JP4580324B2 - Image classification apparatus and image classification method - Google Patents

Description

  The present invention relates to an image classification apparatus and an image classification method for identifying a category to which an object belongs from a plurality of categories by comparing an input image including an object pattern with a template, and in particular, performs discrimination using an image of an object pattern. In particular, the present invention relates to an image classification apparatus and an image classification method that can reduce the time required for processing.

  2. Description of the Related Art Conventionally, an apparatus has been devised that scans an image of an object having a pattern formed on the surface, such as a coin, and determines the type and authenticity of the object by pattern matching between the pattern scan result and a template. For example, in Patent Document 1, a straight line passing through the center of a coin is used as a scan line, and each template prepared with respect to a 360-degree omnidirectional straight line passing through the center of the coin is compared with a scan value on the scan line. It is described that the type of coin is determined by doing so.

  Further, in Patent Document 1, considering that there is a portion having a different pattern even in the same type of coin, such as the year of manufacture of the coin, when the correlation value regarding the scan result of the entire scan line is low, a plurality of scan results are displayed. It is disclosed to divide into parts and to determine that the scan result matches the template if the correlation value is high in any part.

JP 2003-187289 A

  However, when the above-described determination is performed to identify the type or authenticity of an object, there is a problem that the processing time increases as the types of object patterns increase. That is, for example, when there are a large number of coins having different patterns for each denomination and for each country, such as euro coins, in order to distinguish the type or authenticity of one coin, the scan result is a template for all types of patterns. And requires a very long processing time.

  Also, in the comparison of templates related to one type of pattern, as described above, the scan results are compared with all templates prepared for straight lines in all directions of 360 degrees. For this reason, until one coin is finally identified, it is necessary to compare the number of times of the number of templates related to one pattern multiplied by the number of types of object patterns, which requires enormous processing time. Become.

  The present invention has been made in view of the above points, and provides an image classification apparatus and an image classification method capable of reducing the time required for processing when performing discrimination using an image of an object pattern. Objective.

  In order to solve the above-described problem, the present invention is an image classification apparatus for identifying a category to which the object belongs from a plurality of categories by comparing an input image including a pattern of the object with a template, and a plurality of objects belonging to the same category. Storage means for preliminarily storing, as a template for each category, image total data of feature areas common to the patterns, first calculation means for calculating image total data for each feature area of the input image, and the first calculation Second calculation means for calculating a collation value for each template by performing a correlation operation on the image total data calculated by the means and the template stored by the storage means; and the collation value calculated by the second calculation means Output means for outputting the category corresponding to the largest template as the classification result of the object I am characterized in.

  Also, in the present invention according to the above invention, the first calculation means creates a polar coordinate conversion unit that performs polar coordinate conversion on the input image, and a histogram in a feature region in the input image that has been subjected to polar coordinate conversion by the polar coordinate conversion unit. A histogram generation unit; and an autocorrelation calculation unit that calculates an autocorrelation curve of the histogram generated by the histogram generation unit, wherein the second calculation unit includes the autocorrelation curve calculated by the autocorrelation calculation unit. And a collation value for each template is calculated from a normalized correlation between the template and the autocorrelation curve of the template stored by the storage means.

  Further, in the present invention according to the above invention, the first calculation unit includes a histogram generation unit that generates a histogram in a feature region in the input image, and the second calculation unit includes the histogram generation unit. The collation value for each template is calculated from the normalized correlation between the created histogram and the histogram of the template stored by the storage means.

  Further, the present invention is characterized in that, in the above invention, the output means outputs a category corresponding to a template whose collation value calculated by the second calculation means is not less than a predetermined threshold as a classification result of the object. And

  Also, in the present invention according to the above-described invention, the storage unit stores image aggregate data of a plurality of feature regions for one category as a template, and the second calculation unit stores correlations in each of the plurality of feature regions. The comparison result for each template is calculated by weighted addition of the calculation results.

  Further, the present invention according to the above invention further comprises a detecting means for detecting the size of the object from the input image, and the second calculating means is for only the category to which the object of the size detected by the detecting means belongs. A template is used for calculation of a collation value.

  In addition, the present invention is an image classification device that identifies which country the coin is from by comparing an input image including a coin pattern with a template, and a characteristic region characteristic of the coin pattern of each country Storage means for preliminarily storing the image total data as a template for each country, first calculation means for calculating image total data for each feature region of the input image, and image total calculated by the first calculation means A second calculation means for calculating a collation value for each template by correlating the data and the template stored in the storage means, and the collation value calculated by the second calculation means corresponds to the maximum template. Output means for outputting the country as a classification result.

  The present invention is also an image classification method for identifying a category to which the object belongs from a plurality of categories by comparing an input image including a pattern of the object with a template, and is common to a plurality of object patterns belonging to the same category. Calculated in the reading step of reading out the template for each category indicating the image total data of the feature region, the first calculation step of calculating the image total data for each feature region of the input image, and the first calculation step A second calculation step of calculating a collation value for each template by performing a correlation operation between the image total data and the template read in the reading step; and the collation value calculated in the second calculation step is a maximum. And an output step of outputting a category corresponding to the template as a classification result of the object.

  According to the present invention, image aggregate data of feature areas common to a plurality of object patterns belonging to the same category is stored in advance as a template for each category, and image aggregate data for each feature area of the input image is calculated, Since the calculated image aggregation data and the template are subjected to correlation operation to calculate a matching value for each template, and the category corresponding to the template with the largest calculated matching value is output as the classification result of the object, Candidates can be narrowed down by specifying the category to which the object belongs in advance, and the time required for processing can be shortened when performing discrimination using the pattern image of the object.

  Further, according to the present invention, the input image is subjected to polar coordinate conversion, a histogram in the feature region in the input image subjected to polar coordinate conversion is created, an autocorrelation curve of the created histogram is calculated, and the calculated autocorrelation curve and Since the collation value for each template is calculated from the normalized correlation with the autocorrelation curve of the template, it is not necessary to match the rotational direction deviation between the object and the template, and the processing time can be further reduced.

  In addition, according to the present invention, a histogram in the feature region in the input image is created, and the collation value for each template is calculated from the normalized correlation between the created histogram and the histogram of the template. If the shift in the rotation direction is matched, processing in an orthogonal coordinate system can be performed without converting the input image into polar coordinates.

  In addition, according to the present invention, a category corresponding to a template whose collation value is equal to or greater than a predetermined threshold is output as an object classification result. Therefore, flexible classification can be performed by appropriately setting the threshold. .

  Further, according to the present invention, image aggregate data of a plurality of feature areas for one category is stored as a template, and correlation values for each of the plurality of feature areas are weighted and added to calculate a matching value for each template. With this configuration, even if one feature region cannot be used for complete classification, the number of feature regions can be increased to easily obtain a matching value, and an object can be reliably classified.

  In addition, according to the present invention, the size of an object is detected from the input image, and the template of only the category to which the object of the detected size belongs is used for calculation of the collation value. Therefore, a category that cannot be a classification result. The matching value can be calculated by excluding the template, and the processing time can be further reduced.

  In addition, according to the present invention, image aggregate data of feature regions characteristic of the coin pattern of each country is stored in advance as a template for each country, and image aggregate data for each feature region of the input image is calculated and calculated. By calculating the correlation value for each template and calculating the matching value for each template, and outputting the country corresponding to the template with the largest matching value as the classification result, It can be specified whether it is a country, and the time required for processing can be shortened when performing discrimination using an image of a coin pattern.

  Further, according to the present invention, a template for each category indicating image aggregate data of feature areas common to a plurality of object patterns belonging to the same category is read, and image aggregate data for each feature area of the input image is calculated and calculated. Since the calculated image data and the template are correlated to calculate the matching value for each template, the category corresponding to the template with the largest matching value is output as the object classification result. The category to which the user belongs can be specified to narrow down candidates, and the time required for processing can be shortened when performing discrimination using an image of an object pattern.

  Hereinafter, an embodiment of the present invention will be described in detail with reference to the drawings. In the following, a case where euro coins are identified will be described as an example, and a case where the patterns of coins are roughly classified to limit categories such as countries including coin type candidates will be described. Euro coins have different patterns from country to country even with the same denomination, and there are more than 20 patterns. Therefore, in order to discriminate euro coins as usual, it is necessary to collate one coin to be discriminated with each of twenty or more types of templates, which is not efficient. Therefore, in the following embodiment, a case will be described in which the category (ie, country) to which the coin belongs is narrowed down in advance.

  FIG. 1 is a block diagram showing a main configuration of an image classification apparatus according to an embodiment of the present invention. 1 includes an edge detection unit 110, a size detection unit 120, an image cutout unit 130, a coordinate conversion unit 140, a template storage unit 150, a template setting unit 160, a template comparison unit 170, and a maximum value selection unit. 180.

  The edge detection unit 110 performs edge detection on an input image including a coin pattern to be identified, generates an edge image in which the presence / absence of an edge is expressed by shading of the edge strength, and outputs the edge image to the size detection unit 120.

  The size detection unit 120 detects the size of the coin to be identified from the edge image, and notifies the image cutting unit 130 and the template setting unit 160 of the detected size. Specifically, the size detection unit 120 determines that the region where the edge is generated in the edge image corresponds to the size of the coin. For example, the image cutting unit 130 and the template setting unit 160 is notified.

  The image cutout unit 130 cuts out an image of a coin part from the edge image according to the coin size notified from the size detection unit 120, and outputs the cut image to the coordinate conversion unit 140. The coin is assumed to be circular, and the image of the coin portion may be the same shape as the coin circle, or may be a square shape circumscribing the coin circle.

  The coordinate conversion unit 140 converts the image of the coin portion into polar coordinates, and converts the displacement in the rotation direction of the coin into the displacement in the linear direction. That is, the coordinate conversion unit 140 maps the image on the xy coordinate that is an orthogonal coordinate system onto the ρθ coordinate that is the polar coordinate system, thereby converting the displacement in the rotational direction on the xy coordinate to the θ-axis direction on the ρθ coordinate. Convert to linear displacement. Thereby, even if the coin rotates at any angle in the input image, this rotation is expressed as a linear shift.

  The template storage unit 150 stores image aggregate data regarding feature regions common to all denominations in each country as a template for each country. Specifically, the template storage unit 150 adds a band-like region having a width in the ρ-axis direction that is characteristic for each country when the edge image of the coin is arranged in the polar coordinate system (ρθ coordinate system). The autocorrelation curve of the histogram is stored. That is, the template storage unit 150 stores, for example, the autocorrelation curve of the histogram in the addition section in the ρ-axis direction that characterizes the coin pattern of country A as a template of country A, and characterizes the pattern of each country with respect to other countries. The autocorrelation curve of the histogram in the addition section in the ρ axis direction is stored as a template for each country.

  One country template may include auto-correlation curves of a plurality of addition sections. However, if the number of added sections included in the template for one country increases, the number of comparison processes with the template increases for this country, and the overall processing time increases. In addition, since each addition section is a band-shaped region in the ρ axis direction, in a rectangular coordinate system (xy coordinate system) in which the center of the coin is the origin, a donut-shaped region formed from a concentric circle centered on the origin is used for the template. It is none other than being used as an area of

  Here, by using the autocorrelation curve of the histogram in the addition section in the ρ-axis direction as a template, as will be described in detail later, in the comparison between the coin to be identified and the template, the rotational direction shift in the orthogonal coordinate system is ignored. Will be able to. In addition, since a band-shaped addition section in the ρ axis direction is used as a template, it is possible to perform a determination that has an excessively narrow width unlike a case where a coin to be identified is compared with, for example, a linear template. it can. For this reason, it is possible to determine that a plurality of types of coins match one country template like the above-described country-specific templates.

  The template setting unit 160 selects a template to be used for classifying coins according to the size of the coin notified from the size detection unit 120, reads the selected template from the template storage unit 150, and adds the interval in the template comparison unit 170. Set in the instruction unit 171. Euro coins have different coin sizes depending on the denomination, but there are denominations that do not exist depending on the country. Therefore, a template for a country where there is no coin of the size detected by the size detection unit 120 is the coin to be identified. There is no need to use it for classification. That is, the template setting unit 160 sets only the template of the country where the denomination of the size detected by the size detection unit 120 exists in the addition section instruction unit 171. Thereby, the template to be used is limited based on the size of the coin, and as a result, the type of the coin may be narrowed down.

  The template comparison unit 170 includes an addition interval instruction unit 171, a histogram creation unit 172, an autocorrelation calculation unit 173, a normalized correlation calculation unit 174, and a matching value calculation unit 175, and compares an input image with a template. And output a collation value between the input image and the template of each country.

  Specifically, the addition interval instruction unit 171 instructs the histogram creation unit 172 of an addition interval in which the edge intensity for each θ is to be calculated in the image of the coin part that has been subjected to polar coordinate conversion. That is, the addition section instruction unit 171 instructs the histogram creation unit 172 to create a histogram corresponding to each addition section of the template for each country set by the template setting unit 160. At the same time, the addition interval instruction unit 171 outputs the autocorrelation curve (template) of the specified addition interval to the normalized correlation calculation unit 174. In addition, when the addition section instruction unit 171 completes the instruction for all the addition sections of the template for one country, the addition section instruction section 171 instructs the collation value calculation section 175 to calculate the collation value for this country, and completes the instruction for all the templates. In this case, the collation value calculation unit 175 is instructed to output all collation values.

  The histogram creation unit 172 adds the pixel values of the addition section designated by the addition section instruction unit 171 for each θ, and creates a histogram indicating the edge strength of the addition section at each θ. Since the created histogram is a histogram corresponding to the addition section included in the template of any country, if the coins to be identified are from this template country, the histogram is characteristic and the same addition There is a clear difference between the histograms of other countries' coins prepared for the interval.

  The autocorrelation calculation unit 173 calculates the autocorrelation of the histogram created by the histogram creation unit 172, and obtains an autocorrelation curve as the image total data. As described above, since the image of the coin part has been subjected to polar coordinate conversion by the coordinate conversion unit 140, the displacement in the rotation direction of the orthogonal coordinate system (xy coordinate system) is converted into a linear displacement in the θ-axis direction. . For this reason, also in the histogram created by the histogram creation unit 172, the peak position translates in the θ-axis direction in accordance with the rotation of the coin in the orthogonal coordinate system (xy coordinate system). Therefore, the autocorrelation calculation unit 173 calculates the autocorrelation when the histogram is shifted in the θ-axis direction, so that the same coin can be used regardless of the rotation angle of the coin in the orthogonal coordinate system (xy coordinate system). The same autocorrelation curve can be obtained with respect to the histogram.

  As a result, it is possible to remove information on the rotation of the coin in the input image, and an autocorrelation curve that depends only on the pattern of the coin is obtained. When comparing with the template, the rotation of the coin in the orthogonal coordinate system (xy coordinate system) Can be ignored.

  The normalized correlation calculation unit 174 calculates a normalized correlation between the autocorrelation curve obtained by the autocorrelation calculation unit 173 and the template autocorrelation curve output from the addition section instruction unit 171, and the two autocorrelation curves are obtained. Check how similar they are. If the coin to be identified is a coin of a country that is a template, since the two autocorrelation curves are similar, the normalized correlation calculated by the normalized correlation calculation unit 174 increases. On the other hand, if the coins to be identified are not the coins of the country that is the template, the two autocorrelation curves are not similar, and the normalized correlation calculated by the normalized correlation calculation unit 174 becomes small.

The matching value calculation unit 175 calculates a matching value for each country from the normalized correlation for each addition section of the template for each country. Specifically, when the addition value indicating unit 171 is instructed to calculate the reference value, the reference value calculating unit 175 calculates the reference value for each country by weighted addition of the normalized correlations obtained for all the addition regions. To do. In other words, the normalized correlation with respect to the addition interval j in the template for country i is set as s _ij , and the matching value Z _i for country _i is calculated by the following equation (1).

In equation (1), α _ij is a weighting factor for the addition interval j in the template for country i, and is a value determined by a learning algorithm such as boosting. The matching value Z _i for i country obtained by this equation (1) increases as the normalized correlation s _ij for each template in i country increases, and the matching value Z when the coin in the input image is a coin of i country. _i is the maximum.

  The maximum value selection unit 180 selects the maximum value of the collation values for each country, and outputs a classification result that the coins in the input image are coins of the country corresponding to the selected collation values. Since this classification result indicates which country the coin to be identified belongs to, the denomination is not specified. However, the process is completed in a shorter time than by performing detailed discrimination using coins of all denominations of all countries as candidates by performing further detailed discrimination using the coins of countries of the classification results as candidates.

  Next, an image classification method by the image classification apparatus configured as described above will be described with reference to the flowchart shown in FIG.

  First, an edge detection unit 110 performs edge detection on an input image including a coin pattern to be identified, and an edge image is generated. Then, the size detection unit 120 detects the size of the coin from the edge image (step S1000), and notifies the image cutout unit 130 and the template setting unit 160 of this size.

  When the image cutout unit 130 is notified of the size of the coin, the image cutout unit 130 cuts out the image of the coin part from the edge image, and the coordinate conversion unit 140 converts the image of the coin part into polar coordinates (step). S1010). As a result, the pattern of the coins to be identified is arranged on the ρθ coordinate axis, which is a polar coordinate system, and the coin region has a length in the ρ axis direction corresponding to the radius of the coin, and a length in the θ axis direction. Becomes a rectangular region corresponding to 360 degrees. Therefore, the displacement in the rotational direction of the coin on the xy coordinate axis, which is an orthogonal coordinate system, is converted into a linear displacement in the θ-axis direction.

  On the other hand, when the size of the coin is notified to the template setting unit 160, the template setting unit 160 reads the template of the country where the coin of the notified size exists from the template storage unit 150, and sends it to the addition section instruction unit 171. Is set. That is, since the coins to be identified are from any country in which a denomination of the same size as this coin exists, only I countries in which the denomination of the size detected by the size detection unit 120 exists. A template is used (step S1020). These I templates are image total data of coins of respective countries, and are specifically configured from histogram autocorrelation curves in one addition section or multiple addition sections in the ρ-axis direction.

  When I templates are set in the addition section instruction section 171, the addition section instruction section 171 initializes a variable i for designating a template to 1 (step S1030), and the template 1 that is first compared with the input image. Is set (step S1040). As the template 1 is set, the number of addition sections J used for the time being (that is, the number of addition sections of the template 1) is determined.

  Then, the addition section instruction unit 171 initializes a variable j for designating each addition section constituting the template 1 to 1 (step S1050), and corresponds to the addition section 1 of the template 1 for the histogram creation unit 172. An instruction to create a histogram is output (step S1060). At the same time, the autocorrelation curve of addition period 1 of template 1 is output to normalized correlation calculation section 174. When an instruction to create a histogram is output from the addition section instruction unit 171, the histogram creation unit 172 creates a histogram of the coin portion image that has been subjected to polar coordinate conversion by the coordinate conversion unit 140. At this time, the histogram creation unit 172 creates a histogram corresponding to the addition section 1 of the template 1.

  The histogram corresponding to addition section 1 of template 1 is output to autocorrelation calculation section 173, and an autocorrelation curve is obtained by calculating autocorrelation. This autocorrelation curve is an autocorrelation curve of a coin to be identified corresponding to the addition section 1 of the template 1, and information on the rotation direction of the coin in the input image is removed. If the coin to be identified is a coin of the country of template 1, the autocorrelation curve obtained in autocorrelation calculation unit 173 is the template 1 output from addition interval instruction unit 171 to normalized correlation calculation unit 174. It is similar to the autocorrelation curve of the addition interval 1 of

  The autocorrelation curve obtained from the input image is output to the normalized correlation calculation unit 174, and the normalized correlation with the autocorrelation curve of the addition interval 1 of the template 1 already output from the addition interval instruction unit 171 is calculated. (Step S1070). The calculated normalized correlation is an index indicating how similar the two autocorrelation curves are. If the two autocorrelation curves are similar, it will increase, and the two autocorrelation curves must be similar. It will become smaller. This normalized correlation is output to the collation value calculation unit 175 and temporarily held.

  On the other hand, after outputting the histogram creation instruction corresponding to the addition section 1 of the template 1, the addition section instruction section 171 determines whether or not the number of addition sections J of the template 1 is equal to the current variable j specifying the addition section. Determination is made (step S1080). Here, since the instruction for creating the histogram corresponding to the addition section 1 of the template 1 has been output, the value of the variable j is 1. Therefore, if the addition section number J of the template 1 is 2 or more, the addition section number J and the variable j are not equal (No in step S1080), and the variable j is incremented by 1 by the addition section instruction unit 171 (step S1090). An instruction to create a histogram corresponding to 1 addition interval 2 is output (step S1060). At the same time, the autocorrelation curve of the addition section 2 of the template 1 is output to the normalized correlation calculation unit 174.

  Similarly to the case where the variable j is 1, a histogram corresponding to the addition interval 2 of the template 1 is created by the histogram creation unit 172, and the normal correlation between the autocorrelation curve of the histogram and the autocorrelation curve of the addition interval 2 of the template 1 is performed. The normalized correlation is calculated by the normalized correlation calculation unit 174 (step S1070), and is output to and stored in the matching value calculation unit 175.

  Then, the addition section instructing unit 171 again determines whether or not the number of addition sections J of the template 1 is equal to the current variable j (step S1080). Here, since the instruction for creating the histogram corresponding to the addition interval 2 of the template 1 has been output, the value of the variable j is 2. Therefore, if the number of addition sections J of the template 1 is 2, the number of addition sections J and the variable j are equal (Yes in step S1080), and the addition section instruction unit 171 calculates the normalized correlation for all the addition sections of the template 1. The verification value calculation unit 175 is notified of the fact.

Upon receiving this notification, the collation value calculation unit 175 uses the above equation (1) to calculate the collation value for the country of the template 1 (step S1100). That is, here, the collation value Z ₁ is obtained by weighted addition of the normalized correlation s ₁₁ and the normalized correlation s ₁₂ . If this collation value Z ₁ is large, there is a high possibility that the coin to be identified is a coin in the country of template 1, and conversely if the collation value Z ₁ is small, the coin to be identified is a coin in the country of template 1. The possibility is small. The obtained collation value Z ₁ is continuously held by the collation value calculation unit 175.

  On the other hand, after notifying that the normalized correlation for all the addition sections of the template 1 has been calculated, the addition section instructing unit 171 determines whether or not the number of templates I to be used and the current variable i designating the template are equal. Determination is made (step S1110). Here, since it has been notified that the normalized correlation for all the addition sections of template 1 has been calculated, the value of variable i is 1. Therefore, if the number of templates I to be used is 2 or more, the number of templates I and the variable i are not equal (No in step S1110), and the variable i is incremented by 1 by the addition interval instruction unit 171 (step S1120) and compared with the input image. The template 2 to be set is set (step S1040). At this time, as the template 2 is set, the number of addition sections J (that is, the number of addition sections of the template 2) is newly determined.

Thereafter, the normalized correlation is calculated for each added section of template 2 and the template 2 matching is performed in the same manner as the normalized correlation is calculated for each added section of template 1 and the matching value Z _{1 of} template ₁ is calculated. A value Z ₂ is calculated (steps S1050 to S1100). The obtained collation value Z ₂ is continuously held by the collation value calculation unit 175.

Then, the addition section instruction unit 171 again determines whether or not the number of templates I to be used is equal to the current variable i (step S1110). Here, since the collation value Z ₂ of the template 2 has been calculated, the value of the variable i is 2. Therefore, if the number of templates I to be used is 2, the number of templates I and the variable i are equal (Yes at step S1110), and the collation value Z ₁ and collation value Z calculated so far are compared with the collation value calculation unit 175. ₂ is output to maximum value selection section 180.

In response to this instruction, when collation value Z ₁ and collation value Z ₂ are output from collation value calculation unit 175 to maximum value selection unit 180, maximum value selection unit 180 selects the maximum collation value (step). S1130). As described above, the collation value calculated by the collation value calculation unit 175 increases when the autocorrelation curve related to the coin to be identified and the autocorrelation curve of the template are similar. The selected collation value is the collation value corresponding to the template most similar to the coins to be identified. In other words, the coin to be identified is most likely a coin of the template country corresponding to the collation value selected by the maximum value selection unit 180. Therefore, the maximum value selection unit 180 outputs a classification result that the coins to be identified are coins of this country.

  In this way, since the coins to be identified are classified as to which country's coins, it is possible to limit the coins to be identified as being coins of any one country, and only the coins of this country are identified in detail. By performing the above, it is possible to finally identify which denomination coin of which country. At this time, detailed discrimination is performed only for a few types of denominations, so the time is greatly reduced compared to detailed discrimination for all 20 types of coins without country classification. can do.

  Specific methods for detailed discrimination include, for example, the following. That is, a correlation value image is generated from an input image and a template image, and the correlation value image is separated into a positive correlation value image and a negative correlation value image depending on whether the pixel value is equal to or greater than a threshold value, Is separated into a positive template image and a negative template image depending on whether or not the threshold value is equal to or greater than a threshold, and a plurality of positive and negative separated correlation images are generated by combining the positive correlation value image and the negative correlation value image with the positive template image and the negative template image. Then, the collation determination is performed using the positive / negative separated correlation image.

  Next, collation between the input image and the template according to the present embodiment will be described with specific examples.

  FIG. 3A shows a state in which the image of the coin portion cut out by the image cutout unit 130 is arranged on the xy coordinate axis that is an orthogonal coordinate system. On the xy coordinate axis, the coin is represented by a circle having a radius r. This circular radius r is based on the size detected by the size detection unit 120, and by determining the radius r, a country in which no coin with the radius r exists cannot be a classification result of the coins to be identified. I understand. Therefore, the template setting unit 160 reads only the template of the country where the coin with the radius r exists, and sets it in the addition section instruction unit 171. Here, for example, it is assumed that the templates of the two countries of country A and country B are set in the addition section instruction unit 171.

  Further, in FIG. 3A, the black portion in the circle of the coin illustrates a part of the region surrounded by the edge. On the xy coordinate axis, the black portion is displaced in the rotation direction as the coin rotates. Therefore, the coordinate conversion unit 140 converts the image of the coin part into polar coordinates, whereby the circle of the coin is expressed as a rectangle on the ρθ coordinate axis as shown in FIG. On the ρθ coordinate axis, as the coin rotates, the black portion is linearly displaced in the θ-axis direction (left-right direction with respect to the paper surface).

  By the way, the template storage unit 150 stores, as templates for each country, an autocorrelation curve of a histogram in a band-like addition section characterizing the coins of each country. That is, for example, as a template for country A, histogram autocorrelation curves in each of the three addition sections 210 in FIG. 3B are stored, and as a template for country B, 2 in FIG. The autocorrelation curve of the histogram in each of the two addition sections 220 is stored.

  Since each of these addition sections is a band-shaped section having a width in the ρ-axis direction, even if linear displacement in the θ-axis direction occurs as the coin rotates, the autocorrelation curve of the histogram of each addition section Is immutable. Therefore, if the autocorrelation curve of the histogram of the coins to be identified is calculated for these addition sections, the comparison with the template can be easily performed while ignoring the rotation of the coins.

  Therefore, the addition section instructing unit 171 instructs the histogram creating section 172 to create a histogram of the coin portion image in order for each added section of the set template. The histogram creation unit 172 adds the edge strength for each θ in the designated addition section, and creates a histogram as shown in FIG. 3C for each addition section, for example. Further, the autocorrelation calculation unit 173 calculates the autocorrelation when the histogram is shifted in the θ-axis direction, and obtains an autocorrelation curve as shown in FIG. 3D, for example. Note that the scales in the θ-axis direction in FIGS. 3B, 3C, and 3D are not necessarily the same.

The autocorrelation curve corresponding to each addition interval obtained by the autocorrelation calculation unit 173 is compared with the autocorrelation curve stored as the template of the same addition interval, and the normalized correlation calculation unit 174 calculates the normalized correlation. Calculated. That is, as shown in FIG. 4, the autocorrelation curves corresponding to the addition sections 1 to 3 of the country A template are X _A -1 to 3 respectively, and the templates of the country A addition sections 1 to 3 are A-1 respectively. -3, the normalized correlation between X _A -1 and A-1, the normalized correlation between X _A -2 and A-2, and the normalized correlation between X _A -3 and A-3, respectively Calculated. These normalized correlations are output to the collation value calculation unit 175, and the collation value calculation unit 175 uses the above equation (1) to calculate the collation value Z _{A for} country _A. Similarly, for country B, the normalized correlation between X _B -1 and B-1 and the normalized correlation between X _B -2 and B-2 are calculated by the normalized correlation calculation unit 174, respectively. Based on the correlation, the matching value calculation unit 175 calculates the matching value Z _{B for} country _B.

In the example of FIG. 4, the autocorrelation curves X _A -1 and A-1 are greatly different from each other, and thus the normalized correlation is small. On the other hand, the autocorrelation curves X _B -1 and B-1 are similar to each other and thus normal. The correlation is large. The fact that the normalized correlation is large means that the histogram of the input image in this addition section shows a feature similar to the histogram of the coins of the template country. Furthermore, here, since the collation value for each country is calculated by weighted addition of the normalized correlations calculated for a plurality of addition sections (three for Country A and two for Country B), the collation value is large. For example, the possibility that the coin to be identified is a coin of the country of the template increases.

  Therefore, the maximum value selection unit 180 selects the maximum value of the matching value and outputs the country of the template corresponding to the maximum matching value as the classification result. As a result, for example, if a classification result indicating that the coins to be identified are the coins of country B is output, the coins to be identified can be identified by performing detailed discrimination on several types of coins only in country B. .

  As described above, according to the present embodiment, the input image is subjected to polar coordinate conversion, the autocorrelation curve of the histogram is compared with the template for the band-like region having the characteristics for each country, and the individual comparison results regarding each region are The collation value is obtained by weighting and adding together for each, and the coin in the input image is classified as the coin of the country corresponding to the largest collation value. For this reason, it is possible to limit the number of candidates for the discrimination result of the coins to be discriminated to shorten the processing time for performing detailed discrimination, and when performing discrimination using the image of the object pattern, Can be shortened.

  In the above embodiment, the template corresponding to the maximum value of the matching value is used as the classification result. However, the template corresponding to the matching value equal to or higher than the predetermined threshold is obtained by comparing the matching value for each template with a predetermined threshold. May be used as the classification result. By doing so, it is possible to set a threshold value as appropriate, and output two or more countries as classification results, and flexible classification can be performed.

  Moreover, in the said one Embodiment, although the template memory | storage part 150 shall memorize | store the template for every country, you may make it memorize | store the template for every money type, for example. Also in this case, since the band-like region having a width corresponds to the template, the coins of a plurality of countries can be classified into one category (denomination).

  Further, in the above-described embodiment, the image of the coin portion is converted to polar coordinates. However, if the information on the rotation of the coins can be removed, the region on the orthogonal coordinate system is compared with the template without performing the polar coordinate conversion. You may make it do. If the information on the rotation of the coin can be removed, it is not necessary to obtain the autocorrelation curve of the histogram, and the histogram of each region may be stored as a template. Then, the normalized correlation between the histogram of each region and the template is obtained, the collation value is calculated in the same manner as in the above embodiment, and the classification result can be obtained by selecting the maximum collation value or determining the threshold value. At this time, the shape of the region serving as the template may be arbitrary, but by making the region having an area, it is possible to determine whether the region has a width as in the case of a belt-like region having a width in the polar coordinate system.

  Further, in the above-described embodiment, the identification of euro coins has been described as an example, but for example, when coins of the United States with different patterns for each state are classified, or when banknotes and medicine tablets are classified, The present invention can be applied to classify various objects.

  The present invention is suitable for shortening the time required for processing when performing discrimination using an image of an object pattern.

It is a block diagram which shows the structure of the image classification apparatus which concerns on one embodiment of this invention. It is a flowchart which shows the classification | category operation | movement of the input image which concerns on one embodiment. It is a figure which shows an example of the autocorrelation calculation which concerns on one embodiment. It is a figure which shows an example of the collation value calculation which concerns on one Embodiment.

Explanation of symbols

DESCRIPTION OF SYMBOLS 110 Edge detection part 120 Size detection part 130 Image extraction part 140 Coordinate conversion part 150 Template memory | storage part 160 Template setting part 170 Template comparison part 171 Addition area instruction | indication part 172 Histogram creation part 173 Autocorrelation calculation part 174 Normalization correlation calculation part 175 Collation value calculation unit 180 Maximum value selection unit

Claims (4)

An image classification device that identifies which region the money is from by comparing an input image including a pattern of money and a template,
Image summary data pattern to a characteristic feature areas of money for each region storage means for previously storing as a template for each region,
First calculation means for calculating image aggregate data for each feature region of the input image;
Second calculation means for calculating a collation value for each template by performing a correlation operation on the image total data calculated by the first calculation means and the template stored by the storage means;
Image classification apparatus characterized by matching value calculated by the second calculation means and an output means for outputting the resulting classify areas corresponding to the maximum of the template.   An image classification device for identifying which denomination and which region the money is in by comparing a template with an input image including a pattern of money,
  Storage means for preliminarily storing, as a template for each denomination, image aggregate data of a characteristic region characteristic of the money pattern of the same denomination across multiple regions;
  First calculation means for calculating image aggregate data for each feature region of the input image;
  Second calculation means for calculating a collation value for each template by performing a correlation operation on the image total data calculated by the first calculation means and the template stored by the storage means;
  Output means for outputting the denomination corresponding to the template having the maximum matching value calculated by the second calculating means as a classification result;
  With
  An image classification apparatus for identifying an area limited to denominations output by the output means.   An image classification method for identifying which region the money is from by comparing an input image including a pattern of money and a template,
  A storage step of preliminarily storing, as a template for each area, image aggregate data of characteristic areas characteristic to the pattern of money in each area;
  A first calculation step of calculating image aggregate data for each feature region of the input image;
  A second calculation step of calculating a collation value for each template by performing a correlation operation between the image total data calculated in the first calculation step and the template stored in the storage step;
  An output step of outputting, as a classification result, an area corresponding to a template having a maximum matching value calculated by the second calculation step;
  An image classification method characterized by including   An image classification method for identifying which denomination and which region the money is from by comparing an input image including a pattern of money and a template,
  A storage step of storing in advance image aggregated data of characteristic regions characteristic of money patterns of the same denomination across a plurality of regions as templates for each denomination;
  A first calculation step of calculating image aggregate data for each feature region of the input image;
  A second calculation step of calculating a collation value for each template by performing a correlation operation between the image total data calculated in the first calculation step and the template stored in the storage step;
  An output step of outputting, as a classification result, a denomination corresponding to a template having a maximum matching value calculated by the second calculation step;
  Including
  An image classification method characterized by specifying an area limited to the denominations output in the output step.

Images