CN103440332B - A kind of image search method strengthening expression based on relational matrix regularization - Google Patents
A kind of image search method strengthening expression based on relational matrix regularization Download PDFInfo
- Publication number
- CN103440332B CN103440332B CN201310399734.7A CN201310399734A CN103440332B CN 103440332 B CN103440332 B CN 103440332B CN 201310399734 A CN201310399734 A CN 201310399734A CN 103440332 B CN103440332 B CN 103440332B
- Authority
- CN
- China
- Prior art keywords
- image
- matrix
- relational matrix
- feature
- weights
- 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.)
- Active
Links
Landscapes
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The invention discloses a kind of method retrieving image from image instance storehouse based on relational matrix regularization Enhancement Method, comprise the steps of: step 1, input image to be retrieved;Step 2, extracts the feature of image in image to be retrieved and image instance storehouse;Step 3, chooses P image class from image instance feature database, chooses n width image construction sample data X from each image class;Step 4, manifold learning arithmetic based on spectral graph theory, sample data X is built three matrixes;Step 5, preliminary foundation strengthens relational matrix W ';Step 6, calculates regularization and strengthens relational matrix W*;Step 7, calculates generalized characteristic matrix A;Step 8, calculates final graphical representation;Step 9, calculates the graphical representation of image to be retrieved;Step 10, uses Euclidean distance to calculate image to be retrieved and the similarity of all images in image instance storehouse, according to image most like with image to be retrieved in similarity descending output image instance storehouse.
Description
Technical field
The invention belongs to field of image search, a kind of image retrieval based on relational matrix regularization Enhancement Method
Method.
Background technology
In today that Science and Technology Day reaches increasingly, along with the fast development of image acquisition process equipment and Internet technology is with universal
Application, the generation information resource with image as representative has become as has the war of status of equal importance with material, the energy
Slightly resource, its data volume has reached magnanimity scale the most, becomes current information and processes and the main body of construction of information resources.By
There is in image advantages such as containing much information, abundant in content, expressive force is strong, therefore the image of magnanimity scale is carried out effectively
Information processing and application, it has also become the key problem of numerous practical application area.
Due to current image date in magnanimity scale, and constantly increasing, traditional technological means cannot be fitted
Should this demand, this to the tissue of image, analyze, retrieve and the technology such as management is proposed brand-new challenge.Although
The image retrieval research being currently based on content has been achieved for the biggest progress, effectively overcomes literary composition based on manual mark
This information carries out the limitation of image retrieval, but also has a certain distance from the real practical stage, especially to image
High-level semantic understanding aspect.Major part method also only reside within the low-level image feature around image carry out semantic description and
Learn this level, relative to the mankind it will be appreciated that and use colourful semantic concept, bottom data feature
Ability to express still has the biggest limitation, therefore there is bigger gap, the most so-called " language between low-level image feature and high-level semantic
Justice wide gap " (semantic gap), thus cause in the accuracy rate and efficiency of image retrieval, do not reach the most far away actual answering
Needs, especially the multiple abundant semanteme of image is understood and retrieval aspect accurately and effectively.Even to this day,
" semantic gap " problem in image retrieval is not the most well solved, and remains the key of puzzlement researcher
One of property difficult problem.In the middle of the numerous technology solving this difficult problem, image retrieval technologies based on relevant feedback provides
A kind of feasible solution.Relevance Feedback in early days focuses primarily upon information based on relevant feedback, and correction is looked into
Ask vector i.e. characteristics of image, the such as every one dimensional numerical to query vector and redistribute weights, adjust the position of query vector
Put.In recent years, due to the rise of manifold learning, many researchers are diverted through manifold learning technology, by higher-dimension
Image data space dimensionality reduction seeks the immanent structure of image feature space, and its main theory hypothesis is to be regarded as by image
A kind of manifold, target be exactly find in it structural information.Find the lower-dimensional subspace being embedded in high dimensional data
Being the important means of the potential manifold of learning data, the learning method of manifold learning sub-spaces is all based on partial analysis.
Learnt the semantic subspace of its corresponding low-dimensional by the method for manifold learning, this assumes whole data with manifold learning
Collection only meets Euclidean distance in local and matches, and therefore by analyzing the local message of view data, excavates the language of local
Justice manifold structure is more meaningful for image retrieval.
Summary of the invention
Goal of the invention: the present invention is to solve the problems of the prior art, it is proposed that a kind of based on relational matrix regularization
Strengthen the image search method represented, efficiently solve under large-scale data, the quick and precisely search problem of image.
Summary of the invention: the invention discloses a kind of image search method based on relational matrix regularization Enhancement Method, should
Method retrieves image from image instance storehouse, comprises the steps of:
Step 1, inputs image to be retrieved;
Step 2, extracts the feature of image in image to be retrieved and image instance storehouse, by N-dimensional vector description each image,
N=112, obtains image instance feature database U=(u1,…,uM), uiFor the feature of image instance storehouse the i-th width image,
I=1 ... M, M are the picture number included in image instance storehouse, and feature v of image to be retrieved, and described image is real
Example storehouse includes the image class of more than 50, and each image class represents that a semantic category, each image class include 600 width
Above image;
Step 3, chooses P image class from image instance feature database, and P span 20~50, from each image
Class chooses n width image, n span 100~500, and P image class has n × P and open image construction sample data X;
Such as in the embodiment invented, therefrom choosing 30 image classes, each class illustrates different semantic categories, each
Class has 100 width images, has 3000 image construction sample datas X, X=(x1,…,xq), q=n × P, xiFor
The feature of the i-th width image in sample data, q is sample data size, and X is the matrix of 112 × q dimension;
Step 4, manifold learning arithmetic based on spectral graph theory, sample data X is built and strengthens relational matrix W, positive example
Relational matrix WPWith counter-example relational matrix WN;
Step 5, strengthens the relational matrix W built, and preliminary foundation strengthens relational matrix W ';
Step 6, strengthens relational matrix W ' by probability transfer matrix regularization and obtains regularization enhancing relational matrix W*;
Step 7, strengthens relational matrix W according to regularization*Build target equation, calculate generalized characteristic matrix A;
Step 8, utilizes generalized characteristic matrix A all images in image instance feature database to be carried out dimensionality reduction, i.e.
AU=A* (u1,…,uM)=(A*u1,…,A*uM), remember yi=A*xi, i=1 ... M, obtain final figure
As representing Y=(y1,…,yM), yiFor the feature after the i-th width characteristics of image dimensionality reduction of image instance storehouse;
Step 9, utilizes generalized characteristic matrix A to characteristics of image v dimensionality reduction to be retrieved, obtains the image table of image to be retrieved
Show f=A*v;
Step 10, according to the final graphical representation of step 8 and the Euclidean of the graphical representation of the image to be retrieved of step 9
Distance calculates image to be retrieved and the similarity of all images in image instance storehouse, i.e. calculates image dimensionality reduction feature f to be retrieved
With the Euclidean distance of feature after image instance feature database each image Feature Dimension Reduction | | f-yi||2, i=1 ... M, yiFor image
Feature after case library the i-th width characteristics of image dimensionality reduction, according in similarity descending output image instance storehouse with to be retrieved
The image that image is most like.
In step 2, characteristics of image includes color moment, Tamura textural characteristics, Gabor textural characteristics, color histogram.
Step 4 specifically includes following steps: randomly select piece image in sample data X, calculates this image and sample
The Euclidean distance of other images in notebook data X, utilizes relevance feedback retrieval technology, according to the similar figure returned in result
Picture is corresponding with inhomogeneity image sets up positive example set and counter-example set, and uses simple k near neighbor method opening relationships square
Battle array, i.e. belong to k neighbour and be same image class two images between weights be 1, be otherwise 0.
Step 4 use imbeding relation based on feedback technique widen the manifold learning calculation as spectral graph theory of the ARE method
Method, comprises the following steps:
(1) first sample data X is built relational matrix W, from sample data X, randomly draw piece image I, image I
Feature be xi, use k near neighbor method to calculate xiWith the Euclidean distance of other characteristics of image in sample data X, obtain with
K width image most like for image I, wherein k span 5~10;
Arbitrarily taking out piece image T from k width image to belong to, the feature of image T is xt, then between image I and image T
Weights WitBeing 1, the weights between image beyond image I and k width image are 0;I.e. xi∈Nk(xt) or xt∈
Nk(xi), Wit=1, wherein Nk(xi) represent image xiK neighbour set, Nk(xt) represent image xtK neighbour set;
The value obtaining relational matrix W, relational matrix W the i-th row t row is Wit;
Formula is:
The image belonging to same image class in k width image with image I is designated as positive example set Pos, the figure of different images class
As being designated as counter-example set Neg,;
(2) positive example relational matrix W is builtPIf image R and image I belongs to same image class and broadly falls into k width figure
Picture, and the feature of image R is xr, then the weights between image I and image R are 1, the figure beyond image I and k width image
Weights between Xiang are 0;That is,For the weights between image I and image R, xi,xr∈Pos
For representing feature xi,xrBelong to positive example set Pos, positive example relational matrix WPI-th row r row value bePublic
Formula is:
(3) counter-example relational matrix W is builtNIf image H and image I belongs to different images class and broadly falls into k width figure
Picture, the feature of image H is xh, then the weights between image I and image H are 1, the image beyond image I and k width image
Between weights be 0;I.e. xi∈ Pos and xh∈ Neg or xh∈ Pos andTable
Show feature xiBelong to positive example set Pos, xh∈ Neg represents feature xhBelong to counter-example set Neg, xh∈ Pos represents feature
xhBelong to positive example set Pos, xi∈ Neg represents feature xiBelong to counter-example set Neg,For between image I and image H
Weights, counter-example relational matrix WNThe i-th h be classified asFormula is:
Finally build and obtain three relational matrix W, WPAnd WN, it neutralizes the relation used for calculating generalized characteristic matrix to need
Matrix.
Step 5 specifically includes following steps: from relational matrix W, if image z is neighbour's image of image i,
And image z is also neighbour's image of image j, then following formula is used to calculate the weights W ' strengthened between image i and image jij:
W′ij=ΣzWizWjz
Wherein WizFor the weights of image i Yu image z, WjzFor the weights of image j Yu image z, W 'ijIt is enhancing relation
The i row jth train value of matrix W '.
Step 6 specifically includes following steps:
The neighbor relationships repeatedly propagated between image obtains new enhancing relational matrix W ", formula is W "=W ' * W ';
The transfer relationship between image, corresponding transfer matrix are P=[P to utilize transition probability matrix to representij]n×n,
Pij=p (j | i) it is that in sample data X, any image i, to the transition probability of any image j, selects and figure according to Euclidean distance
As the n width image that i is most like, the feature of image j is xj, the computing formula of transition probability P (j | i) is:
Wherein dij=| | xi-xj||2, represent the Euclidean distance of image i and image j feature.
Use the model W that following formula calculated relationship matrix regularization strengthensR:
WR=η P+ (1-η) geT
Wherein, η is the probability that image i transfers to that this event of image j occurs, and (1-η) is the probability that image i redirects at random,
G=(1/n) e, wherein g is a uniformly random distribution vector, and e is that n ties up unit column vector, n the most each image class
Picture number, e=(1,1 ...)T, the i-th row jth of matrix P is classified as P (j | i);
New relation weights between image i and image jComputing formula is:
w″ijFor the weights of image i Yu image j, w "ijFor W " i-th row jth row value,Figure is jumped to for image i
As the probability weights of j,For WRI-th row jth row value;
Finally give regularization and strengthen relational matrix W*, W*The i-th row jth be classified as
Step 7 comprises the steps:
First from sample data X, choose feature x of any two width imagesiAnd xj, the relation weights of two width images are Wij,
The positive example relation weights of two width images areThe counter-example relation weights of two width images areAccording to following target equation
It is calculated generalized characteristic matrix A:
X(LN-γLP)XTA=λXLXTA,
L is the Laplacian Matrix of relational matrix W, LNFor counter-example relational matrix WNLaplacian Matrix, LPFor just
Example relational matrix WPLaplacian Matrix, γ is that the ratio to counter-example image number and positive example image number is directly proportional
Constant, XTRepresenting the transposed matrix of sample data X, λ represents the characteristic value of equation solution.
In the present invention, ARE is for widening relation embedding grammar (Augmented Relation Embedding), and one widens pass
The manifold learning dimension-reduction algorithm that system's figure embeds, ARE mainly utilizes positive example relational matrix to embed the overall situation with counter-example relational matrix
In relational matrix, find projection matrix, i.e. generalized characteristic matrix, thus realize the dimensionality reduction to data characteristics.
The principle of the invention is, sample data X=(x1,…,xN),xi∈Rm, relational matrix W ∈ R between data pointN×N
Representing, entry of a matrix element has weighed the similarity between every pair of data point.Diagonal matrix D and corresponding Laplacian Matrix L by
Following formula defines:
L=D-W
DiiThe i-th row i-th for diagonal matrix D arranges, it is assumed that generalized characteristic matrix is A, completes original data space by projection
Low-dimensional embed, A can be minimized by following formula and try to achieve:
Each column a of matrix AjIndependent role, therefore above formula can be write as argminaΣij(aTxi-aTxj)2Wij, wherein a is to be asked
Characteristic vector.Make yi=aTxi, then have:
Wherein, y represents the projection on this projection vector of a of all data, and y=aTX.Coordinate after conversion is limited,
DiiRepresent the number being connected with i-th point, illustrate this importance degree in a way, and then can increase about
Bundle makes yTDy=1.After this constraint can make the some conversion that importance is high, its coordinate value is more nearly territory initial point, allows
Close quarters is positioned at initial point, and the object function equation finally solved becomes:
From the point of view of derivation, relational matrix W plays leading role in whole process, and data point y after projection also has with W
Close relationship, such as, work as WijTime bigger, represent xiAnd xjSimilarity is relatively big, y after dimensionality reductioniAnd yjBetween distance also should
The smaller the better;If WijLess, represent xiAnd xjSimilarity is less, y after dimensionality reductioniAnd yjBetween distance also should be the bigger the better.
Here similarity relation can represent whether belong to same classification between data, the similarity between homogeneous data is the most very
High;For not having the data of classification information, the similarity between data is just weighed by neighbor relationships, between neighbour's data point
Similarity should be higher;For neither homogeneous data, the most not there is the similarity meeting between the data point of neighbor relationships
Ratio is relatively low, typically makes Wij=0.
Beneficial effect: the present invention utilizes relational matrix regularization enhancing expression that image instance feature is carried out dimensionality reduction, the party
Method can effectively strengthen the relation between similar image, has merged the classification information of data during building relational matrix,
Make it expand to easily in the framework of semi-supervised learning, thus make full use of flag data and Unlabeled data, have
Improving the stability of algorithm and reducing computation complexity of effect, makes image querying have higher accuracy rate simultaneously, because of
This relational matrix regularization strengthens the image search method represented and has higher use value.
Accompanying drawing explanation
Fig. 1 is the flow chart of the present invention.
Fig. 2 bit image case library Feature Dimension Reduction flow chart.
Fig. 3 is characteristics of image dimensionality reduction flow chart to be retrieved.
Fig. 4 is that images relations strengthens schematic diagram.
Fig. 5 is image random walk model schematic.
The regularization of Fig. 6 position strengthens relation schematic diagram.
Fig. 7 is image searching result schematic diagram.
Detailed description of the invention
As it is shown in figure 1, the invention discloses a kind of image search method represented based on regularization enhancing relational matrix;
Comprise the steps of:
Step 1: input image to be retrieved;
As shown in figures 2-3, build image regulation enhancing relational matrix mainly to be carried out, to image by step 2~step 6
Example aspects storehouse dimensionality reduction is carried out by step 8, step 9 carry out characteristics of image dimensionality reduction to be retrieved:
Step 2, extracts image to be retrieved and the characteristics of image of image instance storehouse image, feature includes color moment, Tamura
Textural characteristics, Gabor textural characteristics and color histogram, describe each image, N=112 with the vector of N-dimensional, treat
Retrieval image is v, and image instance feature database is U=(u1,…,uM), M is image instance storehouse total number of images, and U is N × M
Dimension matrix;
Step 3, the character representation each image after extraction, from image instance storehouse, choose 30 image classes, each class
Representing a semantic category, each class has 100 width images, has 3000 images, and as sample data X,
X=(x1,…,x3000), matrix X is 112 × 3000 dimensions;
Step 4, manifold learning arithmetic based on spectral graph theory, sample data X is built and strengthens relational matrix W, positive example
Relational matrix WPWith counter-example relational matrix WN;
Step 5, strengthens relational matrix W, and preliminary foundation strengthens relational matrix W ';
Step 6, strengthens relational matrix W ' by probability transfer matrix regularization and obtains regularization enhancing relational matrix W*;
Step 7, strengthens relational matrix W according to regularization*With positive example relational matrix WPWith counter-example relational matrix WNBuild
Object function, solves generalized characteristic matrix A;
Step 8, utilizes generalized characteristic matrix A that images all in image instance storehouse are carried out dimensionality reduction, i.e. AU=A*
(u1,…,uM)=(A*u1..., A*uM), remember yi=A*xi, i=1 ... M, obtain final graphical representation
Y=(y1,…,yM);
Step 9, as it is shown on figure 3, utilize generalized characteristic matrix A that characteristics of image v to be retrieved is carried out dimensionality reduction, is treated
The graphical representation f=A*v of retrieval image;
Step 10, uses Euclidean distance to calculate image to be retrieved and the similitude of all images in image instance storehouse, i.e. counts
Calculate | | f-yi||2, i=1 ... M, according to most like with image to be retrieved in similarity descending output image instance storehouse
Image.
Step 2 specifically includes following steps:
Extraction each image feature, i.e. iamge description aspect is by color moment (RGB color): 9 dimensions;Color moment (LUV
Color space): 9 dimensions;Tamura textural characteristics: 6 dimensions;Gabor textural characteristics: 24 dimensions;Color histogram (HSV
Color space): 64 dimension compositions.
Step 4 specifically includes following steps: randomly select piece image in sample data X, calculates this image and sample
The Euclidean distance of other images in notebook data X, utilizes relevance feedback retrieval technology, according to the similar figure returned in result
Picture is corresponding with inhomogeneity image sets up positive example set and counter-example set, and uses simple k near neighbor method opening relationships square
Battle array, i.e. belong to k neighbour and be same image class two images between weights be 1, be otherwise 0.
Step 4 use imbeding relation based on feedback technique widen the manifold learning calculation as spectral graph theory of the ARE method
Method, comprises the following steps:
(1) first sample data X is built relational matrix W, from sample data X, randomly draw piece image I, image I
Feature be xi, use k near neighbor method to calculate xiWith the Euclidean distance of other characteristics of image in sample data X, obtain with
K width image most like for image I, wherein k span 5~10;
Arbitrarily taking out piece image T from k width image to belong to, the feature of image T is xt, then between image I and image T
Weights WitBeing 1, the weights between image beyond image I and k width image are 0;I.e. xi∈Nk(xt) or xt∈
Nk(xi), Wit=1, wherein Nk(xi) represent image xiK neighbour set, Nk(xt) represent image xtK neighbour set;
The value obtaining relational matrix W, relational matrix W the i-th row t row is Wit;
Formula is:
The image belonging to same image class in k width image with image I is designated as positive example set Pos, the figure of different images class
As being designated as counter-example set Neg,;
(2) positive example relational matrix W is builtPIf image R and image I belongs to same image class and broadly falls into k width figure
Picture, and the feature of image R is xr, then the weights between image I and image R are 1, the figure beyond image I and k width image
Weights between Xiang are 0;That is,For the weights between image I and image R, xi,xr∈Pos
For representing feature xi,xrBelong to positive example set Pos, positive example relational matrix WPI-th row r row value bePublic
Formula is:
(3) counter-example relational matrix W is builtNIf image H and image I belongs to different images class and broadly falls into k width figure
Picture, the feature of image H is xh, then the weights between image I and image H are 1, the image beyond image I and k width image
Between weights be 0;I.e. xi∈ Pos and xh∈ Neg or xh∈ Pos andxi∈ Pos table
Show feature xiBelong to positive example set Pos, xh∈ Neg represents feature xhBelong to counter-example set Neg, xh∈ Pos represents feature
xhBelong to positive example set Pos, xi∈ Neg represents feature xiBelong to counter-example set Neg,For between image I and image H
Weights, counter-example relational matrix WNThe i-th h be classified asFormula is:
Finally build and obtain three relational matrix W, WPAnd WN, it neutralizes the relation used for calculating generalized characteristic matrix to need
Matrix.
Step 5 specifically includes following steps: from relational matrix W, if image z is neighbour's image of image i,
And image z is also neighbour's image of image j, then following formula is used to calculate the weights W ' strengthened between image i and image jij:
W′ij=∑zWizWjz
Wherein WizFor the weights of image i Yu image z, WjzFor the weights of image j Yu image z, W 'ijIt is enhancing relation
The i row jth train value of matrix W '.
Step 6 specifically includes following steps:
The neighbor relationships repeatedly propagated between image obtains new enhancing relational matrix W ", formula is W "=W ' * W ';
The transfer relationship between image, corresponding transfer matrix are P=[P to utilize transition probability matrix to representij]n×n,
Pij=p (j | i) it is that in sample data X, any image i, to the transition probability of any image j, selects and figure according to Euclidean distance
As the n width image that i is most like, the feature of image j is xj, the computing formula of transition probability P (j | i) is:
Wherein dij=| | xi-xj||2, represent the Euclidean distance of image i and image j feature.
Use the model W that following formula calculated relationship matrix regularization strengthensR:
WR=η P+ (1-η) geT
Wherein, η is the probability that image i transfers to that this event of image j occurs, and (1-η) is the probability that image i redirects at random,
G=(1/n) e, wherein g is a uniformly random distribution vector, and e is that n ties up unit column vector, n the most each image class
Picture number, e=(1,1 ...)T, the i-th row jth of matrix P is classified as P (j | i);
New relation weights between image i and image jComputing formula is:
w″ijFor the weights of image i Yu image j, w "ijFor W " i-th row jth row value,Figure is jumped to for image i
As the probability weights of j,For WRI-th row jth row value;
Finally give regularization and strengthen relational matrix W*, W*The i-th row jth be classified as
Step 7 comprises the steps:
First from sample data X, choose feature x of any two width imagesiAnd xj, the relation weights of two width images are Wij,
The positive example relation weights of two width images areThe counter-example relation weights of two width images areAccording to following target equation
It is calculated generalized characteristic matrix A:
X(LN-γLP)XTA=λXLXTA,
L is the Laplacian Matrix of relational matrix W, LNFor counter-example relational matrix WNLaplacian Matrix, LPFor just
Example relational matrix WPLaplacian Matrix, γ is that the ratio to counter-example image number and positive example image number is directly proportional
Constant, XTRepresenting the transposed matrix of sample data X, λ represents the characteristic value of equation solution.
Embodiment 1
The present embodiment includes with lower part:
1. one image I to be retrieved of input;
2. the dimension of abstract image case library and the characteristics of image of image to be retrieved, each feature and its correspondence is as follows:
Color moment (RGB color): 9 dimensions;Color moment (LUV color space): 9 dimensions;Tamura textural characteristics:
6 dimensions;Gabor textural characteristics: 24 dimensions;Color histogram (hsv color space): 64 dimensions.So each image will
Describing with the vector of 112 dimensions, image to be retrieved is v, and image instance feature database is U=(u1,…,uM), M is image
Case library total number of images, U is that N × M ties up matrix;
3. choosing training sample data from the U of characteristics of image storehouse, each image extraction feature represents, and therefrom chooses 30
Individual image class, each class represents a semantic category, and each class has 100 width images, has 3000 images, and by it
As sample data X, X=(x1,…,x3000), matrix X is 112 × 3000 dimensions;.
4. in sample data X, randomly select piece image, calculate this image and the Europe of other images in sample data X
Formula distance, utilizes relevance feedback retrieval technology, sets up according to the similar image returned in result is corresponding with inhomogeneity image
Positive example set and counter-example set, and use simple k near neighbor method opening relationships matrix, i.e. belong to k neighbour and be
Weights between two images of same image class are 1, are otherwise 0.
Step 4 use imbeding relation based on feedback technique widen the manifold learning calculation as spectral graph theory of the ARE method
Method, comprises the following steps:
(1) first sample data X is built relational matrix W, from sample data X, randomly draw piece image I, image I
Feature be xi, use k near neighbor method to calculate xiWith the Euclidean distance of other characteristics of image in sample data X, obtain with
K width image most like for image I, wherein k value 5;
Arbitrarily taking out piece image T from k width image to belong to, the feature of image T is xt, then between image I and image T
Weights WitBeing 1, the weights between image beyond image I and k width image are 0;I.e. xi∈Nk(xt) or xt∈
Nk(xi), Wit=1, wherein Nk(xi) represent image xiK neighbour set, Nk(xt) represent image xtK neighbour set;
The value obtaining relational matrix W, relational matrix W the i-th row t row is Wit;
Formula is:
The image belonging to same image class in k width image with image I is designated as positive example set Pos, the figure of different images class
As being designated as counter-example set Neg,;
(2) positive example relational matrix W is builtPIf image R and image I belongs to same image class and broadly falls into k width figure
Picture, and the feature of image R is xr, then the weights between image I and image R are 1, the figure beyond image I and k width image
Weights between Xiang are 0;That is,For the weights between image I and image R, xi,xr∈Pos
For representing feature xi,xrBelong to positive example set Pos, positive example relational matrix WPI-th row r row value bePublic
Formula is:
(3) counter-example relational matrix W is builtNIf image H and image I belongs to different images class and broadly falls into k width figure
Picture, the feature of image H is xh, then the weights between image I and image H are 1, the image beyond image I and k width image
Between weights be 0;I.e. xi∈ Pos and xh∈ Neg or xh∈ Pos andxi∈ pos table
Show feature xiBelong to positive example set Pos, xh∈ Neg represents feature xhBelong to counter-example set Neg, xh∈ Pos represents feature
xhBelong to positive example set Pos, xi∈ Neg represents feature xiBelong to counter-example set Neg,For between image I and image H
Weights, counter-example relational matrix WNThe i-th h be classified asFormula is:
Finally build and obtain three relational matrix W, WPAnd WN, for calculating the relational matrix that generalized characteristic matrix needs are used.
5. set up initial relation and strengthen matrix W ', from relational matrix W, if image z is neighbour's figure of image i
Picture, and image z is also neighbour's image of image j, then use following formula to calculate the weights W ' strengthened between image i and image jij:
W′ij=ΣzWizWjz
Wherein WizFor the weights of image i Yu image z, WjzFor the weights of image j Yu image z, W 'ijIt is enhancing relation
The i row jth train value of matrix W '.As shown in Figure 4, image 3 is neighbour's image of image 1 to instantiation, image 3
Be neighbour's image of image 2, connect with the solid line having arrow between image and represent neighbor relationships, image 1 and image 2 it
Between with dotted line connect, the relation between representative image 1 and image 2 need strengthen.
6. build probability transfer matrix WRAnd carry out regularization to strengthening relational matrix W ',
The neighbor relationships repeatedly propagated between image obtains new enhancing relational matrix W ", formula is w "=w ' * w ';
The transfer relationship between image, corresponding transfer matrix are P=[P to utilize transition probability matrix to representij]n×n,
Pij=P(j | it is i) that in sample data X, any image i, to the transition probability of any image j, selects and figure according to Euclidean distance
As the n width image that i is most like, the feature of image j is xj, the computing formula of transition probability P (j | i) is:
Wherein dij=| | xi-xj||2, represent the Euclidean distance of image i and image j feature.
Use the model W that following formula calculated relationship matrix regularization strengthensR:
WR=η P+ (1-η) geT
Wherein, η is the probability that image i transfers to that this event of image j occurs, and η is taken as 0.85, and (1-η) is that image i jumps at random
The probability turned, g=(1/n) e, wherein g is a uniformly random distribution vector, and e is that n ties up unit column vector, and n is the most every
The picture number of individual image class, e=(1,1 ...)T, the i-th row jth of matrix P is classified as P (j | i);
New relation weights between image i and image jComputing formula is:
w″ijFor the weights of image i Yu image j, w "ijFor W " i-th row jth row value,Figure is jumped to for image i
As the probability weights of j,For WRI-th row jth row value;
Finally give regularization and strengthen relational matrix W*, W*The i-th row jth be classified asInstantiation such as Fig. 5~6 institute
Showing, the probability transfer weights relation between Fig. 5 representative image, in Fig. 6, the picture left above 1 represents the enhancing relation square between image
Battle array W ", connect with solid line between two width images is neighbour's image, and dotted line connects that to represent be that enhancing between two width images is closed
System, top right plot 2 represents the transition probability matrix W between imageR, with realizing connecting existence transfer between representative image between image
Relation, the regularization between lower Fig. 3 representative image strengthens relational matrix W*, by W " and WRIt is multiplied and obtains;
7. strengthen matrix W according to the relation after regularization*Build object function, solve generalized characteristic matrix A,
First from sample data X, choose feature x of any two width imagesiAnd xj, the relation weights of two width images are Wij,
The positive example relation weights of two width images areThe counter-example relation weights of two width images areAccording to following target equation
It is calculated generalized characteristic matrix A:
X(LN-γLP)XTA=λXLXTA,
L is the Laplacian Matrix of relational matrix W, LNFor counter-example relational matrix WNLaplacian Matrix, LPFor just
Example relational matrix WPLaplacian Matrix, γ is that the ratio to counter-example image number and positive example image number is directly proportional
Constant, XTRepresenting the transposed matrix of sample data X, λ represents the characteristic value of equation solution.
Mainly utilize generalized characteristic matrix A that view data in image instance feature database is carried out dimensionality reduction and obtain final figure
As representing, i.e. AU=A* (u1,…,uM)=(A*ui,…,A*uM), remember yi=A*xi, i=1 ... M,
Whole graphical representation is Y=(y1,…,yM);
Mainly utilize generalized characteristic matrix A that characteristics of image v to be retrieved is carried out dimensionality reduction, obtain the image of image to be retrieved
Represent f, f=A*v;
10. calculating image to be retrieved and image similarity in image instance storehouse:
Use Euclidean distance to calculate image to be retrieved and the similitude of all images in image instance storehouse, i.e. calculate
||f-yi||2, i=1 ... M, | | f-yi||2The least similarity is the biggest, according to similarity descending output image instance storehouse
In the image most like with image to be retrieved.As it is shown in fig. 7, it is real with image to calculate image to be retrieved according to Euclidean distance
The similitude of all images in example storehouse, according to 4 most like images of the descending output of similarity.
Embodiment 2
Fig. 1 is embodiment 2 retrieval flow figure, and in figure, image sources is public Corel5k database.In figure, 2 is right
Original image pre-processes, with color moment, Tamura textural characteristics, Gabor textural characteristics and color histogram chart
Show piece image, 3 selected characteristic sample in figure, from image instance storehouse, choose 30 image classes, each class illustrates
One semantic category, each class has 100 width images, has 3000 width images, in order to improve calculating speed, only uses result
Concentrate front 400 width images as the overall situation data set, for opening relationships matrix W, positive example relational matrix WP, counter-example
Relational matrix WN.Then relational matrix W is carried out enhancing and obtains W ', and utilize probability transfer matrix WRRegularization increases
Strong relational matrix, obtains W*, then according to the enhancing relational matrix W of regularization*Solve the generalized character square of object function
Battle array A, finally utilizes generalized characteristic matrix A that characteristics of image in image instance storehouse and characteristics of image to be retrieved are carried out dimensionality reduction,
Image to be retrieved is retrieved, utilizes Euclidean distance to calculate image to be retrieved and the similarity of image in image instance storehouse,
According to image most like with image to be retrieved in similarity descending output image instance storehouse.
The invention provides a kind of regularization and strengthen the image search method that relational matrix represents, implement this technical side
The method of case and approach are a lot, and the above is only the preferred embodiment of the present invention, it is noted that for this technology
For the those of ordinary skill in field, under the premise without departing from the principles of the invention, it is also possible to make some improvement and profit
Decorations, these improvements and modifications also should be regarded as protection scope of the present invention.Each part the clearest and the most definite in the present embodiment is equal
Available prior art is realized.
Claims (4)
1. the image search method strengthening to represent based on relational matrix regularization, it is characterised in that the method is from figure
As case library is retrieved image, comprise the steps of:
Step 1, inputs image to be retrieved;
Step 2, extracts the feature of image in image to be retrieved and image instance storehouse, by N-dimensional vector description each image,
N=112, obtains image instance feature database and the feature of image to be retrieved, and described image instance storehouse includes more than 50
Image class, each image class represents a semantic category, and each image class includes the image of more than 600 width;
Step 3, chooses P image class from image instance feature database, and P span 20~50, from each image
Class chooses n width image, n span 100~500, and P image class has n × P and open image construction sample data X;
Step 4, manifold learning arithmetic based on spectral graph theory, sample data X is built relational matrix W, positive example relation
Matrix WPWith counter-example relational matrix WN;
Step 5, strengthens the relational matrix W built, and preliminary foundation strengthens relational matrix W ';
Step 6, strengthens relational matrix W by probability transfer matrix regularization′Obtain regularization and strengthen relational matrix W*:
Step 7, strengthens relational matrix W according to regularization*Build target equation, calculate generalized characteristic matrix A;
Step 8, utilizes generalized characteristic matrix A that all images in image instance feature database are carried out dimensionality reduction, obtains final
Graphical representation;
Step 9, utilizes generalized characteristic matrix A to image dimensionality reduction to be retrieved, obtains the graphical representation of image to be retrieved;
Step 10, according to the final graphical representation of step 8 and the Euclidean of the graphical representation of the image to be retrieved of step 9
Distance calculates image to be retrieved and the similarity of all images in image instance storehouse, according to similarity descending output figure
As image most like with image to be retrieved in case library;
In step 2, characteristics of image includes color moment, Tamura textural characteristics, Gabor textural characteristics, color histogram;
Step 4 specifically includes following steps: randomly select piece image in sample data X, calculates this image and sample
The Euclidean distance of other images in notebook data X, utilizes relevance feedback retrieval technology, according to the similar figure returned in result
Picture is corresponding with inhomogeneity image sets up positive example set and counter-example set, and uses simple k near neighbor method opening relationships square
Battle array, i.e. belong to k neighbour and be same image class two images between weights be 1, be otherwise 0;
Step 4 use imbeding relation based on feedback technique widen the manifold learning calculation as spectral graph theory of the ARE method
Method, comprises the following steps:
(1) first sample data X is built relational matrix W, from sample data X, randomly draw piece image I, image I
Feature be xi, use k near neighbor method to calculate xiWith the Euclidean distance of other characteristics of image in sample data X, obtain with
K width image most like for image I, wherein k span 5~10;
Arbitrarily taking out piece image T from k width image, the feature of image T is xt, then the power between image I and image T
Value WitBeing 1, the weights between image beyond image I and k width image are 0;I.e. xi∈Nk(xt)or xt∈
Nk(xi), Wit=1, wherein Nk(xi) represent image xiK neighbour set, Nk(xt) represent image xtK neighbour set;
The value obtaining relational matrix W, relational matrix W the i-th row t row is Wit;
The image belonging to same image class in k width image with image I is designated as positive example set Pos, the figure of different images class
As being designated as counter-example set Neg;
(2) positive example relational matrix W is builtPIf image R and image I belongs to same image class and broadly falls into k width figure
Picture, and the feature of image R is xr, then the weights between image I and image R are 1, the figure beyond image I and k width image
Weights between Xiang are 0;That is, xi,xr∈Pos, For the weights between image I and image R,
xi,xr∈ Pos is for representing feature xi,xrBelong to positive example set Pos, positive example relational matrix WPThe i-th row r row value i.e.
ForFormula is:
(3) counter-example relational matrix W is builtNIf image H and image I belongs to different images class and broadly falls into k width figure
Picture, the feature of image H is xh, then the weights between image I and image H are 1, the image beyond image I and k width image
Between weights be 0;I.e. xi∈Pos and xh∈ Neg, or xh∈Pos and xi∈Neg,xi∈Pos
Represent feature xiBelong to positive example set Pos, xh∈ Neg represents feature xhBelong to counter-example set Neg, xh∈ Pos represents special
Levy xhBelong to positive example set Pos, xi∈ Neg represents feature xiBelong to counter-example set Neg,For image I and image H it
Between weights, counter-example relational matrix WNThe i-th h be classified asFormula is:
Finally build and obtain three relational matrix W, WPAnd WN。
A kind of image search method strengthening expression based on relational matrix regularization the most according to claim 1, its
Being characterised by, step 5 specifically includes following steps: from relational matrix W, if image z is the neighbour of image i
Image, and image z is also neighbour's image of image j, then use following formula to calculate the power strengthened between image i and image j
Value W 'ij: W 'ij=∑zWizWjz, wherein WizFor the weights of image i Yu image z, WjzPower for image j Yu image z
Value, W 'ijIt is the i-th row jth train value strengthening relational matrix W '.
A kind of image search method strengthening expression based on relational matrix regularization the most according to claim 2, its
Being characterised by, step 6 specifically includes following steps:
The neighbor relationships repeatedly propagated between image obtains new enhancing relational matrix W ", formula is W "=W ' * W ';
The transfer relationship between image, corresponding transfer matrix are P=[P to utilize transition probability matrix to representij]n×n,
Pij=p (j | i) it is that in sample data X, any image i, to the transition probability of any image j, selects and figure according to Euclidean distance
As the n width image that i is most like, the feature of image j is xj, the computing formula of transition probability P (j | i) is:
Wherein dij=| | xi-xj||2, represent the Euclidean distance of image i and image j feature;
Use the model W that following formula calculated relationship matrix regularization strengthensR:
WR=η P+ (1-η) geT
Wherein, η is the probability that image i transfers to that this event of image j occurs, and (1-η) is the probability that image i redirects at random,
G=(1/n) e, wherein g is a uniformly random distribution vector, and e is that n ties up unit column vector, n the most each image class
Picture number, e=(1,1 ...)T, the i-th row jth of matrix P is classified as P (j | i);
New relation weights between image i and image jComputing formula is:
w″ijFor the weights of image i Yu image j, w "ijFor W " i-th row jth row value,Figure is jumped to for image i
As the probability weights of j,For WRI-th row jth row value;
Finally give regularization and strengthen relational matrix W*, W*The i-th row jth be classified as
A kind of image search method strengthening expression based on relational matrix regularization the most according to claim 3, its
It is characterised by, step 7 comprises the steps:
First from sample data X, choose feature x of any two width imagesiAnd xj, the relation weights of two width images are Wij,
The positive example relation weights of two width images areThe counter-example relation weights of two width images areAccording to following target equation
It is calculated generalized characteristic matrix A:
X(LN-γLP)XTA=λ XLXTA
L is the Laplacian Matrix of relational matrix W, LNFor counter-example relational matrix WNLaplacian Matrix, LPClose for positive example
It it is matrix WPLaplacian Matrix, γ is the constant that the ratio to counter-example image number and positive example image number is directly proportional,
XTRepresenting the transposed matrix of sample data X, λ represents the characteristic value of equation solution.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201310399734.7A CN103440332B (en) | 2013-09-05 | 2013-09-05 | A kind of image search method strengthening expression based on relational matrix regularization |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201310399734.7A CN103440332B (en) | 2013-09-05 | 2013-09-05 | A kind of image search method strengthening expression based on relational matrix regularization |
Publications (2)
Publication Number | Publication Date |
---|---|
CN103440332A CN103440332A (en) | 2013-12-11 |
CN103440332B true CN103440332B (en) | 2016-08-17 |
Family
ID=49694025
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201310399734.7A Active CN103440332B (en) | 2013-09-05 | 2013-09-05 | A kind of image search method strengthening expression based on relational matrix regularization |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN103440332B (en) |
Families Citing this family (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN104614411B (en) * | 2015-02-16 | 2017-03-01 | 天津大学 | The electricity tomography Lp regularization reconstruction method being declined based on p vector equal difference |
CN106874268A (en) * | 2015-12-10 | 2017-06-20 | 富士通株式会社 | Image search method and image retrieval apparatus |
CN106055576B (en) * | 2016-05-20 | 2018-04-10 | 大连理工大学 | A kind of fast and effectively image search method under large-scale data background |
CN106886553B (en) * | 2016-12-27 | 2020-07-28 | 浙江宇视科技有限公司 | Image retrieval method and server |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102236717A (en) * | 2011-07-13 | 2011-11-09 | 清华大学 | Image retrieval method based on sketch feature extraction |
CN103064941A (en) * | 2012-12-25 | 2013-04-24 | 深圳先进技术研究院 | Image retrieval method and device |
CN103116762A (en) * | 2013-03-20 | 2013-05-22 | 南京大学 | Image classification method based on self-modulated dictionary learning |
Family Cites Families (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US7603348B2 (en) * | 2007-01-26 | 2009-10-13 | Yahoo! Inc. | System for classifying a search query |
-
2013
- 2013-09-05 CN CN201310399734.7A patent/CN103440332B/en active Active
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102236717A (en) * | 2011-07-13 | 2011-11-09 | 清华大学 | Image retrieval method based on sketch feature extraction |
CN103064941A (en) * | 2012-12-25 | 2013-04-24 | 深圳先进技术研究院 | Image retrieval method and device |
CN103116762A (en) * | 2013-03-20 | 2013-05-22 | 南京大学 | Image classification method based on self-modulated dictionary learning |
Non-Patent Citations (3)
Title |
---|
基于谱图理论的流形学习算法;罗四维等;《计算机研究与发展》;20060730(第07期);第1176页右栏第35-42行、第1177页左栏第1-3行 * |
流形学习方法在Web图像检索中的应用研究;鲁柯;《中国博士学位论文全文数据库 信息科技辑》;20080115(第01期);正文第32页第17-24行 * |
面向语义的图像检索关键技术研究;贺广南;《中国优秀硕士学位论文全文数据库 信息科技辑》;20111015(第10期);正文第21页第14-27行、第22页第1-12行、第24页第16-18行、第25页第13-17行、第27页第1-3行 * |
Also Published As
Publication number | Publication date |
---|---|
CN103440332A (en) | 2013-12-11 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Ye et al. | Remote sensing image retrieval using convolutional neural network features and weighted distance | |
Yu et al. | Beyond bilinear: Generalized multimodal factorized high-order pooling for visual question answering | |
CN107563498B (en) | Image description method and system based on visual and semantic attention combined strategy | |
CN107766324B (en) | Text consistency analysis method based on deep neural network | |
CN106650789A (en) | Image description generation method based on depth LSTM network | |
CN107644235A (en) | Image automatic annotation method based on semi-supervised learning | |
CN105279495A (en) | Video description method based on deep learning and text summarization | |
CN108170736A (en) | A kind of document based on cycle attention mechanism quickly scans qualitative method | |
CN105631479A (en) | Imbalance-learning-based depth convolution network image marking method and apparatus | |
Zhi et al. | Action unit analysis enhanced facial expression recognition by deep neural network evolution | |
Cai et al. | Intelligent question answering in restricted domains using deep learning and question pair matching | |
Wulamu et al. | Multiscale road extraction in remote sensing images | |
Zhang et al. | Structured weak semantic space construction for visual categorization | |
CN103440332B (en) | A kind of image search method strengthening expression based on relational matrix regularization | |
Wang et al. | LaST: Label-free self-distillation contrastive learning with transformer architecture for remote sensing image scene classification | |
He et al. | A multi-attentive pyramidal model for visual sentiment analysis | |
CN109657039A (en) | A kind of track record information extraction method based on the double-deck BiLSTM-CRF | |
CN110263174A (en) | - subject categories the analysis method based on focus | |
Tang et al. | A bidirectional multi-paragraph reading model for zero-shot entity linking | |
Zheng et al. | Knowledge is power: Hierarchical-knowledge embedded meta-learning for visual reasoning in artistic domains | |
Dai et al. | Rethinking the form of latent states in image captioning | |
CN109815478A (en) | Medicine entity recognition method and system based on convolutional neural networks | |
Yang et al. | CLIP-KD: An Empirical Study of Distilling CLIP Models | |
CN107423697A (en) | Activity recognition method based on non-linear fusion depth 3D convolution description | |
Ding et al. | Context-aware semantic type identification for relational attributes |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
C06 | Publication | ||
PB01 | Publication | ||
C10 | Entry into substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
C14 | Grant of patent or utility model | ||
GR01 | Patent grant |