JPH0696285A - Character recognizing device - Google Patents

Abstract

PURPOSE:To provide the character recognizing device being excellent in reliability, which can obtain a correct solution character, even in the case where two adjacent half size characters are segmented erroneously as one full size character, at the time of obtaining character image data by segmenting a character with respect to image data. CONSTITUTION:The character recognizing device is provided with a character segmenting part 2 for segmenting image data to the image data of every one character, a character recognizing part 3 for recognizing the image data and converting it to a code, a language processing part 4 for executing a language processing to a result of recognition of the character recognizing part 3, a suspected word extracting part 5 for extrancing a word being a one-character word, and also, consisting of characters which can be separated into the left and the right as a suspected word from in a word group, a suspected word resegmenting part 6 for segmenting the image data of the suspected word as two half size characters, a suspected word character recognizing part 7 for recognizing its image data and converting it to a code, a language reprocessing part 8 for executing a language processing to a result of recognition after the resegmentation, and a deciding part 9 for comparing the result of recognition by the language processing part and the result of recognition by the language reprocessing part and determining a final solution character-string.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention is used as an input device for a computer or the like, and character-recognizes patterns such as print characters, dot characters and handwritten characters in newspapers, magazines, novels, etc.
The present invention relates to a character recognition device that converts code information such as an S code.

[0002]

2. Description of the Related Art In recent years, with the rapid development of information devices such as computers, digitization of documents and the like has come to be performed for the purpose of improving workability, and patterns such as printed characters are recognized and converted into character codes. The character recognition device has become widely used as an input device such as a computer.

A conventional character recognition device will be described below. A character recognition method of the conventional character recognition device will be described based on the recognition target document shown in (Table 1).

[0004]

[Table 1]

First, a document to be recognized is read by an image reading device to obtain image data.

Next, character image data for each character is cut out from the image data. Next, the character image data is character-recognized to obtain a character recognition result. The result of character recognition for the original sentence shown in (Table 1) is shown in (Table 2).

[0007]

[Table 2]

Here, since the adjacent two half-width characters '(', '3' in the original sentence are erroneously cut out as one full-width character, "(3" in the original sentence is changed to "line", "forward". ",
It is mistakenly recognized as "diameter".

Next, language processing is performed based on the character recognition result of (Table 2). First, each candidate character is combined to form a character string. This character string is queried with the word dictionary and only the matching character string is set as the candidate word. The candidate words thus extracted are shown in (Table 3).

[0010]

[Table 3]

Next, the candidate words are combined to obtain a clause according to grammatical rules. In this example, the longest matching word is adopted as the word composed of a plurality of characters, and the first candidate character is adopted as the candidate word composed of one character.

The adopted clauses are shown in (Table 4).

[0013]

[Table 4]

By the above processing, "line) character recognition" is obtained as a solution character string. In this way, since the adjacent two half-width characters in the original sentence are erroneously cut out as one full-width character, "(3" in the original sentence is erroneously recognized as a "line".

[0015]

However, in the above-mentioned conventional method, when the character data of the image data is cut out to obtain the character image data, two adjacent one-byte characters are accidentally cut out as one full-width character. In addition, since the erroneous recognition cannot be relieved by the subsequent processing, there is a problem that the recognition rate is lowered and the reliability is poor.

The present invention solves the above-mentioned conventional problems. When character data is cut out from image data to obtain character image data, two adjacent one-byte characters are mistakenly cut out as one full-width character. An object of the present invention is to provide a highly reliable character recognition device that can obtain a correct solution character even in the case.

[0017]

In order to achieve this object, a character recognition apparatus of the present invention is a character slicing section for slicing image data read by an image reading section into character image data for each character, and the character slicing section. A character recognition unit that character-recognizes character image data cut out by a unit and converts the character image data into a character code, and character recognition that performs language processing on the character recognition result recognized by the character recognition unit and is divided into character units. A language processing unit that changes the result into word units, and a suspected word extraction unit that extracts, as a suspected word, a word that is a single-character word and is composed of separable left and right words from the word group obtained by the language processing unit. The suspect word extracted by the suspect word extracting unit is likely to have erroneously recognized adjacent half-width characters in the recognition target document as one full-width character. Suspicious word re-cutout unit that cuts out the image data of the suspected word as two half-width characters, and the suspected word character that character-recognizes the character image data cut out by the suspected word re-cutout unit and converts it into a character code. A recognition unit, a relanguage processing unit that performs language processing on the character recognition result after re-cutting by the suspected word character recognition unit, a character recognition result by the language processing unit, and a character recognition result by the relanguage processing unit. And a determination unit that determines the final solution character string by comparing.

[0018]

With this configuration, even if two adjacent one-byte characters are accidentally cut out as one full-width character by extracting the suspected word and re-cutting the suspected word, etc., a correct solution character string is obtained. You can

[0019]

DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the present invention will be described below with reference to the drawings.

FIG. 1 is a block diagram of a character recognition apparatus according to an embodiment of the present invention. 1 is an image reading unit that photoelectrically converts a recognition target document to obtain image data, 2 is a character cutting unit that cuts the image data read by the image reading unit 1 into character image data for each character, and 3 is a character cutting unit 2. The character recognition unit 4 for recognizing the cut out character image data and converting it into a character code is a correct character by creating a clause according to the grammatical rule based on the character recognition result recognized by the character recognition unit 3. The language processing unit 5 selects a suspected word extraction unit that extracts a word that is a single-character word from the word group obtained by the language processing unit 4 and that is composed of characters that can be separated left and right as a suspected word, and 6 is a suspected word. The suspected word recutting unit 7 cuts out the image data of the suspected word extracted by the extraction unit 5 as two half-width characters, and 7 is cut by the suspected word recutting unit 6. A suspicious word character recognizing unit that character-recognizes the output character image data and converts the character image data into a character code; 8 is a re-language processing unit that performs language processing on the character recognition result after re-cutting by the suspicious word character recognizing unit 7; A determination unit 9 compares the character recognition result by the language processing unit 4 with the character recognition result by the re-language processing unit 8 to determine a final solution character string.

The operation of the character recognition apparatus of the present embodiment having the above-mentioned configuration will be described below based on the recognition target document shown in (Table 1) as an example.

FIG. 2 is a flow chart of an image reading unit, a character cutout unit and a character recognition unit of a character recognition apparatus according to an embodiment of the present invention, FIG. 3 is a flow chart of a language processing unit, and FIG. 4 is a suspected word extraction. 6 is a flowchart of the suspected word recutting unit, the suspected word rerecognition unit, and the relanguage processing unit, and FIG.
Is a flowchart of the determination unit.

In FIG. 2, first, the image reading unit 1
At, the recognition target document is photoelectrically converted to obtain image data (S1).

Next, in the character cutting section 2, character image data for each character is obtained from the image data (S2).

Next, the character recognition section 3 performs character recognition based on the character image data (S3).

Here, a concrete character recognition result for the recognition target document shown in (Table 1) is shown in (Table 2).

In Table 2, the upper part shows the recognition result and the lower part shows the similarity. Next, language processing is performed.

In FIG. 3, first, provisional clauses are set by using character type transition (S4). Here, the concrete provisional clause setting result for the character recognition result shown in (Table 2) is shown in (Table 5).
Shown in.

[0029]

[Table 5]

Here, the character type transition is a change point of a character type, and a provisional bunsetsu is set by referring to a predetermined character type transition table as to whether or not it is a break of a bunsetsu. The character type transition table is shown in (Table 6).

[0031]

[Table 6]

In (Table 6), "1" indicates that the provisional phrase set point becomes a break of the phrase, and "0" indicates that the provisional phrase set point does not exist.

Next, the provisional clause number is set to i, and 0 is substituted for i (S5). Next, pay attention to the i-th temporary clause (S6).

Next, the relative character number from the beginning of the provisional phrase is k
Then, 0 is substituted for k (S7). Next, the candidate ranking is j
Then, 0 is substituted for j (S8).

Next, a character string starting from the character in the j-th row and the k-th column in the i-th temporary clause is created (S9). Next, the character string created in S9 is referred to the word dictionary, and only matching ones are extracted as candidate words (S10).

Next, 1 is added to j (S11). Next, it is checked whether j is less than 3 (S12).

If yes, jump to S9. If no, 1 is added to k (S13).

Next, it is checked whether k is less than the number of constituent characters of the i-th temporary clause (S14). If yes, jump to S8.

If the result is no, the candidate words extracted in S10 are combined to form a candidate phrase according to the grammatical rules (S15).

Next, the average similarity of each candidate phrase is obtained (S16). Next, the candidate clause having the highest average similarity is adopted as a solution (S17).

Next, 1 is added to i (S18). Next, it is checked whether i is less than the total number of provisional clauses (S19).

If yes, the process from S6 is performed. If no, the language processing ends.

Here, a specific example of the language processing for the provisional clause setting result shown in (Table 5) is shown below.

First, from the 0th temporary clause to the 1st temporary clause,
Since it consists of one character, the first candidate character is adopted as the solution.

Next, attention is paid to the second provisional bunsetsu "Fun Uori Ori" (S6). Next, a character string having a "sentence" at the beginning is created (S9).

Here, the created character string is shown in (Table 7).

[0047]

[Table 7]

Next, the created character string is referred to the word dictionary and only the matching words are extracted (S10).

Next, the above processing is performed for all the candidate characters (S8 to S14) to obtain the candidate word group shown in (Table 8).

[0050]

[Table 8]

Next, a candidate phrase is constructed by combining candidate word groups (S15). Since'character 'and'recognition' are nouns, they can be connected and the candidate phrase'character recognition 'is obtained. Similarly, since'sentence 'and'recognition' are nouns, they can be connected and the candidate phrase'sentence recognition 'is obtained.

Next, the average degree of similarity of each candidate phrase is calculated, and the larger phrase is adopted as the solution (S16-17).

The average similarity of the candidate phrase “character recognition” is 95.
0. The average similarity of the candidate phrase'sentence recognition 'is 925. Therefore, 'character recognition' is adopted as the solution.

The clauses obtained by the above processing are shown in (Table 9).

[0055]

[Table 9]

Next, the suspected word is extracted. In FIG. 4, first, the phrase number is set to i, and 0 is substituted for i (S2
0).

Next, the relative number from the beginning of the phrase of the words forming the phrase is set to j, and 0 is substituted for j.

Next, pay attention to the j-th word of the i-th clause (S
22). Next, it is checked whether or not the focused word is composed of one letter (S23).

If no, jump to S26. If yes, it is checked whether the focused word is a separated character (S24).

The term "separated character" as used herein refers to a character such as "line", "word", etc., which is composed of "hen" and "make" and can be divided into right and left.

If NO, jump to S26. If yes, the word of interest is the suspected word, and its character image data number is stored in the array doublet (S
25).

Next, 1 is added to j (S26). Next, it is checked whether j is less than the number of constituent words of the i-th clause (S
27).

If yes, jump to S22. If no, 1 is added to i (S28).

Next, it is checked whether i is less than the number of clauses (S
29). If yes, jump to S21.

If no, the extraction of the suspected word ends. Here, a specific example of the suspected word extraction for the language processing result shown in (Table 9) is shown below.

First, the 0th clause is composed only of'lines' (S22). Next, 'line' is a one-letter word (S
23).

Next, "line" is a separation character (S2
4). Next, "line" is the suspected word, and the character image data number of "line", 0, is doublet.

Substitute in [0] (S
25).

Next, the first phrase is not a separating character (S2
4). Therefore, there are no suspect words.

Next, the second phrase is composed of a two-letter word "character" recognition "(S23).

Therefore, there are no suspect words. Next, re-cutting of the suspected word, recognition of the suspected word character, and re-language processing are performed.

In FIG. 5, first, i is the suspected word number, and 0 is substituted for i (S30). Next, the character image data number of the suspected word is set to j, and doubt [i] is substituted for j (S31).

Next, the j-th character image data is re-cut out,
Left and right character image data are obtained (S32).

Next, the two character image data obtained in S32 are recognized as characters (S33).

Next, 1 is added to i (S34). Next, it is checked whether i is less than the number of suspected words (S35).

If yes, jump to S31. If the result is no, the character recognition result before recutting and the character recognition result after recutting are integrated (S36).

Next, the language processing similar to that of FIG. 3 is performed again on the character recognition result integrated in S36 (S37).

Here, a specific example of the re-cutting of the suspected word, the recognition of the suspected word and the re-language process for the suspected word extraction result will be described below.

First, doubt

0 is stored in [0] (S31). This indicates that the 0th character image data is a suspect word.

Next, the 0th character image data is re-cut out to obtain two left and right character image data (S32).

Next, each of the character image data is recognized as a character (S33). The character recognition result at this time is shown in (Table 1
0).

[0081]

[Table 10]

Next, the recognition result before the recutting and the recognition result after the recutting are integrated (S36).

The integration result at this time is shown in (Table 11).

[0084]

[Table 11]

Next, language processing is performed again on this integrated result (S37). The re-language processing result at this time is shown in (Table 1
2).

[0086]

[Table 12]

Next, a judgment is made. In FIG. 6, first,
i is the suspected word number, and 0 is substituted for i (S38).

Next, the image data number of the suspected word is set to j, and doubt [i] is substituted for j (S39).

Next, attention is paid to the bunsetsu before the re-cutout and the bunsetsu after the re-cutout including the recognition result of the j-th character image data (S40).

Next, the certainty of both focused bunsetsus is compared, and the bunsetsu which becomes a solution is determined (S41).

The certainty is determined by using the knowledge that the words that form a phrase appear in a general document and that characters such as parentheses are often used in pairs.

Next, 1 is added to i (S42). Next, it is checked whether i is less than the number of suspected words (S43).

If yes, jump to S39. If the answer is no, all processing is terminated.

Here, a concrete example of the judgment for the re-language processing result shown in (Table 12) is shown below.

First, doubt

0 is stored in [0] (S39). Next, including the 0th character image data,
Pay attention to the phrase "line" before the re-cutout and the phrase "(3" after the re-cutout (S40).

Next, the probabilities of both clauses are compared (S
41). The bunsetsu'line 'has no relation to the adjacent bunsetsu. On the other hand, the clause '(3' includes a left parenthesis.
With the knowledge that brackets are often used in pairs,
When a right parenthesis was not found in other bunsetsu, a right parenthesis ')' was found in the adjacent bunsetsu, and bunsetsu '(3') was related to other bunsetsu.

Therefore, it is judged that the certainty of the bunsetsu after the re-cutout is greater, and that one is adopted as the solution.

By the above processing, "(3) Character recognition" is obtained as the final solution character string, and the recognition target document can be correctly recognized.

[0099]

As described above, according to the present invention, when the suspected word is extracted and the suspected word is re-cut out, two adjacent one-byte characters are accidentally cut out as one full-width character. However, it is possible to realize a highly reliable character recognition device that can correctly recognize the original text.

[Brief description of drawings]

FIG. 1 is a block diagram of a character recognition device according to an embodiment of the present invention.

FIG. 2 is a flowchart showing a control procedure of the image reading unit, the character cutting unit, and the character recognition unit of the character recognition device according to the embodiment of the present invention.

FIG. 3 is a flowchart showing a control procedure of a language processing unit of the character recognition device in the embodiment of the present invention.

FIG. 4 is a flowchart showing a control procedure of a suspected word extraction unit of the character recognition device in the embodiment of the present invention.

FIG. 5 is a flowchart showing a control procedure of a suspected word recutting unit, a suspected word rerecognition unit, and a relanguage processing unit of the character recognition device according to the embodiment of the present invention.

FIG. 6 is a flowchart showing a control procedure of the determination unit of the character recognition device in the embodiment of the present invention.

[Explanation of symbols]

 1 image reading unit 2 character cutting unit 3 character recognition unit 4 language processing unit 5 suspected word extraction unit 6 suspected word recutting unit 7 suspected word character recognition unit 8 relanguage processing unit 9 judgment unit

Claims (1)

[Claims] 1. A character cutout unit for cutting out image data read by an image reading unit into character image data for each character, and a character for recognizing the character image data cut out by the character cutout unit and converting it into a character code. A recognition unit, a language processing unit that performs language processing on the character recognition result recognized by the character recognition unit, and changes the character recognition result divided into character units into word unit divisions; and the language processing unit. A suspect word extracting unit that extracts a word consisting of a character that is a single-character word and that can be separated into left and right from the obtained word group, and the suspect word extracted by the suspect word extracting unit is a recognition target document. It is judged that there is a high possibility that the adjacent half-width characters were mistakenly recognized as one full-width character, and the image data of the suspected word is cut out as two half-width characters. A word recutting unit, a suspected word character recognition unit that character-recognizes the character image data cut out by the suspected word recutting unit and converts it into a character code, and a character recognition result after recutting by the suspected word character recognition unit. A re-language processing unit that performs language processing on the above, and a determination unit that compares the character recognition result by the language processing unit with the character recognition result by the re-language processing unit to determine a final solution character string. A character recognition device characterized by: