A method and system for feature selection of multi-label image data using data reconstruction

Abstract

The present invention relates to the field of image processing technology, and discloses a method and system for feature selection of multi-label image data using data reconstruction, the method comprising the following steps: extracting feature space and label space of multi-label images; reconstructing the feature space using low-rank representation to obtain a reconstructed feature space; transforming the label space using a directional traction matrix to obtain a reconstructed label space; projecting the reconstructed feature space to the reconstructed label space to generate a feature selection matrix, applying L2,1 norm constraints to the sparse regularization of the feature selection matrix to obtain an objective function; obtaining the feature selection matrix by solving the objective function to obtain an optimal feature subset. The present invention reconstructs the feature space using a dynamic feature relationship graph represented by a low-rank representation, introduces a directional traction matrix to perturb the original binary label space to reconstruct the label space, formulates an objective function for feature selection of multi-label image data, and introduces an effective optimization strategy to process the function.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to a method for selecting significant features of multi-label image data by using data reconstruction. Background Art

[0002] In the era of big data, application practice is no longer satisfied with identifying an instance from a single label, because the coexistence of multiple label information can more richly characterize the instance. For example, in a multi-label natural image, it can contain multiple label information: sky, building, tree, lake and pedestrian. As the number of class labels in multi-label classification tasks increases, more features may need to be collected to learn these class labels. However, multi-label data has the disaster of high-dimensional feature space, which significantly increases the risk of overfitting and training time of learning models. Faced with the dimensionality disaster of multi-label image data, we usually resort to two mainstream dimensionality reduction techniques, namely feature extraction and feature selection. These two techniques can help reduce the cost of computing time and storage space, and lay the foundation for improving the accuracy of multi-label learning models. Feature extraction constructs a set of new features whose meanings are not yet clear. Since users cannot obtain the physical meaning of these features, it is difficult to conduct further qualitative analysis on this basis. The goal of feature selection is to eliminate some features, including irrelevant features, redundant features and noise features. In other words, feature selection retains a part of the original features for dimensionality reduction, thereby fully retaining the physical meaning of these features. These features can be used to build a more readable and interpretable model, and can also be used to analyze the feature relationships in the data. Therefore, feature selection is a very important preprocessing before multi-label learning.

[0003] Feature selection methods for multi-label image data usually mine the correlation between features and labels to evaluate the importance of features, but ignore the inappropriate representation mechanism in the data, that is, the initial representation of the feature space and the label space. In the feature space of the original image data, both weak correlation and strong correlation exist. Compared with strong correlation, weak correlation may be a kind of noise information, which interferes with the accurate representation of the original feature space. Therefore, highlighting the expression of strong correlation and suppressing the interference of weak correlation are important issues ignored by existing work. In the label space of multi-label image data, appropriate labels in the label space are assigned through the information in the feature space. However, labels represented by "0" or "1" are logical labels and cannot describe the importance of different labels. Therefore, a method is needed to reconstruct the importance of labels in the logical label space. Summary of the invention

[0004] The object of the present invention is to solve the problems in the prior art.

[0005] The technical solution adopted by the present invention to solve the technical problem is: to provide a method for selecting features of multi-label image data using data reconstruction, comprising the following steps:

[0006] Extract feature space and label space of multi-label images;

[0007] The feature space is reconstructed using low-rank representation to obtain a reconstructed feature space;

[0008] The marker space is transformed by the directional traction matrix to obtain the reconstructed marker space;

[0009] Project the reconstructed feature space to the reconstructed label space to generate a feature selection matrix, and impose an L2,1 norm constraint on the sparse regularization of the feature selection matrix to obtain the objective function;

[0010] The feature selection matrix is ​​obtained by solving the objective function;

[0011] The optimal feature subset is obtained based on the feature selection matrix.

[0012] Preferably, the reconstructing the feature space by using the low-rank representation comprises the following steps:

[0013] The low-rank representation is used for multi-label image data, expressed as:

[0014] ;

[0015] Among them, X is the feature space of multi-label images, Z is the affinity graph to be learned, A is the dictionary, is a trade-off parameter;

[0016] The dictionary A is set to the multi-label image data itself X, and the kernel function is used to extract low-rank information. The low-rank model of the feature relationship is expressed as:

[0017] ;

[0018] Based on the low-rank model, the core feature relationship is extracted to obtain the reconstructed feature space XZ.

[0019] Preferably, the converting of the marking space by the directional pulling matrix comprises the following steps:

[0020] Design a directional traction matrix , expressed as:

[0021] ;

[0022] Among them, each element Represents the relaxation direction; a value of "1" indicates the positive relaxation direction, and a value of "-1" indicates the negative relaxation direction;

[0023] Randomly initialize the non-negative labeled relaxation matrix ;

[0024] Construct the relaxed labeling space as the new labeling space V, expressed as:

[0025] ;

[0026] Where Y represents the label space of multi-label images; Represents the Hadamard product of two matrices.

[0027] Preferably, projecting the reconstructed feature space into the reconstructed label space to generate a feature selection matrix comprises the following steps:

[0028] The mapping projection is designed based on the reconstructed feature space XZ and label space V, which is expressed as:

[0029] ;

[0030] Generate feature selection matrix W based on projection mapping, is a trade-off parameter.

[0031] Preferably, the sparse regularization of the feature selection matrix imposes an L2,1 norm constraint to obtain the objective function , expressed as:

[0032] ;

[0033] in, , and is a trade-off parameter.

[0034] Preferably, the feature selection matrix W is obtained by solving the objective function, Z, W and H are solved by alternating iterative optimization, and the feature selection matrix W obtained after the iteration stops is the final feature selection matrix W; the alternating iterative optimization comprises the following steps:

[0035] Assumptions ,rewrite for:

[0036] ;

[0037] where D is a diagonal matrix and its diagonal elements are ;

[0038] Fix other variables and update Z, remove items irrelevant to Z, and simplify And after the reduction The Z derivative of is expressed as:

[0039] ;

[0040] ;

[0041] in, , and E is a positive definite matrix; according to the singular value decomposition, ;

[0042] Fix other variables and update W, remove items irrelevant to W, and simplify And after the reduction The derivative of W in is expressed as:

[0043] ;

[0044] ;

[0045] Fix other variables and update H, remove items irrelevant to H, simplify And solve the optimal solution of H, expressed as:

[0046] ;

[0047] .

[0048] Preferably, the optimal feature subset is obtained based on the feature selection matrix, specifically: based on the optimal solution H, the feature selection matrix W is obtained, and the feature subset is selected according to Select several features with the highest ranking in descending order as feature subsets; the number of feature subsets is set according to actual needs.

[0049] The present invention also provides a multi-label image data feature selection system using data reconstruction, comprising:

[0050] Extraction module, extracting feature space and label space of multi-label images;

[0051] The first reconstruction module reconstructs the feature space using a low-rank representation to obtain a reconstructed feature space;

[0052] The second reconstruction module transforms the label space through the direction pulling matrix to obtain the reconstructed label space;

[0053] The objective function building module projects the reconstructed feature space into the reconstructed label space to produce a feature selection matrix, and applies an L2,1 norm constraint to the sparse regularization of the feature selection matrix to obtain the objective function;

[0054] The optimization module obtains the feature selection matrix by solving the objective function;

[0055] The selection module obtains the optimal feature subset based on the feature selection matrix.

[0056] The present invention has the following beneficial effects:

[0057] (1) Reconstruct the feature space using the dynamic feature relationship graph represented by low rank to obtain a more accurate feature space;

[0058] (2) A directional traction matrix is ​​introduced to perturb the original binary tag space and reconstruct the tag space, which can adaptively characterize the importance of each tag to the example;

[0059] (3) An objective function suitable for feature selection of multi-labeled images is proposed, and an effective optimization strategy is introduced to solve this function. The proposed method is converted into an objective function, and the objective function is solved through a designed optimization scheme to quickly converge to the final solution.

[0060] The present invention is further described in detail below in conjunction with the accompanying drawings and embodiments, but the present invention is not limited to the embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 A method step diagram of an embodiment of the present invention;

[0062] Figure 2 A schematic diagram of a flow chart of an embodiment of the present invention;

[0063] Figure 3 4 is a system structure diagram of an embodiment of the present invention. DETAILED DESCRIPTION

[0064] See also Figure 1 and Figure 2 The method step diagram and flow chart of an embodiment of the present invention are shown, which include the following steps:

[0065] S101, extracting feature space and label space of multi-labeled images;

[0066] S102, reconstructing the feature space by using a low-rank representation to obtain a reconstructed feature space;

[0067] S103, transforming the label space through a directional pulling matrix to obtain a reconstructed label space;

[0068] S104, projecting the reconstructed feature space into the reconstructed label space to generate a feature selection matrix, and applying an L2,1 norm constraint to the sparse regularization of the feature selection matrix to obtain an objective function;

[0069] S105, obtaining a feature selection matrix by solving the objective function;

[0070] S106, obtaining an optimal feature subset based on the feature selection matrix.

[0071] Specifically, in S101, the multi-label image data is formally composed of a feature space and a label space, so given is a feature space with d-dimensional features, and is a label space with l class labels. Where n represents the number of images in the multi-label image data, and is the feature description of the i-th image. If the i-th image is associated with the jth tag, then ,on the contrary, For any matrix , the trace of U is expressed as , the transpose of U is expressed as , the Frobenius norm of U is defined as , and the L2,1 norm of U is defined as ,in Represents the element in the i-th row and j-th column of matrix U.

[0072] Specifically, the S102 includes the following steps:

[0073] S2-1. Use low-rank representation for multi-labeled image data to better preserve the structural information in the data:

[0074] (1)

[0075] Among them, X is the multi-label image data, Z is the affinity graph to be learned, A is the dictionary, and are trade-off parameters.

[0076] S2-2. In order to learn the affinity graph of feature Z, the dictionary A is set to the multi-label image data itself X, and the kernel function is used to extract low-rank information. Therefore, the low-rank model of the feature relationship is expressed as:

[0077] (2)

[0078] Based on formula (2), the core feature relationship can be extracted. Therefore, the final reconstructed feature space is XZ.

[0079] Specifically, the S103 includes the following steps:

[0080] S3-1. Design a directional traction matrix To ensure the relaxation direction, and Defined as:

[0081] (3)

[0082] Among them, each element Represents the direction of relaxation. In other words, a value of "1" indicates a positive direction of relaxation, while a value of "-1" indicates a negative direction of relaxation.

[0083] S3-2. Randomly initialize the non-negative labeled relaxation matrix .

[0084] S3-3. Construct the relaxed labeling space V:

[0085] (4)

[0086] To illustrate this operation with an example, suppose:

[0087] (5)

[0088] (6)

[0089] Then we can get:

[0090] (7)

[0091] Specifically, the S104 includes the following steps:

[0092] S4-1. Based on the reconstructed feature space XZ and label space V, a mapping projection between the feature space and the label space is constructed to generate a feature selection matrix W. In addition, an L2,1 norm constraint is imposed on the feature selection matrix to enhance the robustness of the feature selection matrix to outliers and noise. Therefore, the mapping projection is designed as:

[0093] (8)

[0094] S4-2. Based on the above steps, we integrate formulas (2), (4) and (8) to form the overall objective function:

[0095] (9)

[0096] in, , and is a trade-off parameter.

[0097] Specifically, the S105 includes the following steps:

[0098] S5-1. Use alternating iterative optimization to solve these variable matrices, assuming , is rewritten as:

[0099] (10)

[0100] where D is a diagonal matrix and its diagonal elements are .

[0101] S5-2. Update Z by fixing other variables: remove items that are irrelevant to Z, so It can be simplified to:

[0102] (11)

[0103] According to the singular value decomposition, Then, taking the derivative of Z in formula (11), we can obtain:

[0104] (12)

[0105] in, , and E is a positive definite matrix.

[0106] S5-3. Fix other variables and update W: remove items that are irrelevant to W, so It can be simplified to:

[0107] (13)

[0108] Taking the derivative of W in formula (13), we can get:

[0109] (14)

[0110] S5-4. Fix other variables and update H: Simplify the terms that are irrelevant to H, and we can get:

[0111] (15)

[0112] Based on this, the optimal solution of H is:

[0113] (17)

[0114] Specifically, S106 calculates the feature selection matrix W according to The top 20% features are selected in descending order as the feature subset.

[0115] Specifically, the quality of the selected features is tested by the MLKNN classifier, and the average performance of the multi-label image dataset Scene on two indicators (Hamming loss and Micro-F1) is recorded through five-fold cross validation; a multi-emotional music dataset is used, and its characteristics are shown in Table 1; the test results are shown in Table 2, where RDMFS is a method of an embodiment of the present invention.

[0116] Table 1 Detailed characteristics of the multi-emotional music dataset:

[0117]

[0118] Table 2 Performance of MFSLL with 5 comparison methods under MLKNN classifier on two indicators:

[0119]

[0120] The lower the Hamming loss value, the better the performance, and the higher the Micro-F1 value, the better the performance. As shown in Table 1 and Table 2, the Hamming loss of LRMFS on the Scene dataset is 0.089, which is better than all the compared methods. On the Arts dataset, the Micro-F1 of LRMFS is 0.726, which is the highest performance among all the compared methods.

[0121] It can be seen that the present invention uses a dynamic feature relationship graph represented by a low rank to reconstruct the feature space, introduces a directional traction matrix to perturb the original binary label space to reconstruct the label space, formulates the objective function of feature selection for multi-label image data, and introduces an effective optimization strategy to process the function.

[0122] See also Figure 3 FIG. 1 is a system structure diagram of an embodiment of the present invention, including:

[0123] Extraction module 301, extracting feature space and label space of multi-labeled images;

[0124] A first reconstruction module 302 reconstructs the feature space using a low-rank representation to obtain a reconstructed feature space;

[0125] The second reconstruction module 303 transforms the label space through the direction pulling matrix to obtain a reconstructed label space;

[0126] An objective function construction module 304 projects the reconstructed feature space into the reconstructed label space to generate a feature selection matrix, and applies an L2,1 norm constraint to the sparse regularization of the feature selection matrix to obtain an objective function;

[0127] The optimization module 305 obtains a feature selection matrix by solving the objective function;

[0128] The selection module 306 obtains the optimal feature subset based on the feature selection matrix.

[0129] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for feature selection of multi-label image data using data reconstruction, characterized in that: The following steps are involved: Extract feature space and label space of multi-label images; The feature space is reconstructed using low-rank representation to obtain a reconstructed feature space; The marker space is transformed by the directional traction matrix to obtain the reconstructed marker space; Project the reconstructed feature space to the reconstructed label space to generate a feature selection matrix, and impose an L2,1 norm constraint on the sparse regularization of the feature selection matrix to obtain the objective function; The feature selection matrix is ​​obtained by solving the objective function; Get the optimal feature subset based on the feature selection matrix; The step of reconstructing the feature space by using the low-rank representation comprises the following steps: The low-rank representation is used for multi-label image data, expressed as: Among them, X is the feature space of multi-labeled images, Z is the affinity graph to be learned, A is the dictionary, and α is the trade-off parameter; The dictionary A is set to the multi-label image data itself X, and the kernel function is used to extract low-rank information. The low-rank model of the feature relationship is expressed as: Based on the low-rank model, the core feature relationship is extracted to obtain the reconstructed feature space XZ; The projecting of the reconstructed feature space into the reconstructed label space to generate a feature selection matrix comprises the following steps: The mapping projection is designed based on the reconstructed feature space XZ and label space V, which is expressed as: Generate feature selection matrix W based on mapping projection, γ is the trade-off parameter; The sparse regularization of the feature selection matrix imposes an L2,1 norm constraint to obtain the objective function Θ, expressed as: Among them, α, β, and γ are trade-off parameters; The feature selection matrix W is obtained by solving the objective function, and Z, W and H are solved by alternating iterative optimization. The feature selection matrix W obtained after the iteration stops is the final feature selection matrix W; the alternating iterative optimization includes the following steps: Assume W = [w1,w1,...,w d ] T , rewrite Θ as: where D is a diagonal matrix and its diagonal elements are Fix other variables and update Z, remove items irrelevant to Z, simplify Θ and take the derivative of Z in the simplified Θ, expressed as: Where A = X T X, C = βWW T +E, E is a positive definite matrix; according to the singular value decomposition, Z = U∑P T ; Fix other variables and update W, remove items irrelevant to W, simplify Θ and take the derivative of W in the simplified Θ, which can be expressed as: Fix other variables and update H, remove items irrelevant to H, simplify Θ and solve to obtain the optimal solution of H, which is expressed as:

2. The method for selecting features of multi-label image data using data reconstruction according to claim 1, characterized in that: The optimal feature subset is obtained based on the feature selection matrix, specifically: based on the optimal solution H, the feature selection matrix W is obtained, and the feature subset is selected according to ||w i ||2(i=1→d) selects several top-ranked features in descending order as feature subsets; The number of feature subsets is set according to actual needs.

3. The method for selecting features of multi-label image data using data reconstruction according to claim 1, characterized in that: The step of transforming the marking space by using the directional pulling matrix comprises the following steps: Design a directional traction matrix It is expressed as: Among them, each element B ij Represents the relaxation direction; a value of "1" indicates a positive relaxation direction, and a value of "-1" indicates a negative relaxation direction; Randomly initialize the non-negative labeled relaxation matrix Construct the relaxed labeling space as the new labeling space V, expressed as: Where Y represents the label space of multi-label images; Represents the Hadamard product of two matrices.

4. A multi-label image data feature selection system using data reconstruction, characterized in that: include: Extraction module, extracting feature space and label space of multi-label images; The first reconstruction module reconstructs the feature space using a low-rank representation to obtain a reconstructed feature space; The second reconstruction module transforms the label space through the direction pulling matrix to obtain the reconstructed label space; The objective function building module projects the reconstructed feature space into the reconstructed label space to produce a feature selection matrix, and applies an L2,1 norm constraint to the sparse regularization of the feature selection matrix to obtain the objective function; The optimization module obtains the feature selection matrix by solving the objective function; Selection module, which obtains the optimal feature subset based on the feature selection matrix; The step of reconstructing the feature space by using the low-rank representation comprises the following steps: The low-rank representation is used for multi-label image data, expressed as: Among them, X is the feature space of multi-labeled images, Z is the affinity graph to be learned, A is the dictionary, and α is the trade-off parameter; The dictionary A is set to the multi-label image data itself X, and the kernel function is used to extract low-rank information. The low-rank model of the feature relationship is expressed as: Based on the low-rank model, the core feature relationship is extracted to obtain the reconstructed feature space XZ; The projecting of the reconstructed feature space into the reconstructed label space to generate a feature selection matrix comprises the following steps: The mapping projection is designed based on the reconstructed feature space XZ and label space V, which is expressed as: Generate feature selection matrix W based on mapping projection, γ is the trade-off parameter; The sparse regularization of the feature selection matrix imposes an L2,1 norm constraint to obtain the objective function Θ, expressed as: Among them, α, β, and γ are trade-off parameters; The feature selection matrix W is obtained by solving the objective function, and Z, W and H are solved by alternating iterative optimization. The feature selection matrix W obtained after the iteration stops is the final feature selection matrix W; the alternating iterative optimization includes the following steps: Assume W = [w1,w1,...,w d ] T , rewrite Θ as: where D is a diagonal matrix and its diagonal elements are Fix other variables and update Z, remove items irrelevant to Z, simplify Θ and take the derivative of Z in the simplified Θ, expressed as: Where A = X T X, C = βWW T +E, E is a positive definite matrix; according to the singular value decomposition, Z = U∑P T ; Fix other variables and update W, remove items irrelevant to W, simplify Θ and take the derivative of W in the simplified Θ, which can be expressed as: Fix other variables and update H, remove items irrelevant to H, simplify Θ and solve to obtain the optimal solution of H, which is expressed as:

Images