JPS60138689A - Character recognizing method - Google Patents

Abstract

PURPOSE:To avoid the misdecision for recognition of characters by retrieving successively plural dictionaries with the feature variable constant of an input character to detect a character most approximate to the feature variable constant and specifying a font from the distance between the variable constant of said detected character and the feature variable constant. CONSTITUTION:The character number for discrimination of font is limited to Nmax, and a dictionary DCalpha is first retrieved when a feature variable constant Pn is supplied. Then a character most approximate to the Pn is detected in the dictionary DCalpha. Then the distance between a variable constant of the retrieved character and the Pn is obtained and added to an adder ADalpha. Then a dictionary DCbeta is retieved to obtain a character having the variable constant most approximate to the Pn. The distance between the variable constant of the obtained character and the Pn is obtained and added to an adder ABbeta. In the same way, the similar distance is obtained also for an adder ADgamma. The contents of these adders ADalpha, ADbeta and ADgamma are divided by the contents of counters CNalpha, CNbeta and CNgamma to obtain values Dalpha, Dbeta and Dgamma respectively. Then a font corresponding to the minimum one of said three values is specified.

Description

DETAILED DESCRIPTION OF THE INVENTION [Technical Field] The present invention relates to multi-font character recognition that automatically performs character recognition for a plurality of types of characters with different fonts, and particularly to prevention of misrecognition.

[Prior art] For English letters, for example, elite letters and biryoku letters.

There are various typefaces such as italic characters and OCR characters. Therefore, in multi-font character recognition in which characters of these various fonts must be recognized, dictionaries corresponding to each font are provided. Then, for each recognized character, dictionaries of all character fonts are searched, and the one with the closest characteristic parameters is selected as a candidate, and the closest one among them is made the final candidate.

However, in this type of multi-font character recognition, there are characters that are similar to each other between different fonts.
Some characters may have similar shapes and differ only in size, or may be completely the same. In such a case, the distance between the recognized character and the data in the dictionary will be the same for a plurality of dictionaries with different fonts, and erroneous recognition that may lead to incorrect font selection is likely to occur.

One of the conventional techniques is to count the number of times an unreadable code has been output for each font and use that result to identify the recognition result.However, between fonts with similar parameters, this value is compared. can't identify the font either.

When erroneous font recognition occurs, for example, in order to shorten search time, erroneous recognition occurs frequently in a character recognition system in which a dictionary to be used thereafter is specified based on the result of initial character recognition.

[Purpose] The present invention aims to reliably identify fonts and prevent misrecognition.

[Structure] If a series of character strings to be recognized is composed of only characters of one type of font, for example, the dictionary containing the data with the smallest distance from the feature parameters of the characters to be recognized is the one that Obviously one. In other words, for multiple character patterns sufficient to identify a font, if you search each dictionary to find the closest data, and find the average distance value for each dictionary, you can find the dictionary with the smallest value. can be determined to be a font of a series of character strings.

The average distance value is 1. For example, it can be obtained by calculating the distance D n (n = 1-N) for each of N characters, calculating the sum, and dividing the result by N.

However, in this case, if a character is erroneously recognized, the result is meaningless and results in a large error with respect to the sum of distances, which is the standard for font determination. Therefore, if a large weight is given to the distance of a font whose recognition result is correct, this type of error can be reduced.

Generally, when searching multiple font p-dictionaries, the same result is often obtained for multiple fonts.
The results are often correct.

Therefore, in one preferred embodiment of the present invention.

From the results obtained by searching the dictionary for each font, that is, the first candidate character, a second candidate character is determined by majority vote, and distance data is calculated for fonts for which the first candidate character that is the same as the second candidate character is obtained. Increase the weighting (that is, the contribution rate to font determination).

Hereinafter, the present invention will be described in detail with reference to the drawings.

FIG. 1 shows a schematic configuration of a character recognition device implementing one aspect of the present invention. Referring to FIG. 1, character information on a document (not shown) is optically read by a scanner SCA, and photoelectrically converted into binary data for each minute pixel. This data is applied to the feature extraction unit PPU after being subjected to a predetermined process in the preprocessing unit PPR for each character pattern. The feature extraction unit PPU extracts various feature parameters of the character pattern as numerical data.

The search control unit CTR controls each unit, and the feature extraction unit PPU
The dictionary is searched for the character closest to the characteristic parameter for each character extracted in , and this is used as the recognition result. In this example, three types of character fonts α, β, and T are handled, so the respective font dictionaries DCα, DCβ, and DC
γ is available. The search control unit CTR includes a distance calculation unit c that calculates the distance between the feature parameters output by the feature extraction unit PPU and the feature parameters obtained from each dictionary.
: PD, buffer memory BFF, three adders ADα,
ADβ and ADγ and three counters CNα, CNβ and CNγ are connected.

FIG. 2 shows a part of the font discrimination procedure, that is, the operation of the search control unit CTR shown in FIG. This will be explained with reference to FIGS. 1 and 2. First, adders ADα, ADβ and AD
γ, and the contents of counters CNα, CNβ, and CNγ. A counter N for counting the number of characters is set to an initial value, and is incremented by 1 each time a feature parameter is input from the feature extraction unit PPU.

Here, the number of characters used for font identification is limited to Nmax (for example, 10). When the feature parameter Pn is input, the dictionary DCα is first searched to find the character closest to Pn (the first candidate character for α). The distance between the parameter of the searched character and Pn is determined and added to the adder ADα.

Next, the dictionary DCβ is searched to find a character whose parameter is closest to Pn (first candidate character for β). The distance between the parameter of the searched character and Pn is determined, and the result is added to the adder ADβ.

Similarly, the dictionary DCγ is searched for the character whose parameter is closest to Pn (the first candidate character for γ), the distance between the parameter and Pn is calculated, and the result is added to the adder ADγ.

From the three first candidate characters obtained in this way,
The second candidate character is determined by majority vote. That is, for example α
If the first candidate characters of α and γ are the same, and the first candidate characters of β are different from each other, the first candidate characters of α and γ are set as the second candidate characters. If the three first candidate characters are all different, it cannot be decided by majority vote, so if the characteristic parameter P is the highest among the three first candidate characters,
The one closest to n is selected as the second candidate.

The second candidate characters obtained in this way are compared with each first candidate character. If the second candidate character and the first candidate character of α are equal, the counter CNα is incremented by 1, but if they are different, the counter CNα is left unchanged. Similarly, the second candidate character and the first candidate character of β are compared, and if they are equal, the counter CNβ is incremented.
1, and if the second candidate character and the first candidate character of γ are equal, the counter CNγ is incremented by 1.

Further, the contents of the feature parameters output from the feature extraction unit PPU are temporarily stored in the buffer memory BFF for later use.

If we continue the above operation until N > N max,
Adders ADα, ADβ, and ADγ store the sum of feature parameter distances between the first candidate character obtained from the dictionaries of fonts α, β, and γ, respectively, and each of Nfflax recognized characters, and counters CNα, CNβ and CNγ store the number of correct characters of fonts α, β, and γ, respectively.

Therefore, when Nmax readings are completed, the value Dα is obtained by dividing the content of the adder ADα by the content of the counter CNα,
A value Dβ obtained by dividing the contents of the adder ADβ by the contents of the counter CNβ and a value Dγ obtained by dividing the contents of the adder ADγ by the contents of the counter CNγ are calculated.

However, if the secret of the counter is 0, it is ignored.

Each value Dα, Dβ and Dγ is the font α
, β and γ, and the parameters of Nmax recognized characters. For example, if the font of the character to be recognized is α, and the same character (correct character) is recognized in three dictionaries, the dictionary that provides data with the smallest distance from the feature parameter Pn is DCα. It is. Therefore, the dictionary font with the smallest average value of Nmax distances matches the font of the recognized character.

If the correct character is obtained in the dictionaries DCα and DCβ, the dictionary,
When DCγ selects a misrecognized character as a candidate, the content of the counter CNγ corresponding to the font γ does not change due to the result of the majority vote.As a result, the divisor for adding the parameter Dγ becomes smaller and Dγ becomes larger. The influence of the distance (error amount) obtained from DCγ on font determination is reduced.

The obtained parameters Dα, Dβ, and Dγ are compared with each other, and the dictionary of the font corresponding to the smallest one is selected. Once the font determination is complete, the already recognized N wa
Character recognition is performed again for x characters using only the selected dictionary. In this case, it takes time to return the scanner to the initial position and perform character data reading, preprocessing, feature extraction, etc.
Perform character igm by reading the feature parameters of each of the Nmax characters stored in Nmax one by one.
After re-recognizing the characters, use the selected dictionary,
Continue character recognition in the normal way.

FIG. 3 shows a modified embodiment of FIG. 2. Referring to FIG. 3, in this embodiment, N
When max characters are recognized, for each character, the feature parameter Pn, the font from which the first candidate that does not match the second candidate is obtained, and the second candidate character are temporarily stored in a buffer memory. Once the font is known, the contents of the papa memory BFF are read and the following processing is performed for each recognized character.

First, it is checked whether the font of the selected dictionary matches the font obtained as the second candidate.

If they do not match, the second candidate stored in the buffer memory BFF is incorrect, so the feature parameter Pn of the character is read and recognition is performed again using the selected dictionary, and the result is the second candidate. Update. If they match, the second candidate stored in the buffer memory BFF is output as is as the recognition result. Therefore, in this embodiment, only characters that have been erroneously recognized are re-recognized, so that the re-recognition time is shortened when the number of characters N1lIax is large.

Next, another embodiment of the present invention will be described. Referring to FIG. 4, in this embodiment, the buffer memory BFF is omitted and three memories Mα2Mβ and Mγ associated with each font are used. With reference to FIG. 5, the recognition procedure of this embodiment will be explained only with respect to the parts that are different from the previous embodiment.

In this embodiment, when recognizing Nmax characters for font identification, the first candidate character of each character is stored in memories Mα2Mβ and Mγ for each font, respectively.

When the font is determined, the content of the memory Mα2Mβ or Mγ associated with the font is read and used as a recognition result. Therefore, in this embodiment, there is no need for re-recognition.

Note that in the above embodiment, another check is performed after the fonts for Nmax characters are determined, but if a sufficient recognition result is obtained by majority judgment or distance judgment, such processing may be performed. may be omitted.

Further, in the embodiment, the characters used for font discrimination are specified as the numerical value Nmax, but the characters may be specified, for example, in one line of character strings or in units of five pages. Also,
In the embodiment, font determination is not performed after first performing font determination for Nmax characters; however, for example, when a page of a manuscript is updated, the font determination may be performed by returning N to the initial value.

Furthermore, in the embodiment, the adder, counter, message, etc. are each made into independent components, but when using a microcomputer, for example, special components can be added by utilizing its own functions. The present invention can be practiced without having to do so.

[Effect] As described above, according to the present invention, for a plurality of characters,
Since the font is determined based on the average distance between the data in each dictionary and the characteristic parameter of the character, misjudgment is eliminated.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing the configuration of a character recognition device implementing one aspect of the present invention. FIG. 2 is a flowchart showing a font discrimination procedure in the apparatus shown in FIG. FIG. 3 is a flowchart showing a modification of the embodiment shown in FIG. FIG. 4 is a block diagram showing the configuration of a character recognition device implementing another embodiment of the present invention. FIG. 5 is a flowchart showing the font discrimination procedure in the apparatus of FIG. 4. DCα, DCβ, DCγ: Dictionaries ADα, ADβ, ADγ: Adders CNα, CNβ, CNγ: Counter BFF: Buffer memory Mα2Mβ2Mγ: Memory

Claims (5)

[Claims] (1) In character recognition where multiple types of character fonts are recognized; multiple characters are recognized, and for each character, a dictionary of different character fonts is used as the first candidate character according to the characteristic parameters of the character to be recognized. At the same time, the distance value between the feature parameter of each first candidate and the feature parameter of the recognized character is calculated and accumulated for each character font, and the most appropriate one is selected from among the first candidates. is selected as the second candidate, calculates a value for each font depending on whether or not there is a match between the first candidate character and the second candidate character, and calculates the value and the cumulative distance value for each font,
A character recognition method characterized by comparing the results between fonts to identify the font. (2) The selection of the second candidate is determined by majority vote from among the first candidates, and if the second candidate cannot be selected, the second candidate is determined to be the one closest to the characteristics of the character to be recognized among the first candidates. A character recognition method according to claim (1). (3) For each of a plurality of characters to be recognized for font identification, characteristic parameters are stored, and after font identification, the characteristic parameters are read and character recognition is performed again for the plurality of characters. A character recognition method according to claim (1). (4) For each of a plurality of characters recognized for font identification, its characteristic parameters, second candidate characters, and whether or not the first candidate character and second candidate character match, and the relationship with the font are stored. If the identified font and the obtained font of the second candidate character do not match, read the feature parameters for that character and perform re-recognition.
A character recognition method according to claim (1). (5) For each of the multiple characters to be recognized for font identification, the first candidate character of each font is memorized, and when the font is identified, the first candidate character of that font is output as the recognition result. , a character recognition method according to claim (1).