JP2000322579A - Image recognizing device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To quickly recognize an image with high accuracy even for an image where the position of a characteristic part is greatly different such as securities. SOLUTION: A sample image is recognized by comparing a prescribed area of the sample image to be compared with a template 3 being the reference. In such a case, the template 3 is composed of the pixels of a plurality of areas 4a to 4o which are selected from the entire image 4 to be recognized and whose respective positions are different, and is also provided with a reference table showing the positions of the pixels constituting the template 3, and the pixels of the sample image are taken out by position designation of the reference table and are compared with the template 3.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

[0001] The present invention relates to an image recognition device. More specifically, the present invention relates to an image recognition device that recognizes an image obtained by reading securities such as coins and bills.

[0002]

2. Description of the Related Art In a pattern recognition device used for discriminating coins, when a recognition target is taken into the device, an image of the recognition target is taken and a comparison pattern is created based on the image. For example, as shown in FIG. 4A, the comparison pattern 2 sets a ring-shaped template range 14 with respect to the read image 13 of the recognition target, cuts out the image, and performs processing such as edge enhancement. B) As shown in B)
× Created by performing coordinate conversion into a rectangular shape of 10 pixels vertically.

[0003] Then, image processing of the image of the object to be recognized,
Alternatively, the material, the outer diameter, the presence or absence of a hole, and the like of the recognition target are determined based on the detection result of another sensor. Based on the determination result of the material, outer diameter, and the like, a template for various kinds of coins prepared in advance, that is, a template that satisfies the conditions among the reference patterns is selected, and a matching operation for performing detailed comparison with the comparison pattern is performed. Like that. As a template,
128x in advance for various coins in the same way as the comparison pattern
A 10-pixel rectangular image that has been created is used.

Here, in the matching operation between the comparison pattern and the template, since the attitude of the coin, that is, the rotation angle is unknown, the comparison operation is performed while shifting both images relatively by one pixel in the longitudinal direction. I do. That is, the matching calculation is performed for one template by shifting it 128 times. For this reason, a long processing time is required for the matching calculation in each template. For example, when it is required to process 30 coins per second, the matching calculation of the template for one recognition target is performed. The number of sheets is limited by the processing speed of the apparatus used, and needs to be suppressed to, for example, about four sheets. On the other hand, a total of four templates are required to discriminate between the front and back of a new clean object and an old dirty object for one coin. Therefore, the outer diameter, material, etc. of one recognition object are required. It is very difficult to determine the same type of coin.

It is to be noted that 128 in one template
The highest value of the similarities obtained by the matching operations is adopted as the similarity of the comparison pattern for this template. Further, the type of the recognition target is recognized by the one having the highest similarity among, for example, four templates on which the matching calculation has been performed.

[0006]

However, in the above-described recognition method using the image recognition apparatus, the template for performing the matching operation is selected based on the outer diameter, size, material, and the like of the recognition target. If there are many types of coins with the same symbols and materials, but different symbols, the number of templates to be subjected to matching calculation will be very large, exceeding four, and it will take a very long time to recognize the recognition target. And the processing capacity is reduced.

Further, since the position to be cut out as a template from a read image is fixed, in the case of securities or the like having different features, the cut-out area of the template becomes large and the processing time becomes long. I will.

Further, a large image may be used as a template to improve the recognition performance. However, since the matching operation requires a long time, the number of templates used in the matching operation is limited.

Accordingly, the present invention provides an image recognition apparatus capable of quickly and accurately recognizing an image even when the position of a characteristic portion such as a securities is greatly different. Aim.

[0010]

In order to achieve the above object, according to the first aspect of the present invention, a predetermined area of a sample image to be compared is compared with a reference template to recognize the sample image. In the image recognition apparatus, the template includes a plurality of pixels in a plurality of regions each having a different position selected from the entire image to be recognized, and further includes a reference table indicating the positions of the pixels forming the template, and specifying the position of the reference table. , The pixels of the sample image are extracted and compared with the template.

Therefore, since the template is composed of pixels in a plurality of regions at different positions, it is possible to create a template by selecting only characteristic portions of the entire image to be recognized. Therefore, the processing speed can be reduced by reducing the size of the template without reducing the recognition accuracy, as compared with the case of using a large template for fixing the template position and improving the recognition accuracy. In addition, if the processing time is the same as that of the related art, it is possible to perform identification determination with more images to be recognized.

According to a second aspect of the present invention, in the image recognition apparatus of the first aspect, the image recognition apparatus has a plurality of templates and has respective reference tables formed corresponding to the plurality of templates. Of these, the pixels of the sample image are extracted for each template using a reference table corresponding to the template.

Accordingly, the type of the sample image can be specified with high accuracy by comparing the sample image with the plurality of templates.

Further, the invention described in claim 3 is the invention according to claim 1.
3. The image recognition device according to 2, wherein the reference table has, as its elements, an address of a pixel included in the template, and further includes an address generation unit for sequentially generating an address of a pixel for performing a correlation operation, A correlation operation unit that performs a correlation operation between the pixels of the template specified by the unit and the pixels of the sample image specified by the reference table.

Therefore, when the address generator generates an address, a pixel of the template corresponding to the address is designated. The address generated by the address generation unit specifies the address of the pixel of the sample image corresponding to the address via the lookup table. Then, a correlation operation is performed by the correlation operation unit between the pixel of the designated template and the pixel of the sample image, and the similarity of the sample image to the template is determined.

According to a fourth aspect of the present invention, in the image recognition apparatus according to any one of the first to third aspects, the template is selected from pixels of a plurality of regions selected from the entire image to be recognized and having different positions. And from
A predetermined process is performed for each of the plurality of pixels to form a template, a neighborhood address generation unit that generates an address of a neighborhood pixel of the pixel specified by the reference table, and a neighborhood address generation unit that is designated by the neighborhood address generation unit. And an arithmetic unit for performing a predetermined process on the data of a plurality of pixels. The output of the arithmetic unit is correlated with the template.

Therefore, when creating a template,
Processing such as edge enhancement and averaging is performed on pixels in a plurality of regions in the image to be recognized. When the address generator generates an address, a pixel of the template corresponding to the address is designated. In addition, an address of a pixel in the vicinity of a pixel of the sample image specified by the address generation unit via the lookup table is generated by the neighborhood address generation unit. Further, the arithmetic unit performs a predetermined process on data of a plurality of pixels of the sample image specified by the neighborhood address generation unit. Then, the correlation between the output of the calculation unit and the template is calculated, and the similarity of the sample image to the template is determined.

Further, the invention according to claim 5 provides the invention according to claim 1.
In the image recognition device described in any one of the above items 4 to 4, the reference table shifts an address designated for extracting a pixel of the sample image by a designated shift amount.

Therefore, by shifting the designated address by the designated shift amount in the reference table, the correlation operation can be performed while shifting the template and the sample image within the range of the designated pixel. Reduction can be suppressed.

[0020]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS The configuration of the present invention will be described below in detail based on an example of an embodiment shown in the drawings. 1 and 2 show an embodiment in which the image recognition device 1 of the present invention is applied to a device for identifying securities such as bills and tickets. This image recognition device 1 is a sample image 2 to be compared.
Is compared with the reference template 3 to recognize the sample image 2. The image recognition apparatus 1 includes the template 3 including a plurality of pixels of different regions selected from the entire image 4 to be recognized and having different positions, as shown in FIG. Is provided, and the pixels of the sample image 2 are extracted by designating the position of the reference table 5 and compared with the template 3. Therefore, the template 3 can be created by selecting pixels in a plurality of regions of the characteristic portions 4a to 4o of the entire image 4 to be recognized. Compared with the case where a large template including a fixed and uncharacteristic part is used, the processing speed can be reduced by reducing the size of the template 3 without lowering the recognition accuracy.

The image recognition apparatus 1 has a plurality of templates 3 as shown in FIG. 2 and lookup tables (hereinafter referred to as LUTs and LUTs) as respective reference tables 5 formed corresponding to the plurality of templates 3. 5). The LUT 5 has an address of the image 4 to be recognized for each pixel of the template 3. Therefore, by referring to the LUT 5, the LU
The address of the image 4 to be recognized corresponding to each pixel of the template 3 corresponding to T5 can be obtained.

Here, since the LUT 5 is uniquely set for each template 3, each LUT 5 is completely different depending on what kind of template 3 is created. For example, when characteristic portions 4a to 4e are present in the image 4 to be recognized shown in FIG. 1A, the template 3 is created by combining these. FIG. 1 (B)
When the characteristic portions 4f to 4i exist in the image 4 to be recognized shown in FIG.
Create Further, when characteristic portions 4j to 4o exist in the image 4 to be recognized shown in FIG. 1C, the template 3 is created by combining these. The position and size of each of the characteristic portions 4a to 4o are determined by, for example, the template creator based on experience.

At this time, the characteristic portions 4a to 4e are substantially square and have the same size as the characteristic portions 4a to 4e, or the characteristic portions 4a to 4e are the same as the characteristic portions 4f to 4i.
The size of f-4i may differ, or the characteristic portions 4j-4
o, the characteristic portions 4j to 4o are made to have the same size as an elongated rectangle, or to an arbitrary shape as shown in FIGS. 9 and 10, or to a circular or polygonal shape. Since the size and size are arbitrarily different, the address of each pixel of the template 3 formed by combining these characteristic portions in the entire image 4 to be recognized, that is, each LUT 5 can be completely different. . Although the size of each template 3 is different, a characteristic portion is selected in each case, so that the processing speed can be reduced by reducing the size of the template 3 without lowering the recognition accuracy. .

Further, as shown in FIG. 2, the image recognition apparatus 1 includes an address generator 6 for sequentially generating the addresses of the pixels for performing the correlation operation, and a pixel of the template 3 designated by the address generator 6. A correlation operation unit for performing a correlation operation with the pixel of the sample image specified by the reference table; Further, the image recognition device 1 includes a determination unit 8 that determines the similarity of the sample image 2 to the template 3 based on the result of the correlation operation.
Further, in the image recognition device 1, the LUT 5 has the sample image 2
And an address shift 12 for shifting an address specified by a specified shift amount in order to take out a pixel.

An operation procedure of the image recognition device 1 will be described.

As a preparation procedure, an image of an existing security to be recognized is taken. Then, several characteristic portions are set for the read image 4, and the image 4 is cut out at the same resolution. Alternatively, processing such as edge enhancement, thinning, and averaging may be performed as necessary. One template 3 is created by connecting the characteristic portions thus obtained. The size of each template 3 can be varied according to the size and characteristics of the read image. For example, increasing the size of the template 3 can improve the accuracy of image recognition, and reducing the size of the template 3 can improve the processing speed of image recognition. Also, an address indicating where each pixel of each template 3 is located in the entire original image is specified in the LUT 5.

Then, when actually performing the identification by reading the sheet, the image is taken of the sheet and the sample image 2 is read.
Create Further, the captured sample image 2 and the template 3 are compared.

Here, the CPU of the image recognition device 1 sequentially selects one from the plurality of templates 3 one by one and selects the LUT 5 corresponding to the template 3. Then, the address generator 6 generates an address.

When the address generator 6 generates an address, a pixel of the template 3 corresponding to the address is specified. On the other hand, the address generated by the address generator 6 is also sent to the selected LUT 5.
In the LUT 5, the sample image 2 corresponding to the address is
Pixel address. Then, the correlation operation unit 7 performs a correlation operation between the designated pixel of the template 3 and the pixel of the sample image 2. The calculation result for this pixel is recorded in the memory.

Further, the next address is generated by the address generator 6. Then, for all the pixels of the template 3, a correlation operation is performed with the corresponding pixels of the sample image 2, and the recorded result is output to the determination unit 8. The determination unit 8 determines the similarity of the sample image 2 to the template 3.

Here, when the address of the pixel of the sample image 2 is designated by one LUT 5, the pixel is designated by being shifted by one pixel within a range of several pixels in the vertical and horizontal directions with respect to the pixel at the position to be originally designated. To do. For example, in the case of specifying within a range that is shifted by 5 pixels in the vertical and horizontal directions with respect to the pixel at the position to be specified, if the coordinates of the original pixel are (0, 0), (−5, +5) to (+5) , -5)
The correlation calculation is performed 121 times in total for all the coordinates within the range. Then, a value having the highest similarity in the correlation calculation is adopted. As a result, it is possible to suppress a decrease in recognition accuracy due to a position shift of an image to be recognized. In particular, it is possible to improve recognition accuracy when the detection accuracy of the end portion of the securities is close.

Then, for the next template 3,
The similarity with the sample image 2 is determined using the LUT 5 in the same manner as the procedure described above. As described above, the similarity is calculated by comparing the sample image 2 with the plurality of templates 3, and as a result, the type of securities of the template 3 having the highest similarity to the sample image 2 is set as the type of the read paper piece. recognize.

The above embodiment is an example of a preferred embodiment of the present invention, but the present invention is not limited to this, and various modifications can be made without departing from the gist of the present invention. For example, in the above-described embodiment, the pixels of the sample image 2 and the pixels of the template 3 are directly compared.
The present invention is not limited to this, and a predetermined process may be performed on the pixels of the sample image 2 and then compared with the pixels of the template 3. In this case, as shown in FIG. 3, a neighboring address generating means 9 for generating an address of a neighboring pixel of a pixel designated by the LUT 5, and a predetermined process for data of a plurality of pixels designated by the neighboring address generating means 9. And an operation unit 10 for performing the following. Then, the output of the calculation unit 10 is transmitted to the correlation calculation unit 7 to perform a correlation calculation with the pixels of the template 3.

In this case, when creating the template 3 from the image 4 to be recognized, processing such as edge enhancement and averaging is performed. When the address generator 6 generates an address, a pixel of the template 3 corresponding to the address is designated, and the neighborhood address generator 9
The address of the pixel 11 in the vicinity of the pixel of the sample image 2 specified by the address generator 6 via the LUT 5 is generated.
Further, the processing unit 10 performs the same processing as the processing performed when the template 3 is created, for example, on the data of the plurality of neighboring pixels 11 of the sample image 2 designated by the neighboring address generation unit 9. Then, this operation unit 1
The output of 0 and the template 3 are correlated, and the similarity of the sample image 2 to the template 3 is determined. According to the image recognition device 1, the arithmetic unit 10 calculates the pixel data output from the template 3 and the sample image 2.
Since the same processing can be performed on any pixel data that is output through the above process, the accuracy of identification can be improved.

For example, when the image recognizing apparatus 1 shown in FIG. 3 performs an operation of reducing the resolution of the sample image 2 in the arithmetic unit 10 in order to perform a correlation operation with the template 3 created by reducing the resolution of the image 4 to be recognized. Prepares an LUT 5 and a neighboring address generating means 9 for the template 3, specifies an address of a center point of a portion to be averaged in the LUT 5, and the neighboring address generating means 9 averages around the center point. Pixels in a power range 11 are sequentially designated. Then, the arithmetic unit 10 averages the image data in the range 11. Also, the range 11 is set by designating the address of the center point of the next averaging portion.
Thereby, all the pixels of the template 3 and the sample image 2
And can be compared.

In the above-described embodiment, the recognition target of the image recognition apparatus 1 is a security such as a banknote. However, the present invention is not limited to this. For example, a coin may be used. in this case,
A template 3 is created in advance for various coins that are assumed to be read. Specifically, as shown in FIG. 5 (A), a ring-shaped cut-out range 14 is set on the read image 13 of the coin, the image is cut out, and processing such as edge enhancement is performed. Further, as shown in FIG. 5B, the cutout range 14 is coordinate-transformed into a rectangular shape of 128 × 10 pixels to create an image 4 to be recognized. Then, the characteristic part 4p of the image 4 to be recognized is selected and cut out, and the template 3 is created by combining with the other characteristic parts. Also, an LUT 5 corresponding to each template 3 is created.

When the coin is actually read and identified, an image of the coin is taken and the coin is read as shown in FIG.
As shown in (A), a ring-shaped cut-out range 14 is set for a read image 13 of a coin, the image is cut out, and processing such as edge enhancement is performed. Further, the cutout area 14 is shown in FIG.
As shown in (B), a sample image 2 is created by performing coordinate conversion into a rectangular shape of 128 pixels horizontally × 10 pixels vertically. Further, the comparison and determination between the sample image 2 and the template 3 are performed in the same procedure as in the above-described embodiment.

According to the image recognition device 1, the template 3 can be created by selecting pixels in a plurality of characteristic regions of the entire image 4 to be recognized. Compared to the case where a large template including a non-characteristic part is used by fixing the template position with respect to 4, the size of the template 3 can be reduced and the processing speed can be reduced without lowering the recognition accuracy. In particular, when discriminating euro coins having the same outer shape and material and a unique design on only one side, use a large template including a non-characteristic part by fixing the template position to improve recognition accuracy. Since the necessity is eliminated, the processing speed can be reduced without lowering the recognition accuracy, and identification determination can be performed with more coins.

Furthermore, in each of the above-described embodiments, the template creator selects a characteristic portion of the entire image 4 to be recognized based on experience or the like. However, the present invention is not limited to this. A feature value representing a change in pixel value in a predetermined extraction region of the entire image 4 to be calculated is calculated, the feature value is evaluated based on a predetermined criterion, and a region satisfying the criterion in the extraction region is defined as a characteristic portion. May be adopted.

For example, an image 4 to be recognized as shown in FIG. 7 is cut out from the read image 18 shown in FIG. 6 and a feature value representing a change in pixel value in the image 4 is calculated. Specifically, all pixels of the image 4 have n × m (for example, 3 × m).
3) A pixel filtering process is performed to calculate a feature amount representing a change in each pixel value of the entire image 4. Further, as the characteristic value, the difference between the maximum luminance and the minimum luminance in the pixel of interest and the neighboring n × m pixel area or the absolute value of the difference,
Values such as a luminance variance, a sum of absolute differences, a differential filter, and a Sobel filter for contour extraction can be used.

Here, with respect to the image 4, the sobel (so
bel) A process of extracting an outline of an image using a filter is performed. The Sobel filter is a filter for extracting the outline of an image from the difference in brightness of image data. When the output is large, it can be determined that the difference in brightness is large and the outline is clear. Is small and the change in the image is small. FIG. 8 shows the result 19 obtained thereby.

Further, a region having a large output by the Sobel filter process is selected, and a noise removing process and a region expanding process are performed. The region 20 thus obtained is shown in the white part of FIG. In other words, it can be said that in this region 20, the difference in brightness is large and the outline is clear. Therefore, the region 20 shown in FIG. 9 may be set as a characteristic portion.

On the other hand, a region having a small output by the Sobel filter process shown in FIG. 8 may be selected, and the noise removal process and the region expansion process may be performed. The region 21 thus obtained is shown in the white part of FIG. That is, this area 2
In the case of 1, it can be said that the difference in brightness is small and the change in the image is small.
Therefore, the region 21 shown in FIG. 10 may be set as a characteristic portion. This is an effective process when the feature is that there is no pattern in the portion.

Another example of a method for setting a characteristic portion of the entire image 4 to be recognized will be described with reference to a flowchart shown in FIG. Here, a case where coins having similar material and outer diameter, for example, a 500 yen coin and a 500 won coin are identified will be described as an example. It is assumed that a normalized correlation coefficient is used as the similarity.

First, in the image collection step of step S1,
Using an image input device such as a CCD camera, a large number of image data of a 500-yen coin to be identified is collected, and an image database is created in a memory of a computer. In this state, since 500-yen coins are taken in a random posture in each image data, conversion is performed in the image alignment step of step S2 so that the images taken in the random posture are aligned in a certain posture. Then, in the layering process based on dirt in step S3, image data is grouped according to the degree of dirt or abrasion. For example, a new currency group, a group of light currencies with light dirt, a group of liquid currencies with medium dirt, a group of liquid currencies with heavy dirt,
Into four groups.

Next, in the pre-processing step of step S4, a characteristic portion of the image 4 to be recognized consisting of each image data is extracted, in other words, an image cut out so as to extract the characteristic portion of the pattern of each image data. Determine the position and size of
Edge enhancement and data compression of the extracted characteristic portion are performed. This process is repeated for each characteristic part of one image 4 to be recognized, and these are combined to output elementary data (image for template creation) which is the basis of the template 3.

In the template output step of step S5, the template 3 is obtained by taking the arithmetic average of the elementary data obtained for each group, and the template 3 for each group is output. In this way, the template 3 for each group is obtained for the 500-yen coin, which is a coin to be identified, and the template 3 is similarly obtained for each 500-yen coin, which is a coin to be excluded.

Thereafter, in the inter-template correlation value calculation step of step S6, the matching operation with each template 3 for the 500 yen coin and the 500 won coin is performed a number of times by changing the angle according to the above-described procedure. The maximum value is stored as a correlation value (similarity). Here, as the correlation value, in the normalized correlation coefficient, the autocorrelation coefficient takes a value of 0 to 1 (−1 to 1 depending on the calculation method), and the cross-correlation coefficient takes a value of −1 to 1. In this embodiment, 500 yen coin and 500 yen
Since the won coin templates 3 are grouped according to the degree of dirt, the same number as the number of groups is created.
Therefore, the distance between templates (data distribution) is determined for each coin on a round robin basis, and the closest one is stored as a correlation value. When the sum of absolute differences is used as the degree of dissimilarity, the sum of absolute differences is used instead of the correlation value between templates, and the sum of absolute differences becomes smaller as the similarity becomes higher. Changes “small” <> “large” to “pull” <> “add”.

Next, in the inter-template correlation value evaluation step of step S7, the correlation value obtained in step S6 is compared with a predetermined threshold value stored. here,
If the correlation value is larger than the threshold value, the distribution of each elementary data for the 500-yen coin is relatively close to the distribution of each elementary data for the 500-won coin, and the template created even when the subsequent processing is continued. It is unlikely that the discrimination performance of No. 3 will reach the level required for coin discrimination. Therefore, in this case, the process returns to step S4 to cut out another part of each image data and recreate each elementary data.

On the other hand, if the correlation value is less than the threshold, 50
The distribution of each elementary data for the 0 yen coin and the distribution of each elementary data for the 500 won coin are relatively far apart,
It is considered that the discrimination performance of the template 3 created by continuing the subsequent processing can reach a level required for coin discrimination. Therefore, when the correlation value is equal to or smaller than the threshold, the process proceeds to the self-recognition step of step S8, and a self-recognition experiment is performed to determine an exclusive threshold. In other words, only the self-recognition experiment is performed, and the exclusion threshold value that is the accuracy required for coin discrimination is determined.

Next, in the exclusive threshold comparing step of step S9, the correlation value obtained in step S6 is compared with the exclusive threshold obtained in step S8. Here, when the correlation value is greater than the exclusion threshold (or a value obtained by subtracting a predetermined value from the exclusion threshold), the discrimination performance of the template 3 created even when the subsequent processing is continued is required for coin discrimination. It is unlikely to reach the level. Therefore, in this case, the process returns to step S4 to cut out another part of each image data and recreate each elementary data.

On the other hand, when the correlation value is equal to or smaller than the exclusion threshold (or a value obtained by subtracting a predetermined value from the exclusion threshold), the discrimination performance of the template 3 created by continuing the subsequent processing is required for coin discrimination. It is believed that the level can be reached. Therefore, in this case, the process proceeds to the statistical calculation step of step S10, and an exclusion experiment using the sample image is performed to estimate whether the coin to be excluded (500 won coin) can be actually identified from the 500 yen coin. Then, the process proceeds to step S11, and if it is estimated that identification is possible, the template creation ends. On the other hand, if it is estimated that the image data cannot be identified, the process returns from step S11 to step S4 to cut out another part of each image data and recreate each elementary data.

As described above, in the present invention, a small feedback loop (step S7 → step S4) is included in a large feedback loop (step S11 → step S4).
4, step S9 → step S4), and the validity of the template 3 is reviewed during the creation of the template 3. Therefore, if the template creation cannot be adopted even if the template creation is continued, wasted labor and time are wasted. Can be saved. That is, the template 3 having low identification performance is quickly determined, the creation of the template 3 is interrupted halfway, and the process returns to the pre-processing step, so that extra labor and wasted time required for template creation can be prevented.

[0054]

As is clear from the above description, according to the image recognition apparatus of the first aspect, a template can be created by selecting only characteristic portions of the entire image to be recognized. Therefore, the processing speed can be reduced by reducing the size of the template without lowering the recognition accuracy. In addition, if the processing time is the same as in the related art, it is possible to perform identification determination with more images to be recognized while maintaining recognition accuracy.

According to the image recognition apparatus of the second aspect, the type of the sample image can be specified with high accuracy by comparing the sample image with a plurality of templates.

Further, according to the image recognition apparatus of the third aspect, the correlation operation is performed by the correlation operation section between the pixels of the template and the pixels of the sample image specified via the reference table, and the similarity of the sample image to the template is obtained. Can be determined.

According to the image recognition apparatus of the present invention, the same processing can be applied to both the pixel data output from the template and the pixel data output from the sample image via the calculation unit. Therefore, the accuracy of identification can be improved.

Further, according to the image recognition apparatus of the fifth aspect, the reference table shifts the designated address by the designated shift amount, so that the correlation operation can be performed while shifting the template and the sample image within the range of the designated pixel. It is possible to suppress a decrease in recognition accuracy due to a displacement of an image to be recognized or the like.

[Brief description of the drawings]

FIG. 1 shows an example of extracting a characteristic portion from an image to be recognized by an image recognition device of the present invention to create a template, and changing the extraction position and size of the characteristic portion in (A) to (C). ing.

FIG. 2 is a block diagram illustrating an image recognition device.

FIG. 3 is a block diagram showing another example of the image recognition device.

4A and 4B are diagrams illustrating a process of creating a sample image from a coin read image, wherein FIG. 4A is a coin read image and FIG. 4B is a sample image obtained by coordinate transformation.

5A and 5B are diagrams showing a process of creating an image to be recognized from a read image of a coin, wherein FIG. 5A is a read image of a coin, and FIG. 5B is an image to be recognized obtained by coordinate transformation. .

FIG. 6 is a diagram illustrating an example of an image to be recognized;

FIG. 7 is a diagram showing an extraction region extracted from an image to be recognized.

FIG. 8 is a diagram illustrating a state in which a Sobel filter process has been performed on an extraction area.

FIG. 9 is a diagram illustrating a result of selecting a region having a large output by the Sobel filter process and performing a noise removal process and a region expansion process.

FIG. 10 is a diagram illustrating a result obtained by selecting a region having a small output by the Sobel filter process and performing a noise removal process and a region expansion process.

FIG. 11 is a flowchart illustrating an example of another embodiment of a template creation method.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 Image recognition apparatus 2 Sample image 3 Template 4 Image to be recognized 4a-4p Multiple areas 5 Look-up table (reference table) 6 Address generator 7 Correlation calculator 9 Neighbor address generator 10 Calculation unit

Claims (5)

[Claims] 1. An image recognition apparatus for recognizing a sample image by comparing a predetermined area of a sample image to be compared with a reference template, wherein the template is selected from the entire image to be recognized. Each of the positions includes a plurality of pixels in different regions, and a reference table indicating the positions of the pixels constituting the template is provided, and the pixels of the sample image are extracted by designating the position of the reference table and compared with the template. An image recognition device, characterized in that: 2. A method according to claim 1, wherein the plurality of templates have a plurality of reference tables, and reference tables are formed corresponding to the plurality of templates, and pixels of the sample image correspond to the templates for each template among the plurality of templates. 2. The image recognition device according to claim 1, wherein the image is retrieved using a reference table. 3. The reference table has, as its elements, an address of a pixel included in the template, and further includes an address generator for sequentially generating an address of a pixel for performing a correlation operation, and an address generator. The image recognition apparatus according to claim 1, further comprising a correlation operation unit configured to perform a correlation operation between the designated pixel of the template and the pixel of the sample image designated by the reference table. 4. The template is composed of pixels in a plurality of regions each having a different position selected from the entire image to be recognized, and performing a predetermined process for each of the plurality of pixels to convert the template. Configure it,
Providing a neighboring address generating means for generating an address of a neighboring pixel of the pixel specified by the reference table, and an arithmetic unit for performing a predetermined process on data of a plurality of pixels specified by the neighboring address generating means, 4. The image recognition device according to claim 1, wherein an output of said operation unit is correlated with said template. 5. The image recognition apparatus according to claim 1, wherein the reference table shifts an address designated for extracting a pixel of the sample image by a designated shift amount.