JP2011014052A - Image identifying device, image identifying method, identifier creating device, identifier creating method and imaging system - Google Patents

Abstract

PROBLEM TO BE SOLVED: To classify images per category regarding an image identifying device, an image identifying method, an identifier creating device, an identifier creating method and an imaging system.SOLUTION: The image identifying device, identifying a given object image to belong to which of predefined categories, includes: a feature quantity calculation part which calculates feature quantity of an image to be input; a shared image removing part which creates processed images by removing a shared image region containing images of a predefined shared object from each of a plurality of learning images classified to the categories beforehand and inputs the processed image in the feature quantity calculation part; and an identifier creation part which creates beforehand an identifier for identifying the image to be input to belong to which of the categories by image identifying processing using the feature quantity calculated by the feature quantity calculation part for each of the processed images. When the identifier creation part identifies the category of the object image after creating the identifier by the identifier creation part, the feature quantity calculation part calculates the feature quantity of the object image and inputs the feature quantity in the identifier.

Description

  The present invention relates to an image identification device, an image identification method, a discriminator creation device, a discriminator creation method, and an imaging system.

  2. Description of the Related Art Conventionally, a method for classifying an image by comparing the characteristics of an image such as a photograph with the characteristics of a learning image that is classified in advance by category is known. For example, Patent Document 1 discloses a method for classifying images based on image feature amounts.

JP 2009-009402 A

  In the method of classifying an image according to the feature amount of the image, the property that the feature amount of the image differs according to the feature of the subject included in the image is used. For example, the feature of the image when the subject is a flower is different from the feature of the image when the subject is a beach. Therefore, by comparing the characteristics of the images to be classified with the characteristics of the images of the respective categories, the target images can be classified into categories with similar characteristics. As an example, the feature of an image is represented using a feature amount calculated using parameters such as chromaticity and luminance of the image.

  However, in the learning image, there may be a subject other than the subject necessary for category classification. For example, even if the image belongs to “flower” as a category, there is a case where a predetermined person is captured together with the flower. In such a case, the feature amount of an image in which a predetermined person is photographed together with a flower is different from the feature amount of an image in which only a flower is photographed. Therefore, when an image in which a predetermined person is photographed together with a flower is used as a learning image, an error in a feature amount used for image classification determination increases, and the image may be classified into an incorrect category. For example, when the learning image includes a family image of a family or the like that is shown regardless of the category, it is preferable that the personal image is not used for calculating the feature amount.

  In order to solve the above-described problem, in the first aspect of the present invention, an input image is an image identification device that identifies which category of an image to be given is a predetermined target image. Generating a processed image by removing a common image area including a predetermined common subject image from each of a plurality of learning images previously classified into categories The image to be input is an image of any category by the common image removal unit input to the feature amount calculation unit and the image identification processing using the feature amount calculated by the feature amount calculation unit for each processed image. A discriminator creating unit that pre-generates a discriminator for identifying whether or not the feature amount calculating unit identifies a category for the target image after the discriminator creating unit generates the discriminator. To provide an image identification apparatus to be input to the discriminator calculates a feature amount.

  In order to solve the above-described problem, in the second aspect of the present invention, the common image removal unit uses a common area obtained by enlarging an area designated as an image area of a common subject by a predetermined enlargement factor. Provided is an image identification device as an image region.

  In order to solve the above-described problem, in the third aspect of the present invention, the common image removal unit sets a region circumscribing the common subject image as a common subject image region and enlarges the region by an enlargement factor. Provided is an image identification device in which a region is a common image region.

  In order to solve the above problem, in the fourth aspect of the present invention, the common image removal unit extracts images of a common subject from a plurality of learning images by performing image matching using a reference image given in advance. An image identification device is provided.

  In order to solve the above problem, in the fifth aspect of the present invention, the common image removal unit removes each of the plurality of learning images by using at least information on pixels adjacent to the removed common image region. Provided is an image identification device that complements image data of a specified region.

  In order to solve the above-described problem, in the sixth aspect of the present invention, the common image removal unit is configured to generate a common based on a size obtained by multiplying the length of one side of the entire learning image by a predetermined enlargement factor. An image identification device for enlarging an image area of a subject is provided.

  In order to solve the above-described problem, in the seventh aspect of the present invention, the common image removal unit has a size obtained by multiplying the length of one side of the entire learning image by the first enlargement factor, or a common size. A size obtained by multiplying the length of one side of the image area of the subject by the second enlargement factor is selected based on the size of the image area of the common subject, and according to the selected size. An image identification device for enlarging an image area of a common subject is provided.

  In order to solve the above-described problem, in the eighth aspect of the present invention, the feature quantity calculation unit can calculate the feature quantity by a plurality of types of methods, and the feature quantity calculation unit is provided for each category by the common image removal unit. Further calculate a common feature amount in a plurality of learning images from which the common image region is not removed, and the discriminator calculates the feature amount from the processed image according to which method the feature amount calculated is compared with An image identification device that selects whether to use a feature amount calculated from a learning image is provided.

  The above summary of the invention does not enumerate all necessary features of the present invention. In addition, a sub-combination of these feature groups can also be an invention.

1 shows a configuration of an image identification device 100 according to the present embodiment. An example of a method for designating a common image area in a learning image is shown. The learning image with which the image data of the removed common image area | region was complemented is shown. An example of the target image to classify is shown. An example of an image when the image data of the removed common image area is not complemented is shown. An example in which the common image area is not correctly specified is shown. An example of an image complemented when the common image area is not correctly specified is shown. An example of a method for generating an enlarged common image region will be described. Another example of a method for generating an enlarged common image region will be described. An example of a reference image is shown. The structure of the image identification apparatus 100 which can select whether a common image area | region is removed is shown. The operation | movement flow which produces the discrimination device 50-1 is shown. The operation | movement flow which produces the discrimination device 50-2 is shown. The operation | movement flow which classify | categorizes a target image is shown. 2 shows a configuration of an imaging system 150 according to another embodiment.

  Hereinafter, the present invention will be described through embodiments of the invention, but the following embodiments do not limit the invention according to the claims. In addition, not all the combinations of features described in the embodiments are essential for the solving means of the invention.

  FIG. 1 shows a configuration of an image identification device 100 according to the present embodiment. The image identification device 100 includes a common image removal unit 20, a feature amount calculation unit 30, a classifier creation unit 40, and a classifier 50. The image identification device 100 identifies which category a predetermined target image belongs to.

  The common image removal unit 20 generates a processed image obtained by removing a common image region including a predetermined common subject image from each of a plurality of learning images classified in advance into a category, and calculates a feature amount. Input to section 30. The feature amount calculation unit 30 calculates the feature amount of the input image.

  The discriminator creating unit 40 creates the discriminator 50 in advance according to the feature amount of the learning image calculated by the feature amount calculating unit 30. The discriminator 50 identifies which category of image the input image is by image discrimination processing using the feature amount calculated by the feature amount calculation unit 30 for each processed image. Prior to the classification of the target image, the image identification device 100 may create the classifier 50 based on a learning image input by the manufacturer or user of the image identification device 100. Hereinafter, the operation of each unit will be described in detail.

  The common image removal unit 20 generates a processed image obtained by removing a common image region including a predetermined common subject image from each of a plurality of learning images classified in advance into categories. Specifically, the common image removal unit 20 acquires a plurality of learning images in which subjects classified into categories such as “cherry blossom viewing”, “seaside”, “festival”, and “wedding” are captured in advance.

  Subsequently, the common image removal unit 20 removes the common image region included in the acquired learning image. The common image area is an area including a predetermined subject included in the learning image regardless of the category of the learning image. For example, the common image area may be an area circumscribing an image showing a predetermined person such as a relative.

  The common image removal unit 20 may acquire common image information indicating a common image region from the outside. For example, the common image removal unit 20 may determine the common image region based on position information designated by the user of the image identification device 100. The common image removal unit 20 may display a learning image so that the user of the image identification device 100 can designate a common image region. The common image removal unit 20 may acquire information indicating the boundary line of the common image area as the position information, and may acquire information indicating the positions of the four corners of the common image area as the position information.

  The feature amount calculation unit 30 receives the processed image generated by the common image removal unit 20. The feature amount calculation unit 30 calculates a feature amount that a plurality of processed images have in common for each category based on the received processed images. For example, the feature amount calculation unit 30 may use information corresponding to the number of edges included in the processed image as a component of the feature amount. Specifically, the feature amount calculation unit 30 divides the processed image area into a plurality of areas each having a predetermined number of pixels, and then the edges included in some or all of the plurality of divided areas. You may accumulate numbers. Note that an edge is a portion where luminance or chromaticity changes by a predetermined amount or more between adjacent pixels.

  The feature amount calculation unit 30 may calculate the component of the feature amount by digitizing the color information of the processed image. For example, the feature amount calculation unit 30 may generate a histogram based on the frequency of occurrence of each color constituting the color space after converting the color space of the learning image into the HSV color space. The feature amount calculation unit 30 may use numerical information on the occurrence frequency of each color as a component of the feature amount.

  The feature amount calculation unit 30 may calculate the component of the feature amount by digitizing the luminance information of the processed image. For example, the feature amount calculation unit 30 may use the average and variance of the brightness of the processed image as the component of the feature amount. The feature amount calculation unit 30 may perform predetermined conversion on the luminance of the processed image, and may use energy and dispersion in the converted space as components of the feature amount. For example, the feature amount calculation unit 30 may convert the color space of the processed image into the HSV color space and then use the energy and dispersion in the converted color space as components of the feature amount.

  The feature amount calculation unit 30 may calculate the component of the feature amount based on the meta information included in the learning image. Specifically, in the case where the learning image is an image generated in conformity with MPEG-7, information on the edge indicated in the video signal feature descriptor, information on the color occurrence frequency, information on the luminance, etc. May be used.

  The discriminator creation unit 40 acquires the feature amount of each learning image calculated by the feature amount calculation unit 30 and the category of each learning image. The discriminator creating unit 40 calculates the boundary value of the region in which the feature amount of each category is distributed based on the correspondence relationship between the category to which the learning image belongs and the feature amount of each learning image.

  For example, the feature amount of the learning image will be described as a vector having n components (that is, an n-dimensional vector, hereinafter referred to as a feature amount vector). The discriminator creation unit 40 converts the boundary between the feature amounts in the first category and the second category into the distance from each feature amount vector of the learning image belonging to the first category to the boundary, and the second category. You may calculate based on the distance from each feature-value vector of the learning image to which it belongs to the said boundary.

  As an example, the classifier creation unit 40 includes the minimum value of the distance from each feature amount vector of the learning image belonging to the first category to the boundary and the feature amount vector of the learning image belonging to the second category to the boundary. The boundary may be selected such that the minimum value of the distance is the largest. Further, the discriminator creating unit 40, at a predetermined value x1 on the x-axis, the distance from the feature vector of the plurality of learning images belonging to the first category and the feature amount of the plurality of learning images belonging to the second category. By calculating the y-axis value y1 that is equal to the distance from the vector, x1 and y1 may be used as boundary values. The discriminator creation unit 40 may determine the boundary between categories by calculating the boundary value for a plurality of values on the x-axis and connecting the calculated boundary values. The discriminator creation unit 40 may calculate the boundary using a so-called support vector machine (Support Vector Machine, SVM).

  Further, the discriminator creating unit 40 may calculate the boundary between the feature amounts of each category using a value obtained by projecting the feature amount vector of each learning image in the axial direction of the maximum eigenvector. Here, the axial direction of the maximum eigenvector may be a direction in which the distribution obtained when each feature vector is plotted in the n-dimensional space is most widely dispersed. That is, the direction may be the direction of the maximum eigenvector of the variance in the distribution. The discriminator creation unit 40 may calculate a boundary between categories by a so-called discriminant analysis method.

  For example, a value obtained by projecting each feature vector of the learning image belonging to the first category in the axial direction of the maximum eigenvector is included in a range of A to B (A and B are real numbers) and belongs to the second category. It is assumed that the value obtained by projecting each feature vector of the learning image in the axial direction of the maximum eigenvector is distributed between C and D (C and D are real numbers). The discriminator creation unit 40 may use the value between B and C as the boundary value of the feature quantity distribution region corresponding to the first category and the second category. For example, the discriminator creating unit 40 may use an intermediate value between B and C as a boundary value.

  The discriminator creation unit 40 may calculate the boundary value for each category based on the probability density distribution function of the value obtained by projecting the feature vector in the axial direction of the maximum eigenvector. For example, in the above example, when B is larger than C, the feature amount distribution region of the learning image belonging to the first category overlaps with the feature amount distribution region of the learning image belonging to the second category. Therefore, the discriminator creating unit 40 has an on-axis where the occurrence probability in the learning image belonging to the first category is equal to the occurrence probability in the learning image belonging to the second category in the region from B to C on the axis. May be the boundary value between the first category and the second category.

  The vector components described above may be, for example, a component based on the chromaticity of the image, a component based on the luminance of the image, or the like. Further, the discriminator creating unit 40 may create a discriminator using a plurality of types of feature quantity vectors. For example, the classifier creation unit 40 may create a lower classifier for each type of feature vector. Further, the discriminator creating unit 40 may create one upper discriminator for a plurality of lower discriminators.

  In this case, the lower classifier performs the above-described category discrimination based on the corresponding type of feature vector in the given target image. The upper classifier determines the category of the target image based on the discrimination result in the lower classifier. The upper classifier may determine the category of the target image by majority decision of the category discrimination result in the lower classifier. Further, the discriminator creating unit 40 may create the discriminator (or lower classifier) described above using a feature quantity vector obtained by combining a plurality of feature quantity vectors.

  The discriminator creating unit 40 inputs the calculated boundary value of the feature quantity between the categories to the discriminator 50. The discriminator 50 stores the category of the learning image, the type of the feature amount, and the boundary value of the feature amount in each category in association with each other in the storage medium.

  When the user of the image identification device 100 classifies the target image, the feature amount calculation unit 30 receives an input of the target image to be classified. The feature amount calculation unit 30 calculates the feature amount of the target image using the same method as that used for calculating the feature amount of the learning image. For example, when the information corresponding to the number of edges is used in the calculation of the feature amount of the learning image, the feature amount calculation unit 30 calculates the feature amount of the target image by integrating the number of edges of the target image. You can do it.

  When the feature amount calculation unit 30 generates a feature amount of a learning image using a plurality of feature amounts, the feature amount calculation unit 30 may select one or more of the plurality of methods used for calculating the feature amount of the learning image. A feature amount of the target image may be calculated by selecting a method. For example, when the feature amount calculation unit 30 calculates the feature amount of the learning image based on the number of edges, the color occurrence frequency, and the luminance, the feature amount calculation unit 30 uses one of the methods to The feature amount may be calculated. The feature amount calculation unit 30 may calculate the feature amount of the target image by combining a plurality of methods.

  The feature amount calculation unit 30 inputs the calculated feature amount of the target image to the classifier 50. The discriminator 50 identifies to which category the inputted feature quantity of the target image belongs. For example, when the boundary between categories is calculated by the discriminant analysis method described above, the feature quantity distribution region corresponding to the first category on the axis of the maximum eigenvector is 10 to 50 and corresponds to the second category. The feature amount distribution region to be set is 50 to 100. In this case, for example, when a target image whose value obtained by projecting a feature vector onto the axis is 30 is given, the classifier 50 classifies the target image into the first category.

  As described above, the image identification device 100 according to the present embodiment calculates the feature amount of the learning image from which the image of the subject commonly included in the learning image is removed regardless of the category of the learning image. The accuracy of the feature amount of the learning image calculated at 30 is improved. As a result, the accuracy of the boundary value of the feature quantity distribution region between categories stored in the classifier 50 is improved. The image identification device 100 can classify the target image into an appropriate category by classifying the target image using the classifier 50.

  FIG. 2 shows an example of a method for designating a common image area in a learning image. The learning image 200 (200-1, 200-2) has a background (shaded pattern) belonging to the first category. The learning image 210 (210-1, 210-2) has a background (horizontal stripe pattern) belonging to the second category. The common image area 220 (220-1, 220-2) and the common image area 230 (230-1, 230-2) indicate areas to be removed from the learning image 200 and the learning image 210, respectively. The common image removal unit 20 may set a region circumscribing the image of the subject captured regardless of the category as the common image region 220. For example, the common image area 220 may be an area circumscribing a predetermined person included in the learning image 200.

  Specifically, the common image removal unit 20 may use the position where the line parallel to the horizontal side of the learning image 200 is in contact with the image of the subject to be removed as the end of the horizontal region of the common image region 220. . Further, the common image removal unit 20 may set the position where the line parallel to the vertical side of the learning image 200 is in contact with the image of the subject to be removed as the end of the common image region 220 in the vertical direction.

  As an example, the common image region 220 may be determined by a user of the image identification device 100 specifying a line indicating the region. Further, the common image area 220 may be determined based on position information of the four corners of the common image area 220 specified by the user of the image identification device 100. The position information may be, for example, coordinates having the origin as one of the four corners of the entire learning image.

  Note that the common image region 220 and the common image region 230 may be located at any position in the learning image 200 and the learning image 210. The common image region 220 and the common image region 230 may have a size having any ratio with respect to the size of the learning image 200 and the learning image 210.

  The common image removal unit 20 uses at least information on the pixels adjacent to the removed common image region 220 and the common image region 230 for each of the plurality of learning images 200 and the learning image 210, and image data of the removed region. May be supplemented. For example, the common image removal unit 20 complements the image data of the removed area by copying the same number of pixels as the predetermined number of pixels adjacent to the common image area 220 into the removed common image area 220. It's okay.

  FIG. 3 shows an example of an image in which the image data of the removed common image region is complemented. The learning image 300 and the learning image 310 are images obtained by complementing the common image region 220 and the common image region 230 shown in FIG. 3 removed by the common image removing unit 20 with the background images of the learning image 200 and the learning image 210.

  The common image region 320-1 and the common image region 320-2 are respectively the common image region 220-1 and the common image region 220-2 shown in FIG. -2 is an area supplemented by image data of pixels adjacent to -2. Similarly, the common image region 430-1 and the common image region 430-2 are respectively the common image region 230-1 and the common image region 230-2, and the common image region 320-1 and the common image region 320-2. This is a region supplemented by image data of pixels adjacent to.

  The feature amount calculation unit 30 calculates the feature amounts of the learning image 300 and the learning image 310. The classifier 50 may store the feature quantities of the learning image 300 and the learning image 310 in association with the names of the categories to which the learning images 300 and 310 belong. For example, when the pattern of the learning image 300 indicates a cherry blossom viewing landscape, the classifier 50 stores the feature amounts of the learning image 300-1 and the learning image 300-2 in association with the category name “hanami”. Good. When the pattern of the learning image 310 indicates a seaside landscape, the classifier 50 may store the feature amounts of the learning image 310-1 and the learning image 310-2 in association with the category name “seaside”.

  The feature amount calculation unit 30 may calculate an average value of the feature amounts of the learning images included in each of the learning image 300 and the learning image 310. The discriminator 50 may store the calculated average value of the feature amounts of the learning images included in each of the learning image 300 and the learning image 310 in association with the name of the category to which each belongs. Specifically, the discriminator 50 may store the average value of the feature values of the learning image 300-1 and the learning image 300-2 in association with “cherry blossom viewing”. Similarly, the discriminator 50 may store the average value of the feature amounts of the learning image 510-1 and the learning image 510-2 in association with the “seaside”.

  By performing the complementing process described above, the area of the image showing the characteristics of the category becomes large. Therefore, the image identification device 100 can improve the accuracy of the calculated learning image feature amount.

  FIG. 4 shows an example of target images to be classified. In the figure, a target image 400-1 is an image including an image of a predetermined person in a background similar to the background of the learning image 200 shown in FIG. The target image 400-2 is an image including an image of a predetermined person in a background similar to the background of the learning image 210.

  The feature amount calculation unit 30 calculates the feature amounts of the target image 400-1 and the target image 400-2. The feature amount calculation unit 30 calculates a feature amount without removing an image of an area including a predetermined person. The discriminator 50 identifies to which category the feature amount of each image calculated by the feature amount calculation unit 30 belongs.

  Specifically, the discriminator 50 determines the distance between the feature amount of the target image 400-1 and the feature amount distribution area of each category determined using the feature amounts of the learning image 300 and the learning image 310 illustrated in FIG. Is calculated. Since the background of the target image 400-1 is similar to the image of the learning image 300, the feature amount of the target image 400-1 is more in the feature amount distribution region of the learning image 300 than the feature amount distribution region of the learning image 310. Has a close value. Therefore, the classifier 50 determines that the feature amount of the target image 400-1 is close to the learning image 300. For example, the classifier 50 classifies the target image 400-1 into the category “cherry-blossom viewing”.

  Further, since the background of the target image 400-2 is similar to the background of the learning image 310, the feature amount of the target image 400-2 is more than the feature amount distribution region of the learning image 300. It has a value close to the region. Therefore, the classifier 50 determines that the feature amount of the target image 400-2 is close to the learning image 310. For example, the classifier 50 classifies the target image 400-2 into the category “seaside”.

  FIG. 5 shows a learning image when the image data of the removed common image region is not complemented. The common image removal unit 20 may complement the image data without the pattern without complementing the image data of the removed common image region. In the learning image 500 and the learning image 510 illustrated in FIG. 5, the area from which the common image area 220 and the common image area 230 illustrated in FIG. 2 are removed is complemented by the common image area 520 and the common image area 530 that have no pattern. ing.

  In the common image area 520 and the common image area 530, the chromaticity and luminance used for calculating the feature amount do not change. Further, since the common image area 520 and the common image area 530 do not include the edge of the image, the influence on the calculation of the feature amount is small. Therefore, the feature amount of the learning image 500 including the common image region 520 and the feature amount of the learning image 510 including the common image region 530 are close to the feature amount of the background region of the learning image 500 and the learning image 510, respectively. As a result, the feature amount calculation unit 30 calculates the feature amount of the learning image 500 and the learning image 510, thereby removing the feature amount from the learning image 200 and the learning image 210 without removing an image of a region including a predetermined person. Compared to the case of calculating, it is possible to calculate a feature amount close to the feature amount of each background region.

  As described above, according to the image identification device 100 according to the present embodiment, it is possible to calculate a feature amount with high accuracy for each category to which an image belongs. As a result, the classification accuracy of the target image is improved. However, when the user of the image identification device 100 designates a common image area, if the image area to be removed is not designated, the accuracy is lowered.

  FIG. 6A shows an example when the common image region is not correctly specified. In the common image area 620 included in the learning image 600 in the figure, a smaller area is specified than the learning image 200-1 in FIG. As a result, part of the head and clothes of a predetermined person to be removed are not removed. If the feature amount is calculated using an image in which a part of the subject to be originally removed remains, an error occurs in the feature amount. In particular, when the common image region 620 is complemented by an image circumscribing the common image region 620, if a portion having characteristics different from the background color and pattern is left, the common image region 620 is complemented by the portion.

  FIG. 6B shows an example of an image when complemented when the common image region is not correctly specified. A partial area of the common image area 630 is complemented by a partial pattern of clothes of a predetermined person shown in FIG. 6A. As a result, the feature amount calculation unit 30 calculates the feature amount based on an image having a feature different from the background feature included in the learning image 600, so that an error included in the feature amount of the learning image 600 increases.

  Therefore, the common image removal unit 20 may set a region obtained by enlarging a region designated as a common subject image region by a predetermined enlargement factor as a common image region. For example, the common image removal unit 20 may enlarge the image area of the common subject based on a size obtained by multiplying the length of one side of the entire learning image by a predetermined enlargement factor. Specifically, the common image removal unit 20 adds a length obtained by multiplying the length of one side of the entire learning image by a predetermined enlargement factor to the four sides of the image area designated by the user of the image identification device 100. By doing so, an enlarged common image region may be generated.

  Here, if the common image area to be removed by the common image removing unit 20 is enlarged, the possibility that the subject to be removed remains low, but the image area that can be used for calculating the feature quantity is reduced. On the other hand, if the common image area to be removed by the common image removal unit 20 is reduced, the possibility that the subject to be removed remains high. Therefore, as an example, the common image removal unit 20 sets the length of about 1/30 of any one side of the entire learning image to four sides of the image region designated by the user of the image identification device 100. By adding, an enlarged common image area may be generated.

  FIG. 7 shows an example of a method for generating an enlarged common image region. In the figure, the length of one side in the horizontal direction of the learning image 700-1 is H, and the length of one side in the vertical direction is V. The common image region 720-1 is a common image region designated by the user of the image identification device 100.

  The learning image 700-2 has a common image area 720-2 obtained by enlarging the common image area 720-1. H1 represents the distance between the vertical left side of the common image region 720-1 and the vertical left side of the common image region 720-2 when the learning image 700-1 and the learning image 700-2 are superimposed. Show. H2 represents the distance between the vertical right side of the common image region 720-1 and the vertical right side of the common image region 720-2 when the learning image 700-1 and the learning image 700-2 are superimposed. Show. H1 and H2 may be a length obtained by multiplying the length of one side of the learning image 700-1 by a predetermined enlargement factor.

  V1 represents the distance between the upper side in the horizontal direction of the common image region 720-1 and the upper side in the horizontal direction of the common image region 720-2 when the learning image 700-1 and the learning image 700-2 are superimposed. Show. V2 represents the distance between the horizontal lower side of the common image region 720-1 and the horizontal lower side of the common image region 720-2 when the learning image 700-1 and the learning image 700-2 are superimposed. Show. V1 and V2 may be a length obtained by multiplying the length of one side of the learning image 700-1 by a predetermined enlargement factor.

  The common image removal unit 20 may generate the common image region 720-2 by adding the length of H1, H2, V1, or V2 to each of the four sides of the common image region 720-1. Each of H1, H2, V1, and V2 may be a value obtained by multiplying H by a predetermined enlargement factor. Further, H1, H2, V1, and V2 may be values obtained by multiplying V by a predetermined enlargement factor. As described above, the common image removal unit 20 generates the enlarged common image region, so that the image of the subject that does not belong to the category of the learning image is not included in the common image region. As a result, the accuracy of the feature amount of the learning image calculated by the feature amount calculation unit 30 is improved.

  FIG. 8 shows another example of a method for generating an enlarged common image region. The common image removal unit 20 may use a region circumscribing a common subject image as a common subject image region, and a region obtained by enlarging the region with an enlargement factor as a common image region. Specifically, when the user of the image identification device 100 designates the common image area 820-1 that substantially circumscribes the image of the common subject, the common image removal unit 20 selects one side of the common image area 820-1. An enlarged common image region 820-2 may be generated by adding a length obtained by multiplying the length of the image by a predetermined enlargement factor to the four sides of the common image region 820-1.

  The common image removal unit 20 adds a length obtained by multiplying the length of the horizontal side of the common image region 820-1 by a predetermined enlargement factor to the horizontal side of the common image region 820-1, thereby The length of the horizontal side of the image area 820-2 may be determined. Further, the common image removal unit 20 adds a length obtained by multiplying the length of the vertical side of the common image region 820-1 by a predetermined enlargement factor to the vertical side of the common image region 820-1. The length of the side of the common image region 820-2 in the vertical direction may be determined. Through the above procedure, the common image removal unit 20 can generate a common image region 820-2 in which the common image region 820-1 is enlarged.

  Specifically, H3 represents the left side in the vertical direction of the common image area 820-1 and the left side in the vertical direction of the common image area 820-2 when the learning image 800-1 and the learning image 800-2 are superimposed. Indicates the distance between. H4 represents the distance between the right side in the vertical direction of the common image region 820-1 and the right side in the vertical direction of the common image region 820-2 when the learning image 800-1 and the learning image 800-2 are superimposed. Show. H3 and H4 may be a length obtained by multiplying the length of the horizontal side of the common image region 820-1 by a predetermined enlargement factor.

  V3 represents the distance between the upper side in the horizontal direction of the common image region 820-2 and the upper side in the horizontal direction of the common image region 820-2 when the learning image 800-1 and the learning image 800-2 are superimposed. Show. V4 represents the distance between the horizontal lower side of the common image region 820-1 and the horizontal lower side of the common image region 820-2 when the learning image 800-1 and the learning image 800-2 are superimposed. Show. V3 and V4 may be a length obtained by multiplying the length of the side in the vertical direction of the common image region 820-1 by a predetermined enlargement factor.

  The common image removal unit 20 has a size obtained by multiplying the length of one side of the entire learning image by the first enlargement factor, or the length of one side of the image area of the common subject. May be selected based on the size of the image area of the common subject, and the image area of the common subject may be enlarged according to the selected size. For example, when the image area of a common subject is larger than a predetermined size, if a size obtained by multiplying the length of one side of the entire learning image by a first enlargement factor is added to the image area, the image belonging to the category In some cases, it is not possible to secure a sufficient image area for calculating the feature amount of the image. Therefore, when the image area of the common subject is equal to or larger than the predetermined size, the common image removal unit 20 sets the size obtained by multiplying the length of one side of the image area of the common subject by the second enlargement factor. A common image area to be removed may be generated by adding to the image area.

  If the image area of the common subject is less than a predetermined size, the feature amount can be obtained even if a size obtained by multiplying the length of one side of the entire learning image by the first enlargement factor is added to the image area. A sufficient image area can be secured for the calculation. Therefore, when the image area of the common subject is less than a predetermined size, the common image removal unit 20 sets the size of the entire learning image by multiplying the length of one side of the entire learning image by the second enlargement factor. By adding, a common image area to be removed may be generated.

  The common image removal unit 20 may extract a common subject image from a plurality of learning images by performing image matching using a reference image given in advance. For example, an image of a predetermined person or the like previously selected by the user of the image identification device 100 may be used as the reference image.

  FIG. 9 shows an example of a reference image. For example, the reference image 900 may be an image equal to the image included in the common image region 220-1 illustrated in FIG. The common image removal unit 20 calculates and stores the feature amount of the reference image 900.

  For example, when extracting the common image area 820-1 in the learning image 800, the common image removal unit 20 selects an area having a size equal to that of the reference image 900 from the predetermined position of the learning image 800 at regular intervals. . The common image removal unit 20 calculates a feature amount in each selected region. The common image removal unit 20 may compare the feature amount in each region with the feature amount of the reference image 900 and determine that the region having the closest feature amount is the common image region 820-1.

  In the above embodiment, the method for calculating the feature amount of the learning image after removing the common image region has been described. However, the image identification device 100 may select whether or not to remove the common image region when calculating the feature amount of the learning image, depending on the condition.

  FIG. 10 shows a configuration of the image identification device 100 that can select whether or not to remove the common image region. In the figure, the feature amount calculation unit 30 includes a feature amount calculation unit 30-1, a feature amount calculation unit 30-2, and a feature amount calculation unit 30-3. The discriminator creating unit 40 includes a discriminator creating unit 40-1 and a discriminator creating unit 40-2. The discriminator 50 includes a discriminator 50-1, a discriminator 50-2, and a selection unit 52.

  The feature amount calculation unit 30-1 calculates the feature amount of the image from which the common image area has been removed by the common image removal unit 20. The discriminator creating unit 40-1 creates the discriminator 50-1 based on the feature amount calculated by the feature amount calculating unit 30-1.

  In the present embodiment, since the calculation capability of the feature amount calculation unit 30-2 is higher than the calculation capability of the feature amount calculation unit 30-1, the feature amount calculation unit 30-2 may include the feature amount calculation unit 30- within a predetermined time. The feature amount can be calculated with high accuracy by more than one type of method. When the feature amount calculation unit 30-2 can calculate the feature amount with high accuracy, it is preferable that the area of the image region that is the target of the feature amount calculation is large. Therefore, the feature amount calculation unit 30-2 calculates the feature amount of the learning image from which the common image region is not removed. The discriminator creating unit 40-2 creates the discriminator 50-2 based on the feature amount calculated by the feature amount calculating unit 30-2.

  The selection unit 52 selects which of the classifier 50-1 and the classifier 50-2 is used to classify the target image. The selection unit 52 may select a classifier to be used based on selection information input by the user of the image identification device 100. The selection unit 52 inputs the feature amount of the target image input from the feature amount calculation unit 30-3 to the selected discriminator. For example, the selection unit 52 selects the discriminator 50-1 when classifying the target image in a short time, and classifies the target image with high accuracy using the feature amount calculated using many types of methods. In this case, the discriminator 50-2 may be selected.

  The feature amount calculation unit 30-3 calculates the feature amount of the target image. The feature amount calculation unit 30-3 calculates the feature amount of the target image using the feature amount calculation method used to create the classifier selected by the selection unit 52. As described above, in the image identification device 100 according to the present embodiment, since a plurality of classifiers can be selected, the target image is selected using an optimal classifier according to the classification time, the accuracy required for classification, and the like. Can be classified.

  Note that the feature amount calculation unit 30-1, the feature amount calculation unit 30-2, and the feature amount calculation unit 30-3 may be configured by using the same hardware. For example, a program that causes a computer to function as the feature amount calculation unit 30-1, a program that causes the computer to function as the feature amount calculation unit 30-2, a program that causes the computer to function as the feature amount calculation unit 30-3, and the like can be used. You may comprise the feature-value calculation part 30-1, the feature-value calculation part 30-2, and the feature-value calculation part 30-3 by making it operate | move.

  Similarly, the discriminator 40-1 and the discriminator 40-2 may be configured by using the same hardware. For example, the discriminator 40-1 and the discriminator 40-2 are configured by operating the computer with a program for causing the computer to function as the discriminator 40-1 and a program for causing the computer to function as the discriminator 40-2. Good.

  FIG. 11A shows an operation flow for creating the discriminator 50-1. The common image removal unit 20 acquires a learning image and common image information (S101), and removes the common image region (S102). The feature amount calculation unit 30-1 calculates the feature amount of the learning image in a state where the common image region is removed (S103). The discriminator creating unit 40-1 creates the discriminator 50-1 when using the learning image from which the common image region is removed based on the calculated feature amount of the learning image (S104).

  FIG. 11B shows an operation flow for creating the discriminator 50-2. The feature amount calculation unit 30-2 acquires a learning image (S111), and calculates a feature amount of the learning image (S112). The discriminator creating unit 40-2 creates the discriminator 50-2 when a learning image that does not remove the common image is used based on the calculated feature amount of the learning image (S113).

  FIG. 11C shows an operation flow for classifying target images. The feature amount calculation unit 30-3 acquires the target image (S121). The selection unit 52 selects a discriminator to be used for classification of the target image based on information input from the outside (S122).

  When the selection unit 52 selects the discriminator 50-1 (S123), the feature amount calculation unit 30-3 uses the feature amount calculation method used by the feature amount calculation unit 30-1 to calculate the feature amount of the target image. Is calculated (S124). The selection unit 52 inputs the feature amount input from the feature amount calculation unit 30-3 to the classifier 50-1. The discriminator 50-1 classifies the target image corresponding to the input feature amount into the optimum category (S125).

  When the selection unit 52 selects the discriminator 50-2 (S123), the feature amount calculation unit 30-3 uses the feature amount calculation method used by the feature amount calculation unit 30-2 to perform the feature amount of the target image. Is calculated (S126). The selection unit 52 inputs the feature amount input from the feature amount calculation unit 30-4 to the discriminator 50-2. The discriminator 50-2 classifies the target image corresponding to the input feature amount into the optimum category (S127).

  FIG. 12 shows a configuration of an imaging system 150 according to another embodiment. The imaging system 150 further includes an imaging unit 120 with respect to the image identification device 100 illustrated in FIG. For example, the imaging system 150 may be a digital camera. The imaging unit 120 inputs a target image obtained by imaging the imaging target to the feature amount calculation unit 30. The discriminator 50 classifies the feature amount of the target image calculated by the feature amount calculation unit 30 into the category having the closest feature amount among the categories whose feature amounts have been calculated in advance. The discriminator 50 may add information corresponding to the classified category to the meta information area of the target image.

  The imaging system 150 may classify and display the target images acquired from the external storage medium according to the category according to the operation of the user of the imaging system 150. Further, the imaging system 150 may attach information corresponding to the category to which the target image belongs to the meta information area of the target image and output the information to an external device. In this manner, by classifying the target images acquired by the imaging system 150 into categories, searching for related images, storing for each related image, or displaying only related images continuously. Can do.

  As mentioned above, although this invention was demonstrated using embodiment, the technical scope of this invention is not limited to the range as described in the said embodiment. It will be apparent to those skilled in the art that various modifications or improvements can be added to the above-described embodiment. It is apparent from the scope of the claims that the embodiments added with such changes or improvements can be included in the technical scope of the present invention.

  The order of execution of each process such as operations, procedures, steps, and stages in the apparatus, system, program, and method shown in the claims, the description, and the drawings is particularly “before” or “prior to”. It should be noted that the output can be realized in any order unless the output of the previous process is used in the subsequent process. Regarding the operation flow in the claims, the description, and the drawings, even if it is described using “first”, “next”, etc. for convenience, it means that it is essential to carry out in this order. It is not a thing.

  20 common image removal unit, 30 feature quantity calculation unit, 40 classifier creation unit, 50 classifier, 52 selection unit, 100 image identification device, 120 imaging unit, 150 imaging system, 200 learning image, 210 learning image, 220 common image Area, 230 common image area, 300 learning image, 310 learning image, 400 target image, 500, learning image, 510 learning image, 520 common image area, 530 common image area, 600 learning image, 620 common image area, 630 common image Area, 700 learning image, 720 common image area, 800 learning image, 820 common image area, 900 reference image

Claims (12)

An image identification device for identifying a predetermined category of an image to be given,
A feature amount calculation unit for calculating a feature amount of an input image;
A processed image obtained by removing a common image area including a predetermined common subject image from each of a plurality of learning images previously classified in the category is generated and input to the feature amount calculation unit An image removal unit;
A discriminator that identifies which category of image the input image belongs to is generated in advance by image discrimination processing using the feature amount calculated by the feature amount calculation unit for each processed image. And a discriminator creation unit,
The feature quantity calculation unit calculates the feature quantity of the target image and inputs it to the classifier when the classifier creation unit identifies a category for the target image after generating the classifier. . The image identification device according to claim 1, wherein the common image removing unit enlarges an area designated as an image area of the common subject by a predetermined enlargement factor as the common image area. 3. The common image removal unit includes a region circumscribing the common subject image as the common subject image region, and a region obtained by enlarging the region by the enlargement factor as the common image region. Image identification device. The image identification device according to claim 3, wherein the common image removal unit extracts images of the common subject from the plurality of learning images by performing image matching using a reference image given in advance. The said common image removal part complements the image data of the removed area | region using at least the information of the pixel adjacent to the removed said common image area | region about each of these learning images. The image identification device according to any one of the above. The said common image removal part expands the image area | region of the said common subject based on the magnitude | size which multiplied the length of one side of the whole learning image by the predetermined said expansion coefficient. Image identification device. The common image removal unit is configured to multiply the length of one side of the entire learning image by a first enlargement factor, or to the length of one side of the image area of the common subject, The size multiplied by the coefficient is selected based on the size of the image area of the common subject, and the image area of the common subject is enlarged according to the selected size. The image identification device described in 1. The feature amount calculation unit further calculates, for each category, a common feature amount in the plurality of learning images in which the common image region is not removed by the common image removal unit,
2. The classifier selects whether to use the feature amount calculated from the processed image or to use the feature amount calculated from the plurality of learning images from which the common image region is not removed. 8. The image identification device according to any one of 1 to 7. An image identification method for identifying a predetermined category of an image to be given,
A common image removal step of generating a processed image obtained by removing a common image region including a predetermined common subject image from each of a plurality of learning images previously classified into the category;
A common feature amount calculating step of calculating a feature amount of the processed image generated in the common image removing step;
A classifier that identifies which category of image the input image belongs to is generated in advance by image identification processing using the feature amount calculated in the feature amount calculation step for each processed image. A target image feature amount calculation unit that calculates a feature amount of the target image and inputs it to the classifier when a category is identified for the target image after generating the classifier in the classifier generation step and the classifier generation step; An image identification method comprising steps. A feature amount calculation unit for calculating a feature amount of an input image;
A common image that is generated from each of a plurality of learning images that have been classified into categories in advance and that has been processed by removing a common image region that includes a predetermined common subject image, and that is input to the feature amount calculation unit A removal section;
Identification for generating a discriminator for identifying which category of image the input image is by image identification processing using the feature amount calculated by the feature amount calculation unit for each processed image A discriminator creating device comprising a device creating unit. A common image removal step of generating a processed image obtained by removing a common image region including a predetermined common subject image from each of a plurality of learning images classified in advance into categories;
A feature amount calculating step of calculating a feature amount of the processed image generated in the common image removing step;
A step of generating a discriminator for identifying which category of image the input image is by image discrimination processing using the feature amount calculated in the feature amount calculation step for each processed image. A classifier creating method comprising: An imaging system for capturing an image,
An imaging unit for acquiring a target image obtained by imaging the imaging target;
An image identification device that identifies which category of image the target image is determined in advance, and
The image identification device includes:
A feature amount calculation unit for calculating a feature amount of an input image;
A processed image obtained by removing a common image area including a predetermined common subject image from each of a plurality of learning images previously classified in the category is generated and input to the feature amount calculation unit An image removal unit;
A discriminator that identifies which category of image the input image belongs to is generated in advance by image discrimination processing using the feature amount calculated by the feature amount calculation unit for each processed image. And a discriminator creation unit,
The said feature-value calculation part is an imaging system which calculates the feature-value of the said target image and inputs it to the said discriminator, when identifying the category about the said target image after the said discriminator preparation part produces | generates the said discriminator.

Images