JP5014398B2 - Search data management device - Google Patents

Description

  The present invention relates to an apparatus for retrieving similar data from a database and registering new data in the database.

  A feature amount extracted from an image, sound, or the like is represented by multidimensional vector data, and similar data is searched using a distance between the vector data. When searching for multi-dimensional vector data, if the number of search target data is large, the distance between the search key data (search data) and the search target data must be calculated one by one. It will be long.

  Patent Document 1 describes the following technique. That is, for all feature quantities in the database, the similarity with other feature quantities is calculated in advance, and data with high similarity is stored in association with each other. Then, the nearest neighbor of the given search key is obtained, and similar data corresponding to the nearest neighbor data is returned as a search result.

JP 2001-52024 A

  By the way, when new data is to be registered in a data group (database) to be searched, an index considering the degree of similarity between the data is assigned in advance, thereby ensuring high speed search. For this reason, it is necessary to search similar data for the registered data from the database.

  However, in the technique of Patent Document 1, in order to speed up the search, it is necessary to store all data in order of similarity for each data, but it is necessary to pre-calculate the similarity to all data in the database. When the search target data is enormous, the time required for the search becomes longer depending on the number of data. Also, every time data is registered, it is necessary to calculate the similarity between the data and review the index. Therefore, as the number of data increases, the construction of the index becomes complicated.

  The present invention has been made in view of the above-described problems, and an object of the present invention is to register data at high speed without complicating a search index in a data group to be searched for similar data. It is.

In order to achieve the above object, the search data management device according to the first invention provides:
A database in which a link for tracing from one data to the other data is set between the data and a plurality of the data is stored;
For the input search target data, select any one of the data stored in the database as a search start point, and sequentially follow the link from the search start point, so that the path following the link Graph search means for searching for and outputting data within a predetermined distance from data to be searched as similar data among certain data,
Similar data acquisition means for acquiring similar data for the registration target data by inputting the data to be registered in the database to the graph search means;
A link that sets a link between the data to be registered and the similar data acquired by the similar data acquisition unit, stores the link in the database, and performs a process of reducing the links set in the similar data Optimization means;
It is characterized by having.

In the second invention, the link optimization means includes:
When the number of links set in the similar data is equal to or greater than a predetermined number, the predetermined number of the links are held in order from the shortest distance between the data set with the links, and other links are released. It is said.

In the third invention, the link optimization means includes:
From the other data linked to the data to be reduced in the link, the holding data for holding the link and the release data for releasing the link are identified,
Based on the distance between the release data and the held data, a link is set between the release data and the held data and stored in the database.

In the fourth invention, the graph search means includes:
Selecting arbitrary data from the database, searching the data within the predetermined distance from the registration target data by sequentially following the link from the data, and selecting the data as a search start point It is said.

In the fifth invention, the graph search means includes:
As a result of following the link from the arbitrary data, if there is no data within the specified distance in the data on the path following the link, data different from the selected arbitrary data is displayed. The search start point is searched by performing a process of newly selecting and sequentially following the links again.

In the sixth invention,
The data stored in the database is further hierarchized by space division, further comprising a hierarchical database for storing for each hierarchy to which the data belongs,
The graph search means includes:
The hierarchy to which the data to be registered belongs is specified, and other data belonging to the specified hierarchy is selected as the search start point.

In the seventh invention, the graph search means includes:
If the distance between the search start point selected based on the hierarchical database and the data to be registered is more than the specified distance, select any data from the database and link to the data. By searching, data within the specified distance is searched and selected as the search start point.

According to an eighth aspect of the present invention, there is provided search data management in which a computer for managing a database in which a plurality of data is stored by setting a link for tracing from one data to the other data as a data group to be searched. In the method
For the input search target data, select one of the data stored in the database as a search start point, and sequentially follow the link from the search start point to A graph search step of searching and outputting data within a predetermined distance from the data to be searched as similar data among the data in
A similar data acquisition step of causing the data to be registered in the database to be searched in the graph search step as the search target, and acquiring similar data for the data to be registered;
A link that sets a link between the data to be registered and the similar data acquired in the similar data acquisition step, stores the link in the database, and performs a process of reducing the links set in the similar data An optimization process;
Is performed by the computer.

  According to a ninth aspect of the invention, there is provided a program for causing a computer to execute processing of a search data management method.

  According to the present invention, when registering data in a database, similar data within a specified distance is searched from data on a route that follows a link set between the data. Since it is not necessary to calculate the distance to the data and similar data is searched from the data group at high speed, data registration can be performed at high speed. In addition, when registering data in the database, the number of links in which similar data is set can be reduced, so that it is possible to sequentially prevent the link structure between data from becoming complicated due to data registration. It is possible to prevent the decrease.

[Apparatus Configuration of First Embodiment]
A search device according to a first embodiment to which the present invention is applied will be described with reference to the accompanying drawings. The contents of the present invention are not limited to the present embodiment, and various modifications can be made to the specific configuration within the scope described in the claims.

  In the present embodiment, an example will be described in which multidimensional vector data representing the feature amount of an image is handled as vector data. However, the vector data may represent features of voice and other multimedia data information, or may be other types of multidimensional data.

  FIG. 1 is a block diagram showing a functional configuration of a search device 1 which is an example of the present invention. The search device 1 does not need to be configured by a single piece of hardware or software. If necessary, a function corresponding to the search device 1 can be provided by a combination of a plurality of hardware or software. The search device 1 of this embodiment can also be configured by a plurality of servers scattered on the network.

  The search device 1 performs data processing in response to various requests transmitted from the client terminal, and returns the processing result to the client terminal. Specifically, when receiving a search query that is a search request, the search device 1 performs a search according to the search query and returns a search result to the client terminal. When a request for registering new data in a database (hereinafter abbreviated as “DB”) that stores data to be searched is received, it is registered in the DB.

[Client terminal configuration]
Although illustration of the client terminal is omitted, the client terminal is configured by a personal computer including an input unit such as a keyboard, a mouse, and a touch pad, an output unit such as a display and a printer, and a CPU. The client terminal is connected to the search device 1 via a network. Further, the client terminal sends search data and registration data designated by the user to the search device 1.

  Here, the search data is data serving as a key for searching for similar data from data stored in the database of the search device 1. Registration data is data newly registered in the database of the search device 1. The vector data extracted from the search data and registration data is also referred to as “search data” and “registration data”.

[Configuration of search device]
According to FIG. 1, the search device 1 includes a registration data input unit 10, a search data input unit 20, a search result output unit 30, a graph search unit 40, a data registration unit 50, an image DB 60, a graph index. DB70 is comprised.

  The registration data input unit 10 is a functional unit that receives registration data transmitted from a client terminal. When registration data is received together with a data registration request, the registration data input unit 10 inputs the registration data to the data registration unit 50.

  The search data input unit 20 is a functional unit for receiving the search data transmitted from the client terminal. When the search data is received together with the search request, the search data input unit 20 inputs the search data to the graph search unit 40.

  The search result output unit 30 is a functional unit for transmitting the search result by the graph search unit 40 to the client terminal according to the input of the search data input unit 20. The search result output unit 30 generates display data for displaying the search result and transmits it to the client terminal.

  The image DB 60 is a database that stores and stores image data to be searched, and stores an image ID, image data, and vector data in association with each other as shown in FIG. The graph search unit 40 extracts and stores multidimensional vector data as a plurality of feature amounts from the image data stored in the image DB 60. The number of dimensions of the vector data is not particularly limited, but is preferably set to a high number of dimensions (for example, 10 dimensions or more) in order to improve search accuracy.

  The graph index DB 70 is a database that stores an index having a graph structure, and stores vector data representing feature amounts of data to be searched. The data registration unit 50 generates a graph structure index based on the vector data extracted from the image data.

  Further, the graph index DB 70 stores information on links set to trace from one vector data to the other vector data. Specifically, as shown in FIG. 1, the graph index DB 70 displays an image ID that is a link source corresponding to the image DB 60 and one or more other link destination image IDs linked to the vector data. Store in association with each other. By establishing a link between the vector data, a graph structure is formed between the vector data a to f in the vector space as shown in FIG.

  The graph structure formed by the link information in the graph index DB 70, that is, the link information is appropriately referred to as “graph index”. A link is information that can trace one data to another data. The link may be unidirectional, but as shown in FIG. 1, the data indicating one link destination and the data indicating the link destination in the opposite direction are stored as a pair and can be traced bidirectionally. Thus, the data search speed can be improved.

  In FIG. 4A, vector data is represented by points (a to f), and shows a state in which vector data is distributed in a multidimensional feature amount space. A link is set between each vector data, and the link is represented by a solid line. In the following description, vector data will be referred to as “points” as appropriate.

  The graph search unit 40 searches for data similar to the search data by circulating the links set in the graph index DB 70. As shown in FIG. 1, the graph search unit 40 includes a vector data generation unit 41, a search start point determination unit 43, a graph circulation unit 45, and a similar data specification unit 47.

  The vector data generation unit 41 extracts feature quantities that become multidimensional vector data from the data input from the search data input unit 20 or the data registration unit 50.

  The search start point determination unit 43 determines the existing vector data on the graph index close to the vector data extracted from the search data as the search start point. The graph search unit 40 determines any one of the plurality of vector data as a search start point, and starts the link circulation starting from the search start point.

  The graph circulator 45 performs a process of sequentially following the links set in the graph index DB 70 from the search start point until a predetermined search end condition is satisfied.

  The similar data specifying unit 47 specifies, as similar data, vector data within a predetermined distance with respect to the search data among vector data on the link route followed by the graph circulating unit 45. The identified similar data or related information is sent to the client terminal as a search result.

  The data registration unit 50 performs processing for registering the registration data input from the registration data input unit 10 in the image DB 60 and the graph index DB 70. The data registration unit 50 assigns a new image ID to the image data input as registration data, and stores it in the image DB 60. In addition, the image data is stored in the image DB 60 in association with the vector data extracted from the image data.

  According to FIG. 1, the data registration unit 50 includes a similar data acquisition unit 51, a link registration unit 53, and a link optimization unit 55.

  The similar data obtaining unit 51 obtains data similar to the registered data by outputting a request for retrieving similar data for the registered data input from the registered data input unit 10 to the graph retrieving unit 40.

  The link registration unit 53 sets a link between the registration data and the similar data acquired by the similar data acquisition unit 51 and stores the link in the graph index DB 70. Specifically, when registering the image ID and vector data for the registered data in the graph index DB 70, the image ID of the similar data is stored in association with the link destination. Further, by adding the image ID of the registration data to the link destination with respect to the image ID of the similar data, a mutual link is formed between the registration data and the similar data.

  The link optimization unit 55 is a functional unit that optimizes the link so that the graph structure of the graph index is not complicated by the new link registered by the link registration unit 53. Although details will be described later, in brief, by adding link destinations to similar data, the number of links from the similar data increases, thereby increasing the number of links that the graph search unit 40 circulates. May lead to a decrease in search speed. Therefore, when adding a link destination from similar data, the link is reviewed so as to keep the number of links appropriate.

  The search device 1 generates a graph index DB 70 that is a search index having a graph structure, and searches for data similar to the search data by following links on the graph structure. In addition, when registration data is input, the registration data is newly added to the search index that circulates at the time of the search, and the registration data can also be searched by making the registration data a circulation route. A graph index is formed in a self-generating manner.

[Operation of search device]
Next, the operation of the graph self-generation process performed when the search device 1 according to the present embodiment registers data will be described in detail with reference to FIGS. This graph self-generation process is executed based on a program stored in advance in the memory of the search device 1. FIGS. 2 and 3 are flowcharts showing an example of the operation of the search device 1, and FIGS. 4 to 6 are views showing a state of data search and link optimization processing on the graph index.

[Graph self-generation processing]
First, the search device 1 performs an initial setting process (step S1). In this initial setting, the number of links set in newly registered data when data registration is performed, and a threshold value (link reduction threshold value) when performing link reduction are set in advance.

  Next, when registration data is input from the registration data input unit 10 to the data registration unit 50 (step S2), the similar data acquisition unit 51 outputs the registration data to the graph search unit 40 to perform graph search processing ( Step S3).

[Graph search processing]
When the graph search unit 40 starts the graph search process, first, the vector data generation unit 41 generates vector data by extracting feature amounts from the registered data (step S41).

[Search start point determination processing]
Next, the search start point determination unit 43 determines vector data close to the registered data as a search start point (step S43). FIG. 3 is a flowchart illustrating an example of a process for determining a search start point.

  The search start point determination unit 43 first selects an arbitrary point (for example, point a in FIG. 4B) from the graph index (step S430), and is linked to the selected point (selected point). Points (for example, points b and c in FIG. 4C) are acquired based on the graph index DB 70 (step S432).

  Next, the distance in the feature amount space from the registered data for the linked point (link destination) is calculated (step S434). The search start point determination unit 43 extracts the point having the shortest distance from the registered data in the calculated distance (step S436), and the distance D2 between the point and the registered data is the distance D1 between the registered data and the selected point. Or less (steps S438 and 440). In FIG. 4D, when the distance D1 between the selected point a and the registered data g is compared with the distance D2 between the selected point a and the link destination c, it is determined that the distance D2 is smaller.

  When the search start point determination unit 43 determines that the distance D2 is smaller than the distance D1 (step S440; Yes), the search start point determination unit 43 selects the point extracted in step S436 (step S442), and the process proceeds to step S432. .

  That is, by performing the loop processing of steps S432 to S442, the link is traced so as to select a point close to the registered data on the graph index from the point arbitrarily selected first.

  When determining that the distance D2 is greater than the distance D1 in step S440 (step S440; No), the search start point determination unit 43 determines the selected point as the search start point (step S444).

  In FIG. 4E, the point closest to the registration data g is sequentially selected from the point a selected first, and the point c → the point b → the point d is circulated so that the point d closest to the registration data becomes the point d. It is determined as a search start point.

  In the search start point determination process described above, even when the search start point is determined on the assumption that the distance D2 is greater than the distance D1, the search start point may be outside the search range centered on the registered data. is there. In this case, the search cannot be started because the search start point is outside the search range. Therefore, if the search start point determined in step S444 is outside the search range, an arbitrary point is newly selected from the graph index, and the search start point determination process is performed a predetermined number of times to start the search within the search range. Determine the point.

  Further, if a point to be arbitrarily selected in step S430 is selected at random, there is a possibility that an adjacent point will be selected again when the search start point determination process is repeated. In this case, even if the link is traced from the adjacent point, the possibility that the link of the same path is traced is high, and the search start point within the search range cannot be determined.

  Therefore, when the search start point determination process is repeatedly performed a predetermined number of times, it is preferable to newly select a point that is a predetermined distance away from a previously selected point. A plurality of points dispersed in the vector space may be specified in advance, and an arbitrary point may be selected from them. Thereby, the search start point within the search range can be determined efficiently.

[Turning the graph]
When the search start point is determined as described above, the graph circulator 45 sequentially follows the links starting from the search start point (step S45). At this time, there is a case where the link is branched from one vector data (that is, there are a plurality of paths). In this case, the vector data on the traced path, or information for specifying it is used. Then, it is stored in the memory of the search device 1 by an appropriate method.

[Search for similar data]
The graph circulator 45 determines whether or not a prescribed search end condition is satisfied while following links on the route. Various search termination conditions can be considered. For example, there may be any of the following search end conditions.

(Search end condition 1)
The result of following the link starting from the search start point is a range obtained by multiplying a predetermined search range (for example, the search range in FIG. 5A) centered on the registered data by α (α> 1) (that is, the search limit range). ). Here, for α, an appropriate value such as 1.5 can be adopted. As α increases, search leaks can be reduced. On the other hand, the search time can be shortened as α is reduced.

  When the number of data searched by following links increases, the search can be terminated early by decreasing the value of α. Thereby, even when the number of data is excessive, a search result can be acquired early. Note that it is preferable to set the value of α so as not to be 1 or less in order to improve the search accuracy. However, if it is desired to end the search at an early stage, it is possible to set α <1.

(Search end condition 2)
The result of following the link starting from the search start point is that the link has been traced a predetermined number of times outside the search range centered on the registered data.

  Here, the predetermined number of times is, for example, five, but an appropriate value can be adopted. For example, in FIG. 5A, when the point a is traced from the point b, the search range is exceeded, but when the link is followed once more after the search range is exceeded, the point c is reached. Thus, the omission of search can be reduced by setting the number of rounds outside the search range.

  When the search end condition is satisfied for one route, the link in the other route is traced. When the search end condition is satisfied for any route, the process proceeds to the next step.

  The similar data specifying unit 47 circulates the graph until the search end condition is satisfied (step S47 → S45). If the search end condition is satisfied (step S47; Yes), the circulated point (vector data) is registered. The distance to the data is calculated, and similar data ranked based on the distance is output (step S49).

  For example, when the search start point is point d in FIG. 5 (a), the links of points b, e, and f which are link destinations are traced and then linked to point e as shown in FIG. 5 (b). Trace to point c. Then, as shown in FIG. 5C, when the point c is traced to the point a, the search range is exceeded. When the search end condition is that the search range is exceeded, the link circulation is ended.

  Then, ranking of similar data is obtained based on the distance between each of the points b to f and the registered data g traced from the search start point d. In FIG. 5C, rankings are obtained in the order of points d, b, c, e, and f.

  When the similar data search by the graph search unit 40 is completed, the link registration unit 53 of the data registration unit 50 sets a link between the registered data and the similar data (step S5). Here, the link registration part 53 is good also as producing | generating the temporary link between registration data and similar data, and memorize | storing it in the graph index DB70 after the link optimization process mentioned later. The temporary link is also for tracing from one vector data to the other vector data, and is temporarily held in the memory.

  Similar data for the number of links set by the initial setting is selected from the top of the ranking of similar data searched, and temporary links are set for the similar data. For example, when the number of links is ‘3’, similar data points b, c, and d are linked to the registered data g as shown in FIG.

[Link optimization processing]
Next, the link optimization unit 55 performs link optimization processing (step S6). In the present embodiment, the link optimization process is performed on the similar data obtained by generating the temporary link from the registered data as the link reduction target. That is, the number of links for the data to be reduced is acquired based on the graph index DB 70, and if the number of links is equal to or greater than the link reduction threshold, the link of the data is reduced.

  At this time, the similar data linked (including the temporary link) to the link reduction target data is reduced in order of increasing distance. For example, when the link reduction threshold is ‘4’, in FIG. 6A, the number of links set at points c, b, and d at which links are generated is 4, which is equal to or greater than the link reduction threshold. Therefore, the link set between the point c and the point b farthest among the links is reduced. Further, for the point d, the link set between the point e and the point e is reduced, and a graph structure as shown in FIG. 6B is obtained. Here, the link reduction target point b does not need to be reduced because the number of links becomes less than the link reduction threshold when the link at the point c is reduced.

  The link optimization unit 55 updates the link information of the graph index DB 70 based on the result of the link reduction process, and the registration data is stored in the image DB 60 (step S7). When further registration data is input from the registration data input unit 10 (step S8; Yes), the process proceeds to step S2 to perform graph search processing, link generation, and link optimization. If there is no new data (step S8; No), the graph self-generation process is terminated.

  Further, although detailed description is omitted, in the case of only the search processing for searching similar data to the search data and returning the search result to the client terminal, the search data input from the search data input unit 20 is processed. The graph search processing by the graph search unit 40 described above is performed. And the search result output part 30 produces | generates the search result which ranked similar data obtained by the similar data specific | specification part 47 based on distance as display data, and returns it to a client terminal.

  As described above, according to the first embodiment, similar data for registered data is searched using a search algorithm that searches for similar data by searching for links between graph index data. By newly generating a link in the meantime, data is registered in the graph index searched by the search device 1 following the link, and the link is formed as data to be circulated after registration. Thus, the search device 1 registers data on the graph index that circulates during the search, and self-generates the graph index.

  In addition, each time data is registered, the search device 1 optimizes the link structure of the graph index when registering data by optimizing the number of links of newly formed data. And self-healing.

  Therefore, similar data can be searched without calculating the distance to all data in the database by a search algorithm using a graph index, so that data can be registered at high speed in a database storing multidimensional vector data. Can do. In addition, since the number of links in the graph index is kept optimal, it is possible to prevent the search speed from being lowered due to a complicated link structure due to an increase in data.

[Second Embodiment]
Next, a second embodiment of the search device 1 will be described. Note that the description of the same functional configuration and the same processing steps as those in the first embodiment will be omitted. The search device 1 in the second embodiment is realized by replacing the link optimization process in FIG. 2 with the link optimization process according to the flowchart shown in FIG. Hereinafter, the link optimization process will be described with reference to the diagram showing the state of the link optimization process of FIG. In the illustration of the link structure in FIG. 8, the illustration of the links between data after FIG. 8B is appropriately omitted by broken lines for the sake of simplicity.

  As described above, a search algorithm using a graph index searches for similar data by circulating links between vector data. However, if a link break occurs, it cannot be performed due to isolation of vector data. End up. The search device 1 according to the second embodiment repairs the link when the link is reduced.

  First, the link optimizing unit 55 specifies a point (link holding point) that holds the link among all the points linked to the point for which the reduction target link is set (step S61). The link holding point may be specified by, for example, a predetermined number of points that are close to the point to be reduced or a point within a predetermined distance. In FIG. 8A, six points b, c, d, e, and f that are close to the point a to be reduced are specified as link holding points. Further, the points to be reduced may be included in the link holding points.

  In FIG. 8A, five points b, c, d, e, and f that are close to the point a to be reduced are specified as link holding points. Then, points other than the link holding point among points linked to the reduction target points are selected one by one from the repair target points as link repair targets (step S62). In FIG. 8A, the point g is selected as a repair target point.

  Next, the link optimization unit 55 calculates the distance between the repair target point and each link holding point (step S63), and the link holding point having the shortest distance among the calculated distances and the repair target point are calculated. A new link is set between them (step S64).

  For example, as shown in FIG. 8B, the distance between the restoration target point g and each of the link holding points b to f is calculated, and the link holding point having the shortest distance as shown in FIG. 8C. A new link is set between e and the repair target point g. When a link has already been set between the repair target point and the link holding point, the link need not be set.

  The link optimization unit 55 releases the link set between the repair target point for which the link has been set and the reduction target point (step S65). In FIG. 8D, the link between the points a and g is released.

  Next, when the points linked to the reduction target points other than the link holding points, that is, the repair process in steps S62 to S65 is performed on all the points among the repair target points (step S66; No). When the link optimization process is finished and there is a point that has not been repaired (step S66; Yes), the process proceeds to step S62.

  Even if the link to the reduction target point is released for points other than the points b to f specified as the link holding point in FIG. 8A, a link is set for each link holding point as shown in FIG. Therefore, it is possible to prevent the link from being broken.

  As described above, according to the second embodiment, even if the link between the link reduction target point and the repair target point is released, a link from the repair target point to the link holding point is secured. It is possible to prevent cutting. Therefore, it is possible to prevent a decrease in search performance even when the graph index link is deleted.

[Third Embodiment]
Next, a third embodiment of the search device 1 will be described. In the above-described embodiment, the search device 1 specifies the search start point by following a link from a point arbitrarily selected on the graph index. However, in the third embodiment, vector data is represented on the vector space. The search start point is specified by clustering and indexing. This index is generated by a known space division technique for a vector space, and is briefly described as follows.

  First, arbitrary vector data (point A) in the graph index DB 70 is selected, and a point B farthest from the point A is selected. In order to select the farthest point, the distance from the point A to each point may be calculated. Next, the distances from the point B to all the points in the graph index DB 70 are obtained and sorted in the order of the distances. The difference between adjacent distances in the sorted order is calculated. For example, it is assumed that the sorted distances are arranged in order as d1, d2, d3,. In this case, the calculation is sequentially performed as | d1-d2 |, | d2-d3 |.

  Then, the cluster is divided between the distances having the maximum difference. For example, if the distance between d1 and d2 is the maximum, division is performed between a point P1 corresponding to the distance d1 and a point P2 corresponding to the distance d2. Thereby, similar data can be divided into clusters of a tree structure.

  An index by a tree-structured cluster is made into a database by, for example, associating the center point of the cluster with the radius of the cluster and the ID of a child cluster having a parent-child relationship. It should be noted that a known technique can be applied as appropriate to the cluster index configuration.

  To search for the search start point from the tree structure index, first specify the cluster to which the registration data belongs from the upper layer of the tree structure based on the center and radius of the cluster, and then specify the cluster to which the registration data belongs from the lower clusters To do. In this way, the lower layer cluster to which the registered data belongs is sequentially identified, and the data belonging to the lowermost layer cluster is determined as the search start point. When a plurality of data belong to the lowermost cluster, each can be set as a search start point. That is, by tracing the tree structure index in the direction approaching the registered data, data relatively close to the registered data can be acquired as a search start point.

  FIG. 9 shows a state where the vector data group is divided into clusters by space division. Thus, by generating an index in which vector data is associated with each divided cluster, vector data in the same hierarchy can be searched at high speed.

  Here, for example, as shown in FIG. 9, it is assumed that registration data is input, and the vector data in the same hierarchy is searched by searching the tree structure index. However, the search range centered on the registration data is the vector data. May be outside. In this case, if the vector data of the same hierarchy is set as the search start point, the search efficiency is deteriorated.

  Therefore, first, vector data in the same hierarchy is searched using an index by clustering, and it is determined whether or not the vector data is within the search range. If it is within the search range, the vector data is specified as the search start point.

  If it is outside the search range, the search start point is determined by selecting and starting vector data of the same hierarchy as an arbitrary point according to the search start point determination process (see FIG. 3) described above. As a result, the search start point can be determined at high speed and the accuracy of the search result can be maintained.

Next, a modification of the above-described embodiment will be described.
[Modification 1]
When the same vector data as the vector data of the registered data is already stored in the graph index DB 70, it is stored only in the image DB 60 without being stored in the graph index DB 70. Can be associated with In this case, the graph index DB 70 is newly provided with a data item for storing the image ID of the image data having the same vector data as the related ID in addition to the image ID.

  When a related ID is associated with data searched as similar data during data search, the image data indicated by the related ID can also be output as a search result. For this reason, it is possible to prevent the search speed from being lowered due to the complicated structure of the graph index due to the registration of the same vector data.

[Modification 2]
Although the search range of the graph search process is determined by a specified distance centered on the registered data, it may be determined by the number of searches that circulate from the search start point. That is, the search start point is determined by the method described above, and the link is traced from the search start point. At this time, data of a predetermined number of searches (for example, 10) is collected by link circulation. Then, the data is sorted in the order of the distance between the collected data and the registered data, and a list of the circulated data is obtained. Vector data corresponding to the number of searches is held in this list, and the distance corresponding to the farthest vector data with respect to the registered data is updated as the search range.

  That is, as long as the search end condition is not satisfied, the link is followed, and if data closer to the registered data than the data in the list is found, the data is added to the list in the order of the distance. Then, the search range is updated according to the farthest data updated in the list by this data addition. As a result, the search range can be narrowed by updating the list with data that is close to the registered data according to the collection of data for the number of searches. For this reason, the search range can be dynamically determined according to the designation of the number of searches. Moreover, since the search range can be narrowed each time near data is searched, the search speed can be improved while maintaining the search accuracy.

[Modification 3]
The range of link circulation for determining the search start point may be determined using the method of the second modification described above. That is, the process of selecting an arbitrary point from the graph index and specifying the nearest point is regarded as a graph search process with one search, and the data closer to the registered data to be searched by link circulation is obtained. By collecting, holding and updating one, the distance between the data and the registered data is dynamically determined as the radius of the search range.

  When the link is visited, data is selected so as not to exceed the dynamically determined search range. Therefore, the search range can be narrowed while following the link to specify the search start point, and the time until the search start point is specified can be shortened.

  In addition, the link is circulated within a range that does not exceed the range (search limit range) obtained by multiplying the search range by α (α> 1), so that the range that circulates even when data close to the registered data may not be linked. As a result, the optimum search start point can be identified.

[Modification 4]
Since the data circulated when determining the search start point may be the final search result, the distance calculated in step S434 in FIG. 3 is stored in the memory as a circulation history, and the circulation is performed in the graph search. The distance calculation process can be omitted by referring to the history.

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

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

The block diagram which shows an example of a function structure of the search device in 1st Embodiment. The flowchart for demonstrating the processing content of the graph self-generation process in 1st Embodiment. The flowchart for demonstrating the processing flow of the search start point determination in 1st Embodiment. The schematic diagram which shows the specific example of the process of the determination of the search start point in 1st Embodiment. The schematic diagram which shows the specific example of the process of searching for similar data in 1st Embodiment. The schematic diagram which shows the specific example of the processing content of the link optimization process in 1st Embodiment. The flowchart for demonstrating the processing flow of the link optimization process in 2nd Embodiment. The schematic diagram which shows the specific example of the processing content of the link optimization process in 2nd Embodiment. The figure which shows an example of the tree structure index in 3rd Embodiment.

DESCRIPTION OF SYMBOLS 1 Search apparatus 10 Registration data input part 20 Search data input part 30 Search result output part 40 Graph search part 41 Vector data generation part 43 Search start point determination part 45 Graph circulation part 47 Similar data specification part 50 Data registration part 51 Similar data acquisition Unit 53 link registration unit 55 link optimization unit 60 image DB
70 Graph Index DB

Claims (9)

A database in which a link for tracing from one data to the other data is set between the data and a plurality of the data is stored;
For the input search target data, select any one of the data stored in the database as a search start point, and sequentially follow the link from the search start point, so that the path following the link Graph search means for searching for and outputting data within a predetermined distance from data to be searched as similar data among certain data,
Similar data acquisition means for acquiring similar data for the registration target data by inputting the data to be registered in the database to the graph search means;
A link that sets a link between the data to be registered and the similar data acquired by the similar data acquisition unit, stores the link in the database, and performs a process of reducing the links set in the similar data Optimization means;
A search data management device comprising: The link optimization means includes:
When the number of links set in the similar data is equal to or greater than a predetermined number, the predetermined number of the links are held in order from the shortest distance between the data set with the links, and other links are released. The search data management device according to claim 1. The link optimization means includes:
From the other data linked to the data to be reduced in the link, the holding data for holding the link and the release data for releasing the link are identified,
3. The search data according to claim 1, wherein a link is set between the release data and the held data based on a distance between the release data and the held data and stored in the database. Management device. The graph search means includes:
Selecting arbitrary data from the database, searching the data within the predetermined distance from the registration target data by sequentially following the link from the data, and selecting the data as a search start point The search data management device according to any one of claims 1 to 3. The graph search means includes:
As a result of following the link from the arbitrary data, if there is no data within the specified distance in the data on the path following the link, data different from the selected arbitrary data is displayed. 5. The search data management apparatus according to claim 4, wherein the search start point is searched by performing a process of newly selecting and sequentially following the links again. The data stored in the database is further hierarchized by space division, further comprising a hierarchical database for storing for each hierarchy to which the data belongs,
The graph search means includes:
The search data according to any one of claims 1 to 3, wherein the hierarchy to which the registration target data belongs is specified, and other data belonging to the specified hierarchy is selected as the search start point. Management device. The graph search means includes:
If the distance between the search start point selected based on the hierarchical database and the data to be registered is more than the specified distance, select any data from the database and link to the data. 7. The search data management apparatus according to claim 6, wherein the search data is searched for and searched for within the specified distance and selected as the search start point. In a search data management method in which a link for tracing from one data to the other data is set between the data, and a computer manages a database in which a plurality of the data is stored as a data group to be searched,
For the input search target data, select one of the data stored in the database as a search start point, and sequentially follow the link from the search start point to A graph search step of searching and outputting data within a predetermined distance from the data to be searched as similar data among the data in
A similar data acquisition step of causing the data to be registered in the database to be searched in the graph search step as the search target, and acquiring similar data for the data to be registered;
A link that sets a link between the data to be registered and the similar data acquired in the similar data acquisition step, stores the link in the database, and performs a process of reducing the links set in the similar data An optimization process;
The search data management method, wherein the computer performs   The program for making a computer perform the process of the search method of Claim 8.

Images