JP5882976B2 - Learning local feature generating apparatus and learning local feature generating method - Google Patents

Description

  The present invention relates to a technique for generating local feature values used for learning for image recognition.

  A technique using a local feature amount for image recognition has been proposed (Patent Document 1 below). The local feature amount is an amount obtained by quantifying local features of an image, and is usually expressed as a vector amount. By using local feature amounts, it is possible to search for similar images or specify the same image.

  By the way, the local feature amount extracted from a certain image is specialized for the shooting state of the image (for example, change in luminance, resolution, etc.). That is, even if the images represent the same content, if the shooting state of the images changes, different local feature amounts are often obtained.

Therefore, a method of extracting a large amount of local feature values for learning by performing image conversion (for example, changing brightness and resolution) on one learning image and extracting local feature values from the obtained converted image is conceivable (see FIG. For example, see Patent Document 1 below). For example, even when one type of image conversion is used, a large number of local feature amounts can be obtained by generating a large number of converted images using different parameters and then generating local feature amounts. If there are a plurality of types of image conversion, the local feature amount can be further increased. However, this method causes the following problems.
In general, since a plurality of substantially the same local feature quantities are generated as feature quantity vectors, a storage area is uselessly required.
-A large amount of computational resources and time are required for processing to generate local feature amounts from a large number of converted images after generating a large number of converted images by performing image conversion with different parameters for one type of image conversion. .

JP2010-079545

  The present invention has been made in view of the above situation. A main object of the present invention is to provide an apparatus or method that can easily and quickly generate many local features for learning.

(Item 1)
An apparatus for generating a local feature for learning using a learning image,
An image converter that converts the learning image to generate a converted image;
For each of the learning image and the converted image, a feature amount acquisition unit that acquires a local feature amount based on a local feature of the image;
A feature amount position determination unit that identifies local feature amounts whose positions on the image space are substantially the same;
Among the local feature quantities having substantially the same position in the image space, the distance in the feature quantity space is between the local feature quantity corresponding to the learning image and the local feature quantity corresponding to the converted image. An interpolation feature value generation unit that generates an interpolation feature value based on the
A learning local feature generating apparatus comprising: a learning data specifying unit that specifies a local feature for learning using at least the interpolation feature.

(Item 2)
Furthermore, it has a feature amount aggregation unit,
The feature amount aggregating unit includes a local feature amount existing within a predetermined distance in the feature amount space among the local feature amounts identified by the feature amount position determining unit and having substantially the same position in the image space. In this case, representative feature amounts are generated from these local feature amounts, and the local feature amounts within the predetermined distance are represented by the representative feature amounts.
Furthermore, when there is a representative feature amount generated by the feature amount aggregation unit, the interpolation feature amount generation unit generates the interpolation feature amount using the representative feature amount as the local feature amount. The learning local feature generating apparatus according to Item 1.

(Item 3)
Furthermore, it has a feature amount aggregation unit,
The feature amount aggregating unit, when there is a local feature amount existing within a predetermined distance in the feature amount space with respect to the interpolation feature amount generated by the interpolation feature amount generation unit, It is configured to generate representative feature values from feature values,
Further, the learning data specifying unit, when there is a representative feature amount generated in the feature amount aggregating unit, specifies the local feature amount for learning using the representative feature amount as the interpolation feature amount, The learning local feature generating apparatus according to Item 1.

(Item 4)
The local feature-value generating device for learning according to any one of items 1 to 3, wherein the interpolation feature-value generating unit is configured to generate the interpolation feature value for each predetermined distance in a feature-value space.

(Item 5)
The learning data specifying unit includes a local feature amount acquired from the learning image, a local feature amount acquired from the converted image, and a local feature amount that is substantially the same in the feature amount space for these local feature amounts. The learning local feature-value generating device according to any one of Items 1 to 4, wherein the learning local feature-value generating apparatus is further specified as a local feature-value for learning.

(Item 6)
Comprising the learning local feature generation device according to any one of items 1 to 5 and a recognition unit for recognizing a target image;
The recognition unit
An image recognition device configured to recognize the target image by searching a local feature amount corresponding to the target image from a local feature amount for learning generated by the local feature amount generation device for learning.

(Item 7)
A method for generating a local feature for learning using a device for generating a local feature for learning using a learning image,
An image conversion step of converting the learning image to generate a converted image;
A feature amount acquisition step for acquiring a local feature amount based on a local feature of the image for each of the learning image and the converted image;
A feature amount position determining step for identifying local feature amounts whose positions on the image space are substantially the same;
Among the local feature quantities having substantially the same position in the image space, the distance in the feature quantity space is between the local feature quantity corresponding to the learning image and the local feature quantity corresponding to the converted image. An interpolation feature generation step for generating an interpolation feature based on the basis;
A learning data specifying step of specifying a local feature quantity for learning using at least the interpolation feature quantity.

(Item 8)
A computer program for causing a computer to execute each step according to item 7.

  This computer program is stored in an appropriate recording medium (for example, an optical recording medium such as a CD-ROM or a DVD disk, a magnetic recording medium such as a hard disk or a flexible disk, or a magneto-optical recording medium such as an MO disk). Can be stored. This computer program can be transmitted via a communication line such as the Internet.

  According to the present invention, it is possible to quickly generate many local feature quantities for learning by a relatively simple method.

It is a block diagram for demonstrating the outline of the local feature-value generator for learning in 1st Embodiment of this invention. It is a flowchart for demonstrating the procedure which produces | generates the local feature-value for learning using the apparatus of FIG. It is explanatory drawing for demonstrating the procedure which produces | generates the local feature-value for learning using the apparatus of FIG. It is a block diagram for demonstrating the outline of the local feature-value production | generation apparatus for learning in 2nd Embodiment of this invention. It is a flowchart for demonstrating the procedure which produces | generates the local feature-value for learning using the apparatus of FIG. It is explanatory drawing for demonstrating the procedure which produces | generates the local feature-value for learning using the apparatus of FIG. It is a block diagram which shows the outline of the image recognition apparatus which concerns on 3rd Embodiment of this invention.

(Configuration of the first embodiment)
The configuration of the learning local feature generating apparatus according to the first embodiment of the present invention will be described with reference to FIG. This apparatus is for generating a local feature amount for learning using a learning image.

  The apparatus according to this embodiment includes an image conversion unit 1, a feature amount acquisition unit 2, a feature amount position determination unit 3, an interpolation feature amount generation unit 4, and a learning data specification unit 5. Furthermore, the apparatus according to the present embodiment includes an image DB 6, a feature amount DB 7, an image acquisition unit 8, and a feature amount aggregation unit 9 as additional elements.

  The image conversion unit 1 is configured to convert a learning image (described later) to generate a converted image (described later).

  The feature amount acquisition unit 2 is configured to acquire local feature amounts based on local features of the image for each of the learning image and the converted image.

  The feature amount position determination unit 3 is configured to identify local feature amounts whose positions in the image space are substantially the same. Here, the image space means a so-called pixel space or pixel coordinates. Assuming a two-dimensional image, the image space is generally expressed in the (x, y) coordinate system, but the coordinate system is not limited.

  The interpolated feature amount generation unit 4 includes a feature amount space between a local feature amount corresponding to the learning image and a local feature amount corresponding to the converted image among local feature amounts having substantially the same position in the image space. The interpolation feature amount is generated based on the above distance.

  Furthermore, when there is a representative feature amount generated by a feature amount aggregating unit 9 described later, the interpolation feature amount generation unit 4 of this example generates an interpolation feature amount using the representative feature amount as a local feature amount. It has become.

  Further, the interpolation feature value generation unit 4 of this example is configured to generate an interpolation feature value for each predetermined distance in the feature value space.

  The learning data specifying unit 5 is configured to specify a local feature amount for learning using an interpolation feature amount.

  Further, the learning data specifying unit 5 of the present example, in addition to the interpolation feature amount, the local feature amount acquired from the learning image, the local feature amount acquired from the converted image, and features for these local feature amounts The configuration is such that any or all of the local feature values that are substantially the same in the quantity space are specified as local feature values for learning.

  The image DB 6 of this embodiment is a database that records an image group to be a learning image. The learning image is, for example, an image of some product, but there are no particular restrictions on the type of image. Further, the image may be acquired by photographing a product or acquired from a network such as the Internet. That is, the image acquisition path is not particularly limited.

  The feature amount DB 7 is a database that stores local feature amounts for learning specified by the learning data specifying unit 5. The feature amount DB 7 is used in query image recognition (collation).

  The image acquisition unit 8 is a functional element that acquires a learning image from the image DB 6. Specifically, the image acquisition unit 8 is an interface that accepts input of image data.

  The feature quantity aggregating unit 9 includes local feature quantities that exist within a predetermined distance in the feature quantity space among local feature quantities that are identified by the feature quantity position determination unit 3 and that have substantially the same position in the image space. In this case, representative feature amounts are generated from these local feature amounts, and local feature amounts within a predetermined distance are represented by the representative feature amounts.

  Details of the apparatus of the present embodiment will be described in more detail in the description of operations described later.

(Operation of the first embodiment)
Next, an example of a method for generating a local feature amount for learning using the above-described generation device will be described with further reference to FIGS.

(Step SA-1 in FIG. 2)
First, the image acquisition unit 8 acquires a learning image (that is, a basic image) 101 (see FIG. 3A) from the image DB 6. Note that the learning image 101 shown in FIG. 3 is merely an example, and there are no particular restrictions on the content thereof. Further, the image acquisition unit 8 can acquire an image acquired from another place by some means, not from the image DB 6, and use it as a learning image.

  Further, in the following description, processing using one learning image 101 will be described, but it is also possible to simultaneously acquire a plurality of learning images and perform the subsequent processing in parallel.

(Step SA-2 in FIG. 2)
Next, the image conversion unit 1 converts the acquired learning image 101 to generate a converted image 102 (see FIG. 3A). Here, the image conversion is, for example, luminance change or resolution change. Of course, it is also possible to perform complex transformation such as affine transformation. However, with the method of this example, even when a relatively simple image conversion is adopted, as will be described later, a large number of learning data can be generated, and therefore the processing load in the image conversion can be reduced. It becomes. In FIG. 3A, one converted image 102 is shown. Usually, a plurality of converted images are generated by changing the conversion parameter or the type of conversion, and the subsequent procedure is performed for them. Execute. There is no particular limitation on the number of converted images to be generated.

(Step SA-3 in FIG. 2)
Next, the feature amount acquisition unit 2 acquires a local feature amount based on the local feature of the image for each of the learning image 101 and the converted image 102. Since an existing method (for example, SIFT) can be used as a local feature acquisition method, detailed description thereof will be omitted.

  Furthermore, the feature quantity acquisition unit 2 of the present embodiment acquires the position of the obtained local feature quantity in the image space. Thereby, a specific local feature-value and the position of the said feature-value on image space can be matched.

(Step SA-4 in FIG. 2)
Next, the feature quantity position determination unit 3 identifies a local feature quantity whose position in the image space is substantially the same among the above-described local feature quantities. FIG. 3A shows an example in which a plurality of local feature amounts having substantially the same position in the image space (in this example, near the tip of the left foot) are arranged in the feature amount space. Note that the feature amount is usually expressed by a multidimensional vector. However, for easy understanding, FIG. 3 shows an example in which the feature amount is arranged on a two-dimensional plane.

  In FIG. 3A, the local feature amount 1011 corresponding to the learning image (that is, the basic image) 101 is represented by white circles, and the local feature amounts 1021 to 1028 corresponding to the plurality of converted images 102 are represented by black circles. Note that one of the local feature amounts 1021 to 1028 basically corresponds to one converted image. In addition, the number of illustrated local feature amounts is merely an example, and the local feature amount corresponding to the i-th converted image is generally denoted by reference numeral 102i.

  Note that “the positions in the image space are substantially the same” means that the distance (for example, the number of pixels) between the local feature amounts in the image space is within a predetermined range (for example, within 2 pixels). This range can be determined experimentally, for example. The number of pixels described above is merely an example, and is not limited thereto.

(Step SA-5 in FIG. 2)
Next, the feature amount aggregating unit 9 determines whether there is a local feature amount existing within a predetermined distance in the feature amount space among local feature amounts (see FIG. 3A) whose positions in the image space are substantially the same. Determine if. Here, the predetermined distance is a distance between local feature amounts in the feature amount space, and can be determined experimentally, for example.

Next, when there are local feature quantities existing within a predetermined distance in the feature quantity space, the feature quantity aggregation unit 9 generates a representative feature quantity from these local feature quantities. For example, in the example of FIG. 3A, the local feature amounts 1011 and 1021 to 1023 exist within a predetermined distance. Therefore, in this example, a representative feature quantity 1001 is generated from these feature quantities (see FIG. 3B). Similarly, in the example of FIG. 3A, the local feature quantities 1024 and 1025 exist within a predetermined distance. Therefore, in this example, a representative feature quantity 1002 is generated from these feature quantities (see FIG. 3B). As a representative feature amount generation method, for example, • select any one of feature amounts within a predetermined distance; or • generate vector data of the centroid (or average) of feature amounts within a predetermined distance as a representative. The method can be considered. The illustrated example uses the latter method. In the latter method, it is also possible to select a feature amount used for calculating the center of gravity on some basis.

  By performing such an aggregation operation, the local feature amount within the predetermined distance can be represented by the representative feature amount. In addition, by performing such an aggregation operation, there is an advantage that generation of substantially the same interpolation feature amount can be restricted as will be described later.

(Step SA-6 in FIG. 2)
Next, the interpolation feature value generation unit 4 generates an interpolation feature value based on the distance in the feature value space between the local feature value corresponding to the learning image and the local feature value corresponding to the converted image. However, in this example, as described above, since there is a representative feature amount 1001 representing the local feature amount 1011 corresponding to the learning image, this representative feature amount 1001 and other local feature amounts (in this example, local features) Interpolated feature amounts are generated using feature amounts 1002, 1027, and 1028). Note that the local feature amount 1002 represents the other two local feature amounts 1024 and 1025.

  More specifically, the interpolation feature value generation unit 4 of the present example generates an interpolation feature value for each predetermined distance in the feature value space. For example, interpolation feature amounts (indicated by ▲ in FIG. 3B) 1101-1103 can be generated for each predetermined distance between the local feature amounts 1001 and 1002. This predetermined distance can also be determined experimentally, for example.

  Similarly, interpolation feature amounts 1104 to 1106 can be generated between the local feature amounts 1001 and 1027, and interpolation feature amounts 1107 to 1109 can be generated between the local feature amounts 1001 and 1028.

  Further, the feature amount aggregating unit 9 of the present embodiment determines whether or not other feature amounts exist within a predetermined distance with respect to the generated interpolation feature amount, and if present, collects those feature amounts. To do. For example, in the example of FIG. 3B, the generated interpolation feature quantity 1105 is within a predetermined distance with respect to the local feature quantity 1026 of the converted image. In this case, a representative feature quantity 1003 can be generated as a representative of these feature quantities.

(Step SA-7 in FIG. 2)
Next, the learning data specifying unit 5 specifies a local feature amount for learning using the interpolation feature amount. More specifically, the learning data specifying unit of the present embodiment
The generated interpolation feature amount, the local feature amount of the converted image, and the local feature amount of the learning image are specified as learning data.

  However, in this example, when there is almost the same local feature quantity in the feature quantity space with respect to these local feature quantities (that is, when a representative feature quantity exists), instead of these feature quantities, a representative feature quantity is used. The feature amount is specified as a local feature amount for learning.

  Further, the learning data specifying unit 5 stores the specified feature quantity in the feature quantity DB 7 as learning data.

  If there is a local feature amount corresponding to another position on the image coordinates, the learning data can be similarly generated by repeating the processing from step SA-4. When other learning images exist, learning data can be generated by repeating the processing from step SA-1.

  In the present embodiment, local feature amounts at substantially the same position on the image coordinates are selected, and learning feature amounts are generated by interpolation between them in the feature amount space. The feature amount generation process is simple, and the amount of calculation resources used can be kept low. That is, if a large number of converted images are generated and a process of extracting local feature amounts from the converted images one by one is performed, the processing requires a lot of calculation resources and time. On the other hand, according to the above-described method, there is an advantage that many local feature quantities for learning can be generated even when a small number of converted images are used.

  Furthermore, in the present embodiment, the interpolation feature amount is generated between the feature amount corresponding to the learning image (that is, the basic image) and the feature amount corresponding to the converted image. Quality can be improved. If an interpolated feature amount is generated between feature amounts corresponding to the converted image, the interpolated feature amount has a low relevance to the learning image, and thus is less useful when collating query images in the future. It seems to be easy to become. Therefore, in the present embodiment, by generating an interpolated feature amount between the feature amount corresponding to the learning image and the feature amount corresponding to the converted image, it is possible to contribute to the improvement of the matching accuracy of the query image. . That is, the method of the present embodiment has an advantage that a high-quality learning local feature can be efficiently generated.

  Further, in this embodiment, representative feature amounts are generated by collecting feature amounts that can be regarded as substantially the same in the feature amount space (that is, within a predetermined distance), and an interpolation feature amount is generated using the representative feature amount. Yes. Here, if an interpolated feature amount is generated between feature amounts without generating a representative feature amount, a large number of substantially the same interpolated feature amounts may be generated. In this case, a storage area is wasted. Become. On the other hand, in this embodiment, since the interpolation feature value is generated using the representative feature value, it is possible to reduce the possibility that the generated interpolation feature value is substantially the same, and to effectively use the storage area. Is possible.

  Therefore, according to the present embodiment, there is an advantage that effective learning data can be generated quickly using a small number of computing resources.

(Second Embodiment)
Next, a learning local feature generating apparatus according to a second embodiment of the present invention will be described with reference to FIG. In the description of the second embodiment, functional elements that are basically the same as those of the first embodiment described above are denoted by the same reference numerals, thereby avoiding complicated description.

  In the first embodiment described above, the feature quantity aggregating unit 9 is arranged upstream of the interpolation feature quantity generation unit 4 (see FIG. 1), and after the feature quantities are aggregated (step SA-5 in FIG. 2), the interpolated feature quantity is obtained. Was generated. On the other hand, in the second embodiment, the feature quantity aggregating unit 9 is arranged on the downstream side of the interpolation feature quantity generation unit 4 (see FIG. 4) and aggregates the feature quantities after generating the interpolation feature quantity.

  That is, the feature amount aggregating unit 9 of the present embodiment, when there are local feature amounts existing within a predetermined distance in the feature amount space with respect to the interpolation feature amount generated by the interpolation feature amount generating unit 4, The representative feature amount is generated from the interpolation feature amount and the local feature amount. Further, when there is a representative feature amount generated by the feature amount aggregating unit 9, the learning data specifying unit 5 of this example specifies a local feature amount for learning using the representative feature amount as an interpolation feature amount. It has become.

  With reference to FIG. 5, a method for generating a local feature amount for learning in the second embodiment will be described.

(Steps SB-1 to SB-4 in FIG. 5)
These steps are basically the same as steps SA-1 to SA-4 in FIG. An example of the generated feature amount space is shown in FIG.

(Step SB-5 in FIG. 5)
Next, in the second embodiment, the interpolation feature value generation unit 4 generates an interpolation feature value as in the first embodiment. However, in the second embodiment, since the interpolation feature amount is generated before the representative feature amount is generated, a large number of interpolation feature amounts having close distances are generated as shown in FIG. In FIG. 6B as well, the generated interpolation feature amount is indicated by ▲.

(Step SB-6 in FIG. 5)
Next, in the second embodiment, the feature quantity aggregating unit 9 aggregates feature quantities. The generation of the representative feature amounts 1001 to 1003 (see FIG. 6B) is the same as in the first embodiment described above. Furthermore, in the second embodiment, the generated interpolation feature value may be close to another interpolation feature value (see the feature value surrounded by a two-dot chain line in FIG. 6B). In that case, these feature quantities are aggregated to generate a representative feature quantity (not shown). The representative feature amount can be used in place of the aggregated feature amount.

  Since the subsequent processing is the same as that of the first embodiment, detailed description thereof is omitted.

  In addition, since the configuration and advantages other than those described above in the second embodiment are the same as those in the first embodiment described above, detailed description thereof is omitted.

(Third embodiment)
Next, an image recognition apparatus according to a third embodiment of the present invention will be described with reference to FIG. In the description of the third embodiment, functional elements that are basically the same as those of the first embodiment described above are denoted by the same reference numerals, thereby avoiding complicated description.

  The image recognition apparatus according to the third embodiment includes a learning unit 100 and a recognition unit 200.

  The learning unit 100 is a functional element corresponding to the learning local feature generating device described in the first embodiment or the second embodiment.

  The recognition unit 200 is a functional element for recognizing a target image (so-called query image). The recognition unit 200 according to the present embodiment is configured to recognize a target image by searching for a local feature amount corresponding to the target image from a learning local feature amount (learning data) generated by the learning unit 100. ing. Since the configuration of such a recognition unit 200 can be the same as that of an existing recognition device, detailed description thereof is omitted.

  The operations of the above-described embodiments can be implemented by incorporating appropriate computer software into the computer.

  The contents of the present invention are not limited to the above embodiment. In the present invention, various modifications can be made to the specific configuration within the scope of the claims.

  For example, each component described above may exist as a functional block, and may not exist as independent hardware. As a mounting method, hardware or computer software may be used. Furthermore, one functional element in the present invention may be realized by a set of a plurality of functional elements, and a plurality of functional elements in the present invention may be realized by one functional element.

  Moreover, the functional element may be arrange | positioned in the position physically separated. In this case, the functional elements may be connected by a network. It is also possible to realize functions or configure functional elements by grid computing or cloud computing.

DESCRIPTION OF SYMBOLS 1 Image conversion part 2 Feature-value acquisition part 3 Feature-value position determination part 4 Interpolated feature-value production | generation part 5 Learning data specification part 6 Image DB
7 Feature DB
8 Image acquisition unit 9 Feature amount aggregation unit 100 Learning unit (local feature amount generation device for learning)
200 recognition unit 101 learning image 102 converted image 1001 to 1003 representative feature amount (local feature amount)
1011 Local feature amount based on learning image 1021 to 1028 Local feature amount based on converted image 1101 to 1109 Generated interpolation feature amount

Claims (9)

An apparatus for generating a local feature for learning using a learning image,
An image converter that converts the learning image to generate a converted image;
For each of the learning image and the converted image, a feature amount acquisition unit that acquires a local feature amount based on a local feature of the image;
From among the feature acquiring the local feature amount acquired by unit, a feature amount position determining unit that identifies a local feature amount which is within the distance specified range in the image space,
Of distance on the image space of the local feature amount is within a predetermined range, and the local feature amount corresponding to the learning image, between the local feature amount corresponding to the transformed image, on the feature amount space An interpolation feature generating unit that generates an interpolation feature based on the distance;
A learning local feature generating apparatus comprising: a learning data specifying unit that specifies a local feature for learning using at least the interpolation feature. The feature amount acquisition unit is configured to further acquire a position on the image space with respect to the local feature amount acquired by the feature amount acquisition unit.
  The learning local feature generating apparatus according to claim 1. Furthermore, it has a feature amount aggregation unit,
The feature amount aggregating unit, the identified by the feature position determining unit, among distances in image space is local feature quantity is within a predetermined range, local features that are present within a predetermined distance in the feature quantity space If there is, a representative feature amount is generated from these local feature amounts, and the representative feature amount is configured to represent the local feature amount within the predetermined distance in the feature amount space ,
Furthermore, when there is a representative feature amount generated by the feature amount aggregation unit, the interpolation feature amount generation unit generates the interpolation feature amount using the representative feature amount as the local feature amount. The learning local feature generating apparatus according to claim 1 or 2 . Furthermore, it has a feature amount aggregation unit,
The feature amount aggregating unit, when there is a local feature amount existing within a predetermined distance in the feature amount space with respect to the interpolation feature amount generated by the interpolation feature amount generation unit, It is configured to generate representative feature values from feature values,
Further, the learning data specifying unit, when there is a representative feature amount generated in the feature amount aggregating unit, specifies the local feature amount for learning using the representative feature amount as the interpolation feature amount, The learning local feature generating apparatus according to claim 1 or 2 . The interpolated feature value generating unit, learning local feature amount generating apparatus according to any one of interpolation claims feature amount has a configuration which generates 1 to 4 for each predetermined distance in the feature quantity space . The learning data specifying unit includes a local feature amount acquired from the learning image, a local feature amount acquired from the converted image, and a distance on the feature amount space with respect to these local feature amounts within a predetermined range. The local feature-value generator for learning according to any one of claims 1 to 5 , wherein the local feature-value is further specified as a local feature-value for learning. The learning local feature generation device according to any one of claims 1 to 6 , and a recognition unit for recognizing a target image,
The recognition unit
An image recognition device configured to recognize the target image by searching a local feature amount corresponding to the target image from a local feature amount for learning generated by the local feature amount generation device for learning. A method for generating a local feature for learning using a device for generating a local feature for learning using a learning image,
An image conversion step of converting the learning image to generate a converted image;
A feature amount acquisition step for acquiring a local feature amount based on a local feature of the image for each of the learning image and the converted image;
From among the local feature amount obtained in the feature amount obtaining step, the feature amount position determination step of specifying a local feature amount which is within the distance specified range in the image space,
Of distance on the image space of the local feature amount is within a predetermined range, and the local feature amount corresponding to the learning image, between the local feature amount corresponding to the transformed image, on the feature amount space An interpolation feature generation step for generating an interpolation feature based on the distance;
A learning data specifying step of specifying a local feature quantity for learning using at least the interpolation feature quantity. A computer program for causing a computer to execute each step according to claim 8 .

Images