JP2000172783A - Character string recognizing method and storage medium - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a character recognizing method capable of performing accurate character segmentation such as even when the sizes of characters are not equal. SOLUTION: In this character string recognizing method, geometrical feature quantity (e.g. the height of a partial character pattern) of a partial character pattern is found, the geometrical feature quantity of the partial character pattern and the average value of the geometrical feature quantity of plural neighborhood partial character patterns are made statistic feature quantity, the width of a formal character pattern is calculated as the geometrical feature quantity of a formal character pattern obtained by integrating adjacent partial character patterns, the formal character pattern is segmented (S3) and a character string candidate is acquired when evaluation value calculated from the geometrical feature quantity of the formal character pattern and the statistic feature quantity of the partial character pattern in the formal character pattern is equal to or less than threshold, neighborhood numbers are further changed (S5) and plural character string candidates are repeatedly acquired, the certainty of each character string candidate is calculated from the character recognition result of each character pattern (S6) and the most probable character string candidate is defined as the recognition result of the character string by comparing the certainty (S7).

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a character string recognition method for extracting a plurality of character patterns from a character string pattern of a character image such as an input form and performing character recognition of the character strings. The present invention relates to a character string recognition method capable of character extraction and character recognition with little recognition error even for a character string having an indeterminate length.

[0002]

2. Description of the Related Art In order to recognize a character in a character string described in a form or the like, a process of cutting out a character from a character string one by one is required. Therefore, in the related art, based on the idea that the width and height of a character take a value close to each other, for example, in the case of horizontal writing, the average value of the height of a black pixel connection pattern that is a pattern of connected black pixels is standardized. The size is obtained, and characters are cut out by combining black pixel connection patterns so as to have a width close to the value. However, if the size of the characters is not constant over one line,
Incorrect clipping based on standard font size. In the conventional character segmentation method, the standard character size is calculated as the average value of the feature amounts (for example, the heights of the partial character patterns) of all the partial character patterns in the character string.
As shown in FIG. 8, small characters are erroneously integrated in some parts. In handwritten characters, the example shown in FIG. 8 often occurs, such as the number in the address part of the address. Therefore, for example, in the conventional technique disclosed in Japanese Patent Application Laid-Open No. 9-50488, when a character string includes characters of different sizes, a black pixel connection pattern (temporary character) is used.
Then, a plurality of standard character sizes are obtained based on the frequency distribution of the height, and the character string is divided into patterns having the same standard character size. That is, a plurality of character sizes are obtained, and character cutout is performed based on the plurality of character sizes. Thus, even when characters having different character sizes are mixed in a character string, character extraction can be performed with few errors.

[0003]

SUMMARY OF THE INVENTION
According to the conventional technique disclosed in Japanese Patent Application Laid-Open No. 9-50488, in the case where the size of characters is not uniform like handwritten characters, or in the case of hiragana or kanji which can be separated into right and left, a small temporary character is used. There is a possibility that erroneous clipping results may be produced due to integration more than necessary. Therefore, the problem to be solved by the present invention is to solve the problems of the conventional technology as described above, and to perform accurate character segmentation even in a case where characters are not uniform in size, and to perform character recognition. It is an object of the present invention to provide a character recognition method and the like that realizes more accurate character extraction by reflecting the result in character extraction and recognition processing, thereby realizing more accurate character recognition.

[0004]

According to a first aspect of the present invention, there is provided a character recognition method comprising the steps of: extracting a plurality of character patterns from a character string pattern of an input character image; In the character string recognition method for performing character recognition of the above, a black pixel connection pattern that is a pattern in which black pixels are connected is extracted, and a partial character pattern is created by combining the extracted black pixel connection patterns. As the geometric feature of the character pattern, the height of the circumscribed rectangle of the partial character pattern in the case of horizontal writing, and the width in the case of vertical writing, are calculated. The average value of the geometric feature values of a plurality of partial character patterns in the vicinity of the pattern is calculated as a statistical feature value, and the temporary character pattern obtained by temporarily integrating the adjacent partial character patterns is calculated. The width of the circumscribed rectangle of the temporary character pattern in horizontal writing and the height in vertical writing are calculated as the geometric feature of the temporary pattern, and the geometric feature of the temporary character pattern and the temporary character are calculated. By calculating the evaluation value of the provisional character pattern from the statistical feature amount of the partial character pattern included in the pattern, and determining whether to cut out the provisional character pattern as a character pattern based on the evaluation value and the threshold value Cut out a plurality of character patterns to obtain one character string candidate, and further obtain a plurality of character string candidates by repeating the above processing by changing the number of neighbors, and for each character string candidate, Calculate the certainty of the entire character string from the certainty of the character recognition result of each character pattern, compare the certainty of each character string candidate, and determine the most likely character string candidate as the character string recognition result. It was to be constant. According to the second aspect of the present invention, in the first aspect of the present invention, a character pattern in which the character cutout result is the same for all of the obtained character string candidates is determined, and one or more undetermined character patterns are determined. For partial character string candidates consisting of multiple character patterns, calculate the partial character string candidate confidence from the confidence of the character recognition result of each character pattern, and compare the confidence of each partial character string candidate to be the most likely. By determining a partial character string candidate as a recognition result of the partial character string, the recognition result of the entire character string is determined. According to the third aspect of the present invention, in the second aspect of the present invention, a character pattern in which the result of character extraction and character recognition is the same for all of the obtained character string candidates is determined, and further determined. For certain substring candidates consisting of one or more character patterns, calculate the substring candidate confidence from the confidence of the character recognition result of each character pattern, and compare the confidence of each substring candidate. Then, the most likely partial character string candidate is determined as the recognition result of the partial character string, so that the recognition result of the entire character string is determined. According to a fourth aspect of the present invention, there is provided a recording medium storing a program for causing an information processing apparatus such as a computer to execute a character string recognition process using the character string recognition method according to the first, second, or third aspect. did.

According to the first aspect of the present invention, a partial character pattern in which black pixel connection patterns are combined is created, and a horizontal character is written as a geometric feature of the created partial character pattern. Calculates the height of the circumscribed rectangle of the partial character pattern, and the width in the case of vertical writing, and calculates the geometric feature of the partial character pattern and the geometrical characteristics of a plurality of partial character patterns near the partial character pattern. The average value with the feature amount is calculated as a statistical feature amount, and the width of the circumscribed rectangle of the temporary character pattern in the case of horizontal writing, as the geometric feature amount of the temporary character pattern in which the adjacent partial character patterns are temporarily integrated, In the case of vertical writing, the height is calculated, and the provisional character pattern is calculated from the geometric feature amount of the provisional character pattern and the statistical feature amount of the partial character pattern included in the provisional character pattern. The evaluation value is calculated, one character string candidate based on the evaluation value and the threshold cut a plurality of character patterns by determining whether cutting the temporary character pattern as a character pattern is obtained,
Further, a plurality of character string candidates are obtained by changing the number of neighbors and repeating the above processing, and for each character string candidate, the confidence of the entire character string is determined from the confidence of the character recognition result of each character pattern. The degree is calculated, and the certainty degree of each character string candidate is compared to determine the most likely character string candidate as a character string recognition result. According to the second aspect of the present invention, in the first aspect of the present invention, a character pattern having the same character cutout result for all the obtained character string candidates is determined, and further, one or more undetermined character patterns are obtained. For partial character string candidates consisting of multiple character patterns, the confidence of the partial character string candidates is obtained from the confidence of the character recognition result of each character pattern, and the confidence of each partial character string candidate is compared. By determining a likely partial character string candidate as a recognition result of the partial character string, the recognition result of the entire character string is determined. Also,
According to a third aspect of the present invention, in the second aspect of the present invention, a character pattern in which the result of character extraction and character recognition is the same for all of the obtained character string candidates is determined.
Further, the confidence of the partial character string candidate is obtained from the confidence of the character recognition result of each character pattern with respect to the partial character string candidate composed of one or more character patterns that have not been determined. Are compared, and the most probable partial character string candidate is determined as the recognition result of the partial character string, whereby the recognition result of the entire character string is determined. According to a fourth aspect of the present invention, a program included in a recording medium is read and executed by a computer or the like to execute the character string recognition method according to the first, second, or third aspect. The device can be realized. Therefore, the recording medium can be easily distributed and sold as a software product independently of the apparatus. In addition, by using this software using a general-purpose computer or other hardware, the character string recognition technology of the present invention can be easily implemented with such hardware.

[0006]

Embodiments of the present invention will be described below in detail with reference to the drawings. FIG. 7 is a block diagram showing a configuration of a main part of an information processing apparatus for implementing the character string recognition method of the present invention. As shown in the figure, the information processing apparatus includes a scanner 1 that reads a character string pattern from a form or the like, a character cutout unit 2 that cuts out a plurality of characters from the character string pattern read (input) by the scanner 1 as a character pattern, A character recognition unit 3 that recognizes individual character patterns cut out by the character cutout unit 2 and converts them into character codes, such as a RAM loaded with a program and a CPU that operates according to the program. The system includes a system control unit 4 for managing and controlling, and an operation unit 5 including a keyboard and display means. The character cutout unit 2 and the character recognition unit 3 are also operated by, for example, a CPU. However, this CPU can be shared with the CPU of the system control unit 4. FIG. 1 is an operation flowchart of the character string recognition method according to the first embodiment of the present invention which operates in the information processing apparatus as described above. As shown in the figure, in the character string recognition method of this embodiment, first, the scanner 1
A character string pattern as shown in (A) is input (S1),
Furthermore, the statistical feature amount of the partial character pattern (described later)
The number of partial character patterns in the vicinity of the partial character pattern to be related to when obtaining is set by the operation unit 5 (S2). This number may be set prior to the operation of this operation flow.

[0007] Subsequently, using a method introduced in, for example, "Latest Trends in Separate Volume OPPlusE Image Processing Algorithm" (ed. By Takagi et al., 1986, New Technology Communications), the character cutout unit 2 converts a character string pattern inputted into A black pixel connection pattern is extracted, and a partial character pattern is created by combining the black pixel connection patterns (S3). For example, each of the rectangles shown in FIG. 2B is a partial character pattern. In this example, most of the partial character patterns consist of one black pixel connection pattern,
The partial character pattern on the left and right sides of "Chiku" consists of two black pixel connection patterns. Next, the height of the circumscribed rectangle of the partial character pattern in the case of horizontal writing and the width of the circumscribed rectangle of the partial character pattern in the case of vertical writing are obtained as the characteristic amount of each partial character pattern (S3). Further, as the statistical feature amount of the partial character pattern, a geometric feature amount (for example, height) of the partial character pattern and a geometric feature amount of a plurality of partial character patterns in the vicinity of the partial character pattern are obtained. Find the average value. As shown in FIG. 2C, when the respective partial character patterns are P1 to P14, for example, if the statistical feature amount of the partial character pattern P5 is obtained by three neighbors (the number of neighbors is three), it is determined by the partial character pattern The average value of the geometric feature amounts of the patterns P4, P5, and P6 is obtained. If it is near 5, the partial character pattern P
3, P4, P5, P6, and P7 are the average values of the geometric feature amounts. The same applies to other neighbor numbers. However, since there is no partial character pattern adjacent to only one side in the partial character pattern at the end, for example, the partial character pattern P1
Is the average of the geometric features of P1 and P2 if it is close to 3, and the average of the geometric features of P1, P2 and P3 if it is near 5. Note that the number of neighbors is the plurality of neighbors set in step S2, and among the plurality of neighbors, first, the statistical feature amount is obtained for the first number of neighbors.

Subsequently, a temporary character pattern is created by temporarily integrating adjacent partial character patterns, and the width of the circumscribed rectangle of the temporary character pattern in horizontal writing and high Is obtained (S3). And
An evaluation value V as a character of the provisional character pattern is calculated from the geometric feature amount of the provisional character pattern and the statistical feature amount of the partial character pattern included therein (the statistical feature amount in the case of the first number of neighbors). (S3). The evaluation value V is represented by S as the geometric feature of the provisional character pattern and Dj as the statistical feature of each partial character pattern included in the provisional character pattern.
Is obtained from the following equation. V = Σ (S−Dj) 2 / N where N is the number of partial character patterns included in the provisional character pattern. Further, the evaluation value V is compared with a predetermined threshold value to determine whether or not the temporary character pattern is cut out as a character (S3). That is, if the geometric feature value S of the provisional character pattern has a value close to the statistical feature value Dj of the partial character pattern, the character-likeness of the provisional character pattern is considered to be high, and the evaluation value V is determined in advance. If it is smaller than the threshold, the provisional character is cut out as a character. Note that the above threshold is obtained by a preliminary experiment using test data or the like.

To explain further in the case of FIG. 2D, first, partial character patterns P1 and P2 are temporarily integrated to obtain a provisional character pattern C1, and this geometric feature S is obtained. The evaluation value V is obtained by substituting 2 into the above equation as the value of the statistical feature values D1, D2, and N of the partial character patterns P1, P2 included in the provisional character pattern. When the evaluation value V is smaller than the threshold value, the provisional character pattern C1 is cut out as a character (S
3). Next, a temporary character pattern C2 in which the partial character pattern P3 adjacent to the integrated temporary character pattern C1 is integrated is obtained, and the evaluation value V is calculated in the same manner as in the case of C1. In this case, N is 3, and the partial character patterns included are P1, P
2, P3. However, in the case of the provisional character pattern C2, the evaluation result V is larger than the threshold value because the geometric feature amount S is large, so that the character is not cut out as a character. Next, the same processing is performed for the provisional character pattern C3 obtained by integrating the partial character pattern P3 that has not been integrated and the adjacent partial character pattern P4. These processes are sequentially executed until the end of the character string to obtain one character string candidate. Subsequently, the character recognition unit 3 performs character recognition for each character pattern of the character string candidate, and calculates a certainty factor (similarity) together with the character code (S4). In addition,
Such a character recognition method is known. For example, “Handwritten Kanji / Hiragana Recognition by Weighted Direction Exponential Histogram Method” (Tsuruoka et al., Transactions of the Institute of Electronics, Information and Communication Engineers, J70-D, 7
, Pp. 1390-1397, 1987).

In this way, when the character recognition of each character constituting the character string candidate acquired for one neighborhood number is completed, it is determined whether or not another neighborhood number is set (S5). (No at S5),
Returning to step S3, step S3 is performed for the next number of neighbors.
Repeat. Then, in step S5, if it is determined that the execution has been performed for all the neighbor numbers (Yes in S5), the calculation for obtaining the certainty of one character string candidate from the certainty of each character constituting the character string candidate is performed. This is performed for all character string candidates (S6). For example, the average of the certainty factors of the individual characters in the character string candidate is set as the certainty factor of the character string candidate. Then, a character string candidate having the highest certainty factor is selected as a character string recognition result (S7) and output (S8). In the above description, the character recognition in step S4 may be performed after step S5, that is, immediately before step S6. FIG. 3A shows an example in which, when the number of neighbors is small, excessive segmentation is performed by character extraction. If the number of neighbors is large, it is cut out correctly as shown in FIG. When character recognition is performed on these two character string candidates, a character string candidate with a larger number of neighbors has a higher degree of certainty, so that a correct result can be obtained. Conversely, FIG. 4B shows an example in which if the number of neighbors is large, characters are erroneously integrated and cut out. In this case, if the number of neighbors is small, it is cut out correctly as shown in FIG. As described above, even in a case where an erroneous result is output with one type of number of neighbors, by executing with a plurality of numbers of neighbors as in this embodiment and selecting using the result of character recognition,
You will get the correct result.

In the second embodiment of the present invention, a character pattern in which the character cutout result is the same for all the character string candidates acquired corresponding to the plurality of neighbors shown in the first embodiment Is determined, and the confidence of the partial character string candidate is determined from the confidence of the character recognition result of each character pattern for the partial character string candidate composed of one or more character patterns that have not been determined. By comparing the certainty factors of the character string candidates, the most likely partial character string candidate is determined as the recognition result of the partial character string, and the recognition result of the entire character string is determined. Hereinafter, the operation of this embodiment will be described with reference to the operation flow shown in FIG. First, a character string pattern is input by the scanner 1 (S11), and a plurality of neighborhoods of a partial character pattern for obtaining a statistical feature of the partial character pattern are set (S12). This number may be set prior to the operation of this operation flow. Subsequently, in the same manner as in the first embodiment, character extraction is performed when the character extraction unit 2 has the first number of neighbors (S13), and one character string candidate is obtained. Furthermore, character string candidates are obtained in the same manner for the second number of neighbors, and character string candidates are obtained for all the numbers of neighbors (Yes in S14), and a temporary character pattern common to all character string candidates is extracted. Then, the character pattern of that part is determined (S15). FIG. 6 shows an example of a temporary character pattern common to the two numbers of neighbors. FIG. 7A illustrates the case of a small number of neighbors, and FIG. 7B illustrates the case of a large number of neighbors. In this example, "Tokyo""Chiku""Ward""Sakae"
Five “towns” are common provisional character patterns, that is, character patterns that are determined first. Note that portions where the cutout results are different are processed as partial character string candidates (described later).

Next, in the same manner as in the first embodiment, character recognition is performed on the common provisional character pattern, the character recognition result of that portion is determined, and each partial character string candidate is constructed. Character recognition is also performed for each character pattern, and the degree of certainty of each character pattern is determined (S16). Then, the degree of certainty of each partial character string candidate is calculated (S17), a partial character string candidate having the highest degree of certainty is selected (S18), and the character recognition result of the partial character string is determined. Such a step S1
7. When the processing in S18 is sequentially performed until there is no unconfirmed character pattern, the character recognizing unit 3 combines the character recognition result with the already-confirmed character recognition result of the common part as a character string recognition result. Output (S19). In the above description, the part that is first determined as a part common to a plurality of character string candidates may not be simply a part where the cut-out character pattern is common, but may be a part where the character recognition result is also common. In the character recognition process, character recognition is performed in consideration of the relationship between the characters before and after the character. Therefore, different character recognition results may be obtained even if the cut-out character pattern is common. As described above, according to this embodiment, a result of a plurality of cutouts and a portion having the same processing result in the character recognition processing is adopted, so that the character string recognition accuracy is improved. that's all,
Although the character string recognition method of the present invention executed in the information processing apparatus shown in FIG. 7 has been described, the recording medium storing the character string recognition program according to the above-described character string recognition method is transferred to another information processing apparatus. The present invention can also be executed, and therefore, a recording medium on which the character string recognition method of the present invention is recorded is also included in the present invention.

[0013]

As described above, the present invention has the following excellent effects. According to the first aspect of the present invention, a partial character pattern combining black pixel connection patterns is created, and the height of a circumscribed rectangle of the partial character pattern in the case of horizontal writing is defined as a geometric feature of the created partial character pattern. In the case of vertical writing, the width is calculated, and the average value of the geometric feature of the partial character pattern and the geometric features of a plurality of partial character patterns near the partial character pattern is calculated as the statistical feature. The width of the circumscribed rectangle of the temporary character pattern in the case of horizontal writing and the height in the case of vertical writing are calculated as the geometric feature amount of the temporary character pattern in which the adjacent partial character patterns are temporarily integrated. The evaluation value of the provisional character pattern is calculated from the geometric feature amount of the provisional character pattern and the statistical feature amount of the partial character pattern included in the provisional character pattern. By determining whether to cut out the temporary character pattern as a character pattern based on the value, a plurality of character patterns are cut out to obtain one character string candidate, and further, the above processing is performed by changing the number of neighbors. By repeating, a plurality of character string candidates are obtained, and for each character string candidate, the certainty factor of the entire character string is calculated from the certainty factor of the character recognition result of each character pattern, and the certainty factor of each character string candidate is calculated. Are compared and the most probable character string candidate is determined as a character string recognition result, so that accurate character segmentation can be performed even when the character size is not uniform, and the character recognition result is Reflected in character extraction and recognition processing, more accurate character extraction and character recognition can be realized. According to the second aspect of the invention,
In the described invention, a character pattern in which the character cutout result is the same in all the obtained character string candidates is determined, and further, a partial character string candidate consisting of one or more character patterns that have not been determined is determined. The degree of certainty of the partial character string candidate is obtained from the degree of certainty of the character recognition result of each character pattern, and the certainty degree of each partial character string candidate is compared. Since the recognition result of the entire character string is determined by determining as, the result of the portion that is the same processing result in a plurality of cutout processes is adopted, and therefore, the character string recognition accuracy is further improved.

According to the third aspect of the present invention, the second aspect
In the described invention, a character pattern in which the result of character extraction and character recognition is the same in all of the obtained character string candidates is determined, and further, a partial character string consisting of one or more undetermined character patterns The certainty of the partial character string candidate is obtained from the certainty of the character recognition result of each character pattern for the candidate, and the certainty of each partial character string candidate is compared. Since the recognition result of the entire character string is determined by being determined as the recognition result of the column, multiple cutouts,
In the character recognition processing, the same processing result is used as the result, so that the character string recognition accuracy is further improved as in the second aspect of the present invention. Further, in the invention according to claim 4, the character string recognition method according to claim 1, 2, or 3 is recorded on a recording medium. By transferring and using another information processing apparatus that is not provided, the above information processing apparatus can also perform the character string recognition according to the present invention.

[Brief description of the drawings]

FIG. 1 is an operation flowchart of a character string recognition method according to a first embodiment of the present invention.

2 (A), 2 (B), 2 (C) and 2 (D) show the first embodiment of the present invention.
It is an explanatory view of a character string recognition method showing the embodiment.

FIGS. 3A and 3B are other explanatory diagrams of the character string recognition method according to the first embodiment of the present invention.

FIGS. 4A and 4B are other explanatory diagrams of the character string recognition method according to the first embodiment of the present invention.

FIG. 5 is an operation flowchart of a character string recognition method according to the second embodiment of the present invention.

FIGS. 6A and 6B are explanatory diagrams of a character string recognition method according to a second embodiment of the present invention.

FIG. 7 is a configuration block diagram of a main part of an information processing apparatus according to the present invention.

FIG. 8 is an explanatory diagram of a character string recognition method showing an example of a conventional technique.

[Explanation of symbols]

1: scanner, 2: character extraction unit, 3: character recognition unit
4: System control unit, 5: Operation unit.

Claims (4)

[Claims] 1. A character string recognition method for extracting a plurality of character patterns from a character string pattern of an input character image and performing character recognition of a character string. A partial character pattern is created by combining the extracted and extracted black pixel concatenated patterns, and the height, length, and height of the circumscribed rectangle of the partial character pattern in the case of horizontal writing are used as the geometric features of the created partial character pattern. In the case of writing, the width is calculated, and the average value of the geometric feature amount of the partial character pattern and the geometric feature amounts of a plurality of partial character patterns near the partial character pattern is calculated as a statistical feature amount. In the case of horizontal writing, the width of the circumscribed rectangle of the temporary character pattern and the vertical writing In this case, the height is calculated, and the evaluation value of the provisional character pattern is calculated from the geometric feature amount of the provisional character pattern and the statistical feature amount of the partial character pattern included in the provisional character pattern. By determining whether or not to cut out the temporary character pattern as a character pattern based on the threshold value and a plurality of character patterns, one character string candidate is obtained by cutting out a plurality of character patterns. By repeating the process, a plurality of character string candidates are obtained, and for each character string candidate, the certainty factor of the entire character string is calculated from the certainty factor of the character recognition result of each character pattern. A character string recognition method characterized by determining the most probable character string candidate as a character string recognition result by comparing certainty factors. 2. The character string recognition method according to claim 1, wherein a character pattern in which the character cutout result is the same for all the obtained character string candidates is determined, and one or more undetermined character patterns are determined. For partial character string candidates consisting of character patterns, obtain the partial character string candidate confidence from the confidence of the character recognition result of each character pattern, and compare the confidence of each partial character string candidate to find the most likely partial character A character string recognizing method comprising: determining a column candidate as a recognition result of a partial character string to determine a recognition result of the entire character string. 3. The character string recognition method according to claim 2, wherein a character pattern in which the result of character extraction and character recognition is the same for all of the obtained character string candidates is determined. For a partial character string candidate consisting of one or more character patterns, the degree of certainty of the partial character string candidate is obtained from the certainty degree of the character recognition result of each character pattern,
Character string recognition characterized by determining the recognition result of the entire character string by determining the most probable partial character string candidate as the recognition result of the partial character string by comparing the certainty factors of the respective partial character string candidates. Method. 4. A machine-readable recording for recording a program for causing an information processing apparatus to execute a character string recognition process according to the character string recognition method according to claim 1, 2 or 3. Medium.