CN105868355A - Large-scale multimedia data spatial index method - Google Patents
Large-scale multimedia data spatial index method Download PDFInfo
- Publication number
- CN105868355A CN105868355A CN201610187012.9A CN201610187012A CN105868355A CN 105868355 A CN105868355 A CN 105868355A CN 201610187012 A CN201610187012 A CN 201610187012A CN 105868355 A CN105868355 A CN 105868355A
- Authority
- CN
- China
- Prior art keywords
- node
- data
- tree
- multimedia data
- index method
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/40—Information retrieval; Database structures therefor; File system structures therefor of multimedia data, e.g. slideshows comprising image and additional audio data
- G06F16/41—Indexing; Data structures therefor; Storage structures
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Software Systems (AREA)
- Multimedia (AREA)
- Data Mining & Analysis (AREA)
- Databases & Information Systems (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The invention discloses a large-scale multimedia data spatial index method. A node is split into three node bodies through a scientific splitting algorithm, m/M is selected to be 30-40%, and data needing to be processed is placed in a processor low in occupancy rate to be processed. By means of the splitting algorithm, the node is split into the three node bodies, overlapped areas of the node bodies are allocated into one node body, multi-path search is reduced during data retrieval, and query efficiency is improved. Meanwhile, the capacity of the node is increased, m/M is selected to be 30-40%, a good node utilization rate can be kept, and the node space utilization rate is fully used. Query of large-scale multimedia data is achieved.
Description
Technical field
The present invention relates to a kind of space index method, a kind of large scale multimedia data
Space index method.
Background technology
For multimedia messages, computer be difficult to judge as the mankind between multimedia object be
No have similarity.Therefore, people, by extracting the feature of multimedia messages, i.e. use feature
The method of vector judges whether have similarity between them.Characteristic vector is typically one group high
The set of vectors of position, in order to effectively extract large scale multimedia information, is generally directed to big
The characteristic vector of scale multimedia data object sets up its spatial index, realizes multimedia messages
The inquiry of greater efficiency.In multimedia information retrieval, multimedia object is often represented as one
The characteristic vector of individual higher-dimension, the similarity between two multimedia objects depends on two corresponding height
Distance (conventional Euclidean distance) between dimensional feature vector.Thus multimedia retrieval turns
Become calculating the calculating of the spacing of set of eigenvectors and characteristic vector to be checked in data base.
Spatial index is a kind of data organized according to data space distribution characteristics and store data
Structure.Spatial index is in large scale multimedia market demand, e.g., and SR-Tree, A-Tree
Deng, there is substantial amounts of overlapping region between index node, and internodal overlapping region be the biggest,
Multichannel during object retrieval can be caused to search increase, thus reduce large scale multimedia data
Search efficiency.
In order to solve the problems referred to above, the present invention proposes the space of a kind of large scale multimedia data
Indexing means (KSR-Tree).
Summary of the invention
It is an object of the invention to, it is provided that the space index method of a kind of large scale multimedia data.
The method uses splitting algorithm can inquire about large scale multimedia data, keeps joint very well
Point utilization rate, and it is effectively improved search efficiency, also make the capacity of node storage data
Increase.
For solving above-mentioned technical problem, the technical scheme that the present invention provides is as follows: a kind of extensive
The space index method of multi-medium data: by utilizing splitting algorithm that one node split is become three
Individual, choose m/M is 30%-40% simultaneously, then need to data to be processed to be placed on occupancy few
Node on process.
In the space index method of aforesaid large scale multimedia data, described splitting algorithm
Step is,
A., when interstitial content reaches M+1 or M+2, whether first decision node is to overflow for the first time,
The most then reinsert splitting algorithm, postpone the division of node, otherwise, then split vertexes;
The spilling of node occurs after B. reinserting again, then node needs division;From node
In M+1 or M+2 data object in, arbitrarily select three objects as in initial clustering
The heart;
C. each data object distance to three initial cluster centers is calculated, initial by three
Cluster centre is respectively divided the center of each data object of its nearest neighbours;
Recalculate the average of the data object divided the most again;
E. iteration C-D two step is until new average is equal with former average or is less than appointment threshold value;With
This is as three intermediate nodes after division.
In the space index method of aforesaid large scale multimedia data, described threshold value is
0.0001。
Compared with prior art, the present invention is by utilizing splitting algorithm that one node split is become three
Individual node, is divided into the overlapping region of node in one node, during minimizing data retrieval
Multichannel is searched, and improves search efficiency, makes the capacity of node increase simultaneously;Choose m/M
For 30%-40%, fine Duty-circle can be kept, the most sufficiently make use of node space profit
By rate;Achieve the inquiry to large scale multimedia data.
KSR-Tree and SR-Tree of the application, A-Tree have been carried out a series of by applicant
Contrast experiment, analysis of experimental data is as follows:
Experimental situation
1. hardware environment: processor AMD 6core, internal memory 16G, hard disk 2T
2. operating system platform: Ubuntu 12.04-64 position
3. programmed environment: Eclipse for C/C++
Testing scheme
Verified the effectiveness of this spatial index algorithm by experiment, with other, there is representative simultaneously
The Spatial Data Index Technology of property, such as the comparison of SR-Tree, A-Tree.Reality in performance evaluation
Test middle employing real data both to be from obtaining micro-in the media computation group of Microsoft Research, Asia
The soft real large scale multimedia image data set that must answer image and video search engine
MSRA-MM2.0 data set.
High dimensional feature vector data used in experiment include 128 dimension Wavelet Texture and
256 dimension RGB feature two kinds, they are both from according to Bing Images photographic search engine
The actual picture data included, data total amount is 1,000,000.100 pictures institutes therein are right
The characteristic vector answered is selected and is used as query set.In an experiment, by comparative analysis data set,
The parameters such as k value, page size index based on KSR-Tree, SR-Tree and A-Tree
KNN retrieve the performance i.e. impact of average response time.
Adjust the impact after data set size
In an experiment, we are dimensioned to 200,000,400,000,800,000,100 data set
Ten thousand, the dimension of data is respectively 128 dimensions and 256 dimensions.Experimental result is as shown in Fig. 2-Fig. 7.
Test result indicate that, A-Tree, when data volume is 200,000, achieves preferably inspection
Suo Xiaoguo.This is because this spatial index introduces characteristic vector approximation table based on compression thought
Show technology, it is hereby achieved that bigger fan-in and fan-out.But A-Tree is owing to have employed
Compress technique, adds the process of a decoding, and needs to tie with reference to its father node and child
The position of point, but when data volume increases, its recall precision is on the contrary not as KSR-Tree.
Therefore, although always A-Tree retrieval effectiveness under low data bulk is pretty good, but in multimedia
In retrieval when data are bigger, the response time to kNN inquiry of A-Tree is not special
Preferable.See again its time of SR-Tree always higher than KSR-Tree, this be due to
In higher dimensional space, the index divided based on space can cause a large amount of weights in search volume is expanded
Folded region and the access of data object and analysis.
On the average response time of inquiry, kNN based on KSR-Tree in this paper
Retrieval achieves best result on response time.Main cause is: (1) is the most
On media data, data volume is big, and dimension is higher.Cause the overlapping region between node more,
And the divisional mode of KSR-Tree makes to which reduce multichannel and searches.(2) cut carrying out kNN
During branch, owing to have employed node split algorithm in this paper, just reduce when kNN beta pruning
Within to hunting zone, respective nodes repeat inquiry.
Adjust the impact after K value size
The impact on kNN based on KSR-Tree retrieval performance of this experimental evaluation k value.Number
Being dimensioned to 1,000,000 according to collection, the dimension of data is respectively 128 dimensions and 256 dimensions.Experiment knot
Fruit is as shown in figures s-11.
From Fig. 8-Fig. 9 it can be seen that although as the increase of k value, three kinds of index technologies
Average response time strains big mutually, but it can be seen that the increase of KSR-tree is minimum
, and its response time is also best.KSR-Tree on response time still
It is better than other two kinds of index technologies.No matter it is at 128 dimensions or 256 dimensions, the sound of A-Tree
It is always the highest between Ying Shi.SR-Tree is the most poor, and this is that search during kNN retrieval is initial
Radius is relatively big and overlap between node is higher causes.
It can be seen that along with the increase of data set, response time also exists from Figure 10-Figure 11
Ceaselessly increase.But under same data set, the amplitude of variation of KSR-Tree is less.With
K value the biggest, its response time there has also been a little to be increased.This is owing to expanding in search volume
Zhang Zhong, result in the regions to a large amount of overlaps and the access of data object and analysis, thus increases
Time of response.
Adjust the impact after Page size size
The change of the experimental evaluation page size impact on the recall precision of KSR-Tree, k
Value is set to 100, and page size is dimensioned to 32k, 64k and 128k.Experimental result
As shown in Figure 12-Figure 13.
For KSR-Tree, increase page size, it is meant that have bigger
Fanout, can accommodate more child's data inside the most each node, thus effectively will weight
Folded region comprises into a node.
Going out from the experimental results, tie up 128, page size is set to retrieval during 64k
Efficiency this be best.Tieing up 256, page size is set to retrieval effect during 128k
Rate this be the most stable, always be two page size less than other.Based on KSR-Tree
KNN retrieval be minimum on response time.
Split vertexes number is selected different numerical value to carry out the impact of response time by applicant
Test, the nodes after division selects the experimental result of (3,4,5,6) as shown in figure 14.
In the case of 128 dimension data, the page size of index node is 64KB, and data set is from 200,000
Bar to 1,000,000.When vertical coordinate position is by inquiry, the response time of inquiry, unit is the second.
Response time curve is the response time that nodes selects (3,4,5,6) the most successively,
As seen from the figure, the when that the nodes after division being chosen as 3, response time is the shortest, and performance is
It is better than 4,5,6.
Split vertexes number is selected different numerical value to carry out the impact of response time by applicant
Test, the nodes after division selects the experimental result of (3,4,5,6) as shown in figure 15,
In the case of 256 dimension data, the page size of index node is 64KB, and data set is from 200,000
Bar to 1,000,000.When vertical coordinate position is by inquiry, the response time of inquiry, unit is the second.
Response time curve is the response time that nodes selects (3,4,5,6) the most successively,
As seen from the figure, the when that the nodes after division being chosen as 3, response time is the shortest, and performance is
It is better than 4,5,6.
Being found out by Figure 14-15, the when that the nodes after splitting being chosen as 3, response time is the shortest.
Accompanying drawing explanation
Fig. 1 is the splitted construction figure of the present invention;
Fig. 2 is 128 dimensions, and when k is 20, data set size is respectively to KSR-Tree, SR-Tree
Impact with A-Tree;Wherein, abscissa is data set size, and vertical coordinate is response time;
Fig. 3 is 256 dimensions, and when k is 20, data set size is respectively to KSR-Tree, SR-Tree
Impact with A-Tree;Wherein, abscissa is data set size, and vertical coordinate is response time;
Fig. 4 is 128 dimensions, and when k is 50, data set size is respectively to KSR-Tree, SR-Tree
Impact with A-Tree;Wherein, abscissa is data set size, and vertical coordinate is response time;
Fig. 5 is 256 dimensions, and when k is 50, data set size is respectively to KSR-Tree, SR-Tree
Impact with A-Tree;Wherein, abscissa is data set size, and vertical coordinate is response time;
Fig. 6 is 128 dimensions, and when k is 100, data set size is respectively to KSR-Tree, SR-Tree
Impact with A-Tree;Wherein, abscissa is data set size, and vertical coordinate is response time;
Fig. 7 is 256 dimensions, and when k is 100, data set size is respectively to KSR-Tree, SR-Tree
Impact with A-Tree;Wherein, abscissa is data set size, and vertical coordinate is response time;
Fig. 8 is 128 dimensions, and data set size is 1,000,000, k value respectively to KSR-Tree,
The impact of SR-Tree and A-Tree;Wherein, abscissa is k value, when vertical coordinate is for response
Between;
Fig. 9 is 256 dimensions, and data set size is 1,000,000, k value respectively to KSR-Tree,
The impact of SR-Tree and A-Tree;Wherein, abscissa is k value, when vertical coordinate is for response
Between;
Figure 10 is k value variation diagram under 128 dimensions;Wherein, abscissa is data set size, vertical
Coordinate is response time;
Figure 11 is k value variation diagram under 256 dimensions;Wherein, abscissa is data set size, vertical
Coordinate is response time;
Figure 12 is 128 dimensions, the different page size impacts on KSR-Tree;Wherein, horizontal
Coordinate is data set size, and vertical coordinate is response time;
Figure 13 is 256 dimensions, the different page size impacts on KSR-Tree;Wherein, horizontal
Coordinate is data set size, and vertical coordinate is response time;
Figure 14 is 128 dimensions, and page size is 64KB, data set from 200,000 to 1,000,000,
Split vertexes number selects the impact on response time of the different numerical value;Wherein, abscissa is data
Collection size, vertical coordinate is response time;
Figure 15 is 256 dimensions, and page size is 64KB, data set from 200,000 to 1,000,000,
Split vertexes number selects the impact on response time of the different numerical value;Wherein, abscissa is data
Collection size, vertical coordinate is response time.
Detailed description of the invention
Embodiment.The space index method of a kind of large scale multimedia data, is divided by utilization
One node split is become three by algorithm, and choose m/M is 30%-40% simultaneously, then by needs
The data of reason are placed on the node that occupancy is few and process.
Described m is the lower limit of object in node, and M is the upper limit of object in node.The value of m
The 30-40% taking M is optimum, the preferable search efficiency of guarantee and insertion efficiency, the most both
Fine Duty-circle can be kept, it is also possible to make full use of the space availability ratio of node.Judgement accounts for
By the method for the few node of rate it is: two index nodes, each node can at most deposit 10
Record, if one of them node A has housed 6 records, another node B, has deposited
Put 8 records, then the occupation rate of node A is less.
The step of described splitting algorithm is,
A., when interstitial content reaches M+1 or M+2, whether first decision node is to overflow for the first time,
The most then reinsert splitting algorithm, postpone the division of node, otherwise, then split vertexes;
The spilling of node occurs after B. reinserting again, then node needs division;From node
In M+1 or M+2 data object in, arbitrarily select three objects as in initial clustering
The heart;
C. each data object distance to three initial cluster centers is calculated, initial by three
Cluster centre is respectively divided the center of each data object of its nearest neighbours;
Recalculate the average of the data object divided the most again;
E. iteration C-D two step is until new average is equal with former average or is less than appointment threshold value;With
This is as three intermediate nodes after division.Described threshold value is smaller floating more than 0
Count, be set as 0.0001 here.
As it is shown in figure 1, the node structure of KSR-Tree is similar with SR-Tree, simply at joint
" one dividing into three " is used in the division of point, if the data object in leaf node C
When reaching M+1, first judge whether that needs reinsert.As need not, this node is carried out point
Split, after division, three new leaf nodes will be produced in intermediate node 1.In judging again
Intermediate node 1, the need of dividing, if do not divided, divides stopping;Then will if desired for division
Node 1 splits into three new intermediate nodes, finally repeats this process to root node root.
In the structure chart of KSR-Tree, if node C overflows, the number in node C
According to object, node C, I and J after carrying out the division of " one dividing into three ", are obtained.Now
Node 1 does not overflows, and division stops, and obtains the structure chart shown in Fig. 1.
Claims (3)
1. the space index method of large scale multimedia data, it is characterised in that: pass through
Utilizing splitting algorithm that one node split is become three, choose m/M is 30%-40% simultaneously, then
Process needing data to be processed to be placed on the node that occupancy is few.
The space index method of large scale multimedia data the most according to claim 1,
It is characterized in that: the step of described splitting algorithm is,
A., when interstitial content reaches M+1 or M+2, whether first decision node is to overflow for the first time,
The most then reinsert splitting algorithm, postpone the division of node, otherwise, then split vertexes;
The spilling of node occurs after B. reinserting again, then node needs division;From node
In M+1 or M+2 data object in, arbitrarily select three objects as in initial clustering
The heart;
C. each data object distance to three initial cluster centers is calculated, initial by three
Cluster centre is respectively divided the center of each data object of its nearest neighbours;
Recalculate the average of the data object divided the most again;
E. iteration C-D two step is until new average is equal with former average or is less than appointment threshold value;With
This is as three intermediate nodes after division.
The space index method of large scale multimedia data the most according to claim 2,
It is characterized in that: described threshold value is 0.0001.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201610187012.9A CN105868355A (en) | 2016-03-29 | 2016-03-29 | Large-scale multimedia data spatial index method |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201610187012.9A CN105868355A (en) | 2016-03-29 | 2016-03-29 | Large-scale multimedia data spatial index method |
Publications (1)
Publication Number | Publication Date |
---|---|
CN105868355A true CN105868355A (en) | 2016-08-17 |
Family
ID=56625121
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201610187012.9A Pending CN105868355A (en) | 2016-03-29 | 2016-03-29 | Large-scale multimedia data spatial index method |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN105868355A (en) |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107391601A (en) * | 2017-06-30 | 2017-11-24 | 安徽四创电子股份有限公司 | A kind of construction method of the high dimensional indexing of face feature vector |
CN116796209A (en) * | 2023-08-24 | 2023-09-22 | 北京安图生物工程有限公司 | Data processing method for monitoring storage environment temperature of detection kit |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101299213A (en) * | 2008-06-17 | 2008-11-05 | 中国地质大学(武汉) | N-dimension clustering order recording tree space index method |
CN101996242A (en) * | 2010-11-02 | 2011-03-30 | 江西师范大学 | Three-dimensional R-tree index expansion structure-based three-dimensional city model adaptive method |
CN102682184A (en) * | 2011-03-08 | 2012-09-19 | 中国科学院研究生院 | Judgment method of fracture-pair intersection in random-distribution three-dimensional fracture network |
CN102831241A (en) * | 2012-09-11 | 2012-12-19 | 山东理工大学 | Dynamic index multi-target self-adaptive construction method for product reverse engineering data |
CN103092926A (en) * | 2012-12-29 | 2013-05-08 | 深圳先进技术研究院 | Multi-level mixed three-dimensional space index method |
-
2016
- 2016-03-29 CN CN201610187012.9A patent/CN105868355A/en active Pending
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101299213A (en) * | 2008-06-17 | 2008-11-05 | 中国地质大学(武汉) | N-dimension clustering order recording tree space index method |
CN101996242A (en) * | 2010-11-02 | 2011-03-30 | 江西师范大学 | Three-dimensional R-tree index expansion structure-based three-dimensional city model adaptive method |
CN102682184A (en) * | 2011-03-08 | 2012-09-19 | 中国科学院研究生院 | Judgment method of fracture-pair intersection in random-distribution three-dimensional fracture network |
CN102831241A (en) * | 2012-09-11 | 2012-12-19 | 山东理工大学 | Dynamic index multi-target self-adaptive construction method for product reverse engineering data |
CN103092926A (en) * | 2012-12-29 | 2013-05-08 | 深圳先进技术研究院 | Multi-level mixed three-dimensional space index method |
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107391601A (en) * | 2017-06-30 | 2017-11-24 | 安徽四创电子股份有限公司 | A kind of construction method of the high dimensional indexing of face feature vector |
CN116796209A (en) * | 2023-08-24 | 2023-09-22 | 北京安图生物工程有限公司 | Data processing method for monitoring storage environment temperature of detection kit |
CN116796209B (en) * | 2023-08-24 | 2023-10-20 | 北京安图生物工程有限公司 | Data processing method for monitoring storage environment temperature of detection kit |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
US8463045B2 (en) | Hierarchical sparse representation for image retrieval | |
CN104199827B (en) | The high dimensional indexing method of large scale multimedia data based on local sensitivity Hash | |
CN106933511B (en) | Space data storage organization method and system considering load balance and disk efficiency | |
CN109086437A (en) | A kind of image search method merging Faster-RCNN and Wasserstein self-encoding encoder | |
US20080071843A1 (en) | Systems and methods for indexing and visualization of high-dimensional data via dimension reorderings | |
US8391355B2 (en) | Method and device for online dynamic semantic video compression and video indexing | |
Cha et al. | The GC-tree: a high-dimensional index structure for similarity search in image databases | |
CN105589938A (en) | Image retrieval system and retrieval method based on FPGA | |
US20110208754A1 (en) | Organization of Data Within a Database | |
US9442950B2 (en) | Systems and methods for dynamic visual search engine | |
CN101866358A (en) | Multidimensional interval querying method and system thereof | |
JPWO2013129580A1 (en) | Approximate nearest neighbor search device, approximate nearest neighbor search method and program thereof | |
Lokoč et al. | Ptolemaic indexing of the signature quadratic form distance | |
CN106095951B (en) | Data space multi-dimensional indexing method based on load balancing and inquiry log | |
CN107545276A (en) | The various visual angles learning method of joint low-rank representation and sparse regression | |
CN107451200B (en) | Retrieval method using random quantization vocabulary tree and image retrieval method based on same | |
Kiranyaz et al. | Hierarchical cellular tree: An efficient indexing scheme for content-based retrieval on multimedia databases | |
WO2010062445A1 (en) | Predictive indexing for fast search | |
JP2002342136A (en) | Device and method for deciding clustering coefficient for database by using block level sampling | |
CN112860937A (en) | KNN and word embedding based mixed music recommendation method, system and equipment | |
CN105868355A (en) | Large-scale multimedia data spatial index method | |
Hua et al. | SamMatch: a flexible and efficient sampling-based image retrieval technique for large image databases | |
CN114972506B (en) | Image positioning method based on deep learning and street view image | |
Zacharatou et al. | Efficient bundled spatial range queries | |
Kriegel et al. | The performance of object decomposition techniques for spatial query processing |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
C06 | Publication | ||
PB01 | Publication | ||
C10 | Entry into substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20160817 |
|
RJ01 | Rejection of invention patent application after publication |