JP3466899B2 - Character recognition device and method, and program storage medium - Google Patents

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a character recognition device and method for recognizing a character entered in a document having a one-character input field, and a program storage medium for storing a program used to implement the character recognition device. In particular, the present invention relates to a character recognition device and method for improving the accuracy of character recognition, and a program storage medium in which a program used for realizing the character recognition device is stored.

[0002] In many cases, a document having a one-character input field is made to be filled with characters such as amount of money. From this, a character recognition device for recognizing a character entered in the one-character input field is required. It is necessary to make such a character recognition device practical by improving the character recognition accuracy.

[0003]

2. Description of the Related Art As shown in FIG. 19, handwritten characters entered in a document having a one-character input field are entered beyond the frame specified by the input field or in a form in which characters are connected to each other. However, it has a problem that recognition is very difficult.

To explain further, this problem occurs when the image quality of printed characters is deteriorated by faxing or copying, and also when printing is performed by a printer with poor accuracy.

Against this background, the present applicant separates the frame forming the one-character input box from the characters entered in the one-character input box, as disclosed in a series of patent applications. A technology that develops a technology to cut out characters, cuts out such separated characters using a cutting out algorithm that assumes that the characters are separated, and recognizes the characters cut out by the character cutting out technology. Has been developed.

Then, the character separated in such a manner is processed, and a continuous character portion between the characters is detected,
We have developed a technology that cuts out characters by deleting them, and has developed a technology that recognizes the characters cut out by the character cutting technology.

[0007] In addition to the applicant, various techniques for recognizing a character entered in a one-character input box have been proposed.

[0008]

Certainly, the character recognition technology developed by the present applicant is characterized in that a character entered in a one-character input box can be read with high accuracy.

However, handwritten characters are very diverse, and the deterioration of the image quality of printed characters is also diverse. From now on, if the specific algorithm is followed, it is unavoidable that the recognition accuracy is limited.

Therefore, according to the prior art, there is a problem that the recognition accuracy of the character entered in the one-character input box may not always be sufficient. The present invention has been made in view of the above circumstances, and has a new character that realizes an improvement in the character recognition accuracy when adopting a configuration for recognizing a character entered in a document having a one-character input field. An object of the present invention is to provide a recognition device and method, and to provide a new program storage medium in which a program used to realize the character recognition device is stored.

[0011]

FIG. 1 shows the principle configuration of the present invention. In the figure, 1 is a character recognition device equipped with the present invention, which recognizes a character entered in a 1-character input field separated by ruled lines, and 2 is an image scanner connected to the character recognition device 1. The document image is read and input to the character recognition device 1.

The character recognition device 1 of the present invention comprises an image memory 10, an extraction means 11, a takeout means 12, a first recognition means 13, a second recognition means 14 and a determination means 15.

The image memory 10 stores a document image (binarized) in which characters are entered in a 1-character input field delimited by ruled lines. The extraction means 11 extracts the ruled lines of the input document image. The extraction unit 12 specifies the contact character portion between the ruled line extracted by the extraction unit 11 and the character of the input document image, and extracts only the character from the input document image according to the specification result.

The first recognizing means 13 cuts out characters one by one from the input document image by determining whether or not the contact character portion specified by the extracting means 12 is valid according to the character size. Perform recognition processing.

The second recognizing means 14 detects the subsequent character portion of the character extracted by the extracting means 12 and deletes it, thereby cutting out the character one by one from the input image and executing the recognition processing.

The deciding means 15 decides a final character recognition result from the recognition result of the first recognizing means 13 and the recognition result of the second recognizing means 14. Here, the function of the character recognition device 1 of the present invention is specifically realized by a program, and this program is stored in a floppy disk or the like, or is stored in a disk or the like of a server, etc. The present invention is realized by being installed in the character recognition device 1 and operating on the memory.

In the character recognition apparatus 1 of the present invention thus constructed, when the extracting means 11 extracts the ruled lines of the input document image, the extracting means 12 extracts the ruled lines and the input document image. In addition to identifying the contact character portion with the possessed character, only the character is extracted from the input document image according to the identification result.

In response to this, the first recognizing means 13 validates the contact character portion specified by the extracting means 12 by using the character size in accordance with the character cutting algorithm which assumes that the characters are separated from each other. By deciding whether or not to do so, the characters are cut out one by one from the input document image and the recognition processing is executed.

On the other hand, the second recognizing means 14 detects the subsequent character portion of the character extracted by the extracting means 12 without considering the character size, and deletes it to delete the character one by one from the input image. Cut out and execute the recognition process.

At this time, when the second recognizing means 14 determines that there is no subsequent character portion of the character, the second recognizing means 14 does not execute the character recognizing process as shown in FIG.

Then, the deciding means 15 receives the recognition result of the first recognizing means 13 and the recognition result of the second recognizing means 14, and finally gives the recognition result of the smaller distance value, for example. Determined as a recognition result.

As described above, in the character recognition device 1 of the present invention, when the character recognition is performed while deleting the ruled line of the one-character input field, the characters are separated from each other as shown in FIG. Recognition target character that is cut out using the character cutting algorithm that assumes that
, The character recognition processing is executed, and the character recognition processing is executed with the character cut out by deleting the subsequent character portion of the character as the recognition target (left side in FIG. 3). Since the final character recognition result is obtained from the recognition result, the character recognition accuracy can be improved as compared with the conventional case.

[0023]

BEST MODE FOR CARRYING OUT THE INVENTION The present invention will be described in detail below with reference to an embodiment in which handwritten numbers (0 to 9) entered in a financial document are recognized.

FIG. 4 shows a character recognition device 1 having the present invention.
1 illustrates an example. The character recognition device 1 of the present invention according to this embodiment includes the image scanner 2 described in FIG.
The image memory 10 described in FIG. 1, the recognition control program 20, the first recognition program 21, the second recognition program 22, the frame extraction program 23, and the contact character portion extraction program 24 are provided.

Here, the program developed in the character recognition device 1 is installed via a floppy disk or a line. This recognition control program 2
When 0 is issued as a recognition request for handwritten characters (written in a 1-character input field delimited by ruled lines) of the document image stored in the image memory 10, as shown in the processing flow of FIG. First, in step 1, the frame extraction program 23 is activated.

When activated in this way, the frame extracting program 23 extracts the frame of the one-character input field (the frame of characters separated by ruled lines) of the document image in accordance with the configuration described later.

Subsequently, in step 2, after waiting for the extraction processing by the frame extraction program 23 to be completed, when this extraction processing is completed, the routine proceeds to step 3 to activate the contact character portion extraction program 24.

When activated in this way, the contact character portion extraction program 24 refers to the extraction result of the frame extraction program 23 according to the configuration described later, while referring to the extraction result of the frame extraction program 23, a handwritten character portion ( (Character portion that touches the frame) is extracted, and only the handwritten character of the document image is extracted according to the extraction result.

Then, in step 4, waiting for the extraction processing by the contact character portion extraction program 24 to end,
When this extraction process is completed, the process proceeds to step 5 and the first
The recognition program 21 is started.

When activated in this manner, the first recognition program 21 refers to the extraction result of the contact character portion extraction program 24 and, in accordance with the configuration described later, a character different from the second recognition program 22. The handwritten characters are cut out one by one and recognized according to the cutout algorithm.

Then, in step 6, the handwriting character recognition process by the first recognition program 21 is completed, and when this recognition process is completed, the process proceeds to step 7 and the second recognition program 22 is executed. to start.

When activated in this way, the second recognition program 22 refers to the extraction result of the contact character portion extraction program 24 and, in accordance with the configuration described later, a character different from the first recognition program 21. The handwritten characters are cut out one by one and recognized according to the cutout algorithm.

Subsequently, in step 8, the handwriting character recognition process by the second recognition program 22 is waited for, and when this recognition process is completed, the process proceeds to step 9 to execute the first recognition program 21. The final recognition result of the handwritten character is determined from the recognition result and the recognition result of the second recognition program 22.

Next, the frame extraction program 23 will be described. As described above, the frame extracting program 23 performs the process of extracting the frame of the one-character input field of the document image.

The frame extracting program 23 may have any structure as long as it extracts the frame of the document image. For example, Japanese Patent Laid-Open No. 6-3 filed by the present applicant.
It is possible to use those having the configurations disclosed in 09498, JP-A-7-28937, JP-A-8-305796, and JP-A-9-50527.

FIG. 6 shows a functional block diagram of the frame extraction program 23 disclosed by the present applicant in Japanese Patent Laid-Open No. 8-305796. The frame extraction program 23 according to the configuration of FIG.
The connection pattern extraction unit 230, the projection unit 231, the straight line detection unit 232, the first four side detection unit 233, and the tracking unit 23.
4, a second four-side detection unit 235, and a frame extraction unit 236.

The connection pattern extraction unit 230 inputs the document image stored in the image memory 10 and extracts an 8-connection connection pattern which is connected in any of eight directions of vertical, horizontal, and diagonal by a labeling process. . The connection pattern obtained at this time is a frame in which characters do not touch,
Either the character that is not in contact with the frame or a part thereof, or the character that is in contact with the frame or a part thereof.

The extracted connection patterns include straight lines that form the character frame of the one-character input field. The projection unit 231 projects the connection pattern extracted by the connection pattern extraction unit 230 in the horizontal direction and the vertical direction. The straight line detection unit 232 detects horizontal lines / vertical lines from the horizontal / vertical projection information obtained by the projection unit 231. The first four-side detecting unit 233 detects the four sides of the rectangle detected from the horizontal line / vertical line detected by the straight line detecting unit 232.

The tracking unit 234 is a straight line detecting unit 232 / first
In order to obtain a straight line with a thin line width that cannot be detected by the four-side detection unit 233, the horizontal line and the vertical line that connect eight within the width of a set number of lines (for example, 2 to 3 lines) are traced. The second four-side detecting unit 235 detects the four sides of the rectangle detected from the horizontal / vertical lines tracked by the tracking unit 234.

The frame extracting section 236 is provided in the first four side detecting section 23.
The frame of the one-character input field is extracted from the four sides detected in 3, and the frame of the one-character input field is extracted from the four sides detected by the second four-side detection unit 235.

According to the configuration of FIG. 6, the frame extracting program 23 can extract the frame of the one-character input field of the document image without prior knowledge about the frame of the one-character input field of the document image. Become.

Next, the contact character portion extraction program 24 will be described. As described above, the contact character portion extraction program 24 extracts the handwritten character portion (the character portion that contacts the frame) overlapping the frame portion of the one-character input field, and only the handwritten character of the document image is extracted according to the extraction result. Perform extraction processing.

As will be described later, the first and second recognition programs 21 and 22 perform recognition processing on the handwritten characters of the document image from which the ruled lines are removed. In the case of contact, if the frame is deleted, the handwritten character portion will be lost. Therefore,
In order to interpolate this missing portion, the contact character portion extraction program 24 is used to perform processing for extracting a handwritten character portion that overlaps the frame portion of the one-character input field.

The contact character portion extraction program 24 is
Any configuration may be used as long as it extracts a handwritten character portion that overlaps the frame portion of the one-character input field.
For example, JP-A-6-309498, JP-A-7-28937, JP-A-8-305796, and JP-A-9-5 filed by the present applicant.
The configuration disclosed in No. 0527 can be used.

FIG. 7 shows a functional block diagram of the contact character portion extraction program 24 disclosed by the present applicant in Japanese Patent Laid-Open No. 8-305796. The contact character portion extraction program 24 according to the configuration of FIG.
And a frame separation unit 241, an intersection calculation unit 242, an intersection correspondence unit 243, and a contact character portion extraction unit 244.

The concatenation pattern attribute adding unit 240, for the concatenation pattern extracted by the concatenation pattern extracting unit 230 of the frame extracting program 23, "frame", "character pattern or part thereof", "frame and character pattern or Any attribute of "contact pattern (contact character pattern) with a part thereof" is added.

The frame separation unit 241 separates the frame from the connection pattern to which the connection pattern attribute adding unit 240 has added the attribute "frame" or "contact character pattern of character and frame". Specifically, the width of the side of the frame portion is calculated, and the frame is removed based on the calculated width. Then, the pattern from which the frame is removed is re-labeled to remove the pattern having a small area as noise, and among the patterns to which the attribute is not added by the concatenated pattern attribute adding unit 240, the patterns that remain after the frame is removed Is performed by adding the attribute of "contact character pattern", and adding the attribute of "frame" only to the pattern in which nothing remains after removing the frame.

The intersection calculating unit 242 first calculates the intersections of the frame and the character in the contact character pattern, and then, at all of these intersections, from the position to the outside of the frame up to about the width of the character. The line segment is searched to calculate an intersection outside the frame, and the area of the searched character line segment is obtained. Subsequently, when the obtained area of the character line segment is equal to or smaller than the threshold value, the character line segment is regarded as noise and removed, and the intersection is determined not to be the intersection of the character and the frame.

The intersection associating unit 243 is used by the intersection calculating unit 24.
Based on the intersection point information obtained in 2, the directionality of the character line segment that is in contact with the frame is obtained. Further, the distance between the two character lines contacting both sides of the frame is calculated. Then, each intersection point between the character and the frame is associated with the obtained directionality and distance and the condition of the continuity of the character line segment based on this directionality. With this association processing, if explained in the example of FIG. 8A, the intersection A and the intersection C are associated with each other, and the intersection B and the intersection D are associated with each other.

The contact character portion extraction unit 244 determines that the image of the frame portion defined by the intersections associated by the intersection association unit 243 is a character component. Then, according to the determination result, only the character pattern is extracted without being influenced by the frame.

In accordance with the configuration of FIG. 7, the contact character portion extraction program 24 uses the contact character pattern shown in FIG. 8 (a) to handwrite as shown in FIG. The character part (dot part in the figure) can be extracted. Then, according to the extraction result of the contact character portion, as shown in FIG. 9, the frame of the one-character input field of the document image is removed so that only the handwritten character entered in the one-character input field can be extracted. Become.

Next, the handwritten character recognition processing executed by the first recognition program 21 will be described. When the first recognition program 21 is activated by the recognition control program 20, the first recognition program 21 refers to the extraction result of the contact character portion extraction program 24 and executes a character cutout algorithm on the assumption that characters are separated from each other. Using it, the handwritten character is cut out from the document image, and the recognition process for it is executed.

As this character segmentation algorithm,
Although any one may be used, for example, the one disclosed in Japanese Patent Application Laid-Open No. 8-305796 by the present applicant can be used. Figure 10
An example of a processing flow executed by the first recognition program 21 is illustrated in FIG. Then, according to this processing flow,
The recognition process executed by the first recognition program 21 will be described.

When the first recognition program 21 is activated by the recognition control program 20, as shown in the processing flow of FIG. 10, first, in step 1, the extraction information of the contact character portion extraction program 24 is obtained. To do.

Then, in step 2, the average character size of handwritten characters written in the document image is calculated. In Japanese Patent Laid-Open No. 8-305796, this average character size is the average size of handwritten characters written in the document image from the size of each character extracted by the contact character part extraction program 24 (approximately obtained by a circumscribed rectangle). Although the structure for calculating the size is adopted, as a simple method, a structure for calculating from the size of the one-character input field extracted by the frame extracting program 23 can be adopted.

Subsequently, in step 3, the size of each character extracted by the contact character part extraction program 24 is compared with the average size calculated in step 2 to extract the frame part extracted by the contact character part extraction program 24. It is determined whether or not the handwritten character portion (contact character portion) overlapping with is the original character portion, and when it is determined that it is not the original character portion, the character is cut out by deleting it.

Since the one-character input field is provided,
The characters "0" to "9" entered in a financial document are basically separated from other characters by being entered in the frame of the one-character input field. Part may be filled out of the frame, or may be filled in by connecting with the adjacent character. Therefore, when the contact character portion extracted by the contact character portion extraction program 24 is valid and does not fit in one character size, the extracted contact character portion is invalidated and deleted, and the contact character portion is converted into one character size. When it fits, the extracted contact character portion is validated so that an accurate character is cut out.

For example, as shown in FIG.
By entering "2", "5", and "3" in succession,
When it crosses the frame of the one-character input box, the contact character portion extracted by the contact character portion extraction program 24 is deleted in consideration of the fact that the character size becomes larger than the average character size. "5" and "3" are cut out accurately.

Further, as shown in FIG. 11B, "1"
If a part of "0" and "0" is out of the range of the 1-character input field, the contact character part extracted by the contact character part extraction program 24 is originally taken into consideration in that the character size fits within the average character size. "1" and "0" are accurately cut out by treating them as characters of.

After the handwritten characters entered in the one-character input box are cut out one by one in this way, the process proceeds to step 4 and the unprocessed characters are cut out from the cut out characters.
Then, in step 5, it is determined whether or not all the characters have been selected.

At this step 5, when it is judged that all the characters have not been selected, that is, step 4
Then, when it is determined that one unprocessed character can be selected from the cut out characters, the process proceeds to step 6, and the feature amount used for character recognition of the selected character is calculated. As the feature amount calculated at this time, any feature proposed in the conventional character recognition processing may be used.

Then, in step 7, the calculated feature amount is used to measure the distance to the registered character to be recognized. In a financial document, 10 characters "0" to "9" are registered characters, and therefore the distance between these 10 characters is measured.

Subsequently, in step 8, as the recognition result of the character selected in step 4, the registered character having the smallest distance measured in step 7 is determined, and then the process returns to step 4 to proceed to recognition of the next character. To go.

Then, by repeating the processing of steps 4 to 8, when it is judged in step 5 that all the characters have been selected, that is, the character recognition of the handwritten characters cut out in step 3 is terminated. When it is determined, the process proceeds to step 9, the total value of the distance between the recognition result and the registered character determined in step 8 is calculated, and the entire process is ended.

In this way, the first recognition program 2
1 is a document image that uses a character segmentation algorithm that presumes that characters are separated, while validating and invalidating the character portion that overlaps the frame portion of the 1-character input field. Characters are recognized by cutting out handwritten characters one by one and performing a recognition process on the cut out characters.

By the recognition processing of the first recognition program 21, for example, the handwritten character "10000" entered in the form of continuous characters shown in the upper part of FIG. 3 is cut out as shown on the right side of FIG. It will be recognized as "18886". Then, with respect to this recognition result, "500" is calculated as the total value of the distances of the recognition results of each character.

Next, the recognition processing executed by the second recognition program 22 will be described. Second recognition program 2
2, when the recognition control program 20 is activated, the contact character portion extraction program 24 detects a handwritten character from which the frame of the one-character input box is removed as a recognition target, and detects a continuous character portion connecting characters. , The handwritten character is cut out from the document image by deleting it, and the recognition process for it is executed.

As the character segmentation algorithm,
Although any one may be used, for example, the one disclosed in Japanese Patent Application Laid-Open No. 7-192094 by the present applicant can be used. 12 and 13 illustrate an example of a processing flow executed by the second recognition program 22. Next, the recognition processing executed by the second recognition program 22 will be described according to this processing flow.

When the second recognition program 21 is started by the recognition control program 20, as shown in the processing flows of FIGS. 12 and 13, first, in step 1, the contact character portion extraction program 24 extracts the second character recognition program 21. By obtaining the extracted information of only the written handwritten characters, the connection pattern forming the handwritten characters is obtained.

Subsequently, in step 2, a horizontally long concatenated pattern is extracted from the concatenated patterns read in step 1 as a candidate for a continuous character in which a plurality of characters are connected. This extraction processing is performed by obtaining a circumscribing rectangle for each connection pattern, calculating a vertical-horizontal ratio of the circumscribing rectangle, and comparing it with a predetermined threshold value.

Subsequently, in step 3, a process of extracting a horizontal continuous line (a line forming a continuous character portion) of the horizontally long pattern extracted in step 2 is performed. In this step 3, first, "character pattern area / area of circumscribing rectangle" is calculated, and the length of the straight line to be extracted is determined according to the value. Specifically, when the value of "character pattern area / area of circumscribed rectangle" is large, a long horizontal line is extracted,
When this value is small, a short horizontal line is extracted.

That is, a large value of "character pattern area / area of circumscribing rectangle" means that the pattern width of the character pattern is large. At times like this,
This is because if a short straight line is extracted, a large number of straight lines will be present in the original character portion and these will be extracted. A small value of "character pattern area / circumscribing rectangle area" means that the pattern width of the character pattern is small. At times like this,
This is because if a long straight line is extracted, the original horizontal line will not be extracted.

In step 3, the process shown in FIG.
As shown in (a), the character pattern is vertically divided according to the length of the straight line to be extracted, and the projection process is performed within the divided range. At this time, considering the existence of diagonal lines,
When the projection values of surrounding lines are added together in the horizontal direction (so-called adjacent projection method) and the projection value is equal to or more than a predetermined threshold value, it is recognized that a straight line exists in that part, and the range Is approximated to a rectangle to form a rectangular straight line.

In step 3, the process shown in FIG.
As shown in (b), a long straight line is extracted by integrating contacting rectangular straight lines, and the longest straight line among them is processed as a horizontal continuation line.

When a horizontal continuation line (a line forming a continuation character portion) is extracted in step 3, then in step 4,
When it is determined whether or not the horizontal continuation line is detected and it is determined that the horizontal continuation line is not detected, step 5
Then, the process proceeds to step S3 and records that the recognition process is not executed, and ends the process.

On the other hand, when it is judged in step 4 that the horizontal continuation line is detected, the process proceeds to step 6 to determine the vertical separation line. As shown in FIG. 15A, this vertical separating line determination processing finds an intersection with the character pattern by tracing the horizontal continuation line from one end point of the lower side of the rectangular continuous horizontal continuation line, The contour search of the character pattern is started, and when the horizontal continuation line is reached, the contour search is temporarily stopped. Subsequently, the horizontal continuation line is traced to find an intersection with the next character pattern, the search for the contour of the character pattern is restarted from that point, and when the horizontal continuation line is reached, the contour search is temporarily stopped. This is repeated until the other end point on the lower side of the horizontal continuous line approximated by a rectangle is reached.

Finally, the number of times the contour search is performed becomes the number of characters, and the region from the start point to the end point of the contour search is one character. The vertical separation line is a vertical line that separates characters from each other, and as shown in FIG. 15B, is a horizontal continuous line between the end point and the start point of the contour search and which is approximated by a rectangle. Determined at the position that has the width value.

When the vertical separation line is determined in step 6, subsequently, zero determination is performed in step 7. As shown in FIG. 16, this zero determination processing is performed in a single character area surrounded by vertical separation lines and horizontal continuation lines, starting from a blank part surrounded by horizontal continuation lines and character patterns, and radially in multiple directions. The search is performed to check whether or not it has a loop structure.

When the zero judgment is made in step 7,
In step 8, the process of separating characters is performed. This character separation processing is performed by deleting unnecessary horizontal continuation lines. That is, in the character determined to be zero, the horizontal continuation line is an unnecessary line and is deleted. In this deletion process, as shown in FIG. 17A, the vertical separation line is removed, and the part where the thickness of the continuation line changes abruptly, the part where the slope of the continuation line and the derivative value thereof change abruptly. By deleting up to. On the other hand, as for the character determined to be not zero, as shown in FIG. 17B, the vertical separation line is separated from other characters, but the continuation line is not deleted.

In this way, the second recognition program 2
2 is an unnecessary continuous character portion (in the case of a financial document, most of them are written when "0" and "0" are continuously written in succession in the case of a financial document) according to the processing of steps 1 to 8. ) Is deleted, the handwritten character is cut out from the document image.

Although not described in detail in this processing flow, even when only a part of characters has a continuous character part (a character already separated is present), the process does not proceed to step 5, The process proceeds from 6 to step 8 to delete the subsequent character portion.

When the handwritten character is cut out from the document image in step 8, the process proceeds to step 9, and one unprocessed character is selected from the cut out characters, and the following step 1
At 0, it is determined whether all the characters have been selected.

In this step 10, when it is judged that all the characters have not been selected, that is, step 9
Then, when it is determined that one unprocessed character can be selected, the process proceeds to step 11 (process flow of FIG. 13),
Using the same calculation method as that used by the first recognition program 21, the feature amount used for character recognition of the selected character is calculated.

Then, in step 12, the calculated feature amount is used to measure the distance to the registered character to be recognized. In a financial document, 10 characters “0” to “9” are registered characters to be recognized, and therefore the distance between these 10 registered characters is measured.

Then, in step 13, as the recognition result of the character selected in step 9, the registered character with the smallest distance measured in step 12 is determined, and then the process returns to step 9 to proceed to recognition of the next character. To go.

When it is judged in step 10 that all the characters have been selected by repeating the processing of steps 9 to 13, that is, step 8
When it is determined that the character recognition of the handwritten character cut out in step is to be ended, the process proceeds to step 14, the total value of the distance between the recognition result and the registered character determined in step 13 is calculated, and the entire process is ended. .

In this way, the second recognition program 2
2 designates the handwritten character with the frame of the one-character input field removed by the contact character portion extraction program 24 as a recognition target,
By detecting a continuous character portion connecting characters and deleting it, a handwritten character is cut out from the document image, and recognition processing for the handwritten character is executed.

By the recognition processing of the second recognition program 22, for example, the handwritten character "10000" entered in the form of continuous characters shown in the upper part of FIG. 3 is cut out as shown on the left side of FIG. It will be recognized as "10000". Then, for this recognition result, “20” is calculated as the total value of the distances of the recognition result.

As described in the processing flow of FIG. 5, the recognition control program 20 performs the processing of determining the final recognition result when the first and second recognition programs 21 and 22 finish the recognition processing.

FIG. 18 shows an example of the determination process executed by the recognition control program 20. Next, this processing flow will be described. Recognition control program 20
When the first and second recognition programs 21 and 22 finish the recognition process, first, in step 1, the second recognition program 22 executes the recognition process as shown in the process flow of FIG. Or not. As described above, the second recognition program 22 does not perform the recognition process when the continuous character portion does not exist, but does not perform the recognition process but only records the fact, so whether or not this record is left. By determining, it is determined whether or not the second recognition program 22 has performed the recognition process.

When it is judged by this judgment processing that the second recognition program 22 has not executed the recognition processing, the routine proceeds to step 2 and the first recognition program 21 is executed.
The recognition result of is output as the final recognition result, and the process ends.

On the other hand, when it is judged by this judgment processing that the second recognition program 22 has executed the recognition processing, the routine proceeds to step 3 and the first recognition program 21 is executed.
And the total distance value output by the second recognition program 22 are compared.

When it is determined by this comparison process that the total distance value output by the first recognition program 21 is smaller, the routine proceeds to step 4, where the first recognition program 2
The recognition result of 1 is output as the final recognition result. On the other hand, when it is determined that the total distance value output by the second recognition program 22 is smaller, the process proceeds to step 2,
The recognition result of the second recognition program 22 is output as the final recognition result.

In this way, the recognition control program 20
When the recognition result of the first recognition program 21 and the recognition result of the second recognition program 22 are received, the one with the smaller total distance value, that is, the recognition result judged to be more similar Is selected and output as the final recognition result.

With this determination processing, as will be described with reference to the example of FIG. 3, the recognition result “10000” of the second recognition program 22 is output as the final recognition result.

The character cutout algorithm used by the first recognition program 21 is based on the premise that the characters are separated from each other, and from this, it may happen that the succeeding character part cannot be removed. In such a case, the recognition result of the second recognition program 22 that cuts out characters by removing the subsequent character portion is more likely to be the correct answer.

On the other hand, the character segmentation algorithm used by the second recognition program 22 does not assume that characters are separated from each other.
Although the characters are originally separated by being filled in the character input field, the succeeding character portion may be erroneously detected. In such a case, the recognition result of the first recognition program 21 that cuts out a character on the premise that the character is separated is more likely to be the correct answer.

In consideration of such a difference in characteristics, in the present invention, the recognition control program 20 can make a more correct answer from the recognition result of the first recognition program 21 and the recognition result of the second recognition program 22. The recognition result of the one having a higher property is used as the final recognition result. By adopting this configuration, it becomes possible to remarkably improve the character recognition accuracy as compared with the conventional case.

Although the present invention has been described with reference to the illustrated embodiment,
The present invention is not limited to this. For example, in the embodiment, the first recognition program 21 and the second recognition program 22 cut out handwritten characters and then perform character recognition using the same recognition algorithm. Different recognition algorithms may be used as long as they can be compared.

Further, in the embodiment, assuming a financial document,
Although it is assumed that "0" to "9" are the recognition targets, the present invention is not limited to the application of "0" to "9" as the recognition targets.

In the embodiment, it is assumed that handwritten characters are to be recognized. However, print characters whose image quality is deteriorated,
The present invention can be directly applied to characters printed by a printer with low accuracy.

[0102]

As described above, in the character recognition apparatus of the present invention, when the character recognition is performed while deleting the ruled line of the one-character input field, the characters are separated from each other. Performs character recognition processing with the characters cut out using the presumed character cutting algorithm as the recognition target, and also performs character recognition processing with the characters cut out by deleting the subsequent characters of the characters as the recognition target. By adopting the configuration described above and obtaining the final character recognition result from the two recognition results, the character recognition accuracy can be improved as compared with the conventional case.

[Brief description of drawings]

FIG. 1 is a principle configuration diagram of the present invention.

FIG. 2 is an explanatory diagram of the present invention.

FIG. 3 is an explanatory diagram of the present invention.

FIG. 4 is an example of the present invention.

FIG. 5 is a processing flow executed by a recognition control program.

FIG. 6 is an explanatory diagram of a frame extraction program.

FIG. 7 is an explanatory diagram of a contact character portion extraction program.

FIG. 8 is an explanatory diagram of a contact character portion extraction process.

FIG. 9 is an explanatory diagram of character extraction processing.

FIG. 10 is a processing flow executed by a first recognition program.

FIG. 11 is an explanatory diagram of character cutting processing.

FIG. 12 is a processing flow executed by a second recognition program.

FIG. 13 is a processing flow executed by a second recognition program.

FIG. 14 is an explanatory diagram of a horizontal continuation line extraction process.

FIG. 15 is an explanatory diagram of a vertical separation line determination process.

FIG. 16 is an explanatory diagram of zero determination processing.

FIG. 17 is an explanatory diagram of character separation processing.

FIG. 18 is a processing flow executed by a recognition control program.

FIG. 19 is an explanatory diagram of handwritten characters entered in the one-character input field.

[Explanation of symbols]

1 character recognition device 2 image scanner 10 image memory 11 Extraction means 12 Means of taking out 13 First recognition means 14 Second recognition means 15 Determining means

Continuation of front page (56) Reference JP-A-9-54813 (JP, A) JP-A-63-251874 (JP, A) JP-A-7-282190 (JP, A) JP-A-6-333089 (JP , A) JP 7-192094 (JP, A) JP 8-305796 (JP, A) JP 11-25220 (JP, A) JP 10-162104 (JP, A) JP 8-202822 (JP, A) JP-A-59-98283 (JP, A) (58) Fields investigated (Int. Cl. ⁷ , DB name) G06K 9/00-9/82

Claims (4)

(57) [Claims] 1. A character recognition device for recognizing a character entered in a one-character input field separated by a ruled line, an extracting means for extracting a ruled line of an input image, and a ruled line extracted by the extracting means and an input image. By identifying the contact character portion with the character and determining the extraction means for extracting only the character according to the identification result and determining whether the contact character portion is valid according to the character size, the character is extracted from the input image. The first recognition means for recognizing the character by cutting it out character by character, and the continuous character part of the character extracted by the extracting means are detected,
A decision to determine the final character recognition result from the second recognition unit that deletes it and recognizes the characters one by one from the input image and the recognition results of the first and second recognition units. A character recognition device characterized by comprising means. 2. The character recognizing device according to claim 1, wherein the second recognizing means performs processing so as not to execute the character recognizing processing when determining that there is no continuous character portion. Character recognition device. 3. A character recognition method for recognizing a character entered in a one-character input field separated by a ruled line, the first process step of extracting a ruled line of an input image, and the ruled line extracted according to the character size. By deciding whether or not to validate the contact character portion with the character that the input image has, the second processing step in which the character is cut out and recognized one by one from the input image, and the extracted ruled lines and the input image A character is extracted from the input image one by one by recognizing the characters that come in contact with the character that it has, by extracting only the character from the input image, detecting the subsequent character part of the character, and deleting it. A character recognition method comprising: the third processing step; and a fourth processing step that determines a final character recognition result from the recognition results obtained in the second and third processing steps. 4. A program storage medium for storing a program used for realizing a character recognition device for recognizing a character entered in a one-character input field separated by a ruled line, and extracting for extracting a ruled line of an input image. Processing, the contact character portion of the ruled line extracted by the extraction processing and the character of the input image is specified, the extraction processing of extracting only the character according to the specification result, and the contact character portion is validated according to the character size. By deciding whether or not to do so, the first recognition processing for recognizing the characters by cutting out the characters one by one from the input image and the succeeding character portion of the characters extracted by the extraction processing are detected
A decision to determine the final character recognition result from the second recognition processing by cutting out the characters one by one from the input image and recognizing them by deleting them and the recognition results of the first and second recognition processing. A program storage medium characterized in that a program for causing a computer to execute processing is stored.