JPH1069522A - Character recognition device, character recognition method and recording medium recording character recognition method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To judge the state of a character string image like thinning or inter-character contact and the difference of the character kind of English sentences/Japanese sentences, to accurately segment character images for respective character strings and to improve a character recognition rate. SOLUTION: A character string state estimation part 104 estimates whether or not the character string image is an abnormal thinned image or the like based on the width of the circumscribing rectangle of a character element image or the like and a first character kind estimation part 105 estimates whether the character string is the English sentence or the Japanese sentence based on the distribution of the interval of the circumscribing rectangle. An English sentence character segmentation part 106 segments the character image of the English sentence character string, a Japanese sentence character segmentation part 108 segments the character element image corresponding to a reference character width decided by a Japanese sentence reference character width decision part 107 and a character recognition part 109 recognizes the character image. A first character segmentation error judgement part 110 judges the correct/error of character segmentation based on the evaluated value of a recognized result and a second character segmentation error judgement part 112 judges it based on the character kind of the character by the estimation of the English sentence or the Japanese sentence by a second character kind estimation part 111. The character recognition part 109 recognizes a segmentation error character again.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a character recognition apparatus for recognizing characters by cutting out a character image from a document image.
More specifically, the present invention relates to a character recognition device that accurately cuts out a character image from a document image in which Japanese and English texts are mixed for each character string as seen in business cards and advertisement documents and recognizes characters.

[0002]

2. Description of the Related Art In recent years, the field of application of character recognition has been expanding, and it has become necessary to extract characters for character recognition from general documents other than conventional dedicated forms and fixed format documents. Hereinafter, a technique described in Japanese Patent Application Laid-Open No. 63-16391, which is an example of a conventional character recognition device, will be described with reference to the drawings.

FIG. 15 shows a configuration diagram of this character recognition device. The character string detection unit 1501 detects a character string to be recognized from the input document image. Character element detector 15
02 detects a character element from the detected character string image. The reference character width determination unit 1503 determines a reference character width from the width or height of each character element. Character cutout part 1
504 cuts out characters based on the reference character width. The character recognizing unit 1505 recognizes the cut out character.

FIG. 16 shows the specific operation of this device.
This will be described with reference to FIG. The character string detection unit 1501 detects an existing area of the character string 1601 to be recognized as shown in FIG. 16A from the input image. Here, the character string 1601
Height Y and width W are required. Character element detection unit 1502
Are the character strings 160 detected by the character string detection unit 1501.
For example, as shown in FIG. 16 (b), the character portion is projected in a direction perpendicular to the character string direction, and a range w where projection is continuous is performed.
_1, w _2, ... range d _1, d ₂ to w ₁₁ and the projection is interrupted, ...
seek and d _10. The range in which the projections are continuous is defined as one character element, and as shown in FIG.
_1, Ps _2, to detect the ... Ps _11.

[0005] Next, the reference character width determination unit 1503 determines the length Y (height in the case of horizontal writing) Y in the character string direction of the character string 1501 and each character element Ps _i in the character string 1501.
Length of the circumscribed rectangle in the character string direction (width for horizontal writing)
w _i is compared with the maximum value, and the larger value is used as the reference character width A.
And FIG. 16B shows that the width w8 is the reference character width A.

The character cutout unit 1504 defines a rectangular length Cw when n adjacent character elements are combined in the character string direction as Cw = | {wi +} (a · di) |. Here, a is a constant. When Cw ≦ A, adjacent n character elements are combined into one character. Thus, the characters P1, P2,..., P5 are detected as shown in FIG.

The character recognizing unit 1505 includes a character extracting unit 15
The individual characters P1, P2... P5 detected at step 04 are recognized.

[0008]

However, in the character recognition apparatus as described above, when there is contact between characters, the character reference unit 1504 cannot obtain a correct reference character width and cannot extract characters one by one. Therefore, correct character recognition cannot be performed. Similarly, it is not possible to distinguish between an English character string and a Japanese character string. Therefore, in the case of an English character string in which the character width varies greatly depending on the character, sufficient character extraction accuracy cannot be obtained.

Furthermore, the image often fades depending on the printing condition of the document and the conditions at the time of reading the image.
In such a case, the estimated reference character width greatly deviates from the actual character width in the same manner as described above, so that correct character extraction cannot be performed. The present invention, even when the document image is in an abnormal state such as blurring or contact between characters, and also when a mixture of Japanese and English character strings, cut out the correct character image, An object of the present invention is to provide a character recognition device and a method thereof with improved character recognition accuracy.

[0010]

In order to solve the above problems, a character recognition apparatus according to the present invention comprises: a character string image extracting means for extracting a character string image and a position of a circumscribed rectangle from a document image; Character element image extraction means for extracting a character element image and a position of a circumscribed rectangle from the character string image, and a first length parallel to the character string image of the circumscribed rectangle of the character element image (ie, a width of a horizontal writing character string) And a second length in a direction perpendicular to the character string image of the character string image (that is, the height of the horizontal writing character string)
A character string state classifying means for classifying whether the state of the character string image is normal or abnormal at a threshold value determined based on the character string image according to the state of the character string image classified by the character string state classifying means Reference character width determining means for determining a reference length (reference character width) parallel to the character string image from which the character image is cut out, and character element images are integrated and separated according to the reference length determined by the reference character width determining means. And a character recognizing means for recognizing a character image as a corresponding character code by collating a characteristic of the character image with a standard characteristic of a recognition dictionary. Thus, the reference character width for cutting out the character element image is distinguished between a normal image and an abnormal image, so that the accuracy of cutting out the character element image is improved, and the recognition efficiency of character recognition is increased.

[0011]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Hereinafter, a character recognition device according to the present invention will be described with reference to the drawings. FIG. 1 is a configuration diagram of an embodiment of a character recognition device according to the present invention. The character recognition device includes an image input unit 101, a character string detection unit 102, a character element detection unit 103, a character string state estimation unit 104, a first character type estimation unit 105, an English character extraction unit 106,
Japanese character width determining unit 107 and Japanese character extracting unit 1
08, a character recognition unit 109, a first character segmentation error determination unit 110, a second character type estimation unit 111, a second character segmentation error determination unit 112, a reference character width correction unit 113,
A modified character cutout unit 114 and a recognition result output unit 115 are provided.

An image input unit 101 is composed of a scanner or the like, and converts a document including characters prepared by an operator into an L / L.
By performing E (Light / Electric) conversion or the like, the document image data, which is a set of pixels (black pixels corresponding to the character image portion and white pixels corresponding to the background portion) formed of binary data, is converted to the character string detection unit 102. Notify. Upon receiving the notification of the document image data from the image input unit 101, the character string detection unit 102 stores the document image data and scans the document image data to detect the character string image from the black pixel existing area. , The coordinates of the character string image in the first column are obtained. Also,
Upon receiving an instruction to detect a character string image from the character recognition unit 109, the coordinates of the character string images in the second and subsequent columns are obtained. The obtained character string coordinates are notified to the character element detection unit 103 and the character string state estimation unit 104.

[0013] Specifically, as shown in FIG. 2 (a) ~ (d) , the rectangle circumscribing a character string image _{L 1, L 2, L 3} , L 4 upper left end point 201, such as a lower left end point 202 Each coordinate (x
_s, y _s), seeking (x _e, y _e). FIG. 2 (a)
To the character string width _Wi and the character string height Y
_i and the coordinate string state estimating unit 104 described later (x _s, y _s), determined from the value of (x _e, y _e).

The character element detecting unit 103 calculates the character string coordinates (x
_s , y _s ) and (x _e , y _e ), the character string images L _1, L ₂ ,... of the document image data stored in the character string detection unit 102 are respectively circumscribed. Scanning is performed to detect a black pixel connection portion where black pixels are connected as a character element image. Thereby, the run information of the character element image (the coordinates of the start position and the end position of the black pixel for each scanning line)
And the coordinates (x _si , y _si ) and (x _ei , y _ei ) of the upper left and lower right points of the circumscribed rectangle of the character element image.
Here, the subscript i is a number assigned from the top of the character element image for each character string image.

[0015] Figure 2 from the character string image L ₁ shown in (a), the 3 character component image S ₁ as shown (a), the S _2, ... S
₅ and the coordinates (x _si , y _si ) and (x _ei , y _ei ) of the upper left corner 301 and the lower right corner 302 of each circumscribed rectangle are obtained, and notified to the character string state estimation unit 104. Is done. Character string image L ₂ ~L ₄ shown similarly in FIG. 2 (b) ~ (d)
, Character element images as shown in FIGS. 3B to 3D are obtained. Here, the numerical values shown in parentheses below the character element image are the coordinates (x _si , y
_si ), (x _ei , y _ei )
04 shows the width of the character element image obtained.

The character string state estimating unit 104 receives the coordinates (x _s ,
y _s), (x _e, notified of y _e), run information and bounding rectangles of coordinates of each character element image from the character element detection unit _{103 (x si, y si)} , and (x _ei, y _ei) Is notified,
The height Y of each character string image is determined from the coordinate values y _s and y _e .
Also, determine the width wi of each character element image coordinates x _si, and a y _ei.

Next, a first threshold value W1 and a second threshold value W2 for determining whether the character element image included in the character string image is a thin character element or a wide character element are set. It is determined by the equations (Equation 1) and (Equation 2). First threshold value W1
Is the formula (Equation 1)

[0018]

(Equation 1) Is required. α1 is a coefficient, which is selected in the range of 0.1 to 0.5, and α1 = 0.2 in the present embodiment. In the formula (Equation 1), the value of Wmin is introduced and the smaller value is set as the first threshold value W1 in order to cope with the case where the height Y of the character string image is very large. Where Wm
in is a value determined by an experiment, and varies depending on the properties of the target image and the performance of the scanner.
A range of 1 mm is appropriate, and in this embodiment, 1 mm
And Since the coordinate values of the circumscribed rectangle are displayed in pixel units, Wmin = 16.

The second threshold value W2 is given by the following equation (Equation 2).

[0020]

(Equation 2) Is required. Here, α2 is a coefficient, and is 1.0 to 1.5 based on a Japanese sentence having many characters in which the aspect ratio of the characters is almost unity.
Α2 = 1.2 in the present embodiment. Next, the character string state estimating unit 104 divides the character element image by the value of the width wi of the circumscribed rectangle for each character string image.
It is classified into a thin character element image having a first threshold W1 or less, a wide character element image having a second threshold W2 or more, and a normal character element image. The number N1 of the classified narrow-width character element images and the number N2 of the wide-width character element images are respectively obtained, and the ratio indicated by the total number N of the character element images is obtained. When the ratio of the narrow character element images is equal to or more than the predetermined ratio R1, it is determined that the character string image is blurred. When the ratio of the wide character element images is equal to or more than the predetermined ratio R2, the character string image is determined. It is determined that there is much inter-character contact, and otherwise, it is determined to be a normal character string image.

Here, the ratios R1 and R2 are selected within a range of about 0.5 or less. In the present embodiment, R1 = 0.33 and R1
2 = 0.17. For example, in the case of the character string image L ₁ (FIGS. 2A and 3A), since the height Y ₁ of the character string image is 50, the first threshold value W1 = 10 and the second threshold value The value W2 is calculated as 60, and the number N1 of narrow character element images is calculated.
= 0, the number N2 of wide character element images becomes 1, and it is determined that the image is a normal character string image.

Character string image L_Two(FIG. 2B, FIG. 3
In the case of (b)), the height Y of the character string image _Two= 43
And the first threshold value W1 = 8 and the second threshold value W2 = 51.
Calculated and required for all characters with the number of narrow character element images N1 = 13
Since the number of elementary images is 0.5, the character string image L_TwoIs shaved
Is determined to be. Character string image L_Three(FIG. 2 (c), FIG.
3 (c)), the height Y of the character string image _Three= 35
Therefore, the first threshold value W1 = 7 and the second threshold value W2 = 42
Is calculated, and all characters are required when the number of wide character element images is N2 = 4.
Since the original image is 0.21, the character string image L_ThreeIs between letters
It is determined that there is a lot of contact.

The character string image L ₄ (FIG. 2D, FIG. 3)
In the case of (d)), it is determined that the image is a normal character string image. If it is determined that the character string image is blurred, or if the character string image is determined to have a lot of character-to-character contact, the Japanese sentence reference character width determination unit 107 notifies the fact, the height Y of the character string image and the character element. The coordinates of the circumscribed rectangle of the image and the run information are notified.

The character string image is determined to be a normal character string image
At this time, the coordinates (x
_s, Y_s), (X_e, Y_e) And the coordinates (x
_si, Y_si), (X_ei, Y_ei) And run information.
The first character type estimating unit 105 outputs the character string state estimating unit 104
When you receive a notification such as a normal character string image,
Distance D between adjacent circumscribed rectangles of the element image_i= X _{S (i + 1)}−
X_eiAnd a distance D between the rectangles_iof
Range D of peak first appearing from short side_RAsk for. D
1 ≦ D_R≤ D2, the distance D between the circumscribed rectangles_iGinger
Range D_RThe number N3 of the intervals between the rectangles belonging to.

Next, the third threshold value D is calculated from the height Y which is the difference between the coordinate values y _s and y _e of the character string image according to the following equation (Equation 3).

[0026]

(Equation 3) Is determined using Here, the coefficient α3 is selected in a range of about 0.33 or less, and in the present embodiment, α3 = 0.33. The value of the maximum value D2 of the peak range is equal to or less than the third threshold value D, and the number N-1 of character element images of the number N3 of the intervals between the circumscribed rectangles belonging to this range D _R (N-1 It is determined whether or not the ratio to the total number of intervals is equal to or greater than a predetermined constant R3. Note that the constant R3 is selected in a range of about 0.8 or more, and R3 = 0.8 in the present embodiment. If affirmative, it is determined that the character string image is in English, and if negative, it is determined that it is in Japanese. When it is determined to be an English sentence, the coordinates of the circumscribed rectangle of the character element image and the like are notified to the English character extraction unit 106, and when it is determined to be a Japanese sentence, the fact that the image is a normal character string image and the coordinates of the character element image The run information is notified to the Japanese-language reference character width determination unit 107.

[0027] For example, in the case of the character string image L ₁ is 4
The histogram shown in (a) is obtained, and is determined to be a Japanese sentence. In the case of the character string image L4, the height Y _{4 of the} character string image
= 31 (FIG. 2 (d)), the third threshold value D = 10, and as shown in FIG. 4 (b), the value of the adjacent character element image belonging to the first peak range 4 ≦ D _R ≦ 7 of the histogram. The number N3 of intervals is 15, and the ratio to the total number of intervals is 0.9.
4, which is equal to or greater than the constant R3, and is therefore determined to be an English sentence.

Upon receiving the notification of the coordinates of the circumscribed rectangle of the character element image and the run information from the first character type estimation unit 105, the English character extraction unit 106 stores the document image data stored in the character string detection unit 102. Is read out, a character image specific to the alphabet is cut out, and the character image is read out along with the coordinates of the circumscribed rectangle of the character element image and the run information.
Notify 9. The details of cutting out the character image are described in JP-A-62-216801.

When the Japanese reference character width determining unit 107 receives the notification of the height Y of the character string image, the coordinates of the circumscribed rectangle of the character element image, and the run information from the character string state estimating unit 104, the expression (Equation 4) is obtained. Using,

[0030]

(Equation 4) The reference character width A1 is calculated from a value obtained by multiplying the height Y of the character string image by a predetermined coefficient α. Here, α is 0.5 for Japanese characters because many characters have an aspect ratio close to 1 in many cases.
It is selected in the range of about 1.5 to 1.5, and is set to 1 in the present embodiment. For character image images that have been slashed in this way or character string images that have a lot of character contact, even if the reference character width is determined based on the width of each character element image, the accuracy does not increase. The reference character width A1 is determined based on the height Y of the image. The determined reference character width, coordinates of the circumscribed rectangle of the character element image, and run information are notified to the Japanese character extraction unit 108.

In the case of the character string image L ₂ (FIG. 2B), since the height Y ₂ of the character string image is 43, the reference character width A1 is determined to be 43. The character string image L3 (FIG. 2)
In the case of (c)), since the height Y ₃ of the character string image is 35, the reference character width A1 is determined to be 35. Further, when receiving the notification of the coordinates of the circumscribed rectangle of the character element image and the run information from the first character type estimating unit 105, the Japanese character reference character width determination unit 107 calculates the width wi of the circumscribed rectangle by the formula (number).

[0032]

(Equation 5) And a histogram having a width wi of the circumscribed rectangle is created. The range W _R1 of the peak to which the width of the most frequent character element image of the obtained histogram belongs is determined. W3 ≦ W _R1 ≦ W4 The maximum value W4 of this range W _R1 is determined as the reference character width A2. The determined reference character width A2, the coordinates of the circumscribed rectangle of the character element image, and run information are notified to the Japanese character extraction unit 108.

In the case of the above character string image L ₁ (FIG. 2A), a histogram as shown in FIG.
≦ _WR1 ≦ 53, and the reference character width A2 = 53 is determined. The Japanese character extracting section 108 is a Japanese character width determining section 1.
When the notification of the reference character width A1 or A2 is received from 07,
The document image data stored in the character string detection unit 102 is read, and the reference character width A1 or A2
Is used to integrate the character element image into the character image, and notifies the character recognition unit 109 of the character image together with the coordinates of the circumscribed rectangle and the run information. The integration of the character images is performed by the equation (Equation 6) described in Japanese Patent Application Laid-Open No. 63-16391 described in the above prior art.

[0034]

(Equation 6) Is used to integrate adjacent n character element images such that Cw ≦ A1 or Cw ≦ A2. The character images cut out from the character string images L1 to L4 as described above are shown in FIGS. 6A to 6D, respectively.

In the case of the character string image L ₁ (FIG. 6A), it is assumed that the image is a normal image.
In practice, "information" in which characters are in contact with each other is regarded as a one-character image. The character string image L ₃ (FIG. 6C)
In the case of, “Hanji”, which should be two characters, is cut out into two abnormal character images C2 and C3 due to contact between characters.

The character recognition unit 109 has a recognition dictionary (not shown), and recognizes a character code corresponding to the character image notified from the English character extraction unit 106 or the Japanese character extraction unit 108. The recognition dictionary registers an N-dimensional standard feature vector Fij of standard characters and a character code as a set. The character recognizing unit 109 extracts the N-dimensional feature vector fj from the notified character image, and calculates the city area distance vi using Expression (Formula 7).

[0037]

(Equation 7) Here, i is a dictionary number, and j = 1, 2,... N. The character code of the dictionary number i with the smallest city area distance vi is used as a recognition result, and the value of this city area distance v is used as an evaluation value v, along with the character code and the coordinates of the circumscribed rectangle of the character image. Notify 110.

This evaluation value is a measure of the likelihood of the character code of the recognition result. That is, if the feature vector fj of the character image to be recognized is the same as the standard feature vector Fij, the urban area distance vi = 0, and the evaluation value is 0. The smaller the evaluation value, the more likely the character code of the recognition result is. The results of recognizing the character images C ₁ , C _2, ... Of the character string images L1 to L4 as character codes in this manner are described below the character images in FIGS. 6 (a) to 6 (d). In FIG. 6, a character code is indicated by a character represented by the character code. Below the character code, an evaluation value of the recognition result is shown.

The character recognizing unit 109 recognizes a character code corresponding to the character image notified from the corrected character extracting unit 114 in the same manner as described above. In this case, if the correction character cutout unit 114 has received a notification as to whether it is a Japanese character or a correction to an English character code, it is compared only with the standard feature vector of the recognition dictionary corresponding to the code. When the character code as the recognition result is recognized, the character image without correction and the corrected character cutout unit 114 previously notified from the English character cutout unit 106 or the Japanese character cutout unit 108
The character code of the character image and the coordinates of the circumscribed rectangle and the evaluation value of the corrected character image are output to the recognition result output unit 115.
When the character recognition of one character string image is completed, the character recognition unit 109 instructs the character string detection unit 102 to detect the next character string image.

The first character segmentation error determination section 110 stores a threshold value R of the evaluation value in advance. Character recognition unit 10
It is determined whether or not the evaluation value v, which is the value of the city distance notified from 9 with the character code, exceeds the threshold value R,
If the result is affirmative, it is determined that a character segmentation error has occurred, and the fact of the error, the character code thereof, and the coordinates of the circumscribed rectangle of the character image are stored. If the result is negative, it is determined that character extraction is not an error, and the character code and the coordinates of the circumscribed rectangle of the character image are stored. The second character type estimating unit 111 calculates a character code in which character extraction is not incorrect and coordinates of a circumscribed rectangle of the character image.
Notify sequentially.

Here, when it is determined whether or not a character segmentation error has occurred with the threshold value R = 800, as shown in FIG. 6, the character image C ₁ “information” of the character string image L ₁ and the character string image L 2 The character image C ₄ “west” and the character images C ₂ and C _{3 of} the character string image L3 are determined to be character extraction errors. The threshold value R is appropriately selected based on the recognition dictionary standard feature vector for calculating the city distance vi and the numerical value of the extracted feature vector.

The second character type estimating section 111 stores a list of character codes and character types (Kanji, Hiragana, Katakana, etc.) and a predetermined constant R4. The character codes notified from the first character segmentation error determination unit 110 are classified into character types, and the number N0 of the notified character codes is counted.
When the processing of one character string image is completed, the number N of Japanese-only characters (kanji, hiragana, katakana) among the character types classified next is N
4 to determine whether N4 / N0 ≧ R4. When the result is affirmative, the one character string image is estimated to be a Japanese character string, and the second character extraction error determination unit 112 is notified. If the result is negative, the one character string image is estimated to be an English character string, and the second character cutout error determination unit 112 is notified.

Here, the value of the constant R4 is selected in a range of about 0.5 or more, and in the present embodiment, R4 = 0.9.
In addition, along with another notification of Japanese and English strings,
The character code and the coordinates of the circumscribed rectangle of the character image notified from the first character segmentation error determination unit 110 are notified to the second character segmentation error determination unit 112.

The second character segmentation error determination unit 112 stores a list of character codes and corresponding character types, similarly to the second character type estimation unit 111. Second character type estimation unit 1
11 receives a notification that the character string is an English character string, the character code, and the coordinates of the circumscribed rectangle of the character image, and searches whether or not the character type of the character corresponding to the character code is a Japanese character type. , The character code of the recognition result is determined to be the result of a character cutout error, and the character code, the coordinates of the circumscribed rectangle of the character image, and the fact that the character code is to be corrected are stored. If it is not a Japanese-only character type, it is determined that the character code of the recognition result is correct.

When the second character type estimating unit 111 notifies that the character string is a Japanese character string, the character code and the coordinates of the circumscribed rectangle of the character image, the character type of the character corresponding to the character code is set to the English character type. Whether the character type corresponding to the preceding and following character codes is a Japanese-only character type.
If so, the notified character code is determined to be the result of a character extraction error, and the character code, the coordinates of the circumscribed rectangle of the character image, and the fact that the character code is to be corrected to the Japanese character code are stored. When the processing of one character string image is completed, the reference character width correction unit 113 is activated.

When activated by the second character type estimating unit 111, the reference character width correcting unit 113 determines that the first character segmentation error determination unit 110 has detected a character segmentation error and stores the character corresponding to the character code stored therein. The coordinates of the circumscribed rectangle of the image or the coordinates of the circumscribed rectangle of the character image corresponding to the character code stored in the second character segmentation error determination unit 112 are read. Next, a character code other than these character codes (the first character cutout error determining unit 1 determines that there is no character cutout error).
10 and the coordinates of the circumscribed rectangle corresponding to the character code of the character image determined by the second character extraction error determination unit 112 are read from the first character extraction error determination unit 110. From the coordinates of the circumscribed rectangle, a histogram of the width of the circumscribed rectangle of the character image is created by a method similar to that of the Japanese text reference character width determination unit 107. The range W _R2 of the peak to which the width of the most frequent character image of the obtained histogram belongs is determined.

W5 ≦ W _R2 ≦ W6 The maximum value W6 of this range W _R2 is determined as the modified reference character width A3. The corrected character cutout unit 114 is notified of the determined correction reference character width A3 and the coordinates of the circumscribed rectangle of the character image read out as a character cutout error. FIG. 7 (a)
Is a histogram of the circumscribed rectangle of the width of the character string image L ₁ in cutout character image is determined not to be erroneous. The correction reference character width A3 = 53 is determined from this histogram.

[0048] Similarly, FIG. 7 (b) is a histogram of the character string image L _2, FIG. 7 (c) is a histogram of the character string image L _3. In modified from the histogram reference character width A3 string image L ₂ A3 = 41 and the character string image L
_In A3, it is determined that A3 = 38. Upon receiving the notification of the correction reference character width A3 from the reference character width correction unit 113, the correction character cutout unit 114 and the correction area and the correction area (width) of the character image obtained from the coordinates of the circumscribed rectangle of the character image notified at the same time. Compare with the reference character width A3. When the corrected reference character width A3 is large, continuous character images are determined to be blurred characters, and a character image that matches the corrected reference character width A3 is referred to the document image data stored in the character string detection unit 102. The cut-out information and the coordinates of the circumscribed rectangle are notified to the character recognition unit 109.

When the modified reference character width A3 is small, it is determined that the character has a character-to-character contact, and the area (width) of the character image is divided by the modified reference character width A3 to obtain the number of characters. Character image detected by character string detector 1
02, referring to the document image data stored in the character recognition unit 10 based on the run information and the coordinates of the circumscribed rectangle.
Notify 9. At this time, the second character segmentation error determination unit 1
Referring to FIG. 12, when it is stored that the character segmentation error is a correction to a Japanese character or English character code, the fact is also notified.

As a result, the character recognition unit 109 described above
And a character string image L as shown in FIG.₁Character painting
Statue C₁(FIG. 6 (a)) shows a correct character image "information" C_C1When
"Report" C_C2And will be recognized. Also, as shown in FIG.
As described above, the character string image L determined to have been cut out error._TwoCharacter image
C (FIG. 6B) indicates that the recognition evaluation value di is again equal to the threshold R
Character image "west" C with a value larger than_C4Recognized
You. This is character image C_FourEvaluation value is large due to
Value, and the recognition result was incorrect.
Because there is no. A character string image as shown in FIG.
L_ThreeCharacter image C _Two, C_Three(FIG. 6 (c)) shows the correct sentence
Character image "Saka" C_C2And "support" C_C3And will be recognized.

The recognition result output unit 115 is provided for the character recognition unit 10.
9 is converted to a corresponding standard character image and displayed on the display screen. Next, the operation of the present embodiment will be described with reference to the flowcharts of FIGS.
First, when an input of a document original is received from an operator, the image input unit 101 converts the input into document image data including black pixels corresponding to a character image and white pixels corresponding to a background part, and notifies the character string detection unit 102 of the image data. (S902).

The character string detecting section 102 stores the document image data, detects a character string image from the document image data,
The character string image is notified to the character element detection unit 103 (S90
4). The character string state estimating unit 104
When the notification of the character string image is received from
Is calculated (S906).

The character element detection unit 103 includes the character string detection unit 1
When the notification of the character string image is received from No. 02, the black pixel connection part is detected as a character element image and the number N of character element images
Is counted (S908). The character string state estimating unit 104
The first threshold value W1 is calculated using the equation (Equation 1). The first threshold value W1 is a value indicating whether or not the character element image is a narrow character element image (S910). The second threshold value W2 is calculated using the equation (Equation 2). The second threshold value W2 is a value indicating whether or not the character element image is a bold character element image (S912). Next, the character string state estimation unit 104 counts the number N1 of character element images in which the width of the circumscribed rectangle of the character element image is equal to or smaller than the first threshold value W1 (S9).
14) Similarly, the number N2 of character element images in which the width of the circumscribed rectangle is equal to or larger than the second threshold value W2 is counted (S916). N1 /
It is determined whether the value of N is equal to or greater than a predetermined ratio R1 (S
918) If above, it is determined that there are many blurred characters, and S9
Move to 22. If less than, it is determined whether or not the value of N2 / N is equal to or greater than a predetermined ratio R2 (S920). If it is more than that, it is determined that there is much inter-character contact, and the flow proceeds to S922.

In S922, the Japanese sentence reference character width determination unit 109 calculates the reference character width A1 using the equation (Equation 4), and proceeds to S1022. If it is determined in S920 that the character image is less than the normal character string image, the first character type estimation unit 105 creates a histogram of the interval between circumscribed rectangles of adjacent character element images (S1002), The range of the first peak of the histogram is obtained (S100
4) Count the number N3 of character element images falling within the range (S1006).

Next, the third threshold value D is calculated using the equation (Equation 3).
Was calculated (S1008), the maximum value D ₂ distance D _R between the circumscribed rectangle of the first peak following the third threshold value D, and, N
It is determined whether the value of 3 / N is equal to or greater than a predetermined ratio R3 (S1010). Thereby, it is determined whether the character string image is an English sentence or a Japanese sentence. If affirmative, it is determined that the character string image is in English, and the English character extraction unit 106
Performs the integration process of the minute character element image based on the run information of the character element image (S1012), cuts out the character image (S1014), and proceeds to S1102.

If a negative determination is made in S1010, the character string image is determined to be a Japanese sentence, and the Japanese sentence reference character width determination unit 107 creates a histogram of the circumscribed rectangle width of the character element image (S1016). The peak range W _R1 to which the most frequent circumscribed rectangle width belongs is determined (S1018). The maximum value W4 of the range W _R1 is determined as the reference character width A2 (S
1020).

The Japanese character extraction unit 108 outputs the reference character width A1 or A
The character element images are integrated and cut out using Expression (Equation 6) so as to obtain a character image having a circumscribed rectangular width _CW of 2 or less (S10).
22). When the character recognition unit 109 receives the notification of the character image cut out from the English character cutout unit 106 or the Japanese character cutout unit 108, the character recognition unit 109 extracts the feature vector and calculates the city distance vi with all the standard feature vectors in the recognition dictionary. It is calculated using the equation (Equation 7), and the character code of the standard feature vector having the minimum value vi is used as the recognition result, and is notified to the first character segmentation error determination unit 110 together with the evaluation value v (S1).
102).

The first character segmentation error determination section 110 stores the value of the notified evaluation value v in a threshold R
Is determined (S1104), and if affirmative, the process moves to S1110. If the result is negative, it is determined that the character recognition is correct, and the second character type estimating unit 111 classifies the character codes notified from the first character segmentation error determining unit 110 into character types (Kanji, Hiragana, Katakana, and others) ( S110
6) Count the number N0 of the notified character codes (S1)
108). If the processing of one character string image has not been completed in S1110, the process returns to S1102. When the processing is completed, the number N4 of Japanese-only characters is calculated (S1112), and N4 / N0 is obtained.
Is determined to be greater than or equal to the constant R4 stored in advance (S1114), and if affirmative, the process moves to S1202.

If not, it is determined that the character string image is an English character string, and the second character segmentation error determination unit 112
Determines whether the character code notified via the second character type estimating unit 111 is a Japanese character only character type (S111).
6) If affirmative, store it as a character segmentation error (S
1118) If not, if the processing of the one character string image has not been completed (S1120), the process returns to S1116, and if completed, the process proceeds to S1210.

In step S1202, the second character segmentation error determination unit 112 determines that the character string image is a Japanese character string, and sets the character code notified via the second character type estimation unit to the English character type (alphabetic, numeric, (S1202), and if not, S1208.
If it is affirmative, it is determined whether any of the preceding and following character codes is a Japanese character type (S1204). If the result is negative, the process proceeds to S1208. Is stored (S1206). In step S1208, it is determined whether the processing of the one character string image has been completed.
It returns to 1202, and if it has been completed, it will transfer to S1210.

At S1210, reference character width correction unit 1
In step S1210, a histogram of the circumscribed rectangle width of the character image determined not to be an extraction error is created (S1210), and a peak range W _R2 to which the most frequent circumscribed rectangle width of the histogram belongs is obtained (S1212). The maximum value W6 of the range W _R2 is notified to the corrected character cutout unit 114 as the corrected reference character width A3 (S1214).

The corrected character cutout unit 114 corrects and cuts out a character image in which a character cutout error has occurred based on the notified correction reference character width A3, and notifies the character recognition unit 109 (S1216). The character recognition unit 1218 determines the character code by comparing the notified character image with the recognition dictionary. At this time, if a notification that the character type is an English character type or a Japanese character type is received, the collation is performed within the target character type (S
1218). It is determined whether the processing of the one character string image has been completed (S1220).
6, if it has been completed, it is determined whether or not processing of all character string images has been completed. If it has not been completed, the process returns to S906. If it has been completed, the recognition result output unit 115 Outputs a character image corresponding to the character code that is the recognition result, and ends the process (S1222).

Although this embodiment has been described by taking a horizontally written character string as an example, it is needless to say that the present invention can be applied to a vertically written character string. In this case, the length in the character string direction is treated as “height”, and the length in the direction perpendicular to the character string is treated as “width”. In this embodiment, the character recognition device according to the present invention is realized by the configuration shown in FIG. 1, but the present invention is realized by a program, which is recorded on a recording medium such as a floppy disk and transferred. By doing so, it can be easily implemented on another independent computer system. FIG. 13 is a diagram for explaining a case where this is carried out using a floppy disk.

Floppy disk 1 as a recording medium body
In the physical format 301, tracks 1, 2,..., 80 are created concentrically from the outer periphery to the inner periphery, and are divided into 16 sectors in the angular direction. The program is recorded according to the allocated area. The floppy disk 1301 is housed in a case 1302, so that the disk can be protected from dust and external impact and can be transported safely.

FIG. 14 is a diagram for explaining recording and reproduction of a program on the floppy disk 1301.
By connecting the floppy disk drive 1401 to the computer system 1403 as shown,
The program can be recorded and reproduced on the disk 1301. The disk 1301 is inserted into and removed from a floppy disk drive 1401 via an insertion slot 1402. When recording, the program is recorded on the disk 1301 by the floppy disk drive 1401 from the computer system 1403. For reproduction, the floppy disk drive 1401 reads the program from the disk 1301 and transfers it to the computer system 1403.

In this embodiment, a description has been given using a floppy disk as a recording medium.
The same can be done using an optical disk. The recording medium is not limited to these, but can be similarly implemented as long as the program can be recorded thereon, such as an IC card or a ROM cassette.

[0067]

As described above, according to the present invention, a character string image extracting means for extracting a character string image and a position of a circumscribed rectangle from a document image, a character element image and a circumscribed rectangle from the character string image Character element image extracting means for extracting the position of the character element image, a first length parallel to the character string image of the circumscribed rectangle of the character element image, and a second length perpendicular to the character string image of the character string image. A character string state classifying means for classifying whether the state of the character string image is normal or abnormal with a threshold value determined by the threshold value, and a character image from the character string image according to the state of the character string image classified by the character string state classifying means. Character width determining means for determining a reference length parallel to the character string image to be cut out, and a character image for integrating and separating the character element images according to the reference length determined by the reference character width determining means to cut out the character image Cutting means and The character string state classifying means determines that the character string state is normal because the character string state classifying means includes character recognition means for recognizing the character of the cut-out character image with the standard characteristic of the recognition dictionary and recognizing the character image as a corresponding character code. The reference character width determination means can determine the cutout width of the character image according to the classification. Therefore, the accuracy of extracting a character image is improved, and the efficiency of character recognition is improved.

The character string state classifying means counts the total number N of character element images for each character string image, and calculates a first length of each character element image for each character string image. A first length calculation unit, and a first threshold value W1 calculated based on a second length Y of a circumscribed rectangle of the character string image, and a circumscribed rectangle of each character element image calculated by the first length calculation unit Is compared with the first threshold value W1, the number N1 of narrow character elements whose first length is equal to or less than the first threshold value W1 is counted, and N1 /
A blurred image determination unit that determines a character string image as an abnormal blurred image when the value of N is equal to or greater than a predetermined ratio R1, and a second threshold value W2 based on a second length Y of a circumscribed rectangle of the character string image And comparing the first length of the circumscribed rectangle of each character element image calculated by the first length calculation unit with a second threshold value W2,
Number N of bold character elements whose first length is greater than or equal to second threshold value W2
2, a contact image determining unit that determines a character string image to be an image having many abnormal character-to-character contacts when the value of N2 / N is equal to or greater than a predetermined ratio R2, and the blurred image determining unit or the contact image determining unit And a normal image determining unit that determines the character string image to be a normal image when the image is not determined to be an abnormal image, so that the first image calculated using the calculation formula from the height of the circumscribed rectangle of the character string image Since the character element images extracted by the character element image extracting means are classified into narrow or wide character elements based on the second threshold value, it is possible to easily classify whether the character string image is normal or abnormal.

Also, based on the distribution of the intervals of the circumscribed rectangles of the character element images of the character string image classified as a normal image by the character string state classifying means, it is estimated whether the character string image is English or Japanese. A first character type estimating unit that extracts a character image by integrating a character element image separated into a character element image when the first character type estimating unit estimates that the character is an English sentence An English character extracting section and the first
When the character type is estimated by the character type estimating means, a Japanese character extracting section is provided which integrates or separates character element images according to the reference character width determined by the reference character width determining means and cuts out the character image. Therefore, it is possible to estimate whether the character string image is an English sentence or a Japanese sentence from the distribution of the intervals of the circumscribed rectangles of the character element image, and cut out the character image. Therefore, even if English sentences and Japanese sentences are mixed in a character string image unit, the accuracy of extracting a character image is further improved, and the efficiency of character recognition is increased.

Further, the first character type estimating means creates a histogram of the interval between circumscribed rectangles of adjacent character element images, and calculates the maximum value of the range of the first peak from the shorter side of the interval of the histogram and the maximum value. An interval number counting section for obtaining the number N3 of intervals between the belonging rectangles, and a third threshold value for calculating the maximum value of the range of the first peak obtained by the interval number counting section based on the second length Y D or less and the number of intervals N
An English-Japanese sentence estimating unit that estimates an English sentence when the ratio of the value of 3 to the number N−1 of intervals between all adjacent rectangles is equal to or greater than a predetermined constant R, and otherwise estimates the sentence as a Japanese sentence Therefore, the accuracy of determining whether the character string image is English or Japanese is improved. Therefore, the cutout of the character image becomes accurate.

Further, when the character string state classifying means classifies the character string image into an abnormal character string image, the reference character width determining means sets the first character string image based on the second length Y of the circumscribed rectangle of the character string image. A first reference character width calculating unit for calculating the length A1, wherein Y is the length of the character string image in the perpendicular direction, α is a predetermined coefficient close to 1, and the character string state classifying means When it is classified as a normal character string image and is estimated as a Japanese sentence by the English-Japanese sentence estimation unit, the first circumscribed rectangle of the character element image
A second reference character width determining unit that creates a length histogram and determines the maximum value of the peak to which the most frequent width belongs as the second reference length width A2; Since the character image is cut out in accordance with the first reference length A1 or the second reference length A2, the first reference character width is obtained even in a character string image in an abnormal state, and the accuracy is further improved in a character string image in a normal state. Is required for the second reference character width. Therefore, the cutout accuracy of the character image is further improved.

Further, the character recognizing means has an evaluation value output section for outputting an evaluation value indicating the certainty of the recognition when the character code corresponding to the character image is recognized, and the evaluation value is set to a predetermined value. A recognition error judging means for judging whether or not the recognition of the character code is erroneous, and a total number of characters N0 of the character code judged to be non-error by the recognition error judging means.
Is counted, and the number of characters N4 of the character code belonging to the character type dedicated to Japanese text is counted. When the ratio of the value N4 to the value N0 is equal to or more than a predetermined value, the character string image is estimated as a Japanese character string, and when the ratio is less than the English character, A second character type estimating means for estimating the character string, and extracting a character image which is solely regarded as an English character from a character string image determined to be a Japanese character character string by the second character type estimating means, to determine an error, and A cutout error determination unit that determines a cutout error in a character image that is a Japanese character in a character string image that is determined to be a character string, and recognition or cutout is determined to be incorrect by the recognition error determination unit or the cutout error determination unit. Character image cutting means for cutting out a corrected character image by alone or integrating the corrected character images, and a character recognition device for re-recognizing the character image cut out by the corrected character image cutting means as a corresponding character code. Since the possible and means, the output of the evaluation value representing the likelihood of the character recognition,
The recognition error can be determined, the character string image is estimated from the character type as an English character string or a Japanese character string, and if there is a character image recognized as a different character type in each character string, it is determined as a character image cutout error, The character image can be corrected and cut out and re-recognized. Therefore, the cutout accuracy and the recognition accuracy of the character image can be further improved.

The modified character image extracting means is a character string image of a Japanese character string, and is a character image other than the character image whose recognition or extraction is determined to be incorrect by the recognition error means or the extraction error determining means. A correction reference character width determination unit that creates a histogram having a length parallel to the character string image of the circumscribed rectangle of the above and determines the maximum value of the peak to which the most frequent length of the histogram belongs as the correction reference character width A3; The number of characters is determined from the area of the character image where the recognition or cutout is determined to be incorrect by the recognition error determination means or the cutout error determination means based on the correction reference character width A3 determined by the correction reference character width determination unit. And a modified cut-out unit for cutting out the character image corresponding to the character image. And it can be re-cut as the number of characters of the character image. Therefore, the accuracy of cutting out the character image is further improved.

The character re-recognizing means re-recognizes the character image cut out by the corrected character image cutting-out means as a corresponding character code. Since the feature is compared only with the standard feature of the English-only character type of the recognition dictionary, and in the case of a character string image of a Japanese character string, the feature of the character image is compared only with the standard characteristic of the Japanese-only character type of the recognition dictionary. The recognition accuracy can be further improved by using a character type dedicated to English characters for recognition of a character image of an English character string and a character type dedicated to Japanese characters for recognition of a character image of a Japanese character string.

A character string image extracting step of extracting a character string image and a position of a circumscribed rectangle from the document image, and a character element image extracting step of extracting a character element image and a position of the circumscribed rectangle from the character string image The character string with a threshold value determined based on a step, a first length parallel to the character string image of the circumscribed rectangle of the character element image, and a second length in a direction perpendicular to the character string image of the character string image A character string state classifying step of classifying whether the state of the image is normal or abnormal, and a reference length parallel to the character string image that cuts out the character image from the character string image according to the state of the character string image classified in the character string state classification step. A reference character width determining step of determining the
Integrating and separating the character element images according to the reference length determined in the reference character width determining step to cut out the character image, and comparing the features of the cut-out character image with the standard features of the recognition dictionary. The character recognition step is performed with a character recognition step of recognizing a character image as a corresponding character code, so that the same effects as those of the character recognition device can be obtained.

Further, the present invention is a recording medium on which a character recognition method is recorded, wherein a recognition dictionary and a program for executing the following specific steps are recorded, and the specific steps include a character string image from a document image and a circumscribed rectangle thereof. A character string image extracting step of extracting a character element image and a position of a circumscribed rectangle from the character string image, and a character string image of a circumscribed rectangle of the character element image. Character string state classification for classifying whether the state of the character string image is normal or abnormal based on a first parallel length and a threshold value determined based on a second length in a direction perpendicular to the character string image of the character string image And a reference character width determining step of determining a reference length parallel to the character string image for extracting a character image from the character string image based on the state of the character string image classified in the character string state classification step. A character image extracting step of integrating and separating character element images according to the reference length determined in the reference character width determining step, and extracting a character image; characteristics of the extracted character image and standard characteristics of the recognition dictionary. And a character recognition step of recognizing a character image as a corresponding character code, so that, in combination with another personal computer having no character recognition function,
The same effects as those of the above character recognition device can be obtained.

[Brief description of the drawings]

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

FIG. 2 is a diagram illustrating a character string image detected by a character detection unit according to the embodiment.

FIG. 3 is a diagram illustrating a character element image detected by a character element detection unit according to the embodiment.

FIG. 4 is a histogram of intervals between circumscribed rectangles of a character element image created by a first character type estimating unit of the embodiment.

FIG. 5 is a histogram of the width of a circumscribed rectangle of a character element image created by the Japanese sentence reference character width determination unit of the embodiment.

FIG. 6 is an explanatory diagram of a character image cut out by an English character cutout unit and a Japanese character cutout unit in the above embodiment, a character first recognized by a character recognition unit, and an evaluation value thereof.

FIG. 7 is a histogram of the width of a circumscribed rectangle of a character image with correct character cutout created by the reference character correction unit of the embodiment.

FIG. 8 is an explanatory diagram of a character image cut out by a corrected character cutout unit according to the embodiment, a character finally recognized by a character recognition unit, and its evaluation value.

FIG. 9 is a flowchart illustrating the operation of the embodiment.

FIG. 10 is a flowchart illustrating an operation of the embodiment.

FIG. 11 is a flowchart illustrating an operation of the embodiment.

FIG. 12 is a flowchart illustrating an operation of the embodiment.

FIG. 13 is an explanatory diagram of a recording medium on which the character recognition method described in the above embodiment is recorded.

FIG. 14 is a diagram for explaining mounting of the recording medium on a computer system.

FIG. 15 is a configuration diagram of a conventional character recognition device.

FIG. 16 is a diagram illustrating character segmentation in a conventional character recognition device.

[Explanation of symbols]

 Reference Signs List 101 Image input unit 102 Character string detecting unit 103 Character element detecting unit 104 Character string state estimating unit 105 First character type estimating unit 106 English character extracting unit 107 Japanese standard character width determining unit 108 Japanese character extracting unit 109 Character recognizing unit 110 1 character extraction error determination unit 111 second character type estimation unit 112 second character extraction error determination unit 113 reference character width correction unit 114 modified character extraction unit 115 recognition result output unit 1301 floppy disk 1401 floppy disk drive 1403 computer system

Claims (10)

[Claims] 1. A character string image extracting means for extracting a character string image and a position of a circumscribed rectangle from a document image, and a character element image extraction for extracting a character element image and a position of a circumscribed rectangle from the character string image Means, and a first parallel to a character string image of a circumscribed rectangle of the character element image
A length and a second perpendicular to the character string image of the character string image
A character string state classifying means for classifying whether the state of the character string image is normal or abnormal with a threshold value determined based on the length, and a character string according to the state of the character string image classified by the character string state classifying means A reference character width determining unit that determines a reference length parallel to a character string image that cuts out a character image from an image; and a character image that integrates and separates character element images according to the reference length determined by the reference character width determining unit. A character image extracting means for extracting a character image, and a character recognizing means for recognizing a character image as a corresponding character code by comparing a characteristic of the extracted character image with a standard characteristic of a recognition dictionary. apparatus. 2. The character string state classifying means includes: a total number counting unit that counts the total number N of character element images in units of character string images; and a second unit that calculates a first length of each character element image in units of character string images. A first length calculation unit, a first threshold value W1 is calculated based on a second length Y of a circumscribed rectangle of the character string image, and a circumscribed rectangle of each character element image calculated by the first length calculation unit Is compared with the first threshold value W1, the number N1 of narrow character elements whose first length is equal to or less than the first threshold value W1 is counted, and the value of N1 / N is a predetermined ratio. A blurred image determination unit that determines a character string image as an abnormal blurred image when the value is equal to or greater than R1, a second threshold value W2 is calculated based on a second length Y of a circumscribed rectangle of the character string image, and the first length is calculated. The first length of the circumscribed rectangle of each character element image calculated by the length calculation unit is compared with a second threshold value W2, and the first length is determined by a second threshold. A contact image determination unit that counts the number N2 of the bold character elements having the value W2 or more, and determines that the character string image is an image having many abnormal inter-character contacts when the value of N2 / N is a predetermined ratio R2 or more; 2. The character recognition device according to claim 1, further comprising a normal image determination unit that determines a character string image as a normal image when the blurred image determination unit or the contact image determination unit does not determine an abnormal image. 3. A method for estimating whether a character string image is English or Japanese based on a distribution of intervals between circumscribed rectangles of character element images of a character string image classified as a normal image by the character string state classifying means. A first character type estimating unit that extracts a character image by integrating a character element image separated into a character element image when the first character type estimating unit estimates that the character is an English sentence An English character extracting section, and when the first character type estimating section estimates a Japanese character, integrates or separates the character element images according to the reference character width determined by the reference character width determining section and cuts out the character image. The character recognition device according to claim 1, further comprising a Japanese character extraction unit. 4. The first character type estimating means creates a histogram of the interval between circumscribed rectangles of adjacent character element images, and calculates the maximum value of the range of the first peak from the shorter side of the interval of the histogram and the range thereof. An interval number counting unit for obtaining the number N3 of intervals between the rectangles belonging to the third threshold value for calculating the maximum value of the range of the first peak obtained by the interval number counting unit based on the second length Y D
If the ratio between the value of the number N3 of the intervals and the number N-1 of the intervals between all adjacent rectangles is equal to or greater than a predetermined constant R, it is estimated that the sentence is an English sentence. The character recognition device according to claim 3, further comprising an English-Japanese sentence estimating unit. 5. The reference character width determining means, when classified into an abnormal character string image by the character string state classification means, based on a second length Y of a circumscribed rectangle of the character string image. A first reference character width calculating unit for calculating the length A1, wherein Y is the length of the character string image in the perpendicular direction, α is a predetermined coefficient close to 1, and When the image is classified as a normal character string image and is estimated as a Japanese sentence by the English-Japanese sentence estimation unit, a histogram of a first length of a circumscribed rectangle of the character element image is created, and a peak to which the most frequent width belongs is created. Second maximum
A second reference character width determining unit that determines a reference length width A2, wherein the Japanese character cutout unit cuts out a character image according to the first reference length A1 or the second reference length A2. The character recognition device according to claim 4, wherein 6. The character recognizing means, when recognizing a character code corresponding to a character image, has an evaluation value output unit for outputting an evaluation value representing the likelihood of the recognition, wherein the evaluation value is a predetermined value. A recognition error judging unit for judging whether or not the recognition of the character code is an error by comparing the number of characters N0 of the character code judged to be no error by the recognition error judging unit, and belonging to a character type dedicated to Japanese text The number of characters N4 of the character code is counted, and when the ratio of the value N4 to the value N0 is a predetermined value or more, the character string image is estimated as a Japanese character string, and when the ratio is less than the predetermined value, the character string image is estimated as an English character string.
The character type estimating means, and a character image that is solely regarded as an English character in the character string image determined to be a Japanese character string by the second character type estimating means is determined to be an error, and is determined to be an English character string. A cut-out error determining unit that determines a cut-out error in a character image that is a Japanese character in the extracted character string image, and a character image whose recognition or cut-out is determined to be incorrect by the recognition error determining unit or the cut-out error determining unit, alone or It is characterized by comprising a corrected character image cutout unit that cuts out a corrected character image by integrating, and a character re-recognition unit that re-recognizes the character image cut out by the corrected character image cutout unit as a corresponding character code. The character recognition device according to claim 1, 2, 3, 4, or 5. 7. The modified character image cutout means is a character string image of a Japanese character string, and is a character image other than the character image whose recognition or cutout is determined to be incorrect by the recognition error means or cutout error determination means. A modified reference character width determination unit that creates a histogram having a length parallel to the character string image of the circumscribed rectangle, and determines the maximum value of the peak to which the most frequent length of the histogram belongs as the modified reference character width A3; The number of characters is determined from the area of the character image where the recognition or cutout is determined to be incorrect by the recognition error determination means or the cutout error determination means based on the correction reference character width A3 determined by the correction reference character width determination unit. 7. The character recognition device according to claim 6, further comprising a correction cutout unit that cuts out a corresponding character image. 8. The character re-recognition unit, when re-recognizing a character image cut out by the corrected character image cutting-out unit as a corresponding character code, when the character image is an English character string, The feature is compared only with the standard feature of the English-only character type of the recognition dictionary, and in the case of a character string image of a Japanese character string, the feature of the character image is compared only with the standard characteristic of the Japanese-only character type of the recognition dictionary. Item 7. The character recognition device according to Item 7. 9. A character string image extracting step of extracting a character string image and a position of a circumscribed rectangle from a document image, and extracting a character element image and a position of a circumscribed rectangle from the character string image And a first step parallel to the character string image of the circumscribed rectangle of the character element image
The length and the second in the direction perpendicular to the character string image of the character string image
A character string state classification step of classifying whether the state of the character string image is normal or abnormal with a threshold value determined based on the length, and a character string according to the state of the character string image classified in the character string state classification step A reference character width determining step of determining a reference length parallel to a character string image from which a character image is cut out of the image; and a character image integrating and separating the character element images according to the reference length determined in the reference character width determining step. And a character recognition step of collating a feature of the cut-out character image with a standard feature of a recognition dictionary and recognizing the character image as a corresponding character code. Character recognition method. 10. A recording medium on which a character recognition method is recorded, in which a recognition dictionary and a program for executing the following specific steps are recorded, and in the specific steps, a character string image and a circumscribed rectangle thereof are converted from a document image. A character string image extracting step of extracting a character element image and a position of a circumscribed rectangle from the character string image; and a character string image of a circumscribed rectangle of the character element image. Parallel first
The length and the second in the direction perpendicular to the character string image of the character string image
A character string state classification step of classifying whether the state of the character string image is normal or abnormal with a threshold value determined based on the length, and a character string according to the state of the character string image classified in the character string state classification step A reference character width determining step of determining a reference length parallel to a character string image from which a character image is cut out of the image; and a character image integrating and separating the character element images according to the reference length determined in the reference character width determining step. A character image extracting step of extracting a character image, and a character recognition step of collating a characteristic of the extracted character image with a standard characteristic of a recognition dictionary and recognizing the character image as a corresponding character code. .