JPH06236455A - Character recognizing device - Google Patents

Abstract

PURPOSE:To provide a character recognizing device which can recognize even such characters that are partly defaced by eliminating the influences of defaced characters without requiring the jobs to previously register the defaced character and with high accuracy, high operability and high performance. CONSTITUTION:A defacement detecting part 160 detects the defacement of an input character pattern and then collates the feature matrix obtained from the input character pattern with the prepared feature matrix of standard characters through an identifying part 180. At the same time, the elements of the feature matrix of defaced parts are excluded from the subjects of collation or the influences of these elements are reduced. Thus even such characters that are partly defaced can be accurately recognized with elimination of the influence of defacement.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a character recognition device capable of accurately recognizing a partially crushed character.

[0002]

2. Description of the Related Art Conventionally, as a character recognition device capable of recognizing a character with high accuracy and dealing with a crushed character pattern, Japanese Patent Application No. 01-11796 (Japanese Utility Model Application No. 03-5875).
There was a character recognition device disclosed in 9). According to this character recognition device, the line width is calculated from the input character pattern,
Extracting a plurality of sub-patterns representing input character patterns and line elements in two or more directions from the line width, dividing the sub-pattern character area into a plurality of sub-areas, and expressing the distribution state of character lines in each direction. The feature matrix is extracted, and the extracted feature matrix is collated with a dictionary matrix prepared in advance to identify the candidate character name. Further, when a predetermined specific character name is present in the candidate characters of the identification result, the pre-calculated character pattern circumscribing frame is divided into small areas based on the character line width of each small area. It is determined whether or not the input character pattern is a collapsed character, and based on the determination result, whether or not to output the identification result as a recognition result, that is, whether or not to reject. It was

[0003]

However, in the above-mentioned conventional recognition method, first, since the possibility of being a crushed character is detected from the candidate character name of the identification result, only the characters registered in advance are as described above. It is possible to test whether or not to reject, and therefore the above processing is not performed for characters that are not registered in advance, so the normal recognition processing is performed for the collapsed characters and the recognition result with very low certainty. Therefore, there is a problem that a character recognition device with high recognition accuracy cannot be realized because the number of misread characters increases.

In addition, in order to perform the above-mentioned test for the crushed characters, it is necessary to investigate and register beforehand the characters that are easily crushed and the names of characters that are liable to cause misreading when crushed. It is a high-level analysis work of a level that can be investigated by using a high-performance computer, using special test data using special test data, for example, a character recognition device user, operator, etc. can easily investigate and register. Not a thing. Therefore, it is extremely difficult for each user to effectively use the process of detecting a collapsed character and rejecting the character.

Even if the registration operation is possible, it requires complicated work requiring specialized knowledge, and it is not possible to realize a character recognition device with good operability that can be easily used even by an amateur. There was a problem that it was possible.

According to the present invention, as described above, only the characters registered in advance can be processed, and the registration work for newly processing is complicated and requires specialized knowledge and cannot be easily carried out.
Furthermore, the difference in the character line width between the small areas created by dividing the character circumscribed frame is so small that the collapse cannot be detected effectively. Preliminary registration work by eliminating the problems that the characters cannot be reduced and the recognition cannot be performed accurately and the problems that complicated operation is required and the operability is good, and that no device can be easily used even by an amateur. It is unnecessary to remove the influence of crushing even for characters with a crushed part, and to provide a high-performance character recognition device with high accuracy and high operability that can be recognized accurately. To do.

[0007]

In order to solve the above-mentioned problems, the present invention provides an image input section for obtaining an input character pattern of characters, symbols, figures, etc., a storage section for storing the input character pattern, and the storage section. A character frame dividing unit that divides the character circumscribing frame of the input character pattern into a plurality of small regions; a feature matrix extracting unit that calculates a feature value for each of the divided small regions and obtains a feature matrix; In a character recognition device comprising a discriminator that performs a recognition process by collating a matrix with a characteristic matrix of a standard character pattern prepared in advance, a collapsing detector that detects a collapsing region of an input character pattern is provided, and the discriminator is And performing a matching process by excluding or reducing the influence degree of the feature value of the small area included in the collapsed area detected by the collapse detecting unit when matching the feature matrix. And it features.

[0008]

According to the present invention, when the crush detecting unit detects the crush of the input character pattern and collates the feature matrix obtained from the input character pattern with the feature matrix of the standard character prepared in advance, the crush is detected. The element of the characteristic matrix of the part where the is present is excluded from the object of the matching or the influence of the matrix element is reduced. Therefore, it becomes possible to reduce the influence of the crushing and accurately recognize even the characters with the crushed part,
The above problems are solved.

[0009]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the character recognition device of the present invention will be described below with reference to FIGS. FIG. 1 is a block diagram showing an embodiment of the character recognition device of the present invention. The character recognition device 100 includes an image input unit 110 and a pattern register 12.
0, a character frame division unit 130, a line width calculation unit 140, a sub pattern extraction unit 150, a crush detection unit 160, an identification unit 180, and an output terminal 190. FIG. 2 is a block diagram showing an example of the crush detection unit 160, which includes a line width calculation unit 161.
Contour tracking unit 162, local line width extraction unit 163, determination unit 16
4 and the collapsed area determination unit 165. FIG. 3 is a diagram showing an example of input character patterns. FIG. 4 is a diagram showing an example of sub patterns. 5 is an operation explanatory diagram of the contour tracking unit 162. FIG. 6 is an explanatory diagram of the operation of the collapse detection unit 130. FIG. 7 is a diagram showing an example of the feature matrix. FIG. 8 is a diagram showing an example of a standard character pattern. FIG. 9 is a diagram showing an example of a characteristic matrix of standard character patterns.

The operation of the character recognition device of this embodiment will be described in detail below with reference to FIGS. Image input unit 110
Photoelectrically converts an optical signal S from a form in which characters, figures, symbols, etc. (hereinafter referred to as characters) are described, and for example, the character line part is quantized into black and white binary in which the background part is expressed by black pixels and the background part is expressed by white pixels. A converted electric signal (hereinafter referred to as form image data) is generated, a character pattern for each character (hereinafter referred to as an input character pattern) is cut out, and the cut out input character pattern is output to the pattern register 120.

The character pattern is cut out by the following method. First, the form image data is scanned in a direction parallel to the row direction to create a black pixel distribution, and row data for each row is extracted from the black pixel distribution. Next, the extracted row data is scanned in a direction perpendicular to the row direction to create a black pixel distribution for each row, and 1 is calculated from the black pixel distribution.
Detect the character spacing for each character. Then, a character area for each character is detected by using the created distribution of black pixels in each row and the detected character spacing, and a character pattern for each character is cut out based on the area. In the present embodiment, the case where an input character pattern as shown in FIG. 3 is obtained will be described as an example and will be specifically described below.

The pattern register 120 is provided in the image input section 11
This is a storage unit that sequentially stores input character patterns input from 0 for each character. In this embodiment, 128 × 128.
It has a memory capacity of pixels. This pattern register 1
The input character pattern for each character stored in 10 can be freely referred to by the character frame division unit 130, the line width calculation unit 140, the sub pattern extraction unit 150, and the collapse detection unit 160.

In the character frame dividing unit 130, the circumscribing frame of the input character pattern is extracted, the circumscribing frame is divided into N × M small regions, and the division result is divided into the feature matrix extracting unit 170.
And output to the identification unit 180. In this embodiment, the circumscribing frame of the input character pattern is divided into 8 × 8 small regions formed by dividing the circumscribing frame of the input character pattern into eight parts in the horizontal and vertical directions. In the example of the input character pattern shown in FIG. It is divided into each small area as indicated by the ruled line 301.

In the line width calculation unit 140, the line width of the line elements constituting the input character pattern is calculated from the input character pattern in the pattern register 120, and the line width value of the calculation result is calculated in the sub pattern extraction unit 150. And to the feature matrix extraction unit 170. In this embodiment, the line width is calculated by the number Q of points at which all the points in the 2 × 2 window are black pixels when the input character pattern is scanned by the 2 × 2 window, and the input character pattern. The total number of black pixels A and
This is performed by calculating the average line width W of the input character pattern based on the approximate expression (1). W = A / (AQ) (1) In the input character pattern shown in FIG. 3, A = 290,
Q = 160, line width W = 290 / (290-16
0) = 2.20.

In the sub-pattern extraction unit 150, the input character patterns stored in the pattern register 120 and the average line widths of the input character patterns input from the line width calculation unit 140 are horizontally, vertically, diagonally left, and diagonally right four. The four sub-patterns obtained by extracting the line elements in the direction are extracted and output to the feature matrix extraction unit 170. In the present embodiment, the process of extracting the sub-pattern is performed as follows. For example, in the case of the horizontal sub-pattern, the input character pattern is scanned in the horizontal direction to detect the succession of black pixels, and when the succession number L of black pixels satisfies the expression (2), the succession of the black pixels is considered as a sub-pattern in the horizontal direction. Is to be extracted as. L> 2 × W (2) where W is the average line width input from the line width calculation unit 140. From the input character pattern shown in FIG. 3, horizontal (H), vertical (V), left diagonal (L), and right diagonal (R) sub-patterns are extracted as shown in FIG.

Next, the operation of the crush detector 160 will be described in detail. The line width calculation unit 161 in the crush detection unit 160 calculates the line width of the line elements forming the input character pattern from the input character pattern in the pattern register 120, and determines the line width value of the calculation result. It is output to the unit 164. In the present embodiment, it is assumed that the line width is calculated by the same processing as the line width calculation unit 140.

A contour tracking section 162 in the crush detecting section 160.
5, the contour of the input character pattern is traced, the coordinate value of the contour pixel and the direction number to the next contour pixel as represented by the number in FIG. 5A are extracted, and the coordinate value and the direction number are calculated. It is output to the local line width extraction unit 163 as a contour series.
The contour of the input character pattern is traced by the following method. (1) While scanning the pattern register from the upper left pixel in the Y direction (hereinafter also referred to as the column direction), it sequentially scans in the X direction to detect black pixels, and the detected black pixel is the start black pixel and the first black pixel. Contour pixels. (2) By sequentially focusing on adjacent pixels in eight directions as shown in FIG. 5A centering on the first black pixel in the order from 1 to 8, the black adjacent pixel is detected, and the pixel is set to the second pixel. Contour pixels. (3) This time, the process of detecting the black pixel from the adjacent pixels in the eight directions with the second contour pixel as the center is performed as the third contour pixel. (4) Thereafter, similarly, the contour pixel detection processing is repeated until the detected contour pixel becomes the same pixel as the start contour pixel. (5) When the detected pixel becomes the same as the start contour pixel and the above processing is completed, the scanning of (1) is performed again from the start pixel to detect a black pixel, and the detected black pixel has already been extracted. If it is a pixel outside the outline, the pixel is set as a start black pixel, and the processes from (2) onward are similarly performed. (6) When the above process is performed for all the pattern register areas, the process ends.

According to the method as described above, for example, FIG.
In the upper left portion of the input character pattern of FIG. 3 as shown in FIG. 3B, the black pixel 501 is the start black pixel,
After that, at the stage where the pixel is tested in the direction 1 in FIG. 5A, it is determined that each pixel is a black pixel, and the contour coordinates 502 are detected.
Further, in the case of the lower left portion of the input character pattern as shown in FIG.
Direction 505, 2 direction 506 are respectively tested,
Since each is a white pixel, 507 pixels in three directions are detected as the contour coordinates, and thereafter, the contour coordinates 508 and 509 are similarly detected. As a result, finally, for the input character pattern as shown in FIG. 3, a contour 602 having 601 as a start pixel and a contour 604 having 603 as a start pixel in FIG. 6 are extracted.

Next, in the local line width detecting section 163 in the crush detecting section 160, each pixel of the contour series input from the contour tracking section 162 is used as a reference in the direction perpendicular to the direction of the contour series. The character pattern is scanned, the number of consecutive black pixels is counted, the number of consecutive black pixels is set as the local line width of the line element in the contour series direction in the pixel portion, and the local line width and the contour series are determined by the determination unit 164. Output. The above process will be described in detail with respect to the contour series shown in FIG. 6. For example, since the pixel 605 has a contour in one direction shown in FIG. The input character pattern is scanned in the direction indicated by 606, the number of consecutive black pixels is counted, and the local line width of the pixel is counted.

Next, the judging section 16 in the crush detecting section 160
4, the average line width W of the input character pattern input from the line width calculation unit 161 and the contour series of each contour pixel of the input character pattern input from the local line width extraction unit 163 and the local line width are It is determined whether or not there is a portion that may be crushed in the input character pattern (hereinafter referred to as a crushed candidate area). If there is a crushed candidate area, the coordinate value of the crushed candidate area is used as the crushed area determination unit. Output to 165. The determination of the collapse candidate area is made by determining whether or not an arbitrary part of the contour sequence satisfies Expression (3),
When the contour series satisfies the expression (3), the contour series and an area having a local line width based on each pixel of the contour series are used as the collapse candidate area, and the contour series satisfies the expression (3). If not, it is determined that there is no collapse candidate area in the contour series portion. L ≧ TL (3) where L has the same contour direction in the contour series and the local line width Wi with reference to the contour pixel is expressed by equation (5).
Is a continuous number of contour pixels satisfying the above condition, and TL is a predetermined fixed value or a calculated threshold value, which is calculated in advance by the formula (4) in this embodiment. TL = kL × W (4) where kL is a predetermined fixed value, which is 2 in this embodiment, and W is the above-mentioned average line width. Wi ≧ TW (5) where TW is a threshold value calculated in advance using the average line width W, and is calculated by the equation (6) in this embodiment. TW = kW × W (6) where kW is a predetermined fixed value and is 2 in this embodiment.

In the input character pattern shown in FIG. 3, the average line width W = 2.
Since it is calculated as 20, the threshold values TW and TL are both 4.40 according to the equations (6) and (4). Therefore, the main determination unit 164 detects a contour portion in which the local line width is 5 pixels or more and the contour pixels which are contours in the same direction are continuous for 5 pixels or more, and the portion is crushed as a crush candidate area. This is output to the area determination unit 165. As shown in FIG. 6, in the input character pattern of FIG. 3, the contour series from 607 to 608 and the portion of the local line width based on the contour series, and the contour series from 608 to 609 and the contour series are used as the reference. The portion having the local line width is output to the collapsed area determining unit 165 as a candidate area for the collapsed area.

The crushed area determination unit 165 determines the crushed portion from the crushed candidate area input from the determination unit 163 and outputs the coordinates of the crushed portion to the identification unit 180. The crushed portion is determined by detecting a pixel commonly included in two or more crushed candidate areas and setting the pixel as the crushed portion. In the two collapse candidate areas input from the determination unit 164 shown in FIG. 6, the pixels in the rectangular area surrounded by the pixels 608, 609, 610, and 611 are commonly used in the two collapse candidate areas. Since it is included, the area is output to the identification unit 180 as a collapsed area.

In the feature matrix extraction section 170,
The number of black pixels in each of the divided small areas of the sub-pattern is counted from the horizontal, vertical, left-oblique, and right-oblique sub-patterns input from the sub-pattern extracting unit 150 and the division coordinates input toward the character frame dividing unit 130. Based on the black pixel count result and the average line width, the horizontal, vertical, left diagonal, and right diagonal characteristic matrices are extracted based on Expression (7). KH (m, n) = BH (m, n) / W KV (m, n) = BV (m, n) / W KL (m, n) = BL (m, n) / W (7) KR ( m, n) = BR (m, n) / W However, KH, KV, KL, and KR are horizontal, vertical, left diagonal, and
The right diagonal feature matrix, BH, BV, BL, and BR are horizontal, vertical, left diagonal, and right diagonal black pixel matrices, respectively (m,
n) is the element number of each matrix, and W is the above-mentioned average line width. A feature matrix as shown in FIG. 7 is extracted from the sub-pattern shown in FIG.

In the identifying unit 180, the collapsed portion corresponds to which part of the divided small area based on the coordinates of the collapsed portion input from the collapse detecting unit 160 and the character frame coordinates input from the character frame dividing unit 130. Of the elements of the feature matrix and the feature matrix of standard characters in the dictionary included in the identification unit 180 in advance, only the elements that do not correspond to the small area of the collapsed portion are based on the equation (8). The character code of the standard character having the smallest distance value D of the expression (8) is output from the output terminal 150 as the recognition result of the input character pattern. D = (Σ (gi − ki) ² ) ^1/2 (8) where gi is an element of the standard character matrix, ki is an element of the input character pattern matrix, and i is a standard character.
And an area number of a portion that does not correspond to the inside-blurred portion of the feature matrix of the input character pattern.

The output terminal 190 is a data output terminal for outputting the recognition result to the outside, and is connected to another system, a medium for recording the recognition result, a communication network, another information processing system, or the like. .

The present invention is not limited to the above embodiment. For example, in the identification unit 180, when the collation is performed according to the equation (8), the collation is performed by excluding the element of the region corresponding to the collapsed portion in the present embodiment, but the invention is not limited to this. For example, the same effect can be obtained by replacing the element value of the collapsed portion of the feature matrix of the input character pattern with a predetermined fixed value or the average value of the element values other than the collapsed portion and performing the matching. It is a thing.

Further, the fixed values kW and kL used in the equations (6) and (4) are not limited to the values of this embodiment, and any suitable value may be used within the scope of the present invention. good. Furthermore, the image input method, the pattern register configuration, the average line width calculation method, the contour tracking method, the feature matrix extraction method, the character frame division method, etc. may be appropriately changed within the scope of the present invention. The operation of parts, the method of processing, the flow of input / output signals, the number of arranged, the position, the shape, the number and other conditions can be arbitrarily changed.

[0028]

As described above in detail, according to the present invention, the element value of the feature matrix corresponding to the crushed portion detected by the crush detecting section is excluded or the degree of influence thereof is reduced to perform the identification. Since it is done, there is no need to investigate and register beforehand characters that are easily crushed or characters that are easily crushed and output as misreading results, no complicated operation, no specialized knowledge is required, and even an amateur can easily A character recognition device that can be operated and has good operability becomes possible.

Further, since the local line width is detected for each contour pixel and the crushing is detected for each continuous portion of the contour pixel, an accurately crushed area can be detected and the feature matrix affected by the area. Elements can be reliably excluded from the distance calculation or their influence can be reduced. Therefore, the recognition accuracy is high, and therefore the misreading rate is low, and therefore the recognition result check,
It is possible to realize a highly accurate and high-performance character recognition device that can shorten the time for correction work and greatly improve the efficiency of work.

[Brief description of drawings]

FIG. 1 is a block diagram showing an embodiment of a character recognition device of the present invention.

FIG. 2 is a block diagram showing an embodiment of a crush detection unit.

FIG. 3 is a diagram showing an example of an input character pattern.

FIG. 4 is a diagram showing an example of a sub pattern.

FIG. 5 is an operation explanatory diagram of a contour tracking unit.

FIG. 6 is an operation explanatory diagram of a crush detection unit.

FIG. 7 is a diagram showing an example of a feature matrix.

FIG. 8 is a diagram showing an example of a standard pattern.

FIG. 9 is a diagram showing an example of a characteristic matrix of standard patterns.

[Explanation of symbols]

 100 Character Recognition Device 110 Image Input Unit 120 Pattern Register 130 Character Frame Division Unit 140 Line Width Calculation Unit 150 Sub Pattern Extraction Unit 160 Collapse Detection Unit 170 Feature Matrix Extraction Unit 180 Discrimination Unit 190 Output Terminal

Claims (1)

[Claims] 1. An image input unit for obtaining an input character pattern of characters, symbols, figures, etc., a storage unit for storing the input character pattern, and a plurality of small areas within a character circumscribed frame of the stored input character pattern. A character frame division unit that divides into regions, a feature matrix extraction unit that calculates a feature value for each of the divided small regions and obtains a feature matrix, and the feature matrix is collated with a feature matrix of a standard character pattern prepared in advance. In the character recognition device including an identification unit that performs recognition processing according to the above, a crushing detection unit that detects a crushed region of an input character pattern is provided, and the identification unit is included in the crushed region detected by the crushing detection unit. A character recognition device characterized in that a characteristic value of a small area is subjected to a collation process by excluding or reducing the influence degree when collating the characteristic matrix.