JPH10240867A - Method and device for character segmentation - Google Patents

Abstract

PROBLEM TO BE SOLVED: To select a word matching a source image out of words even when there are more than one word in a word dictionary among generated words by matching a generated word against the word dictionary and determining a character segmentation area for one character according to the result of the word matching. SOLUTION: A word generation part 25 generates a word by combining candidate character codes of a primary segment and a secondary segment obtained by a character recognition part 21. A word matching part 27 matches the word generated by the word generation part 25 against the word dictionary. A result selection part 33 compares word level evaluated values as to the word generated by the word generation part 25 and selects a preferential word as a word for character area segmentation. A result output part 35 outputs a segmentation candidate position as to the characters constituting the word determined by the result selection part 33 to a control part 11. The control part 11 indicates character segmentation according to the result.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a character extracting method in a character recognition technology and a character extracting apparatus suitable for implementing the method.

[0002]

2. Description of the Related Art A handwritten character string has a greater change in character spacing and character shape than a printed character string. Therefore, if characters are cut out at regular intervals when extracting characters from image data of a handwritten character string, characters cannot be extracted accurately. This causes misrecognition. As a conventional technique for solving this, there is a character cutout method disclosed in, for example, Japanese Patent Application Laid-Open No. 5-35917.

In this prior art, a character block is first cut out from a line image. Here, the character block is a block region of black bits. Note that the character block cut out from the line image is also referred to as a primary segment in this specification. There is no guarantee that this character block is one character. That is, the character block may form a character pattern by itself or may be a part of the character pattern. Next, the character blocks (primary segments) are integrated, and a pattern that can be regarded as the size of one character is generated as a character pattern (also referred to as a secondary segment). The primary segment may be a character pattern. Next, character recognition is performed on the character pattern. Next, when the character pattern does not include the same character block in relation to another character pattern (when there is no overlap of the character blocks), the character pattern is cut out as it is. On the other hand, when there is an overlap of character blocks, a word is generated by combining the recognition result of each of these character patterns and the first candidate of each of the recognition results of several characters before or after the character pattern. The generated words are checked against a word dictionary.

[0004] For a word that matches a registered word in the word dictionary with the generated word, its character pattern is cut out. If they do not match, the evaluation value of each character pattern before word generation and the evaluation value of the word are obtained, and the character pattern with the highest evaluation value is cut out.

[0005]

However, in the prior art, when words generated by combining the results of character pattern recognition are collated with a word dictionary, there are a plurality of words present in the word dictionary among these words. No allowance is given for the case. Therefore, a problem as described below occurs.

For example, consider the case of an original image in which a plurality of cutouts can be considered. Specifically, the first candidate of the recognition result of a certain character pattern A is “arrow”, the first candidate of the recognition result of a certain character pattern B is “knowledge”, and the first candidate of a certain character pattern C is “ Consider an example in which the first candidate of a certain character pattern D is “mouth”. Moreover, character pattern A and character pattern B
Means that the character pattern C and the character pattern D overlap on the image. At this time, the word “Yada” is obtained for the combination of candidate characters “A → C”, and the word “Shiguchi” is obtained for “B → D”. It is assumed that both “Yada” and “Shiguchi” have been registered in the word dictionary. However, in the related art, there is no description as to which of these competing words is to be selected, so that the final result is automatically obtained when the words are compared first. In the character extraction process performed next, character extraction is performed for each character pattern of the word as the final result, so (1) if “Yata” is the final result, “Y” and “ (2) If “Shiguchi” is the final result, “Shi” and “Mouth” are respectively extracted. As can be seen by comparing (1) and (2), there is a great difference in how characters are cut out depending on which word is selected. Therefore, in the related art, the cutout accuracy and the recognition accuracy may be reduced.

Therefore, a word is created by combining the candidate character codes of the primary segment and the secondary segment, and the created word is collated with a word dictionary, and one character is extracted from the original image based on the result of the word collation. This is a character extraction method including a process of determining a character extraction area for each word. Even if there are a plurality of words in the word dictionary among the created words, a word matching the original image is selected from among the plurality of words. There is a need for a character segmentation method that can be selected and a character recognition device suitable for implementing the method.

[0008]

Therefore, according to the invention of the character extracting method of this application, the original image data including the image data of the character string stored in the memory is converted into a primary area which is a block area of black bits. A process of extracting each segment, a process of integrating the extracted primary segments according to a predetermined rule to create a secondary segment, a process of character recognition of each primary segment and each secondary segment, A process of creating a word by combining candidate character codes of the primary segment and the secondary segment obtained by the recognition, a process of matching the created word with a word dictionary, and a process for matching one character based on the result of the word matching And a process of determining a character cut-out area for a minute.

Processing for obtaining a shape characteristic evaluation value based on the shape characteristics of each of the primary segment and each of the secondary segments.

In the above-described character recognition, a process of obtaining a recognition result candidate of each segment up to the Kth position and obtaining a recognition evaluation value represented by a ratio of a character type in the recognition result candidate for each segment. Here, K is a predetermined positive integer.

[0011] For each of the created words, the shape feature evaluation value, the recognition evaluation value, and the number of segments for the constituent segments of each character constituting the word (if the constituent segment is a primary segment, A process of calculating a character level evaluation value calculated based on the number of integrated segments in the case of primary and secondary segments).

When the created word is compared with a word dictionary, a process of obtaining a word level evaluation value indicated by whether or not the word exists in the word dictionary.

In the processing for determining a character cut-out area, the word level evaluation values for each word are compared with each other, and when the word levels are superior, the word is selected, and a plurality of words having the same word level evaluation value are selected. If there is, the character level evaluation values of the plurality of words are compared with each other to select one word, and each segment constituting the selected word is set as the character cutout area.

According to the invention of the character extracting method, first, a shape characteristic evaluation value and a recognition evaluation value for each segment are obtained. Also, words are created by combining the recognition results for each segment. Further, the created character level evaluation value for each word is the shape feature evaluation value,
It is obtained based on the recognition evaluation value and the number of segments.
Here, the shape evaluation value can be said to be an evaluation value as to whether or not the segment has a shape like one character. In addition, the recognition evaluation value indicates that the segment has any character type (for example, any one of the four character types of symbols and numbers, katakana, hiragana, and kanji).
Can be said to be an evaluation value indicating whether or not the property belonging to is high. Therefore, the character level evaluation value is an evaluation value reflecting the shape characteristic evaluation value and the recognition evaluation value. In other words, the character level evaluation value is (1) whether each character constituting the created word is likely to be one character, and (2) whether each character is composed of segments of a valid character type. Evaluation value in consideration of whether or not (in other words, characters that have a low probability of being present, such as characters in which a segment considered to be Hiragana and a segment considered to be Kanji are combined) It can be said that. for that reason,
Using the character level evaluation value, a word including a character that does not have a shape like a single character or a word including a character that is a combination of segments that cannot be normally considered due to the character type is used as a word for determining a character cutout area. It can be difficult to select. Therefore, in the present invention using this character level evaluation value, even if there are a plurality of words in the word dictionary among the created words, it can be said that it is easy to select a word suitable for the original image from the plurality of words. . In addition, since the character cutout region can be determined from the word suitable for the original image, the cutout accuracy and the recognition accuracy can be improved.

In implementing the character extracting method of the present invention, it is preferable that the shape characteristic evaluation value be the aspect ratio of the segment. In this case, the shape corresponding to the circumscribed frame of the character is considered as the shape characteristic of the segment. In general, a character that is likely to be a single character is often close to a square shape. Therefore, if the aspect ratio of a segment is used, it is relatively easy to grasp the characteristics of a single character of a segment.

Further, in implementing the invention of this character segmentation method, the secondary segment has a direction in which characters are arranged in X.
The direction is created by integrating m primary segments that are continuous in the X direction according to a predetermined rule, and
The word is a candidate path represented by the concatenation of the primary segment and / or the secondary segment, and includes the following (a) to
It is preferable to create it based on the candidate path created by the process including the process (d). Here, m is an integer of 2 or more.

(A) Coordinates for dividing each of the m primary segments in the X direction are determined as candidate clipping positions Ci.
When (i = 0 to m), a process of extracting all the segments whose extraction candidate position C0 is the extraction start point from the m primary segments and the created secondary segments.

(B) With respect to each of the segments extracted in the process of (a), the segmentation candidate positions Cj (j = 1 to m) on the end point side of the segment are connected since the segmentation start position. Other segments that can be connected, and other segments that have the same relationship as the above-described segmentation candidate positions to the other segments, have another segment that can be further connected to the segment, and the segmentation candidate position on the end point side is Cm. A process of extracting from the m primary segments and the created secondary segments until another segment appears.

(C) Every time the other segment is extracted in the process of (b), a process of determining whether or not the number of segments of a candidate path including the other segment is within a specified number.

(D) A candidate path having the number of segments within the specified number and a candidate position for extracting the end point of the last segment in the candidate path being Cm is set as a candidate path for word creation. Processing.

According to this preferred embodiment, from the cutout area defined by m primary segments, a cutout candidate position C
All the candidate paths consisting of the segment sequence in which 0 is the starting point, the cutout candidate position Cm is the end point, and the number of segments is within the specified number are extracted. Since the recognition result (candidate character code) of each of the segments constituting the extracted candidate path has already been found in the recognition processing, a word having a specified number of characters or less is generated from the extracted candidate path. You.

Specifically, the candidate path creation process in the preferred embodiment described above is preferably created by a process including the following processes (1) to (9).

(1) Coordinates for dividing each of the m primary segments in the X direction are designated as cutout candidate positions C0
To Cm, the cutout candidate position Ci (i
= 0 to m) is Cm.

(2) A second process which is executed when it is determined that Ci = Cm in the first process, and records the current candidate path in the candidate path memory.

(3) This is executed when Ci 場合 Cm is determined in the first processing, and the segment S sandwiched between the extraction candidate position Ci and the extraction candidate position Cj (j = i + 1)
Third processing for determining whether or not k + 1 exists.

(4) Executed when it is determined in the third processing that a segment exists, and the segment Sk +
A fourth process of determining whether the number of segments of the candidate path does not exceed a specified number when 1 is added to the candidate path.

(5) A fifth process which is executed when it is determined in the fourth process that the number is within the specified number, and adds the segment Sk + 1 to the candidate path.

(6) The sixth processing which is executed subsequent to the fifth processing and is executed again from the first processing by regarding the cut candidate position Cj as the focused cut candidate position Ci.
Processing.

(7) For the candidate path created by executing the fifth processing and the sixth processing, a seventh step is to delete the segment added to the candidate path from the candidate path. processing.

(8) When the determination is negative in the third processing, when the determination is negative in the fourth processing, or when the seventh processing is performed, Eighth processing for changing j defining the candidate position to j = j + 1 and determining whether or not the changed j satisfies j> m in relation to m.

(9) A ninth process that is executed when j ≦ m is determined in the eighth process, and is re-executed from the third process.

According to the processes (1) to (9), from the cutout area defined by the m primary segments, the cutout candidate position C0 is the starting point and the cutout candidate position Cm is the end point. All the candidate paths composed of the segment sequence in which the number of segments is within the specified number are extracted.

Further, in implementing the invention of the character segmentation method, when the direction in which the characters are arranged is set to the X direction as the predetermined rule when the primary segments are integrated, m consecutive m characters are arranged in the X direction. The height H of the segment having the highest height among the primary segments is obtained, and another primary segment existing in the H × N coordinate range in the X direction with respect to the primary segment of interest is focused on. It is preferred to use the rule of merging into primary segments (where N is a predetermined value).

The reason is that the size of a handwritten character, the space between characters, and the like usually change depending on the writer and the circumstances surrounding the writer. Therefore, it is common that the row height also changes. Therefore, by considering the height of the segment, the above variation in the row height is taken into account.

As a result, the secondary segment can be created under the condition in which the size of the character which changes depending on the writer or the situation surrounding the writer can be created, so that an appropriate secondary segment can be created. The value of N is preferably determined empirically (statistically). According to the research of the inventor of the present application, it is known that preferable integration is realized by setting N to 1.2.

In order to carry out the above-described character extracting method, it is preferable to configure a character extracting device as follows.

A segment extraction unit for extracting a primary segment, which is a block area of black bits, from input image data including image data of a character string stored in a memory; A segment integration unit that integrates to create a secondary segment according to rules, a character recognition unit that recognizes each primary segment and each secondary segment as a character, and a primary segment and / or a secondary segment obtained by the character recognition. A word creating unit that creates a word by combining candidate character codes, a word matching unit that matches the created word with a word dictionary, and a result of determining a character cutout area for one character based on the result of the word matching A character segmentation device comprising: A structure for obtaining auxiliary of up to K-position, the word collating unit, a word created in the above is configured to determine a word level evaluation value indicated by the presence or absence in the word dictionary.

A segment shape evaluation value calculation unit for obtaining a shape feature evaluation value based on a shape feature of each of the primary segment and the secondary segment; a recognition result candidate for each segment created by the character recognition unit A segment recognition evaluation value calculation unit that obtains, for each segment, a recognition evaluation value represented by the ratio of the character type in the above, and for each of the created words, the shape feature evaluation for the constituent segments of each character constituting the word value,
A character level evaluation value for calculating a character level evaluation value calculated based on the recognition evaluation value and the number of segments (1 when the constituent segment is a primary segment, the number of integrated segments when the constituent segment is a secondary segment) And a calculation unit. In addition, the result selection unit compares the word level evaluation values for each word, selects the word if the word level is superior, and selects the words if a plurality of words have the same word level evaluation value. The character level evaluation values of the words are compared with each other to select one word, and each of the segments constituting the selected word is determined as the character cutout region. Here, K is a predetermined positive integer.

[0039]

BRIEF DESCRIPTION OF THE DRAWINGS FIG. 1 is a block diagram showing an embodiment of a character extracting method and a character extracting apparatus according to the present invention; However, the drawings used in the description are schematically shown to the extent that the present invention can be understood.

FIG. 1 is a diagram showing a configuration of a character segmenting apparatus according to the embodiment. The character cutout device 10 includes a control unit 11, an image input unit 13, a segment extraction unit 15, a segment integration unit 17, and a segment shape evaluation value calculation unit 1.
9, character recognition unit 21, segment recognition evaluation value calculation unit 2
3. Word creation unit 25, word collation unit 27, word level evaluation value 29, character level evaluation value calculation unit 31, result selection unit 33
And a result output unit 35.

These components 11 to 35 can be constituted by a computer and its peripheral devices, respectively. Hereinafter, the structure and operation of each component will be sequentially described, and the details of the character segmentation method of the present invention will also be described.

(Control Unit) The control unit 11 controls each component 13
To 35 are controlled.

(Image Input Unit) The image input unit 13 includes a memory (not shown), inputs original image data to be subjected to character recognition, and stores it in the memory. Specifically, original image data represented by black and white binary is input.

The image input section 13 can have any suitable configuration. For example, the image input unit may have a configuration in which a scanner is provided and photoelectrically converts an optical signal from a document and takes in the original image data into a memory, or another database that stores the original image data in the first place. .
Of course, a binary image may be obtained from a multivalued image.

(Segment Extraction Unit) Segment Extraction Unit 1
5 extracts a primary segment, which is a block area of black bits, from the original image data stored in the memory of the image input unit 13 in a state where its coordinates are known.

The segment extracting process in the segment extracting section 15 can be easily performed by a method using a horizontal and vertical projection distribution of a black bit (also referred to as a black point) which is well known in the art.

More specifically, a memory in which original image data is stored is scanned in the horizontal direction, and a histogram of black points is obtained. Each minimum point in the histogram is set as a candidate cutting position Ci in the horizontal direction. The same processing is performed by changing the scanning direction to the vertical direction, and the extraction candidate position in the vertical direction is extracted. The primary segment is included in a rectangular area surrounded by the extraction candidate positions.

In order to deepen the understanding of the operation of the segment extracting unit 15, FIG.
The primary segments S0, S1, and S2 extracted from, and the coordinates for segmenting each of the primary segments S0 to S2 in the direction in which the characters are arranged (X direction in this example), that is, cutout candidate positions C0 to C3 are shown, respectively. However, S3 in FIG.
Is the secondary segment. This will be described later.

The X-direction start coordinates Xs, X-direction end coordinates Xe, Y-direction start coordinates Ys, and Y-direction end coordinates Ye of each of the primary segments S0 to S2 are stored in an internal segment coordinate table. (Not shown). FIG. 3A shows the definitions of Xs, Xe, Ys and Ye, and FIG.
The segment coordinate table for 0 to S3 is schematically shown. For example, Xs to Y for the primary segment S1
e is Xs = 37, Xe = 106, Ys = 2, Ye = 8
It turns out that it is 4.

Further, the cutout candidate positions (coordinates) C0 to C3 in the X direction necessary for creating a candidate path later.
Is stored in an internal predetermined memory (not shown).

(Segment Integration Unit) Segment Integration Unit 1
7 creates a secondary segment by integrating the primary segments extracted by the segment extracting unit 15 according to a predetermined rule. Specifically, in consideration of the shape characteristics of a plurality of adjacent primary segments, if there is a possibility that they will be one character even if they are integrated, the primary segments are integrated to create a secondary segment. Here, a secondary segment is created by the following procedure.

First, the height H of the segment having the highest height among the m primary segments continuous in the X direction is obtained. The height H is obtained by calculating the difference between the Ys coordinate and the Ye coordinate for each primary segment. Considering the example of the three primary segments in the example of FIG. 2, the Y coordinate difference of the segment S0 is 89-1 = 88, which is higher than any of the other segments S1 and S2. Therefore, in the example of FIG. 2, the segment having the highest height is the segment S0.

Next, a secondary segment is created by integrating another primary segment present in the H × N coordinate range in the X direction with respect to the primary segment of interest into the primary segment of interest.

This secondary segment creation processing will be described with reference to the explanatory diagram of the primary segment shown in FIG. 2, the segment coordinate table shown in FIG. 3, and the flow chart of the segment integration processing shown in FIG. This will be described more specifically.

First, the processing of the loop 1 starting from each of all the input segments is started. Therefore, as a target segment (referred to as segment A in FIG. 4), the processing of loop 1 starting from the primary segment S0 is first started (steps 41 to 47 in FIG. 4).

That is, the primary segment S0 and the primary segment S1 as the segment B arranged to the right thereof
The distance D in the direction in which the characters are arranged (here, the X direction) is determined (steps 42 and 43 in FIG. 4). This distance D is obtained from, for example, the difference between the Xs coordinates of each of the segments S0 and S1. Then, in this example, the distance D = 37-1 =
That is 36.

Next, it is determined whether or not the distance D is in the range of H × N (step 44 in FIG. 4). Here, N is a predetermined value. Here, N = 1.2. H is 88 here as described above. Therefore, in this case, it is determined whether or not D ≦ 1.2 × 88 = 105.6 is satisfied.

In the case of the example of FIG. 2, the primary segments S0 and S1
The distance D = 36 from the next segment S1 is D ≦ 1.2 × 8.
8, the primary segment S1 is integrated into the primary segment S0, and a secondary segment S3 is created (step 45 in FIG. 4, see FIG. 2).

Next, the loop 2 is executed again (FIG. 4
Steps 46 and 42), this time the primary segment S0
The distance D between the first segment S2 is obtained (step 43 in FIG. 4). This distance D is 173-1 = 172.

Next, it is determined whether or not the distance D is in the range of H × N (step 44 in FIG. 4). In this case, the primary segment S2 is D ≦ 1.2 × with respect to the primary segment S0.
Since the conditions of 87 are not satisfied, they are not integrated.

The primary segment of interest is changed to the primary segment S1, and the processing of loop 1 is started (FIG. 4).
Steps 41 and 42). Therefore, the primary segment S1
The distance D between the first segment S2 is obtained (step 43 in FIG. 4). This distance D is 173-37 = 136.

Next, it is determined whether or not the distance D is in the range of H × N (step 44 in FIG. 4). In this case, the primary segment S2 is D ≦ 1.2 × with respect to the primary segment S1.
Since the conditions of 87 are not satisfied, they are not integrated.

A secondary segment is created by the above procedure. The X direction start coordinate Xs, X direction end coordinate Xe, and Y direction start coordinate Y of each of the created secondary segments.
The segment integrating unit 17 additionally stores each of the s and Y direction end coordinates Ye in the segment coordinate table (step 48 in FIG. 4).

FIG. 5 schematically shows a segment coordinate table for the primary segments S0 to S2 and the secondary segment S3.

Further, with respect to the primary segments S0 to S2 and the secondary segment S3, a table in which what segments are interposed between a certain extraction candidate position and another extraction candidate position (this is referred to as " When a “segment table” is created, the result is as shown in FIG.

This segment table is an adjacency matrix in graph theory. That is, considering an adjacency matrix in which the start cutout point is a row and the end cutout point is a column, the table can be created by giving a segment number (one of S0 to S3 in this case) to the element. However, a blank in FIG. 6 indicates that there is no segment to be sandwiched.

From FIG. 6, the primary segment S0 is a segment sandwiched between the cutout candidate position C0 and the cutout candidate position C1,..., The secondary segment S3
Is a segment sandwiched between the extraction candidate position C0 and the extraction candidate position C2.

(Segment Shape Evaluation Value Calculation Unit) The segment shape evaluation value calculation unit 19 calculates an evaluation value based on the shape characteristics of each primary segment and each of the secondary segments created by the integration (this is a shape feature evaluation value). That)
Ask for. The segment shape evaluation value calculation unit 19 is, for example, a calculation unit that calculates an aspect ratio of a segment as a shape evaluation value. More specifically, when the segment dimension along the X direction is ΔX and the segment dimension along the Y direction is ΔY (see, for example, FIG. 3A), the shape evaluation value is obtained by the following equation (1).

Shape evaluation value = min (ΔX, ΔY) / max (ΔX, ΔY) (1) Here, ΔX = Xe−Xs and ΔY = Ye−Ys.
Also, min (ΔX, ΔY) takes the smaller value of ΔX and ΔY, and max (ΔX, ΔY) is ΔX
And ΔY means the larger value.

The shape evaluation value obtained in this manner is an evaluation value representing the aspect ratio of the segment in the range of 0 to 1.
In this case, a segment closer to a square, that is, a segment whose shape evaluation value is closer to 1, means that the character-likeness is higher.

(Character Recognition Unit) The character recognition unit 21 performs character recognition on the assumption that each primary segment and each secondary segment is one character, and stores a candidate character code as a result. The character recognition processing itself can be performed by any conventionally known method.

However, the character recognition unit 21 in the present invention
Is a character recognizing unit that determines candidate characters up to the Kth position (K is a predetermined positive integer; here, K = 10) for each segment.

In FIG. 7A, the character recognition unit 21 performs a character recognition process on a segment called “bow” corresponding to the primary segment S0. As a result, 10 segments from the first place to the tenth place are obtained.
An example in which the number of candidate characters has been output is shown. Note that the candidate characters are actually given by candidate character codes. Also,
Of course, not ten candidate characters may be output. These candidate characters are calculated by the segment recognition evaluation value calculator 23.
And used by the word creation section 25 (details will be described in the corresponding sections).

(Segment Recognition Evaluation Value Calculation Unit) The segment recognition evaluation value calculation unit 23 calculates the maximum K for each segment.
A recognition evaluation value represented by the ratio of the character type in the recognition result up to the rank is obtained.

In this case, the segment recognition evaluation value calculation unit 2
3 indicates (1) symbols / numbers, (2) katakana, (3) hiragana,
(4) Count the number of each of the four types of kanji, and divide each count value by the number of recognition results K to calculate a recognition evaluation value for each of the four types of characters.

In order to deepen the understanding of the processing in the segment recognition evaluation value calculation section 27, the "bow" shown in FIG.
FIG. 7B shows the recognition evaluation value calculated by the recognition evaluation value calculation unit 27 from the recognition result of the above. This FIG.
In the case of (B), out of the ten recognition results, four results were recognized as symbols and numbers, and two results were recognized as katakana.
Since there are 0 characters and 4 results recognized as Hiragana and 4 results recognized as Kanji, the recognition evaluation values of the four character types are 0.4, 0.2, 0, 0.4 It has become.

(Word Creation Unit) The word creation unit 25 creates a word by combining the candidate character codes of the primary segment and the secondary segment obtained by the character recognition unit 21. More specifically, a word is created based on a form of a candidate path represented by the concatenation of the primary segment and / or the secondary segment. Note that limiting the number of characters when creating a word can prevent the creation of an unnecessarily long word, so here we limit the number of characters in the word (actually limit the number of segments). Create a word as described in. Of course, a word may be created without restricting the number of characters when creating the word.

The configuration and operation of the word generator 25 of this embodiment will be described below with reference to FIG.

The word creating section 25 includes the following means (1) to (8) and creates a candidate path represented by the concatenation of the primary segment and / or the secondary segment.

(1) When the coordinates for dividing each of the m primary segments in the X direction are set as the extraction candidate positions C0 to Cm, the extracted extraction candidate positions Ci (i = 0 to 0)
First means for determining whether or not m) is Cm (step 61 in FIG. 8).

(2) The second means for operating when the first means determines that Ci = Cm and recording the current candidate path in a candidate path memory (not shown) (step 6 in FIG. 8)
2).

(3) The operation is performed when Ci ≠ Cm is determined by the first means, and the segment Sk sandwiched between the extraction candidate position Ci and the extraction candidate position Cj (j = i + 1)
Third means for determining whether or not +1 exists (steps 63 to 65 in FIG. 8).

(4) Segment Sk + 1 in the third process
Is operated when it is determined that the candidate path P exists, and when the segment Sk + 1 is added to the candidate path P, a fourth means for determining whether or not the number of segments of the candidate path is within a specified number (step 71 in FIG. 8).

(5) Fifth means (step 66 in FIG. 8) which operates when the fourth means determines that the number is within the specified number, and adds the segment Sk + 1 to the candidate path.

(6) A sixth means (FIG. 9) which operates following the fifth means and starts the operation of the first means by regarding the cut candidate position Cj as the focused cut candidate position Ci. 8 step 67).

(7) For a candidate path created as a result of the operation of the fifth means and the sixth means, a seventh means for deleting the segment most recently added to the candidate path from the candidate path. (Step 68 in FIG. 8).

(8) If the third means determines that no,
When the fourth means determines that the number exceeds the specified number, or when the seventh means is operated, it is operated, and j defining the cutout candidate position is changed to j = j + 1 and changed. Eighth means for determining whether j satisfies j> m in relation to m (steps 69 and 7 in FIG. 8)
0).

(9) A ninth means which operates when the eighth means determines that j ≦ m, and operates the third means (steps 70 and 64 in FIG. 8).

In order to deepen the understanding of the word creation unit 25, a specific example of the candidate path creation processing will be described. However, since the principle of the candidate path creation processing need only be explained, an example in which the number of segments and the number of extraction candidate positions are reduced will be described. That is, the cutout candidate position is C
There are three segments 0 to C2 and three segments S0 to S2, and the relationship between each of the cutout candidate positions C0 to C2 and each of the segments S0 to S2 is represented by a segment table as shown in FIG. The following describes the candidate path creation processing in the case where the relationship is satisfied.

The cutout candidate position C0 is the first cutout candidate position in the direction in which the characters are arranged, and the cutout candidate position C2 is the last (rightmost) cutout candidate position in the direction in which the characters are arranged.

First, the path P in the candidate path memory (not shown) is cleared, and the process for the function Funct (Ci, P), here (C0, P), is first started (step 60 in FIG. 8).

That is, first, the word creating section 25 determines whether or not the cut candidate position Ci, in this case, C0 is the rightmost cut candidate position (ie, whether or not C0 = C2) (FIG. 8).
Step 61). Here, the control unit 11 transfers the extraction candidate position Ci from the segment extraction unit 15 to the word creation unit 25.

Since C0 is not the rightmost cut-out candidate position, the process proceeds to step 63. That is, j =
The processing of i + 1 = 0 + 1 = 1 is performed. As a result, Cj
Becomes C1.

The segment Sk + 1 sandwiched between the extraction candidate position C0 and the extraction candidate position C1 is sent from the segment extraction unit 15 or the segment integration unit 17 to the control unit 11
Is transferred to the word generator 25. In the case of the example of FIG. 9, the segment S0 is transferred. If there is no corresponding segment, the control unit 11 transfers a signal (NULL) to that effect to the word creation unit 25.

The word creating section 25 determines whether or not the segment Sk + 1 exists (Step 65 in FIG. 8). In this case, since segment S0 exists, step 71
Move on to

In step 71, it is determined whether or not the number of segments of a column (segment column) formed by adding the segment Sk + 1 to the candidate path P is within a specified number. Here, it is assumed that the prescribed number is 4 (of course, an example). In this case, since the segment of the segment row is only S0 and the number of segments is 1, the specified number is satisfied.
Move on to

At step 66, processing for adding the segment S0 to the candidate path P is performed. As a result, the candidate path P = {S0}. Thereafter, the process proceeds to step 67.

In step 67, since Cj is set as the target cutout candidate position this time, the cutout candidate position C1 is regarded as the target cutout candidate position Ci. That is, the function is set to Funct (C1, P). Then, the processing is restarted from the processing of step 61.

Therefore, the word creating section 25 determines whether or not the cutout candidate position C1 is the rightmost cutout candidate position (ie, whether or not C1 = C2) (step 61 in FIG. 8).

Here, since C1 is not the rightmost cut-out candidate position, the process proceeds to step 63. That is, j =
The processing of i + 1 = 1 + 1 = 2 is performed. As a result, C
j becomes C2.

The segment Sk + 1 sandwiched between the extraction candidate position C1 and the extraction candidate position C2 is sent from the segment extraction unit 15 or the segment integration unit 17 to the control unit 11
Is transferred to the word generator 25. In the case of the example of FIG. 9, the segment S1 is transferred.

The word generator 25 determines whether or not the segment Sk + 1 exists (Step 65 in FIG. 8). In this case, since segment S1 exists, step 71
Move on to

In step 71, it is determined whether or not the number of segments in a column (segment column) formed by adding the segment Sk + 1 to the candidate path P is within a specified number. In this case, the number of segments in the segment row is two, S0 and S1, so that the specified number is satisfied.

At step 66, processing for adding the segment S1 to the candidate path P is performed. As a result, the candidate path P = {S0, S1}. Thereafter, the process proceeds to step 67.

In step 67, the extraction candidate position C2 is regarded as the extraction candidate position Ci of interest. That is, the function is set to Funct (C2, P). Then, the processing is restarted from the processing of step 61.

Therefore, the word creation section 25 determines whether or not the cutout candidate position C2 is the rightmost cutout candidate position (ie, whether or not C2 = C2) (step 61 in FIG. 8).

Here, since C2 is the rightmost cut-out candidate position, the process proceeds to step 62. Therefore, the candidate path P = {S0, S1} is recorded in the candidate path memory (not shown). As a result, a candidate path P = {S0, S1} is created as one of the candidate paths whose start point is C0, whose end point is C2, and whose number of segments is equal to or less than the specified number. Also, by the processing up to this point, the function Funct
The processing of (C2, P) ends.

This candidate path P = {S0, S1} is the fifth
And candidate paths created as a result of the operation of the means and the sixth means. That is, it is a candidate path created as a result of the processing of steps 66 and 67. Therefore, the process proceeds to step 68. In this step 68, the candidate path P
= S0, S1}, the segment S1 newly added to this is deleted. As a result, the candidate path P
= {S0}.

Next, j = j + 1 is set (step 69 in FIG. 8). Here, the current j is 2, so j = 2 + 1 =
It becomes 3.

Next, it is determined whether or not j is the maximum clipping candidate position (whether or not j> m) (step 70 in FIG. 8).

Since j = 3 in this case is beyond the maximum clipping candidate position 2, the processing of the function Funct (C1, P) ends. Therefore, this time the original function Fun
The processing from step 68 is performed for ct (C0, P).

Therefore, from the candidate path P = {S0},
Then, a process of deleting the segment S0 added most recently is performed. As a result, the candidate path P = {｛} = 0.

Next, j = j + 1 is set (step 69 in FIG. 8). Here, the current j is 1, so j = 1 + 1 =
It becomes 2.

Next, it is determined whether or not j is the maximum clipping candidate position (whether or not j> m) (step 70 in FIG. 8).

Since j = 2 in this case does not exceed the maximum clipping candidate position 2, the processing from step 64 is performed. Therefore, it is determined whether or not there is a segment Sk + 1 sandwiched between the extraction candidate position C0 and the extraction candidate position C2. Segment Sk in this case
Since the segment S2 exists as +1 (see FIG. 13), it is determined whether or not the number of segments in the segment string obtained by adding the segment S2 to the candidate path P is within a specified number.

Since the segment of this segment row is only S2, it satisfies the specified number 3 or less. Therefore, the segment S2 is added to the candidate path P. As a result, the candidate path P = {S2}. Thereafter, the process proceeds to step 67.

Since Cj is 2 in this case, in this step 67, the extraction candidate position C2 is regarded as the extraction candidate position Ci of interest. That is, the function is set to Funct (C2, P). And step 61
Processing is restarted from the processing of.

Therefore, the word creating section 25 determines whether or not the cutout candidate position C2 is the rightmost cutout candidate position (ie, whether or not C2 = C2) (step 61 in FIG. 8).

Since C2 is the rightmost cut-out candidate position, the process proceeds to step 62. Therefore, the candidate path P = {S2} is recorded in the candidate path memory (not shown). Thus, the starting point is C0, the ending point is C2,
In addition, a candidate path P = {S2} is created as one of the candidate paths whose number of segments is equal to or less than the specified number. Further, the processing for the function Funct (C2, P) is completed by the processing up to this point.

This candidate path P = {S2} is a candidate path created as a result of the operation of the fifth means and the sixth means. That is, it is a candidate path created as a result of the processing of steps 66 and 67. Therefore, the process proceeds to step 68. In this step 68, the candidate path P = ｛S
From 2｝, processing is performed to delete the segment S2 that has been added to this most recently. As a result, the candidate path P = {｛} = 0
become.

Next, j = j + 1 is set (step 69 in FIG. 8). Here, the current j is 2, so j = 2 + 1 =
It becomes 3.

Next, it is determined whether or not j is the maximum clipping candidate position (whether or not j> m) (step 70 in FIG. 8).

Since j = 3 in this case exceeds the maximum clipping candidate position 2, the processing of the function Funct (C0, P) ends.

The process described with reference to FIG. 8 is a process called a recursive algorithm. This is a process capable of recursively tracing from the leftmost segment of the character string to the rightmost segment.

According to this process, all segments starting from C0 are extracted. In addition, other connected segments with the end point of the extracted segment as a starting point are further sequentially extracted. In addition, a candidate path composed of a sequence of segments (including one segment) whose start point is C0 and whose end point is C2 and whose number of segments is equal to or less than a specified number is easily created.

The processing described with reference to FIG. 8 is applied to the primary segment S0 of the original image data 40 shown in FIG.
To S2 and the secondary segment S3 to create a candidate path, the candidate path P0 of “S0 → S1 → S2”
And a candidate path P1 of “S3 → S2” is created. Many words are created by combining the recognition candidates for each of the segments S0, S1, and S2 that constitute the candidate path P0. However, here, for convenience of explanation, as an example of the word based on the candidate path P0, an example of the word “Yumu 3” is considered. Also, S3 constituting the candidate path P1
A number of words are created by combining the recognition candidates for each of the segments S2 and S2. However, here, for convenience of explanation, an example of the word “Kozo” will be considered as an example of the word based on the candidate path P1.

(Word Matching Unit) The word matching unit 27 checks the word created by the word creating unit 25 with a word dictionary (not shown). In the case of the above example,
The word "3" and the word "Kozo" are checked against the word dictionary.

(Word Level Evaluation Value Calculation Unit) The word level evaluation value calculation unit 29 determines an evaluation value for each of the created words in accordance with whether or not the word created by the word creation unit 25 exists in the word dictionary. Ask for. Here, when the created word exists in the word dictionary, the evaluation value of “1” is used. When the word does not exist, the evaluation value of “0” is used. Store as level evaluation value. Note that in this embodiment,
Since the word "3" does not exist in the word dictionary, the word level evaluation value is "0". Since the word "Kozo" exists in the word dictionary, the word level evaluation value is "1". It is assumed that the data is stored in the evaluation value calculation unit 29.

(Character Level Evaluation Value Calculation Unit) The character level evaluation value calculation unit 31 calculates, for each word created by the word creation unit 25, the shape feature evaluation value for the constituent segment of each character constituting the word. , Recognition evaluation value and evaluation value based on the number of segments (this is called character level evaluation value)
Ask for. This will be specifically described with reference to FIGS.

FIG. 10A shows the candidate path P1 created by the word creating section. That is, the word “Kozo” is shown here. As can be seen from FIG. 2, this word is composed of a secondary segment S3 called "Hiro" and a "San"
And a primary segment S2.

The shape evaluation value and the recognition evaluation value in the case of the example of FIG. 10A are obtained for the secondary segment “Hiro” and for the primary segment “III”, respectively. The method of obtaining these evaluation values is as described above. Here, the shape evaluation value and the recognition evaluation value for the segment “Hiro” are shown in FIG.
It is assumed that the shape evaluation value and the recognition evaluation value for the segment “3” are the values shown in FIG.

The constituent segments of the character "Hiro" are:
As described above, this is the secondary segment S3 called “Hiro”.
However, since the secondary segment "Hiro" is a combination of the two primary segments "bow" and "mu" as described above, the character "hiro" is eventually formed. The number of segments is two. Also,
As can be seen from FIG. 2, the constituent segment of the character "3" is only one primary segment "3", so the number of segments forming the character "3" is 1
It turns out that.

Next, the character level evaluation value for the word “Kozo” is obtained based on the shape evaluation value and the recognition evaluation value shown in FIGS. 10B and 10C, and the number of segments. Here, the character level evaluation value is obtained according to the following equation (2).

[0134]

(Equation 1)

Here, in equation (2), P _j is a set of candidate paths, α _i is the number of segments of the constituent segment for the i-th character (2 for “Hiro” and 1 for “3”), K _i is the shape evaluation value of the constituent segment for the _i- th character, and R _i is the recognition evaluation value of the constituent segment for the i-th character.

When the character level evaluation value for the above-mentioned word “Kozo” is obtained by the above equation (2), the result is as follows.

Max [2 × 0.83 × (0,0,0,1) + 0.89 × (0,0.4,0.4,0.2)] = max [(0,0,0, 1.66) + (0,0.356,0.356,0.178) = max [(0 + 0,0 + 0.356,0 + 0.356,1.66 + 0.178)] = max [(0,0.0) 356, 0.356, 1.838)] = 1.838.

That is, max [(0, 0.356,
0.356, 1.838)], the maximum value of 1.838 is the character level evaluation value.

Note that max [(0, 0.356, 0.3
56, 1.838)].
Each of 356 and 1.838 is a recognition evaluation value for the symbol "", a recognition evaluation value for katakana, a recognition evaluation value for hiragana, and a recognition evaluation value for kanji for the word "Kozo" (see later). Of FIG. 10
1).

Further, the character level evaluation values of the other words created by the word creating section 25 are obtained by the above procedure.
Here, the character level evaluation value for the word “Yu-mu-san” in the candidate path P0 is obtained as another word.

The word "bow" is composed of the primary segments S0, S1, and S2, as described above. It is assumed that the shape evaluation value and the recognition evaluation value of each segment S0 to S2 are as described in each column of S0 to S2 in FIG. Then, the bow
The character level evaluation value for the word "3" is obtained by the procedure for "Kozo" described above, and the candidate path P in FIG.
As shown in the column of 0, it is 0.692.

(Result Selection Unit) The result selection unit 33 first compares the word level evaluation values of the words created by the word creation unit, and if there is superiority or inferiority, cuts out the superior word in a character area. If no superiority or inferiority is found, the character level evaluation values are compared with each other, and the superior word is selected as a word for extracting a character area.

More specifically, the above “Kozo” and “Yu
In the case of each of the words "3", the result selection unit 3
No. 3 indicates that the word level evaluation value for “Kozo” is “1” and the word level evaluation value for “Yumu M3” is “0” in the comparison stage between the word level evaluation values.
Select the word

Processing when two or more words having a word level evaluation value of "1" exist in the words created by the word creation unit 25, and no comparison is made between the word level evaluation values. Will be described later (other processing examples).

(Result Output Unit) The result output unit 35 outputs to the control unit 11, for example, the cutout candidate positions for the characters (segments) constituting the word determined by the result selection unit 33. In the above example, "Kozo" is the determined word, "Hiro" and "3" are the characters (segments) constituting this word, and C
0, C2, and C3 are output to the control unit 11. The control unit 11 can instruct character cutout based on the result.

(Other Processing Examples) In the above-described example, when the result selecting unit 33 selects a word for determining a character cut-out area from a plurality of words created by the word creating unit 25, the word level evaluation value Only the comparison made that choice. But,
When the result selecting unit 33 selects a word for determining a character cutout area from a plurality of words created by the word creating unit 25,
There may be a plurality of words whose word level evaluation value is “1”. In that case, a word for determining a character cutout area is selected using the character level evaluation value. Such an example will be described below.

In this case, as shown in FIG.
Let us consider an example of original image data 101. In the following description, segment extraction, segment integration, calculation of each evaluation value, character recognition, word creation, word matching, etc.
Since the processing is performed in each procedure already described in detail using the example of the original image data 40, the details are omitted in the following description.

From the original image data 101 of "Yabuki", the segment extracting section 15 extracts a primary segment. Here, three primary segments S0, S1, and S2 of "arrow", "mouth", and "missing" are respectively extracted. Note that FIG. 12 shows coordinates that separate the primary segments S0 to S2 in the direction in which the characters are arranged (in this example, the X direction), that is, the cutout candidate positions C0 to C3.

Next, the segment integrating section 17 integrates the three primary segments S0, S1, and S2 of "arrow", "mouth", and "missing". As a result, a secondary segment S3 called “knowledge” and a secondary segment S4 called “blown” are created (see FIG. 12).

Further, these primary segments S0 to S2
The segment coordinate table for the secondary segments S3 and S4 is as shown in the column of Xs to Ye in FIG. Further, the segment shape evaluation value calculation unit 19 calculates the primary segments S0 to S2 and the secondary segments S3 and S3.
A shape evaluation value for S4 is obtained based on the above-described equation (1). The obtained shape evaluation values are the values shown in the column of the shape evaluation values in FIGS.

Further, these primary segments S0 to S2,
The segment table created from the secondary segments S3 and S4 is as shown in FIG.

The character recognizing section 21 performs character recognition on each of the primary segments S0 to S2 and the secondary segments S3 and S4, and obtains a recognition result up to the K-th position for each segment. It is assumed that the recognition result of each of the segments S0 to S4 obtained in this manner is a result as shown in FIG.

The recognition evaluation value calculation unit 23 obtains recognition evaluation values for four types of characters for each of the segments S0 to S4. The obtained recognition evaluation value is shown in FIG.
Are the values shown in the column of the recognition evaluation value.

The word creating section 25 determines whether each of the segments S0
A word is created by combining the recognition results of steps S4 to S4. Expressing the created word as a candidate path is as shown in FIG. That is, a candidate path composed of segments S0-S1-S2 as a candidate path P0 and a candidate path P1
, A candidate path composed of segments S3-S2 and a candidate path composed of segments S0-S3 are generated as candidate paths P2. The actual word is a combination of the candidate results for each segment,
There are many candidate paths P0, such as "missing Yaguchi" and "father Yaguchi". There are many candidate paths P1 and P2.

The word collating unit 27 collates many words created by the word creating unit 25 with a word dictionary (not shown). Here, it is assumed that none of the words created by the candidate path P0 exists in the word dictionary. It is also assumed that the word “knowledge” among the words created by the candidate path P1 exists in the word dictionary. It is also assumed that the word “Yabuki” among the words created by the candidate path P2 exists in the word dictionary. Therefore, the word level evaluation value calculation unit 29 stores “0” as the word level evaluation value for the candidate path P0, and also stores the word “knowledge” by the candidate path P1 and the word “Yabuki” by the candidate path P2. "1" is stored as the level evaluation value.

On the other hand, the character level evaluation value calculation unit 31 obtains the character level evaluation values of the candidate paths P0, P1, and P2 based on the shape evaluation value, the recognition evaluation value, and the number of segments as described above. The character level evaluation values of the candidate paths P0, P1, and P2 obtained in this case are 1.81, 1.88, and 2.245. These character level evaluation values are shown in the column of character level evaluation value in FIG.

The result selecting section 33 first compares the word level evaluation values of the words created by the word creating section.
However, for the original image data 101 called "Yabuki",
As described above, the words created therefrom include “knowledge” and “yabuki” as words existing in the word dictionary. Therefore, the comparison between the word level evaluation values does not give any advantage.

Then, the result selecting section 33 compares the character level evaluation values with each other this time. Then, the character level evaluation value for the word “knowledge” is 1.88,
The character level evaluation value for the word "Yabuki" is 2.
Since it is 245, the word “Yabuki” is selected as a word for extracting a character area.

In the case of the prior art, if the word "knowledge" is collated first, the word "knowledge" is selected as a word for extracting a character area. Then, the character is cut out as "knowledge" and "look". However, according to the present invention, "yabuki" is selected without selecting "knowledge" as a word for extracting a character area. As a result, character cutout
It is performed like "arrow" and "blow". Therefore,
Even if there are a plurality of words existing in the word dictionary among the words created by the word creating unit 25, in the present invention, a word suitable for the original image can be selected, and therefore, character extraction suitable for the original image can be performed. I can understand.

The result output unit 35 outputs, for example, to the control unit 11 the extraction candidate positions of the characters (segments) constituting the word determined by the result selection unit 33. In the above example, “Yabuki” is the determined word,
And the characters (segments) that make up this word
And C0, C1, and C
3 are output to the control unit 11. The control unit 11 can instruct character cutout based on the result.

[0161]

As is clear from the above description, according to the invention of the character extracting method of the present application, a process of extracting a primary segment, which is a block region of black bits, from original image data, and A process of combining the next segments to create a secondary segment, a process of character recognition of each of the primary and secondary segments, a process of combining candidate character codes obtained by the character recognition to create a word, A character extraction method that includes a process of comparing the extracted word with a word dictionary and a process of determining a character extraction region for one character based on the result of the word comparison. A process of obtaining: a recognition result candidate of each segment is obtained up to the Kth position, and a recognition evaluation indicated by a ratio of a character type in the recognition result candidate For each segment: a process for obtaining a predetermined character level evaluation value for each of the created words; and a word level evaluation indicated by whether or not the created word exists in a word dictionary. A process of obtaining a value: comparing the word level evaluation values of the respective words, selecting the word if the word is superior, and selecting a plurality of the words if a plurality of words having the same word level evaluation value exist. Comparing the character level evaluation values of the words with each other to select one word, and setting each of the segments constituting the selected word as the character cutout region. Therefore, even when there are a plurality of words in the word dictionary among the created words, it is possible to select a word for extracting a character area based on the character level evaluation value.

Here, the character level evaluation value is determined by (1) whether each character constituting the created word is likely to be one character, and (2) the segment of each character having a valid character type. It can be said that this is an evaluation value in consideration of whether or not it is composed of Therefore, if the character level evaluation value is used,
A word including a character that does not have a shape like a single character or a word including a character that is a combination of segments that cannot be normally considered due to the character type can be hardly selected as a word for determining a character cutout region. Therefore, in the present invention using this character level evaluation value, even if there are a plurality of words in the word dictionary among the created words, it can be said that it is easy to select a word suitable for the original image from the plurality of words. . In addition, since the character cutout region can be determined from the word suitable for the original image, the cutout accuracy and the recognition accuracy can be improved.

Further, according to the invention of the character extracting device of this application, the invention of the above-described character extracting method can be easily implemented.

[Brief description of the drawings]

FIG. 1 is an explanatory diagram of a character cutout device according to an embodiment.

FIG. 2 is an explanatory diagram of a primary segment and a secondary segment.

FIGS. 3A and 3B are diagrams showing examples of a segment coordinate table before integration.

FIG. 4 is an explanatory diagram of segment integration.

FIG. 5 is a diagram showing an example of a segment coordinate table after integration.

FIG. 6 is a diagram showing an example of a segment table after integration.

FIG. 7A is an explanatory diagram of a character recognition unit, and FIG. 7B is an explanatory diagram of a segment recognition evaluation value calculation unit.

FIG. 8 is an explanatory diagram of a word creation unit.

FIG. 9 is an explanatory diagram of a specific example of the candidate path creation processing (the operation of the word creation unit).

FIGS. 10A to 10C are explanatory diagrams of a character level evaluation value calculation unit.

11 is a diagram showing a list of evaluation values for original image data 40. FIG.

FIG. 12 is an explanatory diagram (part 1) of another processing example, illustrating a primary segment extracted from original image data “Yabuki” and a secondary segment obtained by integrating these primary segments. is there.

FIG. 13 is an explanatory diagram (part 2) of another processing example, and is a diagram showing a segment coordinate table for segments S0 to S4.

FIG. 14 is an explanatory diagram (part 3) of another processing example, showing a segment table for segments S0 to S4.

FIG. 15 is an explanatory diagram (part 4) of another processing example, showing the character recognition results for segments S0 to S4.

FIG. 16 is an explanatory diagram (part 5) of another processing example, and is an explanatory diagram of a candidate path created from original image data 101;

FIG. 17 is an explanatory diagram (part 6) of another processing example, showing a list of evaluation values for original image data 101;

[Explanation of symbols]

 11: Control unit 13: Image input unit 15: Segment extraction unit 17: Segment integration unit 19: Segment shape evaluation value calculation unit 21: Character recognition unit 23: Segment recognition evaluation value calculation unit 25: Word creation unit 27: Word collation unit 29: Word level evaluation value calculation unit 31: Character level evaluation value calculation unit 33: Result selection unit 35: Result output unit S0 to S2: Segment (primary segment) S3, S4: Segment (secondary segment) C0 to C3: Cutout candidate positions 40, 101: original image data

Claims (6)

[Claims] 1. A process for extracting primary segments each of which is a black bit lump area from original image data including image data of a character string stored in a memory, and extracting each of the extracted primary segments by a predetermined process. Integrate according to the rules 2
A word is created by combining a process of creating a next segment, a process of character recognition of each primary segment and each secondary segment, and a candidate character code of each of the primary segment and secondary segment obtained by the character recognition. The character segmentation method includes: a process of comparing the created word with a word dictionary; and a process of determining a character segmentation area for one character based on a result of the word collation. And a process of obtaining a shape feature evaluation value based on the shape feature of each secondary segment; and determining the recognition result candidate of each segment up to the Kth position at the time of performing the character recognition. Processing for obtaining a recognition evaluation value to be indicated for each segment; and forming the word for each of the created words. The shape feature evaluation value, the recognition evaluation value, and the number of segments (the primary segment is 1 if the primary segment is a secondary segment, and the number of integrated segments if the secondary segment is a secondary segment) A process of calculating a character level evaluation value calculated based on the word processing, a process of calculating a word level evaluation value indicated by whether or not the word exists in the word dictionary when the created word is compared with the word dictionary; As the processing for determining a character cut-out area, the word level evaluation values of each word are compared with each other, and if the word is superior, the word is selected.If there are a plurality of words having the same word level evaluation value, The character level evaluation values of the plurality of words are compared with each other to select one word, and each of the segments constituting the selected word is selected. And a process for setting the character cutout area (where K is a predetermined positive integer). 2. The character extracting method according to claim 1, wherein the shape characteristic evaluation value is an aspect ratio of a segment. 3. The character segmenting method according to claim 1, wherein the secondary segment is defined as follows: when the direction in which characters are arranged is the X direction, m primary segments that are continuous in the X direction are determined according to a predetermined rule. The word is a candidate path represented by a concatenation of the primary segment and / or the secondary segment.
A character segmentation method (where m is an integer of 2 or more), which is created based on a candidate path created by a process including the process (d). (a) Coordinates for dividing each of the m primary segments in the X direction are determined as candidate clipping positions Ci (i = 0 to 0).
m), from among the m primary segments and the created secondary segment, a cutout candidate position C
A process of extracting all the segments for which 0 is the extraction start point. (b) With respect to each of the segments extracted in the process (a), the segmentation candidate position Cj (j = 1 to m) on the end point side of the segment is the segmentation start position, so that the segments can be connected. Other segments which have the same relationship with the other segment and have the same candidate clipping position as described above, and which can be further connected to each other, are replaced by other segments whose candidate candidate positions on the end point side are Cm. A process of extracting from the m primary segments and the created secondary segments until appears. (c) A process of determining whether or not the number of segments of a candidate path including the other segment is within a specified number every time the other segment is extracted in the process of (b). (d) A process in which a candidate path having a number of segments within the specified number and a candidate position for extracting the end point of the last segment in the candidate path being Cm is set as a candidate path for word creation. . 4. The character segmenting method according to claim 1, wherein the predetermined rule when integrating the primary segments is:
When the direction in which the characters are arranged is the X direction, the height H of the segment having the highest height among the m primary segments continuous in the X direction is obtained, and the primary segment of interest is determined. On the other hand, another one existing in the H × N coordinate range in the X direction
A character segmentation method using a rule of integrating a next segment into the primary segment of interest. 5. A segment extraction unit for extracting a primary segment, which is a block area of black bits, from input image data including image data of a character string stored in a memory, and each extracted primary segment. A segment integrating unit that creates a secondary segment by integrating the primary segment and the secondary segment according to predetermined rules, a character recognizing unit that recognizes each primary segment and each secondary segment, and a primary segment and / or a secondary segment obtained by the character recognition. A word creating unit that creates a word by combining the candidate character codes, a word matching unit that matches the created word with a word dictionary, and determines a character cutout area for one character based on the result of the word matching A character selecting unit comprising: a character selecting unit configured to recognize each of the segments; The word matching unit is configured to obtain a word level evaluation value indicated by whether or not the created word exists in a word dictionary. A segment shape evaluation value calculation unit for obtaining a shape feature evaluation value based on the shape characteristics of each of the next segments; and a recognition evaluation value indicated by a ratio of a character type in a recognition result candidate for each segment created by the character recognition unit. A segment recognition evaluation value calculation unit for each of the words, and for each of the created words, the shape feature evaluation value for a constituent segment of each character constituting the word;
A character level evaluation value for calculating a character level evaluation value calculated based on the recognition evaluation value and the number of segments (1 when the constituent segment is a primary segment, the number of integrated segments when the constituent segment is a secondary segment) A calculation unit, and the result selection unit compares the word level evaluation values for each word, and selects a word when the word level evaluation value is higher or lower, thereby selecting a word having the same word level evaluation value. When a plurality of words exist, the character level evaluation values of the plurality of words are compared with each other to select one word, and each of the segments constituting the selected word is set as the character cutout region. A character cutout device (where K is a predetermined positive integer). 6. The character cutout device according to claim 5, wherein the segment shape evaluation value calculation unit is a calculation unit that obtains an evaluation value based on the aspect ratio of the segment.