KR20030006638A - Apparatus And Method of Cell-based Indexing of High-dimensional Data - Google Patents

Abstract

PURPOSE: A cell based high dimensional data indexing system and method is provided to index high dimensional data based on a cell for preventing a lowering of a search efficiency in searching for high dimensional data. CONSTITUTION: The method comprises several steps. First, an N dimensional feature vector is extracted from a multimedia object via a feature vector extractor(801). A distance signature is generated via a signature generation module by using a distance between a signature on the feature vector and a cell center(802). One signature is generated by concatenating the feature vector signature and the distance signature(803), and then is stored at a signature database(804). At the same time, the feature vector is stored at a feature vector database(805). A user can perform a search operation on the stored feature vectors by using various queries like a point query, a range query or k-nearest query(806).

Description

Apparatus And Method of Cell-based Indexing of High-dimensional Data

The present invention relates to a cell-based high-dimensional data indexing apparatus, a method thereof, and a computer-readable recording medium having recorded thereon a program for realizing the method.

In general, feature vectors extracted from multimedia objects such as images and videos are used for content-based retrieval of multimedia objects. To this end, a number of high-dimensional data indexing techniques have been proposed to efficiently store high-dimensional feature vectors. In order to efficiently search for objects scattered in the data space, the existing high-dimensional indexing technique uses a minimum bounding rectangle (MBR) as a search unit. However, as the data dimension increases, the overlap region between the minimum boundary rectangles is enlarged, thereby causing a dimensional curse problem in which search performance decreases exponentially.

In other words, existing high-dimensional data indexing techniques use a minimum bounding rectangle (MBR) including objects to calculate similar distances between objects. However, as the dimension increases, the overlap occurs between the minimum bounding rectangles so that the search performance is degraded. As such, most of the existing high-dimensional data indexing techniques have excellent search performance for low-dimensional data of 10 dimensions or less, but have a problem in that the search performance decreases exponentially with increasing dimensions.

The present invention has been proposed to solve the above problems, and a cell-based high-dimensional data indexing apparatus for indexing high-dimensional data on a cell-based basis so that performance is not degraded in retrieving high-dimensional data and its It is an object of the present invention to provide a computer-readable recording medium storing a method and a program for realizing the method.

1 is a diagram illustrating an embodiment of a cell-based high-dimensional data indexing system to which the present invention is applied.

2 is a definition of a new minimum distance (MINDIST) and maximum distance (MAXDIST) in accordance with the present invention.

3 is a structural diagram of an embodiment of converting an N-dimensional vector into a signature according to the present invention;

4 is an exemplary structural diagram for the storage of signatures and vectors in accordance with the present invention;

5 is an exemplary structural diagram for signature and vector search according to the present invention.

6 is an exemplary processing diagram for a k-nearest query according to the present invention.

7 is an exemplary process for a range query to retrieve all objects included within a given range in accordance with the present invention.

8 is a flowchart illustrating an embodiment of a cell-based high-dimensional data indexing method according to the present invention.

* Explanation of symbols for the main parts of the drawings

11: Image 12: Object Saver

13: Image Database 14: Feature Vector Extractor

15 N-dimensional feature vector 16 Cell-based indexing device

17: user 101: signature generation module

102: storage module 103: concurrency control module

104: search module 105: signature database

106: Feature Vector Database

The present invention for achieving the above object, in the cell-based high-dimensional data indexing apparatus, using the N-dimensional feature vector extracted from the object, to obtain a cell to which the N-dimensional feature vector belongs and to display the cell Signature generation means for generating a distance signature from a center point of a cell to which a feature vector signature and the N-dimensional feature vector belong to the given N-dimensional feature vector and generating a merge signature in which the feature vector signature and the distance signature are merged. ; Storage means for storing the N-dimensional feature vector and the generated merge signature to correspond to each other; And concurrency control means for supporting a plurality of users by using a locking concept when storing and retrieving the N-dimensional feature vector and the merge signature.

On the other hand, the present invention, a high-dimensional data indexing method used in a cell-based high-dimensional data indexing apparatus, comprising: a first step of extracting an N-dimensional feature vector from a multimedia object through a feature vector extractor; Generating a distance signature using the feature vector signature and a distance value from a cell center to an object from the extracted N-dimensional feature vector; A third step of concatenating the generated feature vector signature and the distance signature into one signature to generate a merge signature; And a fourth step of storing the generated merge signature and the feature vector to correspond to each other.

On the other hand, the present invention provides a high-dimensional data indexing apparatus having a processor, comprising: a first function of extracting an N-dimensional feature vector from a multimedia object through a feature vector extractor; A second function of generating a distance signature using a feature vector signature and a distance value from a cell center to an object from the extracted N-dimensional feature vector; A third function of generating a merge signature by concatenating the generated feature vector signature and the distance signature into one signature; And a computer-readable recording medium having recorded thereon a program for realizing a fourth function of storing the generated merge signature and the feature vector in correspondence with each other.

Disadvantages of the conventional indexing technology supporting high-dimensional data retrieval are the problem of occurrence of overlapping area of the minimum boundary rectangle, which is a set of adjacent objects, and the inefficiency of the concept of distance between the object and the minimum boundary rectangle. In the present invention, by dividing the high-dimensional data space into cells and representing them as signatures, the overlap between the cells is eliminated and the new concept of distance between the object and the cell center is used to increase the filtering effect to maximize the search efficiency for the high-dimensional data.

The above objects, features and advantages will become more apparent from the following detailed description taken in conjunction with the accompanying drawings. Hereinafter, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings.

1 is a configuration diagram of an embodiment of a cell-based high-dimensional data indexing system to which the present invention is applied.

The cell-based high-dimensional data indexing system to which the present invention is applied extracts the feature vector of the object store 12 and the image 11 that stores the image 11 in the image database 13, as shown in FIG. Cell-based for indexing high-dimensional data according to the feature vector extractor 14, the identifier (ID) from the object store 12 and the N-dimensional feature vector 15 extracted through the feature vector extractor 14. A high dimensional data indexing device 16 is provided.

Referring to the cell-based indexing device 16 of the block shown in FIG. 1 in more detail, using a given N-dimensional feature vector 15, a cell to which the N-dimensional feature vector 15 belongs is obtained and this cell is obtained. Signature generation module for converting to a signature represented by a bit pattern of 1s and 0s for display and obtaining a distance value from the center point of the cell to which the N-dimensional feature vector 15 belongs to a given N-dimensional feature vector 15. (101), a storage module 102 for storing the N-dimensional feature vector 15 in the feature vector database 106 and the signature and distance values obtained above in the signature database 105, the N-dimensional feature vector ( 15) and at the time of storing and retrieving signatures, the concurrency control module 103 for supporting a plurality of users using a locking concept and the signature database 105 for user queries can satisfy the user's queries.Filter that there is no possibility, and with a search module 104 to find and output a feature vector and the identity of the object that satisfies the user's query from the feature vector database 106 about the end that is not filtered.

2 is a definition of the new minimum distance (MINDIST) and the maximum distance (MAXDIST) according to the present invention.

In the present invention, a new distance concept is defined and used to perform cell-based filtering. The new distance concept used in the present invention is the minimum distance (MINDIST) 202 and the maximum distance (MAXDIST) 201 shown in FIG.

In the present invention, before storing the object, the distance between the cell center and the object (RADIUS) is calculated and stored in advance. Using this value, the minimum and maximum distances from the user to the given query object (Q) are expressed as follows. >, <Equation 2> is used as defined.

The minimum distance (MINDIST) 202 defined in Equation 1 is the closest distance between the query object and the object stored in the cell, and in the distance between the query object and the cell center (CENTERDIST), the cell center is calculated and stored in advance. Minus the distance between the object and RADIUS.

The maximum distance (MAXDIST) 201 defined in Equation 2 is the farthest distance between the query object and the object stored in the cell, and the distance between the cell center and the object in the distance between the query object and the cell center (CENTERDIST). The distance of RADIUS is added.

By using the minimum and maximum distances obtained from the query objects, you can quickly search for the desired object through filtering without accessing all the objects stored in the database. That is, the newly defined minimum and maximum distances increase the filtering effect by optimizing and expressing an area of objects stored in one cell.

3 is a structural diagram of an embodiment of converting an N-dimensional vector into a signature according to the present invention.

In cell-based filtering, the data space is divided into cells, and each cell is represented by a signature to optimize main memory usage. In this case, the cell refers to a part formed as a result of dividing the interval, and the signature refers to a cell represented by a bit pattern of 1 and 0.

The feature vector of the object in the high dimensional space is converted into the signature of the cell containing the object and stored. In addition, in order to increase the filtering effect, the distance value from the cell center to the object is calculated, and this value is also converted into a signature and stored. 3 shows a process of generating such a signature. Equation 3 is used to convert a feature vector having a dimension of N into a signature.

Where b is the number of signature bits to allocate for each dimension of the feature vector

F is a feature vector with a value between 0 and less than 1.0

s is the signature generated

According to Equation 3, the total size of the N-dimensional signature for the N-dimensional feature vector is bits. In addition, the distance value from the cell center to the object is also expressed as a signature of 1 byte. The signature and the distance signature of the feature vector thus generated are merged into one signature and stored in the signature file.

4 is a structural diagram of an embodiment for storing a signature and a vector according to the present invention.

The merge signature 403 that merges the feature vector signature 401 and the distance signature 402 generated from the N-dimensional feature vector 41 by the signature generation module is sequentially stored in the signature file 42 through the storage module. .

However, if an operation such as deleting or updating an object occurs in the system, the reference file 44 is used to locate the empty record area in the current signature file 42 and store the signature there. .

After storing the signature, the feature vector of the actual object is stored in a data file 43 at the same location as the signature. In this way, by storing the signature and the feature vector in the same position, additional operations in the retrieval can be reduced.

5 is a structural diagram of an embodiment for signature and vector search according to the present invention.

Given a user query for retrieving stored objects, the query signature 501 is obtained from the user query vector 51 through the signature generation module. The signature file 52 is sequentially searched using the user query vector 51 and query signature 501 information. The data file 53 from the feature vector can be retrieved through the candidate cell list 54 obtained by sequentially searching the signature file 52.

At this time, filtering is performed using the minimum distance and the maximum distance newly defined in the present invention. By accessing only the signature record obtained by filtering and the corresponding data records, the desired result can be obtained. In other words, by using the newly defined minimum and maximum distances, unnecessary data access can be reduced, thereby improving search speed.

Hereinafter, the feature vector search method according to the present invention will be described in more detail with reference to FIGS. 6 to 7.

6 is an exemplary processing diagram for a k-nearest query according to the present invention.

The k-nearest query according to the present invention is a method for searching k objects most similar to a user's query. As shown in FIG. 6, the number of dimensions is 2, and it is assumed that a 2-bit signature is used for each dimension.

If the user query Q is given as (0.4, 0.2), first, all the signatures A, B, C, D, and E stored in the signature file are sequentially searched to obtain candidate cells. That is, filtering is performed by comparing the k-th maximum distance (MAXDIST) obtained so far and the minimum distance (MINDIST) of the cell currently being searched while sequentially searching each signature.

In more detail, the k-th maximum distance value obtained while searching for the signatures in FIG. 6 is the distance between the query point Q and the cell D. Thus, cells B, C, and E having a minimum distance greater than this value are excluded from the candidate cell, and cells A and D having a minimum distance less than this value are selected as candidate cells. For the selected candidate cells, the data file is accessed to compare the distance between the feature vector of the object and the query point Q and finally search for and return the object D (0.6, 0.4) having the closest distance.

Again, given the user query Q (0.4, 0.2) (k = 1), the signature files are searched sequentially to store the shortest maximum distance from Q (e.g., the portion of the cell to which the current D in FIG. 6 belongs) and Cells with a long minimum distance are filtered out and are no longer considered (eg, cells including B, C, E of FIG. 6). Therefore, at present, only cells containing A and D are considered, so they are boxed. In the actual data file, only two corresponding cells are selected as considerations and the nearest (1-NN) cell is selected. The final result is object D (0.6,0.4).

7 is an exemplary process for a range query to retrieve all objects included within a given range according to the present invention.

To process a range query that retrieves all objects contained within a given range according to the present invention, the signature file is first searched sequentially to find candidate cells. At this time, cells having a minimum distance value larger than a given distance value (radius) around the user query Q are excluded from the candidate list. That is, in FIG. 7, since the minimum distances of the objects A and E are larger than the radius from the center, they are excluded from the candidate list. The data file is accessed for the selected candidate cells. Finally, the objects B and C included in the given distance value (radius) are searched and reflected as the result value.

As shown in Fig. 7, given the user query Q (0.4, 0.2) (radius = O.3), the signature files are sequentially searched so that only cells within the radius range from the user query Q are considered and the radius range is considered. Outer cells are filtered and no longer considered (eg, cells containing A, E). Therefore, at present, only cells containing B, C, and D are considered, so they are boxed. In the actual data file, only three of them are considered and the range query is found. The dogs are (0.1,0.8) and (0.7,0.85).

8 is a flowchart illustrating an embodiment of a cell-based high-dimensional data indexing method according to the present invention.

First, an N-dimensional feature vector is extracted from a multimedia object through a feature vector extractor (801), and from the extracted N-dimensional feature vector to a signature and a cell center to an object through the signature generation module of the present invention. A distance signature using the distance value is generated (802).

The merged signature is generated by concatenating the generated feature vector signature and the distance signature into one signature (803), and the generated merge signature is stored in a signature database through a storage module (804). . In addition, the feature vector information of the multimedia object is stored in a feature vector database through the storage module (805).

Meanwhile, the user may perform a search for the feature vector of the stored object using various queries (that is, point query, range query, and k-nearest query) through the search module (806). .

As described above, the method of the present invention may be implemented as a program and stored in a recording medium (CD-ROM, RAM, ROM, floppy disk, hard disk, magneto-optical disk, etc.) in a computer-readable form.

The present invention described above is not limited to the above-described embodiments and the accompanying drawings, and various substitutions, modifications, and changes are possible in the art without departing from the technical spirit of the present invention. It will be apparent to those of ordinary knowledge.

The present invention as described above, in order to store the data in the high-dimensional space by dividing the data space into cells by the signature to prevent the overlap between the cells, and more effective filtering by using the distance value between the cell center and the stored object By doing so, it is possible to reduce the loss of performance even in the retrieval of high-dimensional data.

In addition, the present invention has the effect of increasing the efficiency of the search by a variety of search methods according to the user's selection.

Claims (15)

In the cell-based high-dimensional data indexing device, Using a N-dimensional feature vector extracted from an object, a feature vector signature for obtaining and displaying a cell to which the N-dimensional feature vector belongs and the N- given at the center of the cell to which the N-dimensional feature vector belongs Signature generation means for generating a distance signature to a dimensional feature vector and generating a merge signature combining the feature vector signature and the distance signature; Storage means for storing the N-dimensional feature vector and the generated merge signature to correspond to each other; And Concurrency control means for supporting a plurality of users using a locking concept when storing and retrieving the N-dimensional feature vector and the merge signature Cell-based high-dimensional data index device comprising a. The method of claim 1, Retrieval means for filtering out the possibility of satisfying the user's query using the stored signature for the user's query and finding and outputting the stored feature vector for the unfiltered Cell-based high-dimensional data index device further comprising. The method of claim 2, The search means, A cell-based high-dimensional data indexing device which searches for a signature that exactly matches a user's query and outputs a feature vector. The method of claim 2, The search means, A cell-based high-dimensional data indexing device which searches for a predetermined number of objects most similar to a user's query, and comprises a feature vector therefor. The method of claim 2, The search means, A cell-based high-dimensional data indexing device which searches for all objects included in a given range for a user's query, and features a feature vector therefor. The method according to any one of claims 1 to 5, The cell refers to a portion formed as a result of dividing the interval, and the signature is a cell-based high-dimensional data indexing device, characterized in that the representation of the cell in a binary bit pattern. The method according to any one of claims 1 to 5, The storage means, Feature vector storage means for storing the N-dimensional feature vector; And Signature storage means for storing the generated merge signature Cell-based high-dimensional data index device comprising a. In the high-dimensional data indexing method used in the cell-based high-dimensional data indexing apparatus, Extracting an N-dimensional feature vector from the multimedia object through the feature vector extractor; Generating a distance signature using the feature vector signature and a distance value from a cell center to an object from the extracted N-dimensional feature vector; A third step of concatenating the generated feature vector signature and the distance signature into one signature to generate a merge signature; And A fourth step of storing the generated merge signature and the feature vector to correspond to each other Cell-based high-dimensional data indexing method comprising a. The method of claim 8, A fifth step of performing a search using the stored signature with respect to a user's query and outputting the feature vector corresponding to the search result; Cell-based high-dimensional data indexing method further comprising. The method of claim 9, The fifth step, A sixth step of sequentially searching for the signatures stored in the fourth step from the user's query to find a signature that matches the user's query; And A seventh step of outputting a feature vector corresponding to a signature matching the generated user query found as the search result of the sixth step; Cell-based high-dimensional data indexing method comprising a. The method of claim 9, The fifth step, A sixth step of sequentially retrieving the signatures stored in the fourth step from the user's query to obtain candidate cells; Finding a cell having a maximum distance corresponding to a predetermined order from the user query from the candidate cells; An eighth step of filtering a cell having a minimum distance longer than the maximum distance of the cell found in the seventh step; A ninth step of finding the nearest cell among the remaining cells except the cells filtered in the eighth step; And A tenth step of outputting a feature vector corresponding to the cell of the ninth step; Cell-based high-dimensional data indexing method comprising a. The method of claim 9, The fifth step, A sixth step of sequentially searching for the signatures stored in the fourth step from the user's query and searching for cells included in the predetermined range from the user query and filtering cells longer than a predetermined range; A seventh step of finding cells within the fixed range among the remaining cells except the cells filtered in the sixth step; And An eighth step of outputting a feature vector corresponding to the cell of the seventh step Cell-based high-dimensional data indexing method comprising a. The method according to any one of claims 8 to 12, The cell refers to a portion formed by dividing the interval, and the signature is a cell-based high-dimensional data indexing method characterized in that the representation of the cell in a binary bit pattern. In a high-dimensional data indexing device having a processor, A first function of extracting an N-dimensional feature vector from the multimedia object via the feature vector extractor; A second function of generating a distance signature using a feature vector signature and a distance value from a cell center to an object from the extracted N-dimensional feature vector; A third function of generating a merge signature by concatenating the generated feature vector signature and the distance signature into one signature; And A fourth function of storing the generated merge signature and the feature vector to correspond to each other A computer-readable recording medium having recorded thereon a program for realizing this. The method of claim 14, A fifth function of performing a search using the stored signature with respect to a user's query and outputting the feature vector corresponding to the search result A computer-readable recording medium that records a program for further realization.