CN103440332A - Image searching method based on relation matrix regularization enhancement representation - Google Patents

Abstract

The invention discloses a method for retrieving images from an image instance database based on a relational matrix regularization enhancement method, comprising the following steps: step 1, inputting the image to be retrieved; step 2, extracting the features of the image to be retrieved and the image in the image instance database ; Step 3, select P image classes from the image instance feature library, and select n images from each image class to form sample data X; Step 4, based on the manifold learning algorithm of spectrogram theory, construct three Matrix; step 5, initially establish the enhanced relationship matrix W'; step 6, calculate the regularized enhanced relationship matrix W ^* ; step 7, calculate the generalized feature matrix A; step 8, calculate the final image representation; step 9, calculate the image to be retrieved image representation; step 10, using the Euclidean distance to calculate the similarity between the image to be retrieved and all the images in the image instance database, and output the image in the image instance database that is most similar to the image to be retrieved in descending order of similarity.

Description

An Image Retrieval Method Based on Regularized Enhanced Representation of Relation Matrix

technical field

The invention belongs to the field of image retrieval, in particular to an image retrieval method based on a relationship matrix regularization enhancement method.

Background technique

In today's increasingly advanced technology, with the rapid development and popularization of image acquisition and processing equipment and Internet technology, the new generation of information resources represented by images has become a strategic resource with the same importance as materials and energy. It has reached a massive scale and has become the main body of current information processing and information resource construction. Because images have the advantages of large amount of information, rich content, and strong expressiveness, effective information processing and application of massive images has become a core issue in many practical application fields.

Due to the massive scale of current image data and its continuous growth, traditional technical means can no longer meet this demand, which poses new challenges to technologies such as image organization, analysis, retrieval, and management. Although the current research on content-based image retrieval has made great progress, effectively overcoming the limitations of image retrieval based on manually annotated text information, there is still a certain distance from the real practical stage, especially for high-level images. Semantic understanding. Most methods only stay at the level of semantic description and learning around the underlying features of the image. Compared with the rich and colorful semantic concepts that humans can understand and use, the expressive ability of the underlying data features is still very limited. Therefore, the underlying features There is a large gap between the high-level semantics and the so-called "semantic gap", which leads to the fact that the accuracy and efficiency of image retrieval are far from meeting the needs of practical applications, especially for image retrieval. A variety of rich semantics for accurate and effective understanding and retrieval. Today, the "semantic gap" problem in image retrieval has not been well resolved, and it is still one of the key problems that plague researchers. Among the many techniques to solve this problem, image retrieval based on relevance feedback provides a feasible solution. Early relevant feedback technology mainly focused on information based on relevant feedback, modifying the query vector, that is, image features, such as reassigning weights to each dimension value of the query vector, adjusting the position of the query vector, etc. In recent years, due to the rise of manifold learning, many researchers have turned to the manifold learning technology to reduce the dimensionality of the high-dimensional image data space to explore the internal structure of the image feature space. The main theoretical assumption is to regard the image as a A manifold, the goal is to discover its internal structure information. Discovering low-dimensional subspaces embedded in high-dimensional data is an important means of learning the latent manifold of data, and the learning methods of subspaces in manifold learning are all based on local analysis. The corresponding low-dimensional semantic subspace is learned through the manifold learning method, which is consistent with the manifold learning assumption that the entire data set only satisfies the Euclidean distance locally. Therefore, by analyzing the local information of the image data, the local semantics can be discovered. The manifold structure is more meaningful for image retrieval.

Contents of the invention

Purpose of the invention: In order to solve the problems in the prior art, the present invention proposes an image retrieval method based on the regularized enhanced representation of the relationship matrix, which effectively solves the problem of fast and accurate retrieval of images under large-scale data.

SUMMARY OF THE INVENTION: The present invention discloses an image retrieval method based on a relational matrix regularization enhancement method. The method retrieves an image from an image instance database, including the following steps:

Step 1, input the image to be retrieved;

Step 2, extract the features of the image to be retrieved and the images in the image instance library, describe each image with N-dimensional vector, N=112, and obtain the image instance feature library U=(u ₁ ,...,u _M ), u _i is the image The feature of the i-th image in the example library, i=1,...M, M is the number of images contained in the image example library, and the feature v of the image to be retrieved, the image example library includes more than 50 image classes, each An image class represents a semantic class, and each image class includes more than 600 images;

Step 3. Select P image classes from the image instance feature library, and the value range of P is 20-50. Select n images from each image class, and the value range of n is 100-500. There are n×P images in P image classes Image constitutes sample data X; for example, in an embodiment of the invention, 30 image classes are selected therefrom, each class represents a different semantic class, each class has 100 images, and a total of 3000 images constitute sample data X, X= (x ₁ ,...,x _q ), q=n×P, x _i is the feature of the i-th image in the sample data, q is the size of the sample data, and X is a 112×q-dimensional matrix;

Step 4, based on the manifold learning algorithm of spectrogram theory, construct the enhanced relationship matrix W, the positive example relationship matrix W ^P and the negative example relationship matrix W ^N for the sample data X;

Step 5, enhance the constructed relationship matrix W, and preliminarily establish the enhanced relationship matrix W';

Step 6, regularizing the enhanced relationship matrix W' with the help of the probability transition matrix to obtain the regularized enhanced relationship matrix W ^* ;

Step 7. Construct the objective equation according to the regularized enhanced relationship matrix W ^* , and calculate the generalized feature matrix A;

Step 8, use the generalized feature matrix A to reduce the dimensionality of all images in the image instance feature library, that is, AU=A*(u ₁ ,...,u _M )=(A*u ₁ ,...,A*u _M ), Record y _i =A* _xi , i=1,...M, to obtain the final image representation Y=(y ₁ ,...,y _M ), and y _i is the dimensionality-reduced feature of the i-th image in the image instance library;

Step 9, use the generalized feature matrix A to reduce the dimensionality of the image feature v to be retrieved, and obtain the image representation f=A*v of the image to be retrieved;

Step 10, calculate the similarity between the image to be retrieved and all images in the image instance library according to the Euclidean distance between the final image representation in step 8 and the image representation of the image to be retrieved in step 9, that is, calculate the dimensionality reduction feature f and _The Euclidean distance of each image feature in the image instance feature library after dimensionality reduction|| ^fy _i || Output the image most similar to the image to be retrieved in the image instance library from large to small.

Image features in step 2 include color moments, Tamura texture features, Gabor texture features, and color histograms.

Step 4 specifically includes the following steps: randomly select an image in the sample data X, calculate the Euclidean distance between the image and other images in the sample data X, and use the relevant feedback retrieval technology to correspond to images of the same type and images of different types in the returned results Set up a set of positive examples and a set of negative examples, and use the simple k-nearest neighbor method to establish a relationship matrix, that is, the weight between two images belonging to the k-nearest neighbor and the same image class is 1, otherwise it is 0.

In step 4, the embedded relationship widening ARE method based on feedback technology is used as the manifold learning algorithm of spectrogram theory, including the following steps:

(1) First construct a relationship matrix W for the sample data X, randomly select an image I from the sample data X, the feature of the image I is x _i , use the k nearest neighbor method to calculate the Euclidean relationship between x _i and other image features in the sample data X distance to obtain k images most similar to image I, where k ranges from 5 to 10;

An image T is randomly selected from k images, and the feature of image T is x _t , then the weight W it between image I and image T is 1, and the weight W _it between image I and images other than k images The value is 0; that is, x _i ∈ N ^k (x _t ) or x _t ∈ ^{N k} ( _xi ), W _it = 1, where N ^k (xi ₎ represents the k-nearest neighbor set of image x _i , N ^k (x _t ) represents the k-nearest neighbor set of the image x _t ; the relationship matrix W is obtained, and the value of the i-th row and the t-column of the relationship matrix W is W _it ;

The formula is:

The images belonging to the same image class as image I in the k images are recorded as positive example set Pos, and the images of different image classes are recorded as negative example set Neg;

为图像I与图像R之间的权值，x_i,x_r∈Pos为表示特征x_i,x_r属于正例集合Pos，正例关系矩阵W^P的第i行第r列的值即为

公式为：(2) Construct a positive example relationship matrix W ^P , if image R and image I belong to the same image class and both belong to k images, and the feature of image R is x _r , then the weight between image I and image R is 1 , the weight between image I and images other than k images is 0; that is,

is the weight between the image I and the image R, x _i , x _r ∈ Pos is the representation feature x _i , x _r belongs to the positive example set Pos, the value of the i-th row and the r-th column of the positive example relationship matrix W ^P is

The formula is:

表示特征x_i属于正例集合Pos，x_h∈Neg表示特征x_h属于反例集合Neg，x_h∈Pos表示特征x_h属于正例集合Pos，x_i∈Neg表示特征x_i属于反例集合Neg，

为图像I与图像H之间的权值，反例关系矩阵W^N的第i第h列为

公式为：(3) Construct a negative example relationship matrix W ^N , if image H and image I belong to different image categories and both belong to k images, and the feature of image H is x _h , then the weight between image I and image H is 1, and image The weight between I and images other than k images is 0; that is, x _i ∈ Pos and x _h ∈ Neg or x _h ∈ Pos and

Indicates that the feature x _i belongs to the positive example set Pos, x _h ∈ Neg indicates that the feature x _h belongs to the negative example set Neg, x _h ∈ Pos indicates that the feature x _h belongs to the positive example set Pos, x _i ∈ Neg indicates that the feature x _i belongs to the negative example set Neg,

is the weight between image I and image H, and the ith column h of the counterexample relationship matrix W ^N is

The formula is:

Finally, three relationship matrices W, W ^P and W ^N are constructed, and the sum is the relationship matrix needed to calculate the generalized feature matrix.

Step 5 specifically includes the following steps: starting from the relationship matrix W, if image z is the neighbor image of image i, and image z is also the neighbor image of image j, then use the following formula to calculate the weight W between the enhanced image i and image j ′ _ij : W′ _ij = Σ _z W _iz W _jz

Where W _iz is the weight of image i and image z, W _jz is the weight of image j and image z, and W' _ij is the value of row i and column j of the enhanced relationship matrix W'.

Step 6 specifically includes the following steps:

The neighbor relationship among the multi-propagation images gets a new enhanced relationship matrix W″, the formula is W″=W′*W′;

Use the transition probability matrix to represent the transition relationship between images, the corresponding transition matrix is P=[P _ij ] _n×n , and P _ij =p(j|i) is the transition from any image i to any image j in the sample data X Transition probability, select n images most similar to image i according to Euclidean distance, the feature of image j is x _j , the calculation formula of transition probability P(j|i) is:

Where d _ij ＝|| _xi -x _j || ² represents the Euclidean distance between image i and image j features.

The model W ^R enhanced by the regularization of the relationship matrix is calculated by the following formula:

W ^R ＝ηP+(1-η)ge ^T

Among them, η is the probability of image i transferring to image j, (1-η) is the probability of image i jumping randomly, g=(1/n)e, where g is a uniform random distribution vector, e is n-dimensional unit column vector, n is the number of images of each image class, e=(1,1,...) ^T , the i-th row and j-th column of the matrix P is P(j|i);

计算公式为：The new relationship weight between image i and image j

The calculation formula is:

${\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; ij</span>}}^{<span\; class="notranslate">\; *</span>}<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; ij</span>}}^{<span\; class="notranslate">\; \&Prime;</span>}<span\; class="notranslate">\; \&Center\; Dot;</span>{\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; ij</span>}}^{\mathrm{<span\; class="notranslate">\; R</span>}}$

为图像i跳转到图像j的概率权值，

为W^R的第i行第j列的值；w″ _ij is the weight of image i and image j, w″ _ij is the value of row i and column j of W″,

is the probability weight of image i jumping to image j,

is the value of row i and column j of W ^R ;

Finally, the regularized enhanced relationship matrix W ^* is obtained, and the i-th row and j-th column of W ^* are

Step 7 includes the following steps:

根据以下目标方程计算得到广义特征矩阵A：First select the features x _i and x _j of any two images from the sample data X, the relationship weight of the two images is W _ij , and the positive relationship weight of the two images is The weight of the negative relationship between the two images is

The generalized characteristic matrix A is calculated according to the following objective equation:

X(L ^N -γL ^P )X ^T A=λXLX ^T A,

L is the Laplacian matrix of the relationship matrix W, L ^N is the Laplacian matrix of the negative relationship matrix W ^N , L ^P is the Laplacian matrix of the positive relationship matrix W ^P , and γ is the number of negative examples A constant proportional to the ratio of the number of positive images, X ^T represents the transpose matrix of the sample data X, and λ represents the eigenvalue of the equation solution.

In the present invention, ARE is an Augmented Relation Embedding method (Augmented Relation Embedding), a manifold learning dimensionality reduction algorithm for augmenting relational graph embedding. ARE mainly uses the positive example relationship matrix and the negative example relationship matrix to embed into the global relationship matrix to find the projection matrix. That is, the generalized feature matrix, so as to realize the dimensionality reduction of data features.

The principle of the present invention is that the sample data X=(x ₁ ,…,x _N ), x _i ∈ R ^m , and the relationship matrix W ∈ R ^N×N between data points is represented, and the elements of the matrix measure the relationship between each pair of data points similarity. The diagonal matrix D and the corresponding Laplacian matrix L are defined by:

${\mathrm{<span\; class="notranslate">\; D.</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}{\mathrm{<span\; class="notranslate">\; W</span>}}_{\mathrm{<span\; class="notranslate">\; ij</span>}}<span\; class="notranslate">\; ,</span><span\; class="notranslate">\; \&ForAll;</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>$

L=D-W

D _ii is the i-th row and i-th column of the diagonal matrix D. Assuming the generalized feature matrix is A, the low-dimensional embedding of the original data space is completed through projection. A can be obtained by minimizing the following formula:

$\underset{\mathrm{<span\; class="notranslate">\; ij</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; A</span>}}^{\mathrm{<span\; class="notranslate">\; T</span>}}{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; A</span>}}^{\mathrm{<span\; class="notranslate">\; T</span>}}{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}{\mathrm{<span\; class="notranslate">\; W</span>}}_{\mathrm{<span\; class="notranslate">\; ij</span>}}$

Each column a _j of matrix A acts independently, so the above formula can be written as argmin _a Σ _ij (a ^T x _i -a ^T x _j ) ² W _ij , where a is the eigenvector to be sought. Let y _i =a ^T x _i , then:

$\underset{\mathrm{<span\; class="notranslate">\; ij</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}{\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; ij</span>}}<span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; ij</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}{\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}{\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; ij</span>}}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 2</span>}\underset{\mathrm{<span\; class="notranslate">\; ij</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}{\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}{\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}{\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; ij</span>}}<span\; class="notranslate">\; +</span>\underset{\mathrm{<span\; class="notranslate">\; ij</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}{\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}{\mathrm{<span\; class="notranslate">\; W</span>}}_{\mathrm{<span\; class="notranslate">\; ij</span>}}$

$<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 2</span>}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}{\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}{\mathrm{<span\; class="notranslate">\; D.</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 2</span>}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; ij</span>}}{\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}{\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}{\mathrm{<span\; class="notranslate">\; W</span>}}_{\mathrm{<span\; class="notranslate">\; ij</span>}}$

$<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; 2</span>}\mathrm{<span\; class="notranslate">\; the\; y</span>}}^{\mathrm{<span\; class="notranslate">\; T</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; D.</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; W</span>}<span\; class="notranslate">\; )</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; 2</span>}\mathrm{<span\; class="notranslate">\; the\; y</span>}}^{\mathrm{<span\; class="notranslate">\; T</span>}}\mathrm{<span\; class="notranslate">\; Ly</span>}$

Wherein, y represents the projection of all data on the projection vector a, and y=a ^T X . For the coordinate constraints after conversion, D _ii represents the number connected to the i-th point, which explains the importance of the point to some extent, and then can increase constraints so that y ^T Dy = 1. This constraint can make the coordinate values of the highly important points closer to the origin of the domain after conversion, so that the densest area is located at the origin, and the final objective function equation to be solved becomes:

${\mathrm{<span\; class="notranslate">\; a</span>}}^{<span\; class="notranslate">\; *</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; arg</span>}\underset{\mathrm{<span\; class="notranslate">\; a</span>}}{\mathrm{<span\; class="notranslate">\; min</span>}}{\mathrm{<span\; class="notranslate">\; a</span>}}^{\mathrm{<span\; class="notranslate">\; T</span>}}{\mathrm{<span\; class="notranslate">\; XLX</span>}}^{\mathrm{<span\; class="notranslate">\; T</span>}}\mathrm{<span\; class="notranslate">\; a</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; s</span>}<span\; class="notranslate">\; .</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; .</span>{\mathrm{<span\; class="notranslate">\; a</span>}}^{\mathrm{<span\; class="notranslate">\; T</span>}}{\mathrm{<span\; class="notranslate">\; XDX</span>}}^{\mathrm{<span\; class="notranslate">\; T</span>}}\mathrm{<span\; class="notranslate">\; a</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}$

From the point of view of the derivation process, the relationship matrix W plays a leading role in the whole process, and the projected data point y also has a close relationship with W. For example, when W _ij is larger, it means that x _i and x _j are more similar. The distance between y _i and y _j after dimension reduction should be as small as possible; if W _ij is small, it means that the similarity between xi _and x _j is small, and the distance between y _i and y _j after dimension reduction should also be larger and closer. good. The similarity relationship here can indicate whether the data belong to the same category, and the similarity between similar data is naturally high; for data without category information, the similarity between data is measured by the neighbor relationship, and the similarity between neighboring data points The degree of similarity should be high; for data points that are neither the same kind of data nor have a neighbor relationship, the similarity between data points will be relatively low, generally set W _ij =0.

Beneficial effects: the present invention utilizes the regularized enhanced representation of the relational matrix to reduce the dimensionality of image instance features. This method can effectively strengthen the relationship between images of the same type, and the category information of the data is integrated in the process of constructing the relational matrix, making it easy to Extended to the framework of semi-supervised learning, so as to make full use of labeled data and unlabeled data, effectively improve the stability of the algorithm and reduce the computational complexity, and at the same time make the image query have a high accuracy rate, so the regularization of the relationship matrix enhances the representation The image retrieval method has a high use value.

Description of drawings

Fig. 1 is a flowchart of the present invention.

Figure 2. Dimensionality reduction flow chart of bit image instance library features.

Fig. 3 is a flowchart of dimensionality reduction of image features to be retrieved.

Fig. 4 is a schematic diagram of image relationship enhancement.

Figure 5 is a schematic diagram of an image random walk model.

Figure 6. Schematic diagram of bit regularization enhancement relationship.

Fig. 7 is a schematic diagram of image retrieval results.

Detailed ways

As shown in Figure 1, the present invention discloses a kind of image retrieval method based on regularized enhanced relational matrix representation; It comprises the following steps:

Step 1: Input the image to be retrieved;

As shown in Figures 2 to 3, the construction of the image regularization enhancement relationship matrix is mainly carried out by steps 2 to 6, the dimensionality reduction of the image instance feature library is carried out by step 8, and the dimensionality reduction of the image features to be retrieved is carried out by step 9:

Step 2, extract the image features of the image to be retrieved and the image instance library image, the features include color moment, Tamura texture feature, Gabor texture feature and color histogram, use N-dimensional vector to describe each image, N=112, to be The retrieved image is v, the image instance feature library is U=(u ₁ ,...,u _M ), M is the total number of images in the image instance library, and U is an N×M dimensional matrix;

Step 3, the extracted features represent each image, select 30 image classes from the image instance library, each class represents a semantic class, each class has 100 images, a total of 3000 images, and use them as sample data X, X=(x ₁ ,...,x ₃₀₀₀ ), the matrix X is 112×3000 dimensions;

Step 4, based on the manifold learning algorithm of spectrogram theory, construct the enhanced relationship matrix W, the positive example relationship matrix W ^P and the negative example relationship matrix W ^N for the sample data X;

Step 5, enhance the relationship matrix W, and initially establish the enhanced relationship matrix W';

Step 6, regularizing the enhanced relationship matrix W' with the help of the probability transition matrix to obtain the regularized enhanced relationship matrix W ^* ;

Step 7, construct the objective function according to the regularized enhanced relationship matrix W ^* , the positive example relationship matrix ^WP and the negative example relationship matrix ^WN , and solve the generalized feature matrix A;

Step 8, use the generalized feature matrix A to reduce the dimensionality of all images in the image instance library, that is, AU=A*(u ₁ ,…,u _M )=(A*u ₁ ,…,A*u _M ), record y _i =A* _xi , i=1,...M, to obtain the final image representation Y=(y ₁ ,...,y _M );

Step 9, as shown in Figure 3, use the generalized feature matrix A to reduce the dimensionality of the feature v of the image to be retrieved, and obtain the image representation f=A*v of the image to be retrieved;

Step 10, use the Euclidean distance to calculate the similarity between the image to be retrieved and all the images in the image instance library, that is, calculate ||fy _i || ² , i=1,...M, and output the image instance library in descending order of similarity The image most similar to the image to be retrieved.

Step 2 specifically includes the following steps:

Extract the features of each image, that is, the image description is composed of color moments (RGB color space): 9 dimensions; color moments (LUV color space): 9 dimensions; Tamura texture features: 6 dimensions; Gabor texture features: 24 dimensions; color histogram (HSV color space): 64-dimensional composition.

Step 4 specifically includes the following steps: randomly select an image in the sample data X, calculate the Euclidean distance between the image and other images in the sample data X, and use the relevant feedback retrieval technology to correspond to images of the same type and images of different types in the returned results Set up a set of positive examples and a set of negative examples, and use the simple k-nearest neighbor method to establish a relationship matrix, that is, the weight between two images belonging to the k-nearest neighbor and the same image class is 1, otherwise it is 0.

In step 4, the embedded relationship widening ARE method based on feedback technology is used as the manifold learning algorithm of spectrogram theory, including the following steps:

(1) First construct a relationship matrix W for the sample data X, randomly select an image I from the sample data X, the feature of the image I is x _i , use the k nearest neighbor method to calculate the Euclidean relationship between x _i and other image features in the sample data X distance to obtain k images most similar to image I, where k ranges from 5 to 10;

An image T is randomly selected from k images, and the feature of image T is x _t , then the weight W it between image I and image T is 1, and the weight W _it between image I and images other than k images The value is 0; that is, x _i ∈ ^{N k} (x _t ) or x _t ∈ ^{N k} ( _xi ), W _it = 1, where N ^k (xi ₎ represents the k-nearest neighbor set of image x _i , N ^k (x _t ) represents the k-nearest neighbor set of the image x _t ; the relationship matrix W is obtained, and the value of the i-th row and the t-column of the relationship matrix W is W _it ;

The formula is:

The images belonging to the same image class as image I in the k images are recorded as positive example set Pos, and the images of different image classes are recorded as negative example set Neg;

为图像I与图像R之间的权值，x_i,x_r∈Pos为表示特征x_i,x_r属于正例集合Pos，正例关系矩阵W^P的第i行第r列的值即为

公式为：(2) Construct a positive example relationship matrix W ^P , if image R and image I belong to the same image class and both belong to k images, and the feature of image R is x _r , then the weight between image I and image R is 1 , the weight between image I and images other than k images is 0; that is,

is the weight between the image I and the image R, x _i , x _r ∈ Pos is the representation feature x _i , x _r belongs to the positive example set Pos, the value of the i-th row and the r-th column of the positive example relationship matrix W ^P is

The formula is:

x_i∈Pos表示特征x_i属于正例集合Pos，x_h∈Neg表示特征x_h属于反例集合Neg，x_h∈Pos表示特征x_h属于正例集合Pos，x_i∈Neg表示特征x_i属于反例集合Neg，为图像I与图像H之间的权值，反例关系矩阵W^N的第i第h列为

公式为：(3) Construct a negative example relationship matrix W ^N , if image H and image I belong to different image categories and both belong to k images, and the feature of image H is x _h , then the weight between image I and image H is 1, and image The weight between I and images other than k images is 0; that is, x _i ∈ Pos and x _h ∈ Neg or x _h ∈ Pos and

x _i ∈ Pos means the feature x _i belongs to the positive set Pos, x _h ∈ Neg means the feature x _h belongs to the negative set Neg, x _h ∈ Pos means the feature x _h belongs to the positive set Pos, x _i ∈ Neg means the feature x _i belongs to Negative set Neg, is the weight between image I and image H, and the ith column h of the counterexample relationship matrix W ^N is

The formula is:

Finally, three relationship matrices W, W ^P and W ^N are constructed, and the sum is the relationship matrix needed to calculate the generalized feature matrix.

Step 5 specifically includes the following steps: starting from the relationship matrix W, if image z is the neighbor image of image i, and image z is also the neighbor image of image j, then use the following formula to calculate the weight W between the enhanced image i and image j ′ _ij :

W′ _ij ＝∑ _z W _iz W _jz

Where W _iz is the weight of image i and image z, W _jz is the weight of image j and image z, and W' _ij is the value of row i and column j of the enhanced relationship matrix W'.

Step 6 specifically includes the following steps:

The neighbor relationship among the multi-propagation images gets a new enhanced relationship matrix W″, the formula is W″=W′*W′;

Use the transition probability matrix to represent the transition relationship between images, the corresponding transition matrix is P=[P _ij ] _n×n , and P _ij =p(j|i) is the transition from any image i to any image j in the sample data X Transition probability, select n images most similar to image i according to Euclidean distance, the feature of image j is x _j , the calculation formula of transition probability P(j|i) is:

Where d _ij ＝|| _xi -x _j || ² represents the Euclidean distance between image i and image j features.

The model W ^R enhanced by the regularization of the relationship matrix is calculated by the following formula:

W ^R =ηP+(1-η)ge ^T Wherein, η is the probability that image i transfers to image j this event takes place, (1-η) is the probability that image i jumps randomly, g=(1/n)e, Where g is a uniform random distribution vector, e is an n-dimensional unit column vector, n is the number of images of each image class, e=(1,1,...) ^T , the i-th row and j-th column of the matrix P is P( j|i);

计算公式为：The new relationship weight between image i and image j

The calculation formula is:

${\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; ij</span>}}^{<span\; class="notranslate">\; *</span>}<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; ij</span>}}^{<span\; class="notranslate">\; \&Prime;</span>}<span\; class="notranslate">\; \&CenterDot;</span>{\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; ij</span>}}^{\mathrm{<span\; class="notranslate">\; R</span>}}$

为图像i跳转到图像j的概率权值，

为W^R的第i行第j列的值；最终得到正则化增强关系矩阵W^*，W^*的第i行第j列为

步骤7中包括如下步骤：首先从样本数据X中选取任意两幅图像的特征x_i和x_j，两幅图像的关系权值为W_ij，两幅图像的正例关系权值为

两幅图像的反例关系权值为

根据以下目标方程计算得到广义特征矩阵A：X(L^N-γL^P)X^TA=λXLX^TA，w″ _ij is the weight of image i and image j, w″ _ij is the value of row i and column j of W″,

is the probability weight of image i jumping to image j,

is the value of the i-th row and j-column of W ^R ; the regularized enhanced relationship matrix W ^* is finally obtained, and the i-th row and j-column of W ^* is

Step 7 includes the following steps: first select the features x _i and x _j of any two images from the sample data X, the relationship weight of the two images is W _ij , and the positive relationship weight of the two images is

The weight of the negative relationship between the two images is

The generalized characteristic matrix A is calculated according to the following objective equation: X(L ^N -γL ^P )X ^T A=λXLX ^T A,

L is the Laplacian matrix of the relationship matrix W, L ^N is the Laplacian matrix of the negative relationship matrix W ^N , L ^P is the Laplacian matrix of the positive relationship matrix W ^P , and γ is the number of negative examples A constant proportional to the ratio of the number of positive images, X ^T represents the transpose matrix of the sample data X, and λ represents the eigenvalue of the equation solution.

Example 1

This embodiment includes the following parts:

1. Input an image I to be retrieved;

2. Extract the image features of the image instance library and the image to be retrieved. Each feature and its corresponding dimension are as follows:

Color moment (RGB color space): 9 dimensions; color moment (LUV color space): 9 dimensions; Tamura texture feature: 6 dimensions; Gabor texture feature: 24 dimensions; color histogram (HSV color space): 64 dimensions. In this way, each image will be described by a 112-dimensional vector, the image to be retrieved is v, the image instance feature library is U=(u ₁ ,...,u _M ), M is the total number of images in the image instance library, and U is N×M dimensions matrix;

3. Select training sample data from the image feature library U, each image is represented by extracted features, and 30 image categories are selected from it, each category represents a semantic category, each category has 100 images, a total of 3000 images, And take it as sample data X, X=(x ₁ ,...,x ₃₀₀₀ ), matrix X is 112×3000 dimensional;.

4. Randomly select an image in the sample data X, calculate the Euclidean distance between the image and other images in the sample data X, and use the relevant feedback retrieval technology to establish a set of positive examples and corresponding images of the same type and different types in the returned results. Counterexample set, and use the simple k-nearest neighbor method to establish a relationship matrix, that is, the weight between two images belonging to the k-nearest neighbor and the same image class is 1, otherwise it is 0.

In step 4, the embedded relationship widening ARE method based on feedback technology is used as the manifold learning algorithm of spectrogram theory, including the following steps:

(1) First construct a relationship matrix W for the sample data X, randomly select an image I from the sample data X, the feature of the image I is x _i , use the k nearest neighbor method to calculate the Euclidean relationship between x _i and other image features in the sample data X Distance, to obtain k images most similar to image I, where k takes a value of 5;

An image T is randomly selected from k images, and the feature of image T is x _t , then the weight W it between image I and image T is 1, and the weight W _it between image I and images other than k images The value is 0; that is, x _i ∈ ^{N k} (x _t ) or x _t ∈ ^{N k} ( _xi ), W _it = 1, where N ^k (xi ₎ represents the k-nearest neighbor set of image x _i , N ^k (x _t ) represents the k-nearest neighbor set of the image x _t ; the relationship matrix W is obtained, and the value of the i-th row and the t-column of the relationship matrix W is W _it ;

The formula is:

The images belonging to the same image class as image I in the k images are recorded as positive example set Pos, and the images of different image classes are recorded as negative example set Neg;

为图像I与图像R之间的权值，x_i,x_r∈Pos为表示特征x_i,x_r属于正例集合Pos，正例关系矩阵W^P的第i行第r列的值即为

公式为：(2) Construct a positive example relationship matrix W ^P , if image R and image I belong to the same image class and both belong to k images, and the feature of image R is x _r , then the weight between image I and image R is 1 , the weight between image I and images other than k images is 0; that is,

is the weight between the image I and the image R, x _i , x _r ∈ Pos is the representation feature x _i , x _r belongs to the positive example set Pos, the value of the i-th row and the r-th column of the positive example relationship matrix W ^P is

The formula is:

为图像I与图像H之间的权值，反例关系矩阵W^N的第i第h列为

公式为：(3) Construct a negative example relationship matrix W ^N , if image H and image I belong to different image categories and both belong to k images, and the feature of image H is x _h , then the weight between image I and image H is 1, and image The weight between I and images other than k images is 0; that is, x _i ∈ Pos and x _h ∈ Neg or x _h ∈ Pos and x _i ∈ pos means that feature x _i belongs to positive set Pos, x _h ∈ Neg means feature x _h belongs to negative set Neg, x _h ∈ Pos means feature x _h belongs to positive set Pos, x _i ∈ Neg means feature x _i belongs to Negative set Neg,

is the weight between image I and image H, and the ith column h of the counterexample relationship matrix W ^N is

The formula is:

Finally, three relationship matrices W, W ^P and W ^N are constructed, which are the relationship matrices needed to calculate the generalized feature matrix.

5. Establish the initial relationship enhancement matrix W′, starting from the relationship matrix W, if image z is the neighbor image of image i, and image z is also the neighbor image of image j, then use the following formula to calculate the enhanced relationship between image i and image j Weight W′ _ij :

W′ _ij ＝Σ _z W _iz W _jz

Where W _iz is the weight of image i and image z, W _jz is the weight of image j and image z, and W' _ij is the value of row i and column j of the enhanced relationship matrix W'. The specific example is shown in Figure 4. Image 3 is the neighbor image of image 1, and image 3 is the neighbor image of image 2. The solid lines with arrows are connected between the images to represent the neighbor relationship, and the image 1 and image 2 are connected by dotted lines. Represents the relationship between image 1 and image 2 needs to be enhanced.

6. Construct the probability transition matrix W ^R and regularize the enhanced relation matrix W′,

The neighbor relationship between images is propagated multiple times to obtain a new enhanced relationship matrix W", the formula is w"=w'*w';

Use the transition probability matrix to represent the transition relationship between images, the corresponding transition matrix is P=[P _ij ] _n×n , P _ij =P(j|i) is the transition from any image i to any image j in the sample data X Transition probability, select n images most similar to image i according to Euclidean distance, the feature of image j is x _j , the calculation formula of transition probability P(j|i) is:

Where d _ij ＝|| _xi -x _j || ² represents the Euclidean distance between image i and image j features.

The model W ^R enhanced by the regularization of the relationship matrix is calculated by the following formula:

W ^R ＝ηP+(1-η)ge ^T

Among them, η is the probability of image i transferring to image j, η is taken as 0.85, (1-η) is the probability of image i jumping randomly, g=(1/n)e, where g is a uniform random Distribution vector, e is an n-dimensional unit column vector, n is the number of images of each image class, e=(1,1,...) ^T , the i-th row and j-th column of the matrix P is P(j|i);

计算公式为：The new relationship weight between image i and image j

The calculation formula is:

${\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; ij</span>}}^{<span\; class="notranslate">\; *</span>}<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; ij</span>}}^{<span\; class="notranslate">\; \&Prime;</span>}<span\; class="notranslate">\; \&Center\; Dot;</span>{\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; ij</span>}}^{\mathrm{<span\; class="notranslate">\; R</span>}}$

为图像i跳转到图像j的概率权值，为W^R的第i行第j列的值；w″ _ij is the weight of image i and image j, w″ _ij is the value of row i and column j of W″,

is the probability weight of image i jumping to image j, is the value of row i and column j of W ^R ;

具体实例如图5～6所示，图5代表图像间的概率转移权值关系，图6中左上图1表示图像间的增强关系矩阵W″，两幅图像间用实线连接的是近邻图像，虚线连接代表是两幅图像之间的增强关系，右上图2表示图像间的转移概率矩阵W^R，图像间用实现连接代表图像间存在转移关系，下图3代表图像间的正则化增强关系矩阵W^*，由W″和W^R相乘得到；Finally, the regularized enhanced relationship matrix W ^* is obtained, and the i-th row and j-th column of W ^* are

Specific examples are shown in Figures 5-6. Figure 5 represents the probability transfer weight relationship between images. Figure 1 on the upper left of Figure 6 represents the enhanced relationship matrix W" between images. The adjacent images are connected by solid lines between the two images. , the dotted line connection represents the enhanced relationship between the two images, the upper right figure 2 represents the transition probability matrix W ^R between the images, and the realized connection between the images represents the transfer relationship between the images, and the following figure 3 represents the regularized enhanced relationship between the images Matrix W ^* , obtained by multiplying W″ and W ^R ;

7. Construct the objective function according to the regularized relationship enhancement matrix W ^* , and solve the generalized feature matrix A,

两幅图像的反例关系权值为

根据以下目标方程计算得到广义特征矩阵A：First select the features x _i and x _j of any two images from the sample data X, the relationship weight of the two images is W _ij , and the positive relationship weight of the two images is

The weight of the negative relationship between the two images is

The generalized characteristic matrix A is calculated according to the following objective equation:

X(L ^N -γL ^P )X ^T A=λXLX ^T A,

L is the Laplacian matrix of the relationship matrix W, L ^N is the Laplacian matrix of the negative relationship matrix W ^N , L ^P is the Laplacian matrix of the positive relationship matrix W ^P , and γ is the number of negative examples A constant proportional to the ratio of the number of positive images, X ^T represents the transpose matrix of the sample data X, and λ represents the eigenvalue of the equation solution.

8. Mainly use the generalized feature matrix A to reduce the dimensionality of the image data in the image instance feature library to obtain the final image representation, that is, AU=A*(u ₁ ,…,u _M )=(A*u _i ,…,A* u _M ), record y _i =A*xi _i , i=1,...M, and the final image is expressed as Y=(y ₁ ,...,y _M );

9. Mainly use the generalized feature matrix A to reduce the dimensionality of the feature v of the image to be retrieved, and obtain the image representation f of the image to be retrieved, f=A*v;

10. Calculate the similarity between the image to be retrieved and the image in the image instance database:

Use the Euclidean distance to calculate the similarity between the image to be retrieved and all the images in ^the image instance library, that is, calculate ||fy _i || ² , i=1,...M, the smaller ||fy _i || The similarity is from large to small to output the image most similar to the image to be retrieved in the image instance library. As shown in Figure 7, the similarity between the image to be retrieved and all images in the image instance database is calculated according to the Euclidean distance, and the 4 most similar images are output according to the similarity from large to small.

Example 2

Fig. 1 is the retrieval flowchart of embodiment 2, and image source among the figure is public Corel5k database. Figure 2 is to preprocess the original image, using color moment, Tamura texture feature, Gabor texture feature and color histogram to represent an image, Figure 3 selects feature samples, selects 30 image classes from the image instance library, and each One class represents a semantic class, each class has 100 images, a total of 3000 images, in order to improve the calculation speed, only the first 400 images in the result set are used as the global data set, which is used to establish the relationship matrix W, the positive relationship Matrix W ^P , counterexample relationship matrix W ^N . Then the relationship matrix W is enhanced to obtain W′, and the probability transition matrix W ^R is used to regularize the enhanced relationship matrix to obtain W ^* , and then the generalized characteristic matrix A of the objective function is solved according to the regularized enhanced relationship matrix W ^* , and finally the generalized The feature matrix A reduces the dimensionality of the image features in the image instance database and the image features to be retrieved, retrieves the image to be retrieved, and uses the Euclidean distance to calculate the similarity between the image to be retrieved and the image in the image instance database, and the similarity is from large to small Output the image most similar to the image to be retrieved in the image instance database.

The present invention provides an image retrieval method represented by a regularized enhanced relational matrix. There are many methods and approaches to specifically realize the technical solution. The above description is only a preferred embodiment of the present invention. As far as people are concerned, some improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be regarded as the protection scope of the present invention. All components that are not specified in this embodiment can be realized by existing technologies.

Claims (7)

1. the image search method based on relational matrix regularization Enhancement Method, is characterized in that, the method is retrieving images from the image case library, comprises following steps:

Step 1, input image to be retrieved;

Step 2, extract the feature of image in image to be retrieved and image case library, with the N dimensional vector, every width image is described, N=112, obtain image example aspects storehouse and with the feature of retrieving images, described image case library comprises the image class more than 50, and each image class means a semantic category, and each image class comprises the image that 600 width are above;

Step 3 is chosen P image class from image example aspects storehouse, and P span 20～50, choose n width image from each image class, n span 100～500, and P the total n * P of image class opens image construction sample data X; Step 4, the manifold learning arithmetic based on spectral graph theory, build and strengthen relational matrix W, positive example relational matrix W sample data X ^pwith counter-example relational matrix W ⁿ; Step 5, strengthened the relational matrix W built, and preliminary foundation strengthens relational matrix W '; Step 6, strengthen relational matrix W ' by the probability transfer matrix regularization and obtain regularization enhancing relational matrix W ^*step 7, strengthen relational matrix W according to regularization ^*the establishing target equation, calculate generalized characteristic matrix A;

Step 8, utilize generalized characteristic matrix A to carry out dimensionality reduction to all images in image example aspects storehouse, obtains final image representation;

Step 9, utilize generalized characteristic matrix A to treat the retrieving images dimensionality reduction, obtains the image representation of image to be retrieved;

Step 10, calculate the similarity of all images in image to be retrieved and image case library according to the Euclidean distance of the image representation of the image to be retrieved of the final image representation of step 8 and step 9, according to the image the most similar to image to be retrieved in the descending output image case library of similarity.

2. a kind of image search method meaned that strengthens based on the relational matrix regularization according to claim 1, is characterized in that, in step 2, characteristics of image comprises color moment, Tamura textural characteristics, Gabor textural characteristics, color histogram.

3. a kind of image search method strengthen meaned based on the relational matrix regularization according to claim 2, it is characterized in that, step 4 specifically comprises the steps: to choose at random piece image in sample data X, calculate the Euclidean distance of other images in this image and sample data X, utilize the relevance feedback retrieval technology, set up positive example set and counter-example set according to the similar image in returning results and inhomogeneity image correspondence, and adopt simple k near neighbor method opening relationships matrix, belong to the k neighbour and be that weights between two images of same image class are 1, otherwise be 0.

4. a kind of image search method strengthen meaned based on the relational matrix regularization according to claim 3, it is characterized in that, in step 4, adopt the imbeding relation based on feedback technique to widen the manifold learning arithmetic of ARE method as spectral graph theory, comprise the following steps:

(1) at first sample data X is built to relational matrix W, randomly draw piece image I from sample data X, image I be characterized as x _i, adopt the k near neighbor method to calculate x _iwith the Euclidean distance of other characteristics of image in sample data X, obtain the k width image the most similar to image I, wherein the k span 5～10;

Take out arbitrarily piece image T and belong to from k width image, image T is characterized as x _t, the weights W between image I and image T _itbe 1, the weights between the image beyond image I and k width image are 0; Be x _i∈ N ^k(x _t) or x _t∈ N ^k(x _i), W _it=1, N wherein ^k(x _i) presentation video x _ik neighbour set, N ^k(x _t) presentation video x _tk neighbour set; Obtain relational matrix W, the value of the capable t row of relational matrix W i is W _it;

The image that belongs to same image class with image I in k width image is designated as to positive example set Pos, and the image of different images class is designated as counter-example set Neg;

(2) build positive example relational matrix W ^pif image R and image I belong to same image class and all belong to k width image, and image R be characterized as x _r, the weights between image I and image R are 1, the weights between the image beyond image I and k width image are 0; That is,

for the weights between image I and image R, x _i, x _r∈ Pos is representation feature xi, x _rbelong to positive example set Pos, positive example relational matrix W ^pthe value of the capable r of i row be

formula is:

(3) build counter-example relational matrix W ⁿif image H and image I belong to the different images class and all belong to k width image, image H is characterized as x _h, the weights between image I and image H are 1, the weights between the image beyond image I and k width image are 0; Be x _i∈ Posand x _h∈ negor x _h∈ Posand x _i∈ neg,

x _i∈ pos representation feature x _ibelonging to positive example set Pos is, x _h∈ Neg representation feature x _hbelong to counter-example set Neg, x _h∈ Pos representation feature x _hbelong to positive example set Pos, x _i∈ Neg representation feature x _ibelonging to counter-example set Neg is,

for the weights between image I and image H, counter-example relational matrix W ⁿi h classify as

formula is:

finally build and obtain three relational matrix W, W ^pand W ⁿ, the relational matrix that its neutralization need to be used for calculating generalized characteristic matrix.

5. a kind of image search method strengthen meaned based on the relational matrix regularization according to claim 4, it is characterized in that, step 5 specifically comprises the 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, adopt following formula to calculate the weights W strengthened between image i and image j ^' _ij: W ^' _ij∑ _zw _izw _jz, W wherein _izfor the weights of image i and image z, W _jzfor the weights of image j and image z, W ' _ijbe the capable j train value of i that strengthens relational matrix W '.

6. a kind of image search method meaned that strengthens based on the relational matrix regularization according to claim 5, is characterized in that, step 6 specifically comprises the steps:

The neighbor relationships of repeatedly propagating between image obtains new enhancing relational matrix W ", formula is W "=W ' * W ';

Utilize the transfer relationship between the transition probability matrix presentation video, corresponding transition matrix is P=[P _ij] _{n * n}, P _ij=p (j|i) is that in sample data X, arbitrary image i, to the transition probability of arbitrary image j, selects the n width image the most similar to image i according to Euclidean distance, and image j is characterized as x _j, the computing formula of transition probability P (j|i) is:

D wherein _ij=|| x _i-x _j|| ², the Euclidean distance of presentation video i and image j feature;

The model W that adopts the regularization of following formula calculated relationship matrix to strengthen ^r:

W ^R＝ηP+(1-η)ge ^T

Wherein, η is that image i transfers to the probability that this event of image j occurs, and (1-η) is the probability of the random redirect of image i, g=(1/n) e, wherein g is an even stochastic distribution vector, and e is n dimension unit column vector, and n is the picture number of each image class, e=(1,1 ...) ^t, the capable j of i of matrix P classifies P (j|i) as;

The new weights that concern between image i and image j

computing formula is:

$w_{\mathrm{ij}}^{*}=w_{\mathrm{ij}}^{\&Prime;}\&CenterDot;w_{\mathrm{ij}}^R$

W " _ijfor the weights of image i and image j, w " _ijfor W " the value of the capable j of i row,

for image i jumps to the probability weights of image j, for W ^rthe value of the capable j of i row;

Finally obtain regularization and strengthen relational matrix W ^*, W ^*the capable j of i classify as

7. a kind of image search method meaned that strengthens based on the relational matrix regularization according to claim 6, is characterized in that, in step 7, comprises the steps:

At first choose the feature x of any two width images from sample data X _iand x _j, the weights that concern of two width images are W _ij, the positive example of two width images concerns that weights are

the counter-example of two width images concerns that weights are

calculate generalized characteristic matrix A according to following target equation:

X(L ^N-γL ^P)X ^TA＝λXLX ^TA

The Laplacian Matrix that L is relational matrix W, L ⁿfor counter-example relational matrix W ⁿlaplacian Matrix, L ^pfor positive example relational matrix W ^plaplacian Matrix, γ is the constant that is directly proportional to the ratio of counter-example image number and positive example image number, X ^tmean the transposed matrix of sample data X, λ means the eigenwert of equation solution.

Images