JP2010026805A - Character recognition device and character recognition method - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a character recognition device and a character recognition method for accurately recognizing characters to be recognized even though those characters are not clear. <P>SOLUTION: The character recognition device includes: an image acquisition means (2) for acquiring an inspection image obtained by photographing an object to be inspected; a character area detection means (34) for detecting a character area in which characters are projected from the inspection image; a stroke detection means (38) for detecting pixels matched with any of a plurality of strokes representing the partial structure of each of a plurality of characters to be recognized about at least a partial area of the character area, and for detecting the number of strokes included in at least the partial area of the character area by counting the number of the matched pixels; and a pattern identification means (39) for calculating the certainty of the respective characters to be recognized from the number of the strokes, and for deciding that the characters projected in the character area are characters whose certainty is the highest among the characters to be recognized. <P>COPYRIGHT: (C)2010,JPO&INPIT

Description

  The present invention relates to a character recognition device and a character recognition method, and more particularly, to a character recognition device and a character recognition method for recognizing a character from an image obtained by photographing a character written on an inspection object.

  In recent years, methods and apparatuses for recognizing characters by acquiring character information printed or stamped on an inspection object as an image and analyzing the image have been actively developed. In such a character recognition device and character recognition method, a feature amount that is considered to represent a different feature for each character type is extracted from an image, and a character is recognized based on the feature amount. Alternatively, if the font and size of the character to be recognized are known in advance, a character template corresponding to the font and size is prepared, and pattern matching between the image and the template is performed, and the best match is achieved. Recognize characters by specifying a template.

  However, if the surface of the object to be inspected becomes dirty and noise information is superimposed on the character string information, the feature value cannot be extracted or an incorrect pattern matching result can be obtained. It may not be recognized. Also, even when a single character is printed as a set of dots as in a dot matrix printer, the discontinuity between dots acts in the same way as noise, so that the character may not be recognized accurately. Therefore, as a preprocessing, a method of recognizing characters after performing processing that attenuates noise information has been developed.

  For example, in the image data processing method for optical character recognition disclosed in Patent Document 1, a histogram of density values of pixels of image data is created, and the dynamic range of the image data is once reduced by performing 2-D spatial averaging. After that, the range of contrast is expanded with respect to an area that is considered to include character information, edge enhancement processing is further performed, and binarization is performed finally, thereby obtaining image data with reduced noise.

JP-A-5-314315

  However, if the character to be recognized overlaps the scratch or dirt on the surface of the object to be inspected and a part of the character is unclear, information on the image about the character (for example, density, edge strength, etc.) It may not be obtained accurately. In such a case, even if the pre-processing as described above is performed, it is not possible to emphasize only the information related to the characters, and thus it is not possible to expect an improvement in recognition accuracy.

  In view of the above problems, an object of the present invention is to provide a character recognition device and a character recognition method that can accurately recognize a character even if the character to be recognized is unclear.

According to the first aspect of the present invention, there is provided a character recognition device for recognizing characters written on an inspection object. The character recognition apparatus includes an image acquisition unit (2) that acquires an inspection image obtained by photographing an object to be inspected, a character region detection unit (34) that detects a character region in which a character is reflected from the inspection image, and a character region. For at least a part of the region, a pixel that matches one of a plurality of line elements representing the partial structure of each of the plurality of characters to be recognized is detected, and the number of the matching pixels is counted, thereby detecting the character. A line element detecting means (38) for detecting the number of each line element included in at least a part of the area, and a certainty factor of each recognition target character is calculated from the number of each line element, and the character shown in the character area Pattern recognition means (39) for determining that the character has the highest certainty among the recognition target characters.
The character recognition device according to the present invention has such a configuration, so that even if a character to be recognized is overlaid with scratches, dirt, etc., and the character is unclear, the character is accurately detected. Can be recognized.

  According to a second aspect of the present invention, the pattern identifying means (39) is preferably constituted by a support vector machine that receives the number of each line element and outputs the certainty factor of each recognition target character. The character recognition device according to the present invention uses a support vector machine as a discriminator, so that even if the number of line elements corresponding to each recognition target character shows non-linear separation, the detected line Characters on the inspection object can be accurately identified based on the prime number.

According to a third aspect of the present invention, the character recognition device according to the present invention performs template matching between a character region and a plurality of templates corresponding to each recognition target character, and determines the degree of coincidence with each recognition target character. It further has a pattern matching means (36) for calculating and selecting a plurality of recognition target characters in order from the one with the highest degree of coincidence, and the pattern identification means (39) selects the character appearing in the character area. Of the plurality of recognition target characters, it is preferable to determine that the corresponding character has the highest certainty factor.
Since the character recognition device according to the present invention has such a configuration, the recognition accuracy can be further improved because the character on the object to be inspected is recognized using two different types of recognition means.

Furthermore, as described in claim 4, the pattern identifying means (39) is arranged so that at least a part of the plurality of recognition target characters is included in at least a part of the character area. It is preferred to determine the position or size.
Since the character recognition device according to the present invention has such a configuration and performs recognition processing by paying attention to a particularly different portion of each candidate character, the difference is clarified, so that the recognition accuracy is further improved. Can be improved.

  According to the fifth aspect of the present invention, there is provided a character recognition method for recognizing characters written on an inspection object. Such a character recognition method includes a step of acquiring an inspection image obtained by photographing an object to be inspected, a step of detecting a character region in which characters are reflected from the inspection image, and at least a part of the character region. Detecting pixels that match any of a plurality of line elements representing the partial structure of each of the plurality of characters, and counting the number of matching pixels, so that it is included in at least a part of the character region The step of detecting the number of each line element and the certainty of each recognition target character are calculated from the number of each line element, and the character reflected in the character area is the character with the highest certainty among the recognition target characters. Determining.

  In addition, the code | symbol in the parenthesis attached | subjected to each said means is an example which shows a corresponding relationship with the specific means as described in embodiment mentioned later.

Hereinafter, a character recognition device according to the present invention will be described in detail with reference to the drawings.
A character recognition device to which the present invention is applied is, for example, installed in a production line of a pump unit that supplies fuel to an on-vehicle engine, and is stamped on the surface of a part (work) of the pump unit that is an inspection object. It recognizes character strings such as date of manufacture. Then, the character recognition device first performs character recognition by pattern matching, and as a result, a character that is considered to have low recognition accuracy is detected from each character region, and the number of each line element that represents a partial feature of the character is input. It is recognized using a support vector machine that outputs a certainty factor representing the possibility of a specific character as a feature amount.

FIG. 1 shows a block diagram of a character recognition device 1 according to an embodiment of the present invention.
A character recognition device 1 according to an embodiment of the present invention captures a workpiece 5 that is an object to be inspected, an imaging unit 2 that acquires an inspection image, and a character string stamped on the surface of the workpiece 5 based on the inspection image. A processing unit 3 that recognizes and controls the character recognition device.
Hereinafter, each part of the character recognition device 1 will be described in detail.

The imaging unit 2 captures a character string stamped on the surface of the workpiece 5 and acquires an inspection image. The imaging unit 2 captures an image so that the entire character string is included in the inspection image and individual characters included in the character string can be identified on the inspection image. For this purpose, the imaging unit 2 includes a two-dimensional detector composed of a photoelectric converter such as a CCD or C-MOS sensor, and an imaging optical system that forms an image of the surface of the workpiece 5 on the two-dimensional detector. Have. In this embodiment, a 640 × 480 pixel 2/3 inch CCD is used as the two-dimensional detector, and the luminance value of each pixel is represented by 0 to 255. However, you may comprise the imaging part 2 with the two-dimensional detector which has a different pixel count and screen size. The imaging unit 2 adjusts the imaging magnification of the imaging optical system so that the characteristics of each character can be determined on the inspection image. Furthermore, the imaging unit 2 may include an illumination light source that illuminates the workpiece 5.
The imaging unit 2 transmits the acquired inspection image to the processing unit 3.

  FIG. 2 shows a schematic diagram of an inspection image acquired by the imaging unit 2. In FIG. 2, model number information 201 of the workpiece 5 and manufacturing date information 202 are imprinted as character string information on the surface of the workpiece 5 photographed in the inspection image 200. In the present embodiment, the model number information 201 is represented by an 8-digit numerical value. In addition, the manufacturing date information 202 includes a four-digit numerical value indicating a manufacturing year and a one-letter alphabet in order from the left. And the pixel corresponding to each character has a relatively high luminance value compared with the pixel corresponding to a background. However, for example, if the surface of the workpiece 5 is soiled or scratched, as in the vicinity of the numerical value “2” written on the leftmost side of the manufacturing date information 202, the pixel corresponding to the soiled or scratched surface is also changed. , Has a relatively high brightness. For this reason, the presence of such dirt and scratches may make the characters unclear and cause the characters to be recognized incorrectly.

The processing unit 3 includes a so-called personal computer (PC) and peripheral devices thereof, analyzes the inspection image acquired from the imaging unit 2, and recognizes character string information stamped on the surface of the workpiece 5.
FIG. 3 shows a functional block diagram of the processing unit 3. As shown in FIG. 3, the processing unit 3 includes a control unit 31, a communication unit 32, a storage unit 33, a character area detection unit 34, a normalization unit 35, a pattern matching unit 36, a contour extraction unit 37, and a line element detection unit 38. And a pattern identification means 39.
Hereinafter, each part of the processing unit 3 will be described.

The control means 31 is composed of a central processing unit (CPU) of a PC and its peripheral circuits, and operates according to a program read into the CPU, and controls each means of the imaging unit 2 and the processing unit 3. The communication unit 32 is a communication interface that transmits and receives control signals, image data, or data signals between the processing unit 3, the imaging unit 2, and other devices. USB, SCSI, RS232C, Ethernet (registered trademark) Etc. and I / O ports and their drivers. Then, the processing unit 3 receives the inspection image from the imaging unit 2 through the communication unit 32. On the other hand, the control signal generated by the control unit 31 is transmitted to the imaging unit 2 through the communication unit 32.
Further, the processing unit 3 outputs the recognized character information to an external device through the communication unit 32.

The storage means 33 is composed of a semiconductor memory such as a random access memory (RAM) or a read only memory (ROM), or a recording medium such as a magnetic disk or an optical disk and its access device, and temporarily stores an inspection image received from the imaging unit 2. Remember me. Further, the storage unit 33 stores a program for controlling the processing unit 3 and the like.
Further, the storage unit 33 stores a character pattern template corresponding to each character that may be engraved on the workpiece 5 (hereinafter referred to as a recognition target character). This character pattern template was created assuming that a part of the characters could not be identified on the inspection image due to scratches or dirt on the surface of the work 5 in addition to those corresponding to the ideal shape of the characters to be imprinted. Also included are patterns in which some characters are missing.
Further, the storage means 33 stores a line element template corresponding to a line element representing a partial structure of each character. Details of the line element template will be described later.

  The character area detection means 34, normalization means 35, pattern matching means 36, contour extraction means 37, line element detection means 38, and pattern identification means 39 are functional modules implemented by a program executed on the CPU, for example. Alternatively, these means may be mounted as a dedicated processing board including an image processing processor provided separately from the CPU.

The character area detecting means 34 detects a character area in which the character is shown for each character stamped on the work 5 from the inspection image. Various known methods can be used to detect the character area. For example, when the position and size of each character on the inspection image are known in advance, the character region detection unit 34 sets a character region for each character based on the position and size. Further, on the inspection image, the positional relationship between the position of the reference point having a predetermined shape of the workpiece 5 (for example, the end portion, the center of the workpiece 5, or the position of the predetermined marker) and each character is known. If so, the character area detecting means 34 detects the position of the reference point by performing pattern matching using a template corresponding to the shape of the reference point. The character area detecting unit 34 sets a character area for each character according to the position of the reference point and the positional relationship between the characters. In this case, the reference point may be any one of characters to be recognized.
The character area detection means 34 stores the position and size of each character area in the storage means 33 so that other means can also use it.

  The normalizing means 35 enlarges or reduces the character area so that the character size in each character area is constant. For this purpose, for example, the normalizing means 35 calculates the average luminance of the pixels in the character area. And the normalization means 35 detects the pixel located in the uppermost part in a character area among the pixels which have a brightness | luminance higher than the average brightness | luminance. Further, the normalizing means 35 is a pixel located on the lowermost side, a pixel located on the leftmost side, and a pixel located on the rightmost side among the pixels having luminance higher than the average luminance value of the pixels in the character area. Each pixel located is detected. Then, the normalizing means 35 enlarges or reduces the length of the character area in the vertical direction so that the difference between the detected upper pixel and lower pixel becomes a predetermined height (for example, 60 pixels). Normalize. Similarly, the normalizing unit 35 enlarges or reduces the horizontal length of the character area so that the difference between the detected leftmost pixel and rightmost pixel becomes a predetermined width (for example, 30 pixels). Normalize. The normalizing means 35 creates an image corresponding to the character area whose vertical and horizontal sizes are normalized as a normalized character area image. Then, the normalizing means 35 stores the normalized character area image in the storage means 33 so that it can be used by other means.

The pattern matching unit 36 performs pattern matching between the character pattern template read from the storage unit 33 and the normalized character region image to recognize characters included in the character region.
The pattern matching means 36 obtains a matching degree by changing the relative position and angle of the normalized character area image and each character pattern template in the search area by pattern matching, and the character pattern template that maximizes the matching degree. To decide. Note that the degree of coincidence (correlation coefficient) is obtained by, for example, the following equation.
Here, r is the degree of coincidence, and I and M represent the luminance values of the normalized character region image and the character pattern template, respectively. N represents the number of pixels included in the character pattern template. In this expression, when the normalized character area image and the character pattern template completely match, the matching degree r = 1, and when there is no correlation between the normalized character area image and the character pattern template, the matching degree r = 0. .

When the maximum value r _max of the matching degree is higher than a predetermined threshold value Th _pm , the pattern matching unit 36 has a character represented by the character pattern template corresponding to the maximum value r _max stamped in the character area. To do. Note that the predetermined threshold Th _pm is set based on an experimental result or the like so that the possibility that the recognition result becomes an error is very low when the maximum value r _max of the matching degree is higher than the threshold, for example, , 0.95.

On the other hand, when the maximum matching value r _max is equal to or smaller than a predetermined threshold value Th _pm , the pattern matching unit 36 puts characters represented in a predetermined number of character pattern templates in the character area in order from the highest matching score. Select a candidate character that is a candidate for the included character. The predetermined number may be a preset fixed value (for example, 2, 3, 5, etc.) or may be variable. When the predetermined number is variable, the pattern matching means 36 uses a character pattern whose matching degree r is higher than the candidate extraction threshold Th _c (which is lower than the above-mentioned threshold Th _pm , for example, set to 0.7). Select characters that appear in all templates as candidate characters. Further, the pattern matching unit 36 may set all recognition target characters as candidate characters when the maximum matching value r _max is lower than the candidate extraction threshold value Th _c .
The pattern matching unit 36 stores the candidate character in the storage unit 33 in association with the corresponding normalized character region image.

  The contour extracting unit 37, the line element detecting unit 38, and the pattern identifying unit 39 perform processing on the normalized character region image in which the specific character is not recognized and the candidate character is selected as a result of the processing by the pattern matching unit 36. Do.

  The contour extracting unit 37 extracts pixels corresponding to the contour of the character shown in the image from the normalized character region image. Therefore, the contour extracting unit 37 performs noise removal processing as preprocessing. Specifically, the contour extracting unit 37 performs a median filter (for example, a 3 × 3 pixel or 5 × 5 pixel median filter) process on each pixel of the normalized character region image. Alternatively, the contour extracting unit 37 performs an opening process of the morphological operation for pixels having a luminance value equal to or higher than the average luminance of the normalized character region image. In addition, the contour extracting means 37 is not limited to the above processing, or in combination with the above processing, other various noise removal processing (however, a processing that does not blur the contour is preferable for the normalized character region image. ) May be executed.

  Next, the contour extraction unit 37 performs binarization processing on the normalized character region image that has been subjected to noise removal processing, and each pixel of the normalized character region image is represented by a pixel corresponding to a character and other pixels. Divide into Therefore, the contour extracting unit 37 calculates the average luminance value, the maximum luminance value, and the minimum luminance value of the normalized character region image that has been subjected to the noise removal processing. Then, in order to correct the luminance value of the normalized character region image, the contour extracting unit 37 obtains a difference obtained by subtracting the average luminance value from a predetermined reference value (for example, 128) as an offset value, and the offset value is calculated for each pixel. Is added to the luminance value. Further, the contour extracting unit 37 sets the maximum luminance value and the minimum luminance value to which the offset value is added to a predetermined maximum value (for example, 255) and a predetermined value, respectively, in order to make the contrast of each normalized character region image constant. The gradation of the normalized character area image is linearly converted so as to be the minimum value (for example, 0).

  After performing the brightness value correction processing as described above, the contour extracting unit 37 determines a pixel having a brightness value higher than a predetermined reference value (for example, 128) for each pixel of the normalized character area image. 255. A pixel having a luminance value equal to or lower than the predetermined reference value is binarized with a luminance value of 0. As shown in FIG. 2, in the present embodiment, pixels corresponding to characters have a relatively high luminance value as compared with pixels corresponding to other characters. Therefore, each pixel of the normalized character area image is divided into a pixel corresponding to the character (luminance value 255) and other pixels (luminance value 0) by this binarization processing.

  The contour extraction means 37 performs edge detection processing on the binarized normalized character region image to extract pixels corresponding to the character contour. For example, the contour extraction unit 37 uses a first-order or second-order difference filter such as a Sobel filter, a Prewitt filter, or a Labracian filter as edge detection processing, and binarized normalized characters. Filter processing is performed on each pixel of the region image. Then, the contour extracting unit 37 assigns a contour luminance value (for example, 255) indicating that the pixel whose absolute value of the filter processing result is equal to or greater than a predetermined value (for example, 255) as an edge pixel corresponding to the contour. A contour image in which the luminance values of the other pixels are 0 is created.

Finally, the contour extracting unit 37 performs a thinning process on the contour image so that the contour line in which the edge pixels are connected is represented by one pixel width. Further, the contour extracting means 37 may perform a morphological calculation closing process on a pixel having a luminance value corresponding to the contour so that the contour lines separated by the influence of noise or the like are connected.
The contour extracting unit 37 passes the obtained contour image to the line element detecting unit 38.

The line element detection unit 38 obtains the number of line elements representing the partial structure of the character to be recognized, included in the contour image, as the input feature quantity for identification used in the pattern identification unit 39.
4A to 4H show examples of line element templates used in this embodiment. As shown in FIGS. 4A to 4H, each of the line element templates 401 to 408b is represented by 3 × 3 pixels, and a set of pixels represented by white lines is a line element. For example, the line element template 401 corresponds to a line element represented by a set of pixels arranged in a line in the horizontal direction. The line element template 402 corresponds to a line element represented by a set of pixels arranged in a line in the vertical direction. Each pixel corresponding to the line element has the above-described contour luminance value, and the other pixels have a luminance value of 0. The line elements are not limited to these, and may be expressed in a size different from the example of FIG. 4, for example, 5 × 5 pixels. Further, the line element may have different sizes in the vertical direction and the horizontal direction, such as horizontal 3 pixels × vertical 4 pixels or horizontal 5 × vertical 2 pixels.

  The line element detection means 38 calculates the number of pixels matching each line element for a partial region including the entire contour image or only a part of the contour image. In the present embodiment, the line element detection means 38 calculates the number of pixels that coincide with any of the above line element templates for each of the four partial regions obtained by dividing the contour image into two parts in the vertical and horizontal directions. Note that the method for setting the partial area is not limited to the above. For example, the outline image may be divided into three equal parts in the vertical direction, or each partial area may be set to have a different size. .

  Whether or not a specific pixel included in the contour image matches the line element template can be determined by pattern matching. For example, the line element detection means 38, when matching a specific line element template to a predetermined position of the contour image, if the degree of coincidence calculated by the same expression as the expression (1) is 1, It is determined that a line element corresponding to the line element template has been detected. Alternatively, the line element detection unit 38 checks whether or not the luminance values of the arbitrary pixel of interest in the contour image and the surrounding eight neighboring pixels match the luminance value of each pixel of the specific line element template, When the luminance values of all the pixels match, it may be determined that a line element corresponding to the line element template has been detected. Then, each time a specific line element is detected, the line element detection means 38 adds 1 to the number of detections of the specific line element. Since the line element templates 405a and 405b shown in FIG. 4E are symmetric with respect to each other, the line element detection unit 38 uses a line element template (405a, 405b) as one line. Treat as an element. In other words, the line element detection unit 38 sets the total number of detections of pixels matching either the line element template 405a or the line element template 405b as the detection number for one line element. Similarly, for each set of line element templates (406a, 406b), (407a, 407b), (408a, 408b) shown in FIGS. 4 (f) to (h), either one of the respective line element templates is used. The total number of detected matching pixels is taken as the number of detections for one line element. However, these two center line symmetrical line elements may be treated as different line elements.

A method for obtaining the number of detected line elements will be described in more detail with reference to FIGS. 5A shows an example 501 of a normalized character area image binarized, and FIG. 5B shows an example of an outline image showing the outline of a character extracted from the normalized character area image 501. 502 is shown. FIG. 5C is an enlarged view of the partial area 503 in the upper right part of the contour image 502.
In FIG. 5C, the calculation of the number of detections of each line element will be described by taking a region 504 surrounded by a frame line as an example. Here, paying attention to the pixel 511, the luminance values of the pixel 511 and the neighboring eight neighboring pixels match the luminance value of each pixel of the line element template 408b. Therefore, the line element detection means 38 adds 1 to the number of detected line elements corresponding to the set of line element templates (408a, 408b). On the other hand, when attention is paid to the pixel 512, the luminance values of the pixel 512 and the neighboring eight neighboring pixels match the luminance value of each pixel of the line element template 402. Therefore, the line element detection unit 38 adds 1 to the number of detected line elements corresponding to the line element template 402. Similarly, when attention is paid to the pixels 513 to 516, the luminance values of those pixels and the neighboring eight neighboring pixels also match the luminance values of the pixels of the line element template 402. Therefore, the number of detected line elements corresponding to the line element template 402 is 5. In the area 504, there is no place that matches the other line element template, so the line element detection unit 38 sets the number of other line elements detected to 0.
The line element detection means 38 passes the detected number of each line element included in each of the four partial areas divided into two parts, top, bottom, left and right, to the pattern identification means 39 as a feature amount.

The pattern identifying means 39 identifies characters appearing in the character area image from the candidate characters selected by the pattern matching means 36 based on the number of detected line elements included in each partial area. In the present embodiment, the pattern identifying means 39 is configured by a support vector machine.
FIG. 6 shows a conceptual diagram of the support vector machine. When a predetermined identification object belongs to any of a plurality of categories, the support vector machine can identify the identification object based on one or more feature amounts (hereinafter referred to as a feature amount set) obtained from the identification object. It is a discriminator that determines which category an object belongs to. The boundary between the categories is represented by a set of learning data feature amounts belonging to each category having the closest distance from the learning data feature amount set belonging to the adjacent category. A feature amount set representing a boundary between categories is called a support vector. In FIG. 6, each point indicated by a circle is one of feature quantity sets belonging to the category C1, and among these feature quantity sets 601 to 603 are support vectors of the category C1. Each point indicated by a rhombus is one of feature quantity sets belonging to the category C2, and among these feature quantity sets 604 to 606 are support vectors of the category C2. In the support vector machine, in order to improve the identification accuracy, the support vector is determined so that the distance (margin) between the support vector of category C1 and the support vector of category C2 is maximized. Furthermore, in the case of a support vector machine, even if the boundary between categories is non-linear, using a kernel function, the feature vector set of learning data is mapped to a higher dimension, and the support vector is determined to belong to each category. By making the feature quantity set linearly separable, good discrimination performance can be obtained.

In the present embodiment, the support vector machine is trained in advance using the detected number of each line element included in each partial region of the binarized normalized character region image corresponding to each recognition target character as a feature amount set. When learning, as a feature quantity set used as learning data, not only a sample in which the entire recognition target character can be seen completely, but also a part of the recognition target character (for example, about 1/10 to 10th of the number of pixels corresponding to the recognition target character A feature set extracted from a sample in which ¼ pixel) was unclear due to scratches or dirt was also used. Furthermore, a feature set extracted from the normalized character region image including pixels corresponding to scratches and dirt was also used as learning data.
Robustness as a discriminator of a support vector machine can be improved by using a feature amount set extracted from a sample containing scratches and dirt in the image as learning data.
The learned support vector machine is represented by a feature amount set corresponding to a support vector for each recognition target character. These feature quantity sets are stored in advance in the storage means 33 in association with each recognition target character.

When the pattern identification means 39 receives the number of detections of each line element included in each partial area from the line element detection means 38, the pattern identification means 39 uses it as an input feature quantity of the support vector machine, thereby confirming that it is a specific recognition target character. A certainty factor representing the likelihood is output for each candidate character. Then, when the maximum certainty value P _max of the certainty factor is higher than the predetermined threshold Th _pd , the pattern identifying means 39 is a character in which the candidate character corresponding to the maximum certainty value P _max is reflected in the character area. Is determined. On the other hand, when the maximum value P _max of the certainty factor is equal to or less than the threshold value Th _pd , the pattern identifying unit 39 determines that the character shown in the character area cannot be identified. Note that the threshold value Th _pd corresponds to a certainty factor that the reliability of the character recognition result is considered to be sufficient, and is appropriately set according to the recognition accuracy required for the character recognition device 1.

  Finally, the processing unit 3 arranges each character recognized by the pattern matching unit 36 or the pattern identification unit 39 according to the positional relationship of each character region, and obtains character string information stamped on the work 5. Then, the processing unit 3 displays the recognized character string information on the display to notify the user, or outputs it to other devices that are communicably connected via the communication unit 32.

  The operation when the character recognition device 1 to which the present invention is applied recognizes character string information stamped on one work 5 will be described with reference to FIG. The operation of the character recognition device 1 is controlled by the control means 31 of the processing unit 3.

  As shown in FIG. 7, when the inspection is started, the imaging unit 2 captures the surface of the workpiece 5 on which the characters are engraved, and acquires an inspection image (step S101). The acquired inspection image is transmitted to the processing unit 3. When the processing unit 3 receives the inspection image, the character region detection unit 34 detects a character region including the character for each character stamped on the work 5 from the inspection image (step S102). The details of the character area detection processing by the character area detection means 34 are as described above.

  When a character area corresponding to each character is detected, the processing unit 3 sets a character area of interest from the character areas (step S103). Then, the processing unit 3 enlarges or reduces the focused character area so that the character in the focused character area has a predetermined size by the normalizing unit 35, and generates a normalized character area image corresponding to the focused character area. Create (step S104).

Next, the processing unit 3 uses the pattern matching unit 36 to perform pattern matching between the normalized character area image corresponding to the target character area and each character pattern template of the recognition target character, and the matching degree for each character pattern template. Is calculated (step S105). Then, the pattern matching unit 36 compares the maximum value r _{max of the} degree of coincidence with a predetermined threshold value Th _pm (step S106). When the maximum value r _max of the matching degree is higher than a predetermined threshold value Th _pm , the pattern matching unit 36 has the character represented by the character pattern template corresponding to the maximum value r _max stamped in the target character area. Determination is made (step S107).

On the other hand, the pattern matching means 36 selects, as candidate characters, characters represented in a predetermined number of character pattern templates in order from the highest matching level when the maximum matching level r _max is equal to or smaller than a predetermined threshold value Th _pm. To do. Then, the processing unit 3 uses the pattern identification unit 39 to determine that the character stamped in the target character area is one of the candidate characters (step S108).

  After step S107 or step S108, the processing unit 3 determines whether or not all characters stamped on the workpiece 5 have been recognized (step S109). When there is a character that has not been recognized yet, the control is returned to step S103, and a new focused character area is set from the character areas that are not set as the focused character area. And the process part 3 repeats the process of step S104-S109. On the other hand, if it is determined in step S109 that all characters have been recognized, the processing unit 3 obtains character string information by arranging the recognized characters based on the positional relationship of the character regions, Character string information is output (step S110). Then, the processing unit 3 ends the recognition process.

With reference to FIG. 8, a detailed procedure of the process in step S108 will be described.
First, the contour extraction unit 37 of the processing unit 3 performs a noise removal process on the normalized character area image corresponding to the target character area (step S201). As described above, the contour extracting unit 37 executes any one or a combination of median filter processing, morphological operation opening processing, and various other noise removal processing as described above. Next, the contour extracting unit 37 performs binarization processing to distinguish the pixel corresponding to the character from other pixels (step S202). After that, the contour extraction unit 37 performs edge detection processing and creates a contour image in which the contour of the character is extracted (step S203). Further, the contour extracting unit 37 performs a thinning process on the contour image so that the contour line has a width of one pixel.

Next, the line element detecting means 38 of the processing unit 3 detects the number of line elements included in each of the four partial areas obtained by dividing the contour image into two equal parts in the vertical and horizontal directions. (Step S204). The details of the process for detecting the number of line elements are as described above with respect to the line element detection means 38.
When the number of detections of each of the plurality of line elements included in each partial region is obtained, the pattern identification unit 39 of the processing unit 3 calculates the certainty factor of each candidate character using the number of detections as an input feature amount (step S205). Then, the pattern identification unit 39 compares the certainty factor p _max with a predetermined threshold value Th _pd (step S206). If the maximum value p _max of the certainty factor is higher than the predetermined threshold value Th _pd , the pattern identifying unit 39 determines that the candidate character corresponding to the maximum value p _max is imprinted in the target character region (step) S207). On the other hand, if the maximum certainty factor p _max is equal to or smaller than the predetermined threshold value Th _pd , the pattern identifying unit 39 determines that the character in the target character region cannot be identified (step S208).

  As described above, the character recognition device according to an embodiment of the present invention first recognizes a character imprinted on the workpiece surface by pattern matching, and as a result, separately recognizes a character that is considered to have low recognition accuracy. Recognize using prepared support vector machine. As described above, the character recognition apparatus can obtain good recognition accuracy by recognizing characters on the workpiece surface using a plurality of different methods. Furthermore, since the number of detected line elements representing the partial structure of the character extracted from each partial area of the character area is used as the input feature quantity of the support vector machine, a part of the character is scratched or soiled. Even if it becomes unclear due to the above, it is possible to suppress the influence of dirt and the like. Therefore, such a character recognition device can accurately recognize a character even when a part of the character on the inspection object is unclear due to scratches or dirt.

The embodiments described above are for explaining the present invention, and the present invention is not limited to these embodiments.
For example, in the above embodiment, the pattern matching unit 36 may be omitted, and the character recognition may be performed directly by the pattern identification unit 39, that is, the support vector machine. In this case, since the candidate character is not selected, the pattern identifying unit 39 calculates the certainty factor for all the recognition target characters, and determines that the character corresponding to the maximum certainty factor is reflected in the character region. In addition, when it is known in advance that the size of characters written on the object to be inspected is constant, the normalization means 35 and the character area enlargement / reduction processing by the normalization means 35 may be omitted.

In addition, the pattern identifying unit 39 may use only the detected number of each line element in the partial area where all the selected candidate characters are different from each other as the input feature amount. In this case, the position and size of the partial area for obtaining the number of each line element used as the input feature amount can be set according to the combination of candidate characters, such as the upper half, the lower half, the upper left, and the lower right. For example, when the candidate characters are two characters “6” and “8”, the pattern identifying means 39 can set the partial region to be used as the region located in the upper right part of the character region. The correspondence relationship between the combination of candidate characters and the partial area to be used is defined in advance in the lookup table and stored in the storage means 33.
Further, in this case, the support vector machine used in the pattern identifying means 39 is prepared separately for each combination of candidate characters, and each support vector machine inputs the number of detected line elements only in the partial area to be used as an input feature quantity. It is preferable to learn in advance.

Further, the line element detection unit 38 does not perform contour extraction and preprocessing on the character region image, and directly targets pixels having a luminance value equal to or higher than a predetermined threshold (for example, the average luminance value of the character region image). After thinning, the number of line elements may be obtained. In this case, the thinning process is directly performed on the sample character area image and the number of each line element detected for each partial area is used as learning data to learn the support vector machine of the pattern identification means 39. .
Further, the pattern identifying means 39 may be constituted by another nonlinear classifier instead of the support vector machine. For example, the pattern identifying means 39 may be configured by a perceptron type neural network having three or more layers.

  As described above, various modifications can be made within the scope of the present invention according to the embodiment to be implemented.

It is a schematic block diagram of the character recognition apparatus to which this invention is applied. It is the schematic of the test | inspection image which image | photographed the workpiece | work of the to-be-inspected object. It is a functional block diagram of a processing part. (A)-(h) is a figure which shows an example of a line element, respectively. (A) is a figure which shows an example of the normalized character area image binarized, (b) is the outline which shows the outline of the character extracted from the normalized character area image shown to (a) It is a figure which shows an example of an image, (c) is an enlarged view of the partial area | region of the upper right part of the outline image shown to (b). It is a conceptual diagram of a support vector machine. It is a flowchart which shows the whole operation | movement of the character recognition apparatus to which this invention is applied. It is a flowchart which shows the operation | movement of a pattern identification process.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 Character recognition apparatus 2 Imaging part (image acquisition means)
DESCRIPTION OF SYMBOLS 3 Processing part 31 Control means 32 Communication means 33 Storage means 34 Character area detection means 35 Normalization means 36 Pattern matching means 37 Contour extraction means 38 Line element detection means 39 Pattern identification means 5 Workpiece

Claims (5)

Image acquisition means (2) for acquiring an inspection image obtained by photographing an inspection object on which characters are written;
A character region detecting means (34) for detecting a character region in which the character is reflected from the inspection image;
For at least a portion of the character area, detecting pixels that match any of a plurality of line elements representing the partial structures of the plurality of characters to be recognized, and counting the number of the matching pixels A line element detecting means (38) for detecting the number of each of the line elements included in the at least a part of the region,
A pattern identification unit that calculates the certainty factor of each recognition target character from the number of each line element and determines that the character shown in the character region is the character having the highest certainty factor among the recognition target characters ( 39),
A character recognition device comprising:   The character recognition device according to claim 1, wherein the pattern identification unit (39) is configured by a support vector machine that receives the number of each line element and outputs a certainty factor of each recognition target character. Template matching is performed between the character region and a plurality of templates corresponding to each recognition target character to calculate a degree of coincidence with each recognition target character, and a plurality of recognition target characters are selected in order from the highest matching degree. A pattern matching means (36) for selecting;
The said pattern identification means (39) determines with the character reflected in the said character area being the character with the highest said corresponding certainty among the selected several recognition object character, or 2. The character recognition device according to 2.   The said pattern identification means (39) determines the position or size of the said partial area | region so that a different part of the selected several recognition object character may be included in the said at least partial area | region. Character recognition device. Obtaining an inspection image obtained by photographing an object to be inspected with characters;
Detecting a character region in which the character is shown from the inspection image;
For at least a part of the character region, detecting a pixel that matches any of a plurality of line elements representing a partial structure of each of the plurality of characters to be recognized, and counting the number of the matching pixels Detecting the number of each line element included in the at least part of the region;
Calculating the certainty factor of each recognition target character from the number of each line element, and determining that the character shown in the character region is the character having the highest certainty factor among the recognition target characters;
A character recognition method characterized by comprising: