JP5439069B2 - Character recognition device and character recognition method - Google Patents

Description

  The present invention relates to, for example, a character recognition device that recognizes a character string such as a vehicle number, and in particular, even for a character string in which adjacent characters are connected and it is difficult to correctly cut out each character region. The present invention relates to a character recognition device capable of recognizing characters with high accuracy.

  2. Description of the Related Art Conventionally, there has been known an apparatus for automatically recognizing description contents such as a vehicle number described on a license plate of a vehicle. In order to recognize the description content of the license plate, a character recognition method using an image processing technique is used. In such a character recognition method, generally, after inputting a target character image, the character line is emphasized (binarization) to separate the character and the background, and an area that can be determined to be a character is extracted ( Character cutout process). In the region where the character cut-out processing has been performed, the character shape feature (for example, the contour direction feature) is extracted, and the character is recognized by pattern recognition. Here, in the character cut-out process, a plurality of characters included in the character image to be recognized and peripheral structures such as bolts for fixing the license plate are separated from each other, and between each character or between the character and the peripheral structure. It is assumed that there is a certain gap between them.

By the way, the license plate has a different shape, size, written character string, and the like for each country. That is, for example, the characters included in the character image to be recognized are close to each other, the characters appear to be connected to each other, or a peripheral structure such as a bolt that fixes the license plate to the vehicle and the characters are close to each other. There are some license plates that appear to be connected to surrounding structures. In such a case, the conventional character recognition device cannot perform a correct character cut-out process, and a correct recognition result cannot be obtained.
Therefore, Patent Document 1 discloses a technique for maintaining the recognition accuracy even when a character string to be described is a document in which Japanese and English are mixed, using a contour profile of binarized image data. .

JP 2001-22885 A

  However, in the technique described in Patent Document 1, when characters or peripheral structures are close to each other in the vertical direction of the character string to be recognized and the characters and the characters and the peripheral structures appear to be connected, There is a problem that it is difficult to specify the location, and the character recognition accuracy is lowered because the connected location of the characters cannot be correctly separated.

  The present invention has been made to solve the above-described problem, and provides a character recognition device that accurately performs character segmentation processing and improves character recognition accuracy even when a plurality of characters appear to be connected. For the purpose.

  In order to solve the above problems, the present invention employs the following means.

  The present invention is a character recognition device for inputting image data obtained by capturing a character string and recognizing each character included in the character string of the image data, the size of the character for recognizing the character. Storage means for storing a reference character size serving as a reference, estimated character size calculation means for calculating a plurality of patterns of estimated character sizes based on the reference character size, and estimated characters for the plurality of patterns Character cutout means for cutting out the character as a character area from the image data for each estimated character size based on the size, and character recognition processing and character similarity for each character area cut out for each estimated character size A character recognizing means for calculating a degree, and a character area having the highest character similarity is determined based on the character similarity calculated by the character recognizing means; Providing a character recognition apparatus characterized by comprising a determination means the true character area.

  According to the present invention, first, when recognizing each character of a character string included in image data, a plurality of estimated character sizes are calculated based on the reference character size stored in the storage means. Here, the reference character size can be, for example, information including a character height H and a width W. The height H and width W of the character are determined based on the font of the character to be recognized, the resolution of the camera that captures the character string, the distance between the camera and the character to be captured, and the like. For example, if the character to be recognized is a character written on the license plate, the license plate has a different character size, font, etc., depending on the country where the character recognition device is used. Is preferably determined.

  Thus, although the reference character size is stored in the storage unit, the character size of the image data may not necessarily match the reference character size. That is, for example, when the image data is taken due to the resolution of the camera, or when the image data is taken while being tilted with respect to the camera, the characters of the taken image data do not match the reference character size. There is. In particular, when recognizing the letters on the license plate attached to the vehicle, the distance between the camera that shoots the license plate and the license plate, and the angle of the license plate with respect to the camera, the installation position of the license plate relative to the vehicle Since it depends on the speed of the passing vehicle, the character size of the captured image data does not necessarily match the reference character size. Therefore, by calculating a plurality of patterns of estimated character sizes based on the reference character size, it is possible to estimate division candidate points that are boundaries between characters in a character string included in image data.

  Based on the calculated estimated character sizes of the plurality of patterns, characters are extracted from the image data for each of the estimated character sizes as character regions, and character recognition processing is performed on each character region extracted for each estimated character size. And character similarity is calculated. Furthermore, based on the character similarity with respect to each character area, the character area with the highest character similarity is determined, and this is determined as the true character area. In this way, even when a plurality of characters are connected and the reference character size of each character is unclear, by estimating a plurality of character sizes, an appropriate character size is searched and determined. Character cutout processing can be performed accurately, and character recognition accuracy is improved.

The character recognition device includes a split candidate point calculating unit that calculates a split candidate point of the character string based on a luminance gradient of the outline of the character , and the estimated character size calculating unit includes the split candidate point and the split candidate point. It is preferable to calculate a plurality of estimated character sizes obtained by estimating the character size based on the reference character size.

According to the present invention, based on the luminance gradient of the outline of the character, the candidate dividing points for calculating the candidate dividing point of the string, prior to character recognition, for example, on the basis of such a luminance gradient of the outline of the character Determine the character division candidate points. Since the estimated character size is calculated based on the candidate division points in addition to the reference character size, the number of estimated character size patterns to be calculated can be reduced, and the character recognition accuracy can be improved while suppressing the amount of calculation. Can be improved.

  Here, the division candidate point calculation means includes, for example, contour line extraction means for extracting a contour line of the character string from the image data using a differential filter, and vertical contour lines and horizontal contours among the extracted contour lines. For each pixel constituting the line, a code calculating means for calculating a sign of a differential value by the differential filter, and selecting one pixel from the pixels for which the sign is calculated as a target pixel, and including a first pixel including the target pixel Whether or not there is a pixel that sequentially extracts a region from the image data and shows a different code from the code of the pixel of interest based on the code calculated by the code calculation unit for the pixels included in the first region And a code change calculating means for calculating the number of first regions determined to be present in the first region extracted from the image data, and a pixel indicating the different symbol. The determined first region is classified into a region where the pixel of interest indicates a + sign and a region where the pixel of interest indicates a-sign, and as a result of the separation, a smaller number of regions are defined as division points of the character string. And a dividing point determining means to be configured.

  With this configuration, the contour line of the character string is extracted from the image data using a differential filter, and each pixel constituting the vertical contour line and the horizontal contour line among the contour lines is differentiated by the differential filter. Since the sign of the value is calculated, it is possible to grasp the luminance gradient for both the vertical contour line and the horizontal contour line. Then, one pixel is selected as the target pixel from the pixels for which the codes are calculated, and the first region including the target pixel is sequentially extracted from the image data. That is, in order from the top of the image data, one pixel among the pixels whose codes are calculated is selected as a target pixel, and a first region based on the target pixel is extracted. Here, the first area is determined based on the size, thickness, etc. of the character to be recognized. If the character to be recognized is, for example, a character described on a license plate of a vehicle, the first area is 4. It is preferable to define the area of about ˜5 pixels. Then, with respect to the pixels included in the first region, it is determined whether there is a pixel having a different sign from the sign of the pixel of interest based on the sign previously calculated, and the first extracted from the image data. The number of first regions determined to have pixels having different signs in one region is calculated. This process is performed on all the first areas sequentially extracted from the entire image data. By doing so, it is possible to grasp the change in the sign with respect to the target pixel in the first region, and to determine the brightness of the first region based on the calculation result. In the determination of lightness and darkness, the first region determined to have a pixel indicating a sign is classified into a region where the target pixel is positive and a region where the target pixel is negative. An area is set as a division point of the character string. That is, the number of areas where the pixel of interest has a + sign and the area where the pixel has a-sign pixel and the area of interest where the pixel of interest is-sign and the area where the pixel of + sign is in the area In comparison, a larger area is an area that captures a character line, and a smaller area is a segment that divides a character string as an area that captures a background.

  Further, the character recognition device described above determines an estimated character size corresponding to a character region determined to have the highest character similarity by the determination means as a first reference character size for the image data. Among the characters for which the true character area has not yet been determined, the estimated character size calculation means calculates a plurality of patterns of estimated character sizes obtained by estimating the character size based on the first reference character size. It is preferable to do.

  According to the present invention, if there is an estimated character size corresponding to the character region determined to have the highest character similarity by the determining means, this is determined as the first reference character size for the image data. If the recognition result for the same image data already exists, it is clear that the characters included in this image data are taken under the same conditions, so this estimated character size is included in this image data. It is preferable to use it for the character recognition process with respect to this character. Therefore, for a character whose true character area has not yet been determined in the image data, a plurality of patterns of estimated character sizes obtained by estimating the character size are calculated based on a first reference character size. Thus, the character recognition accuracy is further improved.

  Furthermore, when there are a plurality of estimated character sizes corresponding to the character region determined to have the highest character similarity by the determination unit, the character recognition device described above calculates an average value of the plurality of estimated character sizes as the image. Based on the second reference character size by the estimated character size calculation means for a character for which a true character area has not yet been determined in the image data. It is preferable to calculate a plurality of patterns of estimated character sizes obtained by estimating the character size.

  According to the present invention, when there are a plurality of estimated character sizes corresponding to the character region determined to have the highest character similarity by the determination means, an average value of the plurality of estimated character sizes is calculated for the image data. Determined as the second reference character size. Specifically, an average value of character heights and character widths of a plurality of estimated character sizes is calculated, and the average character height and average character width are set as the second reference character size. If multiple recognition results for the same image data already exist, it is clear that the characters included in the image data are taken under the same conditions, so the estimated character size based on the recognition results is set to this image. It is preferably used for character recognition processing for other characters included in the data. Therefore, for the characters for which the true character area has not yet been determined in the image data, a plurality of patterns are calculated for the estimated character size based on the second reference character size. Thus, the character recognition accuracy is further improved.

  Furthermore, the character recognition device described above sets the estimated character size corresponding to the character region determined to have the highest character similarity by the determination unit as the character size of other characters in the character string included in the image data. The character recognition process can be performed based on the character size.

  According to the present invention, when there is an estimated character size corresponding to the character area determined to have the highest character similarity by the determination means, that is, when a recognition result for the same image data already exists. Since it is clear that the characters included in the image data are photographed under the same conditions, it is preferable to use the estimated character size for character recognition processing for other characters included in the image data. For this reason, for characters that have not yet been subjected to character recognition processing in the image data, character estimation processing and processing are performed using the estimated character size corresponding to the character region determined to have the highest character similarity as it is. Character recognition processing can be performed. At this time, since the estimated character size is one, the amount of calculation can be reduced.

The present invention is a character recognition method for inputting image data obtained by capturing a character string and recognizing each character included in the character string of the image data, the size of the character for recognizing the character. A first step of storing a reference character size as a reference; a second step of calculating a division candidate point of the character string based on a luminance gradient of the outline of the character; and the division candidate point and the reference character size A third step of calculating a plurality of estimated character sizes based on the estimated character size, and a character from the image data for each estimated character size based on the estimated character size of the plurality of patterns. a fifth step of calculating a character similarity is performed with a fourth step of cutting out the region, the character recognition process on each of the character areas cut out for each of the estimated character size, the sentence Based on the character similarity calculated by the recognition means, a character recognition the character similarity is determined the highest character region, characterized by comprising this sixth determining the true character area, the Provide a method.

  As described above, according to the present invention, even when a plurality of characters appear to be connected, it is possible to accurately perform the character cutout process and improve the accuracy of character recognition.

It is a block diagram which shows schematic structure of the character recognition apparatus concerning the 1st Embodiment of this invention. It is a flowchart which shows the process of the character recognition process in the character recognition apparatus concerning the 1st Embodiment of this invention. It is explanatory drawing which showed the process of the character recognition process, (a) is image | photographed image data (concatenated label), (b) is an image of the estimated character size, (c) is after a recognition process. The character and character size are shown respectively. It is a block diagram which shows schematic structure of the character recognition apparatus concerning the 2nd Embodiment of this invention. It is a flowchart which shows the process of the character recognition process concerning the 2nd Embodiment of this invention. It is explanatory drawing which showed the process of the character recognition process, (a) is image | photographed image data (connection label), (b) is the image data after an outline emphasis process, (c) is image data. (D) shows an image of the estimated character size, and (e) shows the character and the character size after the recognition processing, respectively. It is a flowchart which shows the process of the division | segmentation candidate point determination process in the character recognition process concerning the 2nd Embodiment of this invention.

Hereinafter, embodiments of a character recognition device according to the present invention will be described with reference to the drawings.
[First Embodiment]
FIG. 1 is a block diagram showing a schematic configuration of a character recognition apparatus according to the first embodiment of the present invention. In the present embodiment, an example will be described in which the character recognition device of the present invention recognizes characters on a license plate of a vehicle.

  The character recognition device 10 according to the present invention includes a camera 11 that captures a license plate of a vehicle and obtains image data, an image input unit 12 that receives input of image data captured by the camera 11, and an image input unit 12 A recognition processing unit 13 that executes character recognition processing based on image data and a recognition result output unit 14 that outputs a recognition result in the recognition processing unit 13 are provided.

  The recognition processing unit 13 performs calculations for executing the character recognition process, and stores a CPU (central processing unit) 21 that executes various calculation processes related to the character recognition process, a program related to the character recognition process, and the like. A ROM (Read Only Memory) 22 which is a read-only memory, a RAM (Random Access Memory) 23 which is a readable / writable memory functioning as a work area of the CPU 21, and a character recognition process based on a program related to the character recognition process are executed. In this case, a storage device (storage means) 24 is stored in which various data necessary for this are stored. Here, the storage device 24 stores a reference character size that serves as a reference for the size of the character to be recognized in the character recognition process. The reference character size is information including the height H and width W of the character, and is determined based on the font of the character to be recognized, the resolution of the camera that captures the character string, the distance between the camera and the character, and the like. It is done. For example, since the license plate has different character sizes and fonts for each country, it is preferable that the reference character size is determined and stored in the storage device 24 in advance according to the country in which the character recognition device is used.

  The recognition processing unit 13 includes an estimated character size calculation unit (estimated character size calculation unit) 25, a character segmentation processing unit (character segmentation unit) 26, a character recognition processing unit (character recognition unit) 27, and a similarity determination unit (determination). Means) 28.

  Based on the reference character size stored in the storage device 24, the estimated character size calculation unit 25 calculates a plurality of patterns of estimated character sizes obtained by estimating the character size. For example, when the character string contained in the license plate is read multiple times, the distance between the camera that shoots the license plate and the license plate, and the angle of the license plate with respect to the camera, the mounting position of the license plate with respect to the vehicle and the camera Because it depends on the speed of the passing vehicle and the like, it does not always have the same distance and the same angle. In some countries, the size of each character included in the character string written on the license plate may not be uniform. For this reason, the reference character size stored in the storage device 24 does not necessarily match the character size of all the image data, and erroneous recognition occurs when character recognition processing is performed uniformly based on the reference character size. There is. Therefore, for the character string included in the input image data, a plurality of patterns of estimated character sizes are calculated according to the font and the like based on the reference character size. Estimated character size is calculated by comparing only the character height H or only the character width W, or both the character height H and the character width W as appropriate, compared with the reference character size. can do. In particular, when the reference character size is 1.0, for example, it is preferable to calculate a plurality of patterns of similar shapes having a size of 0.7 to 2.0 times at a pitch of 0.1 times.

  Based on the estimated character sizes of a plurality of patterns calculated by the estimated character size calculation unit 25, the character cutout unit 26 extracts characters from the image data as character regions for each of the estimated character sizes. That is, based on the estimated character size, a candidate division point that is a boundary between characters in the character string included in the input image data is estimated, and the recognition target is determined from the character string included in the image data for each estimated character size. Cut out the character area of the character.

  The character recognition processing unit 27 performs a character recognition process on each character region cut out for each estimated character size and calculates a character similarity. The character recognition processing unit 27 recognizes a character by performing, for example, a matching process, that is, by comparing a character to be recognized with a pattern provided in advance, as a character recognition process, and a character to be recognized and a pattern provided in advance. The character similarity is calculated. It should be noted that a pattern that serves as a comparison reference for the matching process is preferably stored in advance in the storage device 24 or the like.

  The similarity determination unit 28 determines a character region having the highest character similarity based on the character similarity for each character region calculated by the character recognition processing unit 27, and determines this as a true character region.

  The estimated character size calculation unit 25, the character segmentation processing unit 26, the character recognition processing unit 27, and the similarity determination unit 28 described above all expand the processing program stored in the predetermined ROM 22 by the CPU 21 to the RAM 23 and expand the processing program. A processing unit realized by executing a program.

  Hereinafter, the character recognition process in this embodiment is demonstrated with reference to FIG. 2, FIG. FIG. 2 is a flowchart showing a process of character recognition processing in the character recognition device of this embodiment. Here, a character recognition process when image data including a character string (hereinafter referred to as a “concatenated label”) in which the characters shown in FIG. 3 are concatenated is input will be described.

  As shown in FIG. 2, the license plate image captured by the camera 11 is input to the image input unit 12 as image data, and the recognition processing unit 13 performs stepping on the image data input to the image input unit 12. A predetermined character recognition process is performed in steps S101 to S105.

  First, the estimated character size calculation unit 25 reads the reference character size from the storage device 24, and in step S101, based on the read reference character size, the character “5” included in the concatenated label in FIG. Therefore, the estimated character sizes of a plurality of patterns are calculated in consideration of the estimation error. At this time, as shown in FIG. 3B, the start point coordinates of the character height H and the character width W of the estimated character size to be calculated are fixed, and compared with the reference character size, the character height H and the character The estimated character sizes of a plurality of patterns are calculated by appropriately changing the end point coordinates of the width W.

  In the next step S102, based on the estimated character sizes of a plurality of patterns calculated by the estimated character size calculator 25 in the previous step S101, characters are cut out from the image data as character regions for each of these estimated character sizes. That is, based on the estimated character size, a candidate division point that is a boundary between the character “5” to be recognized and the character “9” adjacent to “5” in the input image data is estimated. For each estimated character size, the character area of the recognition target character “5” is cut out from the concatenated label. Then, the character recognition processing unit 27 performs a matching process (character recognition process) on each clipped character area, that is, by comparing the character to be recognized “5” with a pattern provided in advance. recognize. At this time, the character similarity between the recognition target character “5” and the pattern provided in advance is calculated for all of the recognition results.

  Subsequently, in step S103, based on the character similarity with respect to each character area calculated by the character recognition processing unit 27, the character area with the highest character similarity, that is, the pattern provided in advance among the character areas is the most. It is determined which character area is close, and this character area is determined as a true character area. More specifically, character cutout processing and character recognition processing are performed based on the estimated character sizes of a plurality of patterns as shown in FIG. 3B, and as a result, as shown in FIG. Is determined as the true character region.

  In step S104, the character height H and character width W of the character region determined to be a true character region (see FIG. 3C) are set as reference character sizes for character recognition of this image data. That is, the reference character size stored in the storage device 24 remains stored as the default reference character size. On the other hand, when the recognition process for one character included in the concatenated label is completed for the character recognition process for a certain concatenated label, the estimated character size corresponding to the recognition result is stored as the reference character size for the concatenated label. This is used for character recognition processing for other characters included in this concatenated label.

  More specifically, when the character “5” in the connected label shown in FIG. 3 is recognized, a predetermined process such as calculation of an estimated character size is performed based on the reference character size stored in the storage device 24. At this time, when recognizing the character “9” that is the next recognition target of the character “5”, the estimated character size (character height H and character width W) based on the recognition result of the character “5” is first set. As the reference character size, an estimated character size is calculated based on the first reference character size, and the subsequent processing is performed. Similarly for the next characters “6” and “7”, the recognition result of the estimated character size (character height H and character width W) based on the previous recognition result or “5” which is the first character of this license plate. A plurality of estimated character sizes are calculated using the estimated character size (character height H and character width W) based on the first reference character size, and the subsequent processing is performed. Then, when the character recognition process for the connected label or license plate is completed and the character recognition process for the next connected label or license plate is performed, again based on the default reference character size stored in the storage device 24. Perform each process.

  In step S105, it is determined whether or not there is an area for which the character recognition process has not been completed in the connected label or this license plate. If there is an area for which the character recognition process has not been completed, the process returns to step S101 and described above. Repeat the process. Therefore, when the character recognition process for the character “5” is completed, there is an area in which the character recognition process of “9”, “6”, and “7” has not been completed yet in this concatenated label. The process from step S102 to step S104 is repeated until the process of recognizing these characters is completed. If there is no area in which character recognition processing has not been completed, that is, if all character recognition processing has been completed, it is determined that the character recognition processing for this concatenated label has been completed, and this routine ends.

  Thus, according to the present embodiment, even when a plurality of characters are connected and the reference character size of each character is unclear, an appropriate character size is searched by estimating the plurality of character sizes. Therefore, the character cutout process can be performed accurately, and the character recognition accuracy is improved.

[Second Embodiment]
Next, a second embodiment of the present invention will be described with reference to the drawings.

  The difference between the character recognition process of the present embodiment and the first embodiment described above is that, prior to calculating the estimated character size in the first embodiment, division candidate points between the characters of the character string included in the license plate are determined. In other words, the estimated character size is calculated based on the division candidate point and the reference character size. Therefore, as shown in FIG. 4, the character recognition device 20 in the present embodiment includes, in addition to the components of the character recognition device in the first embodiment, an outline extraction unit 35, a code calculation unit 36, and a code change calculation. A unit 37 and a division point determination unit 38 are provided. Hereinafter, regarding the character recognition processing of the present embodiment, description of points that are common to the first embodiment will be omitted, and different points will be described.

  As illustrated in FIG. 4, the recognition processing unit 13 further includes an outline extraction unit 35, a code calculation unit 36, a code change calculation unit 37, and a division point determination unit 38.

  The contour line extraction unit 35 calculates a spatial first derivative of the image data by using a first-order differential filter such as Sobel differentiation with respect to the image data input by the image input unit 12, and calculates the character in the image data. An outline is extracted (outline enhancement processing). In addition, the contour line is thinned by suppressing non-maximum contour pixels in the image data that has undergone the contour line emphasis process (contour line thinning process). The code calculation unit 36 receives image data that has been subjected to contour extraction processing and contour thinning processing, and the pixels constituting the vertical contour and horizontal contour with respect to the processed image data, that is, the left-right direction or The sign (+ or-) of the differential value of the contour is obtained for a pixel having a gradient in the vertical direction. Based on the code calculated by the code calculation unit 36, the code change calculation unit 37 extracts a local region as a first region corresponding to the thickness of the character from the image data, and the differential value of the pixel in the local region Is calculated. The dividing point determination unit 38 determines the lightness / darkness of the character with respect to the background color of the license plate based on the change of the sign calculated by the sign change calculation unit 37, and changes the sign opposite to the lightness / darkness of the character. The center of the local area is obtained, and this center is calculated as a dividing point.

  In the contour line extraction unit 35, the code calculation unit 36, the code change calculation unit 37, and the division point determination unit 38, the CPU 21 expands the processing program stored in the predetermined ROM 22 in the RAM 23 and executes the expanded program. It is a processing part realized by this.

  Hereinafter, the character recognition processing in this embodiment will be described with reference to FIGS. FIG. 5 is a flowchart showing a process of character recognition processing in the character recognition apparatus of this embodiment. Here, a character recognition process when image data including a character string (hereinafter referred to as a “concatenated label”) in which the characters shown in FIG. 6 are concatenated is input will be described.

  As shown in FIG. 5, the license plate image (see FIG. 6A) captured by the camera 11 is input to the image input unit 12 as image data, and the image data input to the image input unit 12 is The recognition processing unit 13 performs a predetermined character recognition process in steps S201 to S206. The processing in step S203 to step S206 is the same as the processing in step S102 to step S105 in the first embodiment described above, and a description thereof will be omitted.

  First, the license plate image data and the reference character size stored in the storage device 24 are input to the recognition processing unit 13 (see FIG. 6C). In step S201, division candidate points between the characters of the concatenated label are calculated based on the image data of the license plate (the process of division candidate point calculation processing will be described later). Subsequently, in step S202, a plurality of estimated character sizes are calculated based on the determined division candidate points and the reference character size input from the storage device 24 (see FIG. 6D). Here, since the division candidate points are already calculated, the character height H and the character width W are estimated based on the division candidate points, and the estimated character size is calculated. For this reason, the number of patterns of the estimated character size to be calculated can be reduced as compared with the estimated character size calculation process in the first embodiment described above. Therefore, when the size of the reference character size is 1.0, for example, when the character width of the division candidate point is 1.2 times the width of the reference character size, this size is 1.2 times larger. In addition to the estimated character size, it is possible to calculate estimated character sizes of 1.15 times and 1.25 times the reference character size which is the front and back size.

  In step S203 and subsequent steps, based on the estimated character size calculated in step S202, characters are extracted from the image data as character regions for each estimated character size, character recognition processing is performed on the extracted character regions, and the recognized characters are detected. The character area most similar to the pattern provided in advance is determined as the true character area (see FIG. 6E). Furthermore, the character height H and the character width W of the character region determined to be a true character region are set as the reference character size for character recognition of this concatenated label (see FIG. 6E), and character recognition of the concatenated label is still finished. For the area that has not been processed, the process returns to step S202 and the above-described processing is repeated.

  Next, linked label division candidate point determination processing will be described with reference to FIG. FIG. 7 is a flowchart showing the process of the division candidate point determination process.

  As shown in FIG. 7, the license plate image captured by the camera 11 is input to the image input unit 12 as the image data shown in FIG. 6A, and the recognition processing unit 13 performs predetermined processing in steps S301 and S302. Perform the process. That is, in step S301, the contour line extraction unit 35 calculates a spatial first derivative of the image data by using a differential filter such as a Sobel filter for the input image data, and as shown in FIG. A contour line of characters in image data is extracted (contour emphasis processing). In step S302, the contour extraction unit 35 thins the contour line by suppressing the non-maximum contour pixel with respect to the image data subjected to the contour line emphasis processing in step S301 (contour line thinning process). ).

  That is, for the image data that has undergone the contour enhancement processing in step S301, pixels having a gradient in the horizontal direction are extracted to calculate the direction of the contour line (vertical edge image generation), and in the direction perpendicular to the contour line. A point (maximum point) at which the contour line luminance value becomes the peak value is searched, and this is extracted as a point on the true contour line. Similarly, a pixel having a gradient in the vertical direction is extracted from the image data subjected to the contour emphasis processing in step S301 to calculate the direction of the contour line (horizontal edge image generation), and is perpendicular to the contour line. A point (maximum point) at which the contour luminance value becomes the peak value in the direction is searched, and this is extracted as a point on the true contour line.

  Subsequently, in step S303, the code calculation unit 36 of the recognition processing unit 13 applies the contour line to the pixel having a gradient in the horizontal direction or the vertical direction with respect to the image data subjected to the contour line emphasis process and the contour line thinning process. Find the sign of the derivative of. More specifically, the code calculation unit 36 pays attention to the gradient direction of the contour line for each pixel of the extracted contour line in the image data that has undergone the contour line emphasis process and the contour line thinning process. The sign of the differential value of the vertical contour of the differential filter is calculated for a pixel having a gradient in the vertical contour, that is, the vertical contour. Similarly, the sign of the differential value of the horizontal contour of the differential filter is calculated for a pixel having a gradient in the vertical direction, that is, the horizontal contour.

  In step S304, the sign change calculation unit 37 extracts a local area corresponding to the thickness of the character centered on the target pixel based on the sign of the differential value calculated in each pixel in step S303, and is included in the local area. It is calculated whether the sign of the pixel to be changed is compared with the sign of the target pixel. Specifically, first, the image data for which the sign of the differential value has been calculated in step S303 is sequentially scanned one line from the upper left pixel in the right direction to extract the pixel for which the differential value has been calculated. . Then, among the pixels included in one scanned row, a pixel whose differential value is calculated is used as a target pixel, and a predetermined local region including pixels corresponding to the character thickness is extracted from the target pixel in the right direction. Subsequently, the local region is searched for a pixel having a sign opposite to the target pixel. If there is a pixel having a sign opposite to the target pixel, the sign changes with respect to the target pixel. As a region where a pixel is located, information on the coordinates of the pixel and the direction of sign change (a sign for the target pixel + code or a + sign for the target pixel-code) is stored. This process is performed over one scanned line, and when the process for this one line is completed, the process for the next line is started. The above processing is sequentially performed for each line, so that the processing is repeated for the entire image data.

  Then, from the calculated code change, the number of regions where the pixel whose code has changed with respect to the target pixel is calculated for each piece of information of the code change direction. That is, the number of areas where there is a pixel that changes to − sign when the target pixel is + sign, and the number of areas where there is a pixel that changes to + sign when the target pixel is − sign, respectively. calculate.

  Here, the character thickness equivalent is often an area corresponding to about 4 to 5 pixels, and the range of the local area is the size, thickness, etc. of the character written on the license plate to be recognized. Is stored in advance in the storage device 24 as data determined in consideration of the above. In addition, when calculating horizontal division points, a local region is determined in the horizontal direction to calculate a change in sign, and when dividing points in the vertical direction are calculated, a local area is determined in the vertical direction and the code is changed. Calculate the change.

  In step S305, the division point determination unit 38 determines whether the local area is bright or dark with respect to the background, that is, whether the local area is an area related to the character line portion or an area related to the background portion, from the calculation result in step S303. To do. Specifically, the number of regions where there is a pixel that has changed to −sign with respect to the target pixel + sign is larger than the number of regions with pixels that have changed to + with respect to the target pixel−. The local region is determined as a local region that captures the character line portion. For example, if there are more regions with pixels that have changed to + sign with respect to the pixel of interest-sign than there are regions with pixels that have changed to-sign with respect to the pixel of interest + code, A region having a pixel that has changed to a + sign with respect to the pixel-code is determined as a character line portion, and a region having a pixel that has changed to a-sign with respect to the target pixel + code is determined as a background portion.

  In the next step S306, the dividing point determination unit 38 determines a dividing point in the local region. Specifically, the dividing point determination unit 38 determines the determination result of the local region related to the character line portion determined in step S305 and the local region related to the background portion, and the target pixel in the local region calculated in step S304. All the local regions determined as background parts are extracted based on the coordinates of the pixels whose sign is changed and information on the direction of the sign change. Then, the central pixel of the extracted local region is determined as a division candidate point.

  Thus, according to the present embodiment, even when a plurality of characters are connected and the reference character size of each character is unclear, an appropriate character size is searched by estimating the plurality of character sizes. Therefore, the character cutout process can be performed accurately, and the character recognition accuracy is improved. In particular, when character size is estimated, character division candidate points determined based on the brightness gradient of the outline of the character are used, so the number of estimated character size patterns can be reduced, and the amount of computation is reduced. However, the character recognition accuracy can be improved.

  In the first and second embodiments described above, a plurality of estimated character sizes are sequentially calculated for all the characters included in the concatenated label, and the character segmentation and character recognition processing are performed. Using the recognition result of the eye, for example, the character height H and the character width W of the estimated character size of the recognition result of the first character are used to perform character extraction and character recognition processing for the second and subsequent characters. You can also.

  In addition, when the recognition target character is the third or later character among the concatenated labels, the average value of the character height H and the character width of the estimated character size based on the recognition result of the characters for which character recognition processing has been completed so far The first reference character size can be used to perform character cut-out and character recognition processing for the third and subsequent characters. That is, when the character to be recognized is the second character, there is only one character for which recognition processing has already been completed, so the estimated character size based on the recognition result of the first character is set as the first reference character size. As a result, in order to cut out the character area of the second character, a plurality of estimated character sizes are calculated, and the subsequent processing is performed. If the recognition target character is the third character or later, there are a plurality of characters that have already been recognized, and therefore, from the character height H and the character width W of the estimated character size based on the recognition result of these characters. An average value is obtained, and a character size composed of an average character height H and an average character width W is set as a second reference character size. Based on the second reference character size calculated from the average value, a plurality of patterns of estimated character sizes are calculated in order to cut out the third and subsequent character areas, and the subsequent processing is performed.

DESCRIPTION OF SYMBOLS 10 Character recognition apparatus 11 Camera 12 Image input part 13 Recognition process part 14 Recognition result output part 20 Character recognition apparatus 25 Estimated character size calculation part 26 Character cut-out process part 27 Character recognition process part 28 Similarity determination part 35 Contour line extraction part 36 Code calculation unit 37 Code change calculation unit 38 Division point determination unit

Claims (6)

A character recognition device that inputs image data obtained by imaging a character string and recognizes each character included in the character string of the image data,
Storage means for storing a reference character size that serves as a reference for the size of the character in order to recognize the character;
An estimated character size calculating means for calculating a plurality of patterns of estimated character sizes based on the reference character size and estimating the size of the characters;
Based on the estimated character size of the plurality of patterns, character cutout means for cutting out the character as a character region from the image data for each estimated character size;
Character recognition means for performing character recognition processing on each character region cut out for each estimated character size and calculating character similarity;
A determination means for determining a character area having the highest character similarity based on the character similarity calculated by the character recognition means, and determining this as a true character area;
Division candidate point calculation means for calculating a division candidate point of the character string based on a luminance gradient of the outline of the character;
Equipped with a,
The estimated character size calculating means calculates a plurality of patterns of estimated character sizes obtained by estimating the size of the character based on the division candidate points and the reference character size . The division candidate point calculation means includes:
Contour extraction means for extracting a contour line of the character string from the image data using a differential filter;
Code calculating means for calculating a sign of a differential value by the differential filter for each pixel constituting a vertical contour line and a horizontal contour line among the extracted contour lines;
One pixel is selected as a target pixel from the pixels for which the code has been calculated, a first area including the target pixel is sequentially extracted from the image data, and the code calculation is performed for pixels included in the first area. Based on the code calculated by the means, it is determined whether or not there is a pixel having a different sign from the sign of the pixel of interest, and there is a pixel having a different sign in the first region extracted from the image data Sign change calculating means for calculating the number of first regions determined as:
As a result of classifying the first area determined to have a pixel having an opposite sign into an area in which the pixel of interest indicates + sign and an area in which the pixel of interest indicates-sign, the number of areas is smaller. A dividing point determining means that defines a character string dividing point as
A character recognition device according to claim 1. The estimated character size corresponding to the character region determined to have the highest character similarity by the determining means is defined as a first reference character size for the image data,
Among the image data, for characters whose true character area has not yet been determined,
3. The character according to claim 1, wherein the estimated character size calculation unit calculates a plurality of patterns of estimated character sizes obtained by estimating the size of the character based on the first reference character size. 4. Recognition device. When there are a plurality of estimated character sizes corresponding to the character region determined to have the highest character similarity by the determination means, an average value of the plurality of estimated character sizes is a second reference character size for the image data. Set
Among the image data, for characters whose true character area has not yet been determined,
3. The character according to claim 1, wherein the estimated character size calculating unit calculates a plurality of patterns of estimated character sizes obtained by estimating the size of the character based on the second reference character size. 4. Recognition device.   The estimated character size corresponding to the character region determined to have the highest character similarity by the determination means is set as the character size of other characters in the character string included in the image data, and character recognition is performed based on the character size. The character recognition device according to claim 1, wherein the character recognition device performs processing. A character recognition method for inputting image data obtained by imaging a character string and recognizing each character included in the character string of the image data,
A first step of storing a reference character size serving as a reference for the size of the character in order to recognize the character;
A second step of calculating candidate division points of the character string based on a luminance gradient of the outline of the character;
A third step of calculating a plurality of patterns of estimated character sizes based on the division candidate points and the reference character size;
A fourth step of cutting out the character as a character region from the image data for each estimated character size based on the estimated character size of the plurality of patterns;
A fifth step of performing a character recognition process on each character region cut out for each estimated character size and calculating a character similarity;
A sixth step of determining a character area having the highest character similarity based on the character similarity calculated by the character recognition means and determining this as a true character area;
A character recognition method comprising:

Images