JP2019049909A - Generation device, generation method, and generation program - Google Patents

Abstract

To generate centroid information that enables efficient search.SOLUTION: A generation device according to the present application comprises an acquisition part and a generation part. The acquisition part acquires correspondence information indicating correspondence between a first centroid with which a vector for a predetermined object is associated according to a distance and the vector associated with the first centroid. The generation part generates a plurality of second centroids with which each of object vectors included in a first vector group is associated according to the distance on the basis of the correspondence information acquired by the acquisition part, when information on the first vector group associated with the first centroid satisfies a predetermined condition.SELECTED DRAWING: Figure 3

Description

  The present invention relates to a generation device, a generation method, and a generation program.

  Conventionally, techniques relating to search of vector data by vector quantization have been provided. For example, techniques are provided for quantizing vectors with a codebook. Moreover, such a technique is used, for example, for image search.

JP 2011-257970 A

Equal frequency binning, IBM, [online], [search on September 4, 2017], Internet <URL: http://www.ibm.com/support/knowledgecenter/SSWLVY_1.0.0/com.ibm.spss.analyticcatalyst .help / analytic_catalyst / equal_frequency_binning.html> Iwasaki Masajiro "Nearby Search by Approximate k-Nearest Neighbor Graph Using Tree-structured Index", Transactions of Information Processing Society of Japan, 2011/2, Vol. 52, No. 2. pp.817-828.

  However, in the above-described prior art, it may be difficult to generate centroid information that enables efficient search. For example, only by quantizing a vector by a codebook, the processing time required for searching for a vector increases when the number of vectors increases, and it is difficult to suppress the processing time for searching. Further, for example, when many vectors are associated with one centroid, it may take time to determine which vector should be the search result. In such a case, it may be difficult to search efficiently for a query.

  The present application has been made in view of the above, and an object of the present application is to provide a generation device, a generation method, and a generation program for generating centroid information that enables efficient search.

  The generation device according to the present application acquires an acquisition unit that acquires correspondence information indicating correspondence between a first centroid in which a vector related to a predetermined target is associated according to a distance, and a vector associated with the first centroid. A vector included in the first vector group when information on a first vector group associated with the first centroid satisfies a predetermined condition based on the correspondence information acquired by the acquisition unit; And a generator configured to generate a plurality of second centroids in which each of the target vectors is associated according to the distance.

  According to an aspect of the embodiment, it is possible to generate centroid information that enables efficient search.

FIG. 1 is a diagram illustrating an example of a generation process according to the embodiment. FIG. 2 is a diagram illustrating an exemplary configuration of a generation system according to the embodiment. FIG. 3 is a diagram illustrating an example of the configuration of the generation device according to the embodiment. FIG. 4 is a diagram showing an example of the first index information storage unit according to the embodiment. FIG. 5 is a diagram illustrating an example of a centroid information storage unit according to the embodiment. FIG. 6 is a diagram illustrating an example of a second index information storage unit according to the embodiment. FIG. 7 is a flowchart illustrating an example of the generation process according to the embodiment. FIG. 8 is a diagram illustrating an example of provision of information corresponding to a query according to the embodiment. FIG. 9 is a schematic diagram of first index information according to the embodiment. FIG. 10 is a diagram illustrating an example of a process of generating second index information according to the embodiment. FIG. 11 is a diagram illustrating an example of use of second index information according to the embodiment. FIG. 12 is a diagram illustrating an example of vector division according to the embodiment. FIG. 13 is a hardware configuration diagram illustrating an example of a computer that implements the function of the generation device.

  Hereinafter, a generation apparatus, a generation method, and a mode for carrying out a generation program (hereinafter, referred to as “embodiment”) according to the present application will be described in detail with reference to the drawings. Note that the generation device, the generation method, and the generation program according to the present application are not limited by this embodiment. Moreover, the same code | symbol is attached | subjected to the same site | part in the following each embodiment, and the overlapping description is abbreviate | omitted.

(Embodiment)
[1. Generation process]
An example of the generation processing according to the embodiment will be described with reference to FIG. FIG. 1 is a diagram illustrating an example of a generation process according to the embodiment. FIG. 1 shows a case where the generating apparatus 100 (see FIG. 3) generates a plurality of second centroids according to the number of vectors associated with the first centroid. That is, FIG. 1 shows a case where the generating apparatus 100 generates a plurality of second centroids for dividing and correlating the vector group associated with the first centroid. The target information may be any information as long as it can be expressed as a vector. In addition, although the vector information which made image information the object below is demonstrated, the object of vector information may be other objects, such as moving image information and audio information.

  Although the details of the centroid will be described later, the first centroid and the second centroid are relative concepts, and the first centroid is a target to which the associated vector group is divided. Lloyd is shown, and the second centroid is a centroid to be divided and associated with the vector group associated with the first centroid. That is, a second centroid may be a centroid for one centroid and a first centroid for another centroid. Also, in the following, when the vector group associated with the second centroid generated from the first centroid is further divided, the centroid to which the divided vector (target vector) is associated is referred to as a third centroid. There is a case. In this case, when the second centroid is taken as the first centroid, the third centroid becomes the second centroid. In the following, the first to third centroids may be described simply as "centroids".

  Although the example of FIG. 1 targets vector information corresponding to image information of millions to hundreds of millions, only a part of the vector information is illustrated in the drawing. Further, in the following, for example, when indicating vector data itself corresponding to image information, it is described as “vector data”, and when indicating information identifying each vector data corresponding to image information, “vector identification information” or “vector identification information” Describe as "vector ID". That is, vector data here means N-dimensional high-dimensional vector, and vector identification information means information for identifying each vector such as vector ID. For example, in the example shown in FIG. 1, when described as vector identification information V112, it means the vector ID "V112". Also, for example, when the vector V112 is described, it means the vector identified by the vector ID "V112". Further, in the case of describing without distinguishing whether it is vector data or vector identification information, it is described as “vector” or “vector information”.

  Further, in the following, when “vector V * (* is an arbitrary numerical value)” is described, it indicates that the vector is a vector identified by vector ID “V *”. For example, when describing as "vector V1", the vector is a vector identified by vector ID "V1". Moreover, when it describes as "centroid C * (* is arbitrary numerical value)", it shows that the centroid is a centroid identified by centroid ID "C *." For example, when "centroid C4" is described, the centroid is a centroid (vector) identified by centroid ID "C4".

  First, the generating apparatus 100 acquires correspondence information indicating correspondence with a vector associated with a centroid (step S11). The example of FIG. 1 shows a case where the generating apparatus 100 generates correspondence information indicating correspondence using vector information respectively corresponding to image information. For example, the generating device 100 acquires vector information corresponding to image information from an external information processing device such as the information providing device 50. In the example of FIG. 1, the generating apparatus 100 acquires a plurality of pieces of vector information identified by the vector identification information V3 to V1100 and the like as indicated by the space information GR11 and the like. For example, the generating device 100 acquires a large number of pieces of vector information including the vector identification information V3 to V1100 and the like.

  In the example shown in FIG. 1, each acquired vector information is expressed by attaching vector identification information to “o”. For example, the vector identified by the vector ID "V55" is expressed as a lower right "o" point in the space information GR11. For example, in the example shown in FIG. 1, each vector data is an N-dimensional real-valued vector. Space information GR11-1 to GR11-4 shown in FIG. 1 and FIG. 9 etc. are diagrams schematically showing a part of the space information, and the space information GR11-1 to GR11-4 is generated by the generation process. Information corresponding to the information to be Further, in the following, the space information GR11-1 to GR11-4 will be described as space information GR11, in the case of describing the information without particular distinction.

  The space information GR11 in FIG. 1 may be Euclidean space. Also, space information GR11 shown in FIG. 1 is a conceptual diagram for explaining the distance between each vector and the like. It becomes a multidimensional space. For example, although the space information GR11 shown in FIG. 1 is illustrated in a two-dimensional manner for illustration on a plane, it is assumed to be a multidimensional space such as a 100-dimensional or 1000-dimensional space, for example.

  In the present embodiment, the distance between the vectors in the space information GR11 is the degree of similarity between the corresponding images. Here, in the example shown in FIG. 1, the similarity between images in which the distance between each vector in the space information GR11 is short is high, and the similarity between images in which the distance between each vector in the space information GR11 is large is low. For example, in the space information GR11 in FIG. 1, the vector information identified by the vector ID "V9" is close to the vector information identified by the vector ID "V34", that is, the distance is small. Therefore, it indicates that the image corresponding to the vector information identified by the vector ID “V9” and the image corresponding to the vector information identified by the vector ID “V34” have high similarity. Also, for example, in the space information GR11 in FIG. 1, the vector information identified by the vector ID "V9" and the vector information identified by the vector ID "V55" are remote, ie, the distance is large. Therefore, it is indicated that the image corresponding to the vector information identified by the vector ID “V9” and the image corresponding to the vector information identified by the vector ID “V55” have low similarity. Note that the index indicating the degree of similarity may be any index as long as it is applicable to the generation processing of the present application, and may be an index targeting distance, direction, and the like. For example, as the index indicating the degree of similarity, various indexes such as Euclidean distance, Mahalanobis distance, and cosine distance may be used if applicable to the generation processing of the present application.

  In the example of FIG. 1, the generating apparatus 100 clusters a plurality of vectors by a predetermined clustering method. For example, the generating apparatus 100 clusters a plurality of vectors by a predetermined clustering method so that the difference in the number of vectors associated with the centroid becomes uniform. For example, the generating apparatus 100 clusters a plurality of vectors by a predetermined clustering method such that the difference in the number of vectors associated with the centroid is within a predetermined value. A plurality of vectors may be clustered using various conventional techniques such as the k-means method as appropriate. For example, the generation device 100 sets a predetermined number of clusters, and by performing constrained clustering in which the number of classified vectors of each cluster is limited, the difference in the number of vectors associated with the centroid becomes within a predetermined value. You may cluster several vectors in.

  Thus, the generating device 100 generates a plurality of pieces of centroid information such that the difference in the number of vectors associated with each centroid is within a predetermined value. For example, the generating apparatus 100 generates a plurality of pieces of centroid information such that the number of vectors associated with each centroid is equal. In the following, in order to explain this point, the number of clusters is 100, ie, the generated centroids are 100 centroids including centroids C1 to C11 and the like, and correspond to each of centroids C1 to C11 and the like. It is assumed that the number of vectors included in the areas AR1 to AR11 and the like is 25 (pieces). In the space information GR11 of FIG. 1, only a part of a centroid or a vector is illustrated. Further, the number of centroids may be selected (determined) based on the number of vectors, or may be selected (determined) based on a predetermined set value.

  Also, the above generation method is an example, and the generation device 100 can generate a plurality of pieces of centroid information so that the difference in the number of vectors corresponding to each centroid is within a predetermined value. Information and techniques may be used to generate centroid information. For example, the generating apparatus 100 may generate a plurality of pieces of centroid information such that the number of vectors associated with each centroid becomes equal by appropriately using or combining various conventional techniques and the like.

  For example, the generating apparatus 100 may generate a plurality of pieces of centroid information such that the difference in the number of vector identification information is within a predetermined value by performing random sampling. For example, the generating apparatus 100 performs random sampling to select a centroid, and associates each vector based on the selected centroid, so that the difference in the number of vectors associated with each centroid is within a predetermined value. And so on may be generated. In addition, as for the processing of clustering before generating the second centroid, the generating apparatus 100 performs a plurality of centroids by any method as long as the difference in the number of vectors associated with each centroid becomes uniform. Information may be generated. For example, the generation device 100 generates a plurality of pieces of centroid information so that the difference in the number of vectors associated with each centroid becomes uniform, using a conventional technique related to hashing such as LSH (Locality Sensitive Hashing). It is also good. In this case, for example, the generating apparatus 100 divides each vector into clusters (groups) based on the hash value of each vector calculated by LSH or the like, and performs a cent for each cluster into which the vector is divided based on the hash value. Lloyd information may be generated. Also, for example, the generation device 100 appropriately uses various conventional techniques such as the technique of Equal frequency binning as disclosed in Non-Patent Document 1 so that the number of vectors associated with each centroid becomes equal. Multiple pieces of centroid information may be generated.

  In the example of FIG. 1, the generating device 100 centroids vector identification information C1, C2, C3, C4, C5, C6, C7, C8, C9, C10, and C11, as indicated by the space information GR11 in FIG. Generate centroid information. In FIG. 1, although five vectors are illustrated in each of the areas AR1 to AR11 and the like corresponding to each centroid, it is assumed that 99 vectors are located in each of the areas AR1 to AR11 and the like. As described above, the generating apparatus 100 is configured such that the vectors associated with each of the vector identification information C1, C2, C3, C4, C5, C6, C7, C8, C9, C10, and C11 that become centroids become equal. Multiple centroid information may be generated.

  A boundary LN shown in space information GR11 in FIG. 1 is a virtual line for indicating which vector is associated with which centroid. For example, vectors (vectors V33, V66, etc.) located in the boundary LN surrounding the area AR2 indicate that they are associated with the centroid C2. In the example in FIG. 1, the generating apparatus 100 divides the space information GR11 into areas AR1 to AR11 and the like by the boundary LN. For example, the area AR1 in the space information GR11 in FIG. 1 includes a vector identified by each of the vector identification information V4, V25, V32, V41, V87 and the like. In the example of FIG. 1, it is shown that the vector identified by each of the vector identification information V4, V25, V32, V41, V87, etc. is associated with the centroid C1. That is, the generating device 100 associates the centroid C1 with the vector identification information V4, V25, V32, V41, V87, and the like. For example, when there is an image corresponding to the centroid C1, an image corresponding to the vector identified by each of the vector identification information V4, V25, V32, V41, V87, etc. is identified by the centroid C1 among the respective centroids. It is similar to the image corresponding to the centroid.

  In the example of FIG. 1, the generating device 100 generates a plurality of pieces of centroid information as shown in the centroid information storage unit 122-1. For example, the generating apparatus 100 stores, in the centroid information storage unit 122-1, the centroid information including the centroid C1 corresponding to the area AR1 in the spatial information GR11 and the vector information thereof. Specifically, the generating device 100 generates centroid information combining the centroid ID “C1” and vector information corresponding to the position of the centroid in the area AR1 in the spatial information GR11, and stores the centroid information storage. It stores in the part 122-1. The position of the centroid corresponding to the area may be the center of the area. For example, the position of the centroid corresponding to the area may be the center of gravity or the center of mass of the area. In the example of FIG. 1, the generating device 100 stores the vector data “10, 24, 54, 2...” Of the centroid C1 in the centroid information storage unit 122-1.

  For example, the generating device 100 stores the centroid ID identifying the centroid corresponding to the area AR2 in the spatial information GR11 and the centroid information storage unit 122-1 as centroid information including vector information corresponding to the centroid. . Specifically, the generation device 100 is a centroid ID “C2”, vector data corresponding to the centroid C2, and N-dimensional vector information “22, 48, 5, 63. Are generated in combination with each other and stored in the centroid information storage unit 122-1.

  As described above, in FIG. 1, the centroid information storage unit 122 is described as the centroid information storage units 122-1 and 122-2 according to the update of the centroid information. That is, FIG. 1 shows that the centroid information storage unit 122-1 is updated to the centroid information storage unit 122-2. The centroid information storage units 122-1 and 122-2 are the same centroid information storage unit 122. Further, hereinafter, the centroid information storage units 122-1 and 122-2 will be described as the centroid information storage unit 122 in the case of describing the centroid information storage units 122-1 and 122-2 without particular distinction.

  Then, the generating apparatus 100 generates second index information based on the vector associated with each centroid. For example, the generating apparatus 100 generates second index information in which information identifying a vector included in each of the areas AR1 to AR11 and the like corresponding to each centroid is associated with a centroid ID.

  For example, the generating device 100 stores the vector identification information V4, V25, V32, V41, V87, and the like in the second index information storage unit 123-1 in association with the centroid C1. For example, the generating device 100 stores the vector identification information vectors V33, V66, V81, V112, V971 and the like in the second index information storage unit 123-1 in association with the centroid C2.

  Hereinafter, the second index information storage unit 123 will be described as the second index information storage units 123-1 and 123-2 according to the update of the second index. That is, FIG. 1 shows that the second index information storage unit 123-2 is updated from the second index information storage unit 123-1. The second index information storage units 123-1 and 123-2 are the same second index information storage unit 123. Further, hereinafter, the second index information storage unit 123 is described as the second index information storage unit 123 when the second index information storage units 123-1 and 123-2 are described without particular distinction.

  As described above, the generation apparatus 100 speeds up the search for a vector similar to that in the case where a centroid is specified by generating so-called inverted index information in which information identifying a vector is associated with a centroid. be able to. Note that when searching for similar vectors, a plurality of nearby centroids may be specified.

  The generating device 100 may also acquire vector information clustered as shown in the space information GR11-1 of FIG. 1 from an external information processing device. For example, the generating device 100 may acquire information as shown in the centroid information storage unit 122-1 and the second index information storage unit 123-1 in FIG. 1 from an external information processing device.

  From here, the case where the vector increases will be described. For example, when the number of vectors associated with one first centroid reaches a predetermined threshold value, the generation device 100 newly generates a predetermined number of second centroids, and generates the second centroid as the second centroid. The vector group associated with the first centroid is divided and associated. In the example of FIG. 1, when the number of vectors (target vectors) associated with one first centroid reaches 100, the generating apparatus 100 generates four second centroids, and target vectors correspond to each other. The case where the attached centroid is divided into four second centroids is shown. That is, in the example of FIG. 1, when the number of vectors (target vectors) associated with one first centroid reaches 100, the generating device 100 sets the number of vectors associated with each centroid. A case is shown in which one hundred object vectors are divided so as to be approximately associated with four second centroids on an average of about 25 so as to be 25 ". That is, the generating apparatus 100 generates four second centroids based on the number "4" obtained by dividing the target vector number "100" by the set number "25". Note that the generation apparatus 100 may generate an appropriate number (two or more) of second centroids based on the number of target vectors, the number of settings, and the like without being limited to four.

  In the example of FIG. 1, the generating device 100 acquires information indicating an increase of a vector (step S12). For example, the generating device 100 acquires vector information corresponding to the image added to the information providing device 50. In the example of FIG. 1, it is assumed that 75 images are newly added to the information providing apparatus 50. For example, the generating device 100 acquires vector identification information V5001 to V5075 corresponding to each of the 75 images.

  In the example of FIG. 1, for the sake of simplicity, it is assumed that all of the added vectors V5001 to V5075 are located in the area AR2, that is, vectors close to the centroid C2. . Therefore, in the example of FIG. 1, the generating device 100 associates all of the added vectors V5001 to V5075 with the centroid C2.

  Thus, 100 vectors such as vector identification information V33, V66, V81, V112, V971, and V5001 to V5075 are associated with the centroid C2. Therefore, in the example of FIG. 1, as indicated by the space information GR11-2, 100 vectors of vector identification information V33, V66, V81, V112, V971, V5001 to V5075, etc. are located in the area AR2. . In the space information GR11-2 in FIG. 1, only a part of 100 vectors is illustrated.

  In the example of FIG. 1, since 100 vectors are associated with the centroid C2, the generating apparatus 100 determines that the threshold value is reached and generates a centroid (step S13). For example, the generating apparatus 100 generates four second centroids corresponding to the first centroid, the centroid C2. For example, the generating device 100 performs clustering so as to divide 100 vectors associated with the centroid C2 into four, thereby forming four second centroids, centroids C2-1 to C2-4. Generate For example, the generating apparatus 100 generates four second centroids C2-1 to C2-4 by performing clustering to divide the area AR2 into four areas.

  In the example of FIG. 1, the generating apparatus 100 clusters 100 vectors in the area AR2 by a predetermined clustering method. For example, the generating apparatus 100 clusters a plurality of vectors by a predetermined clustering method so that the difference in the number of vectors associated with the four second centroids becomes uniform. For example, the generating apparatus 100 clusters a plurality of vectors so that the difference in the number of vectors associated with the four second centroids is uniform. The number of vectors associated with the centroids C2-1 to C2-4 is not limited to the case of being completely uniform (25), but may be from 25 to within a predetermined range (20 to 30), etc. May be

  Thus, the generating apparatus 100 generates centroids C2-1 to C2-4 such that the difference in the number of vectors associated with each centroid is as small as possible. For example, the area AR2-1 in the space information GR11-3 in FIG. 1 includes vector identification information V66, V5001, and the like. That is, the generating device 100 associates the centroid C2-1 with the vectors V66, V5001, and the like.

  Then, the generating apparatus 100 updates the information according to the generation of the second centroid (step S14). In the example of FIG. 1, the generating device 100 updates the information in response to the generation of the second centroid. For example, when the generation device 100 generates a plurality of second centroids, the generation device 100 associates the plurality of second centroids with the first centroid, and associates each vector included in the first vector group with the first centroid. The information in the storage unit 120 (see FIG. 3) is updated to release the attachment.

  In the example of FIG. 1, the generating device 100 updates the centroid information storage unit 122 with the information corresponding to the plurality of generated second centroids. For example, the generating apparatus 100 stores, in the centroid information storage unit 122-2, centroid C2-1 including the centroid C2-1 corresponding to the area AR2-1 in the spatial information GR11 and the vector information thereof. Specifically, the generation device 100 generates centroid information by combining the centroid ID “C2-1” and vector information corresponding to the position of the centroid of the area AR2-1 in the spatial information GR11. It stores in the centroid information storage unit 122-2. In the example of FIG. 1, the generating apparatus 100 stores the vector data “21, 40, 11, 52...” Of the centroid C2-1 in the centroid information storage unit 122-2.

  For example, the generating apparatus 100 may use the centroid information storage unit 122-2 as centroid information including a centroid ID for identifying a centroid corresponding to the area AR2-2 in the spatial information GR11 and vector information corresponding to the centroid. Store. Specifically, the generating device 100 is an N-dimensional vector that corresponds to the centroid ID “C2-2” and the centroid C2-2 and is “19, 51, 9, 65. And centroid information in combination with each other are generated, and stored in the centroid information storage unit 122-2. Similarly, for the remaining two second centroids, centroids C2-3 and C2-4, centroid information is stored in the centroid information storage unit 122-2. Note that the generating device 100 may delete the information of the centroid C2 stored in the centroid information storage unit 122-2, or that the direct association with the vector is canceled using a predetermined flag. May be determined.

  Further, in the example of FIG. 1, the generating device 100 updates the second index information storage unit 123 with the information corresponding to the plurality of generated second centroids. For example, the generating device 100 updates the second index information storage unit 123 based on the vectors associated with the respective centroids C2-1 to C2-4. For example, the generating apparatus 100 corresponds to the centroids C2-1 to C2-4 with information identifying the vector included in each of the areas AR2-1 to AR2-4 corresponding to the respective centroids C2-1 to C2-4. Generate the attached second index information.

  For example, the generating device 100 stores the vector identification information of 25 vectors in the second index information storage unit 123-2 in association with the centroid C2-1. For example, the generating apparatus 100 stores the vector identification information V66, V5001, V5004 and the like in the second index information storage unit 123-2 in association with the centroid C2-1. For example, the generating device 100 stores the vector identification information of 25 vectors in the second index information storage unit 123-2 in association with the centroid C2-2. For example, the generating device 100 stores the vector identification information vectors V971 and V5003 and the like in the second index information storage unit 123-2 in association with the centroid C2-2.

  In addition, as illustrated in the second index information storage unit 123-2, the generating device 100 deletes the second index information corresponding to the centroid C2 to release the association between the centroid C2 and the vector. Note that the generating device 100 sets the predetermined flag to the information corresponding to the centroid C2 in the second index information storage unit 123, thereby disengaging the direct correspondence between the centroid C2 and the vector. It may be distinguishable. Further, the centroid C2 whose direct association with the vector has been released is used as first index information at the time of search, and the details will be described in detail in FIG.

  As described above, when the number of vectors associated with one first centroid (centroid C2) increases, the generation device 100 divides the plurality of second centroids so as to divide the association. By generating, it is possible to generate centroid information that enables efficient search. For example, when the number of vectors associated with a centroid is large, a vector that is not a neighbor may result. Therefore, at the time of searching using centroid information, the generating apparatus 100 adjusts the centroid according to the change in registration of the image so that the number of vectors associated with each centroid becomes as even as possible. Can improve the performance of In particular, in the case of the k nearest neighbor search, since the number of search results is constant, the generating device 100 has the effect of enabling more efficient search by equalizing the number of vectors associated with the centroid. Can be played.

  More specifically, when the vector associated with one first centroid satisfies the predetermined condition as described above, the generating device 100 divides the plurality of second centroids so as to divide the association. By generating, it becomes possible to adjust so that the difference in the number of vectors associated with each centroid is within a predetermined value. As described above, the generating apparatus 100 is efficient by generating a plurality of second centroids for the first centroid satisfying the predetermined condition so that the number of vectors associated with each centroid becomes equal. It is possible to generate centroid information that enables various searches. For example, since the centroid corresponding to the query can be specified by the centroid information generated by the generation device 100, efficient vector search in neighborhood search can be performed by using the vector associated with the centroid as the search result. It can make search possible. In addition, the generation apparatus 100 performs generation processing of other information besides the above-described processing, and performs information provision for a query using the generated information. Details of these points will be described later.

  In addition, as described above, the generation device 100 stores only vector data of those that become centroids among all vector data in the storage unit 120 (see FIG. 3) such as the centroid information storage unit 122, and other vectors Data is not stored in the storage unit 120. In other words, the generating apparatus 100 stores only the vector identification information in the second index information storage unit 123 for vectors other than centroids. As a result, the generating apparatus 100 can generate centroid information that enables efficient search while suppressing the data amount of the vector information to be stored.

[1-1. Re-merging]
In the example described above, the case of generating the second centroid is shown, but the generating device 100 may delete the generated second centroid. That is, the generation device 100 may remerge the divided regions by generating the second centroid. For example, when an image is excluded from the search target due to deletion of an image or the like, the generation device 100 performs the following remerging process when the number of vectors associated with each centroid decreases. May be

  When the number of vectors associated with each of the plurality of second centroids is equal to or less than a predetermined threshold, the generating device 100 sets the vector associated with each of the plurality of second centroids as the first centroid. The association may be released between the first centroid and the plurality of second centroids.

  For example, if the number of vectors corresponding to each of the generated second centroids C2-1 to C2-4 is equal to or less than a predetermined threshold (for example, six), the generating device 100 may generate the centroid. The vectors associated with each of C2-1 to C2-4 may be associated with the centroid C2. For example, if the total number of vectors corresponding to each of the centroids C2-1 to C2-4 is equal to or less than a predetermined threshold (for example, 25), the generation device 100 may generate the centroids C2-1 to C2-4. The vector associated with each of may be associated with the centroid C2.

  In addition, the generating device 100 may delete the plurality of generated second centroids. For example, the generating apparatus 100 may delete the plurality of generated second centroids when re-merging the area as described above. In the example of FIG. 1, the generating device 100 directly associates the centroid C2 with the vector and performs the second index information storage unit 123 on the centroids C2-1 to C2-4 according to the re-merging process. You may delete it from

  For example, when an image is excluded from the search target due to deletion of the image or the like, the number of vectors associated with each centroid fluctuates. Specifically, when an image is excluded from the search target, the number of vectors associated with each centroid decreases. In this case, when the number of vectors associated with a certain centroid is small, when the centroid is extracted, the number of corresponding search images may be less than the desired number. In addition, when there is no vector associated with a certain centroid, that is, when there are no vectors, when the centroid is extracted, there is no corresponding search image. However, it is possible to suppress a decrease in search performance by keeping the number of vectors associated with each centroid as uniform as possible by re-merging the region (centroid) divided once as described above .

[1-2. Recursive processing]
In addition, the generation device 100 may generate a plurality of second centroids to further divide the centroids C2-1 to C2-4. That is, the generating apparatus 100 may generate a plurality of third centroids to further divide the second centroids, centroids C2-1 to C2-4. For example, the generating device 100 will be described by way of example in which the number of vectors associated with the second centroid, centroid C2-1, is 75.

  In this case, for example, as illustrated in the first index information storage unit 121 in FIG. 9, the generating device 100 generates three second centroids, ie, centroids C2-2-1 to C2-C2, from the centroid C2-2. 2-3 may be generated. In the following, the centroid C2-2 is referred to as a second centroid C2-2 in order to clarify that the centroid C2-2 is a centroid generated based on the centroid C2. Also, in order to specify that the centroid C2-2-1 to C2-2-3 is a centroid generated based on the second centroid C2, the centroid C2-2-1 to C2-2-2 Let 3 be the third centroid C2-2-1 to C 2-2-3.

  For example, when the number of vectors associated with the second centroid C2-2 reaches the threshold value “75”, the generating device 100 generates three third centroids C2-2-1 to C2-2-3. Generate a new one. In addition, the generating device 100 divides the second vector group associated with the second centroid C2-2 into the third centroid C2-2-1 to C2-2-3 and associates them. That is, when the number of vectors (target vectors) associated with one second centroid C 2-2 reaches 75, the generating device 100 sets the number of vectors associated with each third centroid “25 In order to obtain “75,” the 75 target vectors are divided so as to correspond to about every 25 third averages.

  For example, in the example illustrated in FIG. 9, the generating device 100 generates the third centroid C2-2-1 to C2-2-3 in which the difference in the number of vectors associated with each third centroid is as small as possible. Generate For example, the area AR2-2-1 in the space information GR11-4 in FIG. 9 includes the vector identification information V971 and the like. That is, the generating device 100 associates the centroid C2-2-1 with about 25 vectors such as the vector V971. Further, for example, the area AR 2-2-2 in the space information GR11-4 in FIG. 9 includes the vector identification information V5003 and the like. That is, the generating device 100 associates the centroid C2-2-2 with about 25 vectors such as the vector V5003. Further, the generation device 100 similarly associates approximately 25 vectors with respect to the centroid C2-2-3.

  As described above, the generation device 100 may perform the generation process using various threshold values such as 75 and 100 as described above as appropriate. In addition, as described above, the generation device 100 may perform the generation processing with the number of divisions being an appropriate number such as three or four. That is, the generation device 100 may generate two or more second centroids, third centroids, and the like.

  Note that the generating device 100 may limit the number of times of generating a plurality of centroids recursively. For example, the generating device 100 may limit the number of times of recursively generating a plurality of centroids to two. In this case, as illustrated in the first index information storage unit 121 in FIG. 9, the generating device 100 may generate centroids (“centroids of the third hierarchy”) up to the third hierarchy. Then, in this case, after the generation apparatus 100 performs division of the area up to the third hierarchy (generation of the centroid), the number of vectors of the area corresponding to the centroid of the third hierarchy reaches a predetermined threshold. In this case, partial regions as shown in FIG. 10 may be generated for the region corresponding to the centroid of the third hierarchy.

[2. Configuration of Generation System]
As shown in FIG. 2, the generation system 1 includes a terminal device 10, an information providing device 50, and a generation device 100. The terminal device 10, the information providing device 50, and the generating device 100 are communicably connected by wire or wireless via a predetermined network N. FIG. 2 is a diagram illustrating an exemplary configuration of a generation system according to the embodiment. The generation system 1 illustrated in FIG. 2 may include a plurality of terminal devices 10, a plurality of information provision devices 50, and a plurality of generation devices 100.

  The terminal device 10 is an information processing device used by a user. The terminal device 10 receives various operations by the user. In the following, the terminal device 10 may be described as a user. That is, the user can be read as the terminal device 10 below. The terminal device 10 described above is realized by, for example, a smartphone, a tablet terminal, a laptop PC (Personal Computer), a desktop PC, a mobile phone, a PDA (Personal Digital Assistant), or the like. The example shown in FIG. 8 shows the case where the terminal device 10 is a notebook PC.

  The information providing apparatus 50 is an information processing apparatus in which information for providing various information to a user or the like is stored. For example, the information providing device 50 stores vector identification information based on character information and the like collected from various external devices such as a web server. For example, the information providing device 50 is an information processing device that provides an image search service to a user or the like. For example, the information providing device 50 stores each piece of information for providing an image search service. For example, the information providing device 50 provides the generating device 100 with vector information corresponding to an image to be an object of the image search service. In addition, the information providing device 50 transmits a query to the generating device 100 to receive, from the generating device 100, vector identification information indicating an image corresponding to the query.

  When the information on the first vector group associated with the first centroid satisfies a predetermined condition, the generation device 100 associates each of the target vectors that are vectors included in the first vector group according to the distance. Generate multiple second centroids. For example, the generation device 100 generates a plurality of second centroids when the number of target vectors of the first centroid is equal to or greater than a predetermined threshold. Furthermore, in the present embodiment, when acquiring a query, the generating device 100 provides vector identification information corresponding to the query based on the generated information. The generating device 100 may generate centroid information based on vector information collected from various external devices such as the information providing device 50. In addition, the generation device 100 may be integrated with the information providing device 50.

[3. Configuration of generator]
Next, the configuration of the generation device 100 according to the embodiment will be described with reference to FIG. FIG. 3 is a diagram showing an example of the configuration of the generation device 100 according to the embodiment. As illustrated in FIG. 3, the generation device 100 includes a communication unit 110, a storage unit 120, and a control unit 130. The generation device 100 includes an input unit (for example, a keyboard or a mouse) that receives various operations from the administrator of the generation device 100 or the like, and a display unit (for example, a liquid crystal display or the like) for displaying various information. May be

(Communication unit 110)
The communication unit 110 is realized by, for example, a network interface card (NIC). The communication unit 110 is connected to a network (for example, the network N in FIG. 2) by wire or wirelessly, and transmits and receives information to and from the terminal device 10.

(Storage unit 120)
The storage unit 120 is realized by, for example, a semiconductor memory device such as a random access memory (RAM) or a flash memory, or a storage device such as a hard disk or an optical disk. As illustrated in FIG. 3, the storage unit 120 according to the embodiment includes a first index information storage unit 121, a centroid information storage unit 122, and a second index information storage unit 123. Although not shown, the first index information storage unit 121 stores various types of information related to the above-described first index.

(First index information storage unit 121)
The first index information storage unit 121 according to the embodiment stores various types of information related to the first index. FIG. 4 is a diagram showing an example of the first index information storage unit according to the embodiment. Specifically, in the example of FIG. 4, the first index information storage unit 121 indicates tree-structured index information based on centroid information. In the example of FIG. 4, the first index information storage unit 121 includes items such as “first hierarchy”, “second hierarchy”, “third hierarchy”, and the like. In addition to the “first hierarchy” to the “third hierarchy”, the “fourth hierarchy”, the “fifth hierarchy”, the “sixth hierarchy”, and the like may be included according to the number of hierarchies of the index.

  The “first hierarchy” stores information identifying (specifying) centroids belonging to the first hierarchy of the first index. The centroid stored in the “first hierarchy” is a node corresponding to the hierarchy directly connected to the root of the first index. For example, in the “first hierarchy”, information for identifying (specifying) centroids generated by clustering before the generation of second centroids is stored.

  The “second hierarchy” stores information identifying (specifying) centroids belonging to the second hierarchy of the first index. The centroid stored in the "second hierarchy" is a node corresponding to the hierarchy immediately below the centroid of the first hierarchy. For example, in the “second hierarchy”, information for identifying (specifying) a centroid generated for the centroid in the first hierarchy is stored. For example, in the “second hierarchy”, information for identifying (specifying) the second centroid in the case of performing generation processing with the centroid of the first hierarchy as the first centroid is stored.

  The “third hierarchy” stores information identifying (specifying) centroids belonging to the third hierarchy of the first index. The centroid stored in the “third hierarchy” is a node corresponding to the hierarchy immediately below the centroid of the second hierarchy. For example, the “third hierarchy” stores information for identifying (specifying) a centroid generated for the centroid of the second hierarchy. For example, in the “third hierarchy”, information for identifying (specifying) the second centroid in the case of performing generation processing with the centroid of the second hierarchy as the first centroid is stored. For example, in the “third hierarchy”, information identifying (specifying) the third centroid in the case where the first hierarchy centroid is the first centroid and the second hierarchy centroid is the second centroid is Stored.

  For example, in the example illustrated in FIG. 4, the first index information storage unit 121 stores information corresponding to the space information GR 11-3 in FIG. 1. For example, the first index information storage unit 121 indicates that the centroids of the first layer are the centroids C1 to C11. Specifically, the first index information storage unit 121 indicates that the centroids generated by the predetermined clustering are centroids C1 to C11.

  Also, for example, the first index information storage unit 121 indicates that the centroids of the second hierarchy are centroids C2-1 to C2-4. Specifically, the first index information storage unit 121 indicates that the second layer centroid corresponding to the centroid C2 is the centroids C2-1 to C2-4.

  The first index information storage unit 121 is not limited to the above, and may store various information according to the purpose.

(Centroid information storage unit 122)
The centroid information storage unit 122 according to the embodiment stores various pieces of information related to the centroid. For example, the centroid information storage unit 122 stores centroid ID and vector data. FIG. 5 is a diagram illustrating an example of a centroid information storage unit according to the embodiment. The centroid information storage unit 122 illustrated in FIG. 5 includes items such as “centroid ID” and “vector data”.

  The "centroid ID" indicates identification information for identifying a centroid. Also, “vector data” indicates vector data corresponding to the centroid (vector) identified by the centroid ID.

  For example, in the example shown in FIG. 5, in the centroid information storage unit 122, vector data corresponding to the centroid (vector) identified by the centroid ID “C1” is “10, 24, 54, 2. Indicates that the vector is an N-dimensional vector of

  Further, for example, in the example illustrated in FIG. 5, in the centroid information storage unit 122, vector data corresponding to the centroid (vector) identified by the centroid ID “C4” is “32, 1, 120, 31 ... "indicates that the vector is an N-dimensional vector.

  The centroid information storage unit 122 is not limited to the above, and may store various information according to the purpose.

(Second index information storage unit 123)
The second index information storage unit 123 according to the embodiment stores various types of information for identifying a vector associated with each centroid stored in the centroid information storage unit 122. FIG. 6 is a diagram illustrating an example of a second index information storage unit according to the embodiment. In the example of FIG. 6, the second index information storage unit 123 includes items such as “centroid ID” and “vector ID”.

  The "centroid ID" indicates identification information for identifying a centroid. Also, “vector ID” indicates a vector associated with a centroid (vector) identified by the centroid ID.

  For example, in the example shown in FIG. 6, in the second index information storage unit 123, the vectors associated with the centroid (vector) identified by the centroid ID "C1" are vector identification information V4, V25, It indicates that the vector is identified by V32, V41, V87, etc.

  Also, for example, in the example illustrated in FIG. 6, the second index information storage unit 123 stores the vector identification information V9, which is associated with the centroid (vector) identified by the centroid ID “C4”, It indicates that the vector is identified by V12, V34, V54, V85, etc.

  The second index information storage unit 123 is not limited to the above, and may store various information according to the purpose.

(Control unit 130)
Returning to the description of FIG. 3, the control unit 130 is a controller, and is stored in a storage device inside the generating device 100 by, for example, a central processing unit (CPU) or a micro processing unit (MPU). Various programs (corresponding to an example of a generation program) are realized by executing the RAM as a work area. The control unit 130 is a controller, and is realized by, for example, an integrated circuit such as an application specific integrated circuit (ASIC) or a field programmable gate array (FPGA).

  As illustrated in FIG. 3, the control unit 130 includes an acquisition unit 131, a generation unit 132, an extraction unit 133, and a provision unit 134, and implements or executes the function or action of the information processing described below. . Note that the internal configuration of the control unit 130 is not limited to the configuration illustrated in FIG. 3, and may be another configuration as long as it performs the information processing described later.

(Acquisition unit 131)
The acquisition unit 131 acquires various types of information. For example, the acquisition unit 131 acquires various types of information from the first index information storage unit 121, the centroid information storage unit 122, the second index information storage unit 123, and the like. Further, the acquisition unit 131 acquires various information from an external information processing apparatus.

  The acquisition unit 131 acquires correspondence information indicating correspondence between a first centroid in which a vector related to a predetermined target is associated according to a distance, and a vector associated with the first centroid. The acquisition unit 131 acquires correspondence information indicating correspondence between a first centroid to which a vector to be searched for data is associated according to a distance and a vector associated to the first centroid. The acquiring unit 131 is also a first centroid to which vectors are associated by a predetermined clustering method, and acquires correspondence information of the first centroid to which a newly added vector is associated according to a distance. . In addition, the acquisition unit 131 is a first centroid in which vectors are associated by a predetermined clustering method such that the difference in the number of associated vectors is within a predetermined value, and the vector to be newly added is a distance The correspondence information of the first centroid associated in accordance with is acquired.

  For example, the acquisition unit 131 acquires a plurality of pieces of vector identification information corresponding to each of search targets. For example, the acquisition unit 131 may acquire information related to each vector identification information from the information provision device 50. For example, the acquisition unit 131 may acquire a plurality of pieces of vector identification information at different timings. For example, the acquisition unit 131 may acquire vector identification information corresponding to the image information at the timing when the image information is added as a search target.

  For example, the acquisition unit 131 acquires information on a search query. For example, the acquisition unit 131 acquires a search query related to an image search. In the example of FIG. 8, the acquisition unit 131 acquires a query from the information providing device 50 that has received a query from the terminal device 10 used by the user U1.

  In the example of FIG. 1, the acquisition unit 131 acquires vector information respectively corresponding to image information. In the example of FIG. 1, the acquisition unit 131 acquires a plurality of pieces of vector information identified by the vector identification information V3 to V1100 and the like as indicated by the space information GR11 and the like. For example, the acquisition unit 131 acquires many pieces of vector information including the vector identification information V3 to V1100 and the like.

  In the example of FIG. 1, the acquisition unit 131 acquires correspondence information indicating correspondence with a vector associated with a centroid. For example, the acquisition unit 131 acquires vector information corresponding to image information from an external information processing apparatus such as the information providing apparatus 50.

  In the example of FIG. 1, the acquisition unit 131 acquires information indicating an increase in vector. For example, the acquisition unit 131 acquires vector information corresponding to the image added to the information providing device 50. For example, the acquisition unit 131 acquires, from the information providing apparatus 50, vector identification information V5001 to V5075 corresponding to each of the 75 images.

(Generation unit 132)
The generation unit 132 generates various types of information. For example, the generation unit 132 generates various information stored in the storage unit 120. For example, the generation unit 132 generates various types of information stored in the first index information storage unit 121, the centroid information storage unit 122, the second index information storage unit 123, and the like.

  When the information on the first vector group associated with the first centroid satisfies a predetermined condition based on the correspondence information acquired by the acquisition unit 131, the generation unit 132 is a vector included in the first vector group. A plurality of second centroids are generated, each of which is associated with a certain object vector according to the distance. The generation unit 132 also generates a plurality of second centroids when the number of target vectors is equal to or greater than a predetermined threshold. In addition, the generation unit 132 generates a plurality of second centroids whose number corresponds to the number of target vectors.

  Further, the generation unit 132 generates a plurality of second centroids of a number based on the number of target vectors and a predetermined setting value. The generation unit 132 also generates a plurality of second centroids by dividing the number of target vectors by a predetermined setting value. When the generation unit 132 generates a plurality of second centroids, the generation unit 132 associates the plurality of second centroids with the first centroid, and associates each vector included in the first vector group with the first centroid. Release the attachment.

  In addition, when the number of vectors associated with each of the plurality of second centroids is equal to or less than a predetermined threshold, the generation unit 132 generates the vector associated with each of the plurality of second centroids as the first cent It associates with the Lloyd and cancels the association of the first centroid and the plurality of second centroids. In addition, the generation unit 132 deletes the plurality of second centroids.

  In addition, when the information on the second vector group associated with one second centroid among the plurality of second centroids satisfies the predetermined condition, the generation unit 132 is a vector included in the second vector group. A plurality of third centroids are generated, each of which is associated with a certain object vector according to the distance.

  For example, the generation unit 132 is a plurality of pieces of centroid information in which each piece of vector information is associated according to a distance in vector quantization based on a plurality of pieces of vector information, and a vector associated with each piece of centroid information Generate multiple sets of centroid information so that the number of pieces of information is even. For example, the generation unit 132 generates a plurality of pieces of centroid information by selecting vector information as centroid information from a plurality of pieces of vector information based on a predetermined condition. For example, the generation unit 132 generates a plurality of pieces of centroid information by selecting vector information selected by random sampling as centroid information from a plurality of pieces of vector information.

  For example, the generation unit 132 generates, for a predetermined search request, first index information used to specify centroid information. For example, the generation unit 132 generates, as first index information, a search index for searching a high-dimensional vector. For example, the generation unit 132 generates a tree structure type index or a graph structure type index related to a plurality of pieces of centroid information as first index information.

  For example, the generation unit 132 generates second index information indicating each centroid information and vector information associated with each centroid information. For example, the generation unit 132 generates, as second index information, a transposed index that specifies vector information associated with each piece of centroid information.

  For example, the generation unit 132 generates a plurality of pieces of centroid information so that the number of pieces of vector information associated with each piece of centroid information is equal based on vector quantization that quantizes each partial vector obtained by dividing each piece of vector information. Generate For example, the generation unit 132 generates partial centroid information corresponding to a plurality of partial regions into which the region is divided, based on vector information included in the region corresponding to each piece of centroid information. For example, the generation unit 132 generates partial centroid information corresponding to a plurality of partial regions into which the region is divided, based on residual vector information on vector information included in the region corresponding to each piece of centroid information. For example, based on each centroid information and residual vector information generated by vector information included in a region corresponding to each centroid information, the generation unit 132 divides the region into a plurality of divided partial regions. Generate corresponding partial centroid information. For example, the generation unit 132 generates association information in which partial centroid information of each partial region is associated with vector information included in each partial region. For example, the generation unit 132 generates correspondence information based on residual vector information generated by each centroid information and each partial centroid information corresponding to each partial region of the respective centroid information. .

  Further, the generation unit 132 may generate centroid information such as a plurality of first centroid information by selecting a vector as a centroid. For example, the generation unit 132 may perform random sampling a predetermined number of times, and generate a plurality of pieces of centroid information based on a combination of vectors (centroids) in which the vectors are clustered most evenly. Also, for example, the generation unit 132 corresponds to each centroid by changing a centroid with a large deviation from the average value to another vector in the vicinity among the combinations of centroids in which the vectors are clustered most evenly. A plurality of pieces of centroid information may be generated such that the number of vectors to be attached is even.

  For example, the generation unit 132 performs random sampling a predetermined number of times, and generates a plurality of pieces of centroid information based on a combination of centroids in which the difference in the number of vectors included in each cluster is the smallest and vectors are clustered. For example, the generation unit 132 performs random sampling a predetermined number of times, and generates a plurality of pieces of centroid information based on a combination of centroids in which vectors are clustered most evenly. Also, for example, the generation unit 132 corresponds to each centroid by changing a centroid with a large deviation from the average value to another vector in the vicinity among the combinations of centroids in which the vectors are clustered most evenly. A plurality of pieces of centroid information may be generated such that the number of vectors to be attached is even. Also, for example, when the difference in the number of vectors included in each cluster becomes equal to or less than a predetermined value (for example, 5 or 10), the generation unit 132 ends the random sampling, and based on the combination of the centroids. , And may generate a plurality of centroid information. For example, when the difference between the cluster including the largest number of vectors and the cluster including the smallest number of vectors becomes equal to or less than a predetermined value (for example, 1 or 50), the generation unit 132 ends the random sampling. Multiple centroid information may be generated based on a combination of Lloyds. Also, for example, when the difference in the number of vectors included in each cluster is “0”, that is, the number of vectors included in each cluster is the same, the generation unit 132 ends the random sampling, and Multiple centroid information may be generated based on the combination.

  For example, the generation unit 132 may generate the second index information based on the vector associated with each centroid. For example, the generation unit 132 generates second index information in which information identifying a vector included in each of the areas AR1 to AR11 and the like corresponding to each centroid is associated with a centroid ID. In addition, the generation unit 132 generates first index information used to specify a centroid corresponding to the time of acquiring a query.

  In the example of FIG. 1, the generation unit 132 clusters a plurality of vectors by a predetermined clustering method. For example, the generation unit 132 clusters the plurality of vectors by a predetermined clustering method so that the difference in the number of vectors associated with the centroid becomes uniform. For example, the generation unit 132 generates a plurality of pieces of centroid information such that the difference in the number of vectors associated with each centroid is within a predetermined value. For example, the generation unit 132 generates a plurality of pieces of centroid information such that the number of vectors associated with each centroid is equal.

  In the example of FIG. 1, as shown by the space information GR11 in FIG. 1, the generation unit 132 centroids the vector identification information C1, C2, C3, C4, C5, C6, C7, C7, C8, C9, C10, and C11. Generate centroid information. The generation unit 132 generates a plurality of pieces of centroid information as shown in the centroid information storage unit 122-1.

  In addition, the generation unit 132 generates second index information based on the vector associated with each centroid. For example, the generation unit 132 generates second index information in which information identifying a vector included in each of the areas AR1 to AR11 and the like corresponding to each centroid is associated with a centroid ID.

  In the example of FIG. 1, since 100 vectors are associated with the centroid C2, the generation unit 132 determines that the threshold is reached, and generates a centroid. For example, the generation unit 132 generates four second centroids corresponding to the first centroid, the centroid C2. For example, the generating unit 132 performs clustering so as to divide the 100 vectors associated with the centroid C2 into four to generate four second centroids, centroids C2-1 to C2-4. Generate For example, the generating unit 132 generates four centroids C2-1 to C2-4, which are four second centroids, by performing clustering so as to divide the area AR2 into four areas.

  Further, in the example of FIG. 1, the generation unit 132 generates information for identifying a vector included in each of the areas AR2-1 to AR2-4 corresponding to the respective centroids C2-1 to C2-4 as the centroid C2-1. The second index information associated with C2-4 is generated.

(Extraction unit 133)
The extraction unit 133 selects various information. For example, the extraction unit 133 selects vector identification information. For example, the extraction unit 133 selects centroid information based on the query acquired by the acquisition unit 131.

  In the example of FIG. 8, the extraction unit 133 extracts a centroid using the query information QE11. For example, the extraction unit 133 specifies and extracts a centroid corresponding to the query information QE11 based on the first index information stored in the first index information storage unit 121. For example, the extraction unit 133 specifies a feature area in the query information QE11 by using various conventional techniques as appropriate, and calculates a feature quantity in the feature area as a local feature quantity. For example, the extraction unit 133 extracts local feature quantities in the query information QE11 using various conventional techniques as appropriate.

  In the example of FIG. 8, the extraction unit 133 extracts centroid information identified by the centroid ID “C2-1” from the centroid information storage unit 122. For example, since the extraction unit 133 is closest to the centroid identified by the centroid ID “C2-1” in the spatial information GR11, the query information QE11 is identified by the centroid ID “C2-1”. Extracted centroid information. For example, the extraction unit 133 extracts a vector based on centroid information. For example, the extraction unit 133 extracts a vector similar to the query information QE11 based on centroid information. For example, the extraction unit 133 extracts a vector from the second index information storage unit 123 based on centroid information.

  In the example of FIG. 8, the extraction unit 133 stores vector identification information V66 and V5001 stored in the second index information storage unit 123 in association with the centroid (vector) identified by the centroid ID “C2-1”. , V5004 and so on.

(Provision unit 134)
The providing unit 134 provides various information. For example, the providing unit 134 provides various information to the terminal device 10 and the information providing device 50. For example, the providing unit 134 provides various information stored in the storage unit 120 to an external information processing apparatus. For example, the providing unit 134 provides various types of information such as space information GR11.

  For example, the providing unit 134 provides vector identification information corresponding to a query as a search result. For example, the providing unit 134 provides the information providing apparatus 50 with the vector identification information selected by the extracting unit 133. In the example of FIG. 8, the providing unit 134 provides the information providing apparatus 50 with the vector identification information selected by the extracting unit 133 as information indicating a vector corresponding to a query. In the example of FIG. 8, the providing unit 134 provides the vector identification information V66, V5001, V5004 and the like to the information providing apparatus 50 as information indicating a vector corresponding to the query. Further, when there is actual data (image data) corresponding to the centroid C2-1, the providing unit 134 may also provide the information providing apparatus 50 with vector information of the centroid C2-1.

  In the example of FIG. 8, the providing unit 134 provides the information providing apparatus 50 with information for identifying the vector extracted by the extracting unit 133. For example, the providing unit 134 provides the information providing apparatus 50 with vector identification information V66, V5001, V5004 and the like located in the area AR2-1 corresponding to the centroid C2-1 extracted as a centroid. For example, the providing unit 134 provides the information providing apparatus 50 with 25 pieces of vector identification information including vector identification information V66, V5001, V5004 and the like corresponding to the centroid C2-1 extracted as a centroid.

[4. Flow of generation process]
Next, a procedure of generation processing by the generation system 1 according to the embodiment will be described using FIG. 7. FIG. 7 is a flowchart illustrating an example of the generation process according to the embodiment.

  As illustrated in FIG. 7, the generating apparatus 100 acquires correspondence information indicating correspondence between the first centroid and the vector associated with the first centroid (step S101). In the example of FIG. 1, the generating device 100 acquires correspondence information on the centroid C2 stored in the second index information storage unit 123. Specifically, the generating apparatus 100 acquires, from the second index information storage unit 123, correspondence information in which the vectors V33, V66, V81, V112, V971 and the like are associated with the centroid C2.

  After that, when the information on the first vector group associated with the first centroid satisfies a predetermined condition, the generation apparatus 100 generates a plurality of second centroids (step S102). For example, when the number of vectors associated with a centroid is equal to or greater than a predetermined threshold, the generating apparatus 100 generates a plurality of second centroids. In the example of FIG. 1, the generating device 100 divides the number of vectors “100” by the setting value “25” because the number of vectors associated with the centroid C2 as the first centroid is “100”. Based on the number "4", four second centroids are generated. Specifically, the generating device 100 generates four second centroids C2-1, C2-2, C2-3, and C2-4.

[5. About provision of information]
Next, information provision by the generation system 1 according to the embodiment will be described using FIG. 8. FIG. 8 is a diagram illustrating an example of provision of information corresponding to a query according to the embodiment. The example of FIG. 8 shows the case where the information providing device 50 provides the terminal device 10 with the search result in response to the query acquired from the terminal device 10. Specifically, the case where the information providing apparatus 50 provides image information as a search result to the user based on the vector identification information acquired from the generating apparatus 100 is shown. In the example of FIG. 8, the case where the information providing apparatus 50 provides the image search service by the k nearest neighbor search will be described as an example. For example, in the k nearest neighbor search, k vectors are searched (extracted) from the side closer to the search query. In addition, below, in order to simplify description, the case where "k" is "6" is shown.

  First, the information providing device 50 acquires a query from the terminal device 10 used by the user U1 (step S21). In the example of FIG. 8, the information providing device 50 acquires the query information QE11 from the terminal device 10. For example, the information providing device 50 acquires image information from the terminal device 10 as the query information QE11.

  Then, the information providing device 50 transmits a query to the generating device 100 (step S22). In the example of FIG. 8, the information providing device 50 transmits the query information QE11 acquired from the terminal device 10 to the generating device 100. Then, the generating device 100 acquires image information as the query information QE11.

  Thereafter, the generating apparatus 100 extracts centroids using the query information QE11 (step S23). For example, based on the first index information stored in the first index information storage unit 121, the generation apparatus 100 identifies and extracts a centroid corresponding to the query information QE11.

  For example, the generation apparatus 100 specifies a feature area in the query information QE11 using various conventional techniques as appropriate, and calculates a feature quantity in the feature area as a local feature quantity. For example, the generation apparatus 100 extracts local feature quantities in the query information QE11 using various conventional techniques as appropriate.

  Also, for example, the generating device 100 generates information on BoF from the query information QE11. Then, based on the first index information stored in the first index information storage unit 121, the generating apparatus 100 specifies a centroid by searching for BoFs similar to the information related to the BoF of the query information QE11.

  For example, the generating apparatus 100 searches for a BoF expression closest to the information on BoF of the query information QE11 using the first index information stored in the first index information storage unit 121. When the distance between the information on BoF of the query information QE11 and the information on BoF of the centroid extracted in the search is equal to or greater than a predetermined threshold, the generating device 100 has no vector information corresponding to the query information QE11. It may be determined that The generation apparatus 100 may appropriately extract centroids using various conventional techniques as described in Patent Document 2. For example, the generation device 100 verifies the degree of certainty of similarity between the query information QE11 and the image corresponding to the extracted centroid by performing matching of feature point coordinates by the process described in Patent Document 2. May be

  In the example of FIG. 8, the generating device 100 extracts centroid information identified by the centroid ID “C2-1” from the centroid information storage unit 122. For example, since the generation device 100 is closest to the centroid identified by the centroid ID “C2-1” in the spatial information GR11, the query information QE11 is identified by the centroid ID “C2-1”. Extracted centroid information.

  Then, the generating apparatus 100 extracts a vector based on the centroid information (step S24). For example, the generating apparatus 100 extracts a vector similar to the query information QE11 based on centroid information. For example, the generating apparatus 100 extracts a vector from the second index information storage unit 123 based on centroid information.

  In the example of FIG. 8, the generating device 100 stores the vector identification information V66 and V5001 stored in the second index information storage unit 123 in association with the centroid (vector) identified by the centroid ID “C2-1”. , V5004 and so on. For example, in FIG. 8, it is assumed that 25 vector identification information (vectors) including the vector identification information V66, V5001, V5004 and the like located in the area AR2-1 corresponding to the centroid C2-1 are extracted.

  Thereafter, the generating device 100 provides the information providing device 50 with information for identifying the extracted vector (step S25). In the example of FIG. 8, the generating device 100 provides the information providing device 50 with 25 pieces of vector identification information including the vector identification information V66, V5001, V5004 and the like corresponding to the VC centroid extracted as a centroid.

  The information providing device 50 that has received the information identifying the vector from the generating device 100 provides the terminal device 10 with the image information associated with the received information identifying the vector (step S26). For example, the information providing device 50 provides the terminal device 10 with the image information associated with the information identifying the vector received from the generating device 100 as a search result corresponding to the query information QE11. In the example of FIG. 8, the information providing device 50 that has received the vector identification information V66, V5001, V5004, etc. from the generation device 100 provides the terminal device 10 with image information corresponding to each of the vector identification information V66, V5001, V5004, etc. Do. For example, the information providing device 50 calculates the degree of similarity between the image information corresponding to each of the vector identification information V66, V5001, V500 etc. and the query information QE11, and the terminal device is the image information arranged in the order according to the degree of similarity. You may provide to ten.

[6. 1st index]
In addition, the generating apparatus 100 may use the first index information to specify the corresponding centroid when acquiring the query. This point will be described with reference to FIG. FIG. 9 is a schematic diagram of first index information according to the embodiment.

  In the example of FIG. 1, the generating apparatus 100 generates first index information as shown in the first index information storage unit 121 of FIG. 4 in order to specify a corresponding centroid when acquiring a query. In addition, the generating device 100 updates the information in the first index information storage unit 121 according to the generation of the centroid by the generation process.

  The first index information storage unit 121 of FIG. 9 shows a case where the generating device 100 further generates a second centroid for the centroid C2-2. For example, in the example illustrated in FIG. 9, the generating device 100 generates the third centroids C2-2-1 to C2-2-3.

  In the “third hierarchy” shown in the first index information storage unit 121 in FIG. 9, the second centroid of the centroid C2-2 is stored. For example, the first index information storage unit 121 in FIG. 9 indicates that the centroid in the third layer is the centroid C2-2-1 to C2-2-3. Specifically, in the first index information storage unit 121 of FIG. 9, the centroid of the third hierarchy corresponding to the centroid C2-2 of the second hierarchy is the centroid C 2-2-1 to C2-2-3. To indicate that

  Further, the inside of the area AR2 corresponding to the centroid C2 in this case is as shown in space information GR11-4 in FIG. For example, the area AR2-2-1 in the space information GR11-4 in FIG. 9 includes the vector identification information V971 and the like. Further, for example, the area AR 2-2-2 in the space information GR11-4 in FIG. 9 includes the vector identification information V5003 and the like.

  For example, the generating apparatus 100 generates tree information TR11 corresponding to the first index information shown in the first index information storage unit 121 in FIG. 9 (step S31). In addition, in FIG. 9, in order to show an index as a figure, although demonstrated as what produces | generates tree information TR11, the information shown in the 1st index information storage part 121 in FIG. 9 contains the information of tree information TR11. In the example of FIG. 9, the generating apparatus 100 generates graph information GR1 including tree information TR11 and space information GR11. The tree information TR11 may be included in the space information GR11. Further, in the example of FIG. 9, in order to simplify the description, illustration of centroids C1 to C6, C8, C9, C11 and the like is omitted, and a part of tree information TR11 is illustrated.

  The tree information TR11 in FIG. 9 includes a route RT, centroids C1 to C11, C2-1 to C2-4, C2-2-1 to C2-2-3, and the like. In the example of FIG. 1, for example, centroids C1 to C11, C2-1 to C2-4, and C2-2-1 to C2-2-3 may be vectors corresponding to actual data.

  As shown in tree information TR11 in FIG. 9, centroids C1 to C11 belonging to the first hierarchy are directly connected to the route RT. For example, the centroid C10 belonging to the first hierarchy is directly connected to the route RT. In addition, since the centroid C10 has no corresponding second layer centroid, the vector is associated. That is, the centroid C10 is a leaf (leaf node) of the tree information TR11. In addition, the centroids C1, C3 to C9, and C11 of the first hierarchy other than the centroid C2 are also leaves (leaf nodes) of the tree information TR11.

  Also, for example, the centroid C2 belonging to the first hierarchy is directly connected to the route RT. In addition, the centroid C2 is connected to the centroids C2-1 to C2-4 because there are the corresponding centroids C2-1 to C2-4 of the second layer.

  Further, since there is no corresponding third layer centroid, the vectors are associated with the second layer centroids C2-1, C2-3, and C2-4 corresponding to the centroid C2. For example, vectors V66, V5001, and the like are associated with the centroid C2-1. Also, for example, vectors V81, V112, etc. are associated with the centroid C2-3. Also, for example, vectors V33, V5002, and the like are associated with the centroid C2-4. Thus, the centroids C2-1, C2-3, and C2-4 of the second hierarchy other than the centroid C2-2 become leaves (leaf nodes) of the tree information TR11.

  Further, since there is no corresponding fourth-layer centroid, the third-layer centroid C2-2-1 to C2-2-3 corresponding to the centroid C2-2 is associated with a vector. For example, a vector V971 or the like is associated with the centroid C2-2-1. Further, for example, a vector V5003 or the like is associated with the centroid C2-2-2. Thus, the centroids C2-2-1 to C 2-2-3 of the third hierarchy are leaves (leaf nodes) of the tree information TR 11.

  As described above, in the example illustrated in FIG. 9, vectors are associated with centroids that are leaf nodes regardless of the level of the hierarchy. Therefore, when the generating device 100 acquires a search query, the generation device 100 can specify a vector corresponding to the search query by specifying a leaf node close to the search query.

  For example, when the generating device 100 acquires the search query QE11 (see FIG. 8), the generation device 100 efficiently searches the tree information TR11 from the top to the bottom to specify a centroid serving as a neighbor candidate of the search query QE11. The centroid corresponding to the search query can be identified in. For example, the generating device 100 may specify a centroid serving as a neighbor candidate of the search query by tracing the tree information TR11 from the root RT to the leaf node. In the example of FIG. 8, for example, the generating device 100 may specify the centroid C2-1 to be a neighbor candidate of the search query QE11 by tracing the tree information TR11 from the route RT to the centroid C2-1.

  Note that the tree structure or graph type search index related to a plurality of pieces of centroid information as described above is an example, and the generating apparatus 100 can identify centroid information corresponding to a query at high speed. The first index information of such data structure may be generated. For example, if the generating apparatus 100 can specify centroid information corresponding to a query at high speed, the generating apparatus 100 appropriately generates the first index information using various conventional techniques such as a technique related to binary space division. It is also good. For example, the generating apparatus 100 may generate the first index information of any data structure as long as the index can correspond to the search of the high-dimensional vector.

  Further, the generating device 100 is not limited to the first index information described above, and may use any first index information as long as it can be used to specify the corresponding centroid when acquiring a query. . For example, the generating apparatus 100 generates a search index for searching for a high-dimensional vector as first index information. The high-dimensional vector referred to here may be, for example, a vector of several hundred to several thousand dimensions, or may be a vector of more dimensions.

  Also, for example, the generating device 100 may generate, as the first index information, information on a graph-type search index as described in Non-Patent Document 2. Also, for example, the generating device 100 may generate a search index related to a tree structure (tree structure) as the first index information. For example, the generating apparatus 100 may generate a search index for k-dimensional trees as first index information. For example, the generating apparatus 100 may generate a search index on a VP-tree (vantage-point tree) as first index information. Thus, the generating apparatus 100 may generate a tree-structured index and a graph-structured index related to a plurality of pieces of centroid information as the first index information. Also, for example, the generating device 100 may generate a transposed index as first index information.

  Also, for example, the generating device 100 may use another index such as a hash type capable of searching for high-dimensional data as the first index information. For example, the generating device 100 may store another index such as a hash type capable of searching for high-dimensional data in the first index information storage unit 121 as the first index information.

  Also, for example, the generating device 100 may generate a graph structure (graph information) in a leaf node by connecting centroids that are leaf nodes of the tree information TR11 with a predetermined edge. The example of FIG. 9 shows a case where the generating device 100 connects the leaf nodes, such as centroids C7, C10, C2-1, C2-3, etc., with the edge ED. The generation apparatus 100 may generate graph information in which each leaf node of the tree information TR11 is connected. In this case, for example, the generation device 100 may specify a centroid by tracing to a leaf node of the tree information TR11 and then tracing with a graph structure between leaf nodes.

[7. Regarding Generation of Second Index]
In the above-mentioned example, although processing based on a field corresponding to each centroid was explained, generating device 100 may divide a field corresponding to each centroid into partial fields. This point will be described below with reference to FIG. FIG. 10 is a diagram illustrating an example of a process of generating second index information according to the embodiment.

  For example, the space information GR11 in FIG. 10 corresponds to the space information GR11 in FIG. 8, and each vector corresponding to the query information QE21 and the centroid C4 and the vectors V9, V12, V34, V54, V85 included in the area AR4. Show the relationship with Thus, as shown in the space information GR11 in FIG. 10, although the distance between the query information QE21 and each vector corresponding to the centroid C4 and the vectors V9, V12, V34, V54, and V85 is different, Does not store vector data corresponding to the vector identification information V9, V12, V34, V54, and V85. In this case, for example, the distance between the query information QE21 and each vector is regarded as the distance between the query information QE21 and the vector corresponding to the centroid C4 of the area AR4. In other words, vector data corresponding to the vector identification information V9, V12, V34, V54, V85 is quantized into vector data corresponding to the centroid C4. In such a case, for example, since the vectors included in each of the areas AR1 to AR11 and the like are regarded as the same distance, when the number of vectors increases, the corresponding data as the search result increases.

  Therefore, the generation device 100 may further divide each of the areas AR1 to AR11 or the like into partial areas in the generation process. Below, the case where area | region AR4 is divided | segmented into a partial area | region further is shown. For example, the generating apparatus 100 generates partial centroid information corresponding to a plurality of partial regions into which the region is divided, based on residual vector information related to vector information included in the region corresponding to each centroid information. It is also good. In the following, the generation device 100 generates a plurality of partial areas into which the area is divided based on each centroid information and residual vector information generated by vector information included in the area corresponding to each centroid information. Shows a case where partial centroid information corresponding to is generated.

  For example, the generating apparatus 100 may determine the number of subspaces based on the number of vectors included in the area AR4, or may determine the number of subspaces based on a predetermined setting value. The example of FIG. 10 shows the case where the generating device 100 divides the area AR4 into nine partial areas AR41 to AR49. For example, the generating device 100 may determine the range of each partial area AR41 to AR49 such that the number of vectors included in each area AR41 to AR49 is equal. Further, for example, the generating device 100 may determine the range of each partial region AR41 to AR49 by appropriately using various conventional techniques related to clustering.

  In the example of FIG. 10, the generating apparatus 100 calculates a residual vector from the vectors and centroids included in each partial area AR41 to AR49, and generates a centroid of the partial area from the residual vector. In this case, even if there are multiple regions, there is not a separate set of centroids, but a common and single set of centroids. However, the generation device 100 may generate a centroid set individually for each of the plurality of regions. For example, if the total number of centroids is equal to or less than a predetermined threshold, the generation device 100 may generate a separate centroid set for each of the plurality of regions. The centroid of the partial region may not be a residual vector but may be an original vector. For example, when the variance of the original vector is equal to or less than a predetermined threshold, the generating device 100 may set the original vector as the centroid of the partial region. For example, the generating device 100 generates the centroid CN41 based on the vector included in the partial area AR41 including the vector identified by the vector identification information V34. Also, for example, the generating device 100 generates the centroid CN 42 based on the vector included in the partial area AR 42 including the vector identified by the vector identification information V 9. The generating apparatus 100 may generate partial centroid information corresponding to a plurality of partial regions into which the region is divided, based on vector information included in the region corresponding to each piece of centroid information.

  For example, the generating apparatus 100 stores information on residual vectors of the centroid of one region and the centroid of each partial region. In the example of FIG. 10, as indicated by the partial centroid information LT1, the generating device 100 includes information on residual vectors of the centroid C4 of the region AR4 and the centroid CN41 of the partial region AR41 and the centroid of the region AR4. Information on the residual vector of C4 and the centroid CN 48 of its partial area AR 48 is stored. The generating apparatus 100 may store the partial centroid information LT1 in the storage unit 120.

  In addition, the generating device 100 may store centroid information of a partial region to which the vector belongs in association with each vector. In the example of FIG. 10, as illustrated in the second index information storage unit 123A, the generating device 100 may store the centroid CN 42 in association with the vector V9 stored in association with the centroid C4. The generation device 100 may include the second index information storage unit 123A in the storage unit 120. Specifically, the generating device 100 may store the centroid ID “CN42” in association with the vector identification information “V9” associated with the centroid ID “C4”. As described above, the generating device 100 stores, in the second index information storage unit 123A, each vector in association with information on the centroid of the partial region to which the vector belongs. Thereby, the generation device 100 can specify which partial area the vector belongs to by referring to the second index information storage unit 123A, and the processing can be further speeded up.

  As described above, when there is no division by partial areas AR41 to AR49, the distance between query information QE21 and each vector is regarded as the distance between query information QE21 and the vector corresponding to centroid C4 of area AR4. . On the other hand, generation apparatus 100 can further subdivide the distance between query information QE21 and each vector by performing division by partial areas AR41 to AR49. This point will be described with reference to FIG. FIG. 11 is a diagram illustrating an example of use of second index information according to the embodiment.

  For example, the generating apparatus 100 can quantize the position of each vector to the position of the centroid of the partial region to which the vector belongs by using the partial centroid information LT1. Thereby, the generating apparatus 100 can subdivide the distance between the query information QE21 and each vector to the distance between the query information QE21 and the vector corresponding to the centroid of the partial area AR41 to AR49 to which each vector belongs.

  In the example of FIG. 11, the generating device 100 has vector data of the vector identified by the centroid ID "C4" as shown in the centroid information storage unit 122, and the centroid as shown in the partial centroid information LT1. It has residual vector information on CN41. Therefore, the generating device 100 can calculate, for example, the distance to the centroid CN 41 with respect to the query information QE 21. Therefore, the generating device 100 can set, for example, the distance between the query information QE21 and the vector V34 to the distance to the centroid CN41 which is more approximate than the centroid C4.

  As described above, the generation device 100 further suppresses the number of vector data to be stored to the number of centroids corresponding to the regions by using information on partial regions obtained by further dividing the regions AR1 to AR11 and the like. Each vector can be subdivided and quantized. For example, even if the number of vectors included in each of the areas AR1 to AR11 is enormous (for example, 10,000, 100,000, etc.), the generating apparatus 100 further divides each area by further dividing the vectors. Can be appropriately quantized. Therefore, the generating device 100 can generate centroid information that enables efficient search.

7-1. About Generation of Second Index at Division]
In the examples of FIGS. 10 and 11, although the case of dividing the area corresponding to the centroid C4 of the first layer into the divided areas is shown as an example, the generating apparatus 100 determines the cents of the layers after the second layer. The region corresponding to Lloyd may be divided into partial regions.

  For example, in the example of FIG. 1, the generating apparatus 100 may further divide each of the centroids C2-1 to C2-4 generated as the second centroid into partial regions. For example, the generating apparatus 100 generates partial centroid information corresponding to a plurality of partial regions into which the region AR2-1 is divided, based on the vector information included in the region AR2-1 corresponding to the centroid C2-1. You may

  In addition, the generation device 100 may divide an area corresponding to a centroid below a predetermined level into partial areas. For example, as illustrated in FIG. 9, the generation apparatus 100 generates the second centroids up to the third hierarchy, and the vectors corresponding to the centroids of the third hierarchy become a predetermined number or more. The region corresponding to the centroid may be divided into partial regions.

[8. Example of vector division]
In the above example, the generation device 100 processes each vector as one vector. However, the generation device 100 may divide each vector into a plurality of partial vectors for processing. For example, the generation device 100 may perform processing using a technique related to so-called direct product quantization. This point will be described with reference to FIG. FIG. 12 is a diagram illustrating an example of vector division according to the embodiment. Although the example of FIG. 12 illustrates the case where the number of dimensions of the vector data QE31 corresponding to the query is 12 in order to simplify the description, the generation device 100 may target vectors of higher dimensions. .

  In the example of FIG. 12, the generating device 100 divides the vector data QE31 into four. For example, the generating device 100 sets one-dimensional to three-dimensional data as partial vector data QE 31-1 and sets four-dimensional to six-dimensional data as partial vector data QE 31-2 in the vector data QE 31. The data of is the partial vector data QE31-3, and the data of 10 dimensions to 12 dimensions is the partial vector data QE31-4. Specifically, the generation device 100 sets “45, 23, 2” as partial vector data QE 31-1, “127, 34, 5” as partial vector data QE 31-2, and “20, 98, 110”. The partial vector data QE31-3 is used, and "12, 45, 4" is used as partial vector data QE31-4.

  Although the case where the query is divided into partial vectors has been described as described above, in the centroid information generation process, vectors corresponding to each piece of image information are divided into partial vectors. Further, FIG. 12 illustrates a case where the area AR4 in the space information GR11 illustrated in FIG. 1 is divided into a plurality of partial areas (hereinafter, also referred to as “partial space”) for simplifying the description. The example shown in FIG. 12 shows the case where the area AR4 in the space information GR11 is divided into a plurality of partial spaces AR4-1 to AR4-4.

  Partial vector data QE31-1 is vector data corresponding to partial space AR4-1, partial vector data QE31-2 is vector data corresponding to partial space AR4-2, and partial vector data QE31-3 is It is assumed that the vector data corresponds to the partial space AR4-3, and the partial vector data QE31-4 is vector data corresponding to the partial space AR4-4. For example, in the example of FIG. 12, the partial vector data QE31-1 are located in the region of the centroid CN43 in the partial space AR4-1.

  Although the partial spaces AR4-1 to AR4-4 are shown in similar shapes in the example of FIG. 12, the shapes of the partial spaces AR4-1 to AR4-4 may be different, and the partial spaces AR4-1 to AR4-4 may be different. The division mode of the area in AR 4-4 may also be different.

  Then, the generating apparatus 100 calculates the distance between each query and the centroid by summing the distances in all the partial spaces AR4-1 to AR4-4. For example, the generating apparatus 100 may calculate the distance between each query and the vector by the following equation (1). In the above-described example, the centroid of the partial region is basically common, but in the case of the partial vector, the generating device 100 may generate a centroid set in vector units. For example, the generating apparatus 100 may generate a centroid set for each of the subspaces AR4-1 to AR4-4. As a result, even if the tendency of the distribution of vector data for each partial vector is different, the generating apparatus 100 can generate centroid information that enables efficient search. In addition, the generating apparatus 100 may generate a centroid set with the partial vectors in common. For example, the generating apparatus 100 may generate a centroid set common to the subspaces AR4-1 to AR4-4. Thereby, the generation device 100 can generate centroid information that enables efficient search even when the vector is divided into partial vectors.

Here, the value on the left side in the above equation (1) indicates, for example, the square distance between the query and the vector. Also, for example, “x” in the above equation (1) corresponds to a query. Also, for example, “y” in the above equation (1) corresponds to a vector. Also, for example, “q _c (y)” in the right side of the above-mentioned formula (1) indicates a centroid of “y”. For example, in the case where “y” in the above equation (1) does not have vector data of vectors, the generation device 100 uses numeric values of vector data of centroids of partial regions to which each vector belongs as shown in FIG. May be Also, for example, “y−q _c (y)” indicates a residual vector. Also, for example, “q _p ” in the right side of the above equation (1) indicates a predetermined quantizer (function).

Also, for example, “j” in the right side of the above equation (1) may be the number of divided spaces. For example, in the example of FIG. 12, “j” in the right side of the above equation (1) may be the number “4” of the divided spaces. Also, for example, “u _j ()” in the right side of the above equation (1) indicates a partial residual vector between the vectors in parentheses. For example, even if the generating device 100 calculates the square distance between the query and the vector in each partial space using the above equation (1), and calculates the distance between each query and the vector by summing. Good. For example, in the example of FIG. 12, the generating device 100 generates partial vector data QE31-1 to QE31-4 in each partial space AR4-1 to AR4-4 and a vector in each partial space AR4-1 to AR4-4. The distance between the query corresponding to the vector data QE31 and the vector may be calculated by calculating the square distance between them and summing.

  As a result, the generating apparatus 100 can further segment and quantize each vector, and can more appropriately extract the vector corresponding to the query. In addition, the generation device 100 may generate centroid information that enables efficient search in proximity search by appropriately combining the above-described processes. For example, the generating apparatus 100 may use each partial vector at the time of direct product quantization as it is, or may calculate and use a residual vector. For example, the generating apparatus 100 may calculate a residual vector from each partial vector and centroid and use the residual vector. In addition, the generating apparatus 100 may generate centroids in common among partial vectors, or may generate centroids individually. For example, the generating apparatus 100 may generate a centroid in common between partial vectors into which one vector is divided, or may separately generate centroids. Note that the above is an example, and the generating apparatus 100 may generate centroid information that enables efficient search in proximity search by combining various information as appropriate or individually. For example, the generating apparatus 100 may generate centroid information that enables efficient search in proximity search by using certain information in common or using certain information in an individualized manner.

[9. effect〕
As described above, the generation device 100 according to the embodiment includes the acquisition unit 131 and the generation unit 132. The acquisition unit 131 acquires correspondence information indicating correspondence between a first centroid in which a vector related to a predetermined target is associated according to a distance, and a vector associated with the first centroid. When the information on the first vector group associated with the first centroid satisfies a predetermined condition based on the correspondence information acquired by the acquisition unit 131, the generation unit 132 is a vector included in the first vector group. A plurality of second centroids are generated, each of which is associated with a certain object vector according to the distance.

  Thus, the generating device 100 according to the embodiment determines the vectors included in the first vector group when the information on the first vector group associated with the first centroid satisfies the predetermined condition based on the correspondence information. By generating a plurality of second centroids in which each of the object vectors that are is associated according to the distance, it is possible to generate centroid information that enables efficient search.

  Further, in the generating device 100 according to the embodiment, the generating unit 132 generates a plurality of second centroids when the number of target vectors is equal to or more than a predetermined threshold.

  As described above, when the number of target vectors is equal to or more than a predetermined threshold value, the generating device 100 according to the embodiment generates centroid information that enables efficient search by generating a plurality of second centroids. Can be generated.

  In addition, in the generation device 100 according to the embodiment, the generation unit 132 generates a plurality of second centroids of the number according to the number of target vectors.

  As described above, the generating device 100 according to the embodiment can generate centroid information that enables efficient search by generating a plurality of second centroids in a number corresponding to the number of target vectors. it can.

  In addition, in the generation device 100 according to the embodiment, the generation unit 132 generates a plurality of second centroids of a number based on the number of target vectors and a predetermined setting value.

  Thus, the generating apparatus 100 according to the embodiment can generate efficient second searching by generating a plurality of second centroids based on the number of target vectors and a predetermined setting value. Can be generated.

  Further, in the generating device 100 according to the embodiment, the generating unit 132 generates a plurality of second centroids by dividing the number of target vectors by a predetermined setting value.

  As described above, the generating apparatus 100 according to the embodiment generates centroid information that enables efficient search by generating a plurality of second centroids by dividing the number of target vectors by a predetermined setting value. Can be generated.

  In the generation device 100 according to the embodiment, the acquisition unit 131 is a first centroid to which a vector is associated by a predetermined clustering method, and a vector to be newly added is associated according to a distance. 1Correspondence information is acquired.

  As described above, the generating device 100 according to the embodiment is the first centroid to which the vector is associated by the predetermined clustering method, and the first centroid to which the newly added vector is associated according to the distance. By acquiring the correspondence information, it is possible to generate centroid information that enables efficient search.

  In addition, in the generation device 100 according to the embodiment, the acquisition unit 131 is a first centroid in which vectors are associated by a predetermined clustering method such that the difference in the number of associated vectors is within a predetermined value. Then, the correspondence information of the first centroid to which the newly added vector is associated according to the distance is acquired.

  As described above, the generating device 100 according to the embodiment is the first centroid in which the vectors are associated by the predetermined clustering method such that the difference in the number of associated vectors is within the predetermined value, By acquiring the correspondence information of the first centroid to which the vector to be added is associated according to the distance, it is possible to generate centroid information enabling efficient search.

  In addition, in the generation device 100 according to the embodiment, when the generation unit 132 generates a plurality of second centroids, each of the plurality of second centroids is associated with the first centroid, and each of the plurality of second centroids is included in the first vector group. Disassociate the vector from the first centroid.

  As described above, when the plurality of second centroids are generated, the generating device 100 according to the embodiment associates the plurality of second centroids with the first centroid, and the vectors included in the first vector group By disassociating one centroid, it is possible to generate centroid information that enables efficient search.

  In the generation device 100 according to the embodiment, the generation unit 132 corresponds to each of the plurality of second centroids when the number of vectors associated with each of the plurality of second centroids is equal to or less than a predetermined threshold. The attached vector is associated with the first centroid, and the association between the first centroid and the plurality of second centroids is released.

  Thus, when the number of vectors associated with each of the plurality of second centroids is equal to or less than a predetermined threshold, the generation device 100 according to the embodiment is associated with each of the plurality of second centroids. By associating the vector with the first centroid and disassociating the first centroid with the plurality of second centroids, centroid information that enables efficient search can be generated.

  In addition, in the generating device 100 according to the embodiment, the generating unit 132 deletes the plurality of second centroids.

  Thus, the generating device 100 according to the embodiment can generate centroid information that enables efficient search by deleting a plurality of second centroids.

  In addition, in the generating device 100 according to the embodiment, when the information regarding the second vector group associated with one second centroid among the plurality of second centroids satisfies the predetermined condition, the generation unit 132 may satisfy the following condition: A plurality of third centroids are generated in which each of the target vectors, which are vectors included in the second vector group, are associated according to the distance.

  As described above, when the information on the second vector group associated with one second centroid among the plurality of second centroids satisfies the predetermined condition, the generating device 100 according to the embodiment determines the second vector. By generating a plurality of third centroids in which each of the target vectors that are vectors included in the group are associated according to the distance, it is possible to generate centroid information that enables efficient search.

[10. Hardware configuration]
The generation device 100 according to the embodiment described above is realized by, for example, a computer 1000 configured as shown in FIG. FIG. 13 is a hardware configuration diagram illustrating an example of a computer that implements the function of the generation device. The computer 1000 includes a CPU 1100, a RAM 1200, a ROM (Read Only Memory) 1300, an HDD (Hard Disk Drive) 1400, a communication interface (I / F) 1500, an input / output interface (I / F) 1600, and a media interface (I / F). ) 1700.

  The CPU 1100 operates based on a program stored in the ROM 1300 or the HDD 1400 to control each part. The ROM 1300 stores a boot program executed by the CPU 1100 when the computer 1000 starts up, a program depending on the hardware of the computer 1000, and the like.

  The HDD 1400 stores a program executed by the CPU 1100, data used by the program, and the like. The communication interface 1500 receives data from another device via the network N, sends the data to the CPU 1100, and transmits data generated by the CPU 1100 to the other device via the network N.

  The CPU 1100 controls an output device such as a display or a printer and an input device such as a keyboard or a mouse via the input / output interface 1600. The CPU 1100 acquires data from an input device via the input / output interface 1600. The CPU 1100 also outputs the generated data to the output device via the input / output interface 1600.

  The media interface 1700 reads a program or data stored in the recording medium 1800 and provides the CPU 1100 with the program via the RAM 1200. The CPU 1100 loads such a program from the recording medium 1800 onto the RAM 1200 via the media interface 1700 and executes the loaded program. The recording medium 1800 is, for example, an optical recording medium such as a digital versatile disc (DVD) or a phase change rewritable disc (PD), a magneto-optical recording medium such as a magneto-optical disk (MO), a tape medium, a magnetic recording medium, or a semiconductor memory. Etc.

  For example, when the computer 1000 functions as the generation device 100 according to the embodiment, the CPU 1100 of the computer 1000 realizes the function of the control unit 130 by executing a program loaded on the RAM 1200. The CPU 1100 of the computer 1000 reads these programs from the recording medium 1800 and executes them, but as another example, these programs may be acquired from another device via the network N.

  Although some of the embodiments of the present application have been described in detail based on the drawings, these are only examples, and various modifications can be made based on the knowledge of those skilled in the art, including the aspects described in the rows of the disclosure of the invention. It is possible to implement the invention in other improved forms.

[11. Other]
Further, among the processes described in the above embodiment, all or part of the process described as being automatically performed may be manually performed, or the process described as being manually performed. All or part of them can be performed automatically by known methods. In addition, information including processing procedures, specific names, various data and parameters shown in the above-mentioned documents and drawings can be arbitrarily changed unless otherwise specified. For example, the various information shown in each figure is not limited to the illustrated information.

  Further, each component of each device illustrated is functionally conceptual, and does not necessarily have to be physically configured as illustrated. That is, the specific form of the distribution and integration of each device is not limited to the illustrated one, and all or a part thereof may be functionally or physically dispersed in any unit depending on various loads, usage conditions, etc. It can be integrated and configured.

  Moreover, it is possible to combine suitably each process described in each embodiment mentioned above in the range which does not contradict process content.

  In addition, the "section (module, unit)" described above can be read as "means" or "circuit". For example, the acquisition unit can be read as an acquisition unit or an acquisition circuit.

Reference Signs List 1 generation system 100 generation device 121 first index information storage unit 122 centroid information storage unit 123 second index information storage unit 130 control unit 131 acquisition unit 132 generation unit 133 extraction unit 134 provision unit 10 terminal device 50 information provision device N network

Claims (14)

An acquisition unit configured to acquire correspondence information indicating correspondence between a first centroid in which a vector related to a predetermined target is associated according to a distance, and a vector associated with the first centroid;
A target object that is a vector included in the first vector group when information on a first vector group associated with the first centroid satisfies a predetermined condition based on the correspondence information acquired by the acquisition unit A generation unit that generates a plurality of second centroids in which each of the vectors is associated according to the distance;
A generator comprising: The generation unit is
The generation apparatus according to claim 1, wherein the plurality of second centroids are generated when the number of target vectors is equal to or greater than a predetermined threshold. The generation unit is
The generation apparatus according to claim 1 or 2, wherein the plurality of second centroids are generated in a number according to the number of the target vectors. The generation unit is
The generation apparatus according to any one of claims 1 to 3, wherein the plurality of second centroids are generated based on the number of target vectors and a predetermined setting value. The generation unit is
The generating apparatus according to any one of claims 1 to 4, wherein the plurality of second centroids are generated by dividing the number of target vectors by a predetermined setting value. The acquisition unit
Obtaining the correspondence information of the first centroid, which is the first centroid associated with a vector by a predetermined clustering method, and in which a newly added vector is associated according to a distance. The production | generation apparatus of any one of Claims 1-5. The acquisition unit
The first centroid in which the vectors are associated by the predetermined clustering method such that the difference in the number of associated vectors is within a predetermined value, and the newly added vector is associated according to the distance The generating apparatus according to any one of claims 1 to 6, wherein the correspondence information of the first centroid is acquired. The generation unit is
When the plurality of second centroids are generated, the plurality of second centroids are associated with the first centroid, and the association between each vector included in the first vector group and the first centroid is generated. The generating device according to any one of claims 1 to 7, characterized in that it is released. The generation unit is
If the vector associated with each of the plurality of second centroids satisfies a predetermined condition, the vector associated with each of the plurality of second centroids is associated with the first centroid, and The generating device according to any one of claims 1 to 8, wherein correspondence between one centroid and the plurality of second centroids is released. The generation unit is
When the number of vectors associated with each of the plurality of second centroids is equal to or less than a predetermined threshold value, the vector associated with each of the plurality of second centroids is associated with the first centroid. The generation apparatus according to any one of claims 1 to 9, wherein the correspondence between the first centroid and the plurality of second centroids is released. The generation unit is
The generation device according to any one of claims 8 to 10, wherein the plurality of second centroids are deleted. The generation unit is
When information on a second vector group associated with one second centroid among the plurality of second centroids satisfies a predetermined condition, each of target vectors which are vectors included in the second vector group The generation device according to any one of claims 1 to 11, wherein a plurality of third centroids corresponding to the distance are generated. A computer-implemented generation method,
An acquiring step of acquiring correspondence information indicating correspondence between a first centroid to which a vector concerning a predetermined object is associated according to a distance, and a vector to be associated with the first centroid;
If the information on the first vector group associated with the first centroid satisfies a predetermined condition based on the correspondence information acquired in the acquisition step, an object that is a vector included in the first vector group Generating a plurality of second centroids in which each of the vectors is associated according to the distance;
A generation method characterized by including. An acquisition procedure for acquiring correspondence information indicating correspondence between a first centroid to which a vector related to a predetermined target is associated according to a distance, and a vector associated to the first centroid;
A target object that is a vector included in the first vector group when information on a first vector group associated with the first centroid satisfies a predetermined condition based on the correspondence information acquired by the acquisition procedure A generation procedure for generating a plurality of second centroids in which each of the vectors is associated according to the distance;
A generation program that causes a computer to execute.

Images