JP4594952B2 - Character recognition device and character recognition method - Google Patents

Description

  The present invention relates to a character recognition device and a character recognition method for recognizing characters from an image of an identification card such as a license.

  When a mobile phone sales agent, etc., makes a mobile phone usage contract, a copy of an identification card such as a license or insurance card is pasted on a part of the application form written by the user. Fax to a contract center.

  At the contract center, the application form received by FAX is digitized into a monochrome image by an optical character reader (hereinafter referred to as OCR) such as a scanner and displayed on a PC or the like. After confirming the contents of the ID card, the user (contractor) data is registered and managed in the database.

  At this time, the text data of the character recognition result obtained by recognizing the characters output by FAX from the FAX, that is, by printing the image of the printed monochrome FAX document by OCR and recognizing the characters, is registered in the database. .

  By the way, in the predetermined identification card pasting column in the application form, the identification card is not necessarily copied in full size and pasted in the normal direction.

  In the application form image where only the part of the ID card in the application form has been copied at an arbitrary magnification or has not been pasted in the correct direction, the image in the ID range is correctly cut out and the correct character recognition result is obtained. Therefore, the operator registers the contents of the identification card by directly inputting the characters that could not be recognized to the PC.

  As a technique for recognizing an identification card, for example, a technique is known in which a license is read by a dedicated scanner, a ruled line is detected from the read license image, and the license letter is recognized based on the ruled line.

Further, as a technique for determining the orientation of an image and reading a character, a reference mark is printed on a predetermined portion of three sheet angles of four sheet angles of a transmission target form, and a form obtained through a facsimile machine is used. There is already a technique for detecting and correcting the orientation of a form image by detecting a reference mark from the image of the image and recognizing characters from the corrected image (see, for example, Patent Document 1).
JP-A-2-12479

  In the case of the above-mentioned prior art, the position of the ID can be detected by scanning the entire image of the form and determining the position of the character. However, characters with a diagonal line in the background are printed in the license number field such as a license. Therefore, there is a problem in that a monochrome image obtained by scanning a monochrome FAX or its FAX paper with a scanner has a high probability that an error will occur in this portion of character recognition.

  The present invention has been made to solve such a problem, and an object of the present invention is to provide a character recognition apparatus and a character recognition method capable of improving the character recognition rate of a character image with a diagonal line as a background.

In order to solve the above-described problem, the character recognition device of the present invention is acquired by an image information acquisition unit that acquires an image from a printing surface on which a character string with a diagonal line as a background is printed, and the image information acquisition unit. An image of a character string with the background of the oblique line is extracted from the image, and the character image cutout means for cutting out each character unit, and the character image cut out by the character image cutout means, the oblique line is substantially horizontal. A character image rotating means that rotates to an angle of, a ray processing means for removing a black pixel component constituting a line in the horizontal direction from the character image rotated by the character image rotating means, and a horizontal direction by the ray processing means. and feature vector extraction means for black pixel component constituting the line to extract a feature vector of the character image that has been removed, the reference character image being rotated at an angle to match the pre said shaded angle JP Matching compares the dictionary storage unit which stores a dictionary associating the vector and text data, and a feature vector stored in the dictionary storage unit and the feature vectors of the extracted the character image by the feature vector extraction means Further, the image processing apparatus includes a character recognition unit that outputs text data having an approximate feature vector.

According to the character recognition method of the present invention, an image information acquisition unit acquires an image from a printing surface on which a character string with a diagonal line as a background is printed, and a character image is selected from the images acquired by the image information acquisition unit. A cutting means extracts a character string image with the diagonal line as a background and cuts out each character unit; a character image cut out by the character image cutting means; a character image rotating means; A step of rotating to an angle that is substantially horizontal, a step of removing black pixel components constituting a line in a horizontal direction from the character image rotated by the character image rotating unit, and a horizontal direction of the ray processing unit a step of a feature vector of the character images black pixel component is removed feature vector extraction means for extracting constituting a line, is rotated at an angle that matches the previously said shaded angle Stores the state of the reference character image feature vector and the dictionary associating the text data in the dictionary storage unit, a feature vector of the character images extracted by the feature vector extraction means, stored in the dictionary storage unit And a step of outputting text data having a feature vector that the character recognition means compares and matches or approximates with the feature vector thus obtained.

  As described above, according to the present invention, the character recognition rate of a character image with a diagonal line as a background can be improved.

Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.
FIG. 1 is a diagram showing the configuration of an identification card recognition system according to one embodiment of the present invention, and FIG. 2 is a diagram showing an example of an application form.
As shown in FIG. 1, the ID recognition system includes an image scanner 2 (hereinafter referred to as “image scanner”) that acquires (generates) image information (image data) by scanning the surface of an application form 1 as a form with a CCD or the like. And a computer 10 connected to the scanner 2 and performing image extraction processing and character recognition processing of the license 23 on the image information (image data) read from the application form 1. Yes.

  The scanner 2 acquires a monochrome image (image data) by scanning the surface of the application form 1 with a CCD sensor. It should be noted that the same applies when monochrome FAX paper sent by FAX is scanned in color. That is, the scanner 2 functions as an image information acquisition unit that acquires monochrome image information from the application form 1.

  The computer 10 includes an operation unit 11, a communication I / F 12, a memory 13 as a storage unit, a display unit 14, a hard disk device 15, and a CPU 16. The operation unit 11 is an input unit that is operated by a user such as a keyboard or a mouse.

  The memory 13 includes a standard dictionary 13a (hereinafter referred to as “first dictionary 13a”) in which a standard character image for recognizing characters (also referred to as a registered pattern) or its feature vector and text data is associated, and a license. Is stored as a hatched character-only dictionary 13b (hereinafter referred to as "second dictionary 13b") for recognizing a predetermined character (characters with a hatched portion of the license number portion M).

  That is, the memory 13 functions as a dictionary storage unit that stores a dictionary (first dictionary 13b) in which the feature vector of the reference character image in a state rotated (tilted) in advance at a predetermined angle and text data are associated with each other. The predetermined angle is an angle that coincides with the angle of the oblique line of the character image to be recognized including the oblique line.

The memory 13 also stores a license format 13c for detecting the position and inclination of the license image in the form image and cutting out the character image in the detected license image.
In this license format 13c, for example, any one of “year”, “month”, and “day” which are characteristic reference characters printed in advance on the license, for example, “day” is used in this example. The data of the positional relationship between the character type and a plurality of reference characters “day” (the X coordinate of the character “day” with the point (the upper left corner etc.) of the license 23 as the reference (X coordinate 0, Y coordinate 0) Y coordinate and distance data between characters) are stored.

  In the license format 13c, area information for cutting out the image of the character string to be recognized from the license image in the form is set. The area information is designated by the X-coordinate / Y-coordinate of each character and distance data between the characters, as in the case of the character “day”. Note that the format, dictionary, and the like may be stored in the hard disk device 15.

For example, in the license format 13c, in addition to the area information to be cut out (specification information of each reading range when the position coordinate of the upper left corner of the license is X coordinate 0, Y coordinate 0), for example, the license number portion M ( As for the character image on the print surface on which a character string with a diagonal line as a background is printed, the first dictionary 13a and the fifth character image for the first to fourth character images out of twelve character images to be cut out. A dictionary designation condition (dictionary designation information) is set such that the second dictionary 13b is used for the character images with diagonal lines from the eighth to the eighth, and the first dictionary 13a is used for the ninth to twelfth character images. Has been.
The display unit 14 displays an image of the application form 1 captured by the scanner 2, a screen for character recognition processing from the extracted image of the license portion, text data of the character recognition result, and the like.

  The hard disk device 15 is installed with an operating system (hereinafter referred to as OS) and control software that causes the CPU 16 to perform control operations of each unit, and these cooperate to realize the operation of the present system. In the operation description, the operation of the CPU 16 after the computer is started will be described.

  That is, the CPU 16 uses at least one kind of characteristic characters printed on the license from the image of the application form 1 acquired by the scanner 2, for example, “year”, “month”, “day”, and the like. Characteristic character detecting means for detecting a plurality of characters, the expansion / contraction rate of the license image based on the positional relationship between the plurality of characters detected by the characteristic character detecting means and the preset reference position of the characters of the license 23 And / or direction (vertical, horizontal, normal direction, upside down direction, etc.), and image extracting means for extracting the image of the license 23 from the image of the application form 1; the license extracted by the image extracting means It functions as a character recognizing unit that performs character recognition on the image of the portion 23.

  The CPU 16 functions as an image information acquisition unit that acquires an image from a printing surface on which a character string with a diagonal line as a background is printed. The CPU 16 functions as a character image cutout unit that extracts a character string image with a diagonal line as a background from the images acquired by the image information acquisition unit and cuts out each character unit. The CPU 16 functions as a character image rotating unit that rotates the character image cut out by the character image cutting unit to an angle at which the oblique line becomes substantially horizontal.

  The CPU 16 functions as a ray processing means for removing (erasing) a black pixel component constituting a line in the horizontal direction from the character image rotated by the character image rotating means. The CPU 16 functions as a feature vector extracting unit that extracts a feature vector of a character image from which a black pixel component constituting a line is removed in the horizontal direction by the ray processing unit.

  The CPU 16 compares the feature vector of the character image extracted by the feature vector extraction means with the feature vector obtained by removing the feature vector of the hatched portion from the feature vector of the second dictionary 13b of the memory 13, and matches or approximates the feature vector. It functions as a character recognition means for outputting the text data it has.

  That is, the CPU 16 follows the license format 13c for reading the image in the license stored in advance in the memory 13, and the character string of a predetermined area in the image of the license 23, for example, name, date of birth, etc. Acquire images such as date, permanent address, address, license number, date of license acquisition, etc., cut out images of individual areas into individual character units, and individual character images by license format 13c for individual character images Character recognition processing is executed with reference to the specified dictionary (either the first dictionary 13a or the second dictionary 13b).

  When recognizing each cut-out character image, the CPU 16 rotates the character image at an angle at which the oblique line is almost horizontal (in this example, the character center is rotated 45 ° clockwise) for the character image having the oblique line in the background. Then, from the rotated character image, a line component (slanted line component) composed of black pixels in the horizontal direction is removed, a feature vector of the character image from which the diagonal line component is removed is extracted, and the second dictionary 13b dedicated to that part is used. Character recognition is performed by comparing feature vectors excluding the shaded portion.

  As shown in FIG. 2, a form, for example, an application form 1 for a mobile phone usage contract, is provided with an entry field 21 for an address, name, application contents, and the like, and an identification confirmation affixing field 22. . In the entry field 21, the applicant himself / herself is filled in with the ballpoint pen or the like.

  Precautions for attachment are pre-printed in the identification confirmation affixing field 22 and, for example, a license 23 or a health insurance card, etc. presented by the applicant himself as the confirmation of the applicant himself / herself is printed thereon. A copy of the ID is enlarged and reduced to a size that fits in the field and attached.

  The application form 1 is applied to a facsimile machine, etc., transmitted to the communication network, and received by the contract center. At the contract center, the received fax paper image, that is, the identification card is pasted on the application form 1 This image cannot be processed so that only the license part is corrected.

  In other words, the license on the fax paper image has a variety of positions in the identity verification item pasting field 22 and is transformed into various forms such as landscape, portrait, scale, and magnification. There may be. In addition, various characters previously printed on the application form 1 often protrude from the background of the license 23.

  The direction of the license 23 to be pasted on the application form 1 is pasted with the license 23 upside down if the normal direction is pasting the license in the license pasting field of the application form 1 A first case to be attached, a second case in which the license is attached in the longitudinal direction (a direction orthogonal to the normal direction), and a third case in which the respective cases are attached in a slightly inclined state are conceivable. The slightly inclined state is, for example, about 3 to 5 degrees in terms of human visual sense.

As shown in FIG. 2, the license 23 is printed with at least five letters (p1... P5) having the same size and the same date.
Therefore, in the present system, the position information of the central points of the five letters “p” (p1... P5), which serve as a reference for detecting the position where the license 23 is pasted, is registered in the memory 13. As shown in Fig. 3, the most probable combination is obtained by examining the position of the center point of the letters "d" (d1... Dn) detected from the image 1a of the application form to be recognized, corresponding to each other. Ask for.

  This problem of finding the optimal combination from among the combinations is called a “combination optimization problem”, and its solution methods include breadth-first search method, depth-first search method, genetic algorithm, and simulated annealing. Many methods are known, such as the method, and this method uses either method.

  If the most probable combination is required, it can be calculated from the relative position where there are items such as name, date of birth, address, license number, etc. on the image. The system cuts out the partial image of the license 23 and the description items therein according to the calculation result, and executes the character recognition process.

Here, an outline operation of the identification card recognition system will be described with reference to the flowchart of FIG.
In the case of this ID recognition system, when the application form 1 is set on the reading stand of the scanner 2 and the scanning start operation is performed, the scanner 2 scans the surface of the application form 1 to generate image data and generates the computer 10. Send to.

  In the computer 10, the CPU 16 scans the image data of the application form 1 received from the scanner 2 in the vertical direction to detect a plurality of characteristic characters “day” (S 101), and detects from the positional relationship of each characteristic character. The combination of a plurality of “days” is optimized (S102), and the expansion ratio and direction of the image of the license portion are detected.

  The CPU 16 cuts out the partial image of the license based on the expansion / contraction rate (size) and direction (tilt, upside down, etc.) of the detected image of the license portion, and further, the license according to the preset license format 13c. The contents of the license 23 affixed to the application form 1 are further cut out from the partial image (S103), and character recognition processing is performed on the individual items (S104). Is converted into text data, stored in the memory 13 and displayed on the display unit 14.

  After that, when the operation unit 11 is stored or output by the user, the CPU 16 associates the text data in the memory 13 with the image of the license portion to the storage location or output destination hard disk device 15 (database). Save or output.

The details of the license image recognition process in this identification card recognition system will be described below with reference to FIGS.
When performing the license image recognition process, the CPU 16 first scans the image of the application form 1 read by the scanner 2 in a certain direction (vertical direction) (S111 in FIG. 5).

  Then, the CPU 16 counts the number of consecutive white pixels and black pixels obtained by scanning the image (S112), and as shown in FIG. A place where the ratio of the numbers is almost black: white: black: white: black = a: b: a: b: a is searched (S113), and a place (image region) where the word “day” should be detected is detected. .

Next, the CPU 16 tracks black pixels on the left and right sides of the detected location, and detects a black pixel range 32 that is continuous in the horizontal direction (S114).
Further, the CPU 16 traces the black pixels vertically from the black pixels at the end of the range 32 of black pixels continuous in the horizontal direction, and detects the range 33 of black pixels continuous in the vertical direction (S115).

  Then, the CPU 16 determines that the detected horizontal range 32 and the vertical range 33 of the black pixel are the size of the reference character set in the memory 13 in advance, that is, 1 mm wide × 2 mm vertical, 5 mm wide × 5 mm vertical. It is determined whether or not the following conditions are satisfied (S116).

  As a result of this determination, if the condition is satisfied (Yes in S116), the CPU 16 performs character recognition processing on the images in the detected horizontal range 32 and vertical range 33, and the character recognition result is An area that is “day” is selected (S117).

  The CPU 16 repeats the above processing until there is no place where the ratio of the number of consecutive black pixels and white pixels is a: b: a: b: a (S118). That is, the CPU 16 detects a characteristic character from an image area where the arrangement ratio of the black pixel a and the white pixel b is a: b: a: b: a. The arrangement ratio between the black pixel a and the white pixel b is the ratio of the pixel arrangement order and the number of pixels.

  In the case of recognizing characters from an image in this way, conventionally, it has been generally performed to recognize all characters from the entire image. However, when all characters of the entire image are recognized, several hundred characters are recognized. In some cases, the process speed may be significantly reduced.

  Therefore, in the present embodiment, by using the fact that the character “day” font is lined up at equal intervals, the range in which the characters “day” are likely to exist is specified in advance from the arrangement of white and black pixels. By recognizing only the specified range, “character”, which is a characteristic character of the license 23, can be detected very quickly.

  The CPU 16 repeats the above process to detect a plurality of “day” characters, calculates a distance (interval) between the characters, and calculates a reference distance between the characters stored in the memory 13 in advance. In comparison with (interval), the expansion / contraction ratio such as the enlargement ratio or reduction ratio of the image of the license portion is obtained. The reference distance may be the actual size of the license or may be a distance (interval) multiplied by a certain magnification.

  When obtaining the expansion / contraction rate of the image of the license portion, the CPU 16 cuts out a range corresponding to the expansion / contraction rate from the image on the FAX sheet and temporarily stores the image of the license portion in the memory 13.

Then, the CPU 16 cuts out a character image from the image of the license portion of the memory 13 in accordance with the license format 13c, and the first dictionary 13a which is a standard dictionary or a character with hatched lines according to the position of each character image. Character recognition processing is executed using the second dictionary 13b, which is a dedicated dictionary, and the result of character recognition is stored in the memory 13.
The CPU 16 associates the text data, which is the result of this character recognition process, with the partial image of the license 23 temporarily stored in the memory 13 and registers it in the database constructed in the hard disk device 15.

Next, an application example of the license recognition process will be described with reference to FIG.
As shown in FIG. 7, strictly speaking, the letters “day” have different shapes in the upper part and the lower part from the center. By detecting this subtle difference, the processing can be further speeded up.

  That is, after recognizing (detecting) “day” which is a characteristic character of the license 23, the CPU 16 investigates the upper or lower pattern of each character “day” and detects the protruding portion 35. . In this process, the CPU 16 determines whether the upper half portion or the lower half portion of the character has the protruding portion 35.

Then, the CPU 16 counts the number of “days” having the protruding portion 35 at the upper part of the character and the number of “days” having the protruding portion 35 at the lower portion, and determines the direction (up and down) of the license 23 with the larger one. To do. That is, the CPU 16 determines the orientation of the image of the license 23 by counting the number of the upper part and the lower part of the detected character “day” on the side where a part of the character protrudes.
In this way, by detecting the direction of the protruding portion 35 and determining the direction of the character (upper: normal direction or lower: reverse direction) in advance, a plurality of “day” characters can be combined and optimized. Since it is no longer necessary to determine whether it is up or down, the license 23 can be recognized at a higher speed.

  Conventionally, a license is read by a dedicated scanner, and a ruled line is detected from the read image of the license 23, and the license is recognized based on the ruled line. In the case of forms such as copies and faxes received by facsimile machines, the ruled lines are often interrupted, and there is often a problem that the license cannot be recognized. However, in the license recognition system of this embodiment, the ruled line Instead, the image of the license 23 can be recognized more stably by detecting a plurality of “days” printed thick on the license pasted on the application form.

  Further, even when the license is arranged on the application form 1 in the normal direction or in the direction opposite to the normal direction (approximately 0 degrees or 180 degrees), the image portion of the license 23 can be recognized.

  Furthermore, when the “day” of the characteristic character is not detected in the initial vertical scanning, the image is rotated by 90 degrees, and the license image recognition process is performed, so that the direction orthogonal to the normal direction is obtained. The image portion of the license 23 arranged at (90 degrees or 270 degrees) can also be recognized.

  Normally, the character recognition function can recognize a character even if the character image is tilted to about ± 5 degrees. Therefore, according to the license recognition method of the above embodiment, when the license 23 is placed in the range of 0 ° ± 5 °, 90 ° ± 5 °, 180 ° ± 5 °, 270 ° ± 5 °. It becomes possible to recognize. If it is a normal application form, most license images can be recognized if this range is targeted.

  Further, if the image is rotated not only at 0 degrees and 90 degrees, but also at 10 degrees, such as 10 degrees, 20 degrees,... 170 degrees, and the license image recognition process is performed, 10 degrees ± 5 degrees, 20 degrees ± 5 degrees… 170 degrees ± 5 degrees can be targeted, and 190 degrees ± 5 degrees, 200 degrees ± 5 degrees… 350 degrees ± 5 degrees are also targeted due to the feature that characters can be recognized even if the day is upside down. It is possible to recognize licenses placed in all directions.

  Further, in the conventional method, there is a case where a recognition result shifted by one line is sometimes obtained because the ruled lines are arranged at equal intervals. In this embodiment, the character “day” is evenly displayed on the image. Since they are not arranged, there is an effect that it is difficult to obtain a shifted recognition result.

  In addition, “date”, which is a characteristic character on the license, can be recognized as “day” even if it is flipped upside down, so it can be detected as “day” without being rotated 180 degrees like other characters. The processing from the image of the application form 1 to the image of the license portion can be made extremely fast.

  Next, the details of the character recognition process (step S104) for the items described in the license image will be described with reference to FIGS.

  In the image of the license part in the image 1a of the application form shown in FIG. 3, there are columns for license conditions in addition to the name, permanent address, date of birth, address, date of license issuance, date of validity, etc. There is a license number M in the column.

  In this example, for example, as shown in FIG. 8, it is assumed that the license number M “9909811351090” is printed. When characters of the license number M are recognized, the number near the center of the 12-digit number is hatched, which is an obstacle to character recognition.

  Therefore, in the present embodiment, the CPU 16 has character images “9”, “0”, “9”, “8”, “1”, “0” having a white background among character images cut out for each character. ”,“ 9 ”,“ 0 ”, and the like, and the character images“ 1 ”,“ 1 ”,“ 3 ”,“ 5 ”, etc., which are shaded in the background, are used for different image processing and different dictionaries. Perform recognition processing.

The CPU 16 refers to the first dictionary 13a for the white character images “9”, “0”, “9”, “8”, “1”, “0”, “9”, “0”. Perform recognition processing. The creation of the first dictionary 13a is the same as a general character recognition dictionary, and the description thereof is omitted here.
In addition, the CPU 16 performs character recognition processing with reference to the second dictionary 13b for the character images “1”, “1”, “3”, and “5” with diagonal lines drawn in the background.
<Dictionary for exclusive use of license number shaded part>

As shown in FIG. 9, in the second dictionary 13b, patterns in which numbers from “0” to “9” are rotated 45 degrees to the right are registered in advance corresponding to the text data. In order to create the second dictionary 13b, first, a pattern to be registered first is scaled vertically and horizontally so as to have a certain size, and is normalized linearly.
Next, the normalized image is divided into a grid pattern vertically and horizontally, and a feature amount obtained by calculating a density feature or the like for each divided cell is registered as a dictionary.

FIG. 9 shows an example in which a numeral “3”, for example, is divided into a 9 × 9 grid. In this case, the feature amount is represented by an 81-dimensional feature vector. The CPU 16 performs character recognition using a character recognition method such as a composite similarity method, using the feature vector of the second dictionary 13b for the cut out character image with diagonal lines.
<Processing of character images with diagonal lines>
Specific processing will be described below.

  As shown in FIG. 8, the license number is a 12-digit number, and the middle four-digit portion is hatched at an angle of 45 degrees from upper right to lower left. The numbers in the license number column are printed at a fixed pitch. In consideration of this, the license format 13c stores information on the cutout position, range, and usage dictionary of the character image. CPU16 calculates | requires the pitch of a character from the part which is not shaded.

  Thereby, the CPU 16 cuts out the numbers in the hatched portion of the license number portion M one by one in order based on the character pitch, and performs character recognition processing with reference to the dictionary specified by the license format 13c.

  When the CPU 16 recognizes the license number one character at a time in sequence, for example, as shown in FIG. 11, when the character image “3” with a four-digit diagonal line is cut out, the CPU 16 reads the cut character image “3”. ”Is rotated 45 degrees to the right (S201 in FIG. 10), and the image is converted so that the diagonal line becomes a horizontal line (see FIG. 12).

Next, the CPU 16 detects the position of the oblique line with respect to the rotated image (S202).
Here, since the oblique line is a collection of black pixel components substantially parallel to the X axis on the rotated image, that is, a horizontal line, the CPU 16 determines the position of the oblique line by a method such as taking a horizontal projection of the black pixels. To detect.

  When the position of the oblique line is detected, the CPU 16 subsequently removes the oblique line from the image (S203) to obtain a thinned image as shown in FIG.

Next, in the same way as when the second dictionary 13b was created, after normalizing the character pattern with the hatched lines removed (aligning the image size with the registered pattern in the dictionary) A feature vector is obtained (S204). Normalization means scaling to the size of a dictionary registration pattern to make it linear. When the image is divided into a 9 × 9 grid as in this example, the feature vector is 81 dimensions.
Here, the features of the squares that are hatched are significantly different from the features of the first dictionary 13a, which is a standard dictionary, and therefore the features of the hatched portions adversely affect the result of character recognition.

  Accordingly, the CPU 16 compares the registered pattern of the second dictionary 13b with the second dictionary 13b using a vector obtained by removing the hatched square from the feature vectors of the input pattern of the shaped image.

That is, the CPU 16 compares the feature vector of the character image from which the hatched portion is removed with the feature vector of the second dictionary 13b using the feature vector excluding the hatched square.
FIG. 14 shows an example of the feature vector from which the hatched portion is removed. In the example shown in FIG. 14, since the hatched squares are the respective rows P (one row) of the second, fourth, and seventh rows, the feature vector for 3 × 9 squares from the 81-dimensional feature vector, That is, since 27 dimensions are removed, the comparison target is a 54-dimensional feature vector.

  In this way, by removing the feature vectors of the hatched portions and performing character recognition using the feature vectors of the other portions (S205), the character recognition process can be executed without the influence of the shaded lines.

  Further, when all the squares corresponding to the shaded part are removed from the feature vector without rotating the character, there are many areas where the shaded line is applied, and there is a possibility that most of the feature amount cannot be used for recognition.

However, by rotating 45 degrees so that the diagonal line is horizontal, the removed mass can be one horizontal row or two horizontal rows for each horizontal line, and a certain number of feature vector dimensions can be secured. There is also an advantage that it becomes possible.
FIG. 15 shows an example in which the feature amount in the shaded portion is removed without rotating. In the example of FIG. 15, 42 squares indicated by hatched squares are affected by the diagonal lines of the number background, and the feature quantity that can be used for character recognition is 39 dimensions. Since the amount of information is obviously reduced, the recognition accuracy is reduced.

  As described above, according to the ID recognition system of this embodiment, a recognition dictionary (second dictionary 3b) dedicated to hatched characters is created and registered with a character pattern rotated in advance by 45 degrees, and an input pattern is set to 45. A feature vector is obtained from the input image rotated by removing the oblique line and the feature vector is compared with the feature vector excluding the feature vector at the oblique line position from the second dictionary 3b to perform character recognition. With respect to the part where the hatched portion of the identification number is drawn, even if the part is a monochrome image, the character of that part can be recognized.

  That is, for the character image with diagonal lines, the second dictionary 13b dedicated thereto is referred to, and the component of the hatched part is excluded from both the obtained input image pattern and the registered pattern of the second dictionary 13b. By performing character recognition using vectors, it is possible to perform character recognition even for hatched characters.

  In addition, this invention is not limited only to the said embodiment. In the above embodiment, the description has been made with the letter “day” of the license, but other than this, for example, an insurance card or a passport can be recognized by the same method. For example, “year”, “month”, “number”, or the like may be used as long as there are a plurality of characters in the identification card to be recognized in addition to “day”.

  Further, in the above embodiment, only the shaded portion of the character image is removed, but it is slightly reduced that the feature vector is reduced by adding several pixels to the black pixel component in the character image, that is, by thickening the character. But it can be improved.

  In the above embodiment, the second dictionary 13b previously stores the character pattern feature vector rotated by 45 degrees and the text data in association with each other. By storing the vectors, it is not necessary to perform the process of removing the feature vectors in the shaded area during character recognition.

The figure which shows the structure of the identification card recognition system of this embodiment. The figure which shows an application form. The figure which shows the image of an application. The flowchart which shows the outline | summary operation | movement of the identification card recognition system of FIG. The flowchart which shows the detail of a license image recognition process. The figure for demonstrating the detection method of the character "day." The figure which shows judging the direction of a license by the difference in the shape of the upper part and the lower part of the character "day". The figure which expanded the image of the license number part. The figure which shows the registration content of the 2nd dictionary only for a license number oblique line part. The flowchart which shows the character recognition process with respect to the character image with a diagonal line of a license number. The figure which shows the character image "3" containing the diagonal line cut out from the license number part. The figure which shows a mode that character image "3" was rotated 45 degrees. The figure which shows a mode that the oblique line was removed from the rotated character image "3". The figure which shows the example of the feature vector which removed the oblique line part. The figure which shows a mode that the information content of a feature vector reduces when the feature vector of a shaded part is removed without rotating.

Explanation of symbols

  DESCRIPTION OF SYMBOLS 1 ... Application form, 2 ... Scanner, 10 ... Computer, 11 ... Operation part, 12 ... Communication I / F, 13 ... Memory, 13a ... 1st dictionary, 13b ... 2nd dictionary, 13c ... License format, 14 ... Display 15, hard disk drive, 16 CPU.

Claims (4)

Image information acquisition means for acquiring an image from a print surface on which a character string with a diagonal line is printed;
A character image extracting means for extracting an image of a character string with the diagonal line as a background from the images acquired by the image information acquiring means, and cutting out each character unit;
A character image rotating means for rotating the character image cut out by the character image cutting means at an angle at which the oblique line becomes substantially horizontal;
Ray processing means for removing a black pixel component constituting a line in a horizontal direction from the character image rotated by the character image rotating means;
Feature vector extracting means for extracting a feature vector of a character image from which black pixel components constituting a line in the horizontal direction are removed by the ray processing means;
A dictionary storage unit that stores a dictionary in which the feature vector of the reference character image and the text data in a state of being rotated at an angle that coincides with the angle of the oblique line in advance,
Character recognition means for outputting text data having a feature vector that matches and approximates the feature vector of the character image extracted by the feature vector extraction means and the feature vector stored in the dictionary storage unit. A character recognition device characterized by that. The character recognition device according to claim 1,
The image information acquisition means
Form image information acquisition means for acquiring an image from a form;
Characteristic character detection means for detecting a plurality of characters of at least one kind of “year”, “month”, and “day” from the image of the form acquired by the form image information acquisition means;
Based on the positional relationship of the plurality of characters detected by the characteristic character detection means and a reference position of the character of the identification card set in advance, the expansion rate and / or direction of the image of the identification card is obtained, The character recognition apparatus according to claim 1, further comprising an image extracting unit that extracts an image of an identification card from the inside. An image information acquisition unit acquiring an image from a printing surface on which a character string with a diagonal line as a background is printed;
A character image cutting means, from the image acquired by the image information acquisition means, extracting a character string image with the diagonal line as a background, and cutting out each character unit;
A step of rotating the character image cut out by the character image cut-out means to an angle at which the oblique line is substantially horizontal;
A ray processing means removing a black pixel component constituting a line in a horizontal direction from the character image rotated by the character image rotating means;
A step of extracting a feature vector of a character image from which a black pixel component constituting a line in a horizontal direction is removed by the ray processing unit;
A dictionary in which the feature vector of the reference character image rotated at an angle that matches the angle of the oblique line and the text data are stored in a dictionary storage unit in advance, and the feature vector extracting unit extracts the dictionary. Character recognition comprising: a step of outputting text data having a feature vector in which a character recognition unit compares and matches or approximates a feature vector of a character image and a feature vector stored in the dictionary storage unit Method. The character recognition method according to claim 3.
A form image information obtaining unit obtaining an image from the form;
A feature character detection unit detecting a plurality of characters of at least one kind of “year”, “month”, and “day” from the image of the form acquired by the form image information acquisition unit;
Based on the positional relationship between the plurality of characters detected by the characteristic character detection means and a preset reference position of the character of the identification card, the image extraction means obtains the expansion rate and / or direction of the image of the identification card, 4. A character recognition method according to claim 3, further comprising the step of extracting an image of an identification card from a form image.

Images