JP2006309402A - Character string recognition method, number plate recognition method and device - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a character recognition technology adaptive to an imaged image of low contrast, and combining reduction in data processing quantity necessary for character recognition with adaptivity to the variety of a character string. <P>SOLUTION: For the whole part of a character string image constituted by including a plurality of character areas, character segmentation position candidates as the candidates of the boundary of the character areas are determined (A), and the character segmentation position candidates are made to correspond to the boundary of the plurality of character areas and the character candidate areas as the candidates of the plurality of character areas can be determined, and the area arrangement candidates expressed as the combination of the character candidate areas are determined (B), and image recognition is performed for each of the character candidate areas, and the optimal area arrangement candidate is selected from among the area arrangement candidates based on the result of the image recognition (C), and a character string constituted of characters recognized by the image recognition (C) for the character candidate areas corresponding to the optimal area arrangement candidate is output as a recognized result character string. <P>COPYRIGHT: (C)2007,JPO&INPIT

Description

  The present invention relates to a character string recognition method and an apparatus for performing the method, and more particularly, to a character string recognition method and apparatus particularly suitable for recognition of a lowercase character string of a license plate.

  Character recognition for recognizing characters from an image captured by an imaging device is applied in various fields. License plate recognition for recognizing characters written on a license plate is one of typical applications of character recognition.

  The most typical character string recognition method is a method in which a captured image is binarized to generate a binarized image, and a character is recognized by image matching between the binarized image and a template.

  One problem with this character string recognition method is that when a captured image has a low contrast, characters are easily recognized by mistake. When a sufficient amount of light cannot be obtained when capturing a captured image, for example, when an illumination device is not provided in the imaging device, the contrast of the captured image may be reduced. If the contrast of the captured image is low, it is difficult to determine a threshold value for separating the character line and the background by binarization. Therefore, the line constituting the character is missing from the binarized image, or originally The two lines to be separated are connected in the binarized image. Undesirable missing or connected lines cause misrecognition of characters.

  From such a background, a technique for performing character recognition without performing binarization processing has been studied. For example, JP-A-8-147414 (Patent Document 1) and JP-A-2003-147414 (Patent Document 2) both detect character positions and recognize characters using normalized cross-correlation processing from image data. The technology to perform is disclosed. Specifically, the character recognition technique disclosed in Patent Document 1 estimates individual character positions constituting a character string by performing a normalized cross-correlation method and DP matching on the entire image. On the other hand, the character recognition technique disclosed in Patent Document 2 employs a method of detecting the position for each character using the normalized cross-correlation method and detecting the next character position based on the detected position. ing.

  One problem of the technology for performing character recognition without performing binarization processing is to achieve both a reduction in the amount of data processing necessary for character recognition and an improvement in handling character string diversity. For example, in the technique disclosed in Patent Document 1, DP matching is employed in addition to the normalized cross-correlation method to detect the character position of a captured image with an indefinite or unknown number of characters. However, since the normalized cross correlation method and DP matching are performed on the entire image, the amount of data processing required for character recognition is large. On the other hand, the technique disclosed in Patent Document 2 has a small range in which the normalized cross correlation method is performed, and can reduce the data processing amount. However, the technique disclosed in Patent Document 2 can be applied only to a captured image in which the number of characters is fixed (see, for example, paragraph [0006] of Patent Document 2).

  The inability to deal with the diversity of character strings, especially the inability to recognize characters with variable number of characters, is particularly problematic in license plate recognition, especially in lower-case character plate recognition of license plates; Is a land branch code (for example, “Kobe” or “Shinagawa”) which is a character for displaying a vehicle-based transportation branch office or automobile inspection registration office, and a vehicle type code (for example, “33” ”,“ 55 ”,“ 501 ”). The vehicle type code is composed of a 2-digit or 3-digit classification number, and therefore, the number of characters in the lower-case letter plate of the license plate is indefinite. Therefore, the character recognition technique that requires that the number of characters in the character string is constant cannot be applied to the recognition of the lowercase character string in the license plate.

  Against this background, there is a need to provide a character recognition technology that can simultaneously reduce the amount of data processing required for character recognition and cope with the diversity of character strings.

JP-A-8-147414 JP 2003-147414 A

  An object of the present invention is to provide a character recognition technology that can cope with a low-contrast captured image while simultaneously reducing the amount of data processing necessary for character recognition and dealing with the diversity of character strings. It is in.

  In order to achieve the above object, the present invention employs the means described below. In the description of technical matters constituting the means, in order to clarify the correspondence between the description of [Claims] and the description of [Best Mode for Carrying Out the Invention] Number / symbol used in the best mode for doing this is added. However, the added number / symbol should not be used to limit the technical scope of the invention described in [Claims].

A character string recognition method according to the present invention includes:
(A) With respect to the whole character string image (11) including a plurality of character regions (14 to 16) in which one character is copied, candidates for boundaries of the character regions (14 to 16) Determining character extraction position candidates (18, 25) that are:
(B) Character candidate regions (21 to 21) that are candidates for the plurality of character regions (14 to 16) by associating the character cutout position candidates (18, 25) with boundaries of the plurality of character regions (14 to 16). 24) and determining a region arrangement candidate expressed as a combination of the character candidate regions (21 to 24);
(C) performing image recognition for each of the character candidate areas (21 to 24), and selecting an optimal area arrangement candidate from the area arrangement candidates based on the result of the image recognition;
(D) outputting a character string made up of characters recognized in the image recognition in step (C) as the recognition result character string for the character candidate regions (21 to 24) corresponding to the optimal region arrangement candidate. It has.

  In the character string recognition method according to the present invention, character recognition is performed not on the entire character string image but on the character candidate areas (21 to 24) determined from the character cutout position candidates (18, 25). The amount of data processing can be reduced. On the other hand, in the character string recognition method according to the present invention, since the character string corresponding to the optimal area arrangement candidate selected from the area arrangement candidates that are combinations of the character candidate areas (21 to 24) is output as the recognition result, Character recognition of a character string having a variable number of characters can be realized. Such a character string recognition method is particularly effective for the recognition of the lowercase character string in the license plate.

  The character cutout position candidates (18, 25) are most typically determined from the luminance of each pixel of the character string image. The character cutout position candidates (18, 25) may be determined from a differential image of the character string image.

The step (C) includes:
(C1) calculating a width of each of the character candidate regions (21 to 24) defined by the character cutout position candidates (18, 25);
(C2) For each of the character candidate areas (21-24) defined by the character cutout position candidates (18, 25), an image of the character candidate areas (21-24) and a template image prepared in advance Calculating the dissimilarity;
(C3) calculating a character likelihood evaluation value based on the width of the character candidate region (21 to 24) and the degree of difference, and determining the optimal region arrangement candidate based on the character likelihood evaluation value It is preferred that To calculate the character likelihood evaluation value from the width of the character candidate area (21 to 24) and select the optimum area arrangement candidate from the character likelihood evaluation value, in order to select the correct optimum area arrangement candidate with simple processing It is valid.

The step (C) further includes:
(C4) calculating a width of a gap between two adjacent character candidate regions (21 to 24) of the character candidate regions (21 to 24) defined by the character cutout position candidates (18 and 25);
(C5) comprising a step of calculating a clearance likelihood evaluation value based on a width of a gap between the two character candidate regions (21 to 24), wherein the optimum region arrangement candidate is determined based on the clearance likelihood evaluation value. It is also preferable.

In another aspect of the present invention, a license plate recognition method according to the present invention comprises:
(E) The entire lowercase character string image (11) in which the lowercase character string of the license plate is copied is a candidate for the boundary of the character region (14-16) in which one character of the lowercase character string is copied. Determining character cutout position candidates (18, 25);
(F) Associating the character cut-out position candidates (18, 25) with the boundaries of the plurality of character regions (14-16), character candidate regions (21-21) that are candidates for the plurality of character regions (14-16) 24) and determining a region arrangement candidate expressed as a combination of the character candidate regions (21 to 24);
(G) performing image recognition for each of the character candidate areas (21 to 24), and selecting an optimal area arrangement candidate from the area arrangement candidates based on the result of the image recognition;
(H) A step of outputting, as at least a part of a recognition result, a character string made up of characters recognized in the image recognition in the step (G) for the character candidate regions (21 to 24) corresponding to the optimum region arrangement candidates. It comprises.

In one embodiment, the license plate recognition method comprises:
Furthermore,
(I) determining a land support code area in which the land support code in the lower case string is copied from the character cutout position candidates (18, 25);
(J) including a step of determining a vehicle type code region in a region other than the land support code region in the lower case sequence image (11), wherein the region arrangement candidate is a vehicle type code determined for the vehicle type code region. It is a region arrangement candidate.

In this case, the step (I)
(I1) defining a region having a predetermined range of width as a land support code candidate region, with the character cutout position candidate (18, 25) as a boundary, touching an end of the lowercase character string image (11), and
(I2) performing image recognition for each of the land support code candidate areas, and determining a land support code area in which the land support code is copied from the land support code candidate areas based on the result of the image recognition The land support code recognized for the land support code candidate region determined as the land support code region can be output as a part of the recognition result.

  In another embodiment, in the image recognition in the step (G), a land code template image is used for the character candidate region (21 to 24) corresponding to the leftmost character in the lowercase character string. Numeral template images are used for the character candidate areas (21 to 24) corresponding to the other characters, and the recognition result includes the recognition result of the land plate code of the license plate and the vehicle type code of the license plate. Including recognition results.

  If the recognition result “1” is obtained by the image recognition performed for a certain character candidate region in the character candidate regions (21 to 24) in the step (G), the certain When the image recognition is performed again for the enlarged region including the periphery of the character candidate region and a recognition result other than “1” is obtained for the enlarged region, the region arrangement candidate including the certain character candidate region is It is preferable that the optimal region arrangement candidate is not selected.

  When the character candidate area (21-24) is a rectangle, the character cutout position candidates (18, 25) are the horizontal character cutout position candidates (18, 25) that determine the left and right ends of the character candidate areas (21-24). 18) and vertical character cutout position candidates (25) for determining the upper and lower ends of the character candidate areas (21 to 24).

In addition, the step (E)
(E1) dividing the lowercase sequence image (11) into an upper part and a lower part;
(E2) It is preferable to include the step of determining the character cutout position candidate (18) from the image of the lower part. According to such a processing procedure, it is possible to eliminate the influence of screws that may appear in the lowercase string image (11).

In still another aspect of the present invention, a character string recognition device according to the present invention includes an imaging device (1) that acquires a character string image including a plurality of character regions in which one character is copied. And an image processing device (2) for performing character recognition of the character string image. The image processing apparatus (2) includes the following steps:
(A) A step of determining character cutout position candidates (18, 25) that are candidates for the boundary of the character region for the entire character string image;
(B) The character candidate regions (21 to 24) which are candidates for the plurality of character regions are determined by associating the character cutout position candidates (18, 25) with boundaries of the plurality of character regions, and the character candidate regions Determining a region arrangement candidate expressed as a combination of (21-24);
(C) performing image recognition for each of the character candidate regions (21 to 24), and selecting an optimum region arrangement candidate from the region arrangement candidates based on the result of the image recognition; and (D) Executing a step of outputting a character string made up of characters recognized in the image recognition in the step (C) as a recognition result character string for the character candidate regions (21 to 24) corresponding to the optimum region arrangement candidate. It is configured.

In still another aspect of the present invention, a license plate recognition device according to the present invention includes an imaging device (1) that acquires a captured image of a license plate, and an image processing device that performs character recognition of the license plate from the captured image. (2). The image processing apparatus (2) includes the following steps:
(K) a step of cutting out a lowercase string image (11) in which the lowercase string of the license plate is copied from the captured image;
(E) ′ determining a character cutout position candidate (18, 25) that is a candidate for a boundary of a character region in which one character of the lowercase character sequence is copied for each of the lowercase character sequence image (11);
(F) associating the character cutout position candidates (18, 25) with boundaries of the plurality of character regions to determine character candidate regions (21 to 24) that are candidates for the plurality of character regions; Determining a region arrangement candidate expressed as a combination of (21-24);
(G) performing image recognition for each of the character candidate areas (21 to 24), and selecting an optimum area arrangement candidate from the area arrangement candidates based on the result of the image recognition; and (H) the above Executing a step of outputting a character string made up of characters recognized in the image recognition in the step (G) as at least a part of the recognition result for the character candidate regions (21 to 24) corresponding to the optimum region arrangement candidate. It is configured.

  According to the present invention, it is possible to provide a character recognition technology that can cope with a captured image with low contrast, while at the same time reducing both the amount of data processing required for character recognition and dealing with the diversity of character strings.

First Embodiment:
1. Outline FIG. 1 shows a configuration of a license plate recognition apparatus 10 according to a first embodiment of the present invention. The license plate recognition device 10 includes an imaging device 1 and an image processing device 2. The imaging device 1 captures a license plate of a vehicle and obtains a license plate image. The image processing apparatus 2 performs license plate recognition on the license plate image.

  The image processing device 2 includes a storage device 3, a calculation device 4, and an interface 5 connected to the imaging device 1. The storage device 3 stores a license plate recognition program 6, a land code template 7, a vehicle type code template 8, a use code template 9a, and a serial number template 9b.

  The license plate recognition program 6 describes an algorithm for license plate recognition. The license plate recognition is performed by the arithmetic device 4 executing the license plate recognition program 6.

  The land code template 7 includes a land image code image, information indicating the land code class of each template image (that is, information indicating which land code each template image corresponds to) Are stored in association with each other. As will be described later, the character recognition of the land code described in the license plate is performed by pattern recognition using the land code template 7.

  The vehicle type code template 8 is associated with a template image of a number constituting the vehicle type code and information indicating a character class of each template (that is, information indicating which number each template image corresponds to). Stored. As will be described later, the character recognition of the vehicle type code described on the license plate is performed by pattern recognition using the vehicle type code template 8.

  The usage code template 9a includes a template image of a usage code (that is, kana characters and English characters that indicate whether it is for business use, etc.) and information indicating the usage code class of each template image (that is, each template image). Are stored in association with each other). Character recognition of the application code described on the license plate is performed by pattern recognition using the application code template 9b.

  Corresponding to the template for serial number 9b is a template image of numbers constituting the serial number and information indicating the character class of each template image (that is, information indicating which number each template image corresponds to). It is stored with. Character recognition of serial numbers described on the license plate is performed by pattern recognition using the serial number template 9b.

  FIG. 2 is a flowchart showing a license plate recognition process performed by the image processing apparatus 2. When the license plate image is acquired from the imaging device 1 (step S01), the image processing device 2 separates the lowercase letter image and the uppercase letter image from the acquired license plate image (step S02); The above-mentioned lowercase character string is an image, and the uppercase character string image is an image in which a usage code and a serial number are copied. FIG. 3 shows an example of a license plate image and a lower case string image. Both the lowercase sequence image and the uppercase sequence image are grayscale images and are not binarized images. Subsequently, character recognition is performed for each of the lowercase character string image and the uppercase character string image, and the characters written on the license plate are recognized (steps S03 and S04). Character recognition of the characters written on the license plate is performed by normalization cross-correlation or the like from the lowercase character sequence image that is a grayscale image and the uppercase character sequence image. In the character recognition of the present embodiment, a binarized image is not generated from a grayscale image.

  The subject of this embodiment is optimization of character recognition of lowercase sequence images. Below, the process which performs character recognition about a small letter sequence image is demonstrated in detail.

The final goal of character recognition in this embodiment is as shown in FIG.
(1) The land support code area 12 and the vehicle type code area 13 are optimally defined in the lower case sequence image 11, and the vehicle type code area 13 is further divided into two or three character areas each having one number: first By optimally dividing the character area 14, the second character area 15, and the third character area 16, and (2) by character recognition, the land branch code described in the land branch code area 12 and the vehicle type code area 13 This is two ways of acquiring the described vehicle type code (that is, a numeric string composed of the numbers described in the first character area 14, the second character area 15, and the third character area 16). Although FIG. 4 illustrates a case where the vehicle type code area 13 is configured by three character areas, it should be noted that the vehicle type code area 13 may be configured by two character areas.

  The lower case sequence image 11, the boundary between the land code area 12 and the vehicle type code area 13, and the boundary between the character area 14 are called a character cutout position 17. Therefore, the above target (1) can be paraphrased as setting an appropriate character cutout position 17 in the lowercase character string image 11. In the following, the procedure for achieving the above two goals will be described in detail.

  FIG. 5 is a flowchart showing an algorithm for character recognition of the lowercase sequence image 11 in the first embodiment. The processing procedure for character recognition of the lowercase sequence image 11 is roughly as follows: extraction of a character extraction position candidate (step S11), recognition of a land code (step S12), and recognition of a vehicle type code (step S13). Composed. In the following, these processes will be described in detail.

2. Extraction of Character Cutout Position Candidate With reference to FIG. 6A, in character recognition of the lowercase character string image 11, first, a character cutout position candidate 18 is extracted (step S11). The character cutout position candidate 18 is a candidate for the character cutout position 17 to be defined in the lowercase character string image 11. In the present embodiment where the lowercase character string image 11 is long in the horizontal direction (x-axis direction), a character cutout position candidate 18 in the horizontal direction is defined. Each character cutout position candidate 18 is defined so as to extend in the vertical direction (y-axis direction) of the lowercase sequence image 11.

  The character cutout position candidate 18 is determined directly from the lowercase sequence image 11 or from the differential image of the lowercase sequence image 11. Specifically, a portion estimated to correspond to the background of the lowercase sequence image 11 is determined from the luminance data value of the pixels of the lowercase sequence image 11 (that is, the brightness value), and both ends of the lowercase sequence image 11 and the lowercase letters are determined. A character cutout position candidate 18 is determined in a portion estimated as the background of the row image 11. Further, a differential image of the lowercase character string image 11 can be used for extraction of the character cutout position candidate 18 instead of the luminance data value. In this case, a differential image of the lower case sequence image 11 is calculated from the luminance data of each pixel of the lower case sequence image 11, and a portion estimated as the background of the lower case sequence image 11 is specified from the differential image. Also in this case, character cutout position candidates 18 are determined at both ends of the lowercase character string image 11 and portions corresponding to the background thereof.

More specifically, when the paint color of the license plate is known by some means, for example, the character cutout position candidate 18 can be extracted by the following procedure. First, the luminance projection in the vertical direction of the lowercase string image 11 is calculated. The luminance projection in the vertical direction is the sum of the luminance values calculated for each of the pixel columns constituting the lowercase column image 11; here, the pixel column is a column of pixels arranged in one column in the vertical direction. That is. The luminance projection F (x) of the pixel row located at the coordinate x is
F (x) = Σσ (x, y),
Is calculated by Here, σ (x, y) is the luminance (or gradation) of the pixel located at the coordinates (x, y) of the lowercase sequence image 11, and Σ is the entire range of the y coordinates of the lowercase sequence image 11. Represents the sum.

  When the license plate paint color is such that the brightness of the characters is higher than the background (for example, when the license plate paint color is white or yellow), the brightness projection F (x) is minimized. When the position and the brightness projection F (x) are smaller than the threshold over a range wider than a predetermined area, the position where the brightness projection F (x) rises in the range (“river” and “1” in FIG. 6A). ”Is extracted as a character cutout position candidate 18.

  On the other hand, as shown in FIG. 6B, when the license plate paint color is such that the brightness of the characters is higher than the background (for example, the license plate paint color is green or black). If the luminance projection F (x) is larger than the threshold over a position where the luminance projection F (x) is maximum, and the luminance projection F (x) is larger than a predetermined area, the luminance projection F (x) rises in that range. The lowering position is extracted as the character cutout position candidate 18.

  On the other hand, when the paint color of the license plate is unknown, the character cutout position candidate 18 can be extracted by the following procedure, for example. First, a differentiation process is performed on the lowercase sequence image 11 to obtain the sum of the differential values in the x and y directions for each pixel. Further, a differential value projection that is the sum of differential values in the x and y directions for each pixel column is obtained. The differential value projection is used in place of the luminance projection, and the same processing as described above is performed, whereby the character cutout position candidate 18 is extracted.

  What should be considered in the extraction of the character cutout position candidate 18 is that the lowercase string image 11 is cut out from the upper part of the license plate, and thus the lowercase string image 11 can include a screw for fixing the license plate to the vehicle body. . The screw may prevent proper extraction of the character cutout position candidate 18.

  In order to eliminate the influence of screws, as shown in FIG. 6C, it is preferable that the character cutout position candidate 18 is determined from the lower part of the lowercase character string image 11 where no screw is shown. More specifically, the lower-case string image 11 is divided into an upper portion 11a in which screws are shown and a lower portion 11b in which screws are not shown. The lower portion is subjected to the same processing as described above, and the character cutout position candidate 18 is extracted. As a result, the influence of the screw can be eliminated and the character cutout position candidate 18 can be appropriately extracted.

3. Recognition of Land Support Code As shown in FIG. 5, the land support code is recognized following the extraction of the character cutout position candidate 18 (step S12).

In the land support code recognition process, first, a land support code candidate area is set (step S12-1). The land support code candidate area is a candidate for the land support code area 12. As shown in FIG. 7, the left boundary of the land code candidate area 19 is fixed at the left end of the lowercase character string image 11, and the right boundary is the character cutout position candidate 18 extracted in step S <b> 11. Selected from. However, the land code candidate area 19 is determined so that the horizontal width W _R (that is, the distance from the left end of the lowercase character string image 11 to the corresponding character cutout position candidate 18) satisfies the following condition. R:
h · k _{min — R} <W _R <h · k _{max — R} , (1)
Here, h is the height of the lower case sequence image 11, and k _{min — R} and k _{max — R} are predetermined parameters. Such a method for determining the land support code candidate area 19 is based on the aspect ratio of the land support code area 12 being within a certain range. The land code candidate area 19 having an aspect ratio that cannot be assumed as the aspect ratio of the land code area 12 is not set. In the example of FIG. 7, four land supporting code candidate regions 19 ₁ to 19 ₄ is set.

Following the setting of the land support code candidate area 19, as shown in FIG. 5, the land support code area 12 is determined and the land support code is recognized (step S12-2). More specifically, as shown in FIG. 8, pattern recognition is performed using the template image stored in the land code template 7 for each of the images of the land code candidate area 19, and the closest The class (that is, the land support code class with the smallest difference S) and the difference S are calculated. Of the land support code candidate areas 19, the one with the smallest difference is determined as the land support code area 12, and the nearest branch class determined for the land support code candidate area 19 is the land support code to be finally obtained. It is determined. In the example of FIG. 7, the most dissimilarity Riku支code candidate regions 19 ₄ to reduce the S it is determined as Riku支coding region 12, closest class "Shinagawa" final corresponding to the Riku支code candidate region 19 ₄ It is determined that the land support code should be obtained.

  Although the land support code is generally composed of a plurality of characters, it should be noted that the image processing apparatus 2 of the present embodiment handles the land support code as one character as a whole. The term “one character” in the present embodiment means a figure that is copied in one template image.

4). Recognition of vehicle type code As shown in FIG. 5, after the land code is recognized, the vehicle type code is recognized (step S13).

  The recognition of the vehicle type code is first started by determining the vehicle type code area 13 (step S13-1). As shown in FIG. 9, the vehicle type code area 13 is determined from the land code area 12 determined in the land code recognition process. More specifically, one character cutout position candidate 18 whose distance from the land code region 12 is within a predetermined range is determined as the left end of the vehicle type code region 13, and the right end of the lowercase character sequence image 11 is the vehicle type code region 13. Is determined as the right end of

  Subsequently, a vehicle type code area arrangement candidate is set for the determined vehicle type code area 13 (step S13-2). The vehicle type code area arrangement candidate is a candidate for arrangement of the character areas 14 to 16 in the vehicle type code area 13. Since the vehicle type code is composed of two or three characters, two types of vehicle type code region arrangement candidates corresponding to the two-character vehicle type code and vehicle type code region arrangement candidates corresponding to the three-character vehicle type code are determined. . As shown in FIG. 10, the vehicle type code area arrangement candidate corresponding to the 3-character vehicle type code is expressed by a combination of the first character candidate area 21, the second character candidate area 22, and the third character candidate area 23. The Here, the first character candidate area 21 is a candidate for the first character area 14 where the first character of the vehicle type code exists, and the second character candidate area 22 exists for the second character of the vehicle type code. The third character candidate area 23 is a candidate for the third character area 16 in which the third character of the vehicle type code exists. Both ends of the character candidate areas 21 to 23 are defined by two character cutout position candidates 18. Similarly, the vehicle type code area arrangement candidate corresponding to the 2-character vehicle type code is expressed by a combination of the first character candidate area 21 and the second character candidate area 22.

Specifically, the determination of the vehicle type code area arrangement candidate is performed by setting the character extraction position candidate 18 as the start point (left end) and end point (right end) of the first character candidate area 21, the second character candidate area 22, and the third character candidate area 23. ). As an example, let us consider a case where there are _six character cutout position candidates 18 ₁ to 18 ₆ for the positions “1” to “6” of the vehicle type code area 13 as shown in FIG. For example, in a certain car model code area arrangement candidate, the character cutout position candidates 18 ₁ and 18 ₂ are associated with the start point (left end) and the end point (right end) of the first character candidate area 21, respectively. ₂ and 18 ₃ are associated with the start point and end point of the second character candidate region 22, respectively, and the character cutout position candidates 18 ₃ and 18 ₄ are associated with the start point and end point of the third character candidate region 23, respectively. In other vehicle type code region arrangement candidates, different associations are performed as shown in FIG. In the example of FIG. 11, the character cutout position candidates 18 ₁ and 18 ₂ are respectively associated with the start point (left end) and the end point (right end) of the first character candidate area 21, and character cutout position candidates 18 ₃ and 18 _4. Are respectively associated with the start point and end point of the second character candidate region 22, and the character cutout position candidates 18 ₅ and 18 ₆ are associated with the start point and end point of the third character candidate region 23, respectively.

However, the vehicle type code area arrangement candidates include the width W ₁ , W ₂ , and W _{3 of} the first character candidate area 21, the second character candidate area 22, and the third character candidate area 23, all of which are the minimum character width. It is determined to be not less than W _{min — N and not} more than the maximum character width W _{max — N.} Here, the minimum character width W _{min — N} and the maximum character width W _{max — N} are constants represented by the following expressions:
W _min — _N = h × k _{min — N} , (2a)
W _max — _N = h × k _{max — N} , (2b)
Here, h is the height of the lowercase sequence image 11, and _{kmin_N} and _{kmax_N} are predetermined parameters. Such a method for determining the combination of the character candidate areas 21 to 23 is based on the fact that the aspect ratio of characters (that is, numbers) that can be a vehicle type code falls within a certain range. A vehicle type code area arrangement candidate including a character candidate area having an aspect ratio that cannot be assumed as the aspect ratio of the characters constituting the vehicle type code is not set. This is effective in reducing the number of vehicle type code area arrangement candidates and thereby reducing the amount of calculation.

  For example, as shown in FIG. 12, if the width is not considered, five character candidate regions can be set as the first character candidate regions; however, the start point is located at the position “1” due to the above width limitation. And the character candidate region (1) whose end point is the position “2” and the character candidate region (2) whose start point is the position “1” and whose end point is the position “3” are the first character candidate regions 21 Permissible.

In addition, the vehicle type code region arrangement candidate is such that, for all of the character candidate regions 21 to 23 included therein, the left end of the i-th character candidate region is a distance (i−1) × W _min — _N from the left end of the vehicle type code region 13. At the above position, the right end of the i-th character candidate area is determined such that the distance from the left end of the vehicle type code area 13 is at a position equal to or less than i × W _{max_N} . For example, as shown in FIG. 13, the second character candidate region 22 is located at a position where the distance from the left end of the vehicle type code region 13 is _{equal to} or greater than W _{min — N} , and the distance is 2 × W. It is determined so that the right end is located at a position that is less than or equal to _{max_N} . This is to eliminate vehicle type code area arrangement candidates in which any one of the first character candidate area 21, the second character candidate area 22, and the third character candidate area 23 is present at an invalid position.

  For example, in the example of FIG. 12, when the vehicle type code area arrangement candidate is determined so as to satisfy these conditions, the character candidate areas (1) and (2) are set as the first character candidate area 21 as the character candidate area ( 3) to (6) are determined as the second character candidate region 22, and the character candidate regions (5) to (9) are determined as the third character candidate region 23. The vehicle type code region arrangement candidate is a combination of the first character candidate region 21, the second character candidate region 22, and the third character candidate region 23 determined in this way.

Subsequently, as illustrated in FIG. 5, the evaluation value φ _string is calculated for each of the vehicle type code area arrangement candidates (step S13-3). The evaluation value φ _string of the vehicle type code region arrangement candidate is a character likelihood evaluation value φ _char representing “character character” of each character candidate region and a character likelihood evaluation value φ representing “gap character” between adjacent character candidate regions. It is expressed as the sum of _gaps . The character-likeness evaluation value φ _char is a value that increases as the image of the corresponding character candidate region “looks like a character” and decreases as “not like a character”. On the other hand, the gap-likeness evaluation value _φgap is a value that decreases as the area between two adjacent character candidate areas is “not like a gap” (for example, too narrow or too wide as a gap). More specifically, evaluation value phi _string models coding region arrangement candidates that correspond to the three letter models code is represented by the following formula:
φ _string = φ ¹ _char + φ ¹² _gap + φ ² _char + φ ²³ _gap + φ ³ _char , (3a)
Is calculated by Here, φ ⁱ _char is a character likelihood evaluation value of the i-th character candidate area, and φ ^ij _gap is a gap likelihood evaluation value calculated for the i-th character candidate area and the j-th character candidate area. On the other hand, the evaluation value φ _string of the vehicle type code area arrangement candidate corresponding to the 2-character vehicle type code is expressed by the following formula:
φ _string = φ ¹ _char + φ ¹² _gap + φ ² _char , (3b)
Is calculated by

The character likelihood evaluation value φ ⁱ _char of the i-th character candidate region is calculated in the following three steps: First, character recognition is performed on the image of the i-th character candidate region, the closest character class, and The dissimilarity S _i is calculated. For example, as illustrated in FIG. 14, when the left end of the first character candidate area 21 is the character cutout position candidate 18 ₁ and the right end is the character cutout position candidate 18 ₂ , the first character candidate area 21 includes “ The character “1” is included in the complete state. In such a case, “1” is selected as the closest character class, and a small difference is given. On the other hand, if the left edge of the first character candidate region 21 character segmentation position candidate 18 _1, right end is a character segmenting position candidate 18 _3, the first character candidate region 21 characters and "8" and "1" Contains the left half of the character. In such a case, an inappropriate class is selected as the closest character class, and a large difference is given.

Further, a character width deduction point P ⁱ _{char of} the i-th character candidate area is calculated. The character width deduction value P ⁱ _char is a numerical value that is increased when the width W _i of the i-th character candidate region is not valid as the width of the character region in which the character exists, and specifically, the following formula:
P ⁱ _char = c ₁ (W _min — _char −W _i ), (W _i <W _min — _char ) (4a)
P ⁱ _char = c ₂ (W _i −W _{max_char} ), (in the case of W _{max_char} <W _i ) (4b)
P ⁱ _char = _{0, (the} case of _{W min_char ≦ W i ≦ W max_char} ) ··· (4c)
Determined by Here, c ₁ and c ₂ are weight parameters, and W _{min_char} and W _{max_char} are parameters respectively _indicating a lower limit value and an upper limit value of the width of a character candidate area including a “character-like” image. Yes, determined by:
W _{min_char} = h · k _{min_char} ,
W _{max_char} = h · k _{max_char} ,
However, h is the height of the lowercase string _{image, k min_char,} k _{max_char} is a predetermined constant. Such a method for determining the character width deduction point P ⁱ _char is based on the fact that the aspect ratio of the character regions 14 to 16 falls within a certain range. A character candidate area having an aspect ratio that cannot be assumed as the aspect ratio of the character areas 14 to 16 is not set.

Referring to FIG. 15, when the width W _i of the i-th character candidate region is not _less than W _{min_char and not} more than W _{max_char} , the character width deduction point P ⁱ _char is 0. On the other hand, if the width W _i of the i-th character candidate region is smaller than W _{min_char} , it is determined that the image of the i-th character candidate region does not _look like a character, and the character width deduction point P ⁱ _char is set to W _{min_char} and the width W _i . It is increased in accordance with the difference. Similarly, if the width W _i of the i-th character candidate area is larger than W _{max_char} , it is determined that the image of the i-th character candidate area does not look like a character, and the character width deduction point P ⁱ _char is set to the widths W _i and W Increased according to the difference from _{max_char} .

The character likelihood evaluation value φ ⁱ _char is determined so as to decrease as the minimum dissimilarity S _i increases and as the character width deduction point P ⁱ _char increases. For example, the character-likeness evaluation value φ ⁱ _char is expressed by the following formula:
φ ⁱ _char = C−S _i −P ⁱ _char , (5)
Can be calculated by: Here, C is a constant. As understood from the equation (5), the character likelihood evaluation value φ ⁱ _char is a value in consideration of the width W _i of the i-th character candidate region. Calculating the character-likeness evaluation value φ ⁱ _char using not only the degree of difference but also the width W _i of the i-th character candidate region simply and appropriately evaluates the “character-likeness” of the i-th character candidate region. It is effective for.

One problem that may occur when calculating the character-likeness evaluation value φ ⁱ _char is that the character extraction position candidate 18 exists in the middle of the numbers such as “0” and “8”. There is an image in which numbers such as “0” and “8” are vertically cut in half. As shown in FIG. 16, such an image may be erroneously recognized as having the closest character class “1”, and the difference may be reduced. In order to avoid such a situation, the procedure for character recognition of the i-th character candidate region performed when calculating the character likelihood evaluation value φ ⁱ _char is performed according to the flow shown in FIG. Is preferred.

First, character recognition of the i-th character candidate area is performed using the vehicle type code template 8, and the closest class and the minimum difference degree _Si are acquired (step S21). When the closest class obtained by character recognition is “1” (step S22), an area obtained by expanding the i-th character candidate area by several pixels to the left and right is set (step S23), and the set area The character recognition is performed again using the vehicle type code template 8 (step S24). As can be understood from FIG. 17, by using an area where the i-th character candidate area is expanded by several pixels to the left and right, the i-th character candidate area really includes “1” or “ This is because it is possible to determine whether numbers such as “0” and “8” include an image that is vertically cut in half. When the closest character class obtained in step S24 is “1” again (step S25), the minimum difference S _i obtained in step S21 is adopted, and the character likelihood evaluation value φ ⁱ _char is determined. Calculation is performed. On the other hand, if the closest class obtained in step S24 is not “1”, the vehicle type code area arrangement candidate including the i-th character candidate area is invalidated (step S26). More specifically, when the closest class obtained in step S24 is not “1”, the minimum difference S _i of the i-th character candidate region is rewritten to a very large predetermined number. As a result, the character likelihood evaluation value φ ⁱ _char decreases, and the evaluation value φ _string of the vehicle type code area arrangement candidate including the i-th character candidate area decreases. As a result, the vehicle type code region arrangement candidate including the i-th character candidate region is invalidated (that is, the vehicle type code region arrangement candidate is not selected as the optimum vehicle type code region arrangement candidate), and erroneous character recognition is performed. Probability can be reduced.

On the other hand, the gap-likeness evaluation value φ ^ij _gap between the i-th character candidate area and the j-th character candidate area is, as shown in FIG. 18, a gap between the i-th character candidate area and the j-th character candidate area. It is calculated from the gap deduction point value P ^ij _gap depending on the width W ^ij _gap of. The gap deduction value P ^ij _gap is a numerical value that is increased when the width W ^ij _gap of the gap between the i-th character candidate area and the j-th character candidate area is not valid as a gap. :
P ^ij _gap = c _gap (W _{min —gap} −W ^ij _gap ),
(W ^ij _gap <W _{min_gap} ) (6a)
P ^ij _gap = c _gap (W _i −W _{max_char} ),
(In the case of _{^{W max_gap <W ij gap) ···}} (6b)
P ^ij _gap = 0, (when W _{min_gap} ≦ W ^ij _gap ≦ W _{max_gap} ) (6c)
Determined by Here, c _gap is a predetermined parameter, and W _{min_gap} and W _{max_gap} are parameters respectively _indicating a lower limit value and an upper limit value of the width of the character candidate area including the “character-like” image. Yes, determined by the following formula:
W _{min_gap} = h · k _{min_gap} ,
W _{max_gap} = h · k _{max_gap} ,
However, h is the height of the lowercase string _{image, k min_gap,} k _{max_gap} is a predetermined constant. Such a method of determining the evaluation value φ ^ij _gap of the gap is based on the fact that the aspect ratio of the gap between the character areas 14 to 16 falls within a certain range. When there is a gap having an aspect ratio that cannot be assumed as the aspect ratio of the gap between the character areas 14 to 16, the gap deduction point value P ^ij _gap is increased.

The gap likelihood evaluation value φ ^ij _gap is determined so as to decrease as the gap deduction point value P ^ij _gap increases. For example, the clearance likelihood evaluation value φ ^ij _gap is given by the following formula:
φ ^ij _gap = −P ^ij _gap , (7)
Can be calculated by: As understood from the equation (7), the gap likelihood evaluation value φ ^ij _gap is a value depending on the width W ^ij _gap of the _gap between the i-th character candidate region and the j-th character candidate region. The calculation of the gap likelihood evaluation value φ ^ij _gap using the gap width W ^ij _gap easily and appropriately evaluates the “gap likelihood” of the gap between the i-th character candidate area and the j-th character candidate area. It is effective to do.

By the above procedure, the character likelihood evaluation value φ ⁱ _char and the gap likelihood evaluation value φ ^ij _gap are calculated, and further, the evaluation value φ _string is calculated using the equations (3a) and (3b), and then the calculated evaluation is calculated. From the value φ _string , the optimal vehicle type code area arrangement candidate among the vehicle type code area arrangement candidates corresponding to the 3-character vehicle type code and the optimal vehicle type code among the vehicle type code area arrangement candidates corresponding to the 2-character vehicle type code A region arrangement candidate is determined (step S13-4). The vehicle type code area arrangement candidate that maximizes the evaluation value φ _string calculated from the equation (3a) is determined to be the optimum vehicle type code area arrangement candidate among the vehicle type code area arrangement candidates corresponding to the three-character vehicle type code. Is done. Similarly, the vehicle type code region arrangement candidate that maximizes the evaluation value φ _string calculated from the expression (3b) is the optimum vehicle type code region arrangement candidate among the vehicle type code region arrangement candidates corresponding to the two-character vehicle type code. Determined to be.

The evaluation value φ _string is calculated individually for all the vehicle type code area arrangement candidates, and the optimal vehicle type code area arrangement candidate (for each of the two-character vehicle code and the three-character vehicle code) is calculated from the calculated evaluation value φ _string. However, it is possible to excessively increase the calculation amount. In order to reduce the amount of calculation required for calculating the evaluation value of the vehicle type code area arrangement candidate, it is preferable to employ DP matching for determining the optimal vehicle type code area arrangement candidate. By adopting DP matching, it is possible to avoid the duplication of calculation and calculate the evaluation value φ _string of the vehicle type code area arrangement candidate with a small calculation amount.

Referring to FIG. 10 (and FIG. 11), in the DP matching according to the present embodiment, vehicle type code region arrangement candidates (that is, character cut-out position candidates 18 ₁ to 18 ₆ and first to third character candidate regions 21 to 23). Is developed on a coordinate system consisting of the A axis corresponding to the character cutout position candidate 18 and the B axis corresponding to the start and end points of the first to third character candidate areas 21 to 23. Expressed as a graph. Elements a _{1 to} a ₆ on the A axis correspond to character cutout position candidates 18 ₁ to 18 ₆ belonging to the vehicle type code area 13, and elements b _{1 to} b ₆ on the B axis correspond to the first character candidate area. 21 corresponds to the start and end points of the second character candidate region 22 and the third character candidate region 23; the number of elements a _j on the A axis is the number of character cutout position candidates 18 belonging to the vehicle type code region 13 Note that they are adjusted accordingly. Each correspondence graph has a route from the origin (a ₁ , b ₁ ) to the coordinates (a _end , b ₄ ) (when it corresponds to a two-character vehicle model code), or coordinates from the origin (a ₁ , b ₁ ) ( a _end , b ₆ ); where a _end is an element on the A axis of the end point of the corresponding graph and is selected as the end point of the second character candidate region 22 or the third character candidate region 23 This is an element corresponding to the character cutout position candidate 18.

Origin _(a 1, _{b 1)} from the coordinates _(a j, _{b i)} up to the maximum value of the evaluation values phi _string models coding region arrangement candidate corresponding to the corresponding graph Φ _(a j, _{b i)} and represents the if,
Φ (a _j , b _i ) = MAX (a _k , b _p ) [Φ (a _k , b _p )
+ Φ _(i) (a _k , b _p , a _j , b _i )], (8)
Is established. Here, _{b p} is the previous element of the _{b i,}
b _p = b _i−1 , (9)
It is. MAX (a _k , b _p ) [X (a _k , b _p )] represents the maximum value of X (a _k , b _p ) when a _k and b _p are changed. Furthermore, φ _(i) (a _k , b _p , a _j , b _i ) is
(Α) When i is an odd number, a gap likelihood evaluation value φ _gap of a gap formed between character cutout position candidates corresponding to the elements a _k and a _j ,
(Β) When i is an even number, the character likelihood evaluation value φ _char of the character candidate area between the character cutout position candidates corresponding to the elements a _k and a _j .
Further, when b _i = b ₁ , it is determined that Φ = 0.

Since this Φ (a _j , b _i ) is a recurrence formula of b _i , starting from the coordinates (a ₁ , b ₁ ) (that is, the origin) to the coordinates (a ₆ , b ₆ ), Φ (a _j , b _i ) can be determined sequentially. In addition, when Φ (a _j , b _i ) is determined, the coordinates (a _k , b _p ) + φ _(i) (a _k , b _p , a _j , b _i ) that maximize Φ (a _k , b _p ) + φ _(i) b _p ) is stored in association with the coordinates (a _j , b _i ). Specifically, a pointer from the coordinates (a _j , b _i ) to the coordinates (a _k , b _p ) is stored in the storage device.

By such a method, Φ (a _j , b _i ) is calculated for each combination of j and i corresponding to the vehicle type code area arrangement candidate determined by the above-described processing. Using the calculated Φ (a _j , b _i ), an optimal correspondence graph, that is, an optimal vehicle type code area arrangement candidate is obtained for each of the 2-character vehicle type code and the 3-character vehicle type code by the following method. It is determined. In the determination of the corresponding graph corresponding to two-letter models code, _{first, Φ (a k, b 4} ) the maximum value of is determined, the corresponding graph [Phi the (a _{k, b 4)} to the maximum, optimum response Selected as a graph. The correspondence graph that maximizes Φ (a _k , b ₄ ) can be determined from the pointers described above. Similarly, in the determination of the optimal response graph corresponding to the three letter models code, _{first, Φ (a k, b 6} ) the maximum value of is determined, the corresponding graph to maximize [Phi the (a _{k, b 6)} Is selected as the optimal correspondence graph. The correspondence graph that maximizes Φ (a _k , b ₆ ) can be determined from the pointers described above.

  As shown in FIG. 5, after the optimal vehicle type code area arrangement candidate for the 3-character vehicle type code and the optimal vehicle type code area arrangement candidate for the 2-character vehicle code are determined, the vehicle type code area The number of characters of 13, that is, which vehicle type code area arrangement candidate is optimal is finally determined (step S13-5).

  In the determination of the optimal vehicle type code area arrangement candidate in step S13-5, first, each of the optimal vehicle type code area arrangement candidate for the character vehicle type code and the optimal vehicle type code area arrangement candidate for the 2-character vehicle type code, respectively. Is determined to be valid. Specifically, the following four items are checked for two vehicle type code area arrangement candidates, and if any one of them corresponds, it is determined that the vehicle type code area arrangement candidate is invalid.

ｎは、対象の車種コード領域配置候補の文字数である。 (1) Character candidate areas with extremely small character-likeness evaluation values exist With reference to FIG. 19A, a first character candidate area 21, a second character candidate area 22, and a third character candidate area 23 if present. In addition, when the character evaluation value φ _char is extremely small, it is determined that the vehicle type code area arrangement candidate is invalid. Specifically, when the ratio of the character likelihood evaluation value φ ⁱ _char of the i-th character candidate region to the character-like average value φ _{mean_char} is smaller than a predetermined value, the character likelihood evaluation value φ ⁱ _{char of} the i-th character candidate region is extremely It is determined that the corresponding vehicle type code area arrangement candidate is invalid; the average value φ _{mean_char} is a value defined by the following formula:

n is the number of characters of the target vehicle type code area arrangement candidate.

More specifically, it is determined that the character-likeness evaluation value φ ⁱ _{char of} the i-th character candidate region is extremely small when the following formula is satisfied:
φ ⁱ _char <α · φ _{mean_char} , (11)
Here, α is a predetermined parameter smaller than 1. In the example of FIG. 19, since the character likelihood evaluation value φ ⁱ _{char of} the second character candidate area 22 is extremely small, it is determined that the vehicle type code area arrangement candidate is invalid.

(2) Poor character width balance Referring to FIG. 19B, character width of first character candidate region 21, second character candidate region 22, and, if present, third character candidate region 23 is poorly balanced. Is determined to be invalid. The balance of the character width is determined based on the ratio between the maximum value of the character width (that is, the width of the character candidate area) and the minimum value of the character width. More specifically, a vehicle type code arrangement candidate corresponding to a two-character vehicle type code is determined to have a poor character width balance when the following conditions are satisfied:
W _MAX / W _MIN > β, (12a)
W _MAX = max (W ₁ , W ₂ ), (12b)
W _MIN = min (W ₁ , W ₂ ), (12c)
Here, _Wi is the width of the i-th character candidate region, and β is a predetermined parameter larger than 1.

Similarly, for the vehicle type code arrangement candidate corresponding to the 3-character vehicle type code, it is determined that the balance of the character width is bad when the following conditions are satisfied:
W _MAX / W _MIN > β, ... (12d)
W _MAX = max (W ₁ , W ₂ , W ₃ ), (12e)
W _MIN = min (W ₁ , W ₂ , W ₃ ). ... (12f)

  For example, in the example of FIG. 19B, it is determined that the width of the first character candidate region 21 is extremely narrow compared to the width of the second character candidate region 22, and the balance of character width is poor.

(3) An extremely large gap exists compared to the character width Referring to FIG. 19C, when the gap between a pair of character candidate areas is extremely large compared to the character width of each character candidate area Is determined that the vehicle type code area arrangement candidate is invalid. The balance between the character width and the gap width is determined based on the ratio between the maximum gap width value and the minimum character width value.

More specifically, a vehicle type code arrangement candidate corresponding to a two-character vehicle type code is determined to have a poor character width balance when the following conditions are satisfied:
W ¹² _gap / W _MIN > γ, (13a)
W _MIN = min (W ₁ , W ₂ ), (13b)
Here, W ¹² _gap is the width of the gap between the first character candidate region 21 and the second character candidate region 22, and γ is a predetermined parameter. It should be noted that there is only one gap for the vehicle type code arrangement candidate corresponding to the 2-character vehicle type code.

On the other hand, for vehicle type code arrangement candidates corresponding to the 3-character vehicle type code, it is determined that the balance of the character width is poor when the following conditions are satisfied:
W _{G — MAX} / W _MIN > γ, (13c)
^{_{W G_MAX = max (W 12 gap}} , W 23 gap), ··· (13d)
W _MIN = min (W ₁ , W ₂ , W ₃ ), (13e)
Here, W ^ij _gap is the width of the gap between the i-th character candidate region and the j-th character candidate region.

For example, in the example of FIG. 19C, it is determined that the gap between the first character candidate region 21 and the second character candidate region 22 is extremely larger than the width W ₁ (= W _MIN ) of the first character candidate region 21. Is done.

(4) An image with a large character-likeness evaluation value exists in the gap area Referring to FIG. 19D, if an image that seems to be a character exists in a gap between a pair of character candidate areas, the vehicle type code It is determined that the region arrangement candidate is invalid. First, for each of the gaps between character candidate areas, a character likelihood evaluation value is calculated by the same processing as that for the character candidate areas. When the character likelihood evaluation value of the gap is large enough to be compared with the above-described character likelihood average value φ _{mean_char} , it is determined that the vehicle type code area arrangement candidate is invalid. More specifically, when the gap between the i-th character candidate area and the j-th character candidate area satisfies the following expression, it is determined that an image having a high character character exists in the gap area:
φ ^ij _{gap_char} > δ · φ _{mean_char} , (14)
Here, φ ^ij _{gap_char} is a character likelihood evaluation value calculated for the gap between the i-th character candidate region and the j-th character candidate region, and δ is a predetermined parameter.

  As a result of checking the above four items, it is determined that only one of the optimal vehicle type code area arrangement candidate for the 3-character vehicle type code and the optimal vehicle type code area arrangement candidate for the 2-character vehicle type code is valid. If it is determined, the effective vehicle type code area arrangement candidate is finally determined to be the optimum vehicle type code area arrangement candidate.

  When it is determined that both the vehicle type code area arrangement candidates are invalid, it is determined that the character recognition of the vehicle type code area 13 has failed.

If it is determined that both the vehicle type code area arrangement candidates are valid, as shown in FIG. 20, the above-described vehicle type code area arrangement candidate having a large character-like average value φ _{mean_char} is finally the optimum vehicle type. It is determined that it is a code area arrangement candidate. For example, in the example of FIG. 23, it is _determined that the optimal vehicle type code area arrangement candidate of the two-character vehicle type code having a large character- _like average value φ _{mean_char} is finally the optimal vehicle type code area arrangement candidate.

  A numeric string consisting of the numbers recognized for each character candidate area of the finally determined vehicle type code area arrangement candidate is determined as the vehicle code to be finally obtained, and finally the vehicle type code area arrangement determined to be optimal The number of character candidate areas included in the candidates is determined as the number of characters in the vehicle type code area 13.

  One advantage of the vehicle type code recognition algorithm of the present embodiment described above is that the amount of data processing required for character recognition can be reduced. This is mainly achieved by performing character recognition only on the first character region candidate 21 to the third character region candidate 23 determined from the character cutout position candidate 18, not the entire vehicle type code region 13. At this time, if DP matching is adopted, an optimal vehicle type code area arrangement can be determined with a smaller amount of calculation, and character recognition of each character can be performed.

  Another advantage is that the character recognition of the vehicle type code can be appropriately performed regardless of whether the recognition target is a two-character vehicle type code or a three-character vehicle type code. This advantage is that the first character region candidate 21 to the third character region candidate 23 with the character cutout position candidate 18 extracted over the entire vehicle type code region 13 as a boundary are determined, and the vehicle type code region that is a combination thereof. It is obtained by outputting a character string corresponding to an optimal vehicle type code region arrangement candidate among the arrangement candidates as a recognition result.

  As described above, the vehicle type code recognition algorithm according to the present embodiment can achieve both reduction in the amount of data processing necessary for character recognition of the vehicle type code and coping with the diversity of character strings.

Second Second Embodiment FIG. 21 is a flowchart showing an algorithm for character recognition of the lowercase sequence image 11 in the second embodiment. The most important difference between the algorithm of the second embodiment and the algorithm of the first embodiment is that in the second embodiment, land code recognition and vehicle type code recognition are performed simultaneously. As in the first embodiment, the determination of the vehicle type code region 13 and the recognition of the vehicle type code are not performed after the determination of the land support code region 12 and the recognition of the land support code.

More specifically, first, the character cutout position candidate 18 is extracted (step S31). The procedure for extracting the character cutout position candidate 18 is as described in the first embodiment. As illustrated in FIG. 22, the following description will be made assuming that ₁₁ character cutout position candidates 18 ₁ to 18 ₁₁ have been determined.

  Subsequently, region arrangement candidates are set for the entire lowercase character string image 11 (step S32). The region arrangement candidate is a candidate for arrangement of the land code area 12, the first character area 14, the second character area 15, and the third character area 16 in the lowercase character string image 11. However, in the present embodiment, the land code region 12 is treated as one character region; that is, in the present embodiment, the lowercase character sequence image 11 is assumed to include three or four character regions. Character recognition is performed. As described above, it should be noted that the land code is handled as one character in the image processing apparatus 2.

  Since the vehicle type code is composed of two or three characters, two types of region arrangement candidates corresponding to the two-character vehicle type code and region arrangement candidates corresponding to the three-character vehicle type code are determined. As shown in FIG. 22, the region arrangement candidate corresponding to the three-character vehicle type code is expressed by a combination of the first to fourth character candidate regions 21 to 24. In the present embodiment, the first character candidate area 21 is a character area candidate in which the first character of the lowercase character string image 11 exists, that is, a candidate for the land code area 12, and the second character candidate area 22 Is a candidate for the first character region 14 in which the first character of the vehicle type code exists, and the third character candidate region 23 is a candidate for the second character region 15 in which the second character of the vehicle type code exists. The fourth character candidate area 24 is a candidate for the third character area 16 in which the third character of the vehicle type code exists. Both ends of the character candidate areas 21 to 24 are defined by two character cutout position candidates 18. Similarly, the vehicle type code region arrangement candidate corresponding to the two-character vehicle type code is expressed by a combination of the first character candidate region 21, the second character candidate region 22, and the third character candidate region 23.

Specifically, the region arrangement candidate is set by associating the character cutout position candidate 18 with the start point (left end) and end point (right end) of the character candidate regions 21 to 24. As an example, let us consider a case where there are ₁₁ character cutout position candidates 18 ₁ to 18 ₁₁ for the positions “1” to “11” of the vehicle type code area 13 as shown in FIG. For example, in a certain area arrangement candidate, the character cutout position candidates 18 ₁ and 18 ₅ are respectively associated with the start point (left end) and the end point (right end) of the first character candidate area 21, and the character cutout position candidates 18 ₆ , 18 ₇ are respectively associated with the start and end points of the second character candidate region 22. Furthermore, the character cutout position candidates 18 ₇ and 18 ₈ are associated with the start point and the end point of the third character candidate area 23, respectively, and the character cutout position candidates 18 ₈ and 18 ₉ are the start point and the end point of the fourth character candidate area 24. Respectively. Different associations are performed in other region arrangement candidates.

  However, as in the first embodiment, an area arrangement candidate that includes a character candidate area that has an invalid width as the width of the character area is not set. Specifically, the first character candidate area 21 corresponding to the land code area 12 is determined so as to satisfy the expression (1), and the remaining character candidate areas 22 to 24 are determined according to the expressions (2a) and (2b). To be satisfied. This is effective for reducing the number of region arrangement candidates and thereby reducing the amount of calculation.

Returning to FIG. 21, following the setting of the region arrangement candidates, the evaluation value φ _string of each region distribution candidate is calculated (S34). The procedure for calculating the evaluation value φ _string of each area allocation candidate is as follows: (1) the number of character area candidates to be recognized is different, and (2) the character recognition of the first character candidate area 21 is for vehicle type code. The land code template 7 is used instead of the template 8, and (3) the parameters used for calculating the character width deduction point P ⁱ _char : W _{min_char} and W _{max_char} are the first character candidate area 21 and the remaining Except for the difference between the character candidate areas 22 to 24, the procedure is the same as the procedure for calculating the evaluation value φ _string of the vehicle type code area distribution candidate performed in the first embodiment.

  Further, the optimum region allocation candidate is determined for each of the 2-character vehicle type code and the 3-character vehicle type code (step S34), and the number of characters of the vehicle type code, that is, the optimal region allocation candidate is finally determined from these two region allocation candidates. (Step S35). The land code, the vehicle type code, and the number of characters corresponding to the area allocation candidate finally determined as the optimum area allocation candidate are output as the recognition result.

  Although the character recognition algorithm used in the second embodiment has a larger amount of computation than the character recognition algorithm of the first embodiment, it can increase the character recognition accuracy. In the first embodiment, when the position of the land support code (that is, the land support code area 12) cannot be determined optimally, the accuracy of character recognition of the vehicle type code is lowered. On the other hand, in the second embodiment, since the land code area 12 and the character areas 14 to 16 can be determined optimally as the whole lowercase character string image 11, the character recognition accuracy can be improved.

Third Embodiment In the third embodiment, as shown in FIG. 23, in addition to the character cutout position candidate 18 in the horizontal direction, the character cutout position candidate 25 in the vertical direction is extracted. From the cutout position candidates 18 and 25, the land code candidate area 19 and the first to third character candidate areas 21 to 23 are determined. In order to clearly distinguish the horizontal and vertical cutout position candidates 18 and 25, in the following, the horizontal character cutout position candidates will be referred to as the horizontal character cutout position candidate 18 and the vertical character cutout position candidates as vertical. It will be called a directional character cutout position candidate 25. The flow of character recognition in the third embodiment is roughly the same as the flow shown in FIG. 5, but is performed at each step as the vertical character cutout position candidate 25 is extracted. The processing that is performed is changed.

Specifically, in the process of extracting character cutout position candidates (step S11), in addition to the horizontal character cutout position candidate 18, a vertical character cutout position candidate 25 is extracted. FIG. 24 is a conceptual diagram showing a processing procedure for extracting character cutout position candidates in the present embodiment. First, the horizontal character cutout position candidate 18 is extracted by the same procedure as in the first embodiment. Subsequently, for each set of adjacent horizontal character cutout position candidates 18, a horizontal luminance projection of an area sandwiched between adjacent horizontal character cutout position candidates 18 is calculated. The horizontal luminance projection of a region is the sum of the luminance values calculated for each row of pixels located in that region; where a pixel row is the horizontal alignment of pixels arranged in one column It is a line. The luminance projection F (y) of the pixel row located at the coordinate x is
F (y) = Σσ (x, y),
Is calculated by Here, σ (x, y) is the luminance (or gradation) of the pixel located at the coordinate (x, y) of the region, and Σ represents the sum of the entire range of the x coordinate of the region. Yes. Subsequently, a position where the horizontal luminance projection F (y) is minimized is detected. Further, out of the positions where the luminance projection F (y) in the horizontal direction is minimized, for example, a total of two positions one by one at the top and bottom, and the upper end and the lower end of the lowercase character string image 11 are extracted as the vertical character cutout position candidates 25. The

As shown in FIG. 5, following the extraction of the character cutout position candidates, the land code candidate area 19 is set (step S12-1). 25A to 25E are conceptual diagrams showing a procedure for determining the land code candidate area 19 in the present embodiment. Referring to FIG. 25A, in the determination of land support code candidate area 19, first, two horizontal character cutout position candidates 18 are selected, whereby the start point (left end) and end point (right end) of land support code candidate area 19 are selected. Positions x _L and x _R are determined.

Subsequently, the left lateral character segmentation position candidate 18 (i.e., lateral character segmentation position candidate 18 corresponding to the starting point x _L) is in contact with the, and the vertical direction character closer to the upper than the lower end of the lower case string image 11 A cutout position candidate 25 is selected, and a contact point between the left side character cutout position candidate 18 on the left side and the selected vertical character cutout position candidate 25 is selected as the top left vertex of the land code candidate area 19; The top left vertex is indicated by a circle in FIG. 25A.

Furthermore, the right lateral character segmentation position candidate 18 (i.e., lateral character segmentation position candidate 18 corresponding to the end point x _R) is in contact with the, and, cut longitudinally character closer to the lower than the upper end of the lower case string image 11 The position candidate 25 is selected, and the contact point between the right lateral character cutout position candidate 18 and the selected vertical character cutout position candidate 25 is selected as the lower right vertex of the land code candidate area 19; The lower right vertex is indicated by an inverted triangle in FIG. 25A.

  A rectangular area defined by the selected upper left and lower right vertices is set as the land support code candidate area 19. Land support code candidate areas 19 are set for all of the combinations of two allowable horizontal character cutout position candidates 18 and all of the vertical character cutout position candidates 25 allowed for each combination of the horizontal character cutout position candidates 18. Is done. For example, four types of land code candidate areas 19 shown in FIGS. 25B to 25E can be set.

  After the land support code candidate area 19 is set, character recognition is performed for each land support code candidate area 19 (step S12-2). The optimal land support code candidate area 19 is determined as the land support code area 12, and the land support code recognized for the optimal land support code candidate area 19 is output as a recognition result.

  Subsequently, a vehicle type code region 13 is determined from the position of the land support code region 12 (S13-1), and a vehicle type code distribution candidate is set for the vehicle type code region 13 (S13-2). As shown in FIG. 26, one vehicle type code distribution candidate is represented by three character candidate areas 21 to 23 as in the first embodiment. The first to third character candidate areas 21 to 23 are the horizontal character cutout position candidate 18 and the vertical character cutout position candidate 25 belonging to the vehicle type code area 13, the left end of the first to third character candidate areas 21 to 23, It is set by associating with the right end, upper end, and lower end. The left and right ends of each of the character candidate areas 21 to 23 are determined by a pair of horizontal character cutout position candidates 18, and the upper and lower ends are determined by a pair of vertical character cutout position candidates 25.

Further, an evaluation value φ _string is calculated for each vehicle type code distribution candidate (step S13-4), and an optimal vehicle type code distribution candidate is determined for each of the two-character vehicle type code and the three-character vehicle type code (step S13-). 5). The calculation method of the evaluation value φ _string is the same as that of the first embodiment.

  Further, according to the same procedure as in the first embodiment, one of the optimal vehicle type code distribution candidates of the 2-character vehicle type code and the 3-character vehicle type code is determined as the optimal vehicle type code distribution candidate (step S13). -3). A numeric string consisting of the numbers recognized for each character candidate area of the finally determined vehicle type code area arrangement candidate is determined as the vehicle code to be finally obtained, and finally the vehicle type code area arrangement determined to be optimal The number of character candidate areas included in the candidates is determined as the number of characters in the vehicle type code area 13.

  In the present embodiment, in addition to the horizontal character cutout position candidate 18, the vertical character cutout position candidate 25 is used for setting the first to third character candidate areas 21 to 23, so the number of vehicle type code distribution candidates increases. To do. An increase in the number of vehicle type code allocation candidates is not preferable because the data processing amount is increased.

In order to suppress an increase in the amount of data processing due to an increase in the number of vehicle type code distribution candidates, it is preferable to use DP matching as in the first embodiment. Referring to FIG. 26, in the DP matching according to the present embodiment, vehicle type code area arrangement candidates (that is, horizontal character cutout position candidate 18 and vertical character cutout position candidate 25 and first to third character candidate areas 21 to 21 are included. 23) is the A axis corresponding to the horizontal character cutout position candidate 18, the B axis corresponding to the start and end points of the first to third character candidate areas 21 to 23, and the vertical character cutout position candidate. 25 is expressed as a correspondence graph developed on a three-dimensional coordinate system composed of the C-axis corresponding to 25. In the example of FIG. 26, the elements a _{1 to} a ₄ on the A axis correspond to the horizontal character cutout position candidates 18 ₁ to 18 ₄ , and the elements b _{1 to} b ₆ on the B axis are the first characters. Corresponding to the start and end points of the candidate area 21, the second character candidate area 22, and the third character candidate area 23, the elements c _{1 to} c ₆ on the C axis are the vertical character cut position candidates 25 _{1 to} 25 _6. It corresponds to. The numbers of the elements a _j on the A axis and the elements _cm on the C axis are adjusted according to the numbers of the vertical character cutout position candidates 18 and the vertical character cutout position candidates 25 belonging to the vehicle type code area 13. Please note that. Each correspondence graph includes a route from the origin (a ₁ , b ₁ , c _start ) to the coordinates (a _end , b ₄ , c _end ) (when corresponding to a two-character vehicle type code), or the origin (a ₁ , b ₁ , c _start ) to coordinates (a _end , b ₆ , c _end ); where c _start is an element on the C axis at the start point of the corresponding graph, and a _end , c _end are The horizontal character cutout position candidate 18 and the vertical character cutout position candidate which are elements on the A axis and C axis of the end point of the corresponding graph and selected as the end point of the second character candidate area 22 or the third character candidate area 23 This is an element corresponding to 25.

Corresponding to the correspondence graph being represented on three-dimensional coordinates, the method of calculating the evaluation value φs _string using the equation (8) is also expanded to three dimensions. In particular,
The maximum value of the evaluation value φ _string of the vehicle type code area arrangement candidate corresponding to the correspondence graph from the _start point (a ₁ , b ₁ , c _start ) of the correspondence graph to the coordinates (a _j , b _i , c _n ) is represented by Φ (a _j , b _i , c _n )
Φ (a _j , b _i , c _n ) = MAX (a _k , b _p , c _m ) [Φ (a _k , b _p , c _m )
+ Φ _(i) (a _k , b _p , c _m , a _j , b _i , c _n )]], (8) ′
Is established. Here, _{b p} is the previous element of the _{b i,}
b _p = b _i−1 , (9)
It is. _{Furthermore, MAX (a k, b p} , c m) [X (a k, b p, c m)] _{is, a k, b p, c} m X (a k when changing _{the, b} p, c _m ) represents the maximum value. _{Further, φ (i) (a k} , b p, c m, a j, b i, c n) is
(Α) When i is an odd number, the gap likelihood formed between the character cutout position candidate corresponding to the element (a _k , c _m ) and the character cutout position candidate corresponding to (a _j , c _n ) Evaluation value φ _gap
(Β) When i is an even number, a character candidate region formed between a character cutout position candidate corresponding to the element (a _k , c _m ) and a character cutout position candidate corresponding to (a _j , c _n ) Characteristic evaluation value φ _char .
Further, when b _i = b ₁ , it is determined that Φ = 0.

Since Φ (a _j , b _i , c _n ) is a recurrence formula of b _i , the coordinates _{start from} the coordinates (a ₁ , b ₁ , c _start ) (that is, the start point of the corresponding graph). Up to (a ₆ , b ₆ , c _end ), Φ (a _j , b _i , c _n ) can be sequentially determined. In addition, when determining Φ (a _j , b _i , c _n ), Φ (a _k , b _p , c _m ) + φ _(i) (a _k , b _p , _cm , a _j , b _i , Coordinates (a _k , b _p , c _m ) that maximize c _n ) are stored in association with the coordinates (a _j , b _i ). Specifically, pointers from the coordinates (a _j , b _i , c _n ) to the coordinates (a _k , b _p , c _m ) are stored in the storage device.

By such a method, Φ (a _j , b _i , c _n ) is calculated for each combination of j, i, n corresponding to the vehicle type code area arrangement candidate. Using the calculated Φ (a _j , b _i , c _n ), an optimal correspondence graph, that is, an optimal vehicle type code area arrangement candidate is a 2-character vehicle type code, a 3-character vehicle type code, by the following method. For each of them. In the determination of the corresponding graph corresponding to two-letter models code, first, [Phi maximum value of _{(a k, b 4, c} end) is obtained, corresponding to [Phi the _{(a k, b 4, c} end) to the maximum The graph is selected as the optimal correspondence graph. A correspondence graph that maximizes Φ (a _k , b ₄ , c _end ) can be determined from the pointers described above. Similarly, in the determination of the optimal response graph corresponding to the three letter models code, first, [Phi maximum value of _{(a k, b 6, c} end) is _{obtained, Φ (a k, b 6} , c end) The corresponding graph that maximizes is selected as the optimal corresponding graph. The correspondence graph that maximizes Φ (a _k , b ₆ , c _end ) can be determined from the pointers described above.

  In the character recognition processing method according to the third embodiment, since the vertical character cutout position is set in addition to the horizontal character cutout position, the character is slightly tilted or there is a gap above and below the character. Also, it is possible to set the character candidate area correctly.

  Although the embodiments of the present invention have been described in detail above, the present invention should not be construed as being limited to those described in the embodiments. Although the present invention is particularly suitable for license plate recognition, it will be obvious that it can also be used for recognition of other character strings by using appropriate template images.

FIG. 1 is a block diagram showing a configuration of a license plate recognition apparatus according to the first embodiment of the present invention. FIG. 2 is a flowchart showing the operation of the license plate recognition apparatus according to the present embodiment. FIG. 3 is a conceptual diagram showing a license plate image and a lowercase letter row image. FIG. 4 is a conceptual diagram showing the configuration of a lowercase string image. FIG. 5 is a flowchart showing a character recognition processing procedure performed for a lowercase string image. FIG. 6A is a conceptual diagram illustrating a procedure for determining a character cutout position candidate. FIG. 6B is a conceptual diagram illustrating a procedure for determining a character extraction position candidate. FIG. 6C is a conceptual diagram illustrating a preferred procedure for determining character cutout position candidates. FIG. 7 is a conceptual diagram showing a preferred procedure for determining a land support code candidate area. FIG. 8 is a conceptual diagram showing a procedure for selecting an optimum land code code area as a land code area. FIG. 9 is a conceptual diagram showing a procedure for determining the vehicle type code area from the land support code area. FIG. 10 is a conceptual diagram illustrating character candidate areas and vehicle code area arrangement candidates determined as vehicle type code areas. FIG. 11 is another conceptual diagram illustrating character candidate areas and vehicle code area arrangement candidates determined as vehicle type code areas. FIG. 12 is a conceptual diagram showing character candidate areas that can be considered in the present embodiment. FIG. 13 is a conceptual diagram illustrating a method for determining a character candidate area in the present embodiment. FIG. 14 is a conceptual diagram illustrating the procedure of character recognition performed for each character candidate region. FIG. 15 is a conceptual diagram illustrating a procedure for calculating the character width deduction point P ⁱ _char . FIG. 16 is a conceptual diagram illustrating an example of misrecognition that may occur in the present embodiment. FIG. 17 is a flowchart showing a preferred processing procedure for preventing erroneous recognition of a character candidate area. FIG. 18 is a conceptual diagram illustrating a procedure for calculating a clearance likelihood evaluation value. FIG. 19A is a conceptual diagram illustrating a procedure for determining the validity of a vehicle type code area arrangement candidate corresponding to a 2-character, 3-character vehicle type code. FIG. 19B is a conceptual diagram illustrating a procedure for determining the validity of a vehicle type code area arrangement candidate corresponding to a 2-character, 3-character vehicle type code. FIG. 19C is a conceptual diagram illustrating a procedure for determining the validity of a vehicle type code area arrangement candidate corresponding to a 2-character, 3-character vehicle type code. FIG. 19D is a conceptual diagram illustrating a procedure for determining the validity of a vehicle type code area arrangement candidate corresponding to a 2-character, 3-character vehicle type code. FIG. 20 is a conceptual diagram illustrating a procedure for selecting an optimal vehicle type code area arrangement candidate from vehicle type code area arrangement candidates corresponding to 2-character and 3-character vehicle type codes. FIG. 21 is a flowchart for explaining character recognition processing of a lowercase string image according to the second embodiment. FIG. 22 is a conceptual diagram illustrating a method for determining a character candidate area in the second embodiment. FIG. 23 is a conceptual diagram illustrating horizontal character cutout position candidates and vertical character cutout positions defined in the third embodiment. FIG. 24 is a conceptual diagram illustrating a processing procedure for determining a horizontal character cutout position candidate and a vertical character cutout position. FIG. 25A is a conceptual diagram illustrating a procedure for determining a land support code candidate area in the third embodiment. FIG. 25B is a conceptual diagram illustrating a procedure for determining a land support code candidate region in the third embodiment. FIG. 25C is a conceptual diagram illustrating a procedure for determining a land support code candidate region in the third embodiment. FIG. 25D is a conceptual diagram illustrating a procedure for determining a land support code candidate region in the third embodiment. FIG. 25E is a conceptual diagram illustrating a procedure for determining land code candidate regions in the third embodiment. FIG. 26 is a conceptual diagram illustrating a method for determining a character candidate area in the third embodiment.

Explanation of symbols

1: Image pickup device 2: Image processing device 3: Storage device 4: Computing device 5: Interface 6: License plate recognition program 7: Land code template 8: Vehicle type code template 10: License plate recognition device 11: Lower case image 12: Land support code area 13: Vehicle type code area 14: First character area (character area)
15: 2nd character area 16: 3rd character area 17: Character cut-out positions 18, 18 ₁ , 18 ₂ , 18 ₃ , 18 ₅ : Character cut-out position candidates (horizontal character cut-out position candidates)
19, 19 ₁ , 19 ₄ : Land support code candidate area 21: First character candidate area (character candidate area)
22: Second character candidate area 23: Third character candidate area 24: Fourth character candidate area 25: Vertical character cutout position candidate

Claims (16)

(A) determining a character cutout position candidate that is a candidate for the boundary of the character region for the entire character string image including a plurality of character regions in which one character is copied;
(B) A character candidate region that is a candidate for the plurality of character regions is determined by associating the character cut-out position candidates with boundaries of the plurality of character regions, and region arrangement candidates expressed as combinations of the character candidate regions are determined. The steps to decide;
(C) performing image recognition for each of the character candidate regions, and selecting an optimal region placement candidate from the region placement candidates based on the result of the image recognition;
(D) A step of outputting a character string made up of characters recognized in the image recognition in the step (C) for the character candidate region corresponding to the optimal region arrangement candidate as a recognition result character string. Method. The character string recognition method according to claim 1,
The character cutout recognition method is a character string recognition method in which the character cutout position candidate is determined from the luminance of each pixel of the character string image. The character string recognition method according to claim 1,
The character cutout position candidate is determined from a differential image of the character string image. The character string recognition method according to claim 1,
The step (C) includes:
(C1) calculating a width of each of the character candidate areas defined by the character cutout position candidates;
(C2) calculating a difference between the image of the character candidate area and a template image prepared in advance for each of the character candidate areas defined by the character cutout position candidates;
(C3) calculating a character-likeness evaluation value based on the width of the character candidate region and the degree of difference,
The optimum region arrangement candidate is a character string recognition method that is determined based on the character likelihood evaluation value. The character string recognition method according to claim 1,
The step (C) further includes:
(C4) calculating a width of a gap between two adjacent character candidate areas of the character candidate area defined by the character cutout position candidates;
(C5) comprising a step of calculating a clearance likelihood evaluation value based on the width of the gap between the two character candidate regions,
The optimum region arrangement candidate is a character string recognition method that is determined based on the gap likelihood evaluation value. (E) determining a character extraction position candidate that is a candidate for a boundary of a character region in which one character of the lowercase character string is copied for each of the lowercase character string images in which the lowercase character string of the license plate is copied; ,
(F) associating the character cutout position candidates with the boundaries of the plurality of character regions, determining character candidate regions that are candidates for the plurality of character regions, and determining region arrangement candidates expressed as combinations of the character candidate regions The steps to decide;
(G) performing image recognition for each of the character candidate regions, and selecting an optimal region placement candidate from the region placement candidates based on the result of the image recognition;
(H) a step of outputting, as at least part of a recognition result, a character string made up of characters recognized in the image recognition in the step (G) for the character candidate region corresponding to the optimum region arrangement candidate. Plate recognition method. The license plate recognition method according to claim 6,
Furthermore,
(I) determining a land code area where a land code of the lowercase character string is copied from the character cutout position candidates;
(J) determining a vehicle type code area in an area other than the land support code area in the lower case sequence image,
The region arrangement candidate is a vehicle type code region arrangement candidate defined for the vehicle type code region. The character string recognition method according to claim 7,
The step (G) includes:
(G1) calculating a width of each of the character candidate areas defined by the character cutout position candidates;
(G2) calculating a difference between the image of the character candidate area and a template image prepared in advance for each of the character candidate areas defined by the character cutout position candidates;
(G3) calculating a character likelihood evaluation value based on the width of the character candidate region and the degree of difference,
The optimum region arrangement candidate is a character string recognition method that is determined based on the character likelihood evaluation value. The character string recognition method according to claim 8,
The step (G) further includes:
(G4) calculating a width of a gap between two adjacent character candidate areas of the character candidate area defined by the character cutout position candidates;
(G5) comprising a step of calculating a clearance likelihood evaluation value based on the width of the gap between the two character candidate regions,
The optimum region arrangement candidate is a character string recognition method that is determined based on the gap likelihood evaluation value. The license plate recognition method according to claim 7,
The step (I) includes:
(I1) a step of defining an area having a predetermined range of width as a land support code candidate area with the character cutout position candidate as a boundary, in contact with an end of the lowercase character string image, and
(I2) performing image recognition for each of the land support code candidate areas, and determining a land support code area in which the land support code is copied from the land support code candidate areas based on the result of the image recognition And
A license plate recognition method in which a land support code recognized for the land support code candidate region determined as the land support code region is output as a part of the recognition result. The license plate recognition method according to claim 6,
In the image recognition in the step (G), a land code code template image is used for the character candidate region corresponding to the leftmost character in the lowercase character string, and the character candidates corresponding to other characters are used. A number template image is used for the area,
The recognition result includes a recognition result of a land code of the license plate and a recognition result of a vehicle type code of the license plate. The license plate recognition method according to claim 6,
When the recognition result of “1” is obtained by the image recognition performed for a certain character candidate region in the character candidate region in the step (G), the surroundings of the certain character candidate region are included. When image recognition is performed again for the enlarged area and a recognition result other than “1” is obtained for the enlarged area, the area arrangement candidate including the certain character candidate area is selected as the optimum area arrangement candidate. No. License plate recognition method. The license plate recognition method according to claim 6,
The character candidate area is a rectangle,
The character cutout position candidates are
A horizontal character cutout position candidate for determining the left end and the right end of the character candidate area;
A license plate recognition method comprising: a vertical character cutout position candidate for determining an upper end and a lower end of the character candidate region. The license plate recognition method according to claim 6,
The step (E)
(E1) dividing the lower-case string image into an upper part and a lower part;
(E2) A license plate recognition method comprising: determining the character cutout position candidates from the image of the lower part. An imaging device for acquiring a character string image including a plurality of character regions each of which one character is copied;
An image processing device that performs character recognition of the character string image,
The image processing apparatus includes the following steps:
(A) a step of determining a character cutout position candidate that is a candidate for the boundary of the character region for the entire character string image;
(B) A character candidate region that is a candidate for the plurality of character regions is determined by associating the character cut-out position candidates with boundaries of the plurality of character regions, and region arrangement candidates expressed as combinations of the character candidate regions are determined. Steps to determine,
(C) performing image recognition for each of the character candidate areas, and selecting an optimum area arrangement candidate from the area arrangement candidates based on the result of the image recognition; and (D) selecting the optimum area arrangement candidate A character string recognition device configured to execute a step of outputting, as a recognition result character string, a character string composed of characters recognized in the image recognition in step (C) for the corresponding character candidate region. An imaging device for obtaining a captured image of a license plate;
An image processing device for performing character recognition of the license plate from the captured image,
The image processing apparatus includes the following steps:
(K) a step of cutting out from the captured image a lower case sequence image in which a lower case sequence of the license plate is copied;
(E) 'determining a character cutout position candidate that is a candidate for a boundary of a character region in which one character of the lowercase character string is copied for each of the lowercase character string images;
(F) associating the character cutout position candidates with the boundaries of the plurality of character regions, determining character candidate regions that are candidates for the plurality of character regions, and determining region arrangement candidates expressed as combinations of the character candidate regions Steps to determine,
(G) performing image recognition for each of the character candidate areas, and selecting an optimum area arrangement candidate from the area arrangement candidates based on the result of the image recognition; and (H) selecting the optimum area arrangement candidate. A license plate recognition device configured to execute a step of outputting, as at least a part of a recognition result, a character string made up of characters recognized in the image recognition of step (F) for the corresponding character candidate region.

Images