JPH03282983A - Method for extracting character - Google Patents

Abstract

PURPOSE:To surely extract characters even when a bar is not straight in the horizontal direction by detecting character information from an end point and detecting a bar based upon the ratio of the shortest distance between respective end points to an effective distance. CONSTITUTION:A block having the prescribed number of continued dots or more and height or the number of dots larger than a prescribed value is extracted from the character information of a check at first. Then, each of obtained blocks BL1 to BL7 is divided into a matrix area of 3X3 and the numbers of vertical, oblique and horizontal masks are found out. Then, reliability including a bar is calculated in each block. After finding out the reliability in all the blocks, the block having the highest reliability is selected and end points in the block are extracted. Then, an effective bus length and a straight distance between respective end points are found out and the bar is detected based upon the ratio of the straight distance to the effective bus length. Thus, respective characters can surely be extracted.

Description

[Detailed Description of the Invention] Purpose of the Invention; (Industrial Application Field) This invention recognizes only characters (numbers) from text information including bars handwritten on checks, etc. by pressure testing and removing the bars. Concerning a character extraction method for.

(Prior Art) Conventionally, when character recognition is performed on character information including bars, a method for detecting bars from this character information is shown in Fig. 14 (A
) and (B) are known.

In this method, as shown in FIG. 14(A), a straight bar 1 is drawn horizontally in advance as a reference line, and characters 2 are written on it. When recognizing character 2 written above bar 1, first count the total number of dots in the horizontal direction of each character group in the vertical direction, and then create a histogram of the number of dots in the vertical direction. The same figure CB
), and the extremely large portion is determined to be bar 1 of the reference line. Then, the upper character 2 is determined based on this reference line.

(Problem to be Solved by the Invention) However, in the above bar detection method, since a straight bar is drawn in advance in the horizontal direction as a reference line, it can always be detected by creating a histogram of the number of dots in the vertical direction. but,
A problem arises when a straight bar is not drawn in advance as a reference line. For example, when the bar 3 itself is handwritten as shown in Figure 15 (A), the bar 3 is not necessarily horizontal, and even if a pistogram is created by the number of dots, it is difficult to distinguish between the characters and the bar. I didn't understand the judgment of the kiss bar.

This invention was made due to the above-mentioned circumstances,
An object of the present invention is to provide a method for reliably detecting bars and extracting characters from character information including bars, even if the bars are not horizontally straight, or even if the bars themselves are handwritten. It is about providing.

Structure of the Invention: (Means for Solving the Problems) The present invention relates to a character extraction method for detecting bars from character information including bars, separating the bars, and extracting characters. detects the end points from the character information, detects the straight-line distance between each end point, calculates the number of dots connecting each end point, and calculates the distance between the end points based on the ratio of the shortest distance between each end point and the effective distance. This is achieved by detecting the bar and separating the detected bar to extract the character.

(Operation) This invention reliably distinguishes between handwritten characters and bars that are preprinted or handwritten for character entry, removes the bars from character recognition targets, and extracts only the characters. That's what I do. When detecting bars,
In relation to the letters, there are 6: under bar, middle bar, upper bar, diagonal C, diagonal under bar, and diagonal upper bar.
The confidence level of each bar is calculated using a type of formula, and the bar with the highest confidence level is detected.

(Example) Hereinafter, an example of the present invention will be described with reference to the drawings. First, this embodiment is provided to clearly distinguish between dollar orders and cent orders when recognizing handwritten amounts on U.S. checks as shown in FIG. 2 or FIGS. 11(A) to (F). This is a method of automatically detecting the cent bar that has been entered, and by detecting the cent bar, it can be used to recognize dollar orders and cent orders.

FIG. 1 is a flowchart showing the operation of this invention,
Broadly speaking, the process consists of detecting an area containing a bar from text information, detecting a bar from within that area, and character recognition based on the positional relationship between the detected bar and each character. The process will be specifically explained based on the check example shown in FIG. 2 and according to the flowchart in FIG. 1.

First, blocks are extracted from the character information of the check shown in FIG. 2 (step Sl). Blocks are extracted for consecutive dots of 4 or more, and blocks of 4 or less are not adopted as they do not constitute character information. If the number of consecutive dots is 4 or more, the average height is EFI, and the height is > FEIXo, and the average height of 9 or more is EF2, the height is EF2X1.
A block larger than /3 or a block whose number of dots is larger than EF2X1/2 is extracted. If blocks are extracted from the character information in FIG. 2 as described above, seven blocks will be extracted as shown in FIG. 3, and numbers BLI to BL7 will be assigned according to the position from the left side. Note that block extraction may be performed by a method such as that disclosed in Japanese Patent Laid-Open No. 1-233585.

When the extraction of blocks is completed, the process moves on to the calculation operation of the confidence that a bar is included in each block. First, each of the obtained blocks BLI to BL7 is divided into 3× blocks as shown in FIG.
3 into nine matrix areas (step S2
). For example, for block 8L4, 9 areas Z fi, jl as shown in FIG. 4 are obtained, and the number of vertical masks Z for each area Z N, Jl as shown in FIG.
V (month i, number of diagonal masks ZS (1, J), number of horizontal masks ZH++, Jl are determined (step S3).

Here, the reason why the block is divided into nine areas of 3×3 is to determine the presence of a bar and the type of bar, as will be described later. That is, based on the probability information of the position of the bar within the block, the bar is located at, for example, the lower part of the block (Z (0, 21).

2 (1,21+Z+2.21), it can be determined as an underbar, and the upper area within the block (Zoo, o, +L1.0+ +Z+2.o
+), it is determined to be the upper bar, and the middle 1st rear in the block (Zoo, n +Zf1.11
, +2. n) There is an area that can be identified as Tireba middle bar and that cuts off the inside of the block diagonally (Z lo,
o,,IN, +1.21.21 ) or (1+2
．． .

Z ++, ++, Z lo, 2)) If there is a tail, then f, J force bar. It can be used to distinguish between diagonal underbars and diagonal upper bars, and can be used to determine the types of bars in blocks such as underbars, upper bars, diagonal mebars, diagonal underbars, and diagonal upper bars.

Next, I will explain the mask. The mask consists of nine dots in a 3 x 3 pattern, and the mask pattern of the dot mask that will constitute the vertical component of some dots of the character information is used as a vertical mask, and the mask pattern that will constitute the horizontal component is used as a vertical mask. A mask pattern that would constitute a horizontal mask and a diagonal component is used as a diagonal mask.

Here, there are seven types of vertical masks as shown in FIGS. 6(A) to (G), two types of diagonal masks as shown in FIGS. 7(A) and (B), and horizontal masks as shown in FIGS. Figure (8) ~fG)
There are seven types shown below. Note that the above mask may be created in a 3×3 matrix or in a matrix of 3×3 or more.

Then, as described above, each vertical mask is divided into each area Z f
By scanning within each area Z fi, Jl, the number of vertical masks in the block 2V(i, jl is determined, and by scanning each diagonal mask within each area Z fi, Jl,
The number of horizontal masks ZH(i,
Find jl. Then, the reliability including the bar is calculated for each block (step S4), and the reliability is calculated as follows.

The confidence level α of the underbar is determined after the next underline is extracted. That is, first ud-hlin
e[o]=ZH(0,2)+2H(1,2)+ZH(2
,2)-(ZS(0,2)+25(1,2)+ZS(2
,2))/2(ZV(0,2)+ZV(1,2)+ZV
(2,2))/2...(1) ud-hline[1]-ZH(0,2)+Z)1(1
,2)+48(2,1)]-(+2 (0,2) +z
s (1, 2) + zs (2, l) )/2-(zv
(o, 2)+zv(1,2)+zv(2,1))/2・
...(2) ud-1ine [2] -ZH(0,1) +ZH(
1,2) +28 (2,2)-(ZS(0,1)+Z
S(1,2)+ZS(2,2) )/2-(ZV(0
,1)+ZV(1,2)+ZV(2,2) )/2・
...(3), and the extracted data of the underline obtained in this way ud-hline[0], ud-hlin
Using e[1]ud-hline[2], the confidence level α of the underbar is determined according to the following formula.

α-hi MAX [ud-hline[o], ud
-hline[1]ud-hline[2]] /(1
total number of black pixels in the block)
・・・・・・・・・(4) Then, 0≦α≦1, is a constant, and the upper bar [(g degree β is the extraction of the upper line ud-hl ine [0] 4H (0,0) +ZH
(1,0) +Z)l (2,0)-(ZS(0,0)
+ZS(1,0)+ZS(2,0) )/2(ZV(
0,0)+ZV(1,0)+ZV(2,0) )/2
......(5) ud-hline [1] -ZH (0, 1) +ZH
(1,0) +ZH(2,0)-(ZS(0,1)+Z
S(1,0)+ZS(2,0) )/2-(ZV(0
,1)+ZV(1,0)+ZV(2,0) )/2・
・−・・・(6) ud-hl ine [2] −ZH(0, O) +Z
H(1,O) +HI(2,1)(ZS(0,0)+Z
S(1,0)+ZS(2,1))/2-(zv(o,o
)+zv(1,o)+zv(z,1))/2...
・After calculating in (7), calculate according to the formula below.

β-to 2・MAX [ud-hline[0], ud
-hline[1]ud-hline[2]] /(1
total number of black pixels in the block)
・・・・・・・・・(8) Add, 0≦β≦1, 2 is a constant, and the confidence level γ of the middle bar is ud hline [0] −48(0 1) +ZH(1 1 ) ÷ ZH (2, 1), then y = 3・ud−hl ine [0] / (number of horizontal pixels in one block) ・・・・・・・・・(9) However, If 0≦γ≦1, 3 is determined by a constant. Confidence level δ3 for the rough and diagonal bars. Confidence level of diagonal underbar δ2. Confidence of diagonal upper bar δ3
is calculated as follows. First, the confidence level δ1 of the diagonal bar is 5-bar[0] 1-3-ZV(0,0)-1-ZV(1,0)+1-Z
V (2,0)1-ZV(0,1)+1・ZV(1,
1)-1-ZV (2,1)+1・ZV(0,2)-
1-ZV (1,2)-3・ZV (2,2)3・ZS
(0,0)-145(1,0)◆1-ZS (2,0
)-1-ZS(0,1)+1-ZS(1,1)-1・Z
S (2,11+1-ZS(0,2)-1-ZS(1
,2)-3・ZS (2,2)3-ZH(0,0)-
1-ZH(1,0)+1-ZH(2,0)-1・ZH(
0,1)+1-ZH(1,1)-1-ZH(2,1)+
1-ZH(0,2)-1-ZH(1,2)-:l・ZH
After determining (2, 2)...(lO), the confidence level δ1 is calculated according to the following formula.

4-s-bar [01 / (number of pixels in width of one block) for δ1- (11) where 0≦δ1≦1, 4 is a constant, and confidence level δ2 of diagonal under par is ud-bar [0] 1+o・zv (o, o)+o・zv (1,0)+1
-ZV (2,0)40・ZV(0,1)+1-ZV(
1,1)-1-ZV(2,1)+1-ZV(0,2)-
1-ZV(1,2)-3・ZV(2,2)10・ZS(
0,0)40・2S(1,0)+1-25(2,0)+
o-zs (0,1)+1-ZS (1,1)-1-Z
S (2,1)41-ZS(0,2)-1-ZS(1,
2)-345(2,2)+0-Z)l (0,0) 10・ZH(1,0)+1-Zll (2,0)40・ZH
(0,1)+1・ZH(1,1)−1・ZH(2,1)
+1・zH(o,2)−1−zH(1,2)−3・ZH
After determining (2, 2) (12), the confidence level δ2 is calculated according to the following formula.

δ2= 5・5−ud bar[o] / (number of pixels in the width of one block) ・・・・・・(13) Where, 0≦62≦1, 5 is a constant, and the diagonal upper par The confidence level δ3 is up-bar [0] -3-ZV (0,0)-1-ZV (1,0) +1
-ZV (2,0)14V(0,1)+1-ZV(1,
1) 40・ZV (2, 1) + 1・ZV (0, 2) 10・ZV (1, 2) 40・ZV (2, 2) 3・ZS (
0,0)-145(1,0)+1・ZS(2,0)1-
ZS(0,1)+1-ZS(+,1)+0-ZS(2,
1) +1-ZS(0,2)40・ZS(1,2)10・
ZS(2,2)-3-IH(0,0)-1-ZH(1,
0)+1-Zl((2,0)-1-2)1 (0,1)
+1・ZH(1,1)+O・ZH(2,1)+148
(0,2)+0-ZH(1,2) 40・ZH(2,2
)...(14) After calculating, the confidence level δ3 is calculated according to the following formula.

δ3-to 6-5-up-bar[o]]/ (+Number of horizontal black pixels in the block) ・・・・・・(
15) Then, O≦δ3≦1, and 6 is a constant. It is determined whether the calculation of confidence including the bar as described above has been calculated for all blocks. Step S5), calculation for all blocks Repeat the above operation until finished.

Then, confidence levels αβ, γ, δ3 for all blocks
．． δ2. When δ3 is determined, the block with the highest reliability is selected (step 510), and the end points within the block are extracted (step 511). However, only two end points whose horizontal distance is equal to or greater than 1/2 of the width of the block are valid, and extremely short end points are excluded. Regarding block BL4 in FIG. 4, as shown in FIG. 9, for example, points a (xi, yl) and b (
x2. y2) , a and c, a and d, a and e 4
The endpoints of the set are extracted, and the effective path length Pk(k−
1 to n, where n is the number of combinations between two end points) (step 512). The effective path length P is the distance (number of dots) connected by the line between the two end points a and 5. When there are multiple buses between endpoints, the shortest path length is taken as the effective path length. For example, for the block in Figure 12, for the end points a and b, Figure 13FA), (
There are two buses shown in B), and for end points a and C, there are two buses shown in (C) and (D) in the same figure, but the shorter one is Pass Me Elow B ((A) in the same figure). and a-Eber C ((C) in the same figure) are adopted. Then, the straight path ll1ID between each end point is Db-X2-XI'' (y
2-yl)' =-- (16) (step 513), and linearity Sk for each bus is 5k-Dk/P.

・・・・・・・・・(17) Find it with (Step 514).

Repeat the above operation until such linearity S is determined for all buses (step 515). Once linearity S is determined for all buses, the bus with the highest linearity among the buses is assumed to be the cent bar. (Step 520).

Then, it is determined whether or not there is a branch shared with the number in the hypothesized cent bar (step 5211). If there is a shared branch, the hypothesized cent bar is removed, excluding the interpolated branch (step 523). This is to deal with cases where the bar and numbers overlap and touch each other. If there is no branch shared with the number, the hypothesized bar is removed from the block (step 522).6 Figure 10 (B) shows the interpolation within the cent bar hypothesized for the block in Figure 10 (A). An example is shown in which items other than branches are removed, and (C) in the same figure shows an example in which a hypothesized cent bar is removed.

Then, the numerical part is segmented by the method described in JP-A-1-121988 (step 524).
, the number is recognized and verified (step 525), and it is determined whether the recognition and verification are OK (step 526). Then, if the recognition verification is not 0, remove the hypothetical path from the hypothesis candidates (step 527), step 5 above.
Return to 20. If the number of times the hypothesis candidate is removed is the second time or more, an error process is performed (step 528).

Although the above-described embodiments have been explained in terms of dollars, the same applies to other currencies such as yen and bonds. Also,
This method can also be used to extract reference lines as shown in FIGS. 14(A) and 15(A).

Effects of the Invention As described above, according to the character extraction method of the present invention, end points are detected from character information including bars, and bars are detected based on the ratio of the shortest distance and effective distance between each end point. Since characters are recognized after being removed from the block, reliable character recognition is possible even for characters that touch the bar. According to this invention, it is also possible to recognize characters as shown in FIGS. 11(A) to 11(F).

[Brief explanation of the drawing]

FIG. 1 is a flowchart showing an example of the operation of this invention, and FIG.
Figures 3 and 4 show examples of handwritten characters.
The figures are diagrams for explaining blocking processing, Figure 5 is a diagram showing divided areas, Figures 6 (A) to (G) are diagrams showing examples of vertical masks, and Figures 7 (A) and (B) is a diagram showing an example of a diagonal mask, Figures 8 (FA) to (G) are diagrams showing an example of a horizontal mask, and Figure 9 is a diagram showing an example of an effective bus between end points, the shortest distance, and linearity. Figure 10 is a diagram for explaining a processing example based on a hypothesis, Figures 11 (A) to (F) are whistle diagrams 12 and 13 (A) to (D) showing examples of handwritten characters.
) is a diagram for explaining the path length of end points, Figure 14 (A)
, (B) and FIGS. 15(A) and (B) are diagrams for explaining conventional character extraction. BLI ~BL7-...Block, a, b, c
, d, e...One end point.

Claims (1)

[Claims] 1. A character extraction method in which a bar is detected from character information including a bar, and characters are extracted by separating the bar, which detects end points from the character information and calculates the linear distance between each end point. detect the bar, calculate the number of dots connecting each of the end points, detect the bar based on the ratio of the shortest distance between the end points and the effective distance, and separate the detected bar to extract the character. A character extraction method characterized by the following. 2. When recognizing characters in character information including bars, in a character extraction method that detects bars from the character information and extracts only the characters, a block is formed by treating one continuous character group from the character information as one block. The extracted blocks are divided into multiple areas Z_(_i_, _j_)(
i = 0 to m, j = 0 to n), and each area Z_(_i_,
Find the number of existing areas for each area Z_(_i_, _
The confidence that the block contains a bar is calculated for each block based on the number of masks in j__), the endpoints in the block that have a high certainty of including the bar are found, and the shortest distance between the detected endpoints is calculated. The distance and the effective distance, which is the number of dots of the line segment between the end points, are determined, and the bar is detected based on the ratio of the determined shortest distance between each end point and the effective distance, and the character is extracted. A character extraction method characterized by