JP2005149395A - Character recognition device and license plate recognition system - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a character recognition device and a license plate recognition system which can perform processing with high recognition accuracy and at high speed. <P>SOLUTION: Even information about a detailed part lost in a low resolution character image is restored and character recognition is performed with high accuracy by performing super resolution processing of the character image and performing the character recognition by using a specified high resolution dictionary image. In the case of the character recognition, direction histogram features of a contour line are calculated by using the specified high resolution dictionary image and performing binarization processing to it and luminance gradient features are used. In addition, a feature pattern used in the case of specifying a low resolution dictionary image corresponding to the character image is clustered into a plurality of groups and stored in a data base and it is effective to specify the low resolution dictionary image corresponding to the character image using a representative pattern of each group. <P>COPYRIGHT: (C)2005,JPO&NCIPI

Description

  The present invention relates to a character recognition device and a license plate recognition system suitable for automatically recognizing the description of a vehicle license plate.

In recent years, many devices for automatically recognizing a license plate of a vehicle, such as a tollgate on a highway, have been used. This is because, for example, the type of vehicle is determined by recognizing the description of the license plate.
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 (binarized) 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 character cut-out area, character shape features (for example, contour direction features) are extracted, and the characters are recognized by pattern recognition.
However, in a character image (low-resolution image) taken with a low-resolution camera, the character line part is likely to be blurred (line skipping) or the character hole part is crushed, so that the character cannot be recognized correctly. There is. For this reason, a character image is usually taken with a camera having sufficient resolution, but such a camera is expensive. Further, in order to obtain an image with high resolution, since the image is taken in an enlarged manner, the area that can be taken with one camera is small, and a plurality of cameras are required to take an image of all the vehicles on the road. As a result, the cost of the entire system increases.

  As means for solving all of the above problems, there is a method of recognizing characters by increasing the resolution of a character image taken at a low resolution. As a conventional technique for recognizing characters by increasing the resolution of an image, linear interpolation or sub-pixel processing is used. These are all high-resolution processes based on the original low-resolution image. Therefore, since the detailed information originally lost cannot be restored, the characters cannot be completely restored, and it is difficult to improve the character recognition accuracy.

  As a method for solving such a problem of information restoration, there is a super resolution method based on an example (hereinafter referred to as a super-resolution processing method) (for example, see Non-Patent Document 1). In this method, a high-resolution image collected in advance and a corresponding low-resolution image are stored in pairs, and the low-resolution input image to be recognized is compared with the pre-stored low-resolution image. Pick out similar ones. Then, a high-resolution image corresponding to the selected low-resolution image is specified, and image processing for character recognition is performed using the high-resolution image. By increasing the resolution of the low resolution image to be recognized in this way, it is possible to restore even detailed information lost in the low resolution image, leading to improved character recognition accuracy.

S. Baker & T. Kanade, "Limits on Super-Resolution and How to Break Them" IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol. 24, No. 9, September, 2002, pp. 1167-1183

  However, since the technique of Patent Document 1 performs brute force matching processing between an input image and a pre-stored low-resolution image, the amount of data to be stored and the amount of computation at the time of processing become enormous and high-speed processing is required. Is difficult. When reading a license plate of an automobile, it is necessary to complete the processing within approximately one second, and not only high recognition accuracy but also high processing speed is very important.

  In the method of Patent Document 1, matching processing between an input image and a pre-stored low-resolution image is performed by normalized correlation calculation. This normalized correlation calculation is applied to deformation and distortion of a target character. There is a problem of being weak. When reading a license plate of an automobile, for example, a vehicle entering a toll gate may enter at an angle rather than a fixed angle. Also, the license plate may be attached at an angle that is not substantially perpendicular to the road surface but is inclined upward or downward. Further, it may be bent or deformed due to an accident or the like. Then, the characters in the input image are often deformed or distorted. In such a case, it is difficult to obtain high recognition accuracy by the normalized correlation calculation.

  The present invention has been made based on such a technical problem, and an object of the present invention is to provide a character recognition device and a license plate recognition system capable of performing high-speed processing with high recognition accuracy.

For this purpose, the character recognition device according to the present invention divides the dictionary image when each of the plurality of characters is imaged at a predetermined resolution and the shape of each characteristic portion of the plurality of characters into a plurality of groups. And a matching processing unit for identifying a dictionary image that matches the character image based on the character database input from the outside, the character image input from the outside, and the shape of the characteristic part stored in the database, and the matching processing unit. And a recognition processing unit that recognizes characters in the character image by performing character recognition processing on the dictionary image, and a result output unit that outputs the recognized character as a recognition result.
In this way, the shape of the characteristic part of the character is used for matching the input character image and the dictionary image. The characteristic part of the character is divided into groups in the database. Data storage amount can be reduced. Also, if you first identify a group with a similar shape of the characteristic part, you only need to match with the shape of the characteristic part belonging to that group, so without matching the round-robin like in the past, Matching efficiency can be improved.

  Here, in the database, a high-resolution dictionary image when each of a plurality of characters is captured at a higher resolution than the resolution of the dictionary image is stored in association with the dictionary image, and the recognition processing unit specifies the matching processing unit. It is also possible to recognize characters of a character image by calling a high resolution dictionary image associated with the dictionary image from the database and performing character recognition processing on the high resolution dictionary image. That is, so-called super-resolution processing can be executed in combination.

The recognition processing unit can also select whether to perform character recognition for the dictionary image or character recognition for the high-resolution dictionary image based on the filling rate or frequency of the dictionary image.
Further, the recognition processing unit binarizes the dictionary image or the high resolution dictionary image to recognize the character of the character image, or based on the brightness gradient vector of the dictionary image or the high resolution dictionary image, Can also be recognized. In addition, the binarization process and the luminance gradient vector can be switched depending on circumstances. That is, the recognition processing unit binarizes the dictionary image or the high-resolution dictionary image based on the reliability of character recognition to perform character recognition, or character recognition based on the brightness gradient vector of the dictionary image or the high-resolution dictionary image. Choose whether to recognize.
In this way, a super-resolution process that performs character recognition for a high-resolution dictionary image, a binarization process that is used as a technique for character recognition, and a luminance gradient vector are used as necessary, thereby increasing the load ( The processing can be reduced as a whole, and the character recognition accuracy can be improved only when necessary.

With this point as a point, the present invention can also be understood.
That is, the character recognition device of the present invention has a dictionary image when each of a plurality of characters is captured at a predetermined resolution, and a high-resolution dictionary image when each of the plurality of characters is captured at a higher resolution than the resolution of the dictionary image. Based on the condition of the dictionary image specified by the specification processing unit for specifying the dictionary image that matches the character image input from the outside, the dictionary image stored in the database, the dictionary image stored in the database, and the specification processing unit A character recognition unit that recognizes characters in a character image by selecting a high-resolution dictionary image associated with the image or dictionary image and performing character recognition processing on the selected dictionary image or high-resolution dictionary image. And a result output unit that outputs the character as a recognition result.
Here, when a dictionary image is used for character recognition processing, it indicates that super-resolution processing is not performed.
Thus, in order to select a dictionary image or a high-resolution dictionary image, the character recognition unit can make a selection based on the filling rate or frequency of the dictionary image specified by the specifying processing unit.

  The character recognition device as described above does not ask for a character image input means, and is not limited to a character image photographed by a camera. Any acquisition means may be used. Moreover, the character image may be described in any thing. That is, the character recognition device is not intended to limit the character recognition target to the license plate.

  On the other hand, the present invention is a camera that captures a license plate of a passing vehicle, a recognition unit that recognizes characters written on the license plate based on a license plate image captured by the camera, and a recognition unit that recognizes characters. Therefore, it can also be understood as a license plate recognition system including a database storing representative patterns of groups generated by clustering a plurality of feature patterns of a plurality of characters. In this case, in the license plate recognition system, the recognition unit recognizes characters written on the license plate based on the image of the license plate taken by the camera and the representative pattern stored in the database. . In this manner, by storing the character feature patterns in the database in groups, it is possible to reduce the data storage amount and improve the matching efficiency.

Here, in order to recognize the characters written on the license plate, the representative pattern can be used as it is.
On the other hand, the representative pattern of each group and the characteristic pattern of the group to which the representative pattern belongs are stored in the database in association with each other, and the recognition unit identifies the representative pattern corresponding to the license plate image captured by the camera, Based on the characteristic pattern of the group to which the identified representative pattern belongs, it is also possible to recognize the characters written on the license plate. In other words, after specifying the representative pattern, the feature pattern that best matches the image of the license plate photographed by the camera is extracted from the group to which the representative pattern belongs, and the group to which the specified representative pattern belongs is used. Character recognition is performed.

  ADVANTAGE OF THE INVENTION According to this invention, it becomes possible to perform the automatic recognition process of the description content of the license plate of a vehicle at high speed with high recognition accuracy.

Hereinafter, the present invention will be described in detail based on embodiments shown in the accompanying drawings.
FIG. 1 is a diagram for explaining the configuration of a license plate recognition system 10 in the present embodiment.
The license plate recognition system 10 is based on a camera 20 installed obliquely downward above the road R and an image captured by the camera 20 in order to photograph the license plate 100 of the vehicle on the road R. And a character recognition device 30 for recognizing the characters written on the license plate 100.

  The character recognition device 30 includes an image input unit 31 that receives an input of an image captured by the camera 20, and a recognition processing unit (matching processing unit, specific process) that executes character recognition processing based on the image input to the image input unit 31. Part, character recognition unit, recognition unit) 32, a database 33 storing data necessary for executing the recognition process in the recognition processing unit 32, and a recognition result for outputting the recognition processing result in the recognition processing unit 32 to the outside An output unit (result output unit) 34.

  As shown in FIG. 2, the database 33 includes characters (including numbers, “−”, “•”, etc.) used for the license plate 100 when the characters are photographed at a resolution equivalent to that of the camera 20. Data of an image (dictionary image: hereinafter referred to as a low resolution dictionary image) L and an image (hereinafter referred to as a high resolution dictionary image) H obtained by photographing the same character at a higher resolution than the camera 20 are stored in association with each other. Yes. The low-resolution dictionary image L is obtained by analyzing this in advance. For example, the luminance value of each pixel of the low-resolution dictionary image L, the differential value of the luminance value (horizontal direction, vertical direction, diagonally lower right) An index indicating the feature amount of the character image (hereinafter referred to as a feature amount) such as a direction or a diagonally upward right direction is further stored in the database 33 in association with each other.

Hereinafter, a character recognition processing method in the recognition processing unit 32 will be described.
FIG. 3 shows an example of the flow of character recognition processing in the present embodiment. The character recognition processing is automatically executed by the recognition processing unit 32 that is functionally expressed by the cooperation of the CPU, the memory, and the like based on a program installed in advance in the character recognition device 30 that is a computer device. Is.
As shown in FIG. 3, when an image is input to the image input unit 31, the recognition processing unit 32 performs predetermined preprocessing such as cutting out an image of only the area corresponding to the license plate 100 (step). S101).
Then, a character cutout process is performed to cut out an image (hereinafter referred to as a character image) I in an area where characters are written on the license plate 100 (step S102). At this time, the region of the character image I to be cut out is cut out for each character.

Then, character recognition processing using super-resolution processing is executed (step S103). In this character recognition processing, pattern matching processing based on character shape feature extraction from a binarized image is used instead of normalized correlation calculation.
Specifically, the obtained character image I is analyzed by the recognition processing unit 32, and an index indicating the feature amount of the character image I is calculated. Then, as shown in FIG. 4, the calculated feature amount of the character image I is compared with the feature amount of the low resolution dictionary image L stored in the database 33 to perform matching, and the low resolution with the highest matching degree is obtained. A dictionary image L is specified (step S111). That is, so-called super-resolution processing is performed.
Subsequently, the high-resolution dictionary image H associated with the specified low-resolution dictionary image L is called from the database 33, and binarization processing is performed by the recognition processing unit 32 (step S112). Further, in the binarized image, the direction histogram feature of the outline is obtained, and the character is recognized by statistical pattern recognition (step S113).
And the recognition result of a character is decided and this is memorize | stored in the memory of the character recognition apparatus 30 (step S114).

The processes in steps S111 to S114 are looped by the number of character images I cut out in step S102 (the number of characters), and the characters of the individual character images I are recognized.
When the recognition of all the character images I is completed, the recognition results of all the characters described in the license plate 100 are obtained from the recognition results of the individual characters stored in the memory of the character recognition device 30. is there.
The obtained character recognition result is output from the recognition result output unit 34 to another external processing device, storage device, or the like, if necessary.

In this way, the character image I is subjected to super-resolution processing and character recognition is performed using the specified high-resolution dictionary image H, so that even detailed information lost in the low-resolution character image I is restored. Character recognition can be performed with high accuracy. Moreover, when character recognition is performed, the specified high-resolution dictionary image H is binarized to obtain a direction histogram characteristic of the contour line, thereby obtaining a normalized cross-correlation technique that is weak against deformation and distortion. Character recognition accuracy can be improved as compared with the case of using.
As a result, the camera 20 can be made to have a lower resolution than the conventional one, and not only the cost of the camera 20 itself can be suppressed, but also the number of cameras 20 itself can be reduced, which greatly increases the cost of the license plate recognition system 10. It becomes possible to reduce it.

In the character recognition process in step S103, the binarization process is used. However, a luminance gradient feature can be used instead.
That is, as shown in FIG. 5, in the character recognition processing (step S103) after the input image preprocessing (step S101) and the character segmentation processing (step S102), the character image is the same as in step S111 of FIG. Super-resolution processing for I is performed to obtain a high-resolution dictionary image H (step S121).
After that, instead of performing the binarization process as described above, a histogram characteristic in the luminance gradient direction based on the density of the obtained high-resolution dictionary image H is obtained, and the character is recognized by statistical pattern recognition (step S122).

And the recognition result of a character is decided and this is memorize | stored in the memory of the character recognition apparatus 30 (step S123).
The processes in steps S121 to S123 are looped by the number of character images I cut out in step S102 (the number of characters), and the characters of each character image I are recognized, and all character images I are recognized. When the recognition is completed, the recognition results of all the characters described in the license plate 100 are obtained from the recognition results of the individual characters stored in the memory of the character recognition device 30. The recognition result output unit 34 outputs the result to another external processing device, storage device, or the like.

  Thus, the same effect as that of the method of FIG. 3 can be obtained by using the luminance gradient instead of the binarization processing. When the luminance gradient is used, character recognition with higher accuracy than binarization processing can be expected.

In addition, character recognition by binarization processing and character recognition using a luminance gradient can be combined.
That is, as shown in FIG. 6, in the character recognition processing (step S103) after the input image preprocessing (step S101) and the character segmentation processing (step S102), the character image is the same as in step S111 of FIG. Super-resolution processing for I is performed to obtain a high-resolution dictionary image H (step S131).
Subsequently, the obtained high resolution dictionary image H is called from the database 33, and binarization processing is performed by the recognition processing unit 32 (step S132). Further, in the binarized image, the direction histogram feature of the outline is obtained, and the character is recognized by statistical pattern recognition (step S133).

Then, as a result of the character recognition by the binarization process, it is determined whether or not the reliability of the character similarity is higher than a preset threshold value (step S134). Here, the Mahalanobis distance is used to determine whether or not the reliability of character similarity is high. That is, as a result of character recognition by binarization processing, the Mahalanobis distance is obtained as an index of character similarity, and it is determined whether or not this Mahalanobis distance is higher than a preset threshold value S1.
At this time, when there are a plurality of character candidates that clear the threshold value S1 as a result of character recognition, the Mahalanobis distances for the respective characters are compared. In this comparison, the smaller Mahalanobis distance is not used as the final character recognition result (the character), but based on the Mahalanobis distance ratio, whether or not the ratio is greater than a preset threshold value S2. It is preferable to judge by.

For example, if the threshold value S1 is 0.40, the Mahalanobis distance for the first candidate character is 0.25, and the Mahalanobis distance for the second candidate character is 0.42, It is determined that the candidate character is highly reliable. When the Mahalanobis distance exceeds the threshold value S1 from the first candidate character, it is determined that the reliability is low.
Also, if the Mahalanobis distance for the first candidate character is 0.10 and the Mahalanobis distance for the second candidate character is 0.30, both characters have cleared the threshold value S1, so in this case The reliability is determined based on whether these ratios (0.30 / 0.10 = 3.0) are larger than the threshold value S2.

  As a result, when it is determined that the similarity of the recognized character is highly reliable, the character is determined as a recognition result and stored in the memory of the character recognition device 30 (step S135).

On the other hand, if it is determined in step S134 that the reliability of the character similarity is low, the histogram feature in the luminance gradient direction based on the density of the obtained high-resolution dictionary image H is continued as in the case of FIG. The character is recognized by statistical pattern recognition (step S136).
Then, as a result of character recognition, it is determined in the same manner as in step S134 whether or not the reliability of character similarity is higher than a preset threshold value (step S137).
As a result, when it is determined that the similarity of the recognized character is highly reliable, the character is determined as a recognition result and stored in the memory of the character recognition device 30 (step S138). On the other hand, if it is determined that the reliability is low, an unknown determination is made to indicate that the recognized character is unknown, that is, the character could not be recognized (step S139).

  When the character recognition result is determined as described above, this is stored in the memory of the character recognition device 30, and then the above-described processing in steps S131 to S139 is performed as the number of character images I cut out in step S102. (Number of characters) is looped to recognize the characters of each character image I, and when the recognition of all the character images I is completed, the characters of the individual characters stored in the memory of the character recognition device 30 are recognized. From the recognition result, the recognition result of all the characters written on the license plate 100 is obtained, and this is output from the recognition result output unit 34 to other external processing devices, storage devices, etc. as necessary. .

  In this way, when a highly reliable result cannot be obtained by binarized character recognition, a character recognition process using a luminance gradient feature with higher character recognition accuracy can be performed to obtain a recognition result with higher accuracy. Can do. In addition, the character recognition process using the luminance gradient feature having a high calculation load can be executed at high speed by performing the character recognition process only when necessary, that is, when the highly reliable result cannot be obtained by the character recognition by binarization. it can.

By the way, in the character recognition process S103 as described above, it is possible to adopt a configuration in which the super-resolution process is not necessarily performed.
That is, as shown in FIG. 7, in the character recognition processing (step S103) after the input image preprocessing (step S101) and the character segmentation processing (step S102), first, the character size and filling rate are evaluated. (Step S141).
For this purpose, the recognition processing unit 32 obtains the size of the character (portion that can be determined) in the extracted character image I, and the background color of the license plate 100 (in Japan, white, green, yellow, The space factor of the area other than black or blue) with respect to the area of the character image I (this is the filling factor) is calculated (evaluated).
Then, it is determined whether or not the character size is equal to or less than a certain value and the filling rate is equal to or greater than a certain value (step S142).

As a result, when it is determined that the character size is equal to or smaller than a certain value and the filling rate is equal to or larger than the certain value, the super-resolution processing is performed on the character image I in the same manner as in step S111 in FIG. A resolution dictionary image H is obtained (step S143).
Subsequently, the obtained high-resolution dictionary image H is called from the database 33, and binarization processing is performed by the recognition processing unit 32 (step S144). Further, in the binarized image, the direction histogram feature of the outline is obtained, and the character is recognized by statistical pattern recognition (step S145).

  On the other hand, when it is determined in step S142 that the character size is not more than a certain value and the filling rate is not less than the certain value, that is, the character size is not less than a certain value. Or the filling rate is less than a certain value, or the character size is greater than or equal to a certain value and the filling rate is less than a certain value, the super-resolution processing in step S143 is skipped, and steps S144 and S145 are performed. In step S145, the direction histogram feature of the contour line in the binarized image is obtained, and the character is recognized by statistical pattern recognition. At this time, since the target image to be binarized in step S144 is not subjected to super-resolution processing, it can be the character image I itself or the low-resolution dictionary image L having a high matching degree with the character image I. In terms of processing speed, it is preferable to perform binarization processing on the character image I itself. However, in terms of character recognition accuracy, the low-resolution dictionary image L having a high matching degree with the character image I is targeted. It is preferable to do this.

And the recognition result of a character is decided and this is memorize | stored in the memory of the character recognition apparatus 30 (step S146).
By looping the processing of steps S141 to S146 as many as the number of character images I cut out in step S102 (the number of characters), recognition results of all characters described in the license plate 100 are obtained. This is output from the recognition result output unit 34 to other external processing devices, storage devices, etc., as necessary.

  At this time, the character binarization step and the character recognition step in steps S144 and 145 are character recognition processing using a luminance gradient as shown in FIG. 5, or binarization as necessary shown in FIG. It is also possible to perform character recognition processing for switching the luminance gradient.

  In this way, in the example of FIG. 7, the super-resolution processing is performed only when the character size is equal to or smaller than a certain value and the filling rate is equal to or larger than the certain value. As a result, when the character is large or the filling rate is low, the super-resolution processing is not performed. The filling rate is low if the obtained character image is not blurred, and is low if it is not a complicated character. That is, when a large character or a character image is clear or an uncomplicated character, the super-resolution processing is not performed, and the processing speed can be increased. In addition, when a small character or an obtained character image is blurred or a complicated character, super-resolution processing can be performed to improve character recognition accuracy. In this way, it is possible to obtain high character recognition accuracy by performing high-precision character recognition only when necessary while speeding up the processing.

In the example of FIG. 7, the filling rate is used to determine whether or not to perform super-resolution processing. However, instead of this, a frequency can also be used.
That is, instead of step S142, the low resolution character image I obtained in step S102 is subjected to a differentiation process, and it is determined whether the obtained differential value is greater than a predetermined value. Only when it is determined that the value is greater than or equal to the value, the process proceeds to step S143 to perform the super-resolution processing. A large differential value means that the frequency of the background color portion and the character portion of the license plate 100 is high, and there are many portions where it is desired to sharpen the edges. For example, as shown in FIG. 8, since the blank portions A1 to A4 of the license plate 100 are only the background color, the differential value is small (close to zero). Further, in the area A5 where large characters are written on the license plate 100, the frequency is low because the characters are large. On the other hand, the frequency is high in the regions A6 to A8 where the characters are small.
Therefore, according to the specification of the camera 20 to be used, a threshold value (a constant value) is set so that super-resolution processing is performed when the frequency is high even in the small character area A6 to A8 or the small character area A6 to A8. Is set.

  Even in this case, as in the case of FIG. 7 described above, the overall processing amount (calculation amount) is greatly reduced, the processing speed is increased, and high-precision character recognition is performed by super-resolution processing only when necessary. As a result, high character recognition accuracy can be obtained.

As described above, the character recognition process is performed in the character recognition device 30. To further increase the speed of the process, the low resolution dictionary image L and the high resolution dictionary image H stored in the database 33 are used. It is preferable to reduce the amount of data as much as possible.
First, there is a method of quantizing the value of each element of the feature amount (feature pattern vector) of the low resolution dictionary image L collected in advance for storage in the database 33. For example, when each value constituting the feature pattern vector has a value of 256 gradations, it is quantized to about 32 gradations. The quantized feature pattern vector is stored in the database 33 as the feature amount of the low resolution dictionary image L. Thereby, the amount of data stored in the database 33 can be substantially compressed.

Other methods include the following.
The characters in the low-resolution dictionary image L can be decomposed into feature patterns (characteristic part shapes) as shown in FIG. In the super-resolution processing, the character image I and the low-resolution dictionary image L can be matched by specifying the types of these feature patterns and combinations thereof.

Such feature patterns are clustered by a technique such as a k-means method or a vector quantum method, and are grouped into a plurality of groups G1, G2,... As shown in FIG. In other words, those with similar feature patterns are collected in the same group.
In each of the groups G1, G2,..., Representative patterns P1, P2,... Are determined, and all the feature patterns belonging to the groups G1, G2,. To store.

  Actually, when character recognition is performed, as shown in FIG. 10, the feature pattern of the character image I is matched with the representative patterns P1, P2,... Stored in the database 33. When a representative pattern to be matched (for example, representative pattern P3) is specified, subsequently, a group G3 associated with the representative pattern P3 is specified. Then, the feature pattern belonging to the group G3 is matched with the feature pattern of the character image I, and the feature pattern having the highest matching degree is specified for the character image I.

  As described above, when matching between the character image I and the low resolution dictionary image L, the use of the representative patterns P1, P2,. Time can be greatly reduced.

There are also the following methods.
Similar to the method shown in FIG. 9, the character feature patterns in the low-resolution dictionary image L are clustered by a technique such as the k-means method or the vector quantum method, and a plurality of groups are obtained as shown in FIG. Group into G1, G2,. Then, representative patterns P1, P2,... Are determined in each of the groups G1, G2,.
As shown in FIG. 11B, in the database 33, only the representative patterns P1, P2,... Are stored as feature patterns of the low resolution dictionary image L in the database 33 (belonging to the groups G1, G2,...). All other feature patterns are not stored in the database 33).

  When character recognition is actually performed, as shown in FIG. 12, the characteristic pattern of the character image I and the representative patterns P1, P2,... Stored in the database 33 shown in FIG. Perform matching. When a representative pattern to be matched (for example, representative pattern P3) is specified, this representative pattern P3 is used as a feature pattern having the highest matching degree in the character image I.

According to such a method, as in the case shown in FIGS. 9 and 10, when matching the character image I and the low resolution dictionary image L, the representative patterns P1, P2,. It is no longer necessary to perform such round robin matching, and the calculation time required for matching can be greatly reduced. Further, in comparison with the cases shown in FIGS. 9 and 10, after specifying the representative pattern to be matched, the group to which the pattern belongs is specified, and the feature pattern belonging to the group is matched with the feature pattern of the character image I. Therefore, it is not necessary to specify the feature pattern having the highest matching degree in the character image I. Thereby, it becomes possible to perform processing at higher speed. On the other hand, in terms of character recognition accuracy, since the representative patterns P1, P2,... Are used as they are for matching with the feature pattern of the character image I, the character recognition accuracy is higher in the methods of FIGS. Which method should be adopted may be determined according to the required processing speed, character recognition accuracy, and the like.
Further, since only the representative patterns P1, P2,... Are stored in the database 33, the amount of data stored in the database 33 can be significantly reduced.

Note that the structures and methods described in the above embodiments can be combined as appropriate.
In addition to this, as long as it does not depart from the gist of the present invention, the configuration described in the above embodiment can be selected or changed to another configuration as appropriate.

It is a figure which shows the structure of the license plate recognition system in this Embodiment. It is a figure which shows storing a low resolution dictionary image and a high resolution dictionary image in a database. It is a figure which shows the flow of a process in the case of character recognition by a binarization process after a super-resolution process. It is a figure which shows obtaining the high resolution dictionary image which matches the input character image in a super resolution process. It is a figure which shows the flow of a process in the case of character recognition using a brightness | luminance gradient after a super-resolution process. It is a figure which shows the flow of a process in the case of switching the character recognition using a binarization process, and the character recognition using a brightness | luminance gradient according to reliability. It is a figure which shows the flow of a process in the case of performing a super-resolution process only when predetermined conditions are satisfy | filled according to conditions. It is a figure for demonstrating regarding the frequency of an image. It is a figure which shows storing the feature pattern of a character in a database. It is a figure which shows performing character recognition using the feature pattern stored by clustering. It is a figure which shows the other example which clusters and stores the characteristic pattern of a character in a database. It is a figure which shows performing character recognition using the feature pattern stored by clustering.

Explanation of symbols

  DESCRIPTION OF SYMBOLS 10 ... License plate recognition system, 20 ... Camera, 30 ... Character recognition apparatus, 32 ... Recognition process part (Matching process part, Specific process part, Character recognition part, Recognition part), 33 ... Database, 34 ... Recognition result output part ( Result output unit), 100 ... license plate, H ... high resolution dictionary image, I ... character image, L ... low resolution dictionary image (dictionary image)

Claims (10)

A dictionary image when each of the plurality of characters is imaged at a predetermined resolution, and a database storing the shape of each characteristic portion of the plurality of characters divided into a plurality of groups,
A matching processing unit that identifies a dictionary image that matches the character image based on the character image input from the outside and the shape of the characteristic portion stored in the database;
A recognition processing unit that recognizes characters in the character image by performing character recognition processing on the dictionary image specified by the matching processing unit;
And a result output unit that outputs the recognized character as a recognition result. The database stores, in association with the dictionary image, a high-resolution dictionary image obtained by capturing each of the plurality of characters at a higher resolution than the resolution of the dictionary image,
The recognition processing unit calls the high-resolution dictionary image associated with the dictionary image specified by the matching processing unit from the database, and performs character recognition processing on the high-resolution dictionary image, whereby the character image The character recognition apparatus according to claim 1, wherein the character is recognized.   The recognition processing unit selects whether to perform character recognition for the dictionary image or character recognition for the high-resolution dictionary image based on the filling rate or frequency of the dictionary image. The character recognition device according to claim 2.   The character recognition according to claim 1, wherein the recognition processing unit recognizes characters of the character image by binarizing the dictionary image or the high-resolution dictionary image. apparatus.   The character recognition device according to claim 1, wherein the recognition processing unit recognizes a character of the character image based on a luminance gradient vector of the dictionary image or the high-resolution dictionary image.   The recognition processing unit binarizes the dictionary image or the high-resolution dictionary image based on the reliability of character recognition to perform character recognition, or uses the brightness gradient vector of the dictionary image or the high-resolution dictionary image. 4. The character recognition device according to claim 1, wherein whether to perform character recognition is selected based on the selection. A database storing a dictionary image when each of a plurality of characters is captured at a predetermined resolution, and a high-resolution dictionary image when each of the plurality of characters is captured at a higher resolution than the resolution of the dictionary image;
A specific processing unit that identifies the dictionary image that matches a character image input from the outside, from the dictionary image stored in the database;
Based on the dictionary image condition specified by the specification processing unit, the dictionary image or the high-resolution dictionary image associated with the dictionary image is selected, and the selected dictionary image or the high-resolution dictionary image is selected. A character recognition unit that recognizes characters in the character image by performing character recognition processing;
A result output unit for outputting the recognized character as a recognition result;
A character recognition device comprising:   The character recognition according to claim 7, wherein the character recognition unit selects the dictionary image or the high-resolution dictionary image based on a filling rate or a frequency of the dictionary image specified by the specification processing unit. apparatus. A camera that shoots the license plate of the passing vehicle,
Based on the image of the license plate photographed by the camera, a recognition unit for recognizing characters written on the license plate,
A database storing representative patterns of groups generated by clustering a plurality of feature patterns of a plurality of characters in order to recognize characters by the recognition unit, and
The recognition unit recognizes characters described in the license plate based on an image of the license plate photographed by the camera and the representative pattern stored in the database. . The database stores the representative pattern of each group in association with the feature pattern of the group to which the representative pattern belongs,
The recognizing unit identifies the representative pattern corresponding to the image of the license plate photographed by the camera, and based on the characteristic pattern of the group to which the identified representative pattern belongs, the character described on the license plate The license plate recognition system according to claim 9, wherein the license plate recognition system is recognized.

Images