CN106649715A - Cross-media retrieval method based on local sensitive hash algorithm and neural network - Google Patents
Cross-media retrieval method based on local sensitive hash algorithm and neural network Download PDFInfo
- Publication number
- CN106649715A CN106649715A CN201611190238.0A CN201611190238A CN106649715A CN 106649715 A CN106649715 A CN 106649715A CN 201611190238 A CN201611190238 A CN 201611190238A CN 106649715 A CN106649715 A CN 106649715A
- Authority
- CN
- China
- Prior art keywords
- hash
- function
- text
- local sensitivity
- neutral net
- 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.)
- Granted
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
- G06F16/43—Querying
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Health & Medical Sciences (AREA)
- Artificial Intelligence (AREA)
- Biophysics (AREA)
- Evolutionary Computation (AREA)
- Biomedical Technology (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Computational Linguistics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Multimedia (AREA)
- Databases & Information Systems (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The invention discloses a cross-media retrieval method based on local sensitive hash algorithm and neural network, and relates to the technical field of cross-media retrieval. The method comprises two stages of local sensitive hash and hash function learning, wherein in the stage of local sensitive hash, the image data is mapped to hash buckets in m hash tables G = [g1, g2,...,gm] (which is an element of a set R<k*m>) by the local sensitive hash algorithm, wherein G is the set of m hash tables, gj is the jth hash table, and k is the length of the hash code corresponding to the hash bucket; and in the stage of hash function learning, the text data is respectively mapped to hash functions Ht = (Ht (1), Ht (2), ..., Ht (m), Ht (j)) in corresponding hash buckets in m hash tables by the neural network algorithm learning, wherein Ht (j), (1<=j<=m) represents the learned hash function Ht corresponding to the jth hash table. After getting the functions of these two phases, all the images and the documents are further coded and indexed for more accurate retrieval.
Description
Technical field
The present invention relates to cross-media retrieval technical field, refer in particular to a kind of based on local sensitivity hash algorithm and neutral net
Cross-media retrieval method.
Background technology
In across the media big data epoch, do not bring in the magnanimity multi-modal information for producing all the time huge across media
Search Requirement, such as searches for image or video with text, and vice versa.For example, an entry on wikipedia is generally comprised
Text is described and example image, and the retrieval of these information needs to build cross media indexing and learning method.With traditional single matchmaker
Health check-up rope is compared, and the key problem of cross-media retrieval is between the identical or related semantic object for how excavating different media representations
Association.
At present worldwide, numerous solutions are proposed for the key problem of the cross-media retrieval.It is existing
Cross-media retrieval method be broadly divided into two classes, a class is based on the method for theme:Document [1] is by theme proportion grading to not
It is modeled with the correlation between the data of mode;Document [2] by CORR-LDA excavate between image and text marking
The relation of theme level;Document [3] is combined Markov random fields with tradition LDA methods, it is proposed that examined with brief word
The built-up pattern (MDRF) of the oriented and undirected probability graph of rope image;Document [4] proposes one kind to using multiple medium types
Micro-blog information come carry out obtain social event visualization summarize multimedia social event autoabstract framework.It is another kind of to be
Method based on subspace:The core of the method for this class is to seek to make the maximized subspace of different modalities data dependence
[5].Sharma et al. proposes a kind of general multi-modal feature extraction framework technology, referred to as the multi-view analysis GMA of broad sense
[6].The viewpoint of semanteme is introduced in the T-V CCA models that document [7] is proposed, to improve subspace in different classes of multi-modal number
According to classification accuracy.Document [8] proposes a kind of Bi-CMSRM methods, constructs from the angle of optimization bi-directional list sequencing problem
Suitable for the computation model of cross-media retrieval.
[1]Blei D M,Ng A Y,Jordan M I.Latent dirichlet allocation[J].the
Journal of machine Learning research,2003,3:993-1022.
[2]Blei D M,Jordan M I.Modeling annotated data[C]//Proceedings of the
26th annual international ACM SIGIR conference on Research and developme nt
in
information retrieval.ACM,2003:127-134.
[3]Jia Y,Salzmann M,Darrell T.Learning cross modality similarity for
multinomial data[C]//Computer Vision(ICCV),2011IEEE International Conference
on.
IEEE,2011:2407-2414.
[4]Bian J,Yang Y,Zhang H,et al.Multimedia Summarization for Social
Events in Microblog Stream[J].IEEE Transactions on Multimedia,2015,17(2):216-
228.
[5]Hardoon D R,Szedmak S,ShaweTaylor J.Canonical correlation
analysis:An overview with application to learning methods[J].Neural
computation,2004,16(12):2639-2664.
[6]Abhishek Sharma,Abhishek Kumar,H Daume,and
DavidWJacobs.2012.Generalized multi-view analysis:A discriminative latent
space.In IEEE Conference on Computer Vision and Pattern Recognition.2160–
2167.
[7]Yunchao Gong,Qifa Ke,Michael Isard,and Svetlana Lazebnik.2013.A
Multi-View Embedding Space for Modeling Internet Images,Tags,and Their
Semantics.International Journal of Computer Vision(2013),1–24.
Wu F,Lu X,Zhang Z,et al.Cross-media semantic representation via bi-
directional learning to rank[C]//Proceedings of the 21st ACM international
conference on Multimedia.ACM,2013:877-886.
There is same technological deficiency in existing cross-media retrieval method, i.e., only only considered cross-media retrieval method
Itself and have ignored the optimization processing feasible to some of document sets, due to exist in document sets in a large number with inquire about incoherent text
Shelves, therefore document sets are pre-processed before accurately being inquired about, relevant documentation proportion is to carrying in raising document sets
It is significant for high recall precision.
The content of the invention
For the technical problem existing for existing cross-media retrieval method, present invention proposition is a kind of can to improve retrieval
The cross-media retrieval method based on local sensitivity hash algorithm and neutral net of accuracy.
The present invention concrete technical scheme be:
A kind of cross-media retrieval method based on local sensitivity hash algorithm and neutral net, the cross-media retrieval method
Comprise the following steps:
1) FCMR (Fast Cross-Media Retrieval, FCMR) model, the training of the FCMR models are set up
Journey includes local sensitivity Hash stage and hash function study stage;
2) hash function arrived using local sensitivity hash function and neural network learning is mapped in all texts and image
Set up to Hamming space and index;
3) cross-media retrieval inquiry is carried out, including text query and image querying.
As the preferred technical solution of the present invention, step 1 of the present invention) in, the local sensitivity Hash stage includes
Hash bucket is mapped the image data to using local sensitivity hash algorithm, is specifically included image by local sensitivity hash algorithm
Data are mapped to m Hash table G=[g1,g2,...,gm]∈Rk×mHash bucket in, wherein G is the set of m Hash table, gj
J-th Hash table is represented, k is the length of Hash bucket correspondence Hash codes.
As the preferred technical solution of the present invention, step 1 of the present invention) in, the hash function study stage includes
Text data is mapped to using neural network algorithm study for the hash function Ht of Hash bucket, is specifically included and is calculated by neutral net
Text data is respectively mapped to calligraphy learning the hash function Ht=(Ht in m Hash table in its corresponding Hash bucket(1),Ht(2),...,Ht(m)), Ht(j), (1≤j≤m) represents the hash function corresponding to j-th Hash table for learning.
As the preferred technical solution of the present invention, step 3 of the present invention) in,
The text query is a given query text, by hash function Ht(j)The query text is mapped to into m
In Hash bucket in Hash table, then the image file of these Hash bucket memory storages just constitutes the arest neighbors of the query text, will
The image pattern fallen in identical Hash bucket with query text as candidate result collection, and then in the arest neighbors of the query text
In the range of accurately retrieved, calculate the distance between image for concentrating with candidate result of query text and simultaneously accurately examined
Rope ranking;
Described image inquiry is mapped to the query image by local sensitivity hash function to give a query image
In Hash bucket in m Hash table, then the text of these Hash bucket memory storages just constitutes the arest neighbors of the query image,
And then carry out precise search in the arest neighbors scope of the query image.
Used as the preferred technical solution of the present invention, local sensitivity hash function of the present invention is defined as follows:
Wherein, hyperplane vectorMeet multi GaussianN (0,1) distributions;
Define a series of hash function h1,h2,...,hnK function component function g (x) therein is randomly selected, if choosing
It is h1To hk, then g (x)=(h1(x),h2(x),...,hk(x)), choose m g (x) function:g1(x),g2(x),...,gm(x),
Then each g (x) function pair answers a Hash table;By m g (x) function by each image pattern p in image spaceiPoint
In not being mapped to m Hash table, so each image pattern piWill occur in certain Hash bucket of m Hash table;So pi
Corresponding Hash bucket can be expressed as in j-th Hash table:
gj(pi)=<h1(pi),h2(pi)...,hk(pi)>, (0 < j≤m, 0 < i≤n) (2)
As the preferred technical solution of the present invention, m neutral net NN arrived used in FCMR models of the present invention(j),(j∈
1,2 ..., m) there is identical structure;Each neutral net NN(j)There are L layers, wherein input layer has dt neuron to correspond to
The dimension of text feature, output layer has k position of the k neuron corresponding to Hash codes, the residue in addition to input layer and output layer
L-2 layers be used for learn hash function;By each ti∈ T are used as NN(j)Input, neutral net each layer can be obtained
OutputL+1 layers withFor input, output
WhereinWithThe respectively feature representation of l layers and l+1 layers;W(l+1)It is transition matrix;f(l+1)It is activation primitive;
The hash function Ht that neural network learning is arrived(j)With tiTo be input into and export length be k Hash codes:
Wherein,It is a k dimension real-valued vectors, will using sign functionIt is converted into Hash codes;
For training sampleHt(j)(ti) withShould be identical, that is,
WithIt is as equal as possible.
Defining loss function based on minimum variance is:
Wherein,It is not put in marks Function Neural Network to tiPredicted value,Represent piCorresponding to jth (0 < j≤
M) Hash codes of the Hash bucket in individual Hash table;
Obtain training the training sample needed for neutral net from the local sensitivity Hash stage(i∈1,2,...,
Nt, j ∈ 1,2 ..., m), by training neutral net NN(j)Its study can be made to arrive tiIt is mapped toHash function.
Used as the preferred technical solution of the present invention, the training of neutral net of the present invention is divided into pre-training and parameter adjustment, tool
Body includes:
(1) stack self-encoding encoder (Stacked AutoEncoder, SAE) is applied to FCMR models sequentially to train
Neutral net NN(j)In each layer with initialization network parameter;
(2) based on the loss function formula (5), neutral net is trained to adjust network parameter by BP algorithm;
(3) variance and SSE based on all samples of text is devised shown in overall loss function such as formula (6):
Compared with prior art, the invention has the beneficial effects as follows:
The present invention is based on local sensitivity hash algorithm and neutral net, by eliminating document content unrelated with inquiry in a large number
And the arest neighbors of a group polling is obtained, finally more efficiently carry out retrieval tasks in the range of the arest neighbors of inquiry document.
Description of the drawings
Fig. 1 is the FCMR block schematic illustrations of the present invention.
Fig. 2 is the FCMR retrieval schematic diagrames of the present invention.
Specific embodiment
The present invention is elaborated in conjunction with Figure of description.
A kind of cross-media retrieval based on local sensitivity hash algorithm and neutral net that the specific embodiment of the invention is provided
Method (Fast Cross-Media Retrieval, FCMR), the cross-media retrieval method mainly comprises the steps:
1) FCMR (Fast Cross-Media Retrieval, FCMR) model, the training of the FCMR models are set up
Journey includes local sensitivity Hash stage and hash function study stage;
2) hash function arrived using local sensitivity hash function and neural network learning is mapped in all texts and image
Set up to Hamming space and index;
3) cross-media retrieval inquiry is carried out, including text query and image querying.
Wherein, in order that symbol and algorithm statement are more succinct, describe by taking two mode of text and image as an example carry below
The FCMR models for going out, model can easily expand to other mode, and the FCMR models include local sensitivity Hash and Kazakhstan
Two stages of uncommon function learning.
In the local sensitivity Hash stage, Hash bucket is mapped the image data to using local sensitivity hash algorithm, concrete bag
Include and m Hash table G=[g is mapped the image data to by local sensitivity hash algorithm1,g2,...,gm]∈Rk×mHash bucket
Interior, wherein R represents real number field, and G is the set of m Hash table, gjJ-th Hash table is represented, k is Hash bucket correspondence Hash codes
Length;
Learn the stage in hash function, text data is mapped to using neural network algorithm study for the Hash letter of Hash bucket
Number Ht, specifically include by neural network algorithm study by text data be respectively mapped in m Hash table text data it is right
Hash function Ht=(Ht in the Hash bucket answered(1),Ht(2),...,Ht(m)), Ht(j), (1≤j≤m) represents the correspondence for learning
In the hash function of j-th Hash table.
The matrix description of text data is:T=[t1,t2,...,tnt]∈Rdt×nt, wherein T is that the matrix of text data is retouched
State.Accordingly, P=[p1,p2,...,pnp]∈Rdp×np, wherein P is the matrix description of view data.Wherein, tiWith piOne a pair
Should, the number of image text pair is n, i.e. nt=np=n, and nt and np is replaced with n in following content.
If obtaining m Hash table with local sensitivity hash algorithm, then need to design m god corresponding with Hash table
Jing networks are so that text data is mapped in m Hash table in the Hash bucket corresponding to these text datas.Based on neutral net
The local sensitivity hash function that the hash function for learning is used with the local sensitivity Hash stage, can set up to multi-modal data
Index, so as to carry out efficient cross-media retrieval task.
After index is set up, a query text is given, by hash function Ht(j)The query text is mapped to into m
In Hash bucket in Hash table, then the image file of these Hash bucket memory storages just constitutes the arest neighbors of the query text, enters
And accurately retrieved in the range of the arest neighbors of the query text;A query image is given, by local sensitivity Hash
In the Hash bucket that function is mapped to the query image in m Hash table, then the text of these Hash bucket memory storages is with regard to group
Into the arest neighbors of the query image, and then precise search is carried out in the arest neighbors scope of the query image.
The following detailed description of the local sensitivity hash algorithm in the specific embodiment of the invention, the local sensitivity hash algorithm
It is mainly used to solve the approximate KNN search problem at higher dimensional space midpoint, local sensitivity hash function is defined as follows:
Wherein, hyperplane vectorMeet multi GaussianN (0,1) distributions.
Define a series of hash function h1,h2,...,hnK function component function g (x) therein is randomly selected, if choosing
It is h1To hk, then g (x)=(h1(x),h2(x),...,hk(x)), choose m g (x) function:g1(x),g2(x),...,gm(x),
Then each g (x) function pair answers a Hash table.By m g (x) function by each image pattern p in image spaceiPoint
In not being mapped to m Hash table, so each image pattern piWill occur in certain Hash bucket of m Hash table.
So piCorresponding Hash bucket can be expressed as in j-th Hash table:
gj(pi)=<h1(pi),h2(pi)...,hk(pi)>, (0 < j≤m, 0 < i≤n) (2)
During inquiry, query text is given, using Ht(j)Functional query text is mapped, and will be fallen identical with query text
Hash bucket in image pattern as candidate result collection, calculate the distance between image that query text is concentrated with candidate result
And accurately retrieved ranking.
By local sensitivity hash algorithm, the sample p of image spacei, during (0 < i≤n) is mapped to m Hash table, and
Each pi, (0 < i≤n) all can and the sample similar to its together with occur in certain Hash bucket of m Hash table.So, it is each
Individual image pattern piAll establish contact with certain Hash bucket of jth (0 < j≤m) individual Hash table.It is simultaneously mentioned above,
Due to p in modeliAnd tiIt is the description of same semantic different modalities, image pattern is one-to-one with samples of text, therefore,
Each samples of text tiAlso contact is established with certain Hash bucket of jth (0 < j≤m) individual Hash table.So far, used
By samples of text t in training neural network learningiIt is mapped to jth (0 < j≤m) individual Hash table Chinese version sample tiCorrespondence Hash
The training sample of the function of bucket:(i ∈ 1,2 ..., n, j ∈ 1,2 ..., m), whereinRepresent piCorresponding to
The Hash codes of the Hash bucket in j (0 < j≤m) individual Hash table.
The following detailed description of the local sensitivity hash algorithm in the specific embodiment of the invention, as shown in figure 1, Fig. 1 gives
Hash function learns stage neural network structure, m neutral net NN arrived used in Fig. 1 models(j),(j∈1,2,...,m)
With identical structure;Each neutral net NN(j)There are L layers, wherein input layer has dt neuron corresponding to text feature
Dimension, output layer has k position of the k neuron corresponding to Hash codes, and remaining L-2 layers are used to learn hash function.By each
ti∈ T are used as NN(j)Input, the output of each layer of neutral net can be obtainedL+1 layers withFor
Input, output
WhereinWithThe respectively feature representation of l layers and l+1 layers;W(l+1)It is transition matrix;f(l+1)It is activation primitive.
The hash function Ht that neural network learning is arrived(j)With tiTo be input into and export length be k Hash codes:
Wherein,It is a k dimension real-valued vectors, will using sign functionIt is converted into Hash codes.
Due to sign function non-differentiability, it is difficult to optimize, therefore eliminated with the stage of neural network learning hash function
Sign function, and add again in test phase.
For training sampleHt(j)(ti) withShould be identical, that is,
WithIt is as equal as possible.
Defining loss function based on minimum variance is:
Wherein,It is not put in marks Function Neural Network to tiPredicted value.
Training sample according to needed for the local sensitivity Hash stage obtains training neutral net(i∈1,
2 ..., nt, j ∈ 1,2 ..., m), by training neutral net NN(j)Its study can be made to arrive tiIt is mapped toHash
Function.
The training of neutral net is divided into pre-training and parameter adjustment, and pre-training preferably initialization network parameter and can be prevented
Only network is absorbed in locally optimal solution, and the training of neutral net specifically includes following steps:
(1) stack self-encoding encoder (Stacked AutoEncoder, SAE) is applied to FCMR models sequentially to train
Neutral net NN(j)In each layer with initialization network parameter.
(2) based on loss function formula (5), network parameter is adjusted come training network by BP algorithm (back-propagation algorithm);
(3) variance and SSE based on all samples of text is devised shown in overall loss function such as formula (6):
In order that neutral net NN(j)The function Ht for learning(j)Well samples of text data can be mapped to into j
In Hash table in its corresponding Hash bucket, the embodiment of the present invention trains neutral net NN using traditional back-propagation algorithm(j), final hash function Ht is obtained eventually through formula (4) in test phase(j)。
Wherein, the algorithmic procedure of the FCMR of the present embodiment is specific as follows:
When Fig. 2 shows only one of which Hash table, FCMR enters the schematic diagram of line retrieval, and multiple Hash tables need to only use all god
Jing e-learnings to hash function map the text to Hamming space.
Obviously, the above embodiment of the present invention is only intended to clearly illustrate example of the present invention, and is not right
The restriction of embodiments of the present invention.For those of ordinary skill in the field, may be used also on the basis of the above description
To make other changes in different forms.There is no need to be exhaustive to all of embodiment.It is all this
Any modification, equivalent and improvement made within the spirit and principle of invention etc., should be included in the claims in the present invention
Protection domain within.
Claims (7)
1. a kind of cross-media retrieval method based on local sensitivity hash algorithm and neutral net, it is characterised in that described across matchmaker
Body search method is comprised the following steps:
1) FCMR (Fast Cross-Media Retrieval, FCMR) model, the training process bag of the FCMR models are set up
Include local sensitivity Hash stage and hash function study stage;
2) hash function for being arrived using local sensitivity hash function and neural network learning is by all text datas and view data
It is mapped to Hamming space and sets up index;
3) cross-media retrieval inquiry is carried out, including text query and image querying.
2. a kind of cross-media retrieval method based on local sensitivity hash algorithm and neutral net according to claim 1,
Characterized in that, the step 1) in, the local sensitivity Hash stage includes adopting local sensitivity hash algorithm by picture number
According to Hash bucket is mapped to, specifically includes and m Hash table G=[g is mapped the image data to by local sensitivity hash algorithm1,
g2,...,gm]∈Rk×mHash bucket in, wherein R represents real number field, and G is the set of m Hash table, gjRepresent j-th Hash
Table, k is the length of Hash bucket correspondence Hash codes.
3. a kind of cross-media retrieval method based on local sensitivity hash algorithm and neutral net according to claim 2,
Characterized in that, the step 1) in, the hash function study stage includes learning textual data using neural network algorithm
According to the hash function Ht for being mapped to Hash bucket, specifically include and text data is respectively mapped to by m by neural network algorithm study
Hash function Ht=(Ht in individual Hash table in its corresponding Hash bucket(1),Ht(2),...,Ht(m)), Ht(j), (1≤j≤m) table
The hash function corresponding to j-th Hash table that dendrography is practised.
4. a kind of cross-media retrieval method based on local sensitivity hash algorithm and neutral net according to claim 3,
Characterized in that, the step 3) in,
The text query is a given query text, by hash function Ht(j)The query text is mapped to into m Hash
In Hash bucket in table, then the image file of these Hash bucket memory storages just constitutes the arest neighbors of the query text, will with look into
The image pattern that falls in identical Hash bucket of text is ask as candidate result collection, and then in the arest neighbors scope of the query text
Inside accurately retrieved, calculate the distance between image that query text is concentrated with candidate result and carry out accurate retrieval row
Name;
The query image is mapped to m by described image inquiry to give a query image by local sensitivity hash function
In Hash bucket in Hash table, then the text of these Hash bucket memory storages just constitutes the arest neighbors of the query image, enters
And carry out precise search in the arest neighbors scope of the query image.
5. a kind of cross-media retrieval method based on local sensitivity hash algorithm and neutral net according to claim 3,
Characterized in that, the local sensitivity hash function is defined as follows:
Wherein, hyperplane vectorMeet multi Gaussian N (0,1) distributions;
Define a series of hash function h1,h2,...,hn, k function component function g (x) therein is randomly selected, if that choosing is h1
To hk, then g (x)=(h1(x),h2(x),...,hk(x)), choose m g (x) function:g1(x),g2(x),...,gm(x), then often
Individual g (x) function pair answers a Hash table;By m g (x) function by each image pattern p in image spaceiReflect respectively
In being mapped to m Hash table, so each image pattern piWill occur in certain Hash bucket of m Hash table;So pi
Corresponding Hash bucket can be expressed as in j Hash table:
gj(pi)=<h1(pi),h2(pi)...,hk(pi)>, (0 < j≤m, 0 < i≤n) (2).
6. a kind of cross-media retrieval method based on local sensitivity hash algorithm and neutral net according to claim 5,
Characterized in that, m neutral net NN arrived used in FCMR models(j), (j ∈ 1,2 ..., m) there is identical structure;Often
One neutral net NN(j)There are L layers, wherein input layer has dimension of the dt neuron corresponding to text feature, and output layer has k
Corresponding to the k positions of Hash codes, remaining L-2 layers are used to learn hash function neuron in addition to input layer and output layer;Will be each
Individual ti∈ T are used as NN(j)Input, the output of each layer of neutral net can be obtainedL+1 layers withFor
Input, output
WhereinWithThe respectively feature representation of l layers and l+1 layers;W(l+1)It is transition matrix;f(l+1)
It is activation primitive;
The hash function Ht that neural network learning is arrived(j)With tiTo be input into and export length be k Hash codes:
Wherein,It is a k dimension real-valued vectors, will using sign functionIt is converted into Hash codes;
Defining loss function based on minimum variance is:
Wherein,It is not put in marks Function Neural Network to tiPredicted value,Represent piIt is individual corresponding to jth (0 < j≤m)
The Hash codes of the Hash bucket in Hash table;
Obtain training the training sample needed for neutral net from the local sensitivity Hash stage By training neutral net NN(j)Its study can be made to arrive tiIt is mapped toHash function.
7. a kind of cross-media retrieval method based on local sensitivity hash algorithm and neutral net according to claim 6,
Characterized in that, the training of neutral net is divided into pre-training and parameter adjustment, specifically include:
(1) stack self-encoding encoder (Stacked AutoEncoder, SAE) is applied to into the training NN that FCMR models carry out order(j)
In each layer with initialization network parameter;
(2) based on loss function formula (5), network parameter is adjusted come training network by BP algorithm;
(3) variance and SSE based on all samples of text is devised shown in overall loss function such as formula (6):
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201611190238.0A CN106649715B (en) | 2016-12-21 | 2016-12-21 | A kind of cross-media retrieval method based on local sensitivity hash algorithm and neural network |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201611190238.0A CN106649715B (en) | 2016-12-21 | 2016-12-21 | A kind of cross-media retrieval method based on local sensitivity hash algorithm and neural network |
Publications (2)
Publication Number | Publication Date |
---|---|
CN106649715A true CN106649715A (en) | 2017-05-10 |
CN106649715B CN106649715B (en) | 2019-08-09 |
Family
ID=58834417
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201611190238.0A Active CN106649715B (en) | 2016-12-21 | 2016-12-21 | A kind of cross-media retrieval method based on local sensitivity hash algorithm and neural network |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN106649715B (en) |
Cited By (22)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107256271A (en) * | 2017-06-27 | 2017-10-17 | 鲁东大学 | Cross-module state Hash search method based on mapping dictionary learning |
CN107273529A (en) * | 2017-06-28 | 2017-10-20 | 武汉图信科技有限公司 | Efficient level index construct and search method based on hash function |
CN107729557A (en) * | 2017-11-08 | 2018-02-23 | 北京大学 | A kind of classification of inventory information, search method and device |
CN107729290A (en) * | 2017-09-21 | 2018-02-23 | 北京大学深圳研究生院 | A kind of expression learning method of ultra-large figure using the optimization of local sensitivity Hash |
CN108229693A (en) * | 2018-02-08 | 2018-06-29 | 徐传运 | A kind of machine learning identification device and method based on comparison study |
CN108229588A (en) * | 2018-02-08 | 2018-06-29 | 重庆师范大学 | A kind of machine learning recognition methods based on deep learning |
CN108280207A (en) * | 2018-01-30 | 2018-07-13 | 深圳市茁壮网络股份有限公司 | A method of the perfect Hash of construction |
CN108319686A (en) * | 2018-02-01 | 2018-07-24 | 北京大学深圳研究生院 | Antagonism cross-media retrieval method based on limited text space |
CN108345943A (en) * | 2018-02-08 | 2018-07-31 | 重庆理工大学 | A kind of machine learning recognition methods based on embedded coding with comparison study |
CN108345942A (en) * | 2018-02-08 | 2018-07-31 | 重庆理工大学 | A kind of machine learning recognition methods based on embedded coding study |
CN108629049A (en) * | 2018-05-14 | 2018-10-09 | 芜湖岭上信息科技有限公司 | A kind of image real-time storage and lookup device and method based on hash algorithm |
CN109947936A (en) * | 2018-08-21 | 2019-06-28 | 北京大学 | A method of based on machine learning dynamic detection spam |
CN110083762A (en) * | 2019-03-15 | 2019-08-02 | 平安科技(深圳)有限公司 | Source of houses searching method, device, equipment and computer readable storage medium |
CN110674333A (en) * | 2019-08-02 | 2020-01-10 | 杭州电子科技大学 | Large-scale image high-speed retrieval method based on multi-view enhanced depth hashing |
CN110998607A (en) * | 2017-08-08 | 2020-04-10 | 三星电子株式会社 | System and method for neural networks |
CN112699676A (en) * | 2020-12-31 | 2021-04-23 | 中国农业银行股份有限公司 | Address similarity relation generation method and device |
CN112784838A (en) * | 2021-01-28 | 2021-05-11 | 佛山市南海区广工大数控装备协同创新研究院 | Hamming OCR recognition method based on locality sensitive hashing network |
CN113393107A (en) * | 2021-06-07 | 2021-09-14 | 东方电气集团科学技术研究院有限公司 | Incremental calculation method for state parameter reference value of power generation equipment |
CN113449849A (en) * | 2021-06-29 | 2021-09-28 | 桂林电子科技大学 | Learning type text hash method based on self-encoder |
CN113515450A (en) * | 2021-05-20 | 2021-10-19 | 广东工业大学 | Environment anomaly detection method and system |
CN114781642A (en) * | 2022-06-17 | 2022-07-22 | 之江实验室 | Cross-media corresponding knowledge generation method and device |
WO2023035362A1 (en) * | 2021-09-07 | 2023-03-16 | 上海观安信息技术股份有限公司 | Polluted sample data detecting method and apparatus for model training |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN104199827A (en) * | 2014-07-24 | 2014-12-10 | 北京大学 | Locality-sensitive-hashing-based high-dimensional indexing method for large-scale multimedia data |
CN104346440A (en) * | 2014-10-10 | 2015-02-11 | 浙江大学 | Neural-network-based cross-media Hash indexing method |
CN106202413A (en) * | 2016-07-11 | 2016-12-07 | 北京大学深圳研究生院 | A kind of cross-media retrieval method |
CN106227851A (en) * | 2016-07-29 | 2016-12-14 | 汤平 | Based on the image search method searched for by depth of seam division that degree of depth convolutional neural networks is end-to-end |
US9760588B2 (en) * | 2007-02-20 | 2017-09-12 | Invention Science Fund I, Llc | Cross-media storage coordination |
-
2016
- 2016-12-21 CN CN201611190238.0A patent/CN106649715B/en active Active
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US9760588B2 (en) * | 2007-02-20 | 2017-09-12 | Invention Science Fund I, Llc | Cross-media storage coordination |
CN104199827A (en) * | 2014-07-24 | 2014-12-10 | 北京大学 | Locality-sensitive-hashing-based high-dimensional indexing method for large-scale multimedia data |
CN104346440A (en) * | 2014-10-10 | 2015-02-11 | 浙江大学 | Neural-network-based cross-media Hash indexing method |
CN106202413A (en) * | 2016-07-11 | 2016-12-07 | 北京大学深圳研究生院 | A kind of cross-media retrieval method |
CN106227851A (en) * | 2016-07-29 | 2016-12-14 | 汤平 | Based on the image search method searched for by depth of seam division that degree of depth convolutional neural networks is end-to-end |
Cited By (36)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107256271A (en) * | 2017-06-27 | 2017-10-17 | 鲁东大学 | Cross-module state Hash search method based on mapping dictionary learning |
CN107256271B (en) * | 2017-06-27 | 2020-04-03 | 鲁东大学 | Cross-modal Hash retrieval method based on mapping dictionary learning |
CN107273529B (en) * | 2017-06-28 | 2020-02-07 | 武汉图信科技有限公司 | Efficient hierarchical index construction and retrieval method based on hash function |
CN107273529A (en) * | 2017-06-28 | 2017-10-20 | 武汉图信科技有限公司 | Efficient level index construct and search method based on hash function |
CN110998607B (en) * | 2017-08-08 | 2024-03-08 | 三星电子株式会社 | System and method for neural networks |
CN110998607A (en) * | 2017-08-08 | 2020-04-10 | 三星电子株式会社 | System and method for neural networks |
CN107729290A (en) * | 2017-09-21 | 2018-02-23 | 北京大学深圳研究生院 | A kind of expression learning method of ultra-large figure using the optimization of local sensitivity Hash |
CN107729557A (en) * | 2017-11-08 | 2018-02-23 | 北京大学 | A kind of classification of inventory information, search method and device |
CN108280207A (en) * | 2018-01-30 | 2018-07-13 | 深圳市茁壮网络股份有限公司 | A method of the perfect Hash of construction |
CN108319686B (en) * | 2018-02-01 | 2021-07-30 | 北京大学深圳研究生院 | Antagonism cross-media retrieval method based on limited text space |
CN108319686A (en) * | 2018-02-01 | 2018-07-24 | 北京大学深圳研究生院 | Antagonism cross-media retrieval method based on limited text space |
CN108345942B (en) * | 2018-02-08 | 2020-04-07 | 重庆理工大学 | Machine learning identification method based on embedded code learning |
CN108229588A (en) * | 2018-02-08 | 2018-06-29 | 重庆师范大学 | A kind of machine learning recognition methods based on deep learning |
CN108229693A (en) * | 2018-02-08 | 2018-06-29 | 徐传运 | A kind of machine learning identification device and method based on comparison study |
CN108345943A (en) * | 2018-02-08 | 2018-07-31 | 重庆理工大学 | A kind of machine learning recognition methods based on embedded coding with comparison study |
CN108229693B (en) * | 2018-02-08 | 2020-04-07 | 徐传运 | Machine learning identification device and method based on comparison learning |
CN108229588B (en) * | 2018-02-08 | 2020-04-07 | 重庆师范大学 | Machine learning identification method based on deep learning |
CN108345943B (en) * | 2018-02-08 | 2020-04-07 | 重庆理工大学 | Machine learning identification method based on embedded coding and contrast learning |
CN108345942A (en) * | 2018-02-08 | 2018-07-31 | 重庆理工大学 | A kind of machine learning recognition methods based on embedded coding study |
CN108629049A (en) * | 2018-05-14 | 2018-10-09 | 芜湖岭上信息科技有限公司 | A kind of image real-time storage and lookup device and method based on hash algorithm |
CN109947936A (en) * | 2018-08-21 | 2019-06-28 | 北京大学 | A method of based on machine learning dynamic detection spam |
CN109947936B (en) * | 2018-08-21 | 2021-03-02 | 北京大学 | Method for dynamically detecting junk mails based on machine learning |
CN110083762B (en) * | 2019-03-15 | 2023-01-24 | 平安科技(深圳)有限公司 | Room source searching method, device and equipment and computer readable storage medium |
CN110083762A (en) * | 2019-03-15 | 2019-08-02 | 平安科技(深圳)有限公司 | Source of houses searching method, device, equipment and computer readable storage medium |
CN110674333A (en) * | 2019-08-02 | 2020-01-10 | 杭州电子科技大学 | Large-scale image high-speed retrieval method based on multi-view enhanced depth hashing |
CN110674333B (en) * | 2019-08-02 | 2022-04-01 | 杭州电子科技大学 | Large-scale image high-speed retrieval method based on multi-view enhanced depth hashing |
CN112699676A (en) * | 2020-12-31 | 2021-04-23 | 中国农业银行股份有限公司 | Address similarity relation generation method and device |
CN112699676B (en) * | 2020-12-31 | 2024-04-12 | 中国农业银行股份有限公司 | Address similarity relation generation method and device |
CN112784838A (en) * | 2021-01-28 | 2021-05-11 | 佛山市南海区广工大数控装备协同创新研究院 | Hamming OCR recognition method based on locality sensitive hashing network |
CN113515450A (en) * | 2021-05-20 | 2021-10-19 | 广东工业大学 | Environment anomaly detection method and system |
CN113393107B (en) * | 2021-06-07 | 2022-08-12 | 东方电气集团科学技术研究院有限公司 | Incremental calculation method for state parameter reference value of power generation equipment |
CN113393107A (en) * | 2021-06-07 | 2021-09-14 | 东方电气集团科学技术研究院有限公司 | Incremental calculation method for state parameter reference value of power generation equipment |
CN113449849B (en) * | 2021-06-29 | 2022-05-27 | 桂林电子科技大学 | Learning type text hash method based on self-encoder |
CN113449849A (en) * | 2021-06-29 | 2021-09-28 | 桂林电子科技大学 | Learning type text hash method based on self-encoder |
WO2023035362A1 (en) * | 2021-09-07 | 2023-03-16 | 上海观安信息技术股份有限公司 | Polluted sample data detecting method and apparatus for model training |
CN114781642A (en) * | 2022-06-17 | 2022-07-22 | 之江实验室 | Cross-media corresponding knowledge generation method and device |
Also Published As
Publication number | Publication date |
---|---|
CN106649715B (en) | 2019-08-09 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN106649715A (en) | Cross-media retrieval method based on local sensitive hash algorithm and neural network | |
Kafle et al. | Dvqa: Understanding data visualizations via question answering | |
CN107832353B (en) | False information identification method for social media platform | |
CN109492099A (en) | It is a kind of based on field to the cross-domain texts sensibility classification method of anti-adaptive | |
CN106855853A (en) | Entity relation extraction system based on deep neural network | |
CN106446526A (en) | Electronic medical record entity relation extraction method and apparatus | |
CN105393265A (en) | Active featuring in computer-human interactive learning | |
CN108959522B (en) | Migration retrieval method based on semi-supervised countermeasure generation network | |
CN111680484B (en) | Answer model generation method and system for visual general knowledge reasoning question and answer | |
CN110928961B (en) | Multi-mode entity linking method, equipment and computer readable storage medium | |
Zhou et al. | Multi-label image classification via category prototype compositional learning | |
CN113761208A (en) | Scientific and technological innovation information classification method and storage device based on knowledge graph | |
Ou et al. | Semantic consistent adversarial cross-modal retrieval exploiting semantic similarity | |
Yusuf et al. | Evaluation of graph convolutional networks performance for visual question answering on reasoning datasets | |
CN116578738B (en) | Graph-text retrieval method and device based on graph attention and generating countermeasure network | |
CN111382243A (en) | Text category matching method, text category matching device and terminal | |
CN111737470A (en) | Text classification method | |
CN113158878B (en) | Heterogeneous migration fault diagnosis method, system and model based on subspace | |
PASBOLA | Text Classification Using Deep learning Methods | |
Nemade | Refinement of CNN based Multi-label image annotation | |
Jony et al. | Domain specific fine tuning of pre-trained language model in NLP | |
Kamal et al. | Geometry-Based Machining Feature Retrieval with Inductive Transfer Learning | |
Zhang et al. | Deep Normalization Cross-Modal Retrieval for Trajectory and Image Matching | |
Yu et al. | Chinese Electronic Medical Record Retrieval Method Using Fine-Tuned RoBERTa and Hybrid Features | |
Alsharif et al. | Remote Sensing Image Retrieval Using Multilingual Texts |
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 | ||
GR01 | Patent grant | ||
GR01 | Patent grant |