JPH0221387A - Word reader - Google Patents

Abstract

PURPOSE:To determine a word with high accuracy even when the noise, etc., of a non-character exist on a document by performing an operation by integrating the rate of the likelihood of the noise into various parameters and selecting words having high coincidence degree under a ranked state by a word determining means. CONSTITUTION:When the noise, etc., of the non-character exist in the picture data of a character, etc., a pseudo noise type character detecting means 6 detects the rate of the likelihood of the noise. The rate of the likelihood of the noise is additionally integrated into conventional various parameters as a parameter, the operation is performed using a prescribed operation expression, a word determining means 59 selects with which word out of words 12 and 13 recognition candidate character strings 10 and 11 can respectively have the highest coincidence degree, and it is determined that the words 12 and 13 having high ranks are the character strings inputted from a document 1. Thus, even when the noise, etc., exist on the entered character frame of the document, the words can be accurately collated with each other.

Description

[Detailed description of the invention]

[Industrial Application Field] This invention relates to a word reading device that reads and recognizes words such as addresses and names, and in particular, recognizes each character that makes up a word and uses the recognition results to modify the word. This invention relates to a word reading device. (Prior art) In recent years, word reading devices have been attracting attention as a means of high-speed, large-volume data entry into computers. It is desired to develop a word reading device that can read words with high accuracy. As this word reading device, an optical character reading device (OCR) is known. )) etc. In the figure, 1 is a form on which characters are written, 2 is a scanning means for photoelectrically converting and reading the form 1, and 3 is a scanning means for reading input characters.
Character recognition means that cuts out and recognizes each character and outputs recognition candidate characters; 4 is a word dictionary that stores reference words; 5 is recognition candidate characters output from character recognition means 3 and words in word dictionary 4; This is a word determining means that determines a word by comparing the words. The conventional word reading device is configured as described above, and outputs words that are ranked higher among the recognition candidate characters and are registered in the word dictionary 4 and are preferentially selected as the word reading results. be. FIG. 3 is an explanatory diagram showing an example of a general form 1 to be read. In the figure, 7 is the input word string "Okayama", 8 is the input word string "Okaya", and 9 is noise (non-character) attached to the form. Examples of recognition candidate characters output from the character recognition means 3 for the characters on this form are shown in FIGS. 4 and 5. FIG. 4 shows the recognition candidate characters 10 output by the character recognition means 3 for each character of the input word character string "Okayama" 7, and for the first character "O" of the input word. ,“
One recognition candidate character for ``o'' is for the second character ``chi,'' one recognition candidate character for ``power'' is for the third character ``ya.''``Ma'' 5 There are two recognition candidate characters for ``ya,'' and for the fourth character ``ma,'' there is one recognition candidate character for ``nu.'' FIG. 5 shows the recognition candidate characters 11 output by the character recognition means 3 for each character of the input word character string "OKAYA" 8 and the noise 9 attached on the form. For the first character "o", one recognition candidate character is "o", and for the second character "power", one recognition candidate character is "power".
For the third character “Ya”, there are two recognition candidate characters, “Ma” and “Ya”, and for Noise 9 attached on the form, there is one recognition candidate character, “No°”. Figure 6 is an explanatory diagram showing an example of words in a general word dictionary 4 regarding katakana surnames in a word reading device.In the figure, 12 is the word "Okaya", 13 is the word "Okayama" 7 is an explanatory diagram of a word determination method using the idea of general dynamic programming used in the word determination means 5 of the word reading device.The character string sl constituting the word
, s□...S, and jl+ t2...1. When performing matching with substring Sl+S! ...S3 and t
, ', t, ... tj are best matched and the amount of deviation is f (i+jl) In the figure, 14 is ' (i-1+j-11+ 15 is f <=-+, j1
．． 16 is 'N,=-t), 17 is f(=,=>, and 18 is f (i1+ j-11 and f (i+'j>
The distance d,,19 between (i-1+j) and f
(Lj) and the distance dg, 20 is f(i+j-
1) The distance d between and f (i+j). Figure 8 shows the input word character string “Okayama 7” and the word dictionary 4.
In Figure 0, which is an explanatory diagram illustrating the matching with the word "Okayama" 12 in , 21 is the input word string "Okayama" 7
The first letter of and the word “okaya” in word dictionary 4
The distance 22 from ``O'' in 2 is the distance between the character power of the second character and the "power" of the word in the word dictionary 4, and 23 is the distance between the character ya of the third character and the "power" of the word in the word dictionary 4. ”, 24 is the distance for character insertion errors, and 25 is the path that gives the shortest distance. Figure 9 shows the input word string “Okayama” 7 and the word dictionary 4.
FIG. 3 is an explanatory diagram illustrating a comparison with the word “Okayama” 13 in the above. In the figure, 26 is the input word string “Okayama”
The first letter of 7 and the word “Okayama 3” in the word dictionary
13 is the distance from “o”, 27 is the distance between the second character power and the word “power” in the word dictionary 4, and 28 is the distance between the third character ya and the word in the word dictionary 4. Distance from “Ya”, 29
is the distance between the fourth character ``ma'' and the word ``ma'' in the word dictionary, and 30 is the route that provides the shortest distance. Figure 10 shows input word character examples “okaya” 8 and noise 9.
FIG. 4 is an explanatory diagram illustrating a comparison between the word "OKAYA" 12 in the word dictionary 4 and the word "OKAYA" 12 in the word dictionary 4. In the figure, 31 is the input word string "
Okaya "8's first letter O and the word in the word dictionary"
32 is the distance between the character power of the second character of the input word string and the "power" of the word in the word dictionary 4, 33 is the distance of the third character of the human word The distance between the character ya and "ya" in the word dictionary 4, 24, is the distance for a single character insertion error as in FIG. 8, and 34 is the path that gives the shortest distance. Column “Okayama” 8, noise attached to the form 9, and word “Okayama” 13 in the word dictionary 4
FIG. In the figure, 35 is the distance between the first character O of the input word string "Okaya" 8 and the "O" of the word "Okayama" 13 in the word dictionary 4, 3
6 is the power of the second character and the “power” of the word in word dictionary 4
37 is the distance between the third letter strength and the word "ya" in the word dictionary, and 38 is the distance between the noise 9 attached to the form and the word "ma" in the word dictionary 4. 39 is the route that provides the shortest distance. Next, the operation of the conventional device will be explained. The word determining means 5 compares the input word character string entered in the form shown in FIG. 3 with the words in the word dictionary 4 shown in FIG. 6, and the ranking of the probability of word identity is determined by dynamic programming. This is done using the following method using the following concept. <Word determination method> Input word string? +83""+, 3m...S,
and the word T in the word dictionary 4 = t1. When performing matching with tt...t, we introduce the concept of the distance between the characters Si and t, and set this as d to create a 0 substring 5InS! ...
The amount of deviation f(i+j) when 3i and "I+"t...tJ are most closely matched is introduced, and this is calculated using the following recurrence formula. f(O・0)=O f(i+J) ”ml n (f 1i-1+j)
"dt *'(i+J-11" dl + f
(i-1+j-1) ” dl ) dl: Letter S
If there is 11 in the recognition candidate characters for i, its rank is 1, if there is no recognition candidate character, P (constant) d! : Constant d, : Constant min (X, Y, Z) : Minimum value among X, Y, Z, that is, as shown in Figure 7, f <=-+
,J,15 to ft=,J+17 is st,si
...3 i-1 and 1. ．． 1. ...If the matching up to tJ has been completed, and the degree of deviation is In this case, the letter S is considered to be an extra character, and therefore a penalty d219 is given, ' (i+jl = f (i-1'
+j+ "dt. f < r, J-+) The route from 16 to f t =, i, 17 is the same, and 1. is treated as an extra character and is not compared with the character in the input word string S. (allowing for a one-letter omission error), and instead giving a penalty of ds20 r Ti+ jl −
Let f TLj-11+d3. The path of f (i-++j-H14 to f mountain,) is the letter S
, and the character 1J, the degree of this matching is d
, and so on. d takes the following values. Input the word character string sl, s, ...S, and the word tI+ in the word dictionary using f (1&+R) obtained by this recurrence formula.
As the amount of deviation from t2...t7, the amount of deviation is calculated for all words in the word dictionary, sorted in descending order of the amount of deviation, and candidate words are output. Using the word determination method, predetermined d,=30,d
,=30. The case where P=20 will be explained. Fourth
The word determining means 5 compares the input word character string "OKAYAMA" 7 shown in the figure with the words in the word dictionary 4 shown in FIG. First, a comparison is made with the word "okaya" 12 in the word dictionary 4. As shown in Figure 8, the input word string S is “Okayama”.
7, and the word T in the word dictionary is "Okaya" 12. When finding the value of f(.+X) (X≧1), the recurrence formula is:
' (On X) ""' (0,X-11+d:
l, and f. n+1 ='dl =39f(
. , z> =2d3 =60 f(.,3) =3ds =90. Similarly f TIl+. , When finding the value of (X≧1), the recurrence formula is f (X,.,=f, yaw1.., +d2, and f(-・o> ” dz =30r,,
,. , =2dz =60f(3,., =3d
z = 90 f(4,.> ””4dt = 120. Next, S, (“o” of human power word 7 “Okayama”) and 1+
(If we compare the word 12 "O" in word dictionary 4), tl is in the first position of 81 recognition candidate characters 10, so d1=1(21) in Figure 8' ( 1, l) = “ ' ” (' (6+ll
+dt + '(iol+1+d,,f,.,
. , +1) =min (30+30.30+30.1)=1. Similarly, when sl and +2 are compared, t is found among the 10 recognition candidate characters of S. does not exist, so dl=20, and f (1,1
) = min (f (11,11+cL +
f11.1)+a,,r(.,i,+20)=min(60+30.1+30°30+20)=31. When comparing 3t and tl, S! 1 out of 10 recognition candidate characters. does not exist, so a, = 20, and' (211
=min(' (I11+ 10dl + '(2
,01" d 2 r' (1,., +20)
=min (1+30.60+30°30+20) =31. When comparing s2 and 1t, +2 is in the first position of the recognition candidate character lO of st, so d in Figure 8 becomes = 1 (22)' (1+ri ``”m'n(' (Inri ten d
l + '12. ll+d,・fu,n+1
) =min (31+30.31+30°1+1) =2. If we compare S and t, we get 31 i! Since +2 does not exist in the identification candidate character lO, d+=20, and ' 1.3
>"m'n(' (a+3) +dt +
' (1+21+d3+f(.,z)+20) = min (90+30. 31 +30°60+2
0) =61. When comparing S, and tl, since t does not exist in the recognition candidate characters lO of S, a+=20, 'II.
+l = m l n (f (t・
+1 + d, l f (310)
+d2+'l! +. , +20)=min (
31+30.90+30°60+20) =61. When comparing 3t and t, among the 32 recognition candidate characters lO

Since [, does not exist, d, = 20, and f (t, 3
)=min (f <+, s> +dz r
f <z, t++dz, f(t, i> +2
0)=min (61+30. 2+30°31+2
0) =32. When comparing S, with +2, since 1t does not exist in the 10 recognition candidate characters of S, a, = 20, and f + z, z +
=min (f(z, t+ +d,, r(3
,. +d2, f (!, 1) +20>=min
(2+30.61+30°31+20) =32. When comparing Syu and t, t is in the second position among the 10 recognition candidate characters of S, so d, = 2 (23), f(3,31-min(' 12・ri +dt +
'+3・21"dz r f tt, z> +
2)=min (32+30.32+30°2+2)=4. When comparing s4 and tl, t is found in recognition candidate character 0 of S4.
, does not exist, so it becomes dl-20, and f (411)
= min (f (:1ll) 1 dl 1
f (41111+d3 + ' +3+., +
20)=min (61+30.120+30°90+
20) =91. When s4 and t are compared, 1 out of 0 recognition candidate characters for S is found.
2 does not exist, so d,=2o,'(412>
””m l n (f 11+21 10 dl r
f(4+1)+ds* f(3, n +20) = min (32+30.91+30°61+20) =62. When we compare s4 and t, we find that t does not exist among the 10 recognition candidate characters for S! dl = 20, '+4
+31 = m1 n (f (3-3) +dt *
f(4+!>+dit f<s, t> +20
)=min (4+30.62+30°32+20) =34. That is, the amount of deviation f(4.3) between the input word character string "Okayama" 7 and the word "Okayama" 12 in the word dictionary is: r (4,31=1 (21 in Figure 8) + 1 ( 22 in Fig. 8) +2 (23 in Fig. 8) +30 (24 in Fig. 8) = 34, and when finding the path that gives the minimum deviation amount f(413>) by backtracking, we get (4 , 3). The route passes through the points (3, 3), (2, 2), (1, 1), (0, 0), and follows 25 routes. Next, the word " As shown in FIG. 9, the input word string S is “Okayama” 7, and the word T in the word dictionary is “Okayama” 13. S, (Input word string “ “O” in Okayama 07) and 1+ (“O°” in word “Okayama” 13 in the word dictionary)
), tl is in the first position of the recognition candidate character lO of the character S, so dl = 1 (26) in Figure 9 becomes f lI+ll ”m 1 n (' l@+1
> +dt * ' (++01+d3・r (
o・o)+1) =min (30+30.30+30.1)=1. Similarly, when comparing S2 and 1t, there is +2 in the first position of the 32 recognition candidate characters lO, so a, = 1 in Fig. 9
(27) Next' +2.11 "ml n (' (Inri +
dl l'<1+1+”dl 1 f (11
11+1)=mirl (31+30.31+30°1+1)=2. Similarly, when comparing S and t, t is in the second position among the recognition candidate characters lO of S, so dl = 2 in Figure 9.
(28) becomes f(2+3) ””m l n (' (t*21
"dl * '(3,ri+d3・f(!・z
＞”2) =min (32+30.32+30°2+2) +4.Similarly, S4 and +4 +4 d in 10 candidate characters, =20 (29) (,4-,=min = m1n matching) If we take , there is no recognition of 34, so we get 'Figure 9 (f+1.4)+d2・[(4・ko)+a, ,
f (3,31+20)(33+30.34+30°4+20) =24. In other words, the input word string "Okayama" 7 and the word dictionary 4
The amount of deviation from the word "Okayama" 13 in '(4, 4) is f, a, -+ = 1 (26 in Figure 9) + 1
(27 in FIG. 9) +2 (28 in FIG. 9) +20 (29 in FIG. 9) =24, and the shortest route is 30. As a result of the above, the amount of deviation between the input word string "Okayama" 7 and the word "Okara" 12 in the word dictionary is 34. The amount of deviation between the input word string "Okayama" 7 and the word "Okayama" 13 in the word dictionary is 24, and the candidate words are output in the order of "Okayama" 13 and "Okara" 12, which have the smallest amount of deviation, and are correct. In addition, the input word string written on the form in Figure 5 “
The noise 9 adhering to the noise 8 and the form is compared with the words in the word dictionary 4 shown in FIG. 6 by the word determining means. By using the word determination method, the word “Okara” in the word dictionary is
As shown in FIG.
l is f (4, 31= 1 (31 in Figure 10) +
1 (32 in FIG. 10) +2 (33 in FIG. 10) +30 (34 in FIG. 10) =34, and the shortest path is 34. Next, when the word “Okayama” 13 in the word dictionary 4 is compared, the human-powered word character string “Okara” is found as shown in FIG.
8 and noise 9, and the word “okayama” 1 in the word dictionary 4
The amount of deviation from 3 f (4+ 41 is f (4,.=1
(35 in FIG. 11) +1 (36 in FIG. 11) +2 (37 in FIG. 11) +20 (38 in FIG. 11) =24, and the shortest path is 39. From the above results, the amount of deviation between the input word string "Okara" 8 and noise 9 and the word "Okara°12" in the word dictionary is 34. The amount of deviation from the word “Okayama” 13 is 24, and the candidate word is 3 Okayama” 13, which has a smaller amount of deviation.
．． “Okara” Output in the order of 12, the correct word “Okara”
12 is not available at the top. [Problem to be Solved by the Invention] Conventional word reading devices are configured as described above, and because the penalty for incorrect insertion of characters when matching words in word determination is fixed, There is a problem in that when there is non-character noise etc. in the text, the correct word cannot be obtained because the noise etc. are also included in character recognition. The present invention has been made to solve such problems, and an object of the present invention is to provide a word reading device that can correctly match words even if there is noise or the like on the entry character frame of a form. [Means for Solving the Problems] The present invention includes a scanning means 2 that photoelectrically converts characters, etc. on a form 1 to generate image data, and a scanner that performs primary character recognition on the image data of the characters, etc. Candidate string 10
．． 11, a word dictionary 4 in which various words are registered in advance, and a character recognition means 3 that outputs a word 12.13 selected from the word dictionary 4 for the recognition candidate character string 10.11. A word reading device comprising a word determining means 59 that calculates using an arithmetic expression and ranks and selects words with a high probability value of identity, detects and outputs the non-character element rate of image data such as characters. A pseudo-noise character detection means 6 is provided, and a word determination means 59 performs calculations incorporating the non-character element ratio into various parameters, and ranks and selects words with a high probability of identity. (Operation) If non-character noise etc. is present in image data such as characters, the pseudo-noisy character detection means 6 detects the non-character element rate.Using this non-character element rate as a parameter, various conventional parameters are determined. , and calculate it using a predetermined calculation formula to determine which word 1 is the recognition candidate character string 10.11.
2. Word determining means 5 determines which word has the highest probability of identity with 13.
9, and the high ranking words 12 and 13 are determined to be the character string input from form 1. [Example] Hereinafter, an example of the present invention will be described based on the drawings. In Fig. 1, 1 is a form on which characters are written, 2 is a scanning means that reads this form l by photoelectric conversion, and 3 is a scanning means that cuts out and recognizes input characters one by one, and selects the first recognition candidate character. 4 is a word dictionary that stores reference words; 6 is a pseudo-noise character that observes an image of each input character written on a form and detects non-character images, such as characters that are predicted to be noise; The detection means 59 uses the information of the pseudo-noise characters outputted from the pseudo-noise character detection means 6 to compare the recognition candidate characters outputted from the character recognition means 3 with the reference words in the word dictionary 4. This is a word determining means that determines the word read by the user. The pseudo-noise character detection means 6 measures the height and width data of each piece of image data input from the scanning means 2 and outputs the measured data. In the word reading device of the present invention, it is assumed that a general form 1 shown in FIG. 3 is input. Further, the character recognition means 3 outputs recognition candidate characters 10 shown in FIG. 4 and input candidate characters 11 shown in FIG. 5 in response to input of the form 1 shown in FIG. Furthermore, words as shown in FIG. 6 are registered in advance in the word dictionary 4. FIG. 2 shows the pseudo-noise character detection means 6 of the word reading device.
This is an explanatory diagram for explaining an example. In Figure 6, which shows the width and height of the image for each character and noise in 7, 8, and 9 in Figure 3 showing a general form 1, 40 are the width and height of the image of the first character "O" of the input word string "Okayama", 41 are the width and height of the image of the second character "Power", 4
2 is the width and height of the image of the third character "Y", 43 is the width and height of the image of the fourth character "Ma", and 44 is the input word string "
The width of the image of the first character "o" in "Okara". Height, 45 is the width, height of the image of the second character "power", 46
are the width and height of the image of the third character "Y", and 47 are the width and height of the image of noise attached to the form. FIG. 10 shows the input word character string "Okara" 8 and noise 9 in FIG. 3 and the word "Okara" in the word dictionary 4 in FIG. 12 is an explanatory diagram illustrating comparison with "12. In the figure, 48 is the path where the penalty for character insertion errors is small;
1 is the distance between the first character “o” of the input word string “Okara” 8 and the “o” of the word “Okara” 12 in the word dictionary, 3
2 is the distance between the second character “Riki” in the input word string and “Riki” in the word dictionary 4, and 33 is the distance between the third character “Ya” in the human word string and “Ya” in the word dictionary 4. 24 is the distance for a single character insertion error, and 34 is the path that provides the shortest distance. FIG. 11 shows the input word character string "Okara" 8 and noise 9 in FIG. 3 and the word "OKAYAMA" 13 in the word dictionary 4 in FIG. 6 when the penalty for character insertion errors for pseudo-noise characters is reduced. In Figure 0, which is an explanatory diagram illustrating matching with ``O'', 48 is a path that reduces the penalty for character insertion errors, as in Figure 10, and 35 is a path with the first character "o" of the input word string "Okara" 8. Word “Okayama” 13 “ in word dictionary 4
36 is the distance between the second character "R" of the input word string and the word "R" in the word dictionary 4, and 37 is the distance between the third character "Y" of the input word string. The distance from the word “ya” in the word dictionary 4, 38, is the noise attached to the form 9
and the word "ma" in the word dictionary, and 39 is the route that gives the shortest distance. The word determining means 59 uses the general dynamics programming method shown in FIG. 7 with reference to the output signal shown in FIG. 2 from the pseudo-noisy character detecting means 6. Next, the operation will be explained. For example, the detection criteria for the pseudo-noise character detection means 6 are as follows. "When the width and height of the character image are both 4 or less, the character is considered a pseudo-noise character." When this criterion is applied to the width and height of the human-powered characters shown in Figure 2, 40-4
The width and height shown in 47 are both larger than 4, and the characters in the input word strings "Okayama" and "Okaya" do not become pseudo-noise characters, but the width and height shown in 47 are both larger than 4. 4 and is determined to be a pseudo-noise character. Next, when the word determining means 59 performs a comparison with the word dictionary 4, it reduces the penalty for character insertion errors as an element rate for pseudo-noise characters, that is, non-characters. In other words, the recurrence formula for the word determination method is as follows. <Word determination method> Input word character string from the form 7.8S=s. S2...S7 and the word T=t+, Lx in the word dictionary 4
. ...When comparing with A9, the characters Si and 1. Introducing the concept of distance from
l+3! . . Si and jl+'m . . . tj are introduced, and the deviation amount f (inj) is introduced and calculated using the following recurrence formula. f (0・11)=0' (j・j)=m 1 n (' 1i-1-J)
10dt +f (l・j−11+d3・f(mu−
1+ J-1) ” d 1) d, if there is 1 degree in the recognition candidate characters for the two characters Si, its rank, if there is no 1j in the recognition candidate characters, P (constant) d2: constant d,: constant ml n (LY4): The minimum value among X, Y, and Z, that is, the path from f + i-t, j + 15 to r tt, > 17 as shown in Fig. 7 is sI + 32...
・The comparison between S i-1 and Lliim...1j has been completed, and the degree of deviation is (allowing insertion errors). In this case, the letter S
. is considered to be an extra character, and therefore the penalty is d2.
19, f(in J) =' (i-1,j)
It shall be ten d2. Similarly, the route from f tt, j-n 16 to r(A1) 17 is 1. does not match a character in the input word string S as an extra character (allowing a missing one character error), but instead gives a penalty d, 20'(i, j) = r(i,
Let j−H+d 2・. Therefore, the word determining means 59 specifies d as follows, and compares the input word character string "OKAYAMA" 7 in FIG. 4 with the words in the word dictionary 4 in FIG. 6. In this case, since none of the characters in the input word character string "Okayama" 7 is a pseudo-noise character, in the recurrence formula, the predetermined dt "
dtb=30. Setting d3 = 30 and P = 20, the operation is exactly the same as the conventional example. Input word string “Okayama 7
The amount of deviation f(a
, 3+ is f 14.3+ "l (21 in Figure 8) + 1 (22 in Figure 8) + 2 (23 in Figure 8) + 30 (24 in Figure 8) = 34, and the minimum If the path that gives the amount of deviation '(4+31) is found by bank tracking, the path is (4, 3). 25 routes are followed. Also, the difference between the input word character string "Okayama" 7 and the word "Okayama" 13 in the word dictionary 4 is ffi f (4, 41
is f (4,41= 1 (26 in Figure 9) + 1 (27 in Figure 9) + 2 (28 in Figure 9) + 20 (29 in Figure 9), and the shortest path is 30. Therefore, the candidate words from the word determining means 59 are "Okayama" 13 in the word dictionary 4 with a small deviation, followed by "Okayama 1".
The words are output in the order of 2, and the correct word is obtained at the top. When the input word character string "OKAYA" 8 and noise 9 in FIG. 5 are compared with the words in the word dictionary 4 in FIG. It is determined that it is a gender character, and in the recurrence formula, by including the non-character element rate, d8=dx-+10.dzb=30.d
3 +30. Let it be P+20. The verification operation will be explained below. Input characters in Figure 5 and word 4 in the word dictionary “Okaya” 8
and noise 9, and the word T- in the word dictionary 4 in FIG.
becomes “Okaya” 12. SL (input word string “O” of “Okaya°8”) and 1
1 (the word "o" in the word "OKAYA°12" in the word dictionary), 1. is in the first position of the 31 recognition candidate characters 11, so d, -1 (31) becomes ' +111 -m 'n(' (lull ” d
tb+' (Ice)+d,,f,. ,. , ＋
1) = min (30+30.30+30.l). Similarly, when comparing S2 and 1t, 1t is in the first position of the 32 recognition candidate characters 11, so d, +1 (32) becomes f mountain ``”' l rl (f (1+t) + d
zb * f 11 ports) + ds * f (1, 1)
+ 1)=min (31+30.31+30゜1
+1) =2. Similarly, when comparing S3 and t, t is in the second position of the recognition candidate character 11 of S, so d, = 2 (33)' (3+ri ``m''n(' (t・2) 10d
Zb+ ' (3+g)+d3* f tz+t>
+2)=min (32+30.32+30°2+
2) =4. Similarly, when comparing S4 and +4, +4 does not exist among the 34 recognition candidate characters ll, so d, +20, and S4
is a pseudo-noise character, so f(4,. When finding (x=0.1, 2.3), the penalty for character insertion error path 48 is dz=dz-+10, so f(4,3)"'m i n (f (*, 3
) +dza, fn, z)+ d ro1 ・
f (=take, ko, +d plate)=m
in (4+10.42+30°32+20) +14. That is, the amount of deviation f(
4, 33 is f < a, s+ = 1 (31 in Figure 10)+
1 (32 in Figure 1O) +2 (33 in Figure 10) +10 (24 in Figure 10) +14, and the shortest path is 34. On the other hand, when the word "Okayama.
As shown in the figure, the input word string S is "Okayama" 8 and noise 9, and the word T in the word dictionary is "Okayama" 13.
becomes. ! , (“o” of input word string “Okaya” 8) and 1+
When comparing (the "o" in the word "Okayama" 13 in the word dictionary), t-o is ranked first among the recognition candidate characters 11 for s, so a, -t (35) becomes ' (lull = m' n (f
(Owl) + dZb + f (1,6
1+d, ・f (O・+1>+1)=
min (30+30.30+30.1)+1. Similarly, s2 and t! When comparing , 1t is in the first position of 33 recognition candidate characters 11, so d, +1 (36)' (!+ Goose 1 "m" (f (
++2) ” d!b+ ' (lull
+ds* f(+, ++ +1) = min (31+30. 31+30°1 +1) =2. Similarly, when comparing S and t, t is in the second position of the 33 recognition candidate characters 11, so d = 2 (37)' (3 + 3> = m1 n (f (1 + 21
+dtb*' +3・2)+d3, f (t
, t+ +2)=min (32+30.32+30
゜2+2) = 4. Similarly, when comparing S4 and +4, there is no +4 among the 54 recognition candidate characters, so d, = 20, and since S4 is a pseudo-noise character, f (4, 1 (x = 0.1, 2
．． 3) The penalty for path 48 for incorrect character insertion is d
,=d,,=to,' (4・41=m1
n (E <3+4+ ” d!11+ ’ (4
,31+d3・f'(3・s)+2'O')=mi
n (33+10. 34+30°4+20) = 24. That is, the amount of deviation f(
4,. is f, 4. . = 1 (35 in Figure 11) + 1 (11th
36 in the figure) +2 (37 in Figure 11) +20 (38 in Figure 11), and the shortest route is 39. From the above results, the amount of deviation between the input word string "OKAYA" 8 and noise 9 and the word "OKAYA" 12 in the word dictionary is 14, while the amount of deviation between the input word string "OKAYA" 8 and noise 9 and the word "OKAYA" 12 in the word dictionary is 14. The amount of deviation from the word “Okayama” in 13 is 2.
4, and the candidate words are output in the order of "Okaya" with the smallest amount of deviation, followed by "Okayama", and the correct word "Okaya 1" is obtained at the top. In this way, by using the pseudo-noise character detection means 6, The image of each input character is observed to detect characters predicted to be noise, and word matching is performed by reducing the penalty for character insertion errors for characters determined to be pseudo-noise characters. Words can be matched accurately even in the presence of noise. In the above embodiment, the width and height of the image were used as detection criteria for the pseudo-noisy character detection means 6, but the number of black dots in the image, the shape of the image, etc. Other detection methods may also be used. In the above embodiment, the word determination means 59 uses a DP matching method using the concept of Guinamisoku programming, but other matching methods may be used. In the above example, when matching the input word string with the characters constituting the word in the word dictionary, the ranking of the recognition candidate characters is used as the degree of matching, but other information (for example, the rank of the recognized character (e.g., the probability that the input character is that character) may also be used.In addition, although the above embodiment deals with katakana surnames, the present invention also applies to words such as kanji and English characters, and words such as addresses and company names. can be applied. Also, although dz = d2 - = 10 is set as the element ratio of non-characters, it is not limited to this and other values may be used. [Effect of the invention] As explained above, this According to the invention, a scanning device generates image data by photoelectrically converting characters, etc. on a form, a character recognition device performs primary character recognition on the image data of the characters, etc., and outputs a recognition candidate character string; A word dictionary in which various words are registered in advance is compared with words selected from the word dictionary for recognition candidate character strings, and calculations are performed using a predetermined calculation formula that includes various parameters, and words with a high probability of identity are ranked. In a word reading device, a pseudo-noise character detection means is provided which detects and outputs a non-character element rate in image data such as characters, and the word determination means detects and outputs a non-character element ratio in image data such as characters. By incorporating the ratio into the various parameters mentioned above, the words with a high probability of identity are ranked and selected, so even if there is non-character noise etc. on the form, the characters can be accurately input. The word can be determined with high accuracy by comparing the string with the reference word. [Brief Description of the Drawings] Fig. 1 is a block diagram of a word reading device according to an embodiment of the present invention, and Fig. 2 is a non-standard diagram. FIG. 3 is an explanatory diagram showing an example of a form for a word reading device; FIGS. 4 and 5 are explanatory diagrams showing examples of input characters and recognition candidate characters in the word reading device. , FIG. 6 is an explanatory diagram showing an example of a word string in a word dictionary in a word reading device, FIG. 7 is an explanatory diagram showing an example of a word determination method in a word reading device, and FIG.
FIG. 9, FIG. 1O, and FIG. 11 are explanatory diagrams showing an example of word matching in a word reading device, and FIG. 12 is a configuration diagram of a conventional word reading device. 1... Form, 2... Scanning means, 3...
...Character recognition means, 4...Word dictionary, 6.
...Suspicious noisy character detection means, 7.8...
... input character, 10.11 ... recognition candidate character string, 12.13 ... reference word, 59 word determining means. Agent Masuo Ohatsu (and 2 others) Toki 40 10th ntZl 'I! 38I! l Figure 9 2, Title of the invention 3. Single word continuation reading correction device for the person making the amendment (spontaneous) & the entire text of the specification to be amended, and the drawing section. Contents of G Amendment + 11 The entire text of the specification will be amended as shown in the attached sheet. (2) The drawings and Figure 2 shall be amended as shown in the attached sheet. Representative Moriya Shiki Specification (Full text correction) 1. Invention title word reading device 2. Scanning means for generating image data by photoelectrically converting the characters, etc. on the claims form, and A character recognition means performs literary pressure 1 on image data and outputs a recognition candidate character string, a word dictionary in which various words are registered in advance, and the recognition candidate character string is compared with words selected from the word dictionary and various parameters are determined. In a word reading device comprising a word determining means for performing calculation using a predetermined calculation formula including a predetermined calculation formula, and ranking and selecting words with a high double match. Pseudo-noise character detection means is provided for detecting and outputting 2M OTS'L and Tate α Wariji in the image data of the above-mentioned characters, and the word determination means detects and outputs the 2L INO○ A word reading device characterized in that the calculation is performed by incorporating various parameters, and the words with the highest dichotomy are ranked and selected. 3. Detailed Description of the Invention [Field of Industrial Application] This invention relates to a word reading device that reads and recognizes words such as addresses and names, and in particular, recognizes each character that constitutes a word, and reads the recognition results. This invention relates to a word reading device that corrects words using . [Prior Art] In recent years, word reading devices have been attracting attention as a means of entering large amounts of high-speed data into computers, etc., and since Japanese has a complex mixture of kanji, kana, alphanumeric characters, etc. It is desired to develop a word reading device that can read words with high accuracy. As this word reading device, an optical character reading device (OCR) is known. )) etc. In the figure, 1 is a form on which characters are written, 2 is a scanning means for photoelectrically converting and reading the form 1, and 3 is a scanning means for reading input characters.
Character recognition means that cuts out and recognizes each character and outputs recognition candidate characters; 4 is a word dictionary that stores reference words; 5 is recognition candidate characters output from character recognition means 3 and words in word dictionary 4; This is a word determining means that determines a word by comparing the words. The conventional word reading device is configured as described above, and outputs words that are ranked higher among the recognition candidate characters and are registered in the word dictionary 4 and are preferentially selected as the word reading results. be. FIG. 3 is an explanatory diagram showing an example of a general form 1 to be read. In FIG. This is attached (non-character) noise. Examples of recognition candidate characters output from the character recognition means 3 for characters on the form are shown in FIGS. 4 and 5. The figure shows the recognition candidate characters lO output by the character recognition means 3 for each character of the input word character string "Okayama" 7, and for the first character "o" of the input word, "
One recognition candidate character for ``o'' is used for the second character ``power'', one recognition candidate character for ``power'' is used for the third character ``ya'', There are two recognition candidate characters, ``ma'' and ``ya'', and one recognition candidate character, ``nu'', exists for the fourth character 173. Figure 5 shows the input word string “
Recognition candidate characters 11 output by the character recognition means 3 are shown for each character of "Okaya" 8 and noise 9 attached on the form, and for the first character "O" of the input word "Okaya". In this example, one recognition candidate character for “o” is used for the second character, “power”, and one recognition candidate character for “power” is for the third character, “9ya”. In this case, there are two recognition candidate characters "73. "Ya", and there is one recognition candidate character "ノ" for noise 9 attached to the form. Figure 6 shows the word reading. 6 is an explanatory diagram showing an example of words in a general word dictionary 4 regarding katakana surnames in the device; FIG.
In the figure, 12 is the word "Okaya", and 13 is the word "Okayama". FIG. 7 is an explanatory diagram of a word determination method using the general dynamic programming concept used in the word determination means 5 of the word reading device. Character string SI that constitutes a word
+! ! ...S, and tl+l! ...When checking with tゎ, partial string SI+3! ...Si and jl+L!
...1. The amount of deviation when the two best match is f(
! , 14 in the figure is f(tt, J-1
> , 15 is '(i-1+j> + 16 is r(i
nj-1) + 17 is f(i, Jl, and
18 is the distance d between f (i-1・j-1) and f Yuha, and 19 is f. -1, 5, and '+! +j) The distance between at, 2o is f(in j-11 and f(!,
j1 is the distance d. Figure 8 shows the input word character string “Okayama” 7 and the word dictionary 4.
In Figure 0, which is an explanatory diagram illustrating the matching with the word "Okayama" 12 in , 21 is the input word string "Okayama" 7
The first letter of and the word “okaya” in word dictionary 4
The distance 22 from the first character in 2 is the distance between the character power of the second character and the "power" of the word in the word dictionary 4, and 23 is the distance between the character y of the third character and the "power" of the word in the word dictionary. ”, 24 is the distance to the character insertion film, and 25 is the path that gives the shortest distance. Figure 9 shows the input word string “Okayama” 7 and the word dictionary 4.
In Figure 0, which is an explanatory diagram illustrating matching with word 1 ``Okayama'' in 13, 26 is the input word string ``Okayama''.
The distance between the first character of 7゛ and the "O" of the word "Okayama" 13 in the word dictionary, and 27 is the distance between the power of the second character and the word "power" in the word dictionary 4. , 28 is the distance between the third character ya and the word "ya" in the word dictionary 4, 2
9 is the distance between the fourth character ``ma'' and the word ``ma'' in the word dictionary, and 30 is the route that provides the shortest distance. Figure 10 shows input word character examples “okaya” 8 and noise 9.
FIG. 4 is an explanatory diagram illustrating a comparison between the word "OKAYA" 12 in the word dictionary 4 and the word "OKAYA" 12 in the word dictionary 4. In the figure, 31 is the input word string "
Okaya "8's first letter O and the word in the word dictionary"
The distance from ``O'' of Okaya'''12, 32 is the character power of the second character of the input word string and the "power" of the word in the word dictionary 4.
33 is the distance between the third character ya of the input word and "ya" in the word dictionary 4, 24 is the distance for a single character insertion error as in FIG. 8, and 34 is the shortest distance. This is the route that gives FIG. 11 shows the input word character string "Okaya" 8, the noise 9 attached to the form, and the word "Okayama" 13 in the word dictionary 4.
FIG. In the figure, 35 is the distance between the first character O of the input word string "Okaya" 8 and the "O" of the word "Okayama" 13 in the word dictionary 4, 3
6 is the power of the second character and the “power” of the word in word dictionary 4
37 is the distance between the third letter strength and the word "ya" in the word dictionary, and 38 is the distance between the noise 9 attached to the form and the word "ma" in the word dictionary 4. 39 is the route that provides the shortest distance. Next, the operation of the conventional device will be explained. The word determining means 5 compares the input word character string entered in the form shown in FIG. 3 with the words in the word dictionary 4 shown in FIG. This is done using the following method. <Word determination method> When comparing the input word character strings 7, 8S-81, S,...S with the words T=tl, tt...+7 in the word dictionary 4, the characters SL and t, Introducing the concept of distance from
0 substring SI+s2...31 with this as dl
and LI+Lffi...1. Introduce the amount of deviation f'(!, j) when the two best match, and calculate this using the following recurrence formula. f(O・0)=Of(1,J)=min(f+i−+,j
+ + dt +f (lid/j-11+a 3
・ f (5-I・j-lightning)+d+)dl:
Character 3! If there is 1° in the recognition candidate character, its rank, t in the recognition candidate character, otherwise P (constant) d2: constant d,: constant m l n (X, Y, Z): X, Y, The minimum value in Z, that is, f (i-+
, j) The route from 15 to f(!, c 17 is Sl+51
...S, -1 and 1. ．． 1. ...The matching up to L has been completed, and the degree of deviation is f (i - 1°) 15, so the character Si and the character in T are not matched (allowing for one character insertion error). It means that. In this case, character 3i is considered to be an extra character, and therefore a penalty of dz19 is given, '(i+j+ =' (i-1+
j) Set to +d2. The route to f(i, j-1, 16 to f., c) is similar, and 1. is an extra character and is not matched with a character in the input word string S (allowing for a missing one character error). ) Instead, a penalty of d320 is given and f(i, j) = f(!+j
−1) +d 3 . The path from f(i-1+j-1) 14 to f(mu+j) is a path for matching the character Sm with the character 1j, and the degree of this matching is defined as d. dl takes the following values. Let f (so al ) obtained by this recurrence formula be the amount of deviation between the input word string sl+ Sl...Sl and the words tIn1t...tll in the word dictionary, and calculate the amount of deviation to all words in the word dictionary. are determined, rearranged in descending order of the amount of deviation, and output candidate words. Using the word determination method, predetermined dt=30, d
3 = 30. The case where P=20 will be explained. The word determining means 5 compares the input word character string "Okayama. 7" shown in FIG. 4 with the words in the word dictionary 4 shown in FIG. Verify. As shown in Figure 8, the input word string S is "Okayama" 7, and the word T in the word dictionary is "Okayama" 12. Value of f(.,X) (X≧1) When calculating, the recurrence formula is
f(.,It)=f(.llll-1)+d, and f,. , +1 = ds = 30 fto, t> = 2d3 = 60 f(., 3) = 3di = 90. Similarly f. ,. ) When calculating the value of (X≧1), the recurrence formula is f IX, +1) ” f (X-1101”
a, and f (1,.*=d*=30 f(!,.> ”=2d* =60f N...>
=3dFu -90 f(4,.) Fog 4a, =120. Next, SI (input word 7 “O” in Okayama°) and 1
+ (1 o of word 12 “Okaya” in word dictionary 4)
tl is ranked first among 31 recognition candidate characters 10.
Therefore, d in Figure 8 becomes +1 (21) and f 1. ,,) xm i n (f
(out> +d, l f tll
ll>+ds, f(o, e> +1) -min (30+30.30+30.1)+1. Similarly, when comparing S and +8, there is no 1 among the 31 recognition candidate characters 10, so dl +20, ' (1,
ffi> −rn l n (f (6
+ri ” d t + f (Inl1
+d3. f(., 1)+20) -min (60+30. t+ao. 30+20) +31. When st and tl are compared, t is found among recognition candidate characters 0 of st.
Since l does not exist, dl=20, and ' (ton
−m l n (f (ill> +a, + f
(zoo)+d3・r<+・o) +20) =min (1+30.60+30°30+20) +31. When st and +8 are compared, t is the first of 32 recognition candidate characters 10! Therefore, d+'=1(22) in Figure 8'<til>""' I n (f (In)
+dt + '(1-11+d3+f(凰1
1) + 1)=min (31+30.31+
30°1+1) =2. When comparing Sl and tj, t is found among 31 recognition candidate characters 0.
, does not exist, so dl=20, and '<In! l
"ml n ('(Or2> 10d! r
'(+-2>+d,, f(., z>+20) = min (90+30.31+30°60+20) +61. When comparing S and tl, t is found in the recognition candidate character O of S.
Since l does not exist, d+=20, '(:l+1
) ==m'n(' (!nil +dt +
' (3+@)+43r f+g, o+ +20
)=min (31+30.90+30°60+20) +61. When comparing 3t and t, among the 32 recognition candidate characters O, (
, does not exist, so dl=20, and f <t's
) = min (f <+,,>
+dt + f (t, t) + ct,,
fH + +20) = min (61 + 30.2 + 30°31 + 20) = 32. When comparing 1t with S, there is no +2 in the 10 recognition candidate characters of S, so a, = 2O, ' (3+2
+ ”fly l n (f (!, ri ” dl
+'<:ilH+d, , f, t, n
+20)=min (2+30.61+30°31+20)=32. When comparing S, and t, there is +3 in the second position of the recognition candidate character lO of S, so d, = 2 (23), and r (3, 31' = min (f < t, x, +d
z, r ts, z. +d3・f +t, t> + 2)=min (3
2+30.32+30°2+2) =4. When s4 and tl are compared, tl does not exist among the 34 recognition candidate characters lO, so a, = 20, and ' (4+1
) ”m'n(' (3+l) +d* +
f <4+@>+ d s * f ll+. )+2
0)=min (61+30.120+30°90+2
0) =91. s4 and t! When t2 does not exist among the 34 recognition candidate characters lO, it becomes d, -20, and f (41ri = m l rl (f (3・ffi+ ”dz
+ f (411) + d, , f +s, t>
+20)=min (32+30.91+30゜6
1+20) =62. When s4 and t are compared, t does not exist among the 34 recognition candidate characters lO, so d=20, and' (4+
31 4 m '''(' (3-3)
+ d t * ' (4, ri + d,
・rts・snow)+20) = mln (4+30.62+30°32+20) =34. That is, the amount of deviation r < a, 3) between the input word character string "Okayama" 7 and the word "Okaya" 12 in the word dictionary is f < a, s+ -1 (21 in Figure 8) + 1 ( 8th
22) in the figure +2 (23 in Figure 8) +30 (24 in Figure 8), and when the path that gives the minimum deviation amount f(41,) is found by backtracking, (4,3) . The route passes through the points (3,3), (2,2), (1,1), and (0,0), resulting in 25 routes. Next, a comparison is made with the word "Okayama" 13 in the word dictionary 4. As shown in FIG. 9, let the input word string S be 1 OKAYAMA"7, and the word T in the word dictionary be "OKAYAMA"13. 3. ("O" in the input word string "OKAYAMA°7") and 1+ (word “Okayama” in the word dictionary 13 “O°”
), tl is in the first position of recognition candidate character 10 of character 31, so d+ = 1 (26) in Figure 9'(1+1>"ml n (' (ll+1
) +d1 * f (roll>+d,・f
(o...) +1) -min (30+30.30+30.1)=1. Similarly, when we check st and t2, we get 3! Since 1t is in the first place among the recognition candidate characters 10, d in FIG. 9 is 1.
(27), f tt+z+ = min (f <r+z>
+d寞+f (!+l)+d,・f(1・I)
+1) =mirl (31+30.31+30°1+1) =m=2. Similarly, when comparing S and t, t is in the second position of the recognition candidate character lO of S, so d+ -2(
28) Then f <s, s> ”min (f (*, x>
+dz, f <z, t)+dco・f (t・
z>+2) -min (32+30.32+30°2+2) =4. Similarly, when comparing S4 and t4, t4 does not exist among the 34 recognition candidate characters 10, so d in FIG. 9 = 20
(29) becomes f <a+a+ =mln (f (14) +a
, l f (413++a, , f 13.s
> +20)-min (33+30.34+30°4+20) =24. In other words, the input word string "Okayama" 7 and the word dictionary 4
The amount of deviation f (4+41
is f < a, a+ -1 (26 in Figure 9) +1 (27 in Figure 9) +2 (28 in Figure 9) +20 (29 in Figure 9) 24, and the shortest path is It will be 30. From the above results, the amount of deviation between the input word character string "Okayama" 7 and the word "Okayama" 12 in the word dictionary is 34. The amount of deviation between the input word string "Okayama°7" and the word "Okayama" 13 in the word dictionary is 24, and the candidate words are output in the order of "Okayama" 13 and "Okayama" 12, which have the smallest amount of deviation, and are correct. In addition, the input word string written on the form in Figure 5 “
The word determination means compares the noise 9 adhering to the word "OKAYA" 8 and the form with the words in the word dictionary 4 shown in FIG. 6. By the word determination method, the word "OKAYA" in the word dictionary is
12, as shown in FIG. 10, the amount of deviation f (4, )
is f, <a, s> = 1 (31 in Figure 10) + 1(
32 in FIG. 10) +2 (33 in FIG. 10) +30 (34 in FIG. 10) =34, and the shortest path is 34. Next, when the word "Okayama" 13 in the word dictionary 4 is compared, the input word string "Okayama" is found as shown in FIG.
8 and noise 9, and the word “Okayama” 1 in the word dictionary 4
The amount of deviation f(4, n) from 3 is f 14.4>=
1 (35 in FIG. 11) +1 (36 in FIG. 11) +2 (37 in FIG. 11) +20 (38 in FIG. 11) =24, and the shortest path is 39. From the above results, the amount of deviation between the input word string "OKAYA" 8 and noise 9 and the word "OKAYA" 12 in the word dictionary is 34. The amount of deviation from the word “okayama” 13 is 24, and the candidate word is “okayama” 13, which has a smaller amount of deviation.
．． ``OKAYA112'', and the correct word ``OKAYA112'' cannot be obtained at the top. [Problem to be solved by the invention] The conventional word reading device is configured as described above, and the word reading device is configured as described above. Since the penalties for incorrectly inserting characters during verification are fixed, there is a problem in that if non-character noise, etc. is present in the input character frame of the form, the correct word cannot be obtained.This invention was developed in order to solve this problem, and the purpose is to obtain a word reading device that can accurately match words even if there is noise etc. on the entry character frame of a form. Means for Solving the Problem] The present invention includes a scanning means 2 that photoelectrically converts characters, etc. on a form l to generate image data, and performs character recognition on the image data of the characters, etc. to generate recognition candidate character strings 10. 11
A character recognition means 3 that outputs a character recognition means 3, a word dictionary 4 in which various words are registered in advance, and a recognition candidate character string 10.11 is compared with a word 12.13 selected from the word dictionary 4 and a predetermined character string including various parameters is compared. In a word reading device comprising a word determining means 59 which calculates using the calculation formula and ranks and selects words with a high degree of political change, pseudo noise is detected and outputted as a noise-like ratio of image data such as characters. Gender character detection means 6 is provided, and word determination means 59 incorporates this noise-likeness ratio into various parameters and performs calculations to rank and select words with a high degree of -political change. [Operation] If non-character noise or the like is present in image data such as characters, the pseudo-noise character detection means 6 detects the noise-likeness ratio. This noise-likeness ratio is added to the conventional various parameters as a parameter, and calculated using a predetermined calculation formula to determine which word 12, 13 the recognition candidate character string 10.11 has the highest degree of matching. The determining means 59 selects and determines the high-ranking word 12.13 to be the character string input from the form l. [Example] Hereinafter, an example of the present invention will be described based on the drawings. In Fig. 1, 1 is a form on which characters are written, 2 is a scanning means that reads this form l by photoelectric conversion, and 3 is a scanning means that cuts out and recognizes input characters one by one, and selects the first recognition candidate character. Characters to output! ! 4 is a word dictionary storing reference words; 6 is pseudo-noise character detection means for observing an image of each input character written on a form and detecting non-character images, such as characters predicted to be noise; 59 is a word read by comparing the recognition candidate character outputted from the character recognition means 3 with the reference word in the word dictionary 4 using the information of the pseudo-noise character outputted from the pseudo-noise character detection means 6; This is a word determining means for determining. The pseudo-noise character detection means 6 measures the height and width data of each piece of image data input from the scanning means 2 and outputs the measured data. In the word reading device of the present invention, it is assumed that a general form 1 shown in FIG. 3 is input. Further, the character recognition means 3 outputs recognition candidate characters 10 shown in FIG. 4 and input candidate characters 11 shown in FIG. 5 in response to input of the form 1 shown in FIG. Furthermore, words as shown in FIG. 6 are registered in advance in the word dictionary 4. FIG. 2 shows the pseudo-noise character detection means 6 of the word reading device.
This is an explanatory drawing for explaining an example. 40 is the width and height of the image of the first character "O" of the input word string "Okayama", 41 is the width and height of the image of the second character "Chikara", 4
2 is the width and height of the image of the third character "Y", 43 is the width and height of the image of the fourth character "Ma", and 44 is the input word string "
Width of the image of the first character “O” in Okaya°. Height, 45 is the width and height of the image of the second character "power", 46
are the width and height of the image of the third character "Y1", and 47 are the width and height of the noise image attached to the form. Figure 10 shows how to reduce the penalty for character insertion errors for pseudo-noise characters This is an explanatory diagram illustrating how the input word character string "OKAYA" 8 and noise 9 in FIG. 3 are matched with the word "OKAYA" 12 in the word dictionary 4 in FIG. 6 when 31 is the distance between the first character "O" of the input word string "OKAYA" 8 and "O" of the word "OKAYA" 12 in the word dictionary, and 32 is the input word character. The second character in the column “power” and “ in the word dictionary 4
33 is the distance between the third character "ya" in the input word string and "ya" in the word dictionary 4, 24 is the distance for a single character insertion error, and 34 gives the shortest distance. Fig. 11 shows that when the penalty for character insertion errors for pseudo-noise characters is reduced, the input word character string "Okaya" 8 and noise 9 in Fig. 3 and the word dictionary 4 in Fig. 6 are It is an explanatory diagram explaining the collation with the word "Okayama" 13. In the figure, 48 is the path where the penalty for character insertion error is reduced as in FIG. 10, and 35 is the first path of the input word string "Okayama" 8.
The distance between the character "o" and the "o" of the word "Okayama" 13 in the word dictionary 4, 36 is the second character of the input word string "
37 is the distance between the third character "Ya" of the input word string and the word "Ya" in the word dictionary 4, and 38 is the distance from the word "Ya" in the word dictionary 4. This is the distance between the noise 9 attached to the word "ma" in the word dictionary, and 39 is the route that gives the shortest distance. The word determining means 59 uses the general dynamic programming method shown in FIG. 7 with reference to the output signal shown in FIG. 2 from the pseudo-noise character detecting means 6. Next, the operation will be explained. For example, the detection criteria for the pseudo-noise character detection means 6 are as follows. "When the width and height of the character image are both 4 or less, the character is considered a pseudo-noise character." When this criterion is applied to the width and height of the input character shown in Figure 2, 40-4
The width and height shown in 47 are both larger than 4, and the characters in the input word strings "Okayama" and "Okaya" do not become pseudo-noise characters, but the width and height shown in 47 are both larger than 4. 4 and is determined to be a pseudo-noise character. Next, the word determining means 59 reduces the penalty for character insertion errors for pseudo-noise characters when performing comparison with the word dictionary 4. In other words, the recurrence formula for the word determination method is as follows. <Word determination method> Input word string from form? , 8S=st. s2...s, 1 and the word T in the word dictionary 4 = t,, t!
. ...When performing a comparison with t7, we introduce the concept of the distance between characters s1 and 1j, and let this be dl. Partial string S1
, 3□...31 and tIn tJ...1j are introduced, and the deviation amount f(mu+j) is calculated using the following recurrence formula. f (.,.) 0 f (direct, J) = min (f (!-hJ
) 10d-f (Todoroki+j-11+d 31
f 1l-11J-11+d Kuzu) dl: Letter S! If there is tj among the recognition candidate characters, its rank; if there is no 1J among the recognition candidate characters, then P (constant) d8: constant d,: constant min (X, Y, Z): of X, Y, Z In other words, as shown in FIG. 7, f (i-1
,115 to f (the route to 1+117 is 31,3□・
...S. and jl+ to wealth...1. The verification has been completed,
The degree of deviation is f (1-11J) 15, which means that the character 3i is not compared with the characters in T (one character insertion error is allowed). In this case, the character S1 is considered to be an extra character, and therefore the penalty d atmosphere is 1.
Given 9, f mountain = r (Toho, ha + d2. f (i + J - 1) The route from 16 to f mountain, 17 is the same, and tJ is an extra character in the input word string S. It does not match the characters (allowing for the missing error of one character), but instead gives a penalty of d, 20.
Let it be J-11+d2. Therefore, the word determination means 59 specifies dO as follows, and matches the input word character string "Okayama" 7 in FIG. 4 with the words in the word dictionary 4 in FIG. 6. In this case, since none of the characters in the input word string "Okayama" 7 is a pseudo-noise character, in the recurrence formula, the predetermined dt=
dz,=30. ds = 30 and P = 20, and the operation is exactly the same as the conventional example. Input word string “Okayama” 7
The amount of discrepancy between and the word “Okaya” 12 in the word dictionary f<a
, x> is f <a, s+ = 1 (21 in Figure 8) + 1 (22 in Figure 8) + 2 (23 in Figure 8) + 30 (24 in Figure 8)! = 34, and if we find the path that gives the minimum amount of deviation '(4+3) by backtracking, we get (4, 3). The route passes through the points (3,3), (2,2), (1,1), and (0,0), resulting in 25 routes. Also, the amount of deviation f(4,) between the input word string "Okayama" 7 and the word "Okayama'13" in the word dictionary 4 is f (4,4) = 1 (26 in Figure 9) + 1.
(27 in FIG. 9) +2 (28 in FIG. 9) +20 (29 in FIG. 9) =24, and the shortest route is 30. Therefore, from the word determination means 59, the candidate word is ``Okayama'' 13 in the word dictionary 4 with a small deviation, and then ``Okayama'' 1.
The words are output in the order of 2, and the correct word is obtained at the top. When the input word character string "OKAYA" 8 and noise 9 in FIG. 5 are compared with the words in the word dictionary 4 in FIG. It is determined that the character is a noise character, and in the recurrence formula, by including the noise-likeness ratio, ds = dz-=10. dzb=3
0. Let d3=30°P=20. The verification operation will be explained below. The input word character string S becomes "okaya" 8 and noise 9 in FIG. 5, and the word T in the word dictionary 4 in FIG. 6 is "okaya".
It becomes 12. st (input word string “O” in “Okaya°8”) and 1
+ (“O” of the word “Okaya” 12 in the word dictionary) is compared, and it is ranked 1st out of 31 recognition candidate characters 11. Therefore, d,=1 (31)' (11) -m'n(' (11+ll +
dtb*' (1+el+d,,f,.,.,+1
) =min (30+30.30+30.1)=1. Similarly, 3. When comparing and t, t2 is in the first position of the 32 recognition candidate characters 11, so d, = 1 (32) and f (!+ri ”m j n (f ll+ri +
dtbr'(!++)+d3・f<r,n
+1) =min (31+30.31+30°1+1) =2. Similarly, when comparing S and t, t is in the second position among the recognition candidate characters 11 of S, so d, = 2 (33) and f + s * s) = min (f
(!+3) +dtb+ f (1
+! 1+d3+f,z+t>+2)=min
(32+30.32+30°2+2) =4. Similarly, 3. When comparing t4 with
are pseudo-noise characters, so f, 4. . (x=0.1.2.3), the penalty for character insertion error path 48 is a! = az, = 10, so '+4+2>"""n('(1+3)"
dta+'4+t>+d3,f<s,
*> +(L)=min (4+10.42
+30°32+20). That is, the amount of deviation f1 between the input word character string "OKAYA" 8 and the noise 9 and the word "OKAYA" 12 in the word dictionary 4 is
4+21 is r t4. *> = 1 (31 in Figure 1O) +
1 (32 in FIG. 10) +2 (33 in FIG. 10) +10 (24 in FIG. 10) The shortest route is 34. On the other hand, when the word "Okayama" 13 in the word dictionary 4 of FIG. 6 is compared with the input characters and noise of FIG. Therefore, the word T in the word dictionary becomes "Okayama" 13. 3. (“O” of input word string “Okaya” 8) and 11
("O" in the word "Okayama" 13 in the word dictionary), t is in the first position of the recognition candidate characters 11 of s, so d, = 1 (35) and ' (1, 11””m 1 n (f (0+ll
+dtb+ 'Tl+61"ds, f (
(1,0>+1)=min (30+30.3
0+30.1)l. Similarly, when comparing S2 and t2, t2 is in the first position of the 32 recognition candidate characters 11, so d, = 1 (36), and f u, t+ = m l n (f n, t, +dt
b* f <t, +. +d3・fil・+1+1) =min (31+30.31+30°1+1)=2. Similarly, when comparing S and Tomo, there is t in the second position of the recognition candidate character 11 of S, so d,=2(37)' (3+31 ``rn l n (' (!+1 )
"dRbr' +3.21+d, f(!
・ri+2) = min (32+30.32+30°2+2). Similarly, when comparing s4 and t4, t4 does not exist among the 34 recognition candidate characters, so d, = 20, and S4 is a pseudo-noise character, so f(4,. (x = 0.1.2 .3), the penalty for character insertion error path 48 is dt=az,=10, and r (4
, 41=min (f <s, a+
+ d211+ r <a, s++d3・f
(3・3)+20) = min (33+10.34+30°4+20) =24. In other words, the amount of deviation between the input word character string "Okaya" 8 and the noise 9 and the word "Okayama" 13 in the word dictionary '(
4+4) is f <a, a> ” 1' (35 in Figure 11)+
1 (36 in FIG. 11) +2 (37 in FIG. 11) +20 (38 in FIG. 11) =24, and the shortest path is 39. From the above results, the amount of deviation between the input word string "OKAYA" 8 and noise 9 and the word "OKAYA" 12 in the word dictionary is 14, while the amount of deviation between the input word string "OKAYA" 8 and noise 9 and the word "OKAYA" 12 in the word dictionary is 14. The amount of deviation from the word “Okayama” in 13 is 2.
4, and the candidate words are output in the order of ``6 OKAYA'' with a small amount of deviation, followed by 1 OKAYAMA.The correct word ``OKAYA'' is obtained at the top.In this way, by using the pseudo-noise character detection means 6. The image of each input character is observed to detect characters predicted to be noise, and word matching is performed by reducing the penalty for character insertion errors for characters determined to be pseudo-noise characters. Words can be matched accurately even in the presence of noise. In the above embodiment, the width and height of the image were used as detection criteria for the pseudo-noisy character detection means 6, but the number of black dots in the image, the shape of the image, etc. Other detection methods may also be used. In the above embodiment, the word determining means 59 uses a DP matching method using the concept of Guinamisoku programming as a word matching method, but other matching methods may be used. In the above embodiment, when matching the input word string with the characters constituting the word in the word dictionary, the ranking of the recognition candidate characters is used as the degree of matching, but other information (for example, when the characters are recognized (e.g., the probability that the obtained input character is that character).In addition, in the above embodiment, katakana surnames were described, but this invention can also be applied to words such as kanji and English characters, and words such as addresses and company names. In addition, although the ratio of noise-likeness is set to d, = d!, = 10, it is not limited to this and other values may be used. [Effects of the Invention] As explained above, this According to the invention, the scanning means generates image data by photoelectrically converting characters, etc. on a form, the character recognition means performs character recognition on the image data such as the characters, and outputs a recognition candidate character string; A word dictionary in which words are registered is compared with words selected from the word dictionary for recognition candidate character strings, and calculations are performed using a predetermined calculation formula including various parameters, and words with high political change are ranked and selected. A word reading device comprising pseudo-noise character detection means for detecting and outputting a noise-likeness ratio of image data such as characters, and a word determining means incorporating this noise-likeness ratio into the various parameters. By performing calculations, we ranked and selected words with high political change, so even if there is non-character noise etc. on the form, the accurately input character strings are compared with the reference words, and the words with high political change are selected. Able to determine words with precision.

[Brief explanation of the drawing]

FIG. 1 is a configuration diagram of a word reading device according to an embodiment of the present invention, FIG. 2 is a diagram showing the value of the noise-likeness ratio, FIG. 3 is an explanatory diagram showing an example of a form of the word reading device, and FIG. 4 5 is an explanatory diagram showing an example of input characters and recognition candidate characters in a word reading device, FIG. 6 is an explanatory diagram showing an example of a word string in a word dictionary in a word reading device, and FIG. 7 is an explanatory diagram showing an example of a word string in a word dictionary in a word reading device. Explanatory diagram showing an example of the word determination method in 8th
9, 10, and 11 are explanatory diagrams showing an example of word matching in a word reading device, and FIG. 12 is a configuration diagram of a conventional word reading device. l...Form, 2...Scanning means, 3...
...Character recognition means, 4...Word dictionary, 6.
...Suspicious noise character detection means, 7, 8...
...Input character, 10.11...Recognition candidate character string, '12.13...Reference word, 59 word determination means.

Claims (1)

[Scope of Claims] A scanning means for photoelectrically converting characters, etc. on a form to generate image data, and a character recognition means for performing primary character recognition on the image data of the characters, etc. and outputting a recognition candidate character string. , the word dictionary in which various words are registered in advance is compared with the words selected from the word dictionary for the recognition candidate character string, and the words with a high probability of identity are calculated using a predetermined calculation formula including various parameters. A word reading device comprising a word determining means for ranking and selecting a word, a pseudo-noise character detecting means for detecting and outputting a non-character element ratio for the image data such as characters, and a word determining means for A word reading device characterized in that the calculation is performed by incorporating a non-character element rate into the various parameters, and words with a high probability value of identity are ranked and selected.