CN108536750A - Based on point to the characteristics of image binary-coding representation method of relational learning and reconstruct - Google Patents
Based on point to the characteristics of image binary-coding representation method of relational learning and reconstruct Download PDFInfo
- Publication number
- CN108536750A CN108536750A CN201810203371.8A CN201810203371A CN108536750A CN 108536750 A CN108536750 A CN 108536750A CN 201810203371 A CN201810203371 A CN 201810203371A CN 108536750 A CN108536750 A CN 108536750A
- Authority
- CN
- China
- Prior art keywords
- matrix
- rec
- dictionary
- indicate
- iteration
- 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
- 238000000034 method Methods 0.000 title claims abstract description 36
- 239000011159 matrix material Substances 0.000 claims abstract description 108
- 230000003595 spectral effect Effects 0.000 claims abstract description 5
- 230000006870 function Effects 0.000 claims description 17
- 238000010606 normalization Methods 0.000 claims description 9
- 238000010276 construction Methods 0.000 claims description 7
- 238000005457 optimization Methods 0.000 claims description 6
- 238000004364 calculation method Methods 0.000 claims description 4
- 230000003190 augmentative effect Effects 0.000 claims description 3
- 239000003102 growth factor Substances 0.000 claims description 3
- 238000005070 sampling Methods 0.000 claims description 2
- 230000011218 segmentation Effects 0.000 description 6
- 238000000354 decomposition reaction Methods 0.000 description 5
- 238000010801 machine learning Methods 0.000 description 5
- 230000008878 coupling Effects 0.000 description 3
- 238000010168 coupling process Methods 0.000 description 3
- 238000005859 coupling reaction Methods 0.000 description 3
- 238000005516 engineering process Methods 0.000 description 3
- 230000008569 process Effects 0.000 description 3
- 230000008901 benefit Effects 0.000 description 2
- 230000006872 improvement Effects 0.000 description 2
- 238000012935 Averaging Methods 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 238000005034 decoration Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000018109 developmental process Effects 0.000 description 1
- 235000013399 edible fruits Nutrition 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 239000000463 material Substances 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000001737 promoting effect Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/23—Clustering techniques
- G06F18/232—Non-hierarchical techniques
- G06F18/2323—Non-hierarchical techniques based on graph theory, e.g. minimum spanning trees [MST] or graph cuts
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/28—Determining representative reference patterns, e.g. by averaging or distorting; Generating dictionaries
Abstract
The invention discloses the characteristics of image binary-coding representation methods based on point to relational learning and reconstruct, including:Step 1, dictionary representation is converted data to, restricted problem is solved and obtains the coefficient matrix in representation;Step 2, the weight matrix that expression figure is constructed to the coefficient matrix obtained in step 1, is divided into k group by spectral clustering by dictionary item on this map, to realize that dictionary is divided;Step 3, new samples are given, its reconstructed residual in all dictionary item groups is calculated, choose the minimum corresponding optimal dictionary item group of reconstructed residual to carry out linear expression, to complete a little study to relationship;Step 4, optimal models are solved, study holding point realizes reconstruct of the point to relationship to the optimal binary-coding of relationship.
Description
Technical field
The invention belongs to characteristics of image coding fields, more particularly to based on point to the characteristics of image two of relational learning and reconstruct
It is worth coded representation method.
Background technology
Current era also increases in high speed with Internet information age high speed development, image data total amount.It is examined in image
In the application of rope, user gives a query image, needs to retrieve image similar with its from large scale database, and
It is returned the result according to similarity ranking.For this application scenarios, a kind of most basic way is:First to query image and
Database images extract feature respectively.Then, according to certain metric form (such as Euclidean distance) calculating query image and often
The distance between a database images.Finally, database images are ranked up according to apart from size, return to forward data
Library image is as retrieval result.The method in practical applications, needs to choose the spy having to picture material compared with high rule complexity
Sign, and this category feature is often higher-dimension.But it can propose prodigious demand, Er Qiegao to database purchase using high dimensional feature
The Real-valued feature of dimension is not high apart from computational valid time rate, and with the increase of database magnitude, common will apart from calculation
As performance bottleneck.Therefore the characteristics of image based on binary-coding can well solve above-mentioned two problems, improve extensive
The efficiency of image retrieval.
In practical applications, for data often without label, unsupervised binary-coding character representation and learning algorithm are logical
Often it is suitable for not having the case where semantic label.And in the case that unsupervised, with traditional characteristic similarity based on Euclidean distance
Representation method is again too simple, cannot obtain satisfactory effect.When data scale and data complexity increase, how
Learn compact and indicate the strong binary-coding feature of ability, is an important problem in being retrieved for large-scale data.
Invention content
Goal of the invention:The present invention is directed in existing unsupervised algorithm coding learning process, hash function and optimization aim
The tightly coupled problem of loss function proposes a kind of being compiled to the characteristics of image two-value of relational learning and reconstruct based on point for loose coupling
Code representation method effectively solves under the data based on Hash binary-coding to improve the performance and accuracy rate of image retrieval,
The quick and precisely search problem of image.
The characteristics of image binary-coding representation method based on point to relational learning and reconstruct built through the invention, it is intended to
Using the means of machine learning and machine vision, the high, recall precision by the characteristic dimension occurred in traditional image retrieval technologies
The problems such as low, and it is tight for hash function in existing unsupervised algorithm coding learning process and the loss function of optimization aim
The problem of coupling, by reaching promotion figure based on the characteristics of image binary-coding representation method put to relational learning and reconstruct
As the purpose of retrieval performance and accuracy rate.
Technical solution:The invention discloses the characteristics of image binary-coding expression sides based on point to relational learning and reconstruct
Method, it is proposed that a kind of automation solutions effectively carrying out binary-coding to image in the case where image lacks and marks,
Specifically include following steps:
Step 1, coefficient matrix is solved:Data are expressed as to the linear combination of dictionary and coefficient matrix, by solving to being
The restricted model of matrix number is to acquire coefficient matrix;
Step 2, dictionary is divided:The weight matrix for indicating the coefficient matrix construction obtained in step 1 a figure, at this
Dictionary item is divided by k by spectral clustering on a figure2A group;
Step 3, character representation:When given new samples, its reconstructed residual in all dictionary item groups is calculated, then
The minimum corresponding optimal dictionary item group of reconstructed residual is therefrom chosen to carry out linear expression, completes study of the point to relationship;
Step 4, point reconstructs relationship:Optimal models are solved, to which the optimal two-value of relationship is compiled in study to holding point
Code realizes reconstruct of the point to relationship.
Step 1 includes:
Step 1-1 gives a matrix D=[x for including n image data1,x2,...,xn]∈Rd×n, wherein D's is every
One row indicate an image data, xnIndicate n-th image data, the dimension of each image data is d, and matrix D can be from table
It is shown as D=DC, the value of each element of matrix belongs to set of real numbers R, wherein C ∈ Rn×nIt is coefficient matrix, each row indicate every
A initial data corresponding reconstruction coefficients on dictionary, the dictionary, that is, data matrix D itself can to coefficient matrix application
It describes the low-rank constraint of the global structure of initial data and the sparsity of the local structure of initial data can be described about
Beam obtains being grouped dictionary such as drag:
Wherein | | | |*With | | | |1Representing matrix nuclear norm and l respectively1Norm, they are respectively by as to coefficient square
Battle array carries out the approximation of low-rank constraint and the approximation of sparsity constraints, and s.t. indicates that constraint, λ are one for balancing sparsity
The balance parameters of importance between low-rank;
Model in step 1-2, solution procedure 1-1 introduces an auxiliary variable J and obtains such as drag:
Augmented Lagrangian Functions L (C, J, the Y of the model1,Y2, μ) be:
Wherein, Y1,Y2, μ represents the introduced auxiliary variable of solving model, by each variable, that is, C, J, Y1,Y2,μ
Iteration update, which is solved, iteration update each variable rule it is as follows:
Jk+1=max (| Ck+1+Y2,k/μk|-λ/μk,0)
Y1,k+1=Y1,k+μk(D-DCk+1)
Y2,k+1=Y2,k+μk(Ck+1-Jk+1)
μk+1=min (μmax,ρμk)
Wherein, Ck+1Indicate the updated value of+1 iteration of variable C kth, JkIndicate that variable J kth time iteration is updated
Value, μkIndicate the updated value of variable μ kth time iteration, Y1,kIndicate variable Y1The updated value of kth time iteration, Y2,kIt indicates to become
Measure Y2The updated value of kth time iteration, μmaxIndicate that the maximum value of variable μ settings, ρ indicate that the growth factor of variable μ, Θ indicate
Singular value shrinkage operation accord with andThe coefficient matrix C learnt is acquired by iteration.
Step 2 includes:
The coefficient matrix C obtained in step 1 is indicated the weight matrix W of a figure, calculated as follows by step 2-1:
W=(C+CT)/2;
Step 2-2, structure n × n is to angle matrix Dw, DwIt is defined as follows:
Wherein diIt is defined as:di=∑ wij,
wijIt refer to the value of the number of the i-th row jth row of matrix W;
Step 2-3 defines Laplacian Matrix L=Dw- W, and according to L, DwCalculate Dw -1/2LDw -1/2;
Step 2-4 calculates Dw -1/2LDw -1/2Minimum k1The corresponding feature vector f of a characteristic value institute, by feature to
(citation is for amount f standardization:《Linear algebra (sixth version)》Higher Education Publishing House of department of mathematics of Tongji University), final group
At n × k1The eigenmatrix F of dimension;
Step 2-5, to every a line in eigenmatrix F as a k1The sample of dimension, total n sample, uses clustering method
Clustered that (citation is:《Machine learning》Publishing house of Zhou Zhihua Tsinghua University), cluster dimensionality k2, belonging to the i-th row
Class is exactly original xiClass belonging to (i.e. i-th of data of the matrix D in step 1-1) finally obtains cluster and divides c, thus by word
Allusion quotation item is divided into k2A group.
Step 3 includes:
Step 3-1, when a given new samples x ' calculates the reconstruction coefficients z of x ' for entire dictionary Di:
zi=(DTD+αI)-1DTX ',
Wherein, α is balance parameters, and I is unit matrix;
Step 3-2 is calculated for each dictionary item group respectively for the normalization residual error of x ':It is calculated by following formula
Normalization residual error r of k-th of dictionary item group for x 'k(x′):
Wherein, φk_d(zi) it is ziA part of coefficient, corresponding dictionary item is kth _ d dictionary group, DkIndicate kth
The dictionary that the data of a dictionary item group are formed;
The reconstruction coefficients z of x ' has been calculated in step 3-3iAfter normalization residual error in all dictionary item groups, selects and calculate
The dictionary item group of the reconstructed residual minimum arrived carries out linear expression as optimal dictionary item group, to optimal dictionary item group, by dilute
Thin coding is reconstructed expression to x ' and (is realized by LAE (Local Anchor Embedding) algorithms, algorithm is with reference to text
It offers:《Large Graph Construction for Scalable SemiSupervised Learning》Wei Liu,
Junfeng He, Shih-Fu Chang), to complete a little study to relationship.
Step 4 includes:
Such as drag is solved, to learn optimal binary-coding to which holding point is to relationship:
s.t. WrecWrec T=I
Wherein WrecIt is linear projection matrix, ZrecBe data matrix (i.e. step 3-3 obtain by sparse coding logarithm
According to the reconstruct data being reconstructed after indicating), B ∈ { -1,1 }c×nIt is a two values matrix, each column representing matrix ZrecMiddle data
Corresponding binary-coding, μrecIt is an offset parameter, s is a zooming parameter, | | | |FThe F norms of representing matrix, ⊙ tables
Show that the multiplication of corresponding element, model are convex, each unknown parameter μ is updated by iterationrec、B、Wrec, s optimize, directly
It is restrained to target function value.
It is described that each unknown parameter is updated to optimize by iteration, until target function value convergence, specifically include:
Step 4-1, random initializtion B and Wrec:
WrIt is to initialize WrecAnd from the matrix of standardized normal distribution stochastical sampling, Wr=U ∑s VTIt indicates to WrMatrix
Carry out singular value decomposition, it is assumed that WrIt is m × n rank matrix, then U is m × m rank matrix after decomposing, and Σ is a m × n
Rank diagonal matrix, and V* is n × n rank matrix.Bibliography is:《Linear algebra (sixth version)》Department of mathematics of Tongji University
Higher Education Publishing House;C refers to the length of binary-coding, i.e. how many position, such as 16,32;
Initialize μrec=0, s=1;
Step 4-2 starts new round iteration, and iterations add one, if iterations are less than or equal to T times, carries out step
4-3, it is no to then follow the steps 4-4;
Step 4-3 updates B:
It enables
It calculates
Update Wrec:
It enables
It calculates
SVD () herein refers to carrying out singular value decomposition, singular value decomposition (Singular to the object in bracket
Value Decomposition) it is a kind of important matrix decomposition in linear algebra, belong to the prior art, citation is:
《U.S. of mathematics》The People's Telecon Publishing Houses Wu Jun;
Sign () in formula refers to carrying out sign function operations to the object in bracket.The meaning of this function of sign
It is:Work as x>0, sign (x)=1;Work as x=0, sign (x)=0;Work as x<0, sign (x)=- 1;
Update μrec:
μrec=column_mean (Zrec-s⊙Wrec TB),
Update s:
It enables
It calculates
Column_mean () in formula refer to averaging to matrix in bracket along the direction of row obtain it is new
Matrix, trace () are the marks for calculating matrix in bracket.In linear algebra, the leading diagonal of a n × n matrix A is (from a left side
The diagonal line of top to lower right) on the summation of each element be referred to as the mark (or mark number) of matrix A, be generally denoted as tr (A).
Bibliography is:《Linear algebra (sixth version)》Higher Education Publishing House of department of mathematics of Tongji University;
Terminate epicycle iteration, return to step 4-2;
Step 4-4, by iterative calculation, the optimal binary-coding matrix B finally learnt completes weight of the point to relationship
Structure.
The present invention is directed to the problems such as characteristic dimension occurred in traditional image retrieval technologies is high, recall precision is low, base
In the model of machine learning and machine vision, the image data of not semantic label is converted to the binary-coding feature of low-dimensional,
So as to improve retrieval rate in the applications such as the retrieval of image, amount of storage and scramble time are reduced.This method includes mainly
Coefficient matrix, dictionary segmentation, character representation, point are solved to relationship four steps of reconstruct.It is by data to solve coefficient matrix step
Dictionary representation is converted to, restricted problem is solved and obtains the coefficient matrix in representation;Dictionary segmentation step is to obtain
Coefficient matrix construction indicates the weight matrix of figure, then realizes dictionary segmentation by spectral clustering;Character representation step is to choose most
Small reconstructed residual corresponding optimal dictionary item group carries out linear expression new data;Point is by asking to relationship reconstruction step
Optimal models are solved, study is maintained a little to the optimal binary-coding of relationship, to realize a little reconstruct to relationship.It obtains
Binary-coding can be used in the applications such as image retrieval.The present invention is based on machine learning and machine vision, devise a kind of base
In point to the characteristics of image binary-coding representation method of relational learning and reconstruct, there is lower optimization complexity, reduce volume
The code time can be used for performance and accuracy rate that image retrieval promotes image retrieval.
The present invention uses above-mentioned technical proposal, has the advantages that:The point provided by the invention that is based on is to relational learning
With the characteristics of image binary-coding representation method of reconstruct, caused by the characteristic dimension height occurred compared to general image retrieval
The features such as height stores, search efficiency is low can reduce memory requirement, improve computational efficiency, and this method is for existing
Hash function and the tightly coupled problem of the loss function of optimization aim in unsupervised algorithm coding learning process are a kind of loose couplings
The image binary-coding feature learning frame of conjunction, its Optimized model are convex, have lower optimization complexity, finally reach
To the purpose for promoting image retrieval performance and accuracy rate.
Description of the drawings
The present invention is done with reference to the accompanying drawings and detailed description and is further illustrated, it is of the invention above-mentioned or
Otherwise advantage will become apparent.
Fig. 1 is the characteristics of image binary-coding expression workflow based on point to relational learning and reconstruct of the embodiment of the present invention
Cheng Tu;
Fig. 2 is the dictionary segmentation step flow chart of the embodiment of the present invention;
Fig. 3 is the point of the embodiment of the present invention to relationship reconstruction step flow chart;
Fig. 4 is the example that the present invention retrieves highest matching degree on CIFAR-10 databases using eight binary-codings.Its
In leftmost picture be retrieving image, the picture on the right is retrieval result, and red frame represents the result of retrieval error.
Fig. 5 is the display for importing 10 data.
Specific implementation mode
The present invention will be further described with reference to the accompanying drawings and embodiments.
As shown in Figure 1, Figure 2 and Figure 3, constructed by the present invention based on point to the characteristics of image two of relational learning and reconstruct
The workflow of value coded representation method is roughly divided into following several stages:First stage put to the relational learning stage, including
Solve coefficient matrix, dictionary segmentation and character representation work;Second stage is put to relationship reconstruction stage, is mainly iterated to calculate
Obtain the work of optimal binary-coding.The characteristics of image two-value of relational learning and reconstruct is compiled based on point in the embodiment of the present invention
The specific construction step of code representation method is as follows:
Step 1, coefficient matrix is solved:Data are expressed as to the linear combination of dictionary and coefficient matrix, by solving to being
The restricted model of matrix number is to acquire coefficient matrix;
Step 2, dictionary is divided:The weight matrix for indicating the coefficient matrix construction obtained in step 1 a figure, at this
Dictionary item is divided by k by spectral clustering on a figure2A group;
Step 3, character representation:When given new samples, its reconstructed residual in all dictionary item groups is calculated, then
The minimum corresponding optimal dictionary item group of reconstructed residual is therefrom chosen to carry out linear expression, to complete a little to relationship
It practises;
Step 4, point reconstructs relationship:Optimal models are solved, to which the optimal two-value of relationship is compiled in study to holding point
Code realizes reconstruct of the point to relationship.
Step 1 includes:
Step 1-1 gives a matrix D=[x for including n image data1,x2,...,xn]∈Rd×n, wherein D's is every
One row indicate an image data, xnIndicate n-th image data, the dimension of each image data is d, and matrix D can be from table
It is shown as D=DC, the value of each element of matrix belongs to set of real numbers R, wherein C ∈ Rn×nIt is coefficient matrix, each row indicate every
A initial data corresponding reconstruction coefficients on dictionary, the dictionary, that is, data matrix D itself can to coefficient matrix application
It describes the low-rank constraint of the global structure of initial data and the sparsity of the local structure of initial data can be described about
Beam obtains being grouped dictionary such as drag:
Wherein | | | |*With | | | |1Representing matrix nuclear norm and l respectively1Norm, they are respectively by as to coefficient square
Battle array carries out the approximation of low-rank constraint and the approximation of sparsity constraints, and s.t. indicates that constraint, λ are one for balancing sparsity
The balance parameters of importance between low-rank;
Model in step 1-2, solution procedure 1-1 introduces an auxiliary variable J and obtains such as drag:
Augmented Lagrangian Functions L (C, J, the Y of the model1,Y2, μ) be:
Wherein, Y1,Y2, μ represents the introduced auxiliary variable of solving model, by each variable, that is, C, J, Y1,Y2,μ
Iteration update, which is solved, iteration update each variable rule it is as follows:
Jk+1=max (| Ck+1+Y2,k/μk|-λ/μk,0)
Y1,k+1=Y1,k+μk(D-DCk+1)
Y2,k+1=Y2,k+μk(Ck+1-Jk+1)
μk+1=min (μmax,ρμk)
Wherein, Ck+1Indicate the updated value of+1 iteration of variable C kth, JkIndicate that variable J kth time iteration is updated
Value, μkIndicate the updated value of variable μ kth time iteration, Y1,kIndicate variable Y1The updated value of kth time iteration, Y2,kIt indicates to become
Measure Y2The updated value of kth time iteration, μmaxIndicate that the maximum value of variable μ settings, ρ indicate that the growth factor of variable μ, Θ indicate
Singular value shrinkage operation accord with andThe coefficient matrix C learnt is acquired by iteration.
Step 2 includes:
The coefficient matrix C obtained in step 1 is indicated the weight matrix W of a figure, calculated as follows by step 2-1:
W=(C+CT)/2;
Step 2-2, structure n × n is to angle matrix Dw, DwIt is defined as follows:
Wherein diIt is defined as:di=∑ wij,
wijIt refer to the value of the number of the i-th row jth row of matrix W;
Step 2-3 defines Laplacian Matrix L=Dw- W, and according to L, DwCalculate Dw -1/2LDw -1/2;
Step 2-4 calculates Dw -1/2LDw -1/2Minimum k1The corresponding feature vector f of a characteristic value institute, by feature to
(citation is for amount f standardization:《Linear algebra (sixth version)》Higher Education Publishing House of department of mathematics of Tongji University), final group
At n × k1The eigenmatrix F of dimension;
Step 2-5, to every a line in eigenmatrix F as a k1The sample of dimension, total n sample, uses clustering method
Clustered that (citation is:《Machine learning》Publishing house of Zhou Zhihua Tsinghua University), cluster dimensionality k2, belonging to the i-th row
Class is exactly original xiClass belonging to (i.e. i-th of data of the matrix D in step 1-1) finally obtains cluster and divides c, thus by word
Allusion quotation item is divided into k2A group.
Step 3 includes:
Step 3-1, when a given new samples x ' calculates the reconstruction coefficients z of x ' for entire dictionary Di:
zi=(DTD+αI)-1DTX ',
Wherein, α is balance parameters, and I is unit matrix;
Step 3-2 is calculated for each dictionary item group respectively for the normalization residual error of x ':It is calculated by following formula
Normalization residual error r of k-th of dictionary item group for x 'k(x′):
Wherein, φk_d(zi) it is ziA part of coefficient, corresponding dictionary item is kth _ d dictionary group, DkIndicate kth
The dictionary that the data of a dictionary item group are formed;
The reconstruction coefficients z of x ' has been calculated in step 3-3iAfter normalization residual error in all dictionary item groups, selects and calculate
The dictionary item group of the reconstructed residual minimum arrived carries out linear expression as optimal dictionary item group, to optimal dictionary item group, by dilute
Thin coding is reconstructed expression to x ' and (is realized by LAE (Local Anchor Embedding) algorithms, algorithm is with reference to text
It offers:《Large Graph Construction for Scalable SemiSupervised Learning》Wei Liu,
Junfeng He, Shih-Fu Chang), to complete a little study to relationship.
Step 4 includes:
Such as drag is solved, to learn optimal binary-coding to which holding point is to relationship:
s.t. WrecWrec T=I
Wherein WrecIt is linear projection matrix, ZrecBe data matrix (i.e. step 3-3 obtain by sparse coding logarithm
According to the reconstruct data being reconstructed after indicating), B ∈ { -1,1 }c×nIt is a two values matrix, each column representing matrix ZrecMiddle data
Corresponding binary-coding, μrecIt is an offset parameter, s is a zooming parameter, | | | |FThe F norms of representing matrix, ⊙ tables
Show that the multiplication of corresponding element, model are convex, each unknown parameter μ is updated by iterationrec、B、Wrec, s optimize, directly
It is restrained to target function value.
It is described that each unknown parameter is updated to optimize by iteration, until target function value convergence, specifically include:
Step 4-1, random initializtion B and Wrec:
W=(C+CT)/2;
Initialize μrec=0, s=1;
Step 4-2 starts new round iteration, and iterations add one, if iterations are less than or equal to T times, carries out step
4-3, it is no to then follow the steps 4-4;
Step 4-3 updates B:
It enables
It calculates
Update Wrec:
It enables
It calculates
Update μrec:
μrec=column_mean (Zrec-s⊙Wrec TB),
Update s:
It enables
It calculates
Terminate epicycle iteration, return to step 4-2;
Step 4-4, by iterative calculation, the optimal binary-coding matrix B finally learnt completes weight of the point to relationship
Structure.
Embodiment
The present embodiment includes with lower part:
10 data are imported, each data are the image data of 784 dimensions.Fig. 5 is that the picture of 10 data is shown.
Passing point after relational learning step (including solving coefficient matrix, dictionary segmentation, character representation) to that can obtain
The original reconstruct for importing data indicates, that is, completes study of the point to relationship.
The reconstruct of this 10 data indicates as follows, that is, next step point is to the input data of relationship reconstruction step:
0 0 0 0 0 1 0 0
0 0 0 0 0 0 1 0
0 1 0 0 0 0 0 0
0 0 0 0 0 0 0 1
0.0397 0 0 0 0.9603 0 0 0
0.8445 0 0 0 0 0.1555 0 0
0 0 1 0 0 0 0 0
0 0 0 1 0 0 0 0
0.8739 0 0 0 0 0 0.1261 0
0 0.0534 0 0 0.9466 0 0 0
Next step is that point reconstructs relationship, can be to μ, and B, W, s variables are iterated update.
For this 10 data, can useCalculate the updated error of iteration each time.
Finally obtain the error of each iteration:
9.9504,6.6467,6.2349,6.1016,1.8460,1.7460,1.4298,1.1086,1.0954,
1.0954,1.0763,1.0600,1.0597,1.0498,1.0427,1.0427,1.0427,1.0378,1.0343,
1.0343;
It can be found that the error of iteration is smaller and smaller, illustrate that iteration update is effective.
The parameter value after iteration, including μ, B, W, s can be finally obtained, wherein B is exactly that the two-value finally to be obtained is compiled
Code:
μ:0 0 0 0 0 0 0 0
s:1
W:0.526380491743404, -0.703854357886537, -0.124134278972108,
0.202417906030414 ,-0.325885663271724,0.149605265605609 ,-0.0364373516095352-
0.203025641935073, -0.320148832680204, -0.296889358189778,0.187789441006430,
0.0556224552702578,0.0543180195432500, -0.700932457499529, -
0.337914469723198, -0.403186999011160, -0.614824643836743, -0.150284503746103, -
0.384163762219469,0.589286622028627, -0.218319098164965,0.134054680569501,
0.0609955279354016,0.187644253112132, -0.299070612273578, -0.225930941875870, -
0.488501528062168, -0.609111515373163,0.137953228767313,0.325901566927202, -
0.263695188929306, -0.234698359561226,0.000636095155782121,
0.0187859733491997,0.295608689874399,0.0392233814298232, -0.178144690543473,
0.227581141177460, -0.818998382462197,0.395520738197212, -0.246529543586242, -
0.169769446463417,0.251859151075042, -0.469358402701577, -0.673284555188396, -
0.130367042685072,0.299342831948313,0.258374510947432,0.255611415159194 ,-
0.0684087178891663, -0.533049300100546, -0.119796699797562,
0.0948826691189857, -0.526146514585480, -0.0854341580367617,
0.581685804423631, -0.163894212872911, -0.555842782524204,0.357982096845213, -
0.0383476153902449,0.574546274335455,0.128493460417319,0.208727206163476,
0.379966747912398;
The data of binary-coding B are as shown in table 1:
Table 1
One | Two | Three | Four | Five | Six | Seven | Eight | Nine | Ten |
0 | 1 | 1 | 0 | 1 | 1 | 1 | 0 | 1 | 0 |
0 | 1 | 1 | 0 | 1 | 1 | 0 | 0 | 0 | 0 |
0 | 0 | 0 | 1 | 1 | 1 | 0 | 0 | 1 | 0 |
1 | 1 | 1 | 0 | 0 | 0 | 0 | 1 | 0 | 1 |
0 | 1 | 1 | 1 | 0 | 0 | 1 | 1 | 0 | 1 |
0 | 0 | 1 | 0 | 0 | 0 | 0 | 1 | 0 | 0 |
1 | 1 | 1 | 1 | 1 | 1 | 1 | 0 | 0 | 0 |
1 | 0 | 0 | 0 | 0 | 1 | 0 | 1 | 0 | 0 |
First row to the tenth row are corresponding in turn to ten initial data in table 1, it can be seen that second and third original number
According to pattern be all number 1, the binary-coding that they are generated is also similar, and the two-value volume that they are generated with other data
Code is just far short of what is expected.
Fig. 4 is the example that the present invention retrieves highest matching degree on CIFAR-10 databases using eight binary-codings.Its
In leftmost picture be retrieving image, the picture on the right is retrieval result, wherein " Ours " indicate the present invention retrieval knot
Fruit, overstriking frame represent the result of retrieval error.The method of the present invention is compared with other methods accuracy rate higher as can be seen from Figure 4, compared with it
He can retrieve more correct result at method.
The present invention provides the characteristics of image binary-coding representation method based on point to relational learning and reconstruct, specific implementations
There are many method and approach of the technical solution, the above is only a preferred embodiment of the present invention, it is noted that for this
For the those of ordinary skill of technical field, without departing from the principle of the present invention, several improvement and profit can also be made
Decorations, these improvements and modifications also should be regarded as protection scope of the present invention.Each component part being not known in the present embodiment is available
The prior art is realized.
Claims (6)
1. based on point to the characteristics of image binary-coding representation method of relational learning and reconstruct, which is characterized in that including following step
Suddenly:
Step 1, coefficient matrix is solved:Image data table is shown as to the linear combination of dictionary and coefficient matrix, by solving to being
The restricted model of matrix number is to acquire coefficient matrix;
Step 2, dictionary is divided:The weight matrix of a figure is indicated the coefficient matrix construction obtained in step 1, on this map
Dictionary item is divided into k by spectral clustering2A group;
Step 3, character representation:When given new samples, its reconstructed residual in all dictionary item groups is calculated, is then therefrom selected
The minimum corresponding optimal dictionary item group of reconstructed residual is taken to carry out linear expression, completes study of the point to relationship;
Step 4, point reconstructs relationship:Optimal models are solved, it is real to which study to holding point is to the optimal binary-coding of relationship
Now put the reconstruct to relationship.
2. the method as described in claim 1, which is characterized in that step 1 includes:
Step 1-1 gives a matrix D=[x for including n image data1,x2,...,xn]∈Rd×n, each row of wherein D
Indicate an image data, xnIndicate n-th image data, the dimension of each image data is d, and matrix D can be expressed as D certainly
The value of=DC, each element of matrix belong to set of real numbers R, wherein C ∈ Rn×nIt is coefficient matrix, each row indicate each original
Data corresponding reconstruction coefficients on dictionary, the dictionary, that is, data matrix D itself can describe coefficient matrix application original
The global structure of data low-rank constraint and can describe initial data local structure sparsity constraints, obtain as
Drag is grouped dictionary:
mCin||C||*+λ||C||1S.t.D=DC, C >=0
Wherein | | | |*With | | | |1Representing matrix nuclear norm and l respectively1Norm, they respectively by as to coefficient matrix into
The approximation of the approximation and sparsity constraints of the constraint of row low-rank, s.t. indicate that constraint, λ are one for balancing sparsity and low-rank
The balance parameters of importance between property;
Model in step 1-2, solution procedure 1-1 introduces an auxiliary variable J and obtains such as drag:
mCin||C||*+λ||J||1S.t.D=DC, C=J, J >=0
Augmented Lagrangian Functions L (C, J, the Y of the model1,Y2, μ) be:
Wherein, Y1,Y2, μ represents the introduced auxiliary variable of solving model, by each variable, that is, C, J, Y1,Y2, μ's changes
Generation update, solves the model, and the rule that iteration updates each variable is as follows:
Jk+1=max (| Ck+1+Y2,k/μk|-λ/μk,0)
Y1,k+1=Y1,k+μk(D-DCk+1)
Y2,k+1=Y2,k+μk(Ck+1-Jk+1)
μk+1=min (μmax,ρμk)
Wherein, Ck+1Indicate the updated value of+1 iteration of variable C kth, JkIndicate the updated value of variable J kth time iteration, μkTable
Show the updated value of variable μ kth time iteration, Y1,kIndicate variable Y1The updated value of kth time iteration, Y2,kIndicate variable Y2Kth time
The updated value of iteration, μmaxIndicate that the maximum value of variable μ settings, ρ indicate that the growth factor of variable μ, Θ indicate that singular value is shunk
Operator andThe coefficient matrix C learnt is acquired by iteration.
3. method as claimed in claim 2, which is characterized in that step 2 includes:
The coefficient matrix C obtained in step 1 is indicated the weight matrix W of a figure, calculated as follows by step 2-1:
W=(C+CT)/2;
Step 2-2, structure n × n is to angle matrix Dw, DwIt is defined as follows:
Wherein diIt is defined as:di=∑wij,
wijIt refer to the value of the number of the i-th row jth row of matrix W;
Step 2-3 defines Laplacian Matrix L=Dw- W, and according to L, DwCalculate Dw -1/2LDw -1/2;
Step 2-4 calculates Dw -1/2LDw -1/2Minimum k1A corresponding feature vector f of characteristic value institute, feature vector f is marked
Standardization finally forms n × k1The eigenmatrix F of dimension;
Step 2-5, to every a line in eigenmatrix F as a k1The sample of dimension, total n sample, is gathered with clustering method
Class, cluster dimensionality k2, the class belonging to the i-th row is exactly original xiAffiliated class finally obtains cluster and divides c, to divide dictionary item
It is cut into k2A group.
4. method as claimed in claim 3, which is characterized in that step 3 includes:
Step 3-1, when a given new samples x ' calculates the reconstruction coefficients z of x ' for entire dictionary Di:
zi=(DTD+αI)-1DTX ',
Wherein, α is balance parameters, and I is unit matrix;
Step 3-2 is calculated for each dictionary item group respectively for the normalization residual error of x ':It is calculated k-th by following formula
Normalization residual error r of the dictionary item group for x 'k(x′):
Wherein, φk_d(zi) it is ziA part of coefficient, corresponding dictionary item is kth _ d dictionary group, DkIndicate k-th of word
The dictionary that the data of allusion quotation item group are formed;
The reconstruction coefficients z of x ' has been calculated in step 3-3iAfter normalization residual error in all dictionary item groups, selects and be calculated
The dictionary item group of reconstructed residual minimum carries out linear expression as optimal dictionary item group, to optimal dictionary item group, passes through sparse volume
Expression is reconstructed to x ' in code, to complete a little study to relationship.
5. method as claimed in claim 4, which is characterized in that step 4 includes:
Such as drag is solved, to learn optimal binary-coding to which holding point is to relationship:
s.t.WrecWrec T=I
Wherein WrecIt is linear projection matrix, ZrecIt is data matrix, B ∈ { -1,1 }c×nIt is a two values matrix, each column indicates square
Battle array ZrecThe corresponding binary-coding of middle data, μrecIt is an offset parameter, s is a zooming parameter, | | | |FRepresenting matrix
F norms, ⊙ indicate the multiplication of corresponding element, and model is convex, updates each unknown parameter μ by iterationrec、B、Wrec, s come into
Row optimization, until target function value is restrained.
6. method as claimed in claim 5, which is characterized in that described by iteration to update each unknown parameter excellent to carry out
Change, until target function value convergence, specifically includes:
Step 4-1, random initializtion B and Wrec:
WrIt is to initialize WrecAnd from the matrix of standardized normal distribution stochastical sampling, Wr=U ∑s VTIt indicates to WrMatrix carries out strange
Different value is decomposed, it is assumed that WrIt is m × n rank matrix, then U is m × m rank matrix after decomposing, and Σ is that m × n rank is diagonal
Matrix, and V* is n × n rank matrix, c refers to the length of binary-coding;
Initialize μrec=0, s=1;
Step 4-2 starts new round iteration, and iterations add one, if iterations are less than or equal to T times, carries out step 4-3,
It is no to then follow the steps 4-4;
Step 4-3 updates B:
It enables
It calculates
Update Wrec:
It enables
It calculatesR=VUT,Wrec=WrecR,
Update μrec:
μrec=column_mean (Zrec-s⊙Wrec TB),
Update s:
It enables
It calculates
Terminate epicycle iteration, return to step 4-2;
Step 4-4, by iterative calculation, the optimal binary-coding matrix B finally learnt completes reconstruct of the point to relationship.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810203371.8A CN108536750B (en) | 2018-03-13 | 2018-03-13 | Image feature binary coding representation method based on point-to-point relation learning and reconstruction |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810203371.8A CN108536750B (en) | 2018-03-13 | 2018-03-13 | Image feature binary coding representation method based on point-to-point relation learning and reconstruction |
Publications (2)
Publication Number | Publication Date |
---|---|
CN108536750A true CN108536750A (en) | 2018-09-14 |
CN108536750B CN108536750B (en) | 2022-03-18 |
Family
ID=63484445
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810203371.8A Active CN108536750B (en) | 2018-03-13 | 2018-03-13 | Image feature binary coding representation method based on point-to-point relation learning and reconstruction |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108536750B (en) |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103093239A (en) * | 2013-01-18 | 2013-05-08 | 上海交通大学 | Mapping method fusing dot pairs and neighborhood information |
US20130236090A1 (en) * | 2012-03-12 | 2013-09-12 | Fatih Porikli | Learning Dictionaries with Clustered Atoms |
CN103489203A (en) * | 2013-01-31 | 2014-01-01 | 清华大学 | Image coding method and system based on dictionary learning |
CN105469096A (en) * | 2015-11-18 | 2016-04-06 | 南京大学 | Feature bag image retrieval method based on Hash binary code |
-
2018
- 2018-03-13 CN CN201810203371.8A patent/CN108536750B/en active Active
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20130236090A1 (en) * | 2012-03-12 | 2013-09-12 | Fatih Porikli | Learning Dictionaries with Clustered Atoms |
CN103093239A (en) * | 2013-01-18 | 2013-05-08 | 上海交通大学 | Mapping method fusing dot pairs and neighborhood information |
CN103489203A (en) * | 2013-01-31 | 2014-01-01 | 清华大学 | Image coding method and system based on dictionary learning |
CN105469096A (en) * | 2015-11-18 | 2016-04-06 | 南京大学 | Feature bag image retrieval method based on Hash binary code |
Non-Patent Citations (2)
Title |
---|
XIANG-YANG WANG 等: "An image retrieval scheme with relevance feedback using feature reconstruction and SVM reclassification", 《NEUROCOMPUTING》 * |
黄秀杰 等: "基于最小重构误差的优化局部聚合描述符向量图像检索算法", 《计算机应用》 * |
Also Published As
Publication number | Publication date |
---|---|
CN108536750B (en) | 2022-03-18 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Rolet et al. | Fast dictionary learning with a smoothed Wasserstein loss | |
Song et al. | Auto-encoder based data clustering | |
CN109871454B (en) | Robust discrete supervision cross-media hash retrieval method | |
CN110222218B (en) | Image retrieval method based on multi-scale NetVLAD and depth hash | |
CN109766469A (en) | A kind of image search method based on the study optimization of depth Hash | |
CN109284411B (en) | Discretization image binary coding method based on supervised hypergraph | |
CN107329954B (en) | Topic detection method based on document content and mutual relation | |
CN109685093A (en) | Unsupervised adaptive features select method | |
CN111062428A (en) | Hyperspectral image clustering method, system and equipment | |
CN112488301A (en) | Food inversion method based on multitask learning and attention mechanism | |
CN114693923A (en) | Three-dimensional point cloud semantic segmentation method based on context and attention | |
US20230350913A1 (en) | Mapping of unlabeled data onto a target schema via semantic type detection | |
Denisova et al. | EM clustering algorithm modification using multivariate hierarchical histogram in the case of undefined cluster number | |
CN110309333B (en) | Depth hash image retrieval method based on cosine measurement | |
Remil et al. | Data‐Driven Sparse Priors of 3D Shapes | |
CN108536750A (en) | Based on point to the characteristics of image binary-coding representation method of relational learning and reconstruct | |
CN116229061A (en) | Semantic segmentation method and system based on image generation | |
CN112925994B (en) | Group recommendation method, system and equipment based on local and global information fusion | |
CN112925934B (en) | Similar image retrieval method, system, device and medium based on Hash coding | |
CN112487231B (en) | Automatic image labeling method based on double-image regularization constraint and dictionary learning | |
CN111275201A (en) | Sub-graph division based distributed implementation method for semi-supervised learning of graph | |
CN111724221A (en) | Method, system, electronic device and storage medium for determining commodity matching information | |
Wang et al. | Hyperspectral image compressed processing: Evolutionary multi-objective optimization sparse decomposition | |
CN117056550B (en) | Long-tail image retrieval method, system, equipment and storage medium | |
CN114462406B (en) | Method for acquiring first-appearing aviation keywords based on multi-head self-attention model |
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 |