WO2014129018A1

WO2014129018A1 - Character recognition device, character recognition method, and recording medium

Info

Publication number: WO2014129018A1
Application number: PCT/JP2013/079265
Authority: WO
Inventors: 洋平小島; 杉本　喜一; 拓馬岡▲崎▼; 健太中尾
Original assignee: 三菱重工業株式会社
Priority date: 2013-02-25
Filing date: 2013-10-29
Publication date: 2014-08-28
Also published as: JP2014164528A; SG11201506568XA; JP6177541B2; MY177406A

Abstract

In a target image, said target image being a grayscale image, this character recognition device detects both a character-height-direction extent, i.e. the extent in a character-height direction perpendicular to a character-width direction within which a string exists, and a character-width-direction extent for each character in said string, i.e. the extent in the character-width direction within which each character exists. For each character in the string, the character recognition device then determines a candidate character-center region that contains the center of the character-height-direction extent in the character-height direction and the center of the character-width-direction extent of that character in the character-width direction, and for each point in that candidate character-center region, the character recognition device sets a plurality of rectangular candidate character regions such that the diagonals of each of said candidate character regions intersect at that point. The character recognition device performs a character recognition process on the part of the target image corresponding to each of the plurality of candidate character regions, thereby obtaining a character recognition result for each of the plurality of candidate character regions for each character in the string, and for each character in the string, the character recognition device finalizes a character recognition result from the character recognition results for the plurality of candidate character regions for that character.

Description

Character recognition device, character recognition method, and recording medium

The present invention relates to a character recognition device, a character recognition method, and a recording medium, and more particularly, to a character recognition device, a character recognition method, and a recording medium for performing character recognition on a grayscale image.

The technology for performing character recognition on a captured image captured by an imaging apparatus is applied to various uses including license plate recognition. In character recognition for a captured image, generally, processing for generating a binarized image from the captured image is performed, and character recognition processing is further performed for the binarized image. Such a technique is disclosed in, for example, Japanese Patent No. 3411785 (Patent Document 1).

One problem with the technique of generating a binarized image from a captured image and performing character recognition processing on the binarized image is that a good binarized image cannot be obtained if the image quality of the captured image is poor. It is. For example, if the contrast is insufficient in the captured image, a binarized image in which adjacent characters are connected in the binarization process is generated, which reduces the accuracy of character recognition. Japanese Patent No. 3798582 (Patent Document 2) discloses a technique for separating characters in contact in a binarized image, but this technique has a problem that a good binarized image cannot be obtained. It is not a fundamental solution.

In order to deal with such a problem, the inventors consider performing character recognition processing without performing binarization processing, that is, performing character recognition processing on a grayscale image (grayscale image). is doing. According to the study by the inventors, one problem in performing character recognition processing without performing binarization is that each character exists in an area candidate (hereinafter referred to as a “character candidate area”). There is an appropriate setting. If character candidate regions are set inappropriately when performing character recognition processing on a grayscale image, an incorrect character recognition result is obtained, that is, erroneous character recognition occurs. In a binarized image, after removing noise, a process of setting a portion where black pixels are continuously present as a character candidate region (when the background is white) may be performed. Such processing cannot be performed on the scale image. Although it is possible to suppress erroneous recognition of characters by appropriately setting a large number of character candidate regions, on the other hand, setting an excessively large number of character candidate regions is an important factor in the amount of data processing in character recognition processing. Increase, that is, increase the processing time.

From such a background, it is required to provide a technique for appropriately setting a character candidate area when performing character recognition processing on a grayscale image.

Japanese Patent No. 3411795 Japanese Patent No. 3798582

Therefore, an object of the present invention is to provide a technique for appropriately setting a character candidate region when performing character recognition processing on a grayscale image.

In one aspect of the present invention, a character recognition device includes a character height direction range detection unit, a character width direction range detection unit, a character center candidate region determination unit, a character candidate region setting unit, a character recognition unit, Character identification result determination means. The character height direction range detecting means is a grayscale image and the character width direction in which the character string exists for a target image that is an image showing a character string including a plurality of characters arranged in the character width direction. The character height direction range which is the range in the character height direction perpendicular to the character height is detected. The character width direction range detecting means detects a character width direction range that is a range in the character width direction in which each character of the character string exists for the target image. The character center candidate area determining means includes, for the target image, a character center for each character of the character string as an area including the center of the character height direction range in the character height direction and the center of the character width direction range in the character width direction. A candidate area is determined. The character candidate area setting means sets, for each point in the character center candidate area, a plurality of rectangular character candidate areas in which the intersections of diagonal lines match each point. The character recognition means performs character recognition processing on each of the plurality of character candidate area portions of the target image, and obtains the respective character recognition results of the plurality of character candidate areas for each character of the character string. The character identification result determination means determines the character identification result for each character of the character string from the character recognition results of each of the plurality of character candidate areas.

In one embodiment, the character height direction range detecting means calculates a first projection histogram, which is a distribution of the sum of luminance values of pixels arranged in the character width direction, calculated for each position in the character height direction of the target image. Generate and smooth the first projection histogram to calculate first smoothed data, and set the upper end of the range in the character height direction as the position where the sign of the difference between the first projection histogram and the smoothed data is inverted. And the lower end are detected.

In one embodiment, the character width direction range detection means generates a second projection histogram that is a distribution of the sum of luminance values of pixels arranged in the character height direction, calculated for each position in the character width direction of the target image. Then, smoothing processing is performed on the second projection histogram to calculate second smoothed data, and the start point candidate and the end point candidate are determined from the position where the sign of the difference between the second projection histogram and the second smoothed data is inverted. Is detected, the start point of the character width direction range of each character of the character string is selected from the start point candidates, and the end point of the character width direction range of each character of the character string is selected from the end point candidates.

As a preferred application example, the present invention is applied when the target image is a plate area image on which a license plate is projected or an image obtained by performing image processing on a plate area image. In one embodiment, the target image may be an image obtained by performing preprocessing on a plate area image on which a license plate is displayed. In the pre-processing, the luminance of the high-brightness portion is suppressed with respect to the plate region image, and linear conversion is performed so that the average value and the standard deviation of the pixel luminances have predetermined values, respectively. A brightness correction process is performed, an image obtained by the brightness correction process is processed by a differential filter to generate an edge image, and a hole filling process is performed to fill the inside of the character outline included in the edge image. , Processing to remove low-frequency components in the character width direction is performed on the image obtained by the filling process

In another aspect of the present invention, the character recognition method is a grayscale image, and there is a character string for a target image that is an image showing a character string including a plurality of characters arranged in the character width direction. A step of detecting a character height direction range that is a range in a character height direction perpendicular to the character width direction, and a character width direction range that is a range in the character width direction in which each character of the character string exists for the target image And a character center candidate area for each character of the character string as an area including the center of the character height direction range in the character height direction and the center of the character width direction range in the character width direction. Determining, for each point of the character center candidate region, a step of setting a plurality of rectangular character candidate regions whose diagonal intersection points coincide with each point, and a plurality of character candidates of the target image Character recognition processing is performed on each part of the region, and for each character of the character string, a step of obtaining each character recognition result of the plurality of character candidate regions, and a character recognition result of each of the plurality of character candidate regions, Determining a character identification result for each character of the character string.

In still another aspect of the present invention, a program recorded on a recording medium causes an arithmetic unit to execute the following steps: a character string that is a grayscale image and includes a plurality of characters arranged in the character width direction A step of detecting a character height direction range that is a range in a character height direction perpendicular to the character width direction, in which the character string exists, and for each character of the character string for the target image A step of detecting a character width direction range that is a character width direction range in which the character string exists, including the center of the character height direction range in the character height direction and the center of the character width direction range in the character width direction for the target image Determining a character center candidate region for each character of the character string as a region, setting a plurality of character candidate regions for each position of the character center candidate region, A step of performing character recognition processing for each of a plurality of character candidate areas of an image to obtain a character recognition result of each of the plurality of character candidate areas, and a character string from the character recognition results of each of the plurality of character candidate areas Determining a character identification result for each of the characters.

According to the present invention, character candidate areas can be appropriately set when character recognition processing is performed on a grayscale image.

It is a block diagram which shows the structure of the character recognition apparatus in one Embodiment of this invention. It is a flowchart which shows the outline | summary of the character recognition method in one Embodiment of this invention. It is a flowchart which shows the content of the pre-process in the character recognition method of this embodiment. It is a flowchart which shows the content of the character height direction range detection process in the character recognition method of this embodiment. It is a conceptual diagram which shows the content of the character height direction range detection process in the character recognition method of this embodiment. It is a conceptual diagram which shows the example of a structure of a license plate. It is a flowchart which shows the content of the character height direction range detection process performed about the number plate of FIG. 5A. It is a flowchart which shows the content of the character width direction range detection process in the character recognition method of this embodiment. It is a conceptual diagram which shows the content of the character width direction range detection process in the character recognition method of this embodiment. It is a flowchart which shows the content of the character center candidate area | region setting process in the character recognition method of this embodiment. It is a conceptual diagram which shows the content of the character center candidate area | region setting process in the character recognition method of this embodiment. It is a flowchart which shows the content of the character area and character recognition result decision process in the character recognition method of this embodiment. It is a conceptual diagram which shows the relationship between the specific point of a character center candidate area | region, and the character candidate area | region set about this specific point. It is a conceptual diagram which shows the character candidate area | region set in the character area and character recognition result confirmation process in this embodiment.

FIG. 1 is a block diagram showing a configuration of a character recognition device 1 according to an embodiment of the present invention. In the present embodiment, the character recognition device 1 is used for license plate recognition. In other words, the character recognition device 1 performs character recognition processing on the plate area image data 21 supplied from the outside to generate plate recognition data 22. Here, the plate area image data 21 is image data of a plate area image, that is, an image in which a license plate (automobile registration number mark or vehicle number mark) is projected. For example, from a captured image obtained by photographing a vehicle, It is obtained by image processing that cuts out a portion corresponding to the license plate. It should be noted that the plate area image data 21 is grayscale image data. The plate recognition data 22 is data indicating a character recognition result (that is, a recognized character) obtained by the character recognition process.

The character recognition device 1 includes an image processing IC (integrated circuit) 2, an external interface 3, an external storage device 4, a memory 5, and a ROM (read only memory) 6. The external interface 3 supplies the plate area image data 21 received from the outside to the image processing IC 2. The external storage device 4 stores data generated in the character recognition process in the character recognition device 1. Data stored in the external storage device 4 includes plate recognition data 22. The memory 5 is used as a working area for arithmetic processing by the image processing IC 2. The ROM 6 stores a program executed by the image processing IC 2. The program stored in the ROM 6 includes character recognition software 6a that is a program for executing character recognition processing.

When a rewritable nonvolatile memory (for example, a flash memory) is used as the ROM 6, the recording medium storing the character recognition software 6a may be used for installing the character recognition software 6a in the ROM 6. The character recognition software 6a may be stored in the external storage device 4. In this case, a recording medium for storing the character recognition software 6a may be used for installing the character recognition software 6a in the external storage device 4.

The image processing IC 2 includes an arithmetic module 11, an image input interface 12, a data input / output interface 13, a memory controller 14, and a ROM controller 15. The arithmetic module 11, the image input interface 12, the data input / output interface 13, the memory controller 14, and the ROM controller 15 are connected by an internal bus 16. The arithmetic module 11 executes the character recognition software 6 a while using the memory 5 as a working area, and performs character recognition processing on the plate area image data 21. The image input interface 12 is an interface used for inputting the plate area image data 21 to the image processing IC 2. The data input / output interface 13 is an interface for accessing the external storage device 4. The memory controller 14 is an interface for accessing the memory 5. The ROM controller 15 is an interface for accessing the ROM 6.

Each process of the character recognition method described below is executed using the hardware shown in FIG. Below, the character recognition method of this embodiment is demonstrated in detail.

FIG. 2 is a flowchart showing a character recognition method in this embodiment. The character recognition software 6a described above is a program code group for executing this character recognition method.

The character recognition method of the present embodiment generally includes pre-processing (step S01), character height direction range detection processing (step S02), character width direction range detection processing (step S03), and character center candidate position region setting. Processing (step S04) and character region / character recognition result determination processing (step S05) are included. Hereinafter, each processing will be described in detail.

1. Preprocessing First, preprocessing (step S01) is performed on the plate area image data 21. The preprocessing is image processing for eliminating disturbance components around the character as much as possible and emphasizing a line (character line) constituting the character. It should be noted that the preprocessing described below is suitable for performing highly accurate character recognition, but is not essential in principle.

In the pre-processing (step S01), first, a luminance correction process is performed (step S11). In the brightness correction process, first, a process of suppressing the brightness of a portion that is high due to halation or the like is performed, and further, the average value and standard deviation of the brightness of the pixels are set to predetermined values, respectively. Linear transformation is performed.

Further, the image data obtained by the brightness correction in step S01 is processed by a Sobel filter or other differential filter, and an edge image is generated (step S12). A character outline appears in the edge image.

A process of filling the inside of the character outline included in the edge image is performed on the obtained edge image, and a filled image is generated (step S13). Such hole filling processing can be realized, for example, by performing processing using a maximum value filter and a minimum value filter on the edge image. In this case, the filter size in the character width direction (that is, horizontal direction) is selected to be, for example, about the thickness of the character line, and the filter size in the character height direction (that is, vertical direction) is selected to be one pixel. . Here, it should be noted that the plate area image and the edge image as a whole have characters arranged in the horizontal direction.

Further, a process of removing low frequency components in the character width direction is performed on the hole-filled image obtained by the process in step S13 (step S14). This processing is performed in order to remove an image other than the letters of the license plate, such as a bumper or trunk of a vehicle, when the image is included in the hole-filled image. More specifically, the processing is performed by performing a process using a minimum value filter, and further performing a process using a maximum value filter, and then calculating a difference from the hole-filled image obtained in step S13. The difference image data is the image data obtained in step S14.

This completes the pre-processing. Image data of the image obtained by the preprocessing is stored in the external storage device 4. Here, it should be noted that the image obtained by the preprocessing is a grayscale image (not a binarized image).

2. Character Height Direction Range Detection Processing Subsequently, character height direction range detection processing is performed on the image data obtained by the preprocessing (step S01) (step S02). In the character height direction range detection process, the upper and lower ends of the character height direction range (character height direction range) in which the character string exists in the image to be processed (that is, the image obtained by the preprocessing) It is a process to detect. FIG. 4A is a flowchart illustrating processing performed in the character height direction range detection processing, and FIG. 4B is a diagram conceptually illustrating the character height direction range detection processing.

First, a y-axis projection histogram is generated (step S21). Here, the y-axis projection histogram is a distribution of the sum of luminance values of pixels arranged in the character width direction, calculated for each position in the character height direction of the image. As shown in FIG. 4B, when an xy coordinate system in which the y axis is defined in the vertical direction of the image and the x axis is defined in the horizontal direction is defined, the luminance of the pixel at the coordinates (x, y) is represented by B ( x, y), the y-axis projection histogram is obtained as S _Σ1 (y) defined by the following equation (1):
S _Σ1 (y) = ΣB (i, y) (1)
Here, Σ is the sum for all pixels whose y-axis coordinates are y. In FIG. 4B, a graph 31 represents a y-axis projection histogram S _Σ1 (y).

Further, the y-axis projection histogram S _Σ1 (y) is processed by a smoothing filter (smoothing process) to calculate smoothed data (step S22). The filter size of the smoothing filter is selected as a size corresponding to the height of one character. In the present embodiment, a process for calculating a moving average is used as the smoothing process. That is, as the smoothed data, moving average S ^ _.SIGMA.1 the y-axis projection histogram _S Σ1 (y) (y) is calculated. In FIG. 4B, a graph 32 represents the moving average S ^ _Σ1 (y).

Furthermore, the upper end and the lower end of the range in the character height direction are detected using the y-axis projection histogram S _Σ1 (y) and the moving average S ^ _Σ1 (y) (step S24). In the detection of the upper and lower ends of the character height direction range in step S24, first, the peak point 33, which is the position in the character height direction where the smoothed data takes the maximum value, is detected. In the present embodiment, the y-coordinate at which the moving average S ^ _Σ1 (y) has the maximum value is detected as the peak point 33. Further, a position where the value of the smoothed data is larger than the value of the y-axis projection histogram S _Σ1 (y) starting from the peak point 33 is searched, and the value of the smoothed data is the y-axis projection histogram S _Σ1 (y). Positions larger than the value are detected as the positions of the upper end and the lower end of the character height direction range. In FIG. 4B, the code | symbol 34 has shown the upper end of the character height direction range detected by step S24, and the code | symbol 35 has shown the lower end of the character height direction range.

In this embodiment in which the moving average S ^ _Σ1 (y) is used as the smoothed data, there is a character at a position where the moving average S ^ _Σ1 (y) is larger than the value of the y-axis projection histogram _SΣ1 (y). It is detected as the position of the upper end 34 and the lower end 35 of the area to be performed. Here, the position where the moving average S ^ _Σ1 (y) becomes larger than the value of the y-axis projection histogram _SΣ1 (y) for the first time may be detected as the positions of the upper end 34 and the lower end 46. Further, when the moving average S ^ _Σ1 (y) is larger than the value of the y-axis projection histogram _SΣ1 (y) over a predetermined number of pixels in the character height direction, any one of the pixels (typically Specifically, the y coordinate of the center pixel) may be detected as the positions of the upper end 34 and the lower end 46.

Further, it is determined whether the plate area image is an image of the first plate or the second plate (step S24). When the plate region image is a two-stage plate image, in the processing of steps S21 to S23, processing for detecting the upper end 34 and the lower end 35 is performed for one of the two character strings included in the license plate. Become. Therefore, when the plate region image is a two-stage plate image, the same processing as the processing in steps S21 to S23 is performed on the other character string on which the processing for detecting the upper end 34 and the lower end 35 is not performed. The upper end and the lower end of the character height direction range of the other character string are detected (step S25).

Here, if it is known in advance whether the license plate transferred to the plate area image is the first plate or the second plate, it is either the first plate or the second plate. The determination (step S24) may not be performed. If the license plate imaged in the plate area image is a one-stage plate, the character height direction range detection processing is completed in the processing in steps S21 to S23 described above. If the license plate transferred to the plate area image is a two-stage plate, the above-mentioned character string for which the processing for detecting the upper end 34 and the lower end 35 is not performed in step S25 is performed again. The same processing as that in steps S21 to S23 is performed.

For example, as shown in FIG. 5A, the characters written on the license plate in Japan are the upper row in which the characters of the place name 53 and the characters of the classification number 54, which are composed of relatively small characters, are arranged in the horizontal direction. , And a lower character string 52 in which a series designation number 55 composed of relatively large characters and a hiragana 56 are arranged. Therefore, in the character recognition process for the license plate in Japan, it is not necessary to determine which image is the first plate or the second plate in step S24.

When character recognition is performed on a plate area image of a license plate in Japan, the upper end 34 and the lower end 35 of the character height direction range are detected for the lower character string 52 in the processing of steps S21 to S23. It will be. In this case, as shown in FIG. 5B, after the processing of steps S21 to S23 is performed, the same processing as the processing of steps S21 to S23 is performed on the upper character string 51, and the upper character string is processed. The upper end and the lower end of 51 in the character height direction range may be detected (step S26).

The upper and lower end data indicating the positions of the upper end and the lower end detected by the above-described character height direction range detection process are stored in the external storage device 4.

3. Character Width Direction Range Detection Processing Further, character width direction range detection processing is performed on the image data obtained by the preprocessing (step S01) (step S03). The character width direction range detection process is a horizontal start point candidate (start point candidate) and end point candidate (end point) of a horizontal range in which each character is considered to exist (hereinafter referred to as “character width direction range”). This is a process for detecting a candidate. It should be noted that the starting point candidates detected by the character width direction range detection process are only “candidates” and are not finally determined as the starting point in the horizontal direction of the area where each character exists. . Similarly, the end point candidate is merely a “candidate”, and is not finally determined as the end point in the horizontal direction of the region where each character exists.

FIG. 6 is a flowchart showing processing performed in the character width direction range detection processing, and FIG. 7 conceptually illustrates the character width direction range detection processing. First, an x-axis projection histogram is generated (step S31). Here, the x-axis projection histogram is a distribution of the sum of luminance values of pixels arranged in the vertical direction, calculated for each position in the horizontal direction of the image. Referring to FIG. 7, when the luminance of a pixel at coordinates (x, y) is B (x, y), the x-axis projection histogram is expressed as S _Σ2 (x) defined by the following equation (2). can get:
S _Σ2 (x) = ΣB (x, j) (2)
Here, Σ is the sum for all pixels whose x-axis coordinates are x. In FIG. 7, a graph 41 represents an x-axis projection histogram S _Σ2 (x).

Further, the smoothing data is calculated by performing processing (smoothing processing) by the smoothing filter on the x-axis projection histogram S _Σ2 (x) (step S32). The filter size of the smoothing filter is selected as a size corresponding to the width of one character. In the present embodiment, a process for calculating a moving average is used as the smoothing process. That is, as the smoothed data, moving average S ^ _.SIGMA.2 the x-axis projection histogram _S Σ2 (x) (x) is calculated. In FIG. 7, a graph 42 represents the moving average S ^ _Σ2 (x).

Further, using the x-axis projection histogram S _Σ2 (x) and the moving average S ^ _Σ2 (x), the start point candidate and the end point candidate of the character width direction range of each character are detected (step S24). In the present embodiment, the start point candidate and the end point candidate are detected by searching from one end of the image, more specifically from the left end. That is, the start point candidate is detected as the start point candidate of the character width direction range of each character, and the end point candidate is detected as the end point candidate of the character width direction range. In the detection of the start point candidate and the end point candidate, the difference between the moving average and the x-axis projection histogram is calculated for each x coordinate, and the sign of the difference between the moving average and the x-axis projection histogram is reversed for the start point candidate and the end point candidate. It is detected as a position. For example, if the difference S ^ _.SIGMA.2 minus x-axis projection histogram from the moving average (x) -S _Σ2 (x) is calculated for each x coordinate, the starting point candidates, the difference _{_{S ^ Σ2 (x) -S Σ2}} ( x) is calculated as an x coordinate in which the sign of “+” is inverted from plus to minus, and the end point candidate is calculated as an x coordinate in which the sign of the difference S _Σ2 (x) −S _Σ2 (x) is inverted from minus to plus. In FIG. 7, the detected start point candidate is indicated by a symbol “◯”, and the detected end point candidate is indicated by a symbol “□”.

Through the above processing, the processing for detecting the start point candidate and the end point candidate is performed for one of the character strings arranged in the horizontal direction. When the plate region image is a two-stage plate image, the above-described processing of steps S31 to S33 is performed for the other character string for which the start point candidate and the end point candidate are not detected (step S34). The positions of the detected start point candidates and end point candidates are stored in the external storage device 4 as start / end point candidate data.

4). Character center candidate position area setting process In the character center candidate position area setting process (step S04), the positions of the upper and lower ends of the character height direction range detected by the above-described character height direction range detection process, and the character width direction Based on the data of the start point candidate and the end point candidate detected by the range detection process, the character center candidate region of each character is determined. The character center candidate region is a region that is a candidate for a position where the center of each character exists. In the character area / character recognition result determination process (step S05) performed later, any position within the character center candidate area is determined as the position where the center of each character exists.

The range in the character height direction (that is, the y-axis direction) of the character center candidate region is determined as a region including the center position of the character height direction range detected by the character height direction range detection process (step S41). . The range in the character height direction of the character center candidate region is common to all characters included in the character string arranged in the horizontal direction. As an example, using the y coordinate y _MAX at the upper end, the y coordinate y _{MIN at the} lower end, and the predetermined value m detected by the character height direction range detection process, (y _MIN + y _MAX ) / 2−m y _MIN + y _MAX ) / 2 + m. In this case, the position candidate in the vertical direction of the center of each character has a degree of freedom of (2m) pixels. The reason why a certain degree of freedom is given to the candidate of the position where the center of each character exists is that it is assumed that the license plate is photographed somewhat obliquely in the plate region image.

Further, for each character, the range in the horizontal direction (that is, the x-axis direction) of the character center candidate region is determined (step S42). In the determination of the horizontal range of the character center candidate region for each character, first, the character width direction range corresponding to each character is determined from the start point candidate and the end point candidate detected by the character width direction range detection process described above. A start and end pair is selected. Here, the start point of the character width direction range is selected from the start point candidates detected by the character width direction range detection process, and the end point of the character width direction range is selected from the end point candidates. In the present embodiment in which the start point candidate and the end point candidate are detected by the search from the left side, basically, for the i th character from the left, the i th start point candidate from the left is the start point, and the i th from the left A certain end point candidate may be determined as the end point. However, there is a range that is considered to be appropriate for the width of the character, and when a start point and an end point that are outside the range are determined, a pair of start point and end point is used by using adjacent start point candidates or end point candidates instead. Is determined. By such a procedure, the pair of the start point and end point of the character width direction range of each character is determined.

The range in the horizontal direction (that is, the x-axis direction) of the character center candidate region is determined as a region including the center of the character width direction range determined for each character. In the present embodiment, from (x _MIN + x _MAX ) / 2−k to (x _MIN + x _MAX ) using the x coordinate x _{MIN of the} start point and the x coordinate x _{MAX of the} end point corresponding to each character and the predetermined value k. It is determined as a range up to / 2 + k. In this case, the position candidate in the horizontal direction of the center of each character has a degree of freedom of (2k) pixels. This is due to the fact that there is some variation in the width of the characters shown in the plate area image.

With the above processing, the processing for determining the character center candidate region for each character is performed for one of the character strings arranged in the horizontal direction. When the plate region image is a two-stage plate image, the above-described processing of steps S41 to S42 is performed for the other character string that has not been subjected to the processing for determining the character center candidate region (step S43). Data indicating the range of the determined character center candidate region (character center candidate region data) is stored in the external storage device 4.

5. Character Area / Character Recognition Result Determination Process Subsequently, a process for determining the area where each character exists and the character recognition result is performed (step S05).

FIG. 10 is a flowchart showing processing performed in the character area / character recognition result determination processing, and FIGS. 11 and 12 are diagrams conceptually explaining the character width direction range detection processing.

Specifically, as shown in FIG. 10, first, a character center candidate region corresponding to the character to be processed first (in this embodiment, the character located at the leftmost) is selected (step S50). .

Further, a plurality of character candidate areas having different heights and / or widths are set for each point (each pixel) inside the selected character center candidate area (step S51). For example, when the character center candidate region is 9 pixels of 3 × 3, nine character candidate regions corresponding to each pixel of the 9 pixels are set. In this embodiment, the character candidate area is a rectangle (that is, a rectangle (including a square)). FIG. 11 is a diagram illustrating a relationship between a specific point 61 inside the character center candidate region and the character candidate region 62 set for the specific point. The character candidate region 62 set for a specific point 61 inside the character center candidate region is determined so that the intersection of the diagonal lines 63 of the character candidate region 62 coincides with the specific point 61. As shown in FIG. 12, the character candidate area has a reference size (reference height and width), and each character candidate area has the same height and width as the reference size. And the width or the height and width increased or decreased from the reference size. The maximum increase / decrease amount of the height and width is set in advance as a parameter.

Subsequently, a character recognition process is performed on the image in each character candidate area (step S52). In this embodiment, character recognition processing is performed by pattern matching using a template (template matching method), and in the character recognition processing, a character recognition result and its evaluation value are determined. As the evaluation value, for example, the similarity calculated in the template matching method may be used.

Further, the character candidate area having the highest evaluation value and the character recognition result of the image of the character candidate area are extracted (step S53). For example, when the character center candidate region is 9 pixels, the character candidate region having the highest evaluation value among the nine character candidate regions and the character recognition result of the image of the character candidate region are extracted.

When the evaluation value is greater than or equal to a predetermined reference value (step S54: Yes), the character candidate area with the highest evaluation value and the character recognition result are the character area and character recognition result in which the character to be processed exists. Confirmed (step S55). Otherwise (step S54: No), the character recognition result is rejected assuming that a region not including characters is given as a character candidate region.

The processing in steps S51 to S55 is repeated until they are executed for all character center candidate regions (step S56). If there is a character center candidate area that has not been subjected to the processes of steps S51 to S55, the character center candidate area corresponding to the next character is selected (step S57), and the processes of steps S51 to S55 are performed again. When the processing of steps S51 to S55 is performed for all the character center candidate regions, the character region in which each character exists and the character recognition result are determined for all characters (step S58). Data indicating the confirmed character region and character recognition result is stored in the external storage device 4 as plate recognition data 22.

As described above, in the character recognition processing of the present embodiment, a character center candidate region is determined for each character based on the y-axis projection histogram and the x-axis projection histogram, and the character center candidate region inside the character center candidate region is determined. A character candidate area is determined for each point. In this embodiment, a character candidate area is set in a state where a certain degree of freedom is given to the center of each character, and erroneous recognition can be suppressed. On the other hand, the character center candidate area is limited to some extent by the detection result of the character height direction range by the character height direction range detection process and the detection result of the character width direction range by the character width direction range detection process. The number of character candidate areas to be set is suppressed to a certain extent. For this reason, the data processing amount of character recognition processing can be suppressed and processing time can be shortened.

The embodiment of the present invention has been specifically described above, but the present invention is not limited to the above embodiment. It will be apparent to those skilled in the art that the present invention can be implemented with various modifications. For example, in the above description, an embodiment in which the present invention is applied to license plate recognition (recognition of characters written on a license plate) is described. However, the present invention generally recognizes characters for a captured image that is a grayscale image. Applicable.

Claims

A character height direction perpendicular to the character width direction in which the character string exists for a target image that is a gray scale image and an image in which a character string including a plurality of characters arranged in the character width direction is displayed A character height direction range detecting means for detecting a character height direction range that is a range of
Character width direction range detection means for detecting a character width direction range that is a range in the character width direction in which each character of the character string exists for the target image;
About the target image, a character center candidate region for each character of the character string as a region including the center of the character height direction range in the character height direction and including the center of the character width direction range in the character width direction A character center candidate region determining means for determining
For each point of the character center candidate region, a character candidate region setting means for setting a plurality of rectangular character candidate regions in which the intersections of diagonal lines match the points,
Character recognition means for performing character recognition processing on each of the plurality of character candidate area portions of the target image and obtaining each character recognition result of the plurality of character candidate areas for each character of the character string; ,
A character recognition apparatus comprising: a character identification result determining unit that determines a character identification result for each character of the character string from a character recognition result of each of the plurality of character candidate regions.
The character recognition device according to claim 1,
The character height direction range detection means generates a first projection histogram that is a distribution of the sum of luminance values of pixels arranged in the character width direction, calculated for each position in the character height direction of the target image. Then, smoothing processing is performed on the first projection histogram to calculate first smoothed data, and the character height direction is set as a position where the sign of the difference between the first projection histogram and the smoothed data is inverted. A character recognition device that detects the top and bottom edges of a range.
The character recognition device according to claim 1 or 2,
The character width direction range detecting means generates a second projection histogram that is a distribution of the sum of luminance values of pixels arranged in the character height direction, calculated for each position in the character width direction of the target image. Then, smoothing processing is performed on the second projection histogram to calculate second smoothed data, and a start point candidate and an end point are determined from the position where the sign of the difference between the second projection histogram and the second smoothed data is inverted. A candidate is detected, the start point of the character width direction range of each character of the character string is selected from the start point candidates, and the end point of the character width direction range of each character of the character string is selected from the end point candidates Select a character recognition device.
A character recognition device according to any one of claims 1 to 3,
The character recognition device, wherein the target image is a plate area image showing a license plate or an image obtained by performing image processing on the plate area image.
A character recognition device according to any one of claims 1 to 3,
The target image is an image obtained by performing preprocessing on a plate region image on which a license plate is projected,
In the pre-processing, the plate region image is subjected to linear conversion so as to suppress the luminance of the portion where the luminance is high, and the average value and the standard deviation of the pixel luminance are respectively predetermined values. The edge correction is performed on the image obtained by the brightness correction process, an edge image is generated, and the inside of the character outline included in the edge image is filled. A character recognition device in which processing is performed and processing for removing low-frequency components in the character width direction is performed on an image obtained by the hole filling processing.
A character height direction perpendicular to the character width direction in which the character string exists for a target image that is a gray scale image and an image in which a character string including a plurality of characters arranged in the character width direction is displayed Detecting a range in the character height direction that is a range of
For the target image, detecting a character width direction range that is a range in the character width direction in which each character of the character string exists;
About the target image, a character center candidate region for each character of the character string as a region including the center of the character height direction range in the character height direction and including the center of the character width direction range in the character width direction A step of determining
For each point of the character center candidate region, setting a plurality of rectangular character candidate regions whose intersections of diagonal lines match the points;
Performing character recognition processing on each of the plurality of character candidate region portions of the target image to obtain each character recognition result of the plurality of character candidate regions for each character of the character string;
Determining a character identification result for each character of the character string from a character recognition result for each of the plurality of character candidate regions.
A recording medium for storing a program, wherein when the program is executed, the following steps are performed:
A character height direction perpendicular to the character width direction in which the character string exists for a target image that is a gray scale image and an image in which a character string including a plurality of characters arranged in the character width direction is displayed Detecting a range in the height direction of a character that is a range of
Detecting a character width direction range that is a range in the character width direction in which each character of the character string exists for the target image;
About the target image, a character center candidate region for each character of the character string as a region including the center of the character height direction range in the character height direction and including the center of the character width direction range in the character width direction Step to determine,
Setting a plurality of character candidate regions for each position of the character center candidate region;
Performing character recognition processing on each of the plurality of character candidate regions of the target image to obtain each character recognition result of the plurality of character candidate regions; and
A recording medium that causes a computing device to execute a step of determining a character identification result for each character of the character string from character recognition results of each of the plurality of character candidate regions.