JP2001134593A - Method and device for neighborhood data retrieval and storage medium stored with neighborhood data retrieving program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a method and a device for close data retrieval, which can efficient ly retrieve neighborhood data by reducing random access to a disk and solving the problem that the bottom layer needs to be reached without fail so as to find the closest vector, and a storage medium stored with a neighborhood data retrieving program. SOLUTION: Feature variable vectors which are present in a multidimensional space are managed with unique identifiers, the multidimensional space is divided hierarchically into partial spaces, and representative vectors of the respective partial spaces are extracted from feature variable vector groups present in the divided partial spaces; and the extracted representative vectors of the partial spaces are managed for each layer with unique identifiers, the distances between the representative vectors and the feature variable vector of a retrieval key are calculated, and according to the calculated distances, a feature variable vector which is similar to the feature variable vector of the retrieval key is outputted as a retrieval result as to a subset of feature variable vectors present in the multidimensional space which are narrowed down with the representative vectors.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a neighborhood data search method and apparatus and a storage medium storing a neighborhood data search program, and in particular, when a search key is given, k pieces of similar data having the highest similarity are obtained. The present invention relates to a method and an apparatus for retrieving nearby data that can be returned at high speed, and a storage medium that stores a neighboring data search program.

[0002]

2. Description of the Related Art For example, in the Transactions of the Information Processing Society of Japan: Database, Vol. 40, No. SIG3 (TOD1), pp. 171-184, "Image Content Retrieval Technology Based on Surface Features such as Colors and Shapes" , Expressing the color and shape features extracted from each image data as a vector in a multidimensional space (multidimensional vector),
It is stored in the database as a feature vector of image data. At the time of the search, the distance between the feature vector of the search key image extracted from the given search key image and the feature vector of each image stored in the database is calculated, and an image having a short distance is similar. Image data retrieval is realized by using images.

In such a system, the distance between a feature vector expressing a search key and a feature vector expressing each stored data is a similarity measure.
Since the calculation amount of the distance calculation increases, many methods for reducing the distance calculation amount have been proposed. One of them is “Similar feature retrieval method and apparatus and similar feature retrieval program” disclosed in Japanese Patent Application No. 11-229559.
There is a method. In this method, each feature vector in the multidimensional space is used as a search key in advance and distance calculation is performed with respect to all remaining feature vectors in the multidimensional space. A feature quantity vector which becomes similar data when used as a search key is held as a predetermined number of cases and a pre-search result list. By recalculating the distance between the feature vector included in the pre-search result list and the feature vector of the search key, starting from the feature vector (quasi-neighbor vector) similar to the given search key at the time of search, Build search results. The search result is corrected by applying this processing in a chained manner. Therefore, if the starting point is a feature amount vector (nearest neighbor vector) closest to the search key, there is no need for correction. If the starting point is a feature vector that is far from the nearest neighbor vector, it may not be possible to correct the nearest neighbor vector even if the correction process is performed in a chain. For this reason,
The feature vector serving as a starting point has an important meaning.

As one method for obtaining a feature vector serving as a starting point, for example, DA White, and R. Jain: "Simila
rity Indexing: Algorithms and Performance ", In Pro
ceedings of the SPIE.Storage Retrieval for Image
and Video Database IV, 1996, San Jose, CA, Vol.267
0, pp.62-75. There is a tree-like indexing method proposed in pp.62-75. This method measures the variance of data for each dimension axis unit for a data group of feature quantity vectors existing in a multidimensional space,
With respect to the axis having the largest variance value, the data is divided into two by the approximate median value of the spread of all data on the selected axis. By sequentially applying this processing to the divided subspaces, a multidimensional space is hierarchically constructed in an n-ary tree.

[0005]

However, in such a tree-like indexing method, the distance between the feature vector of the retrieval key and the subspace (multidimensional rectangle) of the multidimensional space hierarchically formed is determined. As a measure of the degree of similarity, the search proceeds to lower layers, but there is no guarantee that the nearest neighbor vector exists in the multidimensional rectangle closest to the search key in each layer. In addition, the multidimensional rectangle of the middle layer holds only boundary information indicating a subspace of the multidimensional space.

Therefore, in each hierarchy, it is not possible to simply follow the closest hierarchy immediately below the closest multi-dimensional rectangle based on the distance between the feature vector of the search key and the multi-dimensional rectangle existing in the hierarchy. It is necessary to traverse the hierarchy as a whole while leaving candidate multidimensional rectangles that are complete and may have the nearest neighbor vector. In particular, in a high-dimensional multidimensional space,
The number of multidimensional rectangles remaining as candidates increases. Since the information of the feature amount vector is not stored in the multidimensional rectangle of the middle layer, there is a problem that it is necessary to always go to the lowest hierarchy in order to find the nearest feature amount vector. .

In general, such a tree-like index is placed on a disk. However, when the number of data items is large, access to the disk becomes random.
The present invention has been made in view of the above points, and solves a problem that it is necessary to always go to the lowest hierarchy in order to reduce random access to a disk and obtain a nearest neighbor vector. It is an object of the present invention to provide a nearby data search method and apparatus capable of searching data and a storage medium storing a nearby data search program.

[0008]

FIG. 1 is a diagram for explaining the principle of the present invention. According to the present invention (claim 1), the feature quantity of each data is represented as a feature quantity vector in a multidimensional space, and the top k search results are returned in order of decreasing distance from the feature quantity vector represented by the search key. In a neighborhood data search method, a feature vector existing in a multidimensional space is managed by a unique identifier (step 1), and the multidimensional space is hierarchically divided into subspaces (step 2). A representative vector representative of the subspace is extracted from the feature amount vector group existing in the extracted subspace (step 3), and the extracted representative vector of the subspace is managed by a unique identifier for each layer. (Step 4) Calculate the distance between the representative vector and the feature vector of the search key (Step 5), and in the multidimensional space narrowed down by the representative vector based on the calculated distance. Targeting a subset of feature vector are present, similar feature vectors and the feature amount vector of the search key, and outputs as a search result (step 6).

According to the present invention (claim 2), a center point of all points in the subspace indicated by a feature vector existing in the subspace to be searched is determined, and a point closest to the center point is determined in the subspace. Let the representative vector be The present invention (claim 3)
The center of gravity is used instead of the center point. According to the present invention (claim 4), instead of using all the feature amount vectors existing in the target subspace for the feature amount vector to be calculated for the center point or the center of gravity and the feature amount vector to be the representative vector, Then, only the representative vectors extracted in the hierarchy immediately below the subspace are targeted.

The present invention (claim 5) manages a plurality of representative vectors in each layer as a representative vector management list for each layer, and when a request time for a search response is given, a representative vector management list for a plurality of layers. Of the representative vector management lists to be used is determined based on the number of representative vectors included in the representative vector management list and the request time for the search response.

FIG. 2 is a diagram showing the principle of the present invention. According to the present invention (claim 6), the feature amount of each data is represented as a feature amount vector in a multidimensional space, and the top k search results are returned in the order of shortest distance from the feature amount vector represented by the search key. Feature data vector management device 11 for managing a feature vector existing in a multidimensional space with a unique identifier, and a space dividing device for hierarchically dividing the multidimensional space as a subspace. 12 and
A representative vector extracting unit for extracting a representative vector representing the subspace from a feature amount vector group existing in the subspace divided by the space dividing unit; Representative vector management means 14 for managing each of them by a unique identifier, a distance calculation means 15 for calculating a distance between a representative vector managed by the representative vector management means 14 and a feature amount vector of a search key.
Based on the distance calculated by the distance calculating means 15, a feature vector similar to the feature vector of the search key is set for a subset of the feature vectors existing in the multidimensional space narrowed down by the representative vector. , Neighborhood search,
And a search result management means 16 for returning search results.

According to the present invention (claim 7), the feature amount of each data is represented as a feature amount vector in a multidimensional space.
A storage medium that stores a neighborhood data search program that returns the top k search results in the order of distance from the feature vector expressed by the search key, and stores a feature vector existing in a multidimensional space, A feature vector management process for managing with a unique identifier, a space division process for hierarchically dividing a multidimensional space as a subspace, and a feature vector group existing in the subspace divided by the space division process means, A representative vector extraction process for extracting a representative vector representative of the subspace, a representative vector management process for managing a representative vector of the subspace extracted in the representative vector extraction process with a unique identifier for each layer, A distance calculation process that calculates the distance between the representative vector managed in the management process and the feature vector of the search key And, based on the distance calculated in the distance calculation process, for a subset of the feature vector existing in the multidimensional space narrowed down by the representative vector, a feature vector similar to the feature vector of the search key is obtained. And a search result management process for performing a proximity search and returning as a search result.

According to the present invention (claim 8), in the representative vector extraction process, the center points of all the points in the subspace indicated by the feature quantity vectors existing in the subspace to be searched are obtained, and the most significant points are determined from the center points. The process includes making a near point a representative vector of the subspace. According to the present invention (claim 9), in the representative vector extraction process, instead of the center point,
Includes a process that uses the center of gravity.

According to a tenth aspect of the present invention, in the representative vector extraction process, the feature amount vector for which the center point or the center of gravity is to be calculated and the feature amount vector to be the representative vector exist in the target subspace. Instead of using all feature amount vectors, only the representative vectors extracted in the layer immediately below the subspace are targeted.

The present invention (Claim 11) provides a process for managing a plurality of representative vectors in each hierarchy as a representative vector management list for each hierarchy, and a method for managing a plurality of representative vectors in a plurality of hierarchies when a request time for a search response is given. Of the management list, based on the number of representative vectors included in the representative vector management list and the request time for the search response,
Determining which hierarchical representative vector management list to use.

In the above, the feature vector management means manages feature vectors existing in the multidimensional space by using unique identifiers. Therefore, it is possible to uniquely specify all data including the same point. The space dividing means constructs a hierarchical subspace (multidimensional rectangle) from the multidimensional space including all feature vectors based on the number of divisions in each hierarchy, the distance between feature vectors, and the like. Therefore, it is possible to manage a plurality of feature amount vectors that are close to each other to some extent with one multidimensional rectangle. In addition, since the multidimensional rectangle is hierarchically constructed, it is possible to construct a set of representative vectors having a plurality of granularities (representative vector group).

The representative vector extracting means extracts a feature vector (representative vector) which is representative of the subspace from all the feature vectors existing in the subspace. Therefore, it is possible to select a feature vector to be searched. The representative vector management means manages the representative vectors of the subspaces at each hierarchical level extracted by the representative vector extracting means using a unique identifier for each hierarchical level and stores them in a continuous area. Therefore, it is possible to uniquely identify a feature amount vector to be searched for,
A plurality of feature amount vectors to be searched can be collectively read from the storage area.

The distance calculation means calculates the distance between the feature vector of the search key and the feature vector managed by the representative vector management means. For this reason, it is possible to create a reference value that is the similarity of the search result. The search result management means calculates a feature amount vector sequentially read out from the representative vector management means, and uses a distance calculation means to calculate a brute force distance with the feature amount vector of the search key, and according to the obtained distance, Is sorted and returned.

Therefore, it is possible to store and manage in a continuous area a representative vector, which is a subset extracted from all the feature quantity vectors existing in the multidimensional space without distorting the bias, in the continuous area. By sequentially calculating the distance between the search key and the feature amount vector, it is possible to perform a neighborhood search with the search key feature amount vector, which is an object of the present invention.

[0020]

DESCRIPTION OF THE PREFERRED EMBODIMENTS In the following description, the present invention
Although it is a kind of neighborhood search, since the search target is narrowed down by the representative vector, the accuracy including the nearest neighbor is not necessarily guaranteed, so such a search is referred to as a quasi-neighbor search. Therefore, the device described below is referred to as a quasi-neighbor data search device.

FIG. 3 shows the configuration of the quasi-neighbor data retrieval apparatus of the present invention. The quasi-neighbor data search device shown in FIG. 1 includes a feature vector management unit 11, a space division unit 12, a representative vector extraction unit 13, a representative vector management unit 14, and a distance calculation unit 1.
5 and a search result management unit 16. The feature vector management unit 11 manages feature vectors that exist in a multidimensional space using unique identifiers.

The space division unit 12 constructs a hierarchical subspace (multidimensional rectangle) from a multidimensional space including all feature vectors based on the number of divisions in each hierarchy, the distance between feature vectors, and the like. I do. The representative vector extraction unit 13 extracts a feature vector (representative vector) that is representative of the subspace from all the feature vectors existing in the subspace.

The representative vector management unit 14 manages the representative vectors of the subspaces at each layer extracted by the representative vector extraction unit 13 using a unique identifier for each layer,
Store in a continuous area. The distance calculation unit 15 calculates the distance between the feature vector of the search key and the feature vector managed by the representative vector management unit.

The search result management unit 16 calculates a feature amount vector sequentially read from the representative vector management unit 14 and a feature amount vector of the search key by using the distance calculation unit 15 to calculate a brute force distance. According to the distance, the feature amount vectors of the number of items to be returned are sorted and returned. next,
The operation of the quasi-neighbor data search device in the above configuration will be described.

FIG. 4 is a flowchart of the operation of the near-neighbor data search device of the present invention. Step 101) The feature amount vector management unit 11 manages the feature amount vector in advance using a unique identifier. A search key is input to the near-neighbor data search device. Step 102) In the space dividing unit 12, the multidimensional space is hierarchically divided as a subspace.

Step 103) Representative vector extraction unit 1
In step 3, a feature vector (representative vector), which is representative of the space of the part, is extracted from a group of feature vectors existing in the partial space divided in step 102. Step 104) The representative vector management unit 14 manages the representative vectors extracted in step 103.

Step 105) The distance calculation unit 15 calculates the distance between the input characteristic amount vector of the search key and the representative vector managed by the representative vector management unit 14. Step 106) Based on the calculated distance, the search result management unit 16 targets a subset of the feature vectors existing in the multi-dimensional space narrowed down by the representative vector and searches for a feature similar to the feature vector of the search key. The quantity vector is obtained by neighborhood search.

Step 107) The searched feature amount vector is output.

[0029]

Embodiments of the present invention will be described below with reference to the drawings. In the feature vector management unit 11, for example, each feature vector is stored in a file, and the feature vector management unit 11 manages the feature vector using a unique identifier.

The space dividing unit 12 divides a multidimensional space into a plurality of hierarchical subspaces by using a tree-like index construction method.
FIG. 5 shows an example in which a two-dimensional space is divided into two layers by applying one tree-like index in the space dividing unit according to one embodiment of the present invention. In the figure, reference numerals 21 and 22 denote partial spaces in the first hierarchy, and reference numerals 211, 212, 221, and 222 denote subspaces.
7 shows a subspace in the second hierarchy. At the time of division, the upper limit of the number of branches immediately below is set to 2 (binary tree). In the example shown in the figure, rectangles 21 and 22 are partial spaces at the first level (the root layer is the 0th layer). Also, 211,
212, 221 and 222 are subspaces at the second level.

FIG. 6 shows a tree-like index constructed in this manner. This figure schematically shows a final tree-like index obtained by spatially dividing the feature vector in FIG. In a conventional search method using a tree-like index, when a search key is given, 21, 22, 211,
.. And the candidate subspace are calculated, and the search is performed while sorting the subspaces by the distance.

The representative vector extracting unit 13 extracts a representative vector from the divided subspace. For example, in the second hierarchy in FIG. 6, the center points of all the points in the subspace indicated by the feature amount vectors existing in the target subspace are obtained, and the point closest to the center point is defined as the representative vector of the subspace. In the subspace 211, the feature vector b or the subspace 221 is used as the feature vector close to the center point among the feature vectors a, b, and c in the subspace 211.
Then, the feature vector g is extracted from the feature vectors g, h, and i.

On the other hand, if the method using the center of gravity instead of the center point is applied, the feature vector b
However, in the subspace 221, a feature amount vector i is extracted. The representative vector management unit 14 includes the representative vector extraction unit 1
3 is stored for each layer in the same manner as the feature vector management unit 11 for each layer, for example, as a set of a feature vector identifier and a value of each dimension in a file. to manage.

FIG. 7 shows an example of a representative vector management list of the representative vector management unit according to one embodiment of the present invention. The representative vector management list shown in the figure is obtained by calculating the center points of all the points in the subspace in the feature vector set in FIG. 5, and setting the point closest to the center point as the representative vector of the subspace. It schematically shows that a feature amount vector is managed for each layer. FIG. 9A shows a management list of representative vectors of the first hierarchy, and FIG.
5 shows a second-layer representative vector management list.

The search result management unit 16 first determines the required search response time for the feature vector of the given search key and the number of feature vectors included in the representative vector management list of each hierarchy. Then, a hierarchy vector representative list to be used is determined (for example, the second hierarchy in FIG. 6 is selected). Next, in the distance calculation unit 15,
The distance calculation is performed with respect to all the feature amount vectors (each of the representative vectors 211, 212, 221 and 222 in FIG. 6) managed by the representative vector management list from the feature amount vector of the search key. At the same time as calculating the distance,
A search result list is constructed and managed by sorting the search result list in the search result management unit 16.

The quasi-neighborhood search result constructed in this way is returned. Although the above embodiment has been described based on the configuration of FIG. 3, a disk device connected to a computer used as a quasi-neighbor data search device by constructing the components shown in FIG. The present invention can be easily realized by storing the program in a portable storage medium such as a floppy (registered trademark) disk, a CD-ROM, or the like, and installing the program when implementing the present invention.

The present invention is not limited to the above embodiment, but can be variously modified and applied within the scope of the claims.

[0038]

As described above, according to the present invention, a representative vector, which is a subset extracted from all feature quantity vectors existing in a multidimensional space without losing a bias, is stored in a continuous area. This makes it possible to manage, and by sequentially obtaining the distance between the representative vector and the feature vector of the search key, it is possible to perform a quasi-neighbor search with the feature vector of the search key.

[Brief description of the drawings]

FIG. 1 is a diagram for explaining the principle of the present invention.

FIG. 2 is a principle configuration diagram of the present invention.

FIG. 3 is a configuration diagram of a quasi-neighbor data search device of the present invention.

FIG. 4 is a flowchart of the operation of the quasi-neighbor data search device of the present invention.

FIG. 5 is an example in which a two-dimensional space is divided into two layers by applying one tree-like index in a space dividing unit according to an embodiment of the present invention.

FIG. 6 is an example of a tree-like search according to an embodiment of the present invention.

FIG. 7 is an example of a representative vector management list of a representative vector management unit according to an embodiment of the present invention.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 11 Feature-value vector management means, feature-vector management part 12 Space division means, space division part 13 Representative vector extraction means, representative vector extraction part 14 Representative vector management means, representative vector management part 15 Distance calculation means, distance calculation part 16 Search Result management means, search result management unit

 ────────────────────────────────────────────────── ─── Continuing on the front page (72) Inventor Fumika Konishi 2-3-1, Otemachi, Chiyoda-ku, Tokyo F-term (reference) in Nippon Telegraph and Telephone Corporation 5B075 ND35 NR15 PP22 PQ74 PR06 QM08 5L096 DA02 FA60 FA62 FA66 HA09 JA11 KA09 MA07

Claims (11)

[Claims] 1. A feature amount of each data is expressed as a vector in a multidimensional space (hereinafter, referred to as a feature amount vector), and the top k search results are arranged in ascending order of a distance from the feature amount vector represented by the search key. In the neighborhood data search method to be returned, a feature vector existing in the multidimensional space is managed by a unique identifier, and the multidimensional space is hierarchically divided into subspaces. A representative vector representing the subspace is extracted from the feature amount vector group existing in the subspace. The representative vector of the extracted subspace is managed by a unique identifier for each layer. The distance between the feature vectors of the feature vectors in the multidimensional space narrowed down by the representative vector is calculated based on the calculated distance. A method of searching for neighboring data, comprising: outputting, as a search result, a feature vector similar to a feature vector of a search key to an elephant. 2. The method according to claim 1, wherein a center point of all points in the subspace indicated by the feature vector existing in the subspace to be searched is determined, and a point closest to the center point is set as a representative vector of the subspace. 1. A method for searching for nearby data according to item 1. 3. The neighborhood data search method according to claim 2, wherein a center of gravity is used instead of said center point. 4. For the feature amount vector for calculating the center point or the center of gravity and the feature amount vector serving as the representative vector, instead of using all the feature amount vectors existing in the target subspace, 4. The neighborhood data search method according to claim 2, wherein only the representative vector extracted in the hierarchy immediately below the subspace is targeted. 5. A plurality of representative vectors in each layer are managed as a representative vector management list for each layer, and when a request time for a search response is given, the representative vector management list in the plurality of layers is displayed. 2. The neighborhood data search method according to claim 1, wherein a hierarchy vector representative list to be used is determined based on the number of representative vectors included in the list and a request time for the search response. 6. A feature amount of each data is represented as a vector in a multidimensional space (hereinafter, referred to as a feature amount vector), and the top k search results are arranged in ascending order of a distance from the feature amount vector represented by the search key. A neighboring data search device that returns a feature vector managing unit that manages a feature vector existing in a multidimensional space with a unique identifier, and that hierarchically divides the multidimensional space into subspaces Space dividing means; representative vector extracting means for extracting a representative vector representative of the subspace from a group of feature vectors existing in the subspace divided by the space dividing means; representative of the extracted subspace A representative vector management unit that manages a vector with a unique identifier for each layer; a representative vector managed by the representative vector management unit; Distance calculating means for calculating the distance of the collection vector, based on the distance calculated by the distance calculating means, for a subset of feature amount vectors existing in a multidimensional space narrowed down by the representative vector, A neighborhood data search device comprising: a search result management unit that performs a neighborhood search for a feature amount vector similar to a feature amount vector of a search key and returns the search result as a search result. 7. A feature amount of each data is represented as a vector in a multidimensional space (hereinafter, referred to as a feature amount vector), and the top k search results are arranged in ascending order of distance from the feature amount vector represented by the search key. A storage medium storing a neighborhood data search program to be returned, wherein a feature amount vector management process for managing a feature amount vector existing in a multi-dimensional space by a unique identifier; and A space division process of dividing as a subspace; a representative vector extraction process of extracting a representative vector representative of the subspace from a feature vector group existing in the subspace divided by the space division process unit; A representative vector management program that manages the representative vector of the subspace extracted by the representative vector extraction process with a unique identifier for each layer A distance calculation process for calculating a distance between the representative vector managed in the representative vector management process and a feature amount vector of a search key; and the representative vector based on the distance calculated in the distance calculation process. And a search result management process of performing a neighborhood search for a feature vector similar to the feature vector of the search key for a subset of the feature vectors existing in the multidimensional space narrowed down by the search key and returning the search result as a search result. A storage medium storing a neighborhood data search program characterized by having 8. The representative vector extraction process calculates a center point of all points in the subspace indicated by a feature vector present in the subspace to be searched, and determines a point closest to the center point in the subspace. 8. A storage medium storing the neighborhood data search program according to claim 7, including a process of setting a representative vector of the data. 9. The storage medium according to claim 8, wherein the representative vector extracting process includes a process using a center of gravity instead of the center point. 10. The representative vector extraction process includes, for the feature amount vector for which the center point or the center of gravity is to be calculated, and the feature amount vector to be the representative vector, which are all existing in the target subspace. 10. The storage medium storing the neighborhood data search program according to claim 8, wherein instead of using a feature amount vector, only a representative vector extracted in a layer immediately below the subspace is targeted. 11. A process of managing a plurality of representative vectors in each layer as a representative vector management list for each layer, and when a request time for a search response is given, the representative vector management list of the plurality of layers 8. A process for deciding which hierarchical representative vector management list to use based on the number of representative vectors included in the vector management list and the request time for the search response.
A storage medium storing the described neighborhood data search program.