangement type extracting means for extracting, as a block arrangement type candidate, a block arrangement type candidate that approximates the arrangement of the character block extracted by the block extraction means from among a plurality of predetermined block arrangement types; , attribute a for assigning a block attribute candidate to each block based on the block layout type extracted by the block layout type extraction means; This is achieved by providing a verification means for performing the verification.

(Operation) In this invention, after analyzing the line structure of the character group to be recognized and extracting the characteristics of the block, the currency unit type (yen and sen for yen, tor and toru for toru) is applied to the block. Regarding cents 5 bonds, it is a display type in which the currency notation above the small point and the notation below the decimal point are registered in advance in various types depending on how they are displayed, such as Bon 1- and Bence. Here, for convenience, we will use the Sen 1 type as a target for Toru), and by verifying this hypothesis, we will display the recognition results. I try to recognize numbers accurately.

(Embodiment) FIG. 1 shows a configuration example of the present invention, in which a control unit 1 consisting of a CPU etc. controls the whole, and handwritten characters written on a check etc. are inputted to an image input unit 2. It is read optically and processed by the preprocessing unit 3 such as thinning to create a global variable (
memory) 20 times. The global variable 20 includes a block extraction unit 21. A block feature extraction section 22 and a relative position code creation section 23 are connected, as well as a working memory 30 that stores variables such as block features, block attributes, cent type hypotheses, information representing recognition flow, etc. . In addition, the inference mechanism 40 checks the condition part of the Lulu part of the knowledge base 10 and the contents of the working memory 30, and controls the firing of rules.
Condition part 11B-11E and execution part 128-12E of ~E
The results of the execution units 12B to 12H are manually input to the working memory 30, and are displayed on the display unit 4 via the global variable 20.
will be displayed.

The recognition method of this invention is disclosed in Japanese Unexamined Patent Publication No. 1-116 filed by the present applicant.
781 and JP-A-1-116782, FIG. 2 is a flowchart showing an example of the operation of the present invention, and the recognition operation will be explained with reference to this flowchart.

First, the image data of a check as shown in FIG.
As shown in FIG.

Line thinning processing is performed (step 521. Then, as shown in FIG. 6 in line structure analysis, blocks Bl, 1 to B1 are
, 8 are extracted (step 53). This block extraction is performed on valid blocks, and the average height of blocks with a number of dots or 4 or more is set as EFI, and height > EFI
If the average height of blocks of x O, 9 or more is EF2, a block with a height greater than EF2Xl/3 or a block with a number of dots greater than EF2Xl/2 is an effective block. After that, blocks BLI~BL8
(step S4), the block features include block height, block width, block aspect ratio, horizontal direction and center of gravity position, single character recognition result, bar-likeness,
There are under-bar characteristics, upper-bar characteristics, middle-bar characteristics, diagonal under-par characteristics, and diagonal upper-bar characteristics. The degree of height is expressed as block height/EF2-αl, and when αl>1.3, it is extremely high °°, α
1 When > 1.1, “fairly high” α1 When > 0.9, “high degree”, when α1>07, “slightly low degree” α1>
When it is 05, it is “low°゛α1〉When it is 03, it is “quite low°”
゛, when αl≧0, it is assumed to be extremely low. Also, the degree of width is expressed as block width/EF2-β1, and β
When l〉09, it is considered as ``quite wide'', when β1>05, it is considered as ``wide'', and when βl≧0, it is considered as ``slim''.

Furthermore, the aspect ratio is expressed as height/width=γ1, and γ1. <
When 0.67, ``considerably horizontal °°, γl <, 0.91
When ``Landscape'', when γ1 < 1.43, ``Very horizontal''; when γ1 < 0.91, ``Horizontal'', γl < 1.4
In the case of 3, it is ``medium'', and in other cases, it is ``vertically long.
It is expressed as a value of 0°'.

When block feature extraction is completed, the block attribute pre-assignment condition section 11B in the knowledge base 10 applies a block feature description, and the execution section 12B applies a series of rules B in the form of assigning block attributes. 1.1 shown in FIG. 6 is performed (step S5).
-B10 is given the following attributes.

That is, as shown in (8) in FIG.
is a number (Rule B2), block old, 2 is a number (Rule B2), block BL3 is a number (Rule B2), block BLII is a diagonal bar (Rule B10), block BL
Attributes such as 4 is a number (Rule B2), block old, 5 is a number (Rule B2), block old, 6 is a number (Rule B2), block [lL7 is a number (Rule B2), block BL8 is a number (Rule B2) is given. In addition,
The rules will be explained later.

After assigning such block attributes, the steps shown in FIG. 7 (8) to
A cent type hypothesis as shown in (6) is made based on rule group C (step S6). That is, the condition part 11G for setting the layout type hypothesis in the knowledge base IO is a description using block characteristics, block attributes, and relative position codes.
The execution unit 12G sets the cent type hypothesis as shown in FIG.
This is done by applying rule group C as shown in B). Figure 7 (A
) ~ (G), the dollar order is the m currency notation part of 1 decimal point (cent point) or more, and the sen 1 ~ order is the currency notation part of 1 decimal point or less, and in (C) ~
The cent type C-F in (1:) has a cent bar CII or (t-t!), and the cent bar type D-F in Fig. 7 (D) to (F) has a 100°', " A cent signal such as 'XX"' [:hl or
J: A cent type E or B is hypothesized, and a new attribute is updated for the hypothesized cent type (step 510). zu 2t, sen 1
- As a type hypothesis (step 56), cent type E and rule CBI are determined by rule C3d3, which will be described later.
Accordingly, two scales, Sen 1 to Type B, are temporarily constructed, and for Cent type E, diagonal cent bars and cent symbols are given as attributes of blocks BL4 and B1, 6, and for cent type B, blocks old and 7 are added. And a fidget cent order number is given as an attribute of BL8.

Next, the rules are applied and the attributes are updated.

The condition part LID of hypothesis verification 1 is the cent tights hypothesis, block attribute, bronc feature, description by relative position coat,
The execution unit 1.20 updates the block attributes in line 77. If it does not match any of the rule groups and there is a valid block whose attribute as a component of the cent type is not imaged, then the type hypothesis is rejected (hypothesis verification 1). Then, as shown in Figure 3 (C), Sen 1~
As an update of attributes based on type E, block B[
, 7 is Sen 1 symbol (rule DK3), old flock, 8
Sen [・Ki ℃ (Lulu DK4), Block B1.3!
4 cent over digits (Lulu Ddl), block old 5 f
, l Sen I order number (Lulu l'1dl), block 31-1 updates the Truota number (rule D1), block old, 2]-Luota number (rule old).

In addition, the update of attributes based on sen 1 to type B is as follows: block tl 16 is sen 1 - point (rule DPI), block [lL] is 1 - 1 digit (rule DI),
Blocks B1 and 2 are numbers 1 to 1 (rule Dl-1)
), block old 44 is 1-Luota number (rule Dl
-1), block 131.3 updates the genus/1 cow of 1 to 100 digits (rule D1), and block 81.5 updates the genus 1 cow of 1 to 100 digits (rule D1-1). After updating the attributes and verifying the hypothesis (step 510), verification 2 using segment processing and overall placement is performed (step 520). The cent type hypothesis B written in ``-'' is checked by rule E-PO56, probably because the relative positional relationships between blocks old 4 and 146 and blocks old 8 are appropriate and noisy, as shown in Figure 3 (l). and rejected. Verification 2 is terminated by determining whether the rejection rule is met/acceptable, and the recognition result is not displayed on the display unit 4 (step 530).
), perform the necessary check processing (e.g. large amount processing) in line 11 (
Step 531).

Here, to explain the attributes given to the extracted blocks,
Block attributes include ``Numbers''°, °``Contains an underbar'', ``Bar'°, ``Continuous sand letters'', ``Cent order low sequence 1j J numbers'' 2, ``Contains upper bar,''``Don'tknow°''
There is a ``°'', and the attribute ``Number'' is updated to ``)・Luota numeral'', ``° Centauta numeral ゛°, ゛° Senl-symbol ゛°, and ゛° Cent point.'' “
For '° that includes an underscore, it is updated to ``1 sen - 1 orc character that includes paanhaha'', or it is updated to ``2 cent order letters that include an underbar'', and ``2 cent order letters that include an turbulence''. is updated to "cents order 2 characters °" and "cents 1 - bar". ′.The attribute bar “′ is updated to ゛°sen[・
The bar is updated to ``°'', the attribute ``Continuous digits'' is set to ``Truoter continuous digits'' or senl-oak continuous digits, the roughness is subtracted by 1, and the attribute ``Continuous digits'' is changed to ``Dolorder digits'' and ``'1'' Updated, °゛cent order consecutively, number'' is “sen1”
- Updated to order number ゛° and sen 1 to order number ゛'. Attribute "Sen1-Ota" for low consecutive numbers "Sen1~
The order number ゛ and °゛sen1-order number °' are given. Furthermore, if the attribute is not known, the ``1 order digit'' or the percent over digit '' or the ``sen] symbol ``'' or the ゛sen 1~ pointer digit is given.

Detection of the genus +1 based on the relative positional relationship is shown in FIG. 9 as a four-position map, and in FIG. 10 as the position and shape of the block. Also, an example in which L block is created based on block old 4 is shown in Fig. 11, and the horizontal direction of block old 5 is Xm1n.
z center, X center of gravity; center, X1llaX=center, vertical direction is yi1. , - middle, y center of gravity = middle,
YmAx-middle, and from FIG. 12, the relative position coat is expressed as (26, 2F+). Also, is block B1, 7 lateral or ×, 11, - center, × center of gravity -
Light + XmRX- light, vertical direction is ym+n
- Mill 1 ~ Le, y center of gravity - Mill I - Ru, y A, X - Middle, and the relative position court is from Figure 12 (29, 26)
The horizontal direction of the block 8148 is Xm1n-right,
- It is middle, and the relative position court is represented by (33, 25) from FIG. As shown in FIG. 12, there are 35 combinations of relative positions in the vertical and horizontal directions.
It is expressed as (2[1,261) as a relative position code of After being created by the relative value device code creation unit 23 and stored in the working memory 30 via the global variable 20, it is compared with the code in the condition part of the rule written in advance, and the code is given before the block attribute +1 (step S5), hypothesis of cent type (step S6), assigning/updating attributes based on cent type (
It is used in step 5IG) and verification by segment processing and overall placement (step 520). As shown in FIG. 13, the relative position coats with respect to block a are x-26 and y-26 for both blocks b and C. In order to express the ``-lower relationship for block a, if we scan upward or downward from any one point in blocks b and c and intersect at ``2'' as shown in Figure 14, we get u-+1-d, and the lower The time is u-d-u. However, ud is used only when x-26 and y-26.

Next, each side of the rules within the knowledge base IO used in this invention will be explained.

Rule 810゜ Condition part l Valid block 2 Height = ``Very high'' or ``Quite high'' or ``High'' 3. MA direction center of gravity position〉30 4. Likeness of diagonal bar〉72 Block characteristics of 1 to 4 above When there is a block with , the effective section gives that block as an attribute.

Rule B2 conditional part ゛Diagonal bar 1. It is a valid block 2 Height: 1 ``extremely high'' or ``Quite high'' or °“high” or ``Slightly low'' or °low’ or “Quite low.

3. Character-likeness = ° ``may be a number'' or ``is a number'' Effective part attribute ``number °.

Rule C3d3 that gives conditional attribute: There is a diagonal bar block and an effective block at its lower right position. However, the lateral position of the °°diagonal bar'' block〉40 Effective part Assuming that the effective block is the "°cent symbol °", a hypothesis of cent type E is generated.

Rules CBI.

Condition part 1, the attributes of the first (β) and second (α) valid blocks from the right end or the digits °'' or °small numbers `` 2 The heights are both ``slightly low'' or the digits are low °゛ or ° There are blocks α and β, (X center of gravity of α) and (X center of gravity of β) 3 There is an attribute or number near the side of the above α and β blocks.
Or ``There is a block γ of continuous numbers °°, and the X centroid of α −
X centroid of γ (height of γ) × 3 4 The height of γ is 15 times or more the height of α or β. Using the effective part blocks α and β as cent order numbers, a hypothesis of cent type B is generated.

Rule DK3.

The condition part cent type E and the attribute parameter diagonal sensor 1 - Is there a valid block that is the lower right (ichi-sbar) of the block? Let it be an E cent symbol block.

Rule IMI.

Condition part senl-type E, attribute diagonal sen)-bar pablo nk, its left” 2(ichi-s-Ddl
) in the position where the attribute is ``number par or small number''.

, let it be a cent type E senl-orc numeral block.

Rule D1 Conditional part For all Sen I-types, if there are two blocks (assumed to be α and β), the block genus is expressed by the positional relationship or the following inequality.)'! ``There is a block γ of number pa.

X centroid of γ x X centroid of α and β Height of γ〉0.9 times the height of α or β execution part Let γ be a dollar order number.

Rule DI-1 conditional part For all sen[・types] Doruo Ta ``number'', ゛sen 1 ~ order number' Then, the following positional relationship °゛number'' There is δ.

X center of gravity of α〈X center of gravity of δ Relative position of α seen from δ code-y-11,20,2], 26.27β, γ? block 1~ execution part Let δ be ゛[・Ruoder number′°.

Rule DPI.

Condition part sen]・When the type is divided into cent order digits h) 2 blocks (α, β) for all types (the X centroid of α is the X centroid of β), there are two There, (γ
, the X centroid of δγ is the X centroid of δ (assumed to be the X centroid of α). δ
``1°'', ``2°'', and ``6°'', and δ is small and even smaller than γ. The position is ym + of δ
11 > X centroid of γ.

execution part Let δ be the cent point I-.

Rule E-PO56; Condition part Sen1-Type B, °゛Dollar order number''”α
, β, there is a sent order number “°γ, α2 β, γ
or in the following positional relationship.

Relative position code of γ seen from β; 1cni-bar Relative position code of α seen from β; 1cni-Ddl Here, 1chi-s-bar is the positional relationship between the diagonal amber and the block below it, (x-26
, y-26, u-dd) or (x-26, y-27)
or (x-27y-2e) or (x-29, y-26)
It is.

Rejecting the execution part cent type hypothesis Effects of the invention U As mentioned above, according to the character recognition device of the present invention, for handwritten characters whose written format has been determined to some extent, an arrangement type that typifies the handwritten characters is used. Since it hypothesizes which arrangement type it belongs to and proceeds with recognition based on that hypothesis, it can reliably recognize objects that are very similar, such as the number 1' and a diagonal bar, without misrecognition. be able to.

[Brief explanation of drawings]

FIG. 1 is a block diagram showing the basic configuration of this invention, FIG. 2 is a flowchart showing an example of the operation of this invention, and FIGS.
The figure is a diagram for explaining the invention in detail, FIG. 7(A)
~(G) are diagrams showing the types of cent types, Figure 8 is a diagram for explaining the horizontal center of gravity position, Figures 9 to 11 are diagrams for explaining the detection of attributes based on relative positional relationships, 1st
FIG. 2 is a diagram showing a relative position court, and FIGS. 13 and 14 are diagrams for explaining a 1chi-s-bar. DESCRIPTION OF SYMBOLS 1... Control part, 2... Image human power part, 3... Preprocessing part, 4... Display part, 10... Knowledge base, 20...
・Global variable, 21...Block extraction unit, 22...Block feature extraction unit, 23...Relative position code creation unit,
30...working memory, 40...reasoning mechanism. Unexamined Japanese Patent Publication No. 3 291781 (8) Fig. Zent μbu E 6 grooves - Raw O length new CD) Hypothesis O proof 2 No. beam roof 1/t: Not applicable・ ↓ Verification φ Winter・Complete No. 71- , Type fig': Yo 51g LO text 1 catty ↓ The prohibition rule is a skeleton (Ruzushi E-PO 5; 5) ↓ Former F Yamasho name May 1990 for the cattle that are deeply respected. 29th

Claims (1)

[Claims] 1. A character recognition device that recognizes characters using a knowledge base that stores rules necessary for recognition from information regarding a group of characters to be recognized, and an inference mechanism that makes inferences using the knowledge base. block extraction means for extracting a single character block from a group of characters to be recognized and assigning a block candidate to each character block; block arrangement type extraction means for extracting block arrangement type candidates that approximate the arrangement of character blocks; and assigning block attribute candidates to each block based on the block arrangement type extracted by the block arrangement type extraction means. A character recognition device comprising an attribute assigning means and a verification means for verifying a candidate based on the block arrangement type candidate and the block attribute candidate.