JPS603680B2 - pattern recognition device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a pattern recognition device using a pattern matching method for recognizing patterns of characters and figures, etc., which particularly discriminates between similar categories when recognizing patterns of different sizes that belong to the same category. It is something.

A conventional pattern recognition device is shown in FIG.

In Figure 1, 1 is a form with characters and figures written on it, 2 is a photoelectric conversion section, 3 is a quantization section, 4 is a preprocessing section, 5 is a cutting section, 6 is a cut out input pattern, and 7 is a difference. Degree calculation evil part,
8 is the judgment wind section, 9 is the dissimilarity calculation 'B} section, and 10 is the judgment [
Part B) is constructed such that 11 is a table of standard patterns prepared for each category, and 12 is a table of differential patterns prepared for each set of particularly similar categories. The characters written on the form 1 are photoelectrically converted in an analog manner by a photoelectric converter 2, and converted into black and white binary values by a doji converter 3.

The black/white binary pattern is subjected to noise removal, smoothing, etc. using a filter, for example, in a preprocessing section, and is separated into characters by a cutting section 5 and stored in a pattern register 6. Next, the pattern register 6 and each standard pattern in the standard pattern table 11 and the difference calculation unit 7 calculate the difference. The operation of the dissimilarity calculation unit 7 will be explained in more detail. Standard pattern MM
An example is shown in Figure 2.

Figure 2 a and b are standard patterns MM (o) for the ○ category.
), and Figure 2 a is its minimum line standard pattern MM paper)
, b is up to its maximum line standard pattern MM fin. Similarly, Figure 2 e and d show MM cotton) and even MM pictures.
Black bit (bit 1) of minimum line standard pattern MM Minoru
The white bit (bit 0) of the input pattern KK corresponding to the part
) part bit number and maximum line standard pattern MM(o)
Input pattern K corresponding to the white bit (bit 0) part of
The sum of the number of bits of the black bit (bit 0) part of K becomes the dissimilarity R(o). R(o)=2KK [MNA breasts 12K
Kn bribery {11 Next, in the judgment A part 8, the difference between the smallest difference R (CI) and the next smallest difference R (C2) for each category is calculated.

*Compare the difference with the preset SLI. R
(C2)-R(CI)2SLI■ ■When the formula is satisfied, the input pattern is in the category CI
, and the recognition process is terminated.

■If the formula is not satisfied, it is considered undetermined and the next difference RS is made. ，(
C2) = 2(KK[MMSc, -c2) +Z(KKnR
Sc, (c2) and RSc2 (c,) are calculated, and the process proceeds to the next determination 'B) section 10. The conditions for determination in the determination {B} unit 10 are shown in the following equation.

SL2 is a preset value (SL2<0.5). RSc, (c2) MiSL1 [5) RSC, (
C2) 10 RSC2 (C,) RSc2 (cl) Mi SL1 [61 RSC2 (C,) 10 RSC, (C2) (If formula 5' is satisfied, it belongs to category CI *degree calculation'B) Part 9 Difference pattern MMS (CI, C2) between categories CI and C2 in the difference pattern table 12
Focus on the differences using .

Difference patterns are prepared in advance, but they can be created using the following method. Differential pattern MMS (CI, C
2) is the differential pattern MMSc,- for CI to C2.

2 and the difference pattern MMSc2-c for CI of C2
, consists of.

¥;Declaration...=...3≧Cirrus 3=Pan Snow;) (3) As an example, the differential pattern MMS(o, G) is shown in FIGS. 3a and 3b.

In the input pattern KK, the differential dissimilarity RSc,(c2) of the category CI with respect to the category C2 is expressed by the following equation. MSC2-C, )'4} If the expression '6' is satisfied, it is determined that it belongs to the category C2, and if neither the {5} expression nor the {61 expression is satisfied, it is determined that the determination is impossible.

In this manner, in the conventional method, measures are taken to reduce the influence of fluctuations and noise in the line on the degree of dissimilarity, and to improve the reading rate by making the determination in two stages.

However, when characters of the same type but different sizes are input, the difference R and the difference difference RS are greatly affected. For example, consider the case of the input pattern shown in FIG. R(o)=Z(KKnMNA mother 12(KK[M
M Fukitaka) = 36 + 22 = 58R (G) = 2 (KK [M
NA registration +2 (KK [MM Tsuchima = 33 + 25 = 58R (G
)-R(〇)=0 12 Therefore, in the determination section 8, the result is undetermined.

Next, the SL calculates the difference degree using the difference pattern.
I=12). RSC, (C2)=Z(KK[MMSc,-.2)12(KK[MMSc2-c,)=32=5RSC2(
C,)=ヱ(KK[MMSC2-.

, ) 12 (KK[MMSC.-C2) = 23:5RS
c, (c2) =0.5>0.4RSC2(C,)+
RSC, (C2), and it is still undetermined.

(SL2=0.4) As described above, the subordinate method has the drawback that it is difficult to distinguish between characters of different sizes. In order to solve these drawbacks, the purpose of the present invention is to normalize standard patterns according to input patterns, use them to create differential patterns, and use them to effectively discriminate characters of different sizes. It was made to look like Z.
This will be explained in detail below. FIGS. 6, 7, and 8 show the first embodiment. In FIG. 6, 13 is a form, 14 is a photoelectric conversion unit, 15 is a quantization unit, 16 is a preprocessing unit, 17 is an extraction unit, 18 is an input pattern register, 19 is a Z standard pattern table, and 20' is a difference degree. Calculation part, 21 is normalization pattern, 22 is registration pattern table, 23 is registration part, 24 is judgment part, 25 is difference pattern, 26 is difference creation part, 27 is difference calculation [B} part, 28 is Judgment [This is part B. In Figure 7, 19-1 is the maximum 4-line standard pattern, 19-2 is the maximum line standard pattern, 20-1
is address translation table, 20-2 is address translation, 2
0-3-1, 20-3-2, 20-3-3, 20-3-
4 is a difference calculation unit for each area, 20-4-1, 20-4
-2, 20-4-3, 20-4-4 is memory for difference address conversion constants for each area, 20-5-1, 20-5-
2, 20-5-3, 20-5-4 are comparison sections for each area,
20-6 is a summation unit, 20-7 is an address conversion unit, 21-1 is a difference memory, 21-2 is a minimum line normalization pattern, and 21-3 is a maximum line normalization pattern. In FIG. 8, 22-1 is a first registered pattern difference memory;
22-2 and 22-3 are first registered patterns, 22-4 are difference memory for second registered patterns, 22-5 and 22-6 are second registered patterns, and 26-1 and 26-2 are difference creation units. ,
25-1 is a difference pattern between the first registered pattern and the second registered pattern, 25-2 is a difference pattern between the second registered pattern and the first registered pattern, 27-1, 27-2
is a dissimilarity calculation unit. Characters written on the form 13 are processed by a photoelectric conversion unit 14, a quantization unit 15,
A black/white binary pattern separated into characters after passing through a preprocessing section 16 and a cutting section 17 enters an input pattern register 18. The pattern register 18 (standard pattern, normalization pattern, registered pattern, and differential pattern are also the same) is the fifth
As shown in the figure, it consists of 16 bits in the horizontal direction (X direction) and 24 bits in the vertical direction (Y direction), and is addressed as shown in the figure. The address (×, Y) of point P is (10
, 13). Furthermore, the pattern register is divided into four parts, top, bottom, left and right, and these are called area A, area B, area C, and area D, respectively. Enter the address of the input pattern (XK
, YK), and the address of the standard pattern is (×M, Yw).

The address conversion unit 20-2 uses the address conversion table 20
The conversion shown by the formula '7' is performed using the conversion constants Δ× and AY. (XN', YN') is the address after conversion. ×N'=×M+Δ× YM'=YM+ΔY [71 This corresponds to shifting the standard pattern by Δ× and ΔY.

Using the address-converted standard pattern, the input pattern 15 and the dissimilarity calculation unit 20-3 calculate the dissimilarity shown in equation (1) for each area, and store them in the memory 20-4. In this embodiment, the address translation table 20-1
Dissimilarity calculations were performed nine times for the same standard pattern using the nine types of conversion constants △× and △Y shown in
Store in 14. No. The degree of difference for each area by address conversion using a conversion constant of n is expressed as Rn^, RnB, Rnc
, Rno. Memory 20-4-1 has Ro^,
R, ^, ..., R8^, ROB in memory 20-4-2
, R, B, ..., R88, Ro in memory 20-4-3
c, R, c, ..., R8c, memory 20-4-4 has R
oo, R, o, ..., R8. is stored. Next, the comparison unit 20-5 compares the magnitude of the degree of difference in the memory 20-4 for each area, and selects the smallest degree of difference as R^ and R, respectively.
Let them be B, Rc, and Ro.

The summation unit 20-6 calculates the '8' formula to determine the dissimilarity R and stores it in the dissimilarity memory 21-1. RiR＾1R
80Rc+Ro■ Also, the address conversion unit 20-7 converts the standard patterns 19-1 and 19-2 into addresses independently for each area using the conversion constant when the degree of difference is the minimum in the comparison unit 20-5. death,
The normalized patterns 21-2 and 21-3 are stored.

By these operations, standard patterns 19-1, 19-2
is normalized so that the degree of difference from the input pattern is minimized. The degree of difference is calculated for each category, and the minimum degree of difference, the next smallest degree of difference, and the respective normalized patterns are stored in the registered pattern table 22 by the registration unit 23.

Next, the determination section 24 compares the difference between the first registered dissimilarity R (CI) and the second registered dissimilarity R (C2) in the registered pattern table 23 with a preset SLI.

R(C2)-R(CI)ZSL1[9' When the expression {9} is satisfied, the recognition is terminated using the category CI as the determination result, but when the expression [9} is not satisfied, the process proceeds to the next step.

The difference creation unit 26-1 creates a difference pattern MM from the minimum line pattern (MM ship 1)) 22-3 of the first registered character and the maximum line pattern (MM supply line) 22-6 of the second registered character.
Sc,-. 2 according to formula [3', . It is stored in the differential pattern register 25-1.

Similarly, MMSc2-c is created according to the '31 formula and stored in the differential pattern register 25-2. Using the difference patterns 25 created from these normalized registered patterns 22, the difference calculation unit 2
7-1, 27-2, differential dissimilarity RSc, (c2), RS
Calculate c2(c,) (see equation '4).

Next, in the determination (B) section 28, using the same method as the subordinate method,
If formula '5} is satisfied, category CI is put on hold,
If formula '6' is satisfied, category C2 is determined as the result,
If both equation {5) and equation (6) are not satisfied, the determination cannot be made.
As explained above, in the first embodiment, the standard pattern is normalized according to the input pattern, the degree of dissimilarity is calculated, and the difference pattern is created each time from the normalized standard pattern. Even if the sizes are different, the difference pattern will be suitable. This will be explained below using an example. Consider the case of an input pattern as shown in FIG. Table 1 shows the degree of difference between the standard pattern of category ○ and category G shown in FIG. 2 for each conversion constant and area. The △ mark in Table 1 indicates the minimum Aokika in each area.

(In the case of the same degree of dissimilarity, give priority to the /1 size of the conversion constant.) R(o) = 610111 = 8 * * Rio) = 6 (No. 1) R.

)=. KuN. -1) R.

)=・(N.-5)R too.

):・kuN. -5>R(G)=610193=18 R each G)=6(N.

-1)R+G)=.

(N.-1) ** R + G) = 9
(No. 5) R also G): 3 (N.

-5) R(G)-R(0)=10<SLI In this case, the normalized pattern of the standard pattern of ○ and G in FIG. 2 is shown in FIG.

Due to the difference between R(o) and R(G), the judgment leg is still undetermined as 0.
It is.

Next, when a difference pattern is created using the normalized pattern shown in FIG. 9, it becomes as shown in FIG. 10.

R.S.

(G)=2(KK[MMS.-G)12(KKnMMS
c-o)=010=ORSG(.)=Z(KK[MMS
G-. ) 12 (KKnMMS.-G) = 4 16 = 10R
S. (G) - 0RS.

(G) + RSG (〇) Pile Kario = 0 < 0.4 Therefore, this input pattern is correctly determined to be 0 and is 2. As can be seen by comparing the difference pattern in FIG. 3 and the difference pattern in FIG. 10, it can be seen that the difference pattern in FIG. 10 is a pattern that matches the size of the input pattern.

In this way, it is possible to obtain an optimal difference pattern for two input patterns of different sizes, and to make a correct determination.
In this embodiment, the pattern register is divided into four parts, but the number of divisions may be increased or decreased depending on the nature of the characters to be input.

For example, if the input pattern may have different heights, but it is expected that the inside will hardly change, a predetermined effect can be obtained simply by dividing the input pattern into upper and lower halves.

Further, the method of creating the difference pattern is not limited to the '3' formula, and the difference degree is not limited to the [4] formula.

Further, in the determination (B) section, the determination may be made taking into consideration not only the difference degree of difference but also the degree of difference in the degree of difference calculation section.

Further, the number of types of conversion constants in the address conversion table in the first embodiment is not limited to nine types, but may be decreased or increased according to the nature of the input pattern.

Further, target categories may be prepared in advance before implementing the present invention for all categories.

The method may use the dissimilarity degree in the dissimilarity calculation part (2) in the dependent example, or a known method may be used. By doing so, it is possible to shorten the determination time, especially in the case of the first embodiment. The present invention creates a standard pattern and a differential pattern suitable for the size of the input pattern,
Since this is used to make the determination, there is an advantage that variations in font size can be absorbed.

Because dot printers make up characters using several dots, the overall character height is more likely to change than with regular printers. Therefore, the present invention can be used to recognize dot characters.

[Brief explanation of the drawing]

FIG. 1 is a block diagram of a conventional pattern recognition device, and FIGS. 2 and 3 are explanatory diagrams thereof. 6, 7, and 8 are block diagrams of the first embodiment of the present invention, and FIGS. 4, 5, 9, and 10 are explanatory diagrams thereof. 1... Form, 2... Photoelectric conversion section, 3... Quantization section, 4... Pre-processing section,
5... Cutting section, 6... Input pattern register, 7... Difference calculation leg, 8...
...Judgment ■ part, 9...Difference calculation {B} part, 1
0... Judgment leg, 11... Standard pattern table, 12... Differential pattern table,
13... Ledger, 14... Photoelectric conversion section,
15...Quantization unit, 16...Separate processing unit, 17...Cutout unit, 18...Input pattern register, 19...Standard pattern table,
19-1・・・Minimum line medium standard pattern, 19-2
・・・・・・－Maximum line medium standard pattern mountain, 20・・・・・・
Difference calculation section, 20-1... Address translation table, 20-2... Address translation section, 20-
3-1, 20-3-2, 20-3-3, 20-3-4・
...Dissimilarity Calculation Department, 20-4-1, 20-4-2, 2
0-4-3, 20-4-4...Memory, 20-5-1,
20-5-2, 20-5-3, 20-5-4...
・Comparison section, 20-16... Total section, 20-17...
... Address conversion section, 20-8... Character frame calculation section, 21... Normalization pattern section, 21-1...
...Difference memory, 21-2...Minimum line positive parent pattern, 21-3...Maximum line positive parent pattern, 22...Registered pattern section, 22-1 ...First
Registered character difference memory, 22-2...First registered character minimum line pattern, 22-3...First registered character maximum line pattern, 22-4...Second registered character difference degree memory, 22-5... Second registered character maximum 4 in-line pattern, 22-6... Second registered character maximum in-line pattern, 23... Registration section , 24... Judgment section, 25... Difference pattern section, 25-1,
25-2...Difference pattern, 26, 26-1,
26-2... Difference creation section, 27... Dissimilarity calculation [B} section, 27-1, 27-2... Dissimilarity calculation section, 28...・Judgment 'B} part. Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Figure 8 Figure 10 Figure 9

Claims (1)

[Claims] 1 Calculate the degree of difference between the input pattern and each standard pattern whose address has been converted according to a plurality of address conversion constants for each area, select the minimum degree of difference for each area, and calculate the sum of the differences for each standard pattern. The standard pattern is stored once as the degree of pattern difference, and a normalized pattern is created by converting the address of the standard pattern for each area using an address conversion constant that gives the minimum degree of difference for each area. 1 means 18, 19, 20, 21, and the degree of difference between the minimum degree of difference and the category giving the next smallest degree of difference by calculating the degree of difference of each category based on the degree of difference of the standard pattern; and a normalized pattern; and second means 22, 23 for storing a normalization pattern corresponding to the minimum line width of each one of the categories when the degree of dissimilarity between the two categories does not satisfy a predetermined format condition. Create difference patterns between the normalized pattern and the normalized pattern corresponding to the maximum line width of each of the other categories, and calculate the degree of difference between each of the difference patterns and the input pattern as a difference degree. Third means 25
, 26, and 27, and determines one of the two categories as the category of the input pattern when the two differential dissimilarities satisfy a predetermined format condition. .