JP3197464B2 - Character recognition method and character recognition device - 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 method for reading a character or a graphic (hereinafter, also referred to as a character graphic) on a document, a form, or the like to be read (these are also referred to as a medium). The present invention relates to a character recognition device suitable for implementation.

[0002]

2. Description of the Related Art In general, a character recognition device extracts a feature from a character pattern, compares the obtained feature with a feature of a standard pattern prepared in advance, and recognizes the character pattern. However, differences in the quality and density of characters on the medium and the effects of noise, etc. may cause variations in the extracted features even for the same characters, which may cause misreading and misreading. is there. In order to prevent this, conventionally, for example, (a). A method of extracting a feature after correcting a character pattern by filter processing (for example, a document: "Image processing")
Co-authored by Yu Tsuchiya and Yoji Fukada, Corona (199
2)), (b). A method in which a plurality of standard patterns for blurred characters, crushed characters, etc. are prepared in advance for the same character, and features extracted from the recognition target character pattern are compared with the plurality of standard patterns, respectively, (c). (D). A method in which an operator extracts in advance a medium containing low-quality characters that cause misreading or unreading from a medium (for example, a form) in which a character graphic is described, and separately inputs a key; For example, a method of outputting a recognition result to an output device such as a display device, and allowing the operator to search and correct misread characters from the output result has been adopted.

[0003]

However, the above-mentioned conventional countermeasures have the following problems.

(1). (A) above that performs filter processing
In the case of the method described above, a filter is determined and used for each reading area.However, when character patterns having extremely different qualities or extremely different densities compared to other character patterns are mixed in the same reading area, A pattern obtained by performing a filtering process on such a character pattern cannot be said to have been appropriately filtered, and tends to have an adverse effect on the accuracy of the subsequent recognition processing.

(2). In the case of the above-mentioned method (b) in which various types of standard patterns are prepared in advance for the same character, there are problems that a large-capacity memory is required and that the time required for collation increases.

(3). In the case of the above method (c) in which the operator extracts a medium (form) including low-quality characters, the operator needs to visually inspect all the forms, and thus the burden on the operator increases as the number of forms to be processed increases. In addition, problems such as oversight occur.

(4). In the case of the above method (d) for searching for and correcting misread characters from the recognition result output to the output device,
As the number of processes increases, the burden on the operator increases.
In addition, problems such as oversight occur.

[0008]

Therefore, in the first invention of this application, in order to solve the above-mentioned problem (1), a character / figure on a medium is photoelectrically converted, and a pattern obtained by the photoelectric conversion is applied to a pattern obtained by the photoelectric conversion. In the character recognition method for extracting a feature from the filtered pattern and comparing the extracted feature with a feature of a standard pattern to identify the character graphic, a plurality of filters having different filter characteristics are provided. Prepare a filter in advance. Before performing the filtering process, the total number of black points A in the pattern obtained by the photoelectric conversion
, And the pattern is scanned using an n × m window to count the total number Q of windows in which all pixels in the window become black points. Based on these A and Q, a character quality evaluation value is calculated. calculate. Next, one of the plurality of filters is selected based on the calculated character quality evaluation value, and the filter processing is performed by the selected filter. However, the plurality of filters may include a filter without a filtering process. Further, n and m are positive integers, including the case where n = m.

According to the first aspect, a filter suitable for the pattern to be recognized is selected from a plurality of filters prepared in advance, and a filter process is performed based on the selected filter. This means that, for example, even in a medium in which blurred characters and crushed characters are mixed, appropriate filtering is performed on each of them. Therefore, for example, in the case of both blurred characters and crushed characters, the difference between the two is reduced when the filtering process is completed, so that the standard pattern used for collation can be the same for each of the blurred characters and the crushed characters. Benefits are obtained. Therefore, since the standard pattern does not increase, it is possible to reduce the dictionary memory capacity and the collation time as compared with the conventional case. When calculating the character quality evaluation value based on Q and A, for example, Q /
With A, the evaluation value can be calculated. When the character quality evaluation value is calculated by Q / A, a character pattern having many character line breaks has a small value, and a character pattern having a smooth character line has a large value. It is possible to obtain a quality evaluation value that accurately represents cutting or blurring (the same applies to the following second to tenth inventions).

Further, in the second invention of this application, as a device suitable for implementing the first invention, a predetermined photoelectric conversion unit, filter unit, feature extraction unit, identification unit, a plurality of filters having different filter characteristics, Claims a character recognition device with a character quality evaluation unit and a filter selection unit. In addition, even if a plurality of filters having different filter characteristics are separately provided,
It may be provided in the filter section, or may be provided in the filter selection section.

In the third invention of this application, the above (2)
In order to solve the problem described above, the character / graphic on the medium is photoelectrically converted, and features are extracted from a pattern obtained by the photoelectric conversion or a pattern obtained by performing a filtering process on the pattern. In the character recognition method for identifying the character graphic by comparing each with a feature, before performing the comparison, counting the total number of black spots A in the pattern obtained by the photoelectric conversion or the pattern obtained by filtering the pattern, and The pattern is scanned using an nxm window, the total number Q of windows in which all the pixels in the window are black points is counted, and a character quality evaluation value is calculated based on A and Q. next,
Limiting the standard pattern to be used from the plurality of prepared standard patterns based on the calculated character quality evaluation value,
The collation is performed using the limited standard pattern. Here, n and m are positive integers, and n = m
Including the case. According to the third aspect, the number of standard patterns used for matching can be limited, so that the matching time can be reduced.

Further, in the fourth invention of this application, as a device suitable for implementing the third invention, a character conversion device comprising a predetermined photoelectric conversion unit, feature extraction unit, identification unit, character quality evaluation unit, and dictionary selection unit. Claim the recognizer.

In the fifth invention of this application, in order to solve the above-mentioned problem (3), a character / figure on a medium is photoelectrically converted and a pattern obtained by the photoelectric conversion or a pattern obtained by filtering the pattern is used. In a character recognition method for extracting a feature and comparing the extracted feature with a feature of a standard pattern to identify an input character / graphic, the number of black spots A in the photoelectrically converted pattern or the pattern obtained by filtering the pattern is counted. , And the pattern is scanned using an n × m window to count the total number Q of windows in which all pixels in the window are black points, and calculate a character quality evaluation value based on these A and Q. Next, the storage destination of the medium is switched based on the calculated character quality evaluation value. However, n and m are positive integers, including the case where n = m. According to this fifth invention,
The storage location of the medium to be recognized can be automatically determined based on the quality of the characters written on each medium.

According to a sixth aspect of the present invention, as a device suitable for implementing the fifth aspect, a character having a predetermined photoelectric conversion unit, feature extraction unit, identification unit, character quality evaluation unit, and stacker switching unit is provided. Claim the recognizer.

In the seventh invention of the present application, in order to solve the above-mentioned problem (4), a character / graphic on a medium is photoelectrically converted and a pattern obtained by the photoelectric conversion or a pattern obtained by performing a filter process on the pattern is used. In a character recognition method for extracting a feature, comparing the extracted feature with a feature of a standard pattern and identifying the character / graphic figure, and displaying the recognition result, a pattern obtained by the photoelectric conversion or a filtering process The total number of sunspots A in the calculated pattern is counted,
Then, the pattern is scanned using an nxm window, and the total number Q of windows in which all pixels in the window become black points is counted.
Then, a character quality evaluation value is calculated based on Q and Q. Next, when the recognition result is displayed, the character quality evaluation value is reflected therein. However, n and m are positive integers, including the case where n = m. ).

According to the seventh aspect, when the recognition result is displayed, the quality of the character itself can be reflected on the display. For this reason, for example, it is possible to display (for example, underlining) reflecting the fact that the character quality is bad in the recognition result of the recognition target character having the bad character quality. Here, misrecognition and the like are more likely to occur in a character with poor character quality than in a character with good quality, so the above-mentioned display that character quality is bad is an effective aid for an operator to search for a misrecognized character.

Further, in the eighth invention of this application, as a device suitable for implementing the seventh invention, a character having a predetermined photoelectric conversion unit, feature extraction unit, identification unit, result display unit and character quality evaluation unit is provided. Claim the recognizer.

Further, in the ninth invention of this application, in order to solve at least two of the problems (1) to (4), a character / graphic on a medium is photoelectrically converted, and Filter the obtained pattern, extract features from the filtered character pattern, collate the extracted features with the features of the standard pattern, identify the input character / graphic, and display the recognition result. In the character recognition method, the filtering process using the character quality evaluation value according to the first invention, the standard pattern limiting process using the character quality evaluation value according to the third invention, the character quality evaluation value according to the fifth invention At least two of the medium storage processing using the display processing and the display processing reflecting the character quality evaluation value according to the seventh invention.
Perform one process.

Further, in the tenth invention of this application, as a device suitable for implementing the ninth invention, a predetermined photoelectric conversion unit, a filter unit, a feature extraction unit, an identification unit and a result display unit are provided. Total number of black spots A in pattern obtained by conversion
, And the pattern is scanned using an n × m window to count the total number Q of windows in which all pixels in the window become black points. Based on these A and Q, a character quality evaluation value is calculated. A character quality evaluation unit to be calculated; and a filter selection unit according to the second invention, a dictionary selection unit according to the fourth invention, a stacker switching unit according to the sixth invention, and a display control unit according to the eighth invention. Claim a character recognition device characterized by the following.

[0020]

Embodiments of the present invention will be described below with reference to the drawings. However, the drawings used in the description are merely schematic representations so that the present invention can be understood. Also, in each of the drawings, the same components are denoted by the same reference numerals, and duplicate description thereof may be omitted.

1. First Embodiment of First and Second Invention First, an embodiment of the first and second invention is shown in FIGS.
This will be described with reference to FIG. Here, FIG. 1 is a block diagram showing one configuration example of the character recognition device of the second invention, and FIG.
FIGS. 3A and 3B are diagrams showing an example of a character pattern processed to explain the processing procedure of the character recognition method of the first invention, FIG. 3 is an explanatory diagram of an example of filter processing, and FIG. FIG. 14 is an explanatory diagram of a result of performing a filter process on the indicated character pattern.

According to the character recognition device of the second invention, the photoelectric conversion unit
1, the first pattern register 13, the filter unit 15, the second pattern register 17, the feature extraction unit 19, the identification unit 21, the character quality evaluation unit 23, and the filter selection unit 25 can be provided.

Here, the photoelectric conversion unit 11 is for optically reading a character or graphic on a medium (not shown) and converting it into an electric signal, and may be constituted by any suitable device such as a known line sensor or area sensor. The first pattern register 13 stores an electric signal from the photoelectric conversion unit 11 as original character pattern information, and can be constituted by a known memory, for example. The filter unit 15 has a plurality of filters having different filter characteristics, and performs a filtering process on the original character pattern by a filter selected by the filter selection unit 25 among the plurality of filters. Things. However, the filters referred to here may include filters without filtering. Therefore, according to the present invention, a total of two filters, ie, a filter without filtering and a filter that performs some filtering, is used.
A filter unit including one filter also corresponds to a filter unit including a plurality of filters. Also, the second
The pattern register 17 stores the pattern filtered by the filter unit 15 (the filter unit 15 without filtering).
There is also a case where the pattern is output from. ) Is stored, and can be constituted by a known memory, for example. The feature extracting unit 19 extracts a feature from a pattern on the second pattern register 17. In this case, the configuration is such that a feature matrix is extracted as a feature. The identification unit 21 identifies (collates) the character name by comparing (comparing) the feature extracted by the feature extraction unit 19 with the feature of the standard pattern.
In addition, the character quality evaluation unit 23 performs the photoelectric conversion on the pattern and the pattern before the filter processing.
The number A of all black spots in the pattern on the pattern register 13 is counted, and the pattern is scanned using an nxm window to count the total number Q of windows in which all the pixels in the window become black points. The character quality evaluation value is calculated based on A and Q. Further, the filter selection unit 25 selects one of the plurality of filters in the filter unit 15 based on the character quality evaluation value calculated by the character quality evaluation unit 23. Next, the procedure of character recognition will be described.

The photoelectric conversion unit 11 optically reads a character / graphic on a medium (for example, a form), further converts it into an electric signal, and sends it to the first pattern register 13 as original character pattern information. The first pattern register 13 stores the original character pattern. The character pattern information on the first pattern register 13 is input to the character quality evaluation unit 23 and the filter unit 15.

The character quality evaluation unit 23 counts the total number of black spots A on the first pattern register 13 and, when looking at the pattern on the first pattern register 13 from an nxm window, displays the pixels in this window. The total number Q of windows in which all are black bits is counted. Specific values of n and m are positive integers, and in this embodiment, n = m = 2. Next, raster scanning, that is, this window is scanned from the upper left of FIG. 2A to the right (horizontal direction) under the condition of stroke 1 (one pixel at a time). Further, scanning is performed in the right direction (horizontal direction) to stroke 1 (each pixel) and to the lower right. Then, the character quality evaluation value D is obtained by the following equation (1).

D = Q / A (1) The character quality evaluation value D can be used to determine the quality of the pattern such as crushing or blurring.

For example, the character pattern to be recognized is shown in FIG.
In the case of a normal pattern (a pattern without blurring or crushing) as shown in (A), Q and A are each, for example, Q = 15.
3, A = 208, and D = 0.74 from the above equation (1). On the other hand, when the character pattern to be recognized is a faint pattern as shown in FIG.
96, A = 176, and D = 0.55 from the above equation (1).
Becomes

Next, based on the character quality evaluation value D, the filter selection unit 25 selects an optimum filter for the character pattern to be recognized from among a plurality of types of filters in the filter unit 15.

For example, in this embodiment, when the character quality evaluation value D does not satisfy the expression (2) (where THL is a predetermined threshold value and THL = 0.6 in this embodiment), the blurring of the pattern is corrected. Is selected, and the filter unit 15 performs the processing based on the selected expansion filter. The processed pattern is stored in the second pattern register 17.
On the other hand, when the expression (2) is satisfied, no filtering is performed (that is, a filter without the filtering process is selected). Therefore, the pattern on the first pattern register 13 is stored in the second pattern register 17 as it is. Note that TH
L can be, for example, a value obtained in advance by preparing a plurality of samples of blur patterns, calculating their character quality evaluation values, and statistically processing them.

D ≧ THL (2) Here, the filter section 15 will be described in detail.

The filter unit 15 includes a first pattern register 13
The pixel value of a certain upper pixel (hereinafter, this pixel is referred to as a center pixel) is determined using the pixel values of pixels around the center pixel (hereinafter, referred to as observation pixels) as observation information.
The pixel value of the central pixel (hereinafter, referred to as central pixel value) is determined using the pixel value of one or a plurality of observed pixels (hereinafter, referred to as observed pixel value), and each pixel of the first pattern register 13 is determined. The image data obtained by changing the pixel value of (i) to the center pixel value is corrected image data (filtered image data).

The center pixel value is a pixel value meaning an information pixel (hereinafter, referred to as an information pixel value).
It is determined by counting how many 1 "are in the central pixel value and the observed pixel value. When the number of information pixel values is equal to or greater than a predetermined comparison number, for example,
The center pixel value is determined to be an information pixel value, for example, “1”. The number of comparisons is set arbitrarily and suitably. Note that the center pixel value may be determined by counting the number of pixels indicating a ground pixel, for example, “0” in a white background document.

FIG. 3 is a more specific explanatory diagram of the filter processing. Filter unit 15, for example, the center pixel a ₀ as shown in FIG. 3 is arranged in the center, the center pixel a ₀ and the observation pixel a _n 3 rows and 3 columns (the n is an integer from 1 to 8 in this embodiment) Observation windows are arranged. In this example, the filter unit 15 sets the comparison number to four and sets the center pixel a ₀
And the observation pixel a _n information pixels in determines that the center pixel value "1" if four or more.

FIG. 4 shows a pattern obtained by filtering the faint pattern shown in FIG. 2B by the filter described with reference to FIG.

Next, the feature extracting unit 19 extracts a feature from the pattern stored in the second pattern register 17 by any suitable method. In this case, the area corresponding to the character frame in the second pattern register 17 is divided into [j × k] areas for each of vertical, horizontal, right diagonal, and left diagonal sub-patterns, and the number of black bits (B _jk ), and using the line width w, calculate the feature indicating the line length by equation (3), and calculate [j × k] × 4
Create a dimensional feature matrix.

L _jk = B _jk / w (3) It should be noted that the line width w is, for example, when the pattern stored in the second pattern register 17 is scanned between 2 × 2, The number M of windows whose points are black and the total number A of black points of the pattern
Can be calculated by the following equation (4).

[0037] w = A / (A-M ) ··· (4) Then, the identification unit 21, also referred to as a standard pattern (f _m) [standard dictionary mask (f _m). ] And the feature matrix (f _i ) extracted by the feature extraction unit 19, a well-known distance L of the following equation (5) is applied, and the category name of the standard character matrix with the smallest L is used as the identification result. Output.

L = [Σ (f _m −f _i ) ² ] ^1/2 (5) According to the first and second inventions, the Q / A used as the character quality evaluation value is calculated as shown in FIG. A small value is used for a character pattern with many irregularities of the character line edge and a large number of broken character lines as shown in the example of FIG. 2B, and a small normal character pattern of the character line as shown in the example of FIG. Has a property of having a large value, so that it is a character quality evaluation value that accurately represents the cut or blurred character. Further, even if patterns having different qualities, such as crushing and blurring, are mixed in the input character graphic, an appropriate filter is selected based on the quality evaluation value D = Q / A (including a filter without filter processing). , The difference between the quality and the density of each pattern can be reduced. Therefore, it is possible to stably extract features, so that it is possible to suppress an increase in the dictionary memory, and it is possible to further reduce the dictionary matching amount. Therefore, a highly accurate and high-speed character recognition method and apparatus can be provided.

In the above embodiment, an example of processing in which only faint characters are considered has been described.
By setting a plurality of thresholds, it is also possible to select a larger number of filters individually or in combination.

In the above embodiment, a 3 × 3 window is used as a filter. However, the present invention is not limited to this, and a filter of any method may be used.

2. Third and Fourth Embodiment Next, a third and fourth embodiment of the invention will be described with reference to FIGS.
This will be described with reference to FIG. Here, FIG. 5 is a block diagram showing one configuration example of the character recognition device of the fourth invention, and FIG.
FIGS. 7A and 7B are diagrams showing an example of a character pattern processed to explain the processing procedure of the character recognition method according to the third invention; FIGS.
FIG. 7A is a diagram illustrating a result of performing a filtering process on the character pattern illustrated in FIG. 6A, and FIG. 7B is a diagram illustrating a result of performing a filtering process on the character pattern illustrated in FIG. 6B. FIG.

According to the character recognition device of the fourth invention, the photoelectric conversion unit 1
1, a first pattern register 13, a filter unit 31, a second pattern register 17, a feature extraction unit 19, an identification unit 21, a character quality evaluation unit 23, and a dictionary selection unit 33 can be provided. Here, the photoelectric conversion unit 11, the first pattern register 13, the second pattern register 17, the feature extraction unit 19,
As the identification unit 21 and the character quality evaluation unit 23, those described in the character recognition device of the second invention can be used. Here, the filter unit 31 is a filter unit that performs a specific filter process. On the other hand, the dictionary selection unit 33 is unique to the third and fourth inventions, and limits the standard pattern to be used from the plurality of prepared standard patterns based on the character quality evaluation value calculated by the character quality evaluation unit 23. And a dictionary selection unit for outputting to the identification unit for the collation.
A plurality of prepared standard patterns may be separately prepared as a standard dictionary (not shown), or may be stored in the identification unit 21, or may be stored in the dictionary selection unit 33. It may be optional, depending on the design. Here, the identification unit 2
1 shows an example in which many types of dictionaries are stored. Next, the procedure of character recognition will be described.

As described in the first and second embodiments of the present invention, the character graphic on the medium is optically read by the photoelectric conversion unit 11, further converted into an electric signal, and converted into the original character pattern information. It is stored in one pattern register 13. Next, a filtering process by the filter unit 31 is performed on the original character pattern. The pattern after the filter processing is stored in the second pattern register 17. Next, A and Q are counted for the pattern in the second pattern register in this case to calculate the character quality evaluation value. When the pattern to be counted is a pattern of a certain appropriate density including noise, for example, as shown in FIG.
Q and A are, for example, Q = 149 and A = 204, and the character quality evaluation value is D = 0.73 from the above equation (1). On the other hand, when the pattern to be counted is, for example, a faint character pattern as shown in FIG. 6B, Q and A are, for example, Q = 112 and A = 174, respectively.
= 0.64.

Next, based on the character quality evaluation value D, the dictionary selecting section 33 selects an optimum one for the character pattern to be recognized from among a plurality of types of dictionaries in the identification section 21.

For example, in this embodiment, when the character quality evaluation value D does not satisfy the equation (6), a standard pattern (also referred to as a standard dictionary mask) corresponding to a blurred pattern is selected, and a collation operation using this is performed later. Is made. Note that THL is a predetermined threshold value, and in this embodiment, THL is set to 0.7. On the other hand, when Expression (6) is satisfied, a standard pattern (standard dictionary mask) corresponding to the normal pattern is selected, and a collation operation using this is performed later.

D ≧ THL (6) Next, the discriminating unit 21 sets the distance between the selected standard dictionary mask (f _m ) and the feature matrix (f _i ) extracted by the feature extracting unit 19. By applying the well-known distance L in the above equation (5), the category name of the standard character matrix in which L is the minimum is output as the identification result.

According to the third and fourth aspects of the present invention, the Q / A used as the character quality evaluation value is a character with many irregularities in the character line edge and characters with a lot of character line breaks as shown in the example of FIG. Characters that take small values for patterns and large values for normal character patterns with smooth character lines as shown in the example of FIG. It becomes the quality evaluation value. Furthermore,
Even if patterns having different qualities such as crushing and blurring are mixed in the input pattern, an appropriate standard pattern (standard dictionary mask) is selected based on the quality evaluation value D = Q / A, and feature matching is performed. The collation operation can be performed in a state where the masks used for are limited. Therefore, the collation time can be reduced. Therefore, a highly accurate and high-speed character recognition method and apparatus can be provided.

In the above-described third and fourth embodiments, the character quality evaluation value is calculated from the pattern after the filter processing, but is stored in the pattern before the filter processing, that is, in the first pattern register. The character quality evaluation value may be calculated from the pattern. Further, in the above description, an example in which the standard pattern is selected from two types, one corresponding to faint characters and the other corresponding to normal characters, has been described. May be selected.

3. Fifth and Sixth Embodiments Next, fifth and sixth embodiments will be described with reference to FIG. Here, FIG. 8 is a block diagram showing one configuration example of the character recognition device of the sixth invention.

According to the character recognition device of the sixth invention, the photoelectric conversion unit 1
1, first pattern register 13, filter unit 31, second pattern register 17, feature extraction unit 19, identification unit 21, character quality evaluation unit 23, stacker switching unit 41, and stacker 4
3 can be provided. Here, the photoelectric conversion unit 1
1, first pattern register 13, second pattern register 1
7. Feature extraction unit 19, identification unit 21, character quality evaluation unit 2
3. The filter unit 31 described in the second or fourth aspect of the present invention can be used. On the other hand, the stacker switching unit 41 is unique to the fifth invention and the sixth invention, and switches any one of the storage destinations of the media 45a to 45c based on the calculated character quality evaluation value.
The storage destination is a plurality of individual stackers constituting the stacker 43. The individual stacker is, for example, a stacker for faint character forms that stores many portions written in faint characters, a normal character form stacker for storing forms that include many portions written in normal characters, and the like. Next, the operation of the character recognition device of the sixth invention will be described. This description will be made with reference to FIGS.

As described in the first and second embodiments, the character pattern on the medium is optically read by the photoelectric conversion unit 11 and further converted into an electric signal.
Then, the first pattern register 13 is used as the original character pattern information.
Is stored in Next, a filtering process by the filter unit 31 is performed on the original character pattern. The pattern after the filter processing is stored in the second pattern register 17.

Next, the pattern stored in the second pattern register 17 is stored in the character quality evaluation section 23 and the feature extraction section 19.
Respectively. In the character quality evaluation unit 23, the character quality evaluation value D = Q / A, and in the feature extraction unit 19, the characteristic matrix L _ik = B _ik / W is used in the first and second embodiments of the present invention. It is determined by the same procedure as described. Here, when the pattern for which Q and A are counted is a normal pattern as shown in FIG. 2A and is a pattern processed by the filter described with reference to FIG. For example, Q = 15
3, A = 208, and the character quality evaluation value D is D = 0.74 from the above equation (1). On the other hand, when the pattern to be counted is a faint pattern as shown in FIG. 2B and processed by the filter described with reference to FIG. 3, Q and A are, for example, Q = 112, respectively. , A
= 174, and the character quality evaluation value D is calculated by the above equation (1).
Thus, D = 0.64.

Next, based on the character quality evaluation value D, the stacker switching section 41 determines whether the character or graphic to be identified is blurred or normal. If the character graphic is faint, for example, the medium (form) in which the character graphic exists is stored in the stacker 43.
The stacker 43 is controlled so as to be sent to the faint character form stacker in. In this embodiment, the character quality evaluation value D
Does not satisfy the following expression (7), the form is sent to the faint character form stacker. Here, THL is a predetermined threshold, and in this embodiment, THL is set to 0.7. On the other hand, when the formula (7) is satisfied, the form is sent to a normal character form stacker. In determining the storage location of the form, how many characters on the form to refer to the character quality evaluation value is determined, for example, statistically.

D ≧ THL (7) Since the operation of the identification unit 21 can be performed in the same manner as in the first invention, the description thereof is omitted. In the fifth invention and the sixth invention, for example, the identification operation may not be performed on a form including low-quality characters.

According to the fifth and sixth aspects of the present invention, the Q / A used as the character quality evaluation value is a character with many irregularities in the character line edge and characters with a lot of character line breaks as shown in the example of FIG. Characters that take small values for patterns and large values for normal character patterns with smooth character lines as shown in the example of FIG. It becomes the quality evaluation value. Furthermore,
Forms having low-quality characters such as crushing and blurring can be classified into stackers different from other forms based on the quality evaluation value D = Q / A. Therefore, the operator can automatically perform the operation of extracting the medium (form) including the low-quality characters, so that the burden on the operator can be reduced and the extraction error can be prevented.

In the fifth and sixth embodiments of the present invention, the character quality evaluation value is calculated from the pattern after the filter processing, but is stored in the pattern before the filter processing, that is, in the first pattern register. The character quality evaluation value may be calculated from the pattern. In the above description, an example in which a blurred character form and a normal character form are selected has been described.However, a plurality of thresholds for character quality evaluation values are prepared, and in addition to the blurred character form, various forms such as forms having a lot of collapsed characters are prepared. It is also possible to switch the storage destination of a character quality form.

4. Next, an embodiment of the seventh and eighth inventions will be described with reference to FIG. Here, FIG. 9 is a block diagram showing one configuration example of the character recognition device of the eighth invention.

According to the character recognition device of the eighth invention, the photoelectric conversion unit 1
1, first pattern register 13, filter unit 31, second pattern register 17, feature extraction unit 19, identification unit 21, character quality evaluation unit 23, result display unit 51, and display control unit 53
Can be provided. Here, the photoelectric conversion unit 1
1, first pattern register 13, second pattern register 1
7. Feature extraction unit 19, identification unit 21, character quality evaluation unit 2
3. The filter unit 31 described in the second or fourth aspect of the present invention can be used. The result display unit 51 can be constituted by a display medium such as a CRT, LCD, PDP, or the like, which outputs a character name in various character recognition devices, a printer, or the like. On the other hand, the display control unit 53 is unique to the seventh invention and the eighth invention, and causes the result display unit 51 to perform a display reflecting the character quality evaluation value when displaying the recognition result on the result display unit 51. is there. The display reflecting the character quality evaluation value can be any suitable display. For example, a character portion with poor character quality is underlined, a different mark is added for each character quality (eg, faint character, crushed character, etc.), a display character color is changed, or blinking is performed. Various indications can be made, for example, by notifying the operator to that effect.

Next, the operation of the character recognition device according to the eighth invention will be described. This description will be made with reference to FIG. 2, FIG. 3 and FIG.

As described in the first and second embodiments of the present invention, the character pattern on the medium is optically read by the photoelectric conversion unit 11 and further converted into an electric signal.
Then, the first pattern register 13 is used as the original character pattern information.
Is stored in Next, a filtering process by the filter unit 31 is performed on the original character pattern. The pattern after the filter processing is stored in the second pattern register 17.

The pattern stored in the second pattern register 17 is then stored in the character quality evaluation section 23 and the feature extraction section 19.
Respectively. In the character quality evaluation unit 23, the character quality evaluation value D = Q / A, and in the feature extraction unit 19, the characteristic matrix L _ik = B _ik / W is used in the first and second embodiments of the present invention. It is determined by the same procedure as described. Here, when the pattern for which Q and A are counted is a normal pattern as shown in FIG. 2A and is a pattern processed by the filter described with reference to FIG. For example, Q = 15
3, A = 208, and the character quality evaluation value D is D = 0.74 from the above equation (1). On the other hand, when the pattern to be counted is a faint pattern as shown in FIG. 2B and processed by the filter described with reference to FIG. 3, Q and A are, for example, Q = 112, respectively. , A
= 174, and the character quality evaluation value D is calculated by the above equation (1).
Thus, D = 0.64.

Next, the identification unit 21 performs a feature collation process between the pattern to be recognized and the standard pattern, a calculation of the distance L, and the like according to the procedure described in the first embodiment of the present invention and the like. The category name of the character matrix is output to the result display unit 51. At this time, the display control unit 53
Based on the character quality evaluation value calculated for the pattern to be recognized, the result display unit 51 is controlled separately from the identification unit 21 as described below. That is, when the character quality evaluation value D does not satisfy the following expression (8), the pattern to be recognized is determined to be a faint character, and the recognition result displayed on the result display unit 51 is underlined. Here, THL is a predetermined threshold, and in this embodiment, THL is set to 0.7. On the other hand, when Expression (8) is satisfied, the recognition target pattern is determined to be a normal character. The character recognition result displayed on the result display unit 51 is not particularly modified. FIG. 10 shows a display example in the result display unit 51 when the character figure to be recognized on the form is “Saturday is also closed because it is a two-day holiday system” and the character “Saturday” is a faint character. It was shown to. It can be seen that an underline is added to the character portion of “soil”.

D ≧ THL (8) According to the seventh and eighth aspects of the present invention, the Q / A used as the character quality evaluation value is determined based on the character line edge as shown in the example of FIG. A character pattern having many irregularities and broken character lines has a small value, and a character pattern having a smooth character line has a large value as shown in the example of FIG. This is a character quality evaluation value that accurately represents cutting or blurring. Furthermore, even if the input character graphic contains a mixture of low-quality objects such as
This is detected by the quality evaluation value D = Q / A, and reflects that the quality is low on the display of the result display unit. For this reason, the operator can quickly search for low-quality characters that are likely to be erroneously read or unreadable, and it is difficult for the operator to overlook the characters, so that correction work can be reduced, and a character recognition device with excellent operability can be provided.

In the above-described seventh and eighth embodiments of the present invention, the character quality evaluation value is calculated from the pattern after the filter processing, but is stored in the pattern before the filter processing, that is, in the first pattern register. The character quality evaluation value may be calculated from the pattern. Also, in the above description, an example in which a blurred character and a normal character are determined has been described, but a plurality of thresholds for character quality evaluation values are prepared,
The quality of other characters such as blurred characters as well as crushed characters may be determined and displayed to that effect.

5. Embodiments of Ninth and Tenth Inventions In the above description, this application is to solve (1)-
Although the embodiments of each invention for individually solving the problem (4) have been described, the invention in which each invention is combined to solve two or more of the problems (1) to (4). Of course.

[0066]

As is apparent from the above description, according to the first and second aspects of the present invention, the character quality evaluation value of the character pattern to be recognized is determined as predetermined, and the evaluation is performed on the character pattern. A filter suitable for filter processing is selected from a plurality of filters prepared in advance based on the character quality evaluation value. For this reason, the character pattern after the filter processing becomes stable, so that the number of prepared standard patterns can be reduced. Therefore, the memory capacity for storing the standard pattern can be reduced, and the time required for collating the standard pattern can be reduced.

According to the third and fourth aspects of the present invention, the character quality evaluation value of the character pattern to be recognized is determined as predetermined, and a plurality of standard patterns suitable for the character pattern are prepared in advance. It is selected from the standard patterns based on the character quality evaluation value. For this reason, since the standard pattern used for collation can be limited, collation time can be shortened.

According to the fifth and sixth aspects of the present invention, the character quality evaluation value of the character pattern to be recognized is determined as predetermined, and whether the medium from which the character pattern is read has low-quality characters. If it is likely to be determined based on the character quality evaluation value, the medium can be removed. Since a medium having low-quality characters that are easily misread or unreadable can be automatically extracted, the burden on the operator can be reduced.

According to the seventh and eighth aspects of the present invention, the character quality evaluation value of the character pattern to be recognized is determined as predetermined, and if the character that is the source of the character pattern is of low quality, this fact is indicated. Can also be displayed in the recognition result. For this reason, the operator can easily search for characters that are likely to be erroneously read, so that the burden on the operator can be reduced.

Further, according to the ninth and tenth inventions of this application, it is possible to provide a character recognition method and a character recognition device having at least two of the effects obtained in the first to eighth inventions.

[Brief description of the drawings]

FIG. 1 is an explanatory diagram of an embodiment of a first invention and a second invention, particularly an explanatory diagram of one configuration example of a character recognition device of a second invention.

FIGS. 2A and 2B are explanatory diagrams mainly illustrating an embodiment of the first invention and the second invention, in particular, FIG. 2A is a diagram illustrating an example of a normal pattern, and FIG. 2B is a diagram illustrating an example of a faint pattern; is there.

FIG. 3 is an explanatory diagram mainly illustrating an embodiment of the first invention and the second invention, and particularly an explanatory diagram of a filter of a filter unit.

FIG. 4 is an explanatory diagram mainly illustrating an embodiment of the first invention and the second invention, particularly a diagram illustrating a pattern example after a filtering process;

FIG. 5 is an explanatory diagram of an embodiment of the third invention and the fourth invention, particularly an explanatory diagram of a configuration example of the character recognition device of the fourth invention.

FIGS. 6A and 6B are explanatory views of an embodiment of the third invention and the fourth invention, in particular, FIG. 6A shows an example of a pattern having an appropriate density with noise, and FIG. 6B shows an example of a blur pattern; FIG.

FIG. 7 is an explanatory diagram of the embodiment of the third invention and the fourth invention, particularly showing an example of a pattern after filtering processing.

FIG. 8 is an explanatory diagram of an embodiment of the fifth invention and the sixth invention, particularly an explanatory diagram of a configuration example of the character recognition device of the sixth invention.

FIG. 9 is an explanatory diagram of an embodiment of the seventh invention and the eighth invention, particularly an explanatory diagram of a configuration example of the character recognition device of the eighth invention.

FIG. 10 is an explanatory diagram of an embodiment of the seventh invention and the eighth invention, particularly an explanatory diagram of a display example.

[Explanation of symbols]

 11: photoelectric conversion unit 13: first pattern register 15, 31: filter unit 17: second pattern register 19: feature extraction unit 21: identification unit 23: character quality evaluation unit 25: filter selection unit 33: dictionary selection unit 41: Stacker switching unit 43: Stackers 45a to 45c: Medium (form) 51: Result display unit 53: Display control unit

──────────────────────────────────────────────────続 き Continuation of the front page (56) References JP-A-62-198991 (JP, A) JP-A-2-191085 (JP, A) JP-A-56-65278 (JP, A) JP-A-6-1989 301817 (JP, A) JP-A-5-120488 (JP, A) JP-A-4-38585 (JP, A) (58) Fields investigated (Int. Cl. ⁷ , DB name) G06K 9/00-9 / 82

Claims (12)

(57) [Claims] 1. A character / graphic on a medium is photoelectrically converted, a pattern obtained by the photoelectric conversion is subjected to a filtering process, a feature is extracted from the filtered pattern, and the extracted feature is defined as a feature of a standard pattern. A plurality of filters having different filter characteristics are prepared in advance, and before the filtering process, the total number of black spots A in the pattern obtained by the photoelectric conversion is determined. And the pattern is defined as n
× m, and the total number Q of windows in which all pixels in the window become black points is counted, and a character quality evaluation value is calculated based on A and Q. A character recognition method characterized in that any one of the plurality of filters is selected based on the selected filter, and the filter processing is performed by the selected filter (however, the plurality of filters include a filter having no filter processing). In addition, n and m are positive integers, including the case where n = m.) 2. A photoelectric conversion unit for photoelectrically converting a character / graphic on a medium, a filter unit for performing a filter process on a pattern obtained by the photoelectric conversion unit, and a feature extracted from the pattern processed by the filter unit. A character recognition device comprising: a feature extraction unit; and an identification unit that identifies the character graphic by comparing the extracted feature with a feature of a standard pattern, wherein a plurality of filters having different filter characteristics are provided. The total number of black spots A in the obtained pattern is counted,
Then, the pattern is scanned using an nxm window, and the total number Q of windows in which all pixels in the window become black points is counted.
A character quality evaluation unit that calculates a character quality evaluation value based on the character quality evaluation value, and a filter selection unit that selects one of the plurality of filters based on the character quality evaluation value calculated by the character quality evaluation unit. And a character recognition device (however, the plurality of filters may include a filter without a filtering process.
n and m are positive integers, including the case where n = m. ). 3. A character / figure on a medium is photoelectrically converted, a feature is extracted from a pattern obtained by the photoelectric conversion or a pattern obtained by performing a filtering process on the pattern, and the extracted features are respectively compared with features of a prepared standard pattern. In the character recognition method for collating and identifying the character graphic, before performing the collation, counting the total number of black spots A in the pattern obtained by the photoelectric conversion or the pattern obtained by filtering the pattern, and Scanning using an nxm window, counting the total number Q of windows in which all the pixels in the window are black points, calculating a character quality evaluation value based on these A and Q, Limit the standard pattern to be used from among the plurality of prepared standard patterns based on the value,
A character recognition method characterized in that the matching is performed using the limited standard pattern (however, n and m are positive integers, including the case where n = m). 4. A photoelectric conversion unit for photoelectrically converting a character / graphic on a medium, a feature extraction unit for extracting a feature from a pattern obtained by the photoelectric conversion unit or a pattern obtained by filtering the pattern, and a plurality of the extracted features. In a character recognition device comprising an identification unit for identifying the character graphic by comparing with a feature of a prepared standard pattern, a total number of black spots A in a pattern obtained by photoelectric conversion or a pattern obtained by filtering the pattern is counted. Then, the pattern is scanned using an nxm window to count the total number Q of windows in which all pixels in the window become black points, and calculate a character quality evaluation value based on these A and Q. A character recognition unit, further comprising: a character quality evaluation unit; and a dictionary selection unit for limiting a standard pattern to be used from the plurality of prepared standard patterns based on the calculated character evaluation value. And, n, m is a positive integer, including the case of n = m.). 5. A character / graphic on a medium is subjected to photoelectric conversion, a feature is extracted from a pattern obtained by the photoelectric conversion or a pattern obtained by performing a filtering process on the pattern, and the extracted feature is compared with a feature of a standard pattern. In a character recognition method for identifying a figure, the number of all black spots A in a pattern obtained by the photoelectric conversion or a pattern obtained by filtering the pattern is counted, and the pattern is scanned using an nxm window. Counting the total number Q of windows in which all pixels in the window become black points, calculating a character quality evaluation value based on these A and Q, and switching the storage location of the medium based on the calculated character quality evaluation value. (Where n and m are positive integers, including the case where n = m). 6. A photoelectric conversion unit for photoelectrically converting a character / graphic on a medium, a feature extraction unit for extracting a characteristic from a pattern obtained by the photoelectric conversion unit or a pattern obtained by filtering the pattern, and a standard A character recognition device comprising: an identification unit that identifies the character graphic by comparing with a feature of the pattern; counting the total number of black spots A in the pattern obtained by the photoelectric conversion unit or the pattern processed by filtering the pattern; and The pattern is scanned using an nxm window, the total number Q of windows in which all pixels in the window become black points is counted, and a character quality evaluation value is calculated based on A and Q. And a stacker switching unit that switches the storage destination of the medium based on the calculated character quality evaluation value (where n and m are positive integers and n = M No.). 7. A character / figure on a medium is photoelectrically converted, a feature is extracted from a pattern obtained by the photoelectric conversion or a pattern obtained by filtering the pattern, and the extracted feature is compared with a feature of a standard pattern. In a character recognition method for identifying a figure and displaying the recognition result, the number of all black spots A in the photoelectrically converted pattern or the pattern obtained by filtering the pattern is counted, and the pattern is determined using an nxm window. Counting the total number Q of windows in which all pixels in the window become black points, calculate a character quality evaluation value based on these A and Q, and display the character quality when displaying the recognition result. Character recognition method characterized by reflecting the evaluation value (however,
n and m are positive integers, including the case where n = m. ). 8. A photoelectric conversion unit for photoelectrically converting a character / graphic on a medium, a feature extraction unit for extracting a feature from a pattern obtained by the photoelectric conversion unit or a pattern obtained by filtering the pattern, and a standard A character recognition device comprising: an identification unit that identifies the character / graphic figure by comparing it with a pattern feature; and a result display unit that displays a result of the identification. A pattern obtained by photoelectric conversion or a pattern obtained by filtering the pattern , The pattern is scanned using an nxm window, and the total number Q of windows in which all pixels in the window are black points is counted. A character quality evaluation unit that calculates a character quality evaluation value, and a display control unit that causes the result display unit to perform display reflecting the character quality evaluation value when displaying the recognition result. Characterized by Character recognition apparatus (although, n, m is a positive integer, including the case of n = m.). 9. A character pattern on a medium is photoelectrically converted, a filter process is performed on the pattern obtained by the photoelectric conversion, a feature is extracted from the filtered character pattern, and the extracted feature is a feature of a standard pattern. 4. A character recognition method for identifying the character graphic by comparing with a character graphic and displaying a result of the recognition, wherein: a filtering process using the character quality evaluation value according to claim 1; and a character quality evaluation according to claim 3. A standard pattern limiting process using values, a medium storage destination process using the character quality evaluation value according to claim 5, and a display process reflecting the character quality evaluation value according to claim 7. And a character recognition method. 10. A photoelectric conversion unit for photoelectrically converting a character / graphic on a medium, a filter unit for performing a filtering process on a pattern obtained by the photoelectric conversion unit, and a feature for extracting a feature from the filtered pattern. A character recognition device comprising: an extraction unit; an identification unit that identifies the character graphic by comparing the extracted feature with a feature of a standard pattern; and a result display unit that displays a result of the recognition. The number A of all black spots in the pattern is counted, and the pattern is scanned using an nxm window to count the total number Q of windows in which all the pixels in the window become black points. 7. A filter selection unit according to claim 2, further comprising a character quality evaluation unit that calculates a character quality evaluation value based on the reference value, a dictionary selection unit according to claim 4, and a stacker switching unit according to claim 6. And, and according to claim 8. Character recognition device, characterized in that it comprises a Display controller. 11. A character / figure on a medium is photoelectrically converted, a filter process is performed on the pattern obtained by the photoelectric conversion, a feature is extracted from the filtered character pattern, and a plurality of the extracted features are prepared. 4. A character recognition method for identifying the character graphic by comparing the character pattern with a characteristic of a standard pattern, wherein: a filtering process using the character quality evaluation value according to claim 1; and a character quality evaluation value according to claim 3. A character recognition method (where n and m are positive integers and n = m
Including the case. ). 12. A photoelectric conversion unit for photoelectrically converting a character / graphic on a medium, a filter unit for filtering a pattern obtained by the photoelectric conversion unit, and extracting a feature from the pattern processed by the filter unit. 3. A character recognition device comprising: a feature extraction unit; and an identification unit that identifies the character graphic by comparing the extracted features with features of a plurality of prepared standard patterns, wherein the filter selection unit according to claim 2. And a dictionary selection unit according to claim 4 (where n and m are positive integers, including the case where n = m).