JPH08329191A - Character string segmenting method - Google Patents

Abstract

PURPOSE: To provide a character string segmenting method capable of accurately and efficiently extracting a certain character string even when plural character strings concentrically exist. CONSTITUTION: In the character string segmenting method, character candidates are extracted from an input picture (24), two extracted character candidates e.g. are combined to prepare a partial character string to be a constitutional element of a character string (26), a character string candidate is prepared only by combining plural prepared partial, character strings (28), a character string is determined from plural prepared character string candiates (30). Preferably a character string candidate is prepared only when one character candidate is superposed between partial character strings to be combined or between a partial character string and already prepared character string candidate.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method for extracting character string candidates from an image.

[0002]

2. Description of the Related Art In a device for recognizing an image written on a document such as a drawing recognition device or a text reader, the process of cutting out a character string is important. In the conventional character string cutout method, the following procedure is usually taken.

All character candidates are extracted. Whether or not the extracted character candidates can be integrated is checked by character recognition or the like, and if they can be integrated, they are integrated. It is checked whether or not another character candidate can be integrated with a character string candidate composed of n (n is 2 or more) characters, and if they can be integrated, the character string candidate composed of n + 1 character candidates is integrated. To generate. When there are no more character candidates that can be integrated, the process ends.

Here, as a result of repeatedly performing the above-described processing for each character candidate and each character string sequence, various character string candidates are generated. At that time, in the above procedure, when a character candidate can be integrated into a plurality of character string candidates,
A plurality of character candidate strings having the same character candidate as a component may be generated.

In this case, it becomes necessary to determine which character string candidate is to be integrated with the character candidate, that is, which character string candidate is adopted. At the time of determination, for example, each of a plurality of character candidate strings having the same character candidate as a component is evaluated for the likelihood as a character string by comparing it with a preset evaluation criterion, and the likely character candidate is obtained according to this evaluation result. Adopt columns.

In addition, there are cases where there are a plurality of interpretations for the combination of characters that make up a character string. In such a case, all interpretations are performed by the method disclosed in Japanese Patent Laid-Open No. 28305/1993. There is a method in which a character string candidate related to is evaluated and a character string candidate related to a proper interpretation is selected according to the evaluation value.

However, in the conventional method, the character is cut out on the assumption that the direction of the character string is limited only to the vertical and horizontal directions. Therefore, when the character strings in all directions exist intensively. It was impossible to cut out a character string. Further, when the conventional method is expanded to support character strings in all directions, the number of character string candidates to be selected is explosively increased, which is not realistic.

[0008]

According to the conventional character string cutout method, it is difficult to correctly cut out a character string for an image in which character strings having various directions are concentrated. Further, conventionally, since all combinations of character candidates are considered as character string candidates, it takes time to process.

The present invention has been made in view of the above circumstances, and an object thereof is to provide a character string segmentation method capable of extracting a character string accurately and efficiently even when the character string is concentrated. To do.

[0010]

A character string cutout method according to the present invention extracts character candidates from an input image and combines the extracted character candidates to create a partial character string that can be a constituent element of the character string. The character string candidates are created only by combining the created partial character strings, and the character string is determined from the created character string candidates.

Preferably, the determination of the character string is performed by determining the character string based on the reliability obtained at the time of creating the partial character string and at the time of creating the character string candidate. Further, preferably, the reliability of the partial character string is obtained based on the similarity as a character obtained by recognizing the character candidates forming the partial character string, and the reliability of the character string candidate is It is characterized in that it is obtained based on the similarity of the partial character strings forming the character string candidates.

Further, it is preferable that the partial character string is created by combining the two extracted character candidates. Further, preferably, when creating a partial character string by combining the character candidates, the distance between the character candidates is shorter than a predetermined threshold, the size of the character candidates are substantially uniform, and When a partial character string is created from the character candidates of and the similarity as a character satisfies a predetermined threshold value when the partial character string is recognized, the partial character string composed of the character candidates is adopted. To do.

Preferably, in creating the character string candidates, partial character strings or a combination of the partial character strings and the already created character string candidates are created to create longer character string candidates, and the longer character string candidates are created. It is characterized in that each of the character string candidates is examined to determine whether or not each character string candidate has a characteristic as a character string and whether or not to adopt it as a character string candidate.

Preferably, when creating the character string candidates, the character string candidates are created only when the partial character strings to be combined or the partial character strings and the already created character string candidates have overlapping character candidates. It is characterized by

Preferably, the character string candidates are created when the character candidates overlap only one character. In addition, preferably, when checking whether or not the character string candidates have characteristics as a character string, it is checked whether or not the heights of the character candidates forming the character string candidates are aligned and whether the character candidates are lined up on a straight line, It is characterized in that a character string candidate is created only when the result satisfies a predetermined condition.

Further, preferably, in determining the character string from the character string candidates, the character string candidates are sorted in order of reliability for each character string candidate having each number of characters, and the character string is determined based on this result. It is characterized by.

[0017]

In the present invention, from the input image, for example, the size
Character candidates are extracted based on the shape and the like. Then, the extracted character candidates are appropriately combined based on, for example, the distance, the size, the recognition processing result, and the like to create a partial character string of, for example, two characters. At that time, preferably, the reliability of each partial character string is obtained.

Then, character string candidates are created based on the created partial character strings. At that time, it is preferable to integrate those including the same character candidates. Further, preferably, the reliability of each character string candidate is obtained.

Finally, a character string is determined from the character string candidates (which may include partial character strings). For example, it is determined based on the reliability or the position, direction, length, etc. of the character string in addition to the reliability.

According to the present invention, among the character candidates first extracted from the input image, the character candidates determined to be able to be integrated are combined to create a partial character string, and then this partial character string is created. We will create character string candidates by combining them with appropriateness, and finally determine the character string from among the character string candidates according to a predetermined evaluation standard, so many character candidates will be concentrated and Even if a character string cannot be determined only by its features, the character string can be cut out correctly, and it is much more efficient than the conventional method that considers all character candidate combinations as character string candidates. Can be cut out.

Further, since the character string can be determined after enumerating all the character string candidates having a high possibility, it is easy to find the optimum combination of the character strings. In addition, since all the character string candidates that are highly likely to be available at the time of character string determination are available, it is easy to create combination candidates for multiple character strings, and advanced character string determination such as semantic processing that cannot be judged only by shape and recognition. A combination of a plurality of character string candidates can be easily presented to the process.

Further, the similarity by the recognition process indicating the reliability of the character string is obtained only when the partial character string is created, and the reliability as the character string when creating the character string candidate is the portion that constitutes the character string candidate. The processing time is shortened because the reliability is calculated based on the character string. As described above, according to the present invention, a character string can be accurately and efficiently extracted even when the character string is concentrated.

[0023]

Embodiments will be described below with reference to the drawings. FIG. 1 shows the configuration of a character clipping device according to an embodiment of the present invention. As shown in FIG. 1, the character slicing device according to the present embodiment is suitable for use in a device for recognizing an image described in a document such as a drawing recognition device or a text reader. Candidate extraction unit 2
4, a partial character string creating unit 26, a character string candidate creating unit 28, and a character string determining unit 30.

A character candidate is a pattern extracted as a character from an input image, and means a candidate for a character that constitutes a character string. A character string candidate is a character string candidate composed of two or more character candidates. A character string candidate composed of two character candidates is called a partial character string for convenience.
A character string candidate composed of three character candidates is called a quasi-character string candidate for convenience.

Reference numeral 32 in the drawing is a character string candidate table. As will be described in detail later, for each character string candidate, it is a table that roughly shows the character candidates that make up the character string candidate and the reliability of the character string candidate.

The image input section 22 is for inputting an image. In the case of optically reading and inputting an image described in a document, etc., it is configured by using a photoelectric conversion device such as an image scanner, and when the image is already digitized and stored in an electronic file, etc. It is configured by using a reading device or the like.

The character candidate extraction unit 24 extracts a character from the input image and uses it as a character candidate for the partial character string creation unit 2
Pass to 6. The partial character string creating unit 26 creates a partial character string from the character candidates.

The character string candidate creating unit 28 integrates the partial character strings and the partial character strings to create a quasi-character string candidate, and further, to the extent possible, a character string candidate composed of three or more character candidates. The partial character strings of are integrated to create long character string candidates one after another.

The character string determination unit 30 determines a character string candidate to be adopted as a character string from all the created character string candidates. Next, the character string cutout processing (character string candidate table creation processing) of this embodiment will be described.

In the character string cut-out processing of this embodiment, after the image is input, the character candidate extracting section 24, the partial character string creating section 26, the character string candidate creating section 28 perform the following steps. The processing by the character string determination unit 30 is sequentially performed.

FIG. 2 is a flow chart showing the flow of the main part of the character string cutout processing of this embodiment. Hereinafter, the character string cutout processing according to the present embodiment will be described in more detail with reference to FIG.

(Processing by Character Candidate Extraction Unit 4) First, the character candidate extraction unit 24 extracts characters included in the image input from the image input unit 22 as character candidates.
At that time, information such as the position and size of the center of the character candidate is added to the extracted character candidate.

As a process of extracting characters from an image, for example, a method disclosed in Japanese Patent Publication No. 63-63951 for distinguishing character areas from graphic areas can be used.

After the character candidates are extracted, the positions of the character candidates in the entire image are searched, and the character candidates having a close distance in the image are put together to form a character candidate group. The process from here is performed in each group created as described above.

The character candidates included in the same group are sequentially numbered from 0. There is no particular standard for numbering. FIG. 3A shows an example of the input image. It is assumed that all the patterns in the figure belong to the same group.
(B) is an example in which serial numbers are assigned to the character candidates obtained from (a), and character candidate numbers 0 to 16 are assigned to the character candidates, respectively.

In this embodiment, as described above, each group is provided with the character string candidate table 32 for representing character string candidates obtained by integrating a plurality of character candidates in the same group.

The character string candidate table 32 has, for example, a matrix structure in which character string candidates are arranged vertically and character candidate numbers are arranged horizontally. In the character string candidate table 32, by setting the bit corresponding to the position of the character candidate number forming a certain character string candidate to 1 and setting the other bits to 0, which character candidate the character string candidate is composed of Represents Further, each character string candidate is provided with a field for storing “reliability” indicating how likely the character string candidate is to be a character string from the viewpoint of character recognition. FIG. 4 shows an example of the character string candidate table 32.

In this processing, the character string candidate table 3
In 2, a character string candidate composed of one character candidate in the group is registered (see the column for the number of characters 1 in FIG. 4). A group of n character candidates can have n character string candidates. For the character candidates, the reliability as a character string candidate is not obtained.

When the group of character candidates is the example of FIG.
Since there are character candidates from No. 16 to No. 16, the horizontal length of the character string candidate table 32 is 17 as shown in FIG. As for the line corresponding to the character string candidates composed of the character candidates of “A” in FIG. 3, only the first column is 1 as shown in FIG. 4,
Other columns are 0.

<Process of Partial Character String Creating Unit 6> Next, the process by the partial character string creating unit 6 will be described. Two character candidates are selected from the same group, and the distance between the character candidates is calculated. It is checked whether the distance is shorter than a certain threshold value, that is, the character candidates are not so far apart from each other. Also, it is checked whether the two character candidates have the same size.

If the two character candidates are close to each other and have the same size (step S603), the two character candidates are subjected to recognition processing (step S604). For the recognition process, a known technique such as a pattern matching method or a composite similarity method can be used.

In the present embodiment, the recognition process is performed in two vertical directions, with the direction of the character string composed of two character candidates (for example, the direction connecting the centers of the character candidates) being the lateral direction of the character. As the processing result, the similarity as the character of the character candidate is output with a maximum of 1000 points, for example.

In this embodiment, a partial character string formed by these character candidates is stored in the character string candidate table 32 only when the similarity between the two character candidates is higher than the threshold value in at least one of the vertical direction. Register (step S605).

At this time, the reliability of the partial character string is calculated based on the similarity of the character candidates included in the character string candidate. For example, the similarity of two characters in the direction exceeding the threshold is averaged.

However, in the part corresponding to the partial character string (the number of characters is 2) in the character string candidate table 32, the reliability is not registered, and instead the similarity of the two characters forming the partial character string is unchanged. Register This is to obtain the reliability of a character string candidate having a longer length of 3 characters or more.

If the upper and lower similarities exceed the threshold value, the higher average is taken as the reliability. This process is performed by combining all the character candidates in the group (step S
601, S602).

In this way, the character string candidate table 32
The character candidates and the partial character strings are partially completed. <Processing of Character String Candidate Creating Unit 8> Next, the process proceeds to the processing by the character string candidate creating unit 28 that creates a character string candidate of three or more characters by combining partial character strings. Here, first, two partial character strings having a common character candidate are combined to create a quasi-character string candidate consisting of three characters (step S60).
6, S609-S612).

First, two partial character strings are selected. Then, it is checked whether or not those partial character strings have an overlap of one character candidate. If one character candidate has an overlap, a quasi-character string candidate that is a character string candidate that integrates them is created (step S610). In this case, the quasi-character string candidate is composed of three character candidates.

Next, it is checked whether or not this quasi-character string candidate has characteristics as a character string (satisfies the geometrical property) (step S611). Here, the condition having a characteristic as a character string is, as an example, that the longest side and the next longest side when the candidate character string is surrounded by the convex closure have substantially the same length and are parallel to each other. Of (about the height of the letters)
It is conceivable that there is a distance and the centers of all the character candidates that form the quasi-character string candidates are arranged on a straight line within a predetermined allowable range.

If the quasi-character string candidate has a feature as a character string, it is registered in the character string candidate table 32 as a character string candidate (3 characters) (step S612).
At this time, the reliability of the quasi-character string candidate is calculated based on the similarity between the character string candidates and the partial character strings that form the candidate. For example, the average or the weighted average of the similarities of the character candidates forming the character candidate is used.

However, as this similarity, the one obtained when the partial character string was created is used, and new recognition processing is not performed. The above-described character string candidate creation process of three characters (quasi character string candidate creation process) is performed for all combinations of partial character strings.

Here, the character candidate (1 character), the partial character string (2 characters), and the quasi character string candidate (3 characters) of the character string candidate table 32 are completed. Next, a character string candidate creation process of four or more characters is performed (steps S607 to S61).
2).

The process of creating a character string candidate of four or more characters is basically the same as the process of creating a character string candidate of three characters.
However, the character string candidates to be combined are not partial character string candidates, but 1 from the partial character string and the character string candidate to be created.
The combination of character string candidates with few characters is performed. That is, the 4-character character string candidates are obtained by combining the partial character strings and the 3-character character string candidates.

Also in this case, the quasi-character string candidates are executed only when the character string candidates to be combined have an overlap of one character. Similar to the creation of the three-character string candidates, the character strings of these quasi-character string candidates are examined, and if the conditions are satisfied, they are registered in the character string candidate table 32 as character string candidates.
The reliability at this time is also obtained as an average of the similarities of the character candidates that form the character string candidates.

For example, if the partial character string "53" shown in FIG. 5A and the quasi character string candidate "347" are integrated,
As shown in (b), a 4-character character string candidate “5347” can be created. Then, as shown in FIG. 6, the corresponding bit is set and the reliability is calculated.

When the character string candidates of 4 characters are registered by all the combinations, the same method is used to sequentially create 5 characters, 6 characters, ... (Step S607).
~ S612).

As described above, the character string candidates are created in order from the shortest one in the same group, and the character string candidate table 32 is completed by repeating the character string candidates and the partial character strings to be combined (step S613).

In this character string candidate table 32, as shown in FIG. 4, the uppermost character string candidates are one character, and the character strings that are longer toward the bottom are registered. <Process of Character String Deciding Unit 10> When the character string candidate table 32 in the group is completed, next, the process proceeds to the process by the character string deciding unit 30 which determines the character string using the character string candidate table 32.

The character string candidates registered in the character string candidate table 32 created as described above have the reliability created based on the similarity calculated by the recognition process. -The best combination of character strings may be obtained by using the direction and length. When determining a character string, a partial character string is also treated as a character string candidate.

In this embodiment, first, character string candidates of the same length in the character string candidate table 32 are sorted. The sorting is performed so that the character string candidates with low reliability are on the top and the character string candidates with high reliability are on the bottom.

If the character string "has a higher possibility of being a character string as long as it is" and "has a higher reliability of the same length as a character string is more likely", this character string candidate In the table 32, a lower character string candidate is more likely to be a character string, and an upper table is less likely to be a character string.

Therefore, the character string candidate table 32 is sequentially searched from the bottom to determine the character string. The partial character string is treated as a character string candidate of 2 characters. Since the one-character string candidate table 32 has no reliability, no sorting is performed.

Here, the character string determining unit 30 has a buffer having a serial number of character candidates in the group (not shown). This is called a character candidate buffer. First, the character string candidate at the bottom of the character string candidate table 32 is determined as a character string. Then, among the character candidates included in the group, the character candidates included in this character string are removed from the character candidate buffer. The character string candidate table 32 is searched in order from the bottom, and if there is a character string candidate configured only by the character candidates registered in the character candidate buffer, it is determined as a character string and included in the character string. Remove character candidates from the character candidate buffer.

This process is continued until the character candidate buffer becomes empty or a partial character string of two characters. This is the end of the character string determination process within the group.

When the above character string cutout processing is performed for all groups, the character string cutout for the entire image is completed. In the following, the character string cutout process of the present embodiment will be described more specifically by using the image example shown in FIG. 3 as an input.

First, the character candidate extracting section 24 is used to execute the process shown in FIG.
It is assumed that character candidates as shown in (a) are collected as a group. In the case of FIG. 3, it is assumed that the arrow is also a character candidate. In this case, the partial character string obtained from FIG. 3 by the partial character string creating unit 26 is shown in FIG. In FIG. 7, for example, 0
The character pairs of number "← (arrow)" and number 1 "A" are close to each other and have the same size, but they correspond to number "← arrow" when they are recognized. Since there is no dictionary candidate and the similarity is low, the character string candidate table 32 is not created. No. 4 "7" and No. 5 "→ (arrow)", 1
No. 1 "6" and No. 12 "6", No. 13 "D" and 14
It is also determined that the number “H” and the like cannot be integrated. Nos. 7 "1" and 9 "B" do not have the same height from the center line, and are therefore repelled before the recognition process.

Next, the processing by the character string candidate creating section 28 is started. First, as a character string candidate of three characters, a character string candidate having a character string candidate of (1,2,3) as an element is obtained from a pair of character candidate numbers (1,2) (2,3).

Similarly, (2, 3, 4), (6, 7,
Character string candidates having elements 8), (9, 10, 11), and (13, 15, 16) are created, respectively. Next, as a character string candidate of 4 characters, character candidate number (1, 2, 3)
A combination of character string candidates (1, 4, 2) or (1, 2)
3, 4) can be done. Since the same character string is not created, only one character string candidate (1, 2, 3, 4) is registered.

In the case of FIG. 3, character string candidates of 5 characters or more are not created, so the character string candidate table 32 is as shown in FIG. When the character string candidate table 32 is completed, the processing moves to the character string determination unit 30. For the sake of simplicity, assume that there is no change from the state of FIG. 4 even if the character string candidate table 32 is sorted by reliability.

Initially, the character candidate numbers 0 to 16 are registered in the character candidate buffer. First, the lowest character string candidate (1, 2, 3, 4) in the character string candidate table 32 is determined as a character string. The character candidates of character candidate numbers 1, 2, 3, and 4 are removed from the character string candidate buffer.

When the character string candidate table 32 is searched for a character string composed of only character candidates registered in the character candidate buffer in order from the bottom, the character string candidates (13,
15, 16) is the character string to be determined next.

Similarly, the character strings (9, 10, 11) (6,
7, 8) are decided. Character string candidates (1, 2, 3)
(2, 3, 4) is not selected because it has the character candidates that have deviated from the character string buffer as elements.

Therefore, the character strings of the group in FIG.
2, 3, 4) "A007", (6, 7, 8) "C1"
2 "," B86 "of (9, 10, 11), (13, 1
It is decided to be four of "D53" of 5, 16).

The following advantages can be obtained by this embodiment as described above. Even if a large number of character candidates are concentrated and the character string cannot be determined only by the external characteristics, the character string can be cut out correctly. Moreover, the processing efficiency is far better than the conventional method in which the combination of all character candidates is considered as a character string candidate.

Further, since the character string can be determined after listing all the character string candidates having a high possibility, it is easy to find the optimum combination of the character strings. In addition, since all the character string candidates that are highly likely to be available at the time of character string determination are available, it is easy to create combination candidates for multiple character strings, and advanced character string determination such as semantic processing that cannot be judged only by shape and recognition. A combination of a plurality of character string candidates can be easily presented to the process.

Further, as a more specific point, when the partial character strings are formed by the partial character string creating unit 26, the sizes are not uniform, or the two characters are similar in the direction of the partial character strings in the recognition process. If the degree is low, the partial character string is not created, so it is difficult to cut out the correct character string with the conventional method An image in which character strings with various directions are concentrated (see Fig. 3) Even in the case of
Only those shown in are created.

Further, when the character string candidate creating unit 28 creates the character string candidates, the character string candidates are created based on this partial character string, so that the characters are different in direction or are uneven in size. No column candidates are created.

Further, in the partial character string creating section 26, the character recognition processing is performed after the partial character string is created, so that the processing amount can be small. Further, the similarity by the recognition process indicating the reliability of the character string is obtained only by the partial character string creating unit 26, and the reliability of the character string created by the character string candidate creating unit 28 is the partial character string forming the character string candidate. Since it is obtained based on the reliability of, it is not necessary to perform recognition processing every time a character string candidate is created, and the processing time is shortened.

<Modification> In the processing by the partial character string creating unit 6 described above, when the two character candidates are subjected to the recognition processing, the recognition processing uses the direction of the character string formed by the two character candidates as the character direction. , 2 directions in the vertical direction.

Considering that the character strings include vertically arranged strings, recognition processing is performed not only in the upper and lower two directions but also in the left and right directions, and the reliability is based on the obtained four similarities. You may ask for the degree.

Furthermore, recognition processing may be performed in four directions of 45 degrees in addition to the two directions of up, down, left and right, and may be appropriately set according to the image to be processed. Further, the present invention is not limited to the above-described embodiments, and various modifications can be carried out without departing from the scope of the invention.

[0082]

According to the present invention, a partial character string is created by combining character candidates that have been determined to be integrated among character candidates extracted from an input image first, and then this partial character string is created. We will create character string candidates by combining the character strings with validity, and finally, for example, determine the character string from among the character string candidates according to a predetermined evaluation standard.
Even if a large number of character candidates are concentrated and the character string cannot be determined only by the external features, the character string can be cut out correctly and compared with the conventional method that considers all combinations of character candidates as character string candidates. Then, the character string can be cut out much more efficiently.

Further, since the character string determination can be carried out after enumerating all the character string candidates having a high possibility, it is easy to find the optimum combination of the character strings. In addition, the similarity by recognition processing that represents the reliability of a character string is obtained only when creating a partial character string, and the reliability as a character string when creating a character string candidate is the reliability of the partial character string that constitutes the character string candidate. The processing time is shortened because it is calculated based on the reliability. As described above, according to the present invention, a character string can be accurately and efficiently extracted even when the character string is concentrated.

[Brief description of drawings]

FIG. 1 is a diagram showing a configuration of a character string clipping device according to an embodiment of the present invention.

FIG. 2 is a flowchart showing a flow of a character string cutout process of the embodiment.

FIG. 3 is a diagram showing an example of an image which is a target of a character string cutout process.

FIG. 4 is a diagram showing an example of a character string candidate table.

FIG. 5 is a diagram for explaining creation of a 4-character string candidate.

FIG. 6 is a diagram for explaining creation of character string candidates of four characters.

FIG. 7 is a diagram showing an example of a partial character string created by a partial character string creating unit of the embodiment.

[Explanation of symbols]

22 ... Image input unit, 24 ... Character candidate extraction unit, 26 ... Partial character string creation unit, 28 ... Character string candidate creation unit, 30 ... Character string determination unit, 32 ... Character string candidate table

Claims (10)

[Claims] 1. Only by extracting character candidates from an input image, combining the extracted character candidates to create a partial character string that can be a constituent element of the character string, and combining the created partial character strings. A character string segmentation method characterized by creating character string candidates and determining a character string from the created character string candidates. 2. The character string slicing method according to claim 1, wherein in the character string determination, the character string is determined based on the reliability obtained when the partial character string is created and when the character string candidate is created. . 3. The reliability of the partial character string is obtained based on the similarity as a character obtained by recognizing the character candidates forming the partial character string, and the reliability of the character string candidate is The character string cutout method according to claim 2, wherein the character string candidate is obtained based on the similarity of the partial character strings forming the character string candidate. 4. The character string cutout method according to claim 1, wherein the partial character string is created by combining the two extracted character candidates. 5. When a partial character string is created by combining the character candidates, the distance between the character candidates is shorter than a predetermined threshold value, the character candidates are substantially equal in size, and
When a partial character string is created from the character candidates and the similarity as a character satisfies a predetermined threshold value when the partial character string is recognized, the partial character string consisting of the character candidates is adopted. The character string cutout method according to claim 1. 6. When creating the character string candidates, the longer character string candidates are created by combining the partial character strings or by combining the partial character strings with the already created character string candidates, and the created character strings are created. The character string cutout method according to claim 1, wherein it is determined whether or not each candidate has a characteristic as a character string and whether or not to adopt it as a character string candidate. 7. When creating the character string candidates, the character string candidates are created only when the partial character strings to be combined or the partial character strings and the already created character string candidates have overlapping character candidates. The character string cutout method according to claim 6. 8. The character string cutout method according to claim 7, wherein a character string candidate is created when the character candidates overlap only one character. 9. When checking whether or not the character string candidates have characteristics as a character string, it is checked whether or not the heights of the character candidates forming the character string candidates are uniform and whether or not the character candidates are arranged in a straight line. The character string cutout method according to claim 6, wherein the character string candidates are created only when the result satisfies a predetermined condition. 10. When determining a character string from the character string candidates, the character string candidates are sorted in order of reliability for each character string candidate of each number of characters, and the character string is determined based on this result. The method for cutting out a character string according to claim 2.