JP2006092345A - Equipment, method, and program for character recognition - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide equipment, a method, and a program for character recognition which can improve a recognition rate of hand-written letters in documents in which printed characters and hand-written characters mixedly exist. <P>SOLUTION: After an image input part 11 creates an input image of a document in which printed characters and hand-written characters mixedly exist, a binarization part 12 binarizes the input image. A character cut-out part 15 performs a cut-out to the binarized data, character by character, and a characteristic amount calculation part 16 calculates two or more different characteristic amounts for each character. A recognition rate of hand-written characters can be improved thanks to a printed, hand-written characters separation part 19 separating printed characters and hand-written characters using the obtained characteristic amount. <P>COPYRIGHT: (C)2006,JPO&NCIPI

Description

  The present invention relates to a character recognition device, a character recognition method, and a character recognition program for a document, and in particular, a character recognition device, a character recognition method, and a character capable of increasing the recognition rate of handwritten characters in a document in which type characters and handwritten characters are mixed. It relates to recognition programs.

  In recent years, the distribution of electronic documents such as e-mail has increased. On the other hand, there are many documents output as paper. One of the reasons for this is that it is easy to add by handwriting on paper. For example, it is often the case that additional correction to a draft manuscript created by a personal computer (PC) or the like, or additional writing to a document distributed at a meeting or the like is performed by handwriting. Furthermore, a document with handwritten characters added is scanned with a scanner, etc., and character recognition is performed using OCR (Optical Character Reader) software, which is used to reconstruct a document that includes the recognition results of these handwritten characters. There is also a need to do.

  However, for character information written by hand, a practical recognition rate cannot be obtained unless conditions such as grid designation and numbers are strictly limited, which has hindered online / offline information conversion. Therefore, in order to improve the recognition accuracy of both type / handwritten characters, the type part and the handwritten part are separated and OCR corresponding to each is executed.

  As a conventional technique for performing character recognition by separating a type part and a handwritten part, character data is cut out from read data in a field (character string) unit and stored in a field buffer, and a character type of the character data is determined by a character type determination unit. An optical character reader that recognizes character data in a field buffer by referring to a handwritten dictionary or a print dictionary based on the determination result is known (for example, see Patent Document 1).

  There is also known an optical character reader that includes a printed character recognition unit and a handwritten character recognition unit, which individually perform OCR of character data and adopts the one with higher accuracy (probability) (for example, patents). Reference 2).

  In addition, a white frame for each dot is added around the binarized character information to form a white framed pattern, and this white framed pattern is composed of 4 pixels of 2 × 2 dots. 16 types of 2 × 2 patterns consisting of different combinations of pixels and black pixels are made to correspond to each other, and the frequency of occurrence of each 2 × 2 pattern of white framed patterns is counted, and the non-linear component and linear component of each 2 × 2 pattern There is also known a character type discriminating apparatus that discriminates the character type from the ratio of the above and always allows recognition using a dictionary suitable for the character type (see, for example, Patent Document 3).

In addition, handwritten and typed characters are distinguished in the state of the character line before character extraction, and after character type identification, image extraction is performed for each character, and the character recognition method is changed by distinguishing between vertical writing and horizontal writing. In addition, there is also known an addressed character recognition method that can ignore errors in character segmentation and can distinguish between handwritten characters and printed characters at high speed and with high accuracy (see, for example, Patent Document 4).
Japanese Unexamined Patent Publication No. 5-189604 ([0008] to [0011], FIG. 1) JP 7-37034 A ([0019] to [0038], FIGS. 1 to 3) JP-A-7-93466 ([0036] to [0053], FIG. 4) JP-A-9-212579 ([0006] to [0008], FIG. 1)

  However, according to the conventional character recognition device, in the case of Patent Document 1, the determination threshold value varies depending on the type of font or the writing of a person, and the recognition rate of handwritten characters decreases. Further, according to Patent Document 2, since two types of character recognition are performed, the processing takes time. Furthermore, according to Patent Documents 3 and 4, depending on the type of font, the straight line ratio of type / handwritten characters varies greatly, so the recognition rate of handwritten characters decreases.

  Accordingly, an object of the present invention is to provide a character recognition device, a character recognition method, and a character recognition program capable of increasing the recognition rate of handwritten characters in a document in which printed characters and handwritten characters are mixed.

  In order to achieve the above object, the present invention provides a character cutout unit that cuts out characters on a character-by-character basis for an input image of a document in which printed characters and handwritten characters are mixed, and a feature amount in units of one character for the cut-out characters A feature amount calculation unit that calculates the feature amount, a feature amount aggregation unit that aggregates the calculated feature amount of each character unit over a predetermined range of the input image, and a type based on the aggregation result by the feature amount aggregation unit A separation coefficient calculation unit for calculating a separation coefficient for separating a handwritten character and a handwritten character that separates a character related to the input image into the type and the handwritten character based on the calculated separation coefficient There is provided a character recognition device including a separation unit.

  By summing up the feature amount of one character unit over a predetermined range, it is possible to obtain a separation coefficient for separating a character related to the input image into a printed character and a handwritten character.

  The feature amount calculation unit may be configured to calculate feature amounts for at least uniformity of character density and linearity of character line segments. Since the character density uniformity and the linearity feature amount of the character line segment are large for printed characters and handwritten characters, a high-accuracy separation coefficient can be obtained by using these feature amounts.

  Further, the feature amount calculation unit calculates the character size, height, width uniformity, character density uniformity, density histogram, character line segment linearity, character line width uniformity, and character inclination. The feature amount may be calculated for at least one of the plurality. By using many feature amounts, a highly accurate separation coefficient can be obtained.

  The feature amount totaling unit may total feature amounts in units of one character for the entire input image. By increasing the number of samples, a highly accurate separation factor can be obtained.

  The type handwritten character separation unit preferably separates the character relating to the input image into a type and a handwritten character, and then re-determines the type / handwritten character using a separation result around each separated character. In this case, the re-determination may be performed for each chapter or paragraph of the sentence.

  The type handwritten character separating unit may separate the typed characters and the handwritten characters from the input image or an image obtained by binarizing the input image.

  The feature amount calculation unit calculates a plurality of types of feature amounts in units of one character, and the separation coefficient calculation unit performs a plurality of intermediate evaluations on the plurality of types of feature amounts as a separation coefficient, each of which seems to be a type or a handwritten character. In addition, the printed handwritten character separating unit may separate the printed characters from the handwritten characters based on a plurality of intermediate evaluations. As a result, even if it is not possible to clearly separate the printed characters or the handwritten characters according to the threshold, the separation is possible. For example, when a plurality of intermediate evaluations that are likely to be handwritten characters appear from the result of counting a plurality of types of feature amounts, it can be determined as a handwritten character.

  In order to achieve the above object, the present invention cuts out characters in units of one character from an input image of a document in which type letters and handwritten characters are mixed, calculates a feature amount in units of one character for the cut out characters, Separation coefficients for separating the calculated feature values in units of one character over a predetermined range of the input image and separating the printed characters from the handwritten characters based on the total results of the feature amounts in units of one character. A character recognition method is provided that calculates and separates a character related to the input image into the printed character and the handwritten character based on the calculated separation coefficient.

  In order to achieve the above object, the present invention provides a character cutout means for cutting out characters in units of characters from an input image of a document in which printed characters and handwritten characters are mixed, and a feature amount in units of characters for the cut out characters. Based on the result of counting by the feature amount calculating means, the feature amount calculating means for calculating the calculated feature amount for each character over a predetermined range of the input image, A separation coefficient calculation means for calculating a separation coefficient for separating a handwritten character from a handwritten character, and a typed handwritten character that separates a character related to the input image into the type letter and the handwritten character based on the calculated separation coefficient There is provided a character recognition program for causing a computer to execute the separating means.

  ADVANTAGE OF THE INVENTION According to this invention, the recognition rate of the handwritten character in the document in which the type | mold and the handwritten character were mixed can be improved.

(Configuration of character recognition device)
FIG. 1 shows a character recognition apparatus according to an embodiment of the present invention. An image input unit 11 that reads a document such as a document, a binarization unit 12 that binarizes image data from the image input unit 11, a labeling unit 13 that collects binarized outputs in units of continuous pixel groups, and a labeling A label deletion integration unit 14 that identifies one character with respect to the result of the unit 13, a character extraction unit 15 that extracts characters in units of one character from the result of the label deletion integration unit 14, and a feature amount of the extracted character of 1 A feature amount calculating unit 16 that calculates character units, a feature amount totaling unit 17 that totals feature amounts over the entire image, a separation coefficient calculating unit 18 that calculates a separation factor of type / handwritten characters from the feature amount totaling results, and a separation Based on the calculation result of the coefficient calculation unit 18, the original handwritten character separation unit 19 that separates the original image input from the image input unit 11 or the binarized image into a type part and a handwriting part, and a separated type letter Type recognition unit 20 that performs OCR processing on the part, handwriting recognition unit 21 that performs OCR processing on the separated handwriting unit, type dictionary 22 used for recognition processing of the type recognition unit 20, and handwriting It has the handwriting dictionary 23 used for the recognition process of the recognition part 21, and the synthetic | combination part 24 which synthesize | combines the recognition result of the type recognition part 20, and the recognition result of the handwriting recognition part 21 to one file or one image.

  The image input unit 11 includes an original platen made of transparent platen glass on which an original is placed, and a solid-state image sensor such as a CCD (Charge Coupled Device) that optically reads an image from the original placed on the original platen. Prepare.

  The binarization unit 12 has a function of binarizing the image data output from the image input unit 11 using, for example, a threshold set for the density.

  The labeling unit 13 has a function of performing labeling that numbers the binarized outputs in units of continuous pixel groups in order to facilitate the extraction of character units.

  In addition, the label deletion integration unit 14 has a function of determining which one forms a set of characters in consideration of unevenness, creation, and the like, and can accurately cut out characters. Therefore, it also has a function of performing a process of removing blocks that are separated from characters as dust.

The feature amount calculation unit 16 calculates feature amounts for the following elements.
(1) The size according to the width and height of the extracted character. In the case of type characters, the height and width are generally constant unless the number of points is changed, whereas handwritten characters are scattered, and thus become one of the feature values.
(2) Pixel value variation (histogram). This is one of the feature quantities because it is generally constant for printed characters, but varies for handwritten characters.
(3) Linearity of line segment. A handwritten character is one of the feature quantities because a horizontal or vertical line often does not become a straight line due to writing.
(4) Line width. A type character is one of features because the line width is almost constant if the font is the same, whereas handwritten characters vary depending on the writing pressure or the like.
(5) Character inclination. The typeface has an inclination in italic characters and the like, but the inclination is constant. On the other hand, the handwritten character tends to be inclined due to the wrinkle of the person who wrote it, and the inclination tends to vary. Become one.

  The feature amount totaling unit 17 has a function of totaling the feature amounts acquired by the feature amount calculating unit 16 over the entire image. In addition, a configuration in which layout analysis is performed in advance and tabulated in a predetermined unit such as a sentence unit may be used.

  The separation coefficient calculation unit 18 has a function of calculating a separation coefficient for determining whether the feature amount aggregated by the feature amount aggregation unit 17 is a type or a handwritten character.

(Principle of separation factor calculation)
FIG. 2 shows the principle of the separation coefficient calculation in the separation coefficient calculation unit 18. Here, the line width among the feature amounts will be described. In general, the size of characters used in documents and books is Mincho in the range of 10.5 to 12 points, and the line width is relatively thin. On the other hand, handwritten characters tend to be large, and since the thickness of normally used ballpoint pens and mechanical pencils is 0.5 mm or more, the line widths of printed characters and handwritten characters often clearly differ. Therefore, if both are totaled for the entire image, the characteristic has two peaks as shown in FIG. However, when the line widths of the printed characters and the handwritten characters are approximated, as shown in FIG. 2B, there is a characteristic that there is one gentle mountain and the printed characters and the handwritten characters cannot be distinguished. Therefore, for the evaluation of the feature quantity, a feature quantity having characteristics as shown in FIG. 2A is adopted, and a feature quantity having characteristics as shown in FIG. 2B is not adopted.

  Further, in FIG. 2A, the width s of the valley portion (gray zone) between the peaks is digitized, and the boundary between the printed character side and the handwritten character side is fuzzy to determine whether it is a printed character or a handwritten character. It is possible to determine whether the characters in the gray zone that cannot be determined by the quantity are printed or handwritten. Thus, since the gray zone is quantified regardless of the threshold value, it is referred to as a separation coefficient here. Note that other feature quantities other than the line width can be obtained in the same manner as the line width. For example, although the total result of each feature amount exists in the gray zone, if a result that any feature amount is good is obtained, it can be determined as a type.

  The printed handwritten character separation unit 19 uses the feature amount calculation unit 16 to generate an image from the image input unit 11, the label deletion integration unit 14, or the binarization unit 12 or the labeling unit 13 (not shown). Based on the calculated feature amount and the separation coefficient calculated by the separation coefficient calculation unit 18, the type part and the handwriting part are separated.

  The type recognition unit 20 compares the extracted type character pattern and the type pattern registered in the type dictionary 22 by, for example, a pattern matching method, and determines the type having the highest similarity as a result of the type recognition (character code). ).

  The handwriting recognition unit 21 compares the extracted handwritten character pattern and the handwritten character pattern registered in the handwriting dictionary 23 by, for example, a pattern matching method, and recognizes the type with the highest similarity as a result of recognition of the type ( Character code).

  As the type dictionary 22 and the handwriting dictionary 23, for example, an electronic dictionary stored in a storage medium such as a hard disk can be used.

  The synthesizing unit 24 synthesizes the recognition result of the type recognition unit 20 and the recognition result of the handwriting recognition unit 21 into, for example, one file or one image.

(Operation of character recognition device)
FIG. 3 schematically shows processing in each part of FIG. 1, and FIG. 4 shows a partially enlarged view of a document, and a printed image and a handwritten image obtained by separation. The operation of the character recognition device 1 will be described below with reference to FIGS.

  The image input unit 11 reads a manuscript in which handwritten characters are written in a printed document. Here, color conversion processing is performed depending on whether the content of the document is a color image or a gray image. In the case of a color image, processing for converting R, G, and B into gray is performed so that time is not required for processing in a later process. When converting R, G, and B into gray, one of the following is employed.

(A) The gray value is set to the G value of the color image.
(B) Gray value = 0.30r + 0.59g + 0.11b. Here, r, g, and b are the R, G, and B values of the color image, respectively.
(C) Gray value = α × r + β × g + γ × b. Here, α, β, and γ are coefficients, and α, β, and γ are changed according to the determination color associated with the color determination.

Next, binarization is performed by the binarization unit 12. For example, the following method is used for the binarization process.
(I) Fixed threshold binarization process: binarization with a predetermined threshold.
(Ii) Dynamic threshold binarization processing: binarization is performed using an average value of surrounding pixels.

  Next, a labeling process is performed by the labeling unit 13. Here, the image binarized by the binarization unit 12 is labeled for each continuous pixel group having pixel values corresponding to the foreground such as characters.

  Next, the label deletion / integration unit 14 deletes and integrates the labels. Here, removal of a label corresponding to noise, integration of labels of the same character divided into a plurality of labels, and the like are performed.

  Next, the character cutout unit 15 cuts out characters in units of one character, as shown in FIG. 3A, excluding regions other than characters, for example, figures / photos. Here, the printed characters 31 and the handwritten characters 32 in the document image 30 are labeled with a label 33.

  Next, the feature amount calculation unit 16 calculates the feature amount in character units for character size (width, height), pixel value variation (histogram), line segment linearity, line width, and character inclination. Calculated.

  Next, the feature amount totaling unit 17 totals the feature amount in units of one character for the entire image for each chapter or paragraph of the sentence determined from the layout analysis result or for the entire image.

  Next, the separation coefficient calculation unit 18 grasps the characteristics as shown in FIG. 2A based on the feature amounts totaled by the feature amount totalization unit 17 and calculates a type / handwritten character separation coefficient.

  The printed handwritten character separation unit 19 determines whether each character is a printed character or a handwritten character based on the printed character / handwritten character separation coefficient, and performs separation as shown in FIG. FIG. 3 shows that a hatched character is recognized as a handwritten character. Further, as shown in FIG. 3C, the original image (label image) is developed. For this result, based on the result determined for each character, re-determination is performed based on the determination result for each sentence, for each paragraph, or around the character, and correction as shown in FIG. carry out.

  For example, the “,” part of the second line in FIG. 3C is determined to be a handwritten character, but before and after that, it is a determination of a type character, and there is only one character in the type string. Is unnatural. Therefore, “,” is re-determined as a type and is corrected as shown in FIG. Similarly, “no” on the third line in FIG. 3C is determined to be a type, but it is determined to be a handwritten character before and after that, and it is unnatural to enter only one type. Therefore, “no” is re-determined as a handwritten character, not a type, and is corrected as shown in FIG. In this correction, layout analysis may be performed and correction may be performed in units of text areas. In this case, if a large number of characters in the text area are type determination, the character for handwriting determination is a type, and similarly, if a large number of characters in the text area are handwriting determination, the type determination character is a handwritten character. To do.

  Next, the type recognition unit 20 and the handwriting recognition unit 21 use the type dictionary 22 and the handwriting dictionary 23 to perform OCR processing on the typed characters and the handwritten characters on the image that has been corrected as shown in FIG. Each recognition result or recognition image is generated.

  For example, if a manuscript 40 including type 41 and handwritten characters 42 as shown in FIG. 4A is input from the image input unit 11, only the type 52 shown in FIG. 4 is obtained, and the handwriting recognition unit 21 obtains a handwritten image 53 including only the handwriting unit 54 as shown in FIG. In FIG. 4C, the recognition / output result of the handwriting recognition unit 21 is output in a font type similar to a handwritten character like the handwriting unit 54, but is output using a general type font. May be. The printed image 51 and the handwritten image 53 are combined by the combining unit 24, and a recognition image having the same layout as that of the original 40 in FIG. 4A is generated. If this recognition image is output by a printer or the like, a document in which handwritten characters are converted into letters can be obtained. The recognition results in the type recognition unit 20 and the handwriting recognition unit 21 may be output as character information (code information), and the combining unit 24 may combine them.

(Effect of embodiment)
According to this embodiment, the following effects can be obtained.
(A) Since the feature amount is calculated for each character and the separation determination of type / handwritten characters is performed based on the feature amount, the character recognition rate can be increased.
(B) Since a plurality of feature quantities having different contents are used, the feature quantities can be calculated with certainty, and the type / handwritten character separation determination can be reliably performed.
(C) Since the feature values are aggregated over the entire image or in units of a predetermined range, and the separation coefficient is calculated based on this, the character / handwritten character separation can be performed in units of characters, so that the character recognition rate of handwritten characters is increased. Can do.
(D) In the handwritten character separation unit 19, the character segmentation result obtained by the character segmentation unit 15 is used as a result of determination of the surroundings of characters such as up and down, left and right, or for each sentence or paragraph included in the character. / By re-determining handwritten characters and performing correction based on the results, it is possible to separate the printed characters from the handwritten characters with high accuracy, thereby increasing the OCR recognition rate.

[Other embodiments]
In addition, this invention is not limited to the said embodiment, A various deformation | transformation is possible within the range which does not change the summary. For example, the image input to the printed handwritten character separation unit 19 may include an output image of the binarization unit 12. This makes it possible to separate type / handwritten characters from black and white images.

It is a block diagram which shows the character recognition apparatus which concerns on embodiment of this invention. 1 shows the principle of calculation of a separation coefficient in the separation coefficient calculation unit of FIG. 1, (a) is a line width-frequency characteristic diagram for calculating a separation coefficient when the line widths of type characters and handwritten characters are clearly different, and (b) It is a line width-frequency characteristic figure when the line width of a type letter and a handwritten character is unclear. 1 schematically shows the processing in each part of FIG. 1, (a) is an enlarged view showing an example of a document image, (b) is a document image diagram after separation of type / handwritten characters, and (c) is an image of (b). The image figure which shows the state expand | deployed to the original image, (d) is an image figure after correct | amending the image of (c). FIG. 4 shows a processing process in a type letter hand separating unit, (a) is a plan view of a document read by an image input unit, (b) is an image diagram of a type image obtained by being separated by a type hand letter separating unit, (C) is an image diagram of a handwritten image obtained by separation by a printed handwritten character separation unit.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 Character recognition apparatus 11 Image input part 12 Binarization part 13 Labeling part 14 Label deletion integration part 15 Character extraction part 16 Feature-value calculation part 17 Feature-value totaling part 18 Separation coefficient calculation part 19 Type-letter hand-character separation part 20 Type recognition Unit 21 Handwriting recognition unit 22 Type dictionary 23 Handwritten dictionary 24 Composition unit 30 Original image 31 Type 32 Handwritten character 33 Label 40 Original 41 Type 42 Handwritten character 51 Type image 52 Type unit 53 Handwritten image 54 Handwritten unit

Claims (10)

A character cutout unit that cuts out characters on a character-by-character basis for an input image of a document in which printed characters and handwritten characters are mixed;
A feature amount calculation unit that calculates a feature amount in units of one character for the extracted character;
A feature amount totaling unit that totalizes the calculated feature amount in units of one character over a predetermined range of the input image;
A separation factor calculation unit that calculates a separation factor for separating a type letter and a handwritten character based on a counting result by the feature amount counting unit;
A character recognition apparatus, comprising: a printed handwritten character separating unit that separates a character related to the input image into the printed character and the handwritten character based on the calculated separation coefficient.   The character recognition device according to claim 1, wherein the feature amount calculation unit calculates the feature amount at least with respect to uniformity of character density and linearity of character line segments.   The feature amount calculation unit includes a character size, height, width uniformity, character density uniformity, density histogram, character line segment linearity, character line width uniformity, and character slope. The character recognition device according to claim 1, wherein the feature amount is calculated for at least one of a plurality.   The character recognition device according to claim 1, wherein the feature amount totaling unit totalizes the feature amounts in units of one character for the entire input image.   The type handwritten character separation unit separates the character related to the input image into the type and the handwritten character, and then performs re-determination of the type / handwritten character using a separation result around each separated character. The character recognition device according to claim 1.   The character recognition apparatus according to claim 5, wherein the character handwriting character separation unit performs the re-determination for each chapter or paragraph of the sentence.   The character recognition apparatus according to claim 1, wherein the printed handwritten character separating unit separates printed characters and handwritten characters from the input image or an image obtained by binarizing the input image. The feature amount calculation unit calculates the plurality of types of feature amounts in units of one character,
The separation factor calculation unit includes a plurality of intermediate evaluations that are likely to be printed or handwritten characters for the plurality of types of feature quantities as the separation factor,
The character recognition apparatus according to claim 1, wherein the printed handwritten character separating unit separates printed characters and handwritten characters based on the plurality of intermediate evaluations. For the input image of a document containing both type and handwritten characters, cut out the characters in units of one character,
Calculate the feature value of each character for the extracted characters,
The calculated feature amount of each character unit is totaled over a predetermined range of the input image,
Calculating a separation coefficient for separating a type letter and a handwritten character based on the total result of the feature amount of one character unit;
A character recognition method, wherein a character related to the input image is separated into the printed character and the handwritten character based on the calculated separation coefficient. A character cutout unit that cuts out characters in units of characters for an input image of a document in which printed characters and handwritten characters are mixed;
A feature amount calculating means for calculating a feature amount in character units for the cut out character;
A feature amount totalizing unit that totalizes the calculated feature amount of each character unit over a predetermined range of the input image;
A separation coefficient calculating means for calculating a separation coefficient for separating a type letter and a handwritten character based on a counting result by the feature amount counting means;
A character recognition program for causing a computer to execute type handwritten character separating means for separating a character relating to the input image into the type and the handwritten character based on the calculated separation coefficient.