JPH03260886A - Character recognizing method - Google Patents

Abstract

PURPOSE:To improve the rate of recognition by calculating a ratio to a standard character in a document including a character to be recognized and comparing this ratio with a reference value calculated in advance concerning at least one of the character width, character height, area of circumscribed rectangle and aspect ratio of the character to be recognized. CONSTITUTION:Remaining difference Hn of the normalized character height to be defined by expressions I and II, and remaining difference Wn of the normalized character width is calculated. In the expressions I and II, H and W show the character height and character width to be measured from the character picture, Hs and Ws show the standard character height and standard character width in the II document, Hmuj and Hdeltaj show the average value and standard deviation of the ratio to the standard character height for each character form of a character type (j), and Wmuj and Wdeltaj show the mean value and standard deviation of the ratio to the standard character height for each character form of this character type (j). Next, remaining difference NVn of a normalized longitudinal discrete number and remaining difference NHn of a lateral discrete number is calculated. Namely, concerning one of the character width, character height, area of the circumscribed rectangle and aspect ratio of the character to be recognized at least, the ratio to the respectively corresponding amount of the standard character in the document including the character to be recognized is estimated from the character size. Thus, recognition is enabled with high accuracy.

Description

[Detailed Description of the Invention] [Industrial Application Field] The present invention relates to a character recognition method for recognizing characters, marks, etc. (hereinafter simply referred to as characters, etc.), and in particular a character recognition method capable of highly accurate character recognition. Regarding the method.

[Conventional technology]

Conventionally, as a method of this kind, for example, after cutting out a character line or character string (hereinafter simply referred to as a character line) from a document image, what looks like characters is extracted as a temporary character, and a predetermined character size is set as a standard from this temporary character. as full-width characters (kanji, etc.)
If a character cannot be determined to be a full-width character based on the font size, it can be checked to see if it can be effectively used as a half-width character, and then recognized as a half-width character. However, with this method, depending on the character, a full-width character may be mistakenly treated as two or more half-width characters. There is a risk of being made 21. Therefore, the applicant has proposed the following method (Japanese Patent Application No. 63-292445; also referred to as Proposed Method 1).
.

FIG. 6 is a flowchart for explaining proposed method 1.

This shows the processing procedure of an image processing device (not shown). First, document image data is input (see (2)), and each character line is cut out by taking its horizontal projection value (see (2)). As a result, the width of the line is determined, and the amount (character size) corresponding to the size of a full-width character is obtained. Note that although horizontal writing is assumed here, the same applies to vertical writing.

Next, by examining the projection value in the direction perpendicular to each line and taking the character size into account, we cut out what appears to be characters, that is, a group of temporary characters, from each line of characters (see ■). Select full-width characters using the above font size (see ■).

The conditions for selecting characters as full-width characters are as follows.

b) The font size is full-width font size by itself, that is, there is no room for combining with other temporary characters.

oral) punctuation.

c) It is a half-width size by itself, but becomes a full-width size when combined with other adjacent half-width temporary characters.

2) Although the size alone is smaller than the full-width size,
There is too much distance between adjacent half-width temporary characters, and if you forcefully combine them, the size will exceed the full-width character size.

After selecting all full-width characters according to the above conditions, merge or separate the remaining temporary characters (refer to 1 integrated character 1 separated character (0)), and then combine these integrated characters 1 separated character. Characters are recognized using an OCR (character reading device) based on their similarity to dictionary patterns (see ■).

Next, the following contradiction processing is performed on the recognition result (see ■).

a) For example, even though the target to be recognized is a separated character, if the recognition result by OCR indicates a full-width kanji character, this recognition result is not adopted because it contradicts each other.

b) Contrary to the above, when the recognition result by OCR shows half-width size characters such as alphabets and numbers, even though the target to be recognized is an integrated character.

Then, regarding the last remaining character, whether it should be used as an integrated character or a separated character is determined by OCR using relative similarity (see ■).

Note that the degree of similarity xl is defined as the ratio of the degree of similarity X and the average value m of the degrees of similarity multiplied by a certain constant (for example, 1024). That is, xl=x/m constant number (1024
).

In the above method, recognition results are determined based on the degree of similarity, so for example, when characters are very similar, such as ``u'' and ``ro'', or ``te'' and ``chij'', slight differences in the character images may be detected. , for example, there is a risk of misrecognition due to differences in the thickness and font of the characters, crushed or blurred characters, etc. Also, for the same reason, the relative similarity of ``(j and yo'' is more important than the relative similarity of ``wa''). The relative MifJ degree of the part may be higher and f4 may be recognized.This will be explained in detail below.

FIG. 7 shows the similarity X and relative similarity xi of the input character "ha" when it is processed as an integrated character 1 and when it is processed as a separated character 2.3. That is, assuming that the similarity X and the average value m of the similarity are obtained as shown in the figure, the relative similarity obtained by multiplying the ratio of the similarity and the average value by a constant number (1024) is the integrated For character 1, r954. , in the case of separator characters 2 and 3, respectively "10283."898.The average value is "96.
3", and since 954<963, the input character "wa" will be misrecognized as consisting of "(" (front parenthesis) and "hayo".The constant is not limited to "r1024". Of course.

In order to solve these problems, the applicant has further proposed the following method (Japanese Patent Application No. 1-39308; also referred to as Proposed Method 2).

FIG. 8 is a flowchart for explaining proposed method 2, in which the shape feature matching step [
It is characterized by the addition of [phase] to [phase].

The key point of this method is that the aspect ratio of the character to be recognized and the shape features that indicate how many parts can be separated in the vertical and horizontal directions are also stored in advance as a dictionary, and the shape features of the input character are used as the recognition results. The present invention is characterized in that it is determined whether or not to adopt a recognition result based on the degree of similarity by comparing the shape feature amount of the corresponding candidate character with the previously stored shape feature amount.

This example will be specifically explained below.

In Figure 9, the input image (or input character) is ``u'', and the shape feature obtained from the human image is ``vertically long and separated into two vertically'', and the recognition result candidate character This is an example in which the first place among the first to tenth places is "ro". In this case, the shape characteristics of "ro" stored in the dictionary in advance are "neither vertically nor horizontally long, nor separated vertically and horizontally."

Therefore, the step of vertical/horizontal check (Fig. 8 [phase]
), the second candidate character is checked.The second candidate character is "u", and since the shape features match, this is adopted as the recognition result.

Figure 10 shows an example where the input character "wa" is processed as an integrated character and when it is processed as a separate character r (,, ryo). In other words, it is processed as an integrated character as shown in (a) of the same figure. In this case, the first candidate "ha" in the recognition result is also suitable from the point of view of the shape feature, so "haj" is adopted.Similarly, the separated character "(" is the first candidate as shown in (b) of the same figure. Position candidate character ``
However, as shown in the same figure (c), the separated character ``yo'' does not match the first candidate character ``ha'' due to its shape features, and in the end, the second candidate character ``maJ'' Compatible with

As shown in the same figure (d), the relative similarity when recognized as "wa" is r954, and the relative similarity when it is separated into "(" and "ma" is rc+46). , 946<954, then 1 is adopted.In other words, the input character is recognized as ``ha'', and the possibility of it being mistakenly recognized as ``('' and ``ma'' can be eliminated).

[Problem to be solved by the invention]

Although the above shape features are effective under limited conditions, −
It is difficult to use it for boat fishing. For example, since the aspect ratio of a character differs depending on the font, the aspect ratio must be twice or more, which limits the types of characters that can be applied, such as "-" and "1". For example, the character images of the katakana ``口'' and the kanji ``mouth'' are similar, but when comparing the sizes, the kanji ``日二'' is larger; ``The same is true for the kanji ``ko j'' and the kanji ``r''.

Also, for lowercase and uppercase letters, r(, and rc4.r
K'' and rk,, r□, and 70;, rp.

and 7PJ, 'Sj and 'SJ, 'UJ and rtB, rV
Goto'v2, rJ, and rWJ, 7XJ and "X-1"Z
: and zB are similar. Furthermore, the English letter ``I'', the kanji ``工'', and the kanji ``82'' and ``r日2'' become similar when the aspect ratio is changed.

On the other hand, regarding the number of separations, ``2j is definitely 1@2 in one direction.
"to separate into two", but "wa" does not always mean "to separate into two laterally". In other words, it is not possible to reliably specify the number of separations depending on the character type.

In this way, the number of separations is affected not only by noise in the character image, but also by the inclination of the characters, the size (series) of each character, the deformation rate, etc., and thus the applicable character types are limited.

Therefore, an object of the present invention is to improve the shape features so that they can be widely applied, eliminate the above-mentioned drawbacks, and improve the recognition rate.

[Means to solve the problem]

When recognizing characters by image processing a document, cutting out individual characters, and determining the similarity of each character with a dictionary pattern, we calculate the character width, character height, circumscribed rectangle area, and aspect ratio of the character to be recognized. By determining the ratio of at least one of the characters to standard characters in the document containing the character to be recognized and comparing this with a predetermined reference value, the glyph shape of a certain character can be determined from the glyph shape of another character. To facilitate discrimination when similar to characters obtained by expanding or compressing in the vertical direction, horizontal direction, or both vertical and horizontal directions. At this time, find the average value, standard deviation, or amount derived from these for each ratio, and
Reference values for each of these can be determined in advance and stored in a predetermined memory as a dictionary.

In addition, when recognizing characters by image processing a document, cutting out individual characters, and determining the similarity of each character with a dictionary pattern, we can calculate the number of vertical and horizontal separations of the characters to be recognized based on the character size. By predicting and comparing this with a reference value, it is easy to determine when a certain character has a shape that is likely to become a separated character. At this time, the average value of the number of vertical separations and the number of horizontal separations.

In addition to determining standard deviations or quantities derived therefrom, reference values for each of them can be determined in advance and stored in a predetermined memory as a dictionary.

[Effect]

In general, for example, it is difficult to distinguish between the katakana ``kuchi'' and the kanji ``ichikuchi'' based on the characters alone, except in Mincho fonts. However, if we can compare the two characters ``guchi'' in katakana and ``mouth'' in kanji, the smaller one is the katakana ``guchi j'' and the larger one is the kanji ``mouth''. When reading a general document, such direct comparison is not possible, but it is possible to determine whether a character belongs to a larger or smaller size by comparing it with other characters. That is, by knowing the standard character size (including the deformation rate) of the character and quantitatively comparing the sizes of character images, it is possible to accurately determine similar characters.

Also, as the font size decreases and approaches 0, the character image of any character becomes crushed and lumped, so the number of separations converges to 1, and as the font size increases, the characters become less likely to be crushed and the number of separations increases. As the number of separations increases further, the number of separations converges to the maximum value that can be separated for that character type. Therefore, by regarding the separated characters as a continuous function of the character size and determining how much the number of separated characters deviates from the expected value, it is possible to accurately determine whether the integrated character is a separated character or not.

In other words, the character width and height of the character to be recognized.

Predict the area of a circumscribing rectangle, the ratio of at least one of the aspect ratios to the corresponding amount of standard characters in a document containing the character to be recognized, or the number of vertical and horizontal separations from the character size. This enables more accurate recognition.

〔Example〕

FIG. 1 is a flowchart showing an embodiment of the invention. Incidentally, steps ① to ① are the same as those of the conventional method, so steps ② to [phase] will be explained below.

First, the normalized character height residual Hn defined by the following formula
, and find the residual Wn of the normalized character width (■, [
phase]).

22. H and W are the character height and 5 character width measured from the character image. FIG. 2(a) shows sample data of the input image. Hs and Ws are the standard character height and standard character width of the input document (two paragraphs of text, one line), and the data is shown as document data in FIG. 2 (b). The standard character height and standard character width of a certain font are the maximum character height and maximum character width of other fonts, and can be thought of as the maximum character height and maximum character width of Kanji characters of that font. Note that the standard font size is used to absorb errors caused by differences in font, deformation rate, etc., and if a stable value can be obtained, it is not limited to the maximum value, but can also be the average value or mode. Of course it's a good thing. The standard character height and standard character width of the input document can be specified externally before reading the document, or can be statistically determined from surrounding characters during reading, so their details will be omitted.

In addition, for the standard character height Hs and standard character width Ws, the character series is Q1, the vertical deformation rate is αV, the horizontal deformation rate is αh, and the ratios between the letter face and the original standard character height and standard character width are βV and βh, respectively. Then, we have the following relationship.

Hs=β7. Q, αV...(3)W
s = βh− to lαh = (4jHμj
, Hσj represent the average value and standard deviation of the ratio of character type j to the standard character height for each font, and W, czzj.

Wσj represents the average value and standard deviation of the ratio of character type j to the standard character height for each font. How to obtain these values will be explained below.

For example, when comparing Gothic and Mincho fonts with the same character type in the same series, Mincho fonts are generally larger. Katakana “
In order to distinguish between 口 j and the kanji character ``口'' based on their size, it is desirable to correct this difference in character size between fonts. So, for example, "a" in Gothic grade 13,
Find the value divided by the standard character height of Gothic grade 13.

Next, calculate the value of "A" in the 16th grade Mincho typeface divided by the standard character height of the 16th grade Mincho typeface. In this way, one after another,
Find the ratio of standard character height for each character cent (font) for ``A'', and calculate the average value ``HμA'' and the standard deviation ``A'' from these.
Hσa: is required. These Huj, Hσj, Huj
, Wσj are calculated while creating a reading dictionary prior to reading the document, and are stored in a table as shown in FIG. 3 so that they can be easily retrieved by character type j. Note that although FIG. 3 shows the feature amount for character type 1 in detail, the same applies to other character types. Furthermore, although character height and one character width are used in the above example, a circumscribing rectangle and aspect ratio equivalent to this may be used.

Next, the residual difference NVn of the normalized number of vertical separations and the residual difference NHn of the normalized number of horizontal separations are determined as defined by the following equation (see ■, a2).

NV σ j Second, NV and NH are the number of vertical separations and the number of horizontal separations measured from the input character image, and these are the average of the number of vertical separations 2Nμj=F1 (L H) Standard of the number of vertical separations Deviation2 NVσj=F2 (j, H) Average number of horizontal separations: NHμj=F3 (j, W) Standard deviation of number of horizontal separations: NHtjj=F・4 (j, W) for character type j and character width W1 is a function of character height H.

For example, the horizontal separation number of "ha" in Gothic grade 10 is "2".
'', the number of horizontal separations for ``ha'' in the 10th grade textbook is ``1'', and the number of horizontal separations for ``ha'' in the 10th grade Mincho font is F2'', then the average value and standard deviation are, NHμ is =F3(is, 10
class) =, (2+1+2)/3 = 1.667 NHσ = F4 (is, 10th class) = [(4+1+4)/3- (1,667)”] ”
2=0.47. In this way, rNHμ is determined to be J, and rNHσ is determined to be C for more fonts in grade 10. Next, 1
Similarly for the secondary level, rNHμ is j + rNHσ
After that, repeat the same operation and use the interpolation method to obtain Fl (ha, H), F2 e ma H), F3 (ha, W)
, F4 (is W).

Then, if you perform this operation for all character types,
Function Fl (j, H), F2 (j, H), F3 (j
, W) , F4 (j, W) will be found.

Figure 4 shows the average function of the number of vertical separations and its standard deviations F1 and F2.
and the average function of the number of lateral separations and its standard deviation F3. An example of F4 is shown.

The average function Fl (j, H) of the number of vertical separations is a function of the character type j1 character height H as shown in the same figure (a), and (a 3
It is expressed as (a 2 ≦H) (7). The standard deviation function F2 (j, H) is obtained from the same figure (b) as [6 (H≦bl) NVσ=J, (b5-b3) (H-b2)/ (b
...(8) It is expressed as: Similarly, it is expressed as ... (9) ... (10).

In the next step @, the norm (corresponding to the length of the pattern) defined by the following equation is determined.

D j = [Hn”+Wnz+NVn”+NH
n2 co l/2 (11) In the above, the norm is expressed using four quantities: Hn, Wn, NVn, and NHn, or it can be expressed using one or more of them.

In step [phase], the result of similarity determination is integrated to obtain a comprehensive determination result Tj of character type j. That is, the similarity Rj, which is a correlation value, is combined with the shape feature result Dj as shown in the following equation.

Tj=wl ・Rj+w2・ [max (A"Dj"
, O) ko 1″ ... (
12) Here, wl and w2 are constants indicating the weight of the similarity Rj and the shape feature result Dj, and A is a constant for giving a limit to the effect of the shape feature. Note that the above equation (12) can be simplified as Tj=W1.Rj+w2.Dj (13).

When using shape features as an aid to similarity, in other words, when there are no similar glyphs, and the similarity alone is enough to determine the character type, and shape features are not required, the value of the large number of standard deviations mentioned above should be set larger than the actual value. The value of Dj becomes very small, and the effect of the shape feature can be weakened.

Note that in order to express the separated characters in more detail, the characters can also be expressed as a combination of partial figures and blank spaces. For example, as shown in Figure 5 (a), the kanji ``zu'' is expressed as a list of six partial figure sizes and five blank sizes as shown in (b) and (c) in the same figure, and the standard character size is If the size of the blank space becomes less than 1 pixel when normalized with , it will collapse, and pattern matching may be performed between the lists. However, since it is a probabilistic problem as to which blank space is to be collapsed, the pattern matching method becomes complicated.

〔Effect of the invention〕

With this invention, standard character height Hs, standard character width Ws
is the amount to correct the difference in character size between fonts, and 1
Since the two documents use the font ``Rima-'', it is possible to determine the difference in character size between the fonts in detail and accurately. This allows the glyph shape of one character to vertically change the glyph shape of another character.

Even when similar characters are obtained by expanding or compressing in the horizontal direction or both vertical and horizontal directions, it is possible to distinguish between the two. Furthermore, by incorporating the average value and standard deviation for each character type into the evaluation criteria for evaluating size variations, the determination becomes more accurate.

As a result, the character images of the kanji ``mouth'' in katakana are similar, but when comparing the sizes, the kanji ``mouth'' is larger; ``Technology''
The same is true for the kanji ``工''. In addition, lowercase and uppercase letters are "C", "Cl, rK", rk,, rO"
and ro'', rp'' and 'PJ, r3. and "S", "U"
The same applies to "and ru", rV, and "■", "W" and "W", "X" and "X", and "Z" and "z". In addition, the alphabet “
The kanji ``I'' and the kanji ``工'' and the kanji ``日'' and ``日'' become similar when the aspect ratio is changed, but it becomes easier to distinguish the similar characters as described above.

In addition, the method of separating characters can be predicted based on the font size.
If the prediction result is compared with a reference value for determination, it is possible to more accurately determine whether characters are integrated into two or separated. In addition, we incorporated the average value and standard deviation of the separation method for each type of character into the evaluation criteria for evaluating the variation in size.
If this is used as an approximation formula or function for character size, the determination will be more accurate.

As a result, for character types that are not always separated into two horizontally such as ``r'', as well as ``j'' that is reliably separated into two vertically, such as ``ni''. Also, it becomes easy to make an integrated and two-separate judgment.

[Brief explanation of drawings]

FIG. 1 is a flowchart showing an embodiment of the invention, and FIG.
The figure is an explanatory diagram for explaining sample data and document data, FIG. 3 is an explanatory diagram for explaining dictionary data,
Figure 4 is an explanatory diagram to explain the average and standard deviation of the number of vertical separations and the number of horizontal separations; Figure 5 is an explanatory diagram to explain an example of a certain character and its character separation list; Figure 6 is a flowchart for explaining proposed method 1, FIG. 7 is an explanatory diagram for concretely explaining proposed method 1, and FIG. 8 is a flowchart for explaining proposed method 1.
The flow chart for explaining the proposed method 2, and FIGS. 9 and 10 are explanatory diagrams for specifically explaining the proposed method 2. Code explanation 1.--Integrated character, 2.3.--Separation character.

Claims (1)

[Claims] 1) When recognizing characters by image processing a document to cut out individual characters and determining the degree of similarity with a dictionary pattern for each character, the character width, character height of the character to be recognized, area of circumscribed rectangle,
For at least one of the aspect ratios, the ratio of each corresponding amount of standard characters in the document containing the character to be recognized is determined, and this is compared with a predetermined reference value to determine the character of a certain character. A character recognition method, characterized in that it facilitates discrimination when a character shape is similar to a character obtained by expanding or compressing the shape of another character vertically, horizontally, or both vertically and horizontally. 2) An average value, a standard deviation, or a quantity derived from these of each of the ratios is determined, and a reference value for each of the ratios is determined in advance and is stored in a predetermined memory as a dictionary. ) Character recognition method described in 3) Image processing a document to extract individual characters and calculate the similarity of each character with a dictionary pattern to predict the number of vertical and horizontal separations of characters to be recognized based on the character size. A character recognition method characterized in that by comparing this with a reference value, it is easy to determine when a certain character has a shape that is likely to become a separated character. 4) The average value and standard deviation of the number of vertical separations and the number of horizontal separations, or quantities derived from these are determined, and reference values for each of them are determined in advance and stored as a dictionary in a predetermined memory. The character recognition method according to claim 3).