JP4229521B2 - Character recognition method and apparatus - Google Patents

Description

[0001]
BACKGROUND OF THE INVENTION
In recent years, demand for character recognition technology as an input peripheral device has increased. The present invention relates to a character recognition method and apparatus used for character recognition, and in particular, the present invention relates to character recognition capable of obtaining a recognition result at a high speed with higher recognition accuracy in a document in which handwritten characters and printed characters are mixed. It relates to a method and a device.
[0002]
[Prior art]
In character recognition of a document in which handwritten characters and type letters are mixed, usually, two processes of handwritten character recognition and type recognition are performed, and the recognition results of both are compared to adopt a result with high reliability. This method requires processing time because it requires recognition of handwritten characters and type recognition for recognition.
In view of this, a method has been proposed in which it is determined in advance whether a character (sequence) to be recognized is a handwritten character or a type, and handwritten character recognition or type recognition is performed according to the result. According to this method, since it is not necessary to perform the character recognition process twice, the processing time can be shortened.
[0003]
As a handwritten character / type discrimination method, for example, a document image is obtained from a document or form by a scanner or the like, a character string is extracted from the document image, characters included in the character string are further extracted, pitches for all characters, Obtain the regularity of the character arrangement such as the shift of the center position, and when the character arrangement of the character string is regular, it is regarded as a printed character string, and when the arrangement is not regular, the character string is determined as a handwritten character string The method is used.
[0004]
[Problems to be solved by the invention]
As a method of expressing the regularity of the arrangement of characters in the character string by the regularity of the center position in the height direction of the characters of all characters in the character string, for example, as described above, the characters are cut out from the character string, The center position of the character in the height direction is obtained. If the variation in the center position is small, the type is used, and if the variation is large, the discrimination method is a handwritten character.
This method is characterized by the fact that in many cases in the case of kanji or numbers, the size of the characters in the type is constant and regularly arranged, so there is little variation in the center position, and it can be used for discrimination from handwritten characters. is there.
[0005]
However, as shown in FIG. 14, in the case of a character string in which the character seen in katakana includes a cloud point, the center position of the character is shifted upward due to the cloud point (for example, “G” in the drawing), The position variation becomes large.
If the character string includes lowercase letters, the center position of the character shifts in the downward direction, so that the variation in the center position becomes large even in the typed characters (“ョ” in the figure).
For the above reasons, the conventional method could not be applied to the case where the character type having dakuten and lowercase letters such as katakana and hiragana was included.
[0006]
In the handwriting / type discrimination method described above, the layout information is not used for handwriting / type discrimination processing even if the target is a form whose layout information is known.
That is, in a form in which layout information is registered in advance, for example, it is clear that an item such as an amount field is written with a number in the character entry area that follows this, but in the conventional method, characters that can be distinguished from the layout information The handwriting / type discrimination is performed without using the information on the character type written in the entry column, and the discrimination accuracy of the handwritten / type is poor.
In addition, in the form, the text entry field following the item such as the amount field is mostly unified with handwritten characters or printed characters in one form, but this information is not used and each character is entered. The handwriting / type discrimination results for each entry field were used as they were. For this reason, it has been difficult to improve the discrimination accuracy of handwritten characters / types.
The present invention has been made in consideration of the above circumstances, and is to improve the accuracy of handwriting / type discrimination in character recognition processing, shorten the processing time by character recognition, and improve the character recognition accuracy.
[0007]
[Means for Solving the Problems]
FIG. 1 is a diagram illustrating the principle of the present invention. As shown in FIG. 1, the present invention solves the above problem as follows.
As shown in FIG. 1 , a character string is extracted, handwriting / type discrimination processing is performed, and handwriting recognition processing or type recognition processing is performed based on the result.
In the handwriting / type discrimination process, a handwritten character and type are discriminated as follows. (a) Characters are extracted from the character string without including a cloud point, the center positions of all the extracted characters are obtained, and handwritten characters and printed characters are discriminated from their regularity.
Thereby, it is possible to solve the problem that the variation of the center position becomes large in the type including the muddy point.
(b) The center positions of all characters extracted without including lowercase letters included in the character string are obtained, and handwritten characters and type characters are discriminated based on their regularity.
Thereby, it is possible to solve the problem that the variation in the center position becomes large in a type including lower case letters.
As a method for obtaining the regularity of the center position of the character without including the lower case letters, for example, one character as a reference is extracted from the character string, the size is small with respect to the character, and the reference is used. It is possible to use a method of obtaining the regularity of the center position of the character by removing the character located below the character.
In addition, when the regularity of the center position in the height direction of the character in the character string is used for handwriting / printing discrimination, if the number of characters included in the character string is small, the handwriting / printing discrimination may be suspended. Good. As a result, even when the number of characters is small, there is no mistake in handwriting / printing discrimination. When the handwriting / type discrimination is suspended, both the handwritten character recognition process and the type recognition process are performed in the character recognition process, and a highly reliable result is adopted.
[0008]
DETAILED DESCRIPTION OF THE INVENTION
FIG. 2 is a diagram showing a configuration example of a system to which the present invention is applied. In the figure, reference numeral 11 denotes a CPU, and a bus 10 is connected to the CPU 11. The bus 10 includes a RAM 12, a ROM 13, a communication interface 14, an external storage device 15, a storage medium reading device 16, and a scanner for reading an image. 17, an input / output device 18 such as a display and a keyboard is connected. The ROM 13 and RAM 12 store an OS, a program for performing the processing of this embodiment, data, and the like.
The scanner 17 reads an image in which characters and the like that are targets for character recognition are recorded. The input / output device 18 includes input devices such as a keyboard and a mouse, and output devices such as a display monitor and a printer, and inputs data for performing data processing according to the present invention and outputs processed data.
[0009]
The program, data, and the like for performing the character recognition processing of the present invention are recorded on a storage device 15 such as a hard disk or a recording medium such as a CD-ROM or a floppy disk. Then, it is read from the recording medium by the recording medium reading device 17, stored in the RAM 12, and executed by the CPU 11. In addition, a communication interface 14 is connected to the bus 10, and data and programs can be exchanged via a communication line.
[0010]
FIG. 3 is a diagram showing a processing flow of character recognition processing according to the embodiment of the present invention.
As shown in the figure, an image of a document or a form is acquired by an image input means such as a scanner 17, and preprocessing such as noise removal and inclination correction is performed on the acquired image (step S1). Next, one character string is extracted from the image (step S2), and handwriting / printing according to the present invention is identified (step S3). When the extracted character string is determined as a handwritten character, the flag is set to 1, and when it is determined as a printed character, the flag is set to 2. Further, when it is impossible to discriminate between handwritten characters and printed characters, the flag is set to 3 (step S4).
[0011]
In the character recognition process, the flag is referred to and character recognition is performed as follows (steps S5 to S6).
(1) When the handwritten character is determined, a handwritten character recognition process is performed.
(2) When it is determined as a type, type recognition processing is performed.
(3) When it is impossible to discriminate between handwritten characters and characters, a handwritten character recognition process and a character recognition process are performed, and the one with higher reliability is adopted.
As described above, it is possible to improve the reliability of the character recognition process by determining whether the character is a handwritten character or a character in advance and performing the character recognition process. Since character recognition can be performed without performing processing, the processing speed can be increased.
[0012]
Next, an embodiment of the present invention will be described with respect to the handwriting / type discrimination process in the character recognition process.
FIG. 4 is a diagram showing a processing flow of handwritten type discrimination according to the embodiment of the present invention.
As shown in FIG. 4, first, a connected component is extracted from the characters of the extracted character string (step S1). For example, if the extracted character string is “modification” shown in FIG. 14, the first character “MO” is connected to all the pixels constituting the character, so that as shown in FIG. Extracted as a connected component. For example, for the next character “di”, five connected components are extracted.
Next, the muddy point is removed from the extracted connected components (step S2), and the connected components are integrated (step S3). That is, the connected components remaining in the muddy point removal process are integrated under appropriate conditions and separated into units constituting characters. As the integration condition, for example, circumscribed rectangles of connected components are compared and integrated when there is an overlap.
In the case of the example in FIG. 14, as shown in FIG. 4, the “d” cloud point is removed and the remaining connected components are integrated.
[0013]
In step S4, lowercase letters are removed. In the case of the example in FIG. 14, the lowercase “yo” is removed.
In step S5, the remaining number of characters is compared with a predetermined threshold value. If the remaining number of characters is smaller than the threshold value, the process goes to step S6 and the flag is set to 3 because the handwriting / printing type cannot be determined. End.
[0014]
That is, when the number of extracted characters is small, there is a possibility of erroneous determination because the reliability of the variation value of the center position of the characters described later is poor. Therefore, when the number of extracted characters is small, the determination is suspended, and both handwriting / printing recognition processing is performed as described in (3) above, and the one with higher reliability is adopted.
If the number of remaining characters is larger than the threshold, the process goes to step S7 to calculate the regularity of the character center position. That is, as shown in FIG. 5, each character is separated and extracted, and the center position in the height direction of each character is obtained. Then, the variation in the center position of each character is digitized by dispersion or the like.
Next, in step S8, it is determined whether the character is a handwritten character or a printed character according to the variation in the center position. That is, if the variation in the center position is large, the flag is set to 1 because it is a handwritten character string, and if the variation is small, the flag is set to 2 because it is a character string.
[0015]
FIG. 6 is a processing flow showing a method for removing the muddy point. As a muddy point removal method, as shown in the process flow of FIG. 6, first, in step S1, a small connected component that is above a predetermined value and is extracted is extracted. For example, as shown in FIG. 7, in the case of “di”, connected components c, d, e larger than a predetermined value are extracted.
Next, in step S2, the distance between the extracted connected components is calculated. In the case of the example in FIG. 7, the mutual distance between the connected components c, d, e is calculated.
In step S3, it is checked whether there is an object whose distance between connected components is smaller than a threshold value. Then, if there is no distance between connected components that is smaller than the threshold value, the process is terminated because there is no cloud point. If the distance between the connected components is smaller than the threshold value, the process goes to step S4 to remove the connected component pair whose distance is smaller than the threshold value as a muddy point. In the case of FIG. 7, since the distance between the connected components d and e is smaller than the threshold value, it is removed as a muddy point. As a result, the muddy point candidates 1 and 2 are removed as shown in FIG.
[0016]
After removing the cloud point as described above, the connected components are integrated as described above and separated into units constituting characters.
FIG. 9 shows a process flow for removing lowercase letters from characters separated and extracted by this process.
First, in step S1, a reference character is extracted, and in step S2, the height and lower coordinates of the reference character are extracted.
[0017]
For example, the first character (the leftmost character in the horizontally written character string in FIG. 10) is used as a reference character, and the height and the lower coordinate in the height direction are extracted. In addition, the character with the maximum height in the character string may be used as the reference character.
In step S3, the reference character is compared with the lower coordinates of each character, and the difference (or ratio) between the height of the character whose difference in the lower coordinates is equal to or less than a predetermined value and the height of the reference character is calculated.
For each remaining character, a character whose height is smaller than the height of the reference character and whose lower coordinate is close to the lower coordinate of the reference character is removed as a lower case letter (step S4). For example, in the example of FIG.
Hereinafter, as described above, the regularity of the center position of the character with respect to the remaining character after removing the lowercase letters is obtained by quantifying the character by, for example, the distribution of the coordinates of the center position in the height direction of the character. Depending on the size, it is determined whether it is handwritten or typed.
[0018]
As described above, in this embodiment, handwriting / type discrimination is performed based on the regularity of the center position of the character in the character string from which the dakuten and lowercase characters are removed. Can be performed with high accuracy even if it is included in the target.
In addition, although the discrimination processing based on the regularity of the center position of the character shown in the present embodiment may be used alone, handwriting / printing discrimination may be performed. It is also possible to extract a plurality and collectively determine these features. In this case, as a discrimination method, not only discrimination based on the magnitude of the feature amount but also a method such as discriminant analysis can be used.
[0019]
Next, an example in which layout information is used for handwriting / type discrimination processing will be described.
Shows a process flow in this case is shown in FIG. 11 also shows the processing conceptual diagram of this case is shown in FIG. 12. First, in step S1 of FIG. 11, the layout of the form is analyzed with respect to the form converted into image data, and in step S2, the item and the subsequent character entry area are divided into groups.
For grouping, a method such as creating a group for each column of a form can be used. For example, in the case of FIG. 12, the items are grouped into item characters such as “bank name”, “branch name”, “account number”, “money amount”, and subsequent character entry fields.
In step S3, character recognition is performed on the item area of the form to recognize the item character.
[0020]
Next, in step S4, the character type of the character entry area is determined from the information obtained as a result of character recognition. For example, when “amount” is written in the item area, the character type of the character entry column belonging to the same group as this item is estimated as a number. As a method of estimating the character type, for example, as shown in FIG. 11, the character type of the character entry area can be determined by referring to the correspondence data of the character type in the item and the subsequent character entry field.
In step S5, the feature amount is extracted only for the character type obtained from the determination result, and the handwritten character / type determination process is performed. As the feature amount, variations in the center position of characters included in the character string shown in the first embodiment, variations in character height, and the like can be used.
Further, as a determination method, a threshold value in which a feature amount is set in advance is compared, and if the amount of variation is equal to or less than the threshold value, it can be determined as a printed character, and if it is equal to or greater than the threshold value, it can be determined as handwritten characters. It is also possible to use a plurality of feature amounts to make a comprehensive discrimination on the feature space by discriminant analysis or the like.
[0021]
As described above, the handwriting / type discrimination processing is performed by limiting the character type in the character entry field based on the character recognition result of the item area using the layout information of the form. Can be improved.
In the above, for unknown forms, the character type of the character entry area belonging to the same group as this item is limited based on the result of recognizing the characters in the item area, and handwriting / type discrimination is performed. However, if the layout information of the form and the character type information of the character entry area are registered in advance, the layout information of the form to be processed is extracted first, the form is collated using the layout information, and registered. In the case of a match with a registered form, it is possible to perform handwriting / type discrimination by limiting the character type for each character entry field using the character type information for the character entry field of the registered form.
[0022]
Next , another example in which layout information is used for handwriting / type discrimination processing will be described. The processing flow is shown in FIG.
In step S1 of FIG. 13, the form structure is analyzed for the form converted into image data, and layout information is acquired.
Next, in step S2, the character entry areas of the form are grouped. For grouping, a method such as creating a group for each column of a form can be used. For example, in the case of the form shown in FIG. 12, the text entry field (group 1) of the item “bank name”, the text entry field (group 2) of the item “branch name”,..., The item “amount” (group) Group them into 4) character entry fields.
[0023]
In step S3, handwriting / type discrimination is performed for each character string in each character entry field. That is, in the case of FIG. 12, handwriting / type discrimination in the character entry column is performed for groups 1 to 4.
In step S4, the handwritten / type discrimination results are totaled for each group, and in step S5, the discrimination result of the larger number is the correct answer and the discrimination result of the lesser number is the larger one. Correct the discrimination result.
For example, in the example of FIG. 12, when there are many results for group 1 that are determined as handwritten characters and there are few results that are determined as printed characters, it is determined that the group is a handwritten character, and the results determined as printed characters are Correct the discrimination result. On the contrary, when there are many results determined as a type and few results as a handwritten character, the result determined as a handwritten character is corrected to a result of type determination. Similarly, the discrimination result is corrected for each of the groups 1 to 4.
As described above, using the layout information of the form, the handwriting / type discrimination results of each group are totaled, and the handwriting / type discrimination processing results of the character entry fields belonging to the same group are corrected based on the tabulation results. As in the second embodiment, it is possible to improve the accuracy of handwritten character / type discrimination processing.
[0024]
【The invention's effect】
As described above, Oite the present invention can obtain the following effects.
(1) Since the regularity of the center position of the character obtained without including the cloud point is extracted from the character string to distinguish between handwritten characters and printed characters, any characters such as katakana and hiragana that may contain a cloud point It is possible to distinguish between handwritten / printed columns.
(2) Since the regularity of the center position of the character is not included in the character string and the regularity of the center position of the character is extracted to distinguish between handwritten characters and printed characters, any characters such as katakana and hiragana that may contain lowercase characters It is possible to distinguish between handwritten / printed columns.
In addition, when the number of characters in the target character string is small, the determination result of the determination hold (reject) is output, and the recognition is left to both handwriting and type recognition processing, so that the character recognition accuracy is not impaired.
[Brief description of the drawings]
FIG. 1 is a principle diagram of the present invention.
FIG. 2 is a diagram illustrating a configuration example of a system to which the present invention is applied.
FIG. 3 is a diagram illustrating a processing flow of character recognition processing according to an embodiment of the present invention.
FIG. 4 is a diagram showing a processing flow of handwritten type discrimination according to an embodiment of the present invention.
FIG. 5 is a diagram for explaining a concept of handwriting / type discrimination processing according to the embodiment.
FIG. 6 is a processing flow showing a method for removing muddy spots.
FIG. 7 is a diagram illustrating a method for removing muddy points.
FIG. 8 is a diagram for explaining the concept of muddy point removal processing.
FIG. 9 is a diagram showing a processing flow for removing lowercase letters.
FIG. 10 is a diagram for explaining the concept of lowercase removal processing.
FIG. 11 is a processing flow of an example in which layout information is used for handwriting / type discrimination processing .
12 is a diagram for explaining the processing concept of the example of FIG . 11 ;
FIG. 13 is another example of a processing flow in which layout information is used for handwriting / type discrimination processing .
FIG. 14 is an example of katakana type characters including dakuten and lower case, which are problematic in the conventional method.
[Explanation of symbols]
11 CPU
10 buses 12 RAM
13 ROM13
14 Communication Interface 15 External Storage Device 16 Storage Medium Reading Device 17 Scanner 18 Input / Output Device

Claims (3)

A character string is extracted from a document image obtained by converting the document into image data, the center position in the height direction of each character in the character string is obtained, and the regularity of the center position in the height direction of all characters A character recognition method for determining whether a character string is a handwritten character or a type, and performing character recognition based on the determination result,
A character recognition method characterized by discriminating a handwritten character from a printed character based on the regularity of the center position of the character obtained without including a cloud point included in the character string. Extracting a character string from a document image obtained by converting the document into image data, obtaining a center position in the height direction of each character in the character string, and by regularity of the center position in the height direction of all characters, A character recognition method for determining whether the character string is a handwritten character or a type, and performing character recognition based on the determination result,
A character recognition method characterized by discriminating between handwritten characters and printed characters based on the regularity of the center position of the characters obtained without including lowercase letters included in the character string. A character string extracting means for extracting a character string from a document image obtained by converting the document into image data;
A handwriting / type discrimination unit that obtains a center position in the height direction of each character in the character string and discriminates whether the character string is a handwritten character or a type by regularity of the center position in the height direction of all characters;
A character recognition apparatus comprising character recognition means for performing character recognition based on the determination result, wherein the handwriting / printing type determination means obtains a center position of each character by removing a cloud point from the character in the character string. The second means for obtaining the center position of each character by removing lowercase letters from the character string, the regularity of the center position of the character obtained by the first and second means, and the size of the variation A character recognition apparatus comprising a third means for determining whether a character string is a handwritten character or a printed character.

Images