CN109739999A - A kind of efficient drosophila neural network Hash Search WMSN data method - Google Patents
A kind of efficient drosophila neural network Hash Search WMSN data method Download PDFInfo
- Publication number
- CN109739999A CN109739999A CN201910039907.1A CN201910039907A CN109739999A CN 109739999 A CN109739999 A CN 109739999A CN 201910039907 A CN201910039907 A CN 201910039907A CN 109739999 A CN109739999 A CN 109739999A
- Authority
- CN
- China
- Prior art keywords
- data
- matrix
- fjlt
- data set
- search
- 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
- 238000000034 method Methods 0.000 title claims abstract description 119
- 241000255581 Drosophila <fruit fly, genus> Species 0.000 title claims abstract description 37
- 238000013528 artificial neural network Methods 0.000 title claims abstract description 20
- 239000011159 matrix material Substances 0.000 claims abstract description 103
- 230000035945 sensitivity Effects 0.000 claims abstract description 40
- 238000004422 calculation algorithm Methods 0.000 claims description 53
- 230000008569 process Effects 0.000 claims description 25
- 230000007704 transition Effects 0.000 claims description 12
- 238000000605 extraction Methods 0.000 claims description 8
- 238000009826 distribution Methods 0.000 claims description 6
- 210000002569 neuron Anatomy 0.000 claims description 6
- 230000000717 retained effect Effects 0.000 claims description 4
- 238000012512 characterization method Methods 0.000 claims description 3
- 238000004364 calculation method Methods 0.000 claims description 2
- 210000000196 olfactory nerve Anatomy 0.000 abstract description 9
- 238000004088 simulation Methods 0.000 abstract 1
- 238000002474 experimental method Methods 0.000 description 25
- 229910002056 binary alloy Inorganic materials 0.000 description 9
- 230000014759 maintenance of location Effects 0.000 description 9
- 230000008447 perception Effects 0.000 description 8
- 238000011160 research Methods 0.000 description 8
- 210000005036 nerve Anatomy 0.000 description 6
- 238000012545 processing Methods 0.000 description 6
- 238000006243 chemical reaction Methods 0.000 description 5
- 238000013480 data collection Methods 0.000 description 5
- 238000005516 engineering process Methods 0.000 description 5
- 230000009467 reduction Effects 0.000 description 5
- 238000012360 testing method Methods 0.000 description 5
- 230000001149 cognitive effect Effects 0.000 description 4
- 230000006870 function Effects 0.000 description 4
- 230000009466 transformation Effects 0.000 description 4
- 230000003321 amplification Effects 0.000 description 3
- 230000000694 effects Effects 0.000 description 3
- 238000013507 mapping Methods 0.000 description 3
- 230000015654 memory Effects 0.000 description 3
- 238000003199 nucleic acid amplification method Methods 0.000 description 3
- RTZKZFJDLAIYFH-UHFFFAOYSA-N Diethyl ether Chemical compound CCOCC RTZKZFJDLAIYFH-UHFFFAOYSA-N 0.000 description 2
- 241000545744 Hirudinea Species 0.000 description 2
- 230000005856 abnormality Effects 0.000 description 2
- 230000009471 action Effects 0.000 description 2
- 230000004913 activation Effects 0.000 description 2
- 238000013459 approach Methods 0.000 description 2
- 230000008901 benefit Effects 0.000 description 2
- 230000008859 change Effects 0.000 description 2
- 238000001514 detection method Methods 0.000 description 2
- 238000011161 development Methods 0.000 description 2
- 238000005259 measurement Methods 0.000 description 2
- 230000007246 mechanism Effects 0.000 description 2
- 238000012544 monitoring process Methods 0.000 description 2
- 238000007781 pre-processing Methods 0.000 description 2
- 238000012827 research and development Methods 0.000 description 2
- 230000004044 response Effects 0.000 description 2
- 238000010845 search algorithm Methods 0.000 description 2
- NAWXUBYGYWOOIX-SFHVURJKSA-N (2s)-2-[[4-[2-(2,4-diaminoquinazolin-6-yl)ethyl]benzoyl]amino]-4-methylidenepentanedioic acid Chemical compound C1=CC2=NC(N)=NC(N)=C2C=C1CCC1=CC=C(C(=O)N[C@@H](CC(=C)C(O)=O)C(O)=O)C=C1 NAWXUBYGYWOOIX-SFHVURJKSA-N 0.000 description 1
- 241000406668 Loxodonta cyclotis Species 0.000 description 1
- 230000002159 abnormal effect Effects 0.000 description 1
- 238000004458 analytical method Methods 0.000 description 1
- 230000003796 beauty Effects 0.000 description 1
- 239000012141 concentrate Substances 0.000 description 1
- 238000007418 data mining Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 230000007613 environmental effect Effects 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 230000004927 fusion Effects 0.000 description 1
- 230000005764 inhibitory process Effects 0.000 description 1
- 230000010354 integration Effects 0.000 description 1
- 239000000203 mixture Substances 0.000 description 1
- 230000008555 neuronal activation Effects 0.000 description 1
- 238000010606 normalization Methods 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 238000003909 pattern recognition Methods 0.000 description 1
- 238000012913 prioritisation Methods 0.000 description 1
- 238000003672 processing method Methods 0.000 description 1
- 230000001629 suppression Effects 0.000 description 1
- 238000009827 uniform distribution Methods 0.000 description 1
Landscapes
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The present invention discloses a kind of efficient drosophila neural network Hash Search WMSN data method, first, data set is pre-processed, the feature of data set is set to be converted into numeric data, secondly utilize Fast Johnson-Lindenstrauss Transform (FJLT) projection matrix by data projection to more high-dimensional metric space, to ensure the content similarity between original data object, be conducive to accurately scan for;Finally, reducing the dimension of data set using the feature selecting strategy of winner-take-all in the local sensitivity hash method of drosophila olfactory nerves, search efficiency is improved.When the present invention is using drosophila olfactory nerves simulation local sensitivity Hash procedure, it is answered with preferably pervasive, the accuracy for improving search result efficiently solves the problems, such as the approximate NN Query of higher-dimension big data, is efficiently applied to based on WMSN data search affairs in WMSN application system.
Description
Technical field
The present invention relates to wireless multimedia sensor more particularly to a kind of efficient drosophila neural network Hash Search WMSN
Data method.
Background technique
Wireless multimedia sensor (Wireless Multimedia Sensor Networks, WMSN) is wirelessly to pass
The novel wireless sensor network with multimedia messages such as video, audio, images to grow up on sensor network (WSN).
Up to the present, WMSN is very extensive is applied to security monitoring, intelligent transportation, environmental monitoring etc., WMSN multimedia data query
It is an important core technology in the research and development of WMSN application system.WMSN multi-medium data usually have it is high-dimensional, extensive,
The characteristics of polymorphic type, conventional neighbor search algorithm are not able to satisfy the demand of system, and in recent years, research WMSN data are effectively searched
Suo Fangfa is to promote search performance, it has also become the hot issue that industrial circle is paid close attention to jointly with academia.
In general, people need to carry out dimension-reduction treatment to WMSN multi-medium data, generally there are two class dimension reduction methods: first is that feature
Selection, second is that feature extraction.Feature selecting is according to the correlation of feature and label during supervised learning, and selection most can generation
The feature of table data point indicates data.This method can effectively handle the problem of classification, but simultaneously as need to compare
The size of contribution of each feature of data object in assorting process in data set selects the feature most influenced relatively, should
Method time complexity usually with higher, and need data set with label, therefore this method is in proximity search problem
Upper application it is fewer.And feature extraction [1] is different with feature selection approach, maximum difference is that feature extraction passes through projection
It is converted on the feature of data object, creates a completely new feature set to indicate data, which does not need label
Data, and the original dimension of data set can be suitably reduced in projection process, therefore, this method is widely used in neighbour
Search problem also commonly uses this method in the neighbor search problem of multi-medium data.Most common feature extracting method have it is main at
Divide analytic approach (PCA), Linear Discriminant Analysis (LDA), Locally linear embedding
(LLE), the methods of Laplacian Eigenmaps (LE).
Currently, there are many research based on extensive high dimensional data searching method, and achieve it is significant it is significant into
Exhibition.2017 University of California Nian8Yue Sanjoy Dasgupta, CharlesF.Stevens and SaketNavlakha exist
" Science " proposes a kind of novel local sensitivity Hash FLSH (Fly Local Sensitive based on accidental projection
Hash, FLSH) method [2], with the Hash procedure of the olfactory nerves simulated data sets of drosophila, the method that they are proposed is to perception
The processing method of the local sensitivity Hash LSH (Local Sensitive Hash, LSH) of the perception sum of neuron provides
Very effective integration program, this method breach the yoke that scholars study LSH for a long time, the research to related fields
Produce great influence.Before the proposition of FLSH method, people are accustomed to being studied in the general field of LSH to generate novelty
Searching method.
In general, extensive high dimensional data nearest neighbor search method can generate a large amount of time and space expense, need
Dimension-reduction treatment achievees the purpose that reduce query time complexity.The Johnson of Texas ,Usa A&M university in 1984 and
Lindenstrauss proposes famous J-L theorem [3], provides the support that theorizes for the correlation technique of high dimensional data dimensionality reduction.
Piotr Indyk and the Rajeev Motwani of Stanford University in 1998 propose local sensitivity on the basis of J-L theorem
Hash algorithm (Locally sensitive hash, LSH) [4], gives and provides for extensive high dimensional data nearest neighbor search
Landmark solution.Since LSH method proposes, because its superior function in high dimensional data neighbor search field obtains
The extensive concern for having arrived academia and industry, by it is vicennial development and it is perfect, LSH method neighbor search field,
Many successful examples are produced in the related research of many application fields such as pattern-recognition, abnormality detection.But LSH also has
Significant limitation, firstly, inevitably loss section divides similitude when the feature to data set carries out Hash processing,
Secondly, Hash procedure is also required to part-time consumption.How under the premise of guaranteeing time complexity, reduce as far as possible similar
The loss of property, is always the target that scholars study.Finally, how to construct between biology perception nerve and data set Hash procedure
Bionical relationship, equally the development of local sensitive hash algorithm is had a major impact.
The FLSH method that the scholars such as Sanjoy in 2017 are proposed on Science periodical is searched in extensive high dimensional data
Although the advantages of there are many in terms of rope, the local sensitivity Hash Search based on accidental projection that FLSH method tundish contains
Process has certain problems.It is mainly manifested in following three aspects:
Problem 1 (the choice problem of search result accuracy and time efficiency): during high dimensional data neighbor search, I
Be frequently necessary to tradeoff search result accuracy rate and the precision of search time on weighed, although the office based on accidental projection
Portion's sensitive hash strategy reduces the time complexity of neighbor search, but is lost part precision simultaneously;
Problem 2 (is hidden between the processing and nerve perception of data set and is contacted): traditional based on the close of local sensitivity Hash
It is to handle in search target (data set) data object like neighbor search, initiates the main body identification data of query actions
Often studied personnel ignore for the activity of perception neuron when object.
Problem 3 (Problem of Universality of the J-L theorem in High Dimensional Data Set dimensionality reduction): J-L theorem is reducing data dimension
Degree field has obtained extensive reference, but about FJLT logm according to the new feature for collecting progress space translated data collection
The regularity of distribution, there are also to be studied at present for influence of the new feature to data point.
The neighbor search problem of high dimensional data is always a research hotspot of the field of data mining.Early in 1984, beauty
Scholar the William B.Johnson and Joram.Lindenstrauss of Texas A&M university, state have proposed J-L theorem
[3], theory support is provided for the dimensionality reduction conversion of high dimensional data.But they not provide High Dimensional Data Set effective
Neighbor search method.Until 1998, the scholar Piotr Indyk of Stanford Univ USA et al. on the basis of J-L theorem,
The local sensitivity hash method [4] based on Hamming space is proposed, provides effective solution for the neighbor search of High Dimensional Data Set
Certainly method, so-called " dimension curse " puzzlement problem have obtained effective solution scheme.Since then, local sensitivity hash method starts
The neighbor search field of High Dimensional Data Set is shown up prominently, and the extensive concern of academia and industry is obtained.But first generation office
Portion's sensitive hash is mainly manifested in there is also some defects: when handling data, needing for data to be mapped to Hamming sky one by one
Between, greatly reduce the efficiency of our neighbor search;In addition, first generation local sensitivity hash method there is also space hold compared with
Big problem.
Scholar MayurDatar, the Nicole Immorlica and Piotr Indyk et al. of the Massachusetts Institute of Technology in 2004
It is proposed the local sensitivity hash method [5] stable based on p, this method avoid the process that data set is mapped to Hamming space,
Reduce the space complexity of first generation local sensitivity Hash strategy simultaneously, and on time complexity and inquiry precision and with
Past algorithm, which is compared, to be improved.Since then, this method is in image recognition, and paper duplicate checking, the fields such as abnormality detection have obtained extensively
General application.But this method the problems such as there is still a need for a large amount of memory spaces, the index of large-scale data is not suitable for, secondly, base
The result and unstable of index coding is generated in probabilistic model.Although number of encoding bits increase, the raising ten of accuracy rate is inquired
Divide slow.
The scholar AlexandrAndoni and PiotrIndyk of Massachusetts Institute Technology in 2006 propose a kind of new
Local sensitivity hash method [6], their method combine accidental projection and Leech Lattice, so that it is multiple to reduce the time
Miscellaneous degree and space complexity.Reduce data set and project the time it takes, improves the execution efficiency of algorithm.But the party
The search accuracy rate of method is largely influenced by lattice selection and accidental projection, therefore accuracy rate is difficult to be protected
Card.Scholar NirAilon and the Bernard Chazelle [7] of Princeton university in 2006 are proposed based on J-L theorem
FJLT move on to image method, and carried out theoretical proof.This method combines sparse matrix and the Heisenberg of Fourier transformation is former
Reason, improves the speed of the projection of data set, meanwhile, the effective similitude guaranteed in data set between data object makes to count
It being capable of similarity as far as possible before retaining projection in data set between data object according to the data set after collection projection.But
NirAilon et al. only theoretically demonstrates the superiority of this method, and there is no further prove that this method exists by experiment
The superiority of actual application.In this stage, the research for local sensitivity hash algorithm is had been greatly developed, very
More scholars, which are absorbed in, optimizes processing to data, such as the scholar QinLv and Wiliam of Princeton University in 2007
Josephson et al. proposes the local sensitivity hash method [8] that one kind detects more, and this method is by being equipped with part number to multiple
It is detected simultaneously according to the bucket of collection, further improves the accuracy rate of neighbor search;California, USA university Berkeley point in 2009
The scholar BrianKulis in school and the scholar KristenGrauman of university texas at austin are proposed based on kernel function
Core Hash (Kernelized Local Sensitive Hashing, KLSH) algorithm [9], the algorithm is with the random of input vector
Main support [10] are projected as, the sublinear approximation neighbor search time is realized, reduces the time cost of search process;2012
Year, scholar VenuSatuluri and the Srinivasan Parthasarathy of Ohio State University were proposed based on Bayes
Quick similarity query method BayesLSH [11], this method is easy to tuning, and does not need manual setting Hash coding
Length.
Research for local sensitivity hash algorithm in nearly 2 years has many prioritization schemes, such as Colombia in 2016
The scholar of scholar AlexandrAndoni, IBM graduate the researcher Thijs Laarhoven and the Massachusetts Institute of Technology of university
Ilya Razenshteyn et al. has weighed time complexity and space complexity [12] in approximate neighbor search, it is intended to find out most
Excellent solution;The scholar Ilya Razenshteyn of the Massachusetts Institute of Technology in 2017 and the scholar of Columbia University
The LSH-FOREST algorithm [13] that AlexandrAndoni is proposed changes the structure of traditional local sensitivity Hash;The U.S. in 2017
The printenv that the scholar TeresaNicoleBrooks and RaniaAlmajalid of Pei Si university is proposed on the basis of [8] it is more
Face ball-type local sensitivity Hash hashes the time complexity [14] for optimizing neighbor search;In 2018 University of Illinois, the U.S.
Person KarthekeyanChandrasekaran et al. proposes the local sensitivity hash algorithm [15] based on grid, by draw
The process of divided data collection applies control condition to improve the inquiry precision of neighbor search.This paper seminar Fujian China in 2018
Normal university Xiao for example good equal [16] is directed to approximate neighbour's search problem of large-scale data, establishes two end number mixing mould using LSH
The high-efficiency search method of state data.Undeniably, the optimization proposed in this serial of methods, including document [17,18,19,20]
Method, further improves the performance of local sensitivity Hash, also extends the ability of local sensitivity Hash.But in this hair
During exhibition, few scholars consider bionical class Intelligentized method and part sensitive hash process to combine, with into one
Step is expanded.
Scholar the Sanjoy Dasgupta, CharlesF.Stevens of 2017 Universities of California Nian8Yue and
A kind of novel local sensitivity hash method [2], the party are proposed in the article that SaketNavlakha is delivered on Science
Method for the first time gets up the perception nerve connections of the process of local sensitivity Hash and cognitive subject, is simulated by the olfactory nerves of drosophila
The process of local sensitivity Hash, and using random sparse binary system projection matrix, by accidental projection to data set
Intrinsic dimensionality amplifies, the activation of imictron, this point is very different with previous local sensitivity hash method, finally
After reducing projection by the neuron inhibition tactful (Anterior Paired Lateral, APL) of winner-take-all (WTA)
The intrinsic dimensionality of data, their experimental result show that their used random sparse binary system projection matrixes are opposite
In the experiment effect that the intensive projection matrix of Gauss has had, show that the bionical Hash of drosophila smell can be calculated with local sensitive hash
Method is effectively combined.The maximum contribution of their local sensitivity Hash strategy is to change traditional local sensitivity Hash pair
The mode of data processing, it is established that the perception nerve of cognitive subject and approximate neighbor search connection construct neuronal activation
Relationship between mode and data projection, to provide a new research direction for approximate neighbor search.
Local sensitivity hash method (Fly Local Sensitive Hashing, FLSH) based on drosophila olfactory nerves
There are still some problems.We have found that its sparse binary system random matrix used is not suitable for the data projection [7] of low distortion, though
The speed of data projection so is improved, but sacrifices the part similitude between data object simultaneously as cost, it is thus impossible to
Meet requirement of the WMSN application system to inquiry precision;The stability that we also identify FLSH algorithm under study for action needs into one
Step is promoted, and the inquiry precision of the algorithm on different data sets differs larger, the enquiry machine not being suitable in WMSN application system
System;Finally, the feasibility of the algorithm lacks strong theory support, although sparse binary matrix can preferably simulate mind
Activation pattern through member, but lack corresponding theory support in the treatment process of data set.
Summary of the invention
The purpose of the present invention is to provide a kind of efficient drosophila neural network Hash Search WMSN data methods.
The technical solution adopted by the present invention is that:
A kind of efficient drosophila neural network Hash Search WMSN data method comprising following steps:
Step 1, feature extraction is carried out to multi-medium data, multi-medium data is converted into feature vector number by characterizing
According to;
Step 2, building search indexes after being quickly converted matrix projection using FJLT to the characteristic vector data on data set;
Step 3, query object is mapped in search index structure;To the given data object for needing to inquire using FJLT
The index for inquiring data is formed after being quickly converted matrix projection;
Step 4, the index based on inquiry data carries out approximate neighbor search in search index, searches and inquiry point data
Most like data object.
Further, text data is converted under theorem in Euclid space using TF-IDF method or word frequency method in step 1
Feature vector;Characterization is carried out to image data by extracting SIFT feature value.
Further, step 2 specifically includes the following steps:
Step 2.1, matrix is quickly converted by FJLT and each data object in data set is projected to new measurement sky
Between,
Step 2.2, neuron winner-take-all is enlivened in conjunction in drosophila neural network local sensitivity Hash FLSH algorithm
(WTA) strategy accepts or rejects the new feature of data object each after projection respectively, using the feature retained as the data pair
The index of elephant.
Further, the calculation formula for being quickly converted matrix projection using FJLT of step 2 is as follows:
Y=P × H × D × X
Wherein, Y is data after projection, and X is the characteristic vector data in data set,
P matrix is one using J-L theorem as the sparse matrix of theory support, and it is 0 that each element in P matrix, which meets mean value,
And variance is q-1Be just distributed very much, i.e., each element in P matrix meets Pij- N (0, q-1) probability be q, be not inconsistent in matrix
Close Pij- N (0, q-1) distribution element value be 0, wherein the calculating process of q is as follows:
Wherein, d is the original dimension of data set, and n is the size of data set, and ε is to protect away from performance parameter, and p is used
Normal form;
H is a hadamard matrix related with original dimension d, each element H thereinijValue can be calculated by following formula
Go out,
Wherein, d is the original dimension (meanings of parameters in supplementary type) of data set.
D is a diagonal matrix, and for the value on diagonal line between -1 and 1, -1 and 1 respectively has 1/2 probability, dimension d
D indicate are as follows:
Wherein, d is the original dimension of data set.
The invention adopts the above technical scheme, replaces original FLSH local sensitivity using the PHD transition matrix in FJLT
The simple sparse matrix of hash method;" winner-take-all " strategy of drosophila smell and the FJLT based on J-L theorem are converted and carried out
It is effectively combined, has obtained more accurate experimental result on the basis of not increasing time loss;FJLT is quickly converted matrix
It is optimized on time complexity.The present invention changes the prior art and only knows in the similitude that the object of identification is handled
Not, it is realized using the similitude identification of drosophila olfactory nerves from identification object to the transformation of cognitive subject, has found similitude
Retrieval is potentially contacted with cognitive subject, on the basis of controlling time complexity, further promotes high dimensional data neighbor search
Accuracy.
Detailed description of the invention
The present invention is described in further details below in conjunction with the drawings and specific embodiments;
Fig. 1 is FJLT-FLSH method overall framework of the present invention;
Fig. 2 is the figure compared with flshon propagation size keeps at a distance performance of the fjlt-flsh based on data set mnist;
Fig. 3 is the figure compared with flshon propagation size keeps at a distance performance of the fjlt-flsh based on data set sift;
Green line and blue line indicate that data set features are expanded to M and tieed up by FJLT-FLSH algorithm and FLSH algorithm, and in all spies
Sign is lower to keep inquiry precision.
Fig. 4 is that the projection time of FJLT matrix and sparse binary matrix based on data set mnist compares figure;
Fig. 5 is that the projection time of FJLT matrix and sparse binary matrix based on data set sift compares figure;
Fig. 6 is data set sift FJLT-FLSH (WTA), FLSH (WTA), FJLT- in Hash code length M=100
FLSH (LTA), FLSH (LTA) algorithm inquiry precision compare figure;
Fig. 7 is data set sift FJLT-FLSH (WTA), FLSH (WTA), FJLT- in Hash code length M=200
FLSH (LTA), FLSH (LTA) algorithm inquiry precision compare figure;
Fig. 8 is data set sift FJLT-FLSH (WTA), FLSH (WTA), FJLT- in Hash code length M=300
FLSH (LTA), FLSH (LTA) algorithm inquiry precision compare figure;
Fig. 9 is data set sift FJLT-FLSH (WTA), FLSH (WTA), FJLT- in Hash code length M=600
FLSH (LTA), FLSH (LTA) algorithm inquiry precision compare figure;
Figure 10 is data set sift FJLT-FLSH (WTA), FLSH (WTA), FJLT- in Hash code length M=1000
FLSH (LTA), FLSH (LTA) algorithm inquiry precision compare figure;
Figure 11 is that data set sift fjlt projection matrix in Hash code length M=100 uses wta, lta, random spy
Sign keeps the inquiry precision of strategy to compare figure;
Figure 12 is that data set sift fjlt projection matrix in Hash code length M=200 uses wta, lta, random spy
Sign keeps the inquiry precision of strategy to compare figure;
Figure 13 is that data set sift fjlt projection matrix in Hash code length M=300 uses wta, lta, random spy
Sign keeps the inquiry precision of strategy to compare figure;
Figure 14 is that data set sift fjlt projection matrix in Hash code length M=600 uses wta, lta, random spy
Sign keeps the inquiry precision of strategy to compare figure;
Figure 15 is that data set sift fjlt projection matrix in Hash code length M=1000 uses wta, lta, random spy
Sign keeps the inquiry precision of strategy to compare figure.
Specific embodiment
As shown in one of Fig. 1-15, the invention discloses a kind of data sides efficient drosophila neural network Hash Search WMSN
Method.In WMSN Internet of Things application system, the block linkwork based on current hotspot block chain technology and IPFS technology building WMSN
System [25,26,27,28] is a very promising technical solution, can be to very sensitive crucial number in opening and shares
According to being protected.The program is typically based on ether mill and IPFS to construct, and WMSN data are after storing to IPFS distributed frame
Cochain processing is carried out, inquiry operation is carried out to the data after cochain on IPFS by intelligent contract.Thus, WMSN multimedia number
According to search inquiry by be WMSN block catenary system research and development in a basic design.Because multi-medium data usually have compared with
High characteristic dimension, therefore, conventional neighbor search algorithm are not able to satisfy the demand of system.We combine [29,30,31,32]
What is proposed solves the search inquiry scheme in the application of block chain, proposes a kind of new WMSN for being known as FJLT-FLSH method
High dimensional data searching method, based on this, we can construct intelligent contract in subsequent work, use for WMSN block chain
Family provides query search application.
The FJLT-FLSH method incorporates the main thought of drosophila neural network Yu FJLT matrixing, is a kind of new
The search process key step of the method for local sensitivity Hash Search WMSN data, high dimensional data is constituted, as shown in Figure 1.
(1) carry out feature extraction to multi-medium data: this step is data preprocessing phase, in the original more of input WMSN
After categorical data, multi-medium data is converted into feature vector by characterizing;In this process, text data is carried out special
Sign is extracted, and generallys use TF-IDF method or word frequency method, text data is converted to the feature vector under theorem in Euclid space;It is right
When image data carries out feature extraction, characterization is carried out by extracting SIFT feature value.
(2) to the characteristic building search index on data set: after the completion of feature extraction, the characteristic dimension of data object
Excessively high, a large amount of memory can be occupied and take a substantial amount of time by directly carrying out force search, it is therefore desirable to be thrown data set
Shadow, and Hash Index Structure is constructed on data set in the projected, so as to quickly execute function of search.It is walked by two
Suddenly the operation: the first step is completed, by projection matrix FJLT, it is empty that each data object in data set is projected into new measurement
Between, second step will enliven neuron winner-take-all (WTA) strategy knot in drosophila neural network local sensitivity Hash FLSH algorithm
It is incorporated, the new feature of data object each after projection is accepted or rejected respectively, using the feature retained as the data object
Index.
(3) query object is mapped in search index structure: inquiry phase, the given data object for needing to inquire, warp
After crossing numerical characteristicsization processing, equally query object is projected using FJLT matrix, using the feature retention strategy of WTA,
The new feature of inquiry data to be accepted or rejected, the feature of reservation is to inquire the index of data,
(4) in new metric space, approximate neighbor search is carried out by the index of generation, is found and inquiry point data most phase
As data object.
Next, we are unfolded stage by stage for proposed FJLT-FLSH algorithm frame.
Accidental projection based on FJLT matrix: FJLT transition matrix appears in Princeton University scholar NirAilon earliest
It is published in the article on STOC [7] with Bernard Chazelle2006, the major function of the algorithm is that Fourier is become
The projection of heisenberg principle and sparse matrix in changing combines, and reduces the distortion level during data projection, adds simultaneously
The projection process of fast data set." low distortion " and " fast projection " of FJLT transition matrix is theoretically proven
[7], proof [34] have also been obtained in superiority in use.The input of this method is image or text data warp
Data after pre-processing are crossed, the feature of the data is numerical value, and output is one group of new numeric data feature, these new numerical value numbers
According to representing original data set in the mapping of FJLT projector space.When selecting the projection matrix of local sensitivity Hash (LSH), square is projected
The guarantor of battle array is away from property, and the speed of projection is to determine whether feasible two deciding factors of local sensitivity hash method.FJLT
The property of matrix can satisfy local sensitivity Hash and be protected in projection process away from the requirement with low time complexity, use FJLT square
Battle array improves FLSH method, can promote the search efficiency of inquiry mechanism in WMSN.
The FJLT-FLSH local sensitivity hash method of drosophila olfactory nerves and FJLT fusion: proposed by the present invention to be based on
The drosophila algorithm of FJLT is pre-processed by data set, and the projection transform and data characteristics of data set choose three parts composition.In number
According to collection pretreatment stage, groundwork is by image data set, and lteral data collection and other nonnumeric type data sets are converted into
Digital data collection.In the second step stage, data set is projected, Sanjoy Dasgupta et al. is used in document [1]
Random, sparse binary matrix projects raw data set, and sparse binary matrix improves data set projection
Speed, but part precision is sacrificed simultaneously, we are quickly converted the sparse binary system that matrix replaces them using FJLT herein
Logm is projected according to collection.
The hash algorithm of nerve based on drosophila smell has higher requirement to projection matrix, firstly, because data set
The phenomenon that will appear " distortion " in projection process, therefore, transition matrix, need the phase after protection data projection to greatest extent
Adjust the distance, secondly as need to amplify the intrinsic dimensionality of data object in experimentation, therefore, transition matrix
Time complexity needs in projection process meet the needs of experiment.FJLT transition matrix is in " protect away from " performance and time complexity
On have excellent performance [7,34].In conclusion the FLSH algorithm and FJLT of present invention combination drosophila olfactory nerves are quickly thrown
Shadow algorithm proposes a kind of new local sensitivity hash method FJLT-FLSH.
During using FJLT transition matrix, the parameter required special attention has ε and p, and the former projected
The fluctuation range of journey relative distance can derive by formula (8), data of the data set after FJLT is projected, after conversion
The fluctuation range for collecting relative distance is [(1- ε) αp, (1+ ε) αp], whereinα2=k, k indicate that data object is thrown
The dimension of movie queen, according to J-L theorem, it is only related with the size of data set and ε, unrelated [3] with the former dimension d of data set;P table
Demonstration formula, usual value range are { 1,2 }.
FJLT-FLSH algorithm is embedded in FJLT transformation matrix, which meets two following properties:
It (1), include two below set X, ε < 1and p ∈ { 1,2 } Φ~FJLT. of the Rd of n point for any one
A event at least 2/3 can occur.
1, for any x ∈ X, (1- ε) αp||x||2≤||Φx||p≤(1+ε)αp||x||2 (12)
In formulaThe formula is the deformable body of J-L theorem
2, mapping φ: Rd → RkTime complexity are as follows:
O(dlogd+min{dε-2Logn, εp-4logp+1n}) (13)
The above property is similar to scholar NirAilon and the Bernard Chazelle of Princeton university in 2006
[7] it proposes the FJLT based on J-L theorem and moves on to image method.
(2), in FJLT-FLSH algorithm projection matrix time response:
Projection operator in FJLT-FLSH algorithm determines by FJLT, matrix Dd×dTime for diagonal matrix, D (x) is complicated
Degree is O (d), and matrix H is Walsh-Hadamard matrix, and the time complexity of H (Dx) is O (dlogd), finally, P (HDx)
Time complexity is O (| P |), | P | in representing matrix P for 0 element number, | P | obey B (nk, q) distribution, can push away
Export
E [| P |]=O (εp-4logp+1n) (14)
Due in P matrixTherefore, it can be deduced that the time complexity of matrix P is
min{dε-2Logn, εp-4logp+1N }, so the time complexity of FJLT transition matrix is O (dlogd+min { d ε-2Logn, εp- 4logp+1n}).Markov boundary, which demonstrates FJLT lemma, ensure that the projection time of FJLT
The above property is similar to the scholar AlexandrAndoni and PiotrIndyk of Massachusetts Institute Technology in 2006
A kind of new local sensitivity hash method [6] is proposed, their method combines accidental projection and Leech Lattice, from
And reduce time complexity and space complexity.The property is theoretically illustrated in use FJLT matrix to high dimensional data pair
When as carrying out dimensionality reduction operation, original dimension d, the size n of data set are protected away between performance parameter ε and used normal form p
Relationship.It is of great significance for the parameter selection during dimension transformation of the present invention.
The proof of time response property:
(Without loss of generality) without loss of generality, if ε be one close enough 0 and be greater than 0 number,
Define variable h=HD (x)=(h1..., hd)T, it is assumed that W.1.o.g set up, then | | x | |2=1, by H, the generation of D matrix
Formula (10), (11) can be derivedai=± d-1/2, pass through H, the property of D matrix, it can be deduced that knot
By: aiIt is independent and uniform distribution.
Chernoff-type shows that the probability of formula (15) at least 19/20 is set up;
Obviously, by general mathematical proof, formula (16) is centainly set up;
Therefore, it by markov injustice scheduling theory, can derive;
Formula (14) defers to joint circle rule (The union bound:P (A1∪A2∪...∪Ak)≤P(A1)+P(A2)
+…+P(Ak)), thenFor arbitrary n and d, nd < n2, it was therefore concluded that:
Since matrix H and matrix D are all isometric matrixes, so h=HD (x) is equidistantly to change, i.e., | | h | |2=| | x | |2,
If y=(y1, y2..., yk)T=Ph=Φx, available according to the definition of three matrixes in FJLTIts
In, cjThe probability that value is 1 is q, and obeys random independent same distribution.Another parameter rjObey N (0, q-1), if By 2 it is stable be just distributed very much, obtain (y1| Z=z)~N (0, q-1z).Due to y1, y2..., ykIt is to obey independently
With distribution, therefore obtain Z=Z1, Z2..., ZkSame obey is independently distributed, i.e. E [Z]=q.Princeton in 2006
Scholar NirAilon and the Bernard Chazelle [7] of university, which propose the FJLT based on J-L theorem and moves on to image method, to be provided
FJLT lemma is in l2Proof under normal form.
The present invention uses two datasets, and first data set is that handwritten word identifies Mnist data set, second data set
For image data set Sift, wherein the intrinsic dimensionality of Mnist data set is the intrinsic dimensionality dimension 128 of 784, Sift data set.?
In experimentation, the subset that the size for choosing each data set is 10000 is tested.
Our experimentation is divided into three groups:
Experiment one: respectively expanding to two datasets M dimension (M=100,200,300,600,1000) first, extends
M dimension afterwards all retains, directly progress neighbor search, is respectively compared FJLT and sparse binary matrix in different data collection
Guarantor when extension dimension is away from performance;
Experiment two: after data set extension is to different dimensions M, using different projection matrix FJLT and sparse binary system
Projection matrix, uses two kinds of feature retention strategies: winner-take-all (WTA), and weak person covers all (LTA).Retain identical Hash coding length
The intrinsic dimensionality of k (k=cM, c=0.01,0.02,0.03,0.05,0.1,0.20.3,0.5) is spent, respectively to FJLT-WTA,
FLY-WTA, the inquiry precision of FJLT-LTA, FLY-LTA method compare, and judge that two kinds of algorithms are excellent in inquiry precision
It is bad;
Experiment three: the projection property in order to further analyze FJLT projection matrix, we choose different features and protect
Stay strategy: WTA, LTA, RANDOM analyze the feature of the data set by the conversion of FJLT projection matrix, to find its spy
Sign rule.
During the experiment, the main judgment criteria for using searching accuracy and time (second) to test as the present invention, it is first
First, set the scale N of data set, initial data concentrate find out and query point in Euclidean space closest to S it is close
Adjoint point, then the S point closest in Euclidean space with the data set after projection carries out the operation that seeks common ground, with weight
The number of multiple Neighbor Points is obtained divided by S, the inquiry precision of the query point.As shown in formula (18).Wherein x ∈ X, X indicate number
According to collection;P indicates to inquire the probability of Neighbor Points;F is represented in data set in luv space, and query point x is in theorem in Euclid space next week
S nearest point is enclosed, R is nearest S point around query point x under theorem in Euclid space in the projected;| | return is set
Length.
In order to keep experimental result more convincing, in experimentation of the invention, each parameter repeats 20 times
1000 points of random challenge are tested in experiment every time, then calculate the average value of this 20 times 1000 points inquiry precision every time, as
Experimental evaluation standard of the invention.
Experiment and interpretation of result:
Firstly, we compared us and project square using FJLT in the case that dimensional characteristics after expansion all retain
Inquiry of the sparse binary system random matrix on Mnist data set and Sift data set used in battle array and drosophila algorithm is accurate
Degree.
Experiment one: the purpose for designing the experiment is to judge that two kinds of projection matrixes expand to data set intrinsic dimensionality
Guarantor when exhibition is away from performance, during the experiment, uses two different projection matrix FJLT and sparse binary matrix first
Data set features dimension is extended to M dimension, all intrinsic dimensionalities of the data object after retaining extension directly carry out neighbour and look into
It askes." protect away from " performance when comparing two kinds of matrix-expand dimensions.In the specific experiment stage, S value 200 calculates mapping front and back
Similitude of 200 points as us around query point.
As shown in figures 2 and 3, the amplitude amplified with data set dimensional characteristics is increasing, and the FJLT that the present invention uses is fast
Advantage of the fast transition matrix in inquiry precision is increasing.This shows that FJLT transition matrix is in data set used in us
There is better experiment effect during the amplification of dimensional characteristics.The core strategy of drosophila algorithm is exactly to simulate the smell of drosophila
Neural network first amplifies the signal of input, then, utilizes the plan of the winner-take-all (WTA) in drosophila algorithm FLSH
Slightly, make the relatively small feature of maximum remaining characteristic value of k feature suppression of characteristic value, to obtain the Kazakhstan of local sensitivity Hash
Uncommon coding, as shown in figures 2 and 3, FJLT algorithm proposed by the present invention has higher inquiry precision in the amplified space of feature,
Therefore our FJLT projection matrix is more suitable the simulation algorithm of drosophila olfactory neural network.
The comparison of FJLT-FLSH and FLSH hold in range performance in extension dimension in 1. two datasets of table.
Experiment two: experiment purpose is to compare the time it takes when two kinds of projection matrixes project data set.
Compare the time efficiency and time trend of two kinds of matrixes.During the experiment, firstly, it is necessary to which data set is amplified to M
Then dimension directly intercepts projection process the time it takes.Matrix and data set are converted into array in projection process by we,
The process is necessary, and can greatly shorten projection time;It is tested every time in experimentation and runs a program only to prevent
Because other programs occupy computer resource, the experimental result of the secondary experiment is influenced;Average value is calculated using many experiments, to reduce
Interference of the abnormal experimental result to normal experimental result;Our experimental situation is 10 system of Window, processor i5-
6500,8G memories, PyCharm compiler.
As shown in fig. 4 or 5, the experimental results showed that, FJLT of the present invention is quickly converted matrix and drosophila article
Used sparse binary system random matrix [2] compare have better time performance, this expression: FJLT matrix more suitable for
The treatment process of large-scale dataset.Meanwhile from experimental result, we can also conclude that the increasing with extension dimension M
Add, the projection time of two kinds of projection matrixes used by testing linearly increases, this is consistent with actual conditions.But FJLT
The rate of rise of projection matrix is far below sparse binary matrix.Since the growth rate of FJLT projection time is smaller, because
This FJLT matrix, which is more suitable, is extended the dimension of data set.
The projection time of FJLT matrix and sparse binary matrix compares in 2. two datasets of table
Experiment three: experiment purpose is the inquiry precision in order to compare two kinds of algorithms under identical Hash coding, first when experiment
First data set is pre-processed using the division normalization [23] of nerve perception, the M of Sift data set amplification is tieed up (M=
100,200,300,600,1000) inquiry that two kinds of algorithms are covered all under (LTA) in winner-take-all (WTA) and weak person, is then calculated
Precision.Retain after two kinds of algorithms are converted respectively in M dimensional feature value in maximum k feature (k=cM) and M dimensional feature value most
K small feature, as Hash code length.Compare inquiry precision of two kinds of algorithms under different characteristic retention strategy.
As shown in one of Fig. 6-10, during the experiment, the present invention compares FJLT matrix and 2012 years Ohio first
Scholar VenuSatuluri and the Srinivasan Parthasarathy of state university propose the quick phase based on Bayes
Like inquiry precision of the sparse binary matrix in property querying method BayesLSH [11] under WTA (winner-take-all), by right
Than discovery, in the case where k value takes minimum, sparse binary system projection matrix is quickly converted matrix with higher compared to FJLT
Accuracy, still, under this section of code length, the inquiry precision of two kinds of algorithm entirety is not high, with the increase of code length,
FJLT-FLSH algorithm queries precision starts to surpass in reverse FLSH algorithm, and highest can achieve 83% inquiry precision,
It is compared by experimental result, it can be deduced that conclusion, with the increase of the amplitude of data set amplification dimension, FJLT-
The superiority of FLSH algorithm is more prominent.
When the hash code length of table 3. takes m=600, fjlt-flsh (wta), flsh (wta), fjlt-flsh (lta), flsh
(lta) the inquiry precision of algorithm
Comparing two kinds of algorithms after the inquiry precision under winner-take-all (WTA) feature retention strategy, we are also to warp
Inquiry precision of the new feature of data set after crossing two kinds of algorithm conversions under other feature retention disciplines compares, first
First, we cover all to two kinds of algorithms in weak person and have done the comparison of precision on (LTA), it has been found that inquiry of the FLY algorithm on LTA
Precision is relatively low, and inquiry precision curve of the FJLT matrix that the present invention uses on LTA be basic and the inquiry precision on WTA
Curve essentially coincides, this shows when carrying out neighbor search, the feature tribute having the same that the feature and WTA that LTA chooses are chosen
It offers, guesses that the feature after FJLT is converted all has characteristic of equal value one by one to verify us, we devise a reality
It tests, in the data set after FJLT is converted, randomly selects k feature and encoded as Hash, then retained with WTA and LTA
The inquiry precision of feature be compared,
Experiment four: it is distributed to verify the contribution of the feature of the data set after FJLT is converted, we use three kinds of spies
Retention strategy is levied, the first is winner-take-all (WTA), that is, chooses maximum k feature as our Hash coding, be for second
Weak person covers all (LTA), and the Hash for choosing the smallest k characteristic value as us encodes, the third is to randomly select k characteristic value
Hash as us encodes (k=cM, c=0.01,0.02,0.03,0.05,0.1,0.13,0.15,0.2,0.3,0.5), so
Progress neighbor search calculates the inquiry precision between them and compares and analyzes afterwards.Firstly, testing according to upper one, FJLT is used
Transition matrix ties up the dimension of data set to M, then, retains its maximum k characteristic value respectively, the smallest k characteristic value, and
K random characteristic value carries out neighbor search, calculates inquiry precision.In order to keep experimental result closer to the truth.It generates
Different query points calculate the inquiry precision of three kinds of different characteristic retention strategies.
As shown in one of Figure 11-15, it can be deduced that conclusion is converted by FJLT-FLSH algorithm in the data set after projection,
Influence of the feature of data to inquiry precision be it is identical, i.e., arbitrarily take k feature after projection in data set, can obtain closely
Like identical inquiry precision, this characteristic can bring us and inspire, if a kind of completely new feature retention strategy can be found,
Under this policy, inquiring precision and query time can further be promoted.
When the hash code length of table 4. takes m=600, fjlt projection matrix keeps rule in (wta, lta, random) three features
Inquiry precision under then
Experiments have shown that FJLT projection matrix is more applicable for the data characteristics dimension in WMSN block catenary system inquiry mechanism
It is extended, under identical and moderate Hash code length, FJLT-WTA algorithm proposed by the present invention has in inquiry precision
There is superior performance, meanwhile, time of the FJLT matrix spent by projection process is much smaller than general sparse, binary system square
Battle array, this shows that FJLT matrix is more suitable in WMSN system large-scale data neighbor search.Finally, due to which FJLT matrix is certainly
The characteristics of band standardization, the data characteristics after the conversion of FJLT projection matrix is to data contribution having the same.Therefore, it is selecting
When taking characteristic dimension, we can use the lower strategy of time complexity, reduce the time in our query process, further
Improve time efficiency.
Of the invention is more efficient compared with prior art, and WMSN application system is improved in the case where guaranteeing time complexity
The inquiry precision of high dimensional data in system meets us using the query demand of multi-medium data in WMSN block catenary system.It is real
Test the result shows that, Hash strategy of the invention drosophila olfactory nerves identify smell during to odiferous information intrinsic dimensionality carry out
When extension, has and preferably protect away from performance, higher inquiry precision, while greatly reducing projection time, therefore be more suitable
Data intrinsic dimensionality is extended in the bionical Hash procedure of data, meanwhile, retain phase choosing identical feature retention strategy
With length characteristic dimension when, the present invention has higher inquiry precision, meets WMSN block catenary system to inquiry precision
It is required that and the projecting method that uses make the feature of the data set after projection to inquiry precision contribution having the same, can be with
It is random to retain specified Hash code length, secondary treatment of the strategy of winner-take-all (WTA) to data characteristics is needed not move through,
To improve the time efficiency in query process.
Claims (7)
1. a kind of efficient drosophila neural network Hash Search WMSN data method, it is characterised in that: itself the following steps are included:
Step 1, feature extraction is carried out to multi-medium data, multi-medium data is converted into characteristic vector data by characterizing;
Step 2, building search indexes after being quickly converted matrix projection using FJLT to the characteristic vector data on data set;
Step 3, query object is mapped in search index structure;It is quick using FJLT to the given data object for needing to inquire
The index of inquiry data is formed after transition matrix projection;
Step 4, the index based on inquiry data carries out approximate neighbor search in search index, searches and inquiry point data most phase
As data object.
2. a kind of efficient drosophila neural network Hash Search WMSN data method according to claim 1, feature exist
In: text data is converted to in step 1 using TF-IDF method or word frequency method the feature vector under theorem in Euclid space;By mentioning
SIFT feature value is taken to carry out characterization to image data.
3. a kind of efficient drosophila neural network Hash Search WMSN data method according to claim 1, feature exist
In: step 2 specifically comprises the following steps:
Step 2.1, matrix is quickly converted by FJLT and each data object in data set is projected into new metric space,
Step 2.2, neuron winner-take-all strategy pair is enlivened in conjunction in drosophila neural network local sensitivity Hash FLSH algorithm
The new feature of each data object is accepted or rejected respectively after projection, using the feature retained as the index of the data object.
4. a kind of efficient drosophila neural network Hash Search WMSN data method according to claim 1, feature exist
In: the calculation formula for being quickly converted matrix projection using FJLT of step 2 is as follows:
Y=P × H × D × X
Wherein, Y is data after projection, and X is the characteristic vector data in data set, and P matrix is one and is propped up with J-L theorem to be theoretical
The sparse matrix of support, H are a hadamard matrixs related with original dimension d, and D is a diagonal matrix, on the diagonal line of D
Value between -1 and 1, and -1 and 1 respectively have 1/2 probability.
5. a kind of efficient drosophila neural network Hash Search WMSN data method according to claim 4, feature exist
Meet that mean value is 0 and variance is q in: each element in P matrix-1Be just distributed very much, i.e., each element in P matrix meets
Pij-N(0,q-1) probability be q, do not meet P in matrixij-N(0,q-1) value of element of distribution is 0, the wherein calculating process of q
It is as follows:
Wherein, d is the original dimension of data set, and n is the size of data set, and ε is to protect away from performance parameter, and p is used normal form.
6. a kind of efficient drosophila neural network Hash Search WMSN data method according to claim 4, feature exist
In: each element H of H-matrixijValue can by following formula calculate go out:
Wherein, d is the original dimension of data set.
7. a kind of efficient drosophila neural network Hash Search WMSN data method according to claim 4, feature exist
In: the D matrix that dimension is d indicates are as follows:
Wherein, d is the original dimension of data set.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910039907.1A CN109739999A (en) | 2019-01-16 | 2019-01-16 | A kind of efficient drosophila neural network Hash Search WMSN data method |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910039907.1A CN109739999A (en) | 2019-01-16 | 2019-01-16 | A kind of efficient drosophila neural network Hash Search WMSN data method |
Publications (1)
Publication Number | Publication Date |
---|---|
CN109739999A true CN109739999A (en) | 2019-05-10 |
Family
ID=66365008
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910039907.1A Pending CN109739999A (en) | 2019-01-16 | 2019-01-16 | A kind of efficient drosophila neural network Hash Search WMSN data method |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109739999A (en) |
Cited By (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN113313212A (en) * | 2021-06-29 | 2021-08-27 | 福建师范大学 | Fruit fly olfactory neural network-based recessive anomaly detection method and device |
CN113364884A (en) * | 2021-06-28 | 2021-09-07 | 福建师范大学 | Industrial Internet of things recessive anomaly detection method based on local sensitive bloom filter |
CN113535717A (en) * | 2021-06-28 | 2021-10-22 | 福建师范大学 | Retrieval method and system based on Laplacian operator and LSH technology |
CN114003635A (en) * | 2021-09-27 | 2022-02-01 | 中国科学院自动化研究所 | Recommendation information acquisition method, device, equipment and product |
CN114003266A (en) * | 2021-10-14 | 2022-02-01 | 红石阳光(深圳)科技有限公司 | Method and device for generating multiple differential packets based on Android flash firmware |
CN114020839A (en) * | 2021-10-29 | 2022-02-08 | 哈尔滨工业大学 | Academic achievement publishing and right authentication system and method based on block chain |
CN115085941A (en) * | 2022-07-25 | 2022-09-20 | 深圳麦客存储科技有限公司 | Computer data information processing method based on block chain network |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102609441A (en) * | 2011-12-27 | 2012-07-25 | 中国科学院计算技术研究所 | Local-sensitive hash high-dimensional indexing method based on distribution entropy |
US20170161271A1 (en) * | 2015-12-04 | 2017-06-08 | Intel Corporation | Hybrid nearest neighbor search tree with hashing table |
-
2019
- 2019-01-16 CN CN201910039907.1A patent/CN109739999A/en active Pending
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102609441A (en) * | 2011-12-27 | 2012-07-25 | 中国科学院计算技术研究所 | Local-sensitive hash high-dimensional indexing method based on distribution entropy |
US20170161271A1 (en) * | 2015-12-04 | 2017-06-08 | Intel Corporation | Hybrid nearest neighbor search tree with hashing table |
Non-Patent Citations (2)
Title |
---|
NIR AILON等: "Approximate nearest neighbors and the fast Johnson-Lindenstrauss transform", 《PROCEEDINGS OF THE THIRTY-EIGHTH ANNUAL ACM SYMPOSIUM ON THEORY OF COMPUTING》 * |
SANJOY DASGUPTA等: "A neural algorithm for a fundamental computing problem", 《SCIENCE》 * |
Cited By (13)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN113364884A (en) * | 2021-06-28 | 2021-09-07 | 福建师范大学 | Industrial Internet of things recessive anomaly detection method based on local sensitive bloom filter |
CN113535717A (en) * | 2021-06-28 | 2021-10-22 | 福建师范大学 | Retrieval method and system based on Laplacian operator and LSH technology |
CN113535717B (en) * | 2021-06-28 | 2023-07-18 | 福建师范大学 | Retrieval method and system based on Laplacian operator and LSH technology |
CN113364884B (en) * | 2021-06-28 | 2023-06-30 | 福建师范大学 | Industrial Internet of things recessive anomaly detection method based on local sensitive bloom filter |
CN113313212B (en) * | 2021-06-29 | 2023-06-06 | 福建师范大学 | Implicit abnormality detection method and device based on drosophila olfactory neural network |
CN113313212A (en) * | 2021-06-29 | 2021-08-27 | 福建师范大学 | Fruit fly olfactory neural network-based recessive anomaly detection method and device |
CN114003635A (en) * | 2021-09-27 | 2022-02-01 | 中国科学院自动化研究所 | Recommendation information acquisition method, device, equipment and product |
CN114003266B (en) * | 2021-10-14 | 2022-05-27 | 红石阳光(深圳)科技有限公司 | Method and device for generating multiple differential packets based on Android flash firmware |
CN114003266A (en) * | 2021-10-14 | 2022-02-01 | 红石阳光(深圳)科技有限公司 | Method and device for generating multiple differential packets based on Android flash firmware |
CN114020839B (en) * | 2021-10-29 | 2022-07-22 | 哈尔滨工业大学 | Academic achievement publishing and right authentication system and method based on block chain |
CN114020839A (en) * | 2021-10-29 | 2022-02-08 | 哈尔滨工业大学 | Academic achievement publishing and right authentication system and method based on block chain |
CN115085941A (en) * | 2022-07-25 | 2022-09-20 | 深圳麦客存储科技有限公司 | Computer data information processing method based on block chain network |
CN115085941B (en) * | 2022-07-25 | 2022-11-15 | 深圳麦客存储科技有限公司 | Computer data information processing method based on block chain network |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109739999A (en) | A kind of efficient drosophila neural network Hash Search WMSN data method | |
CN111198959B (en) | Two-stage image retrieval method based on convolutional neural network | |
Gong et al. | Iterative quantization: A procrustean approach to learning binary codes for large-scale image retrieval | |
CN103116762B (en) | A kind of image classification method based on self-modulation dictionary learning | |
Raginsky et al. | Locality-sensitive binary codes from shift-invariant kernels | |
Yin et al. | Incomplete multi-view clustering via subspace learning | |
Cevikalp et al. | Semi-supervised dimensionality reduction using pairwise equivalence constraints | |
Qin et al. | Kernel neural gas algorithms with application to cluster analysis | |
Sun et al. | Feature fusion using locally linear embedding for classification | |
Savelonas et al. | Fisher encoding of differential fast point feature histograms for partial 3D object retrieval | |
Deng et al. | Adaptive multi-bit quantization for hashing | |
Cao et al. | Local information-based fast approximate spectral clustering | |
Tan et al. | L1-norm latent SVM for compact features in object detection | |
Zhang et al. | Learning all-in collaborative multiview binary representation for clustering | |
Mishra et al. | Kohonen self organizing map with modified k-means clustering for high dimensional data set | |
Shang et al. | Object recognition using rotation invariant local binary pattern of significant bit planes | |
Hare et al. | Efficient clustering and quantisation of SIFT features: exploiting characteristics of the SIFT descriptor and interest region detectors under image inversion | |
Acosta-Mendoza et al. | Improving graph-based image classification by using emerging patterns as attributes | |
Remil et al. | Data‐Driven Sparse Priors of 3D Shapes | |
Shao et al. | FJLT-FLSH: More efficient fly locality-sensitive hashing algorithm via FJLT for WMSN IoT search | |
Chen et al. | Research on optimized R-tree high-dimensional indexing method based on video features | |
Liang et al. | Facial feature extraction method based on shallow and deep fusion CNN | |
Asif et al. | Composite description based on color vector quantization and visual primary features for CBIR tasks | |
Liu et al. | Perceptual uniform descriptor and Ranking on manifold: A bridge between image representation and ranking for image retrieval | |
Tan et al. | “Feature fusion multi-view hashing based on random kernel canonical correlation analysis |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190510 |
|
RJ01 | Rejection of invention patent application after publication |