Polysemy object metric learning method and system based on implicit mark importance recovery and utilization
By restoring the implicit marking importance of polysensible objects and constraining them in discriminative measurement learning, the problem of marking important hypothesis in polysensible object measurement learning in the prior art is solved, and the portrayal of similarity of polysensible objects is significantly improved. ability.
Patent Information
- Application Number
- CN202411415777.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-11
- Publication Date
- 2025-01-17
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
The existing polysensory object metric learning method assumes that each associated marker has the same marker importance and cannot effectively characterize the similarity of polysensory objects in the real world, resulting in limited ability to characterize the similarity of polysensory objects.
By locally linearly reconstructing the latent manifold structure of feature space, the implicit mark importance of polysense objects is restored, and the prediction output of the constrained discriminant model is aligned with the implicit mark importance in discriminant measurement learning, and the solution is obtained using iterative optimization strategy Discriminant model coefficients and discriminant measures, explicitly portray the similarity of polysensical objects.
It significantly improves the ability of metric learning to characterize the similarity of polysense objects, and can more accurately draw samples with similar marker importance closer and samples with large differences in marker importance farther away.
Smart Images
Figure CN119322940A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of machine learning and data mining, and in particular relates to a polysemous object metric learning method and system based on implicit label importance recovery and utilization. Background Art
[0002] As auxiliary knowledge to indicate the semantic information of objects, the similarity between objects plays a vital role in human cognition and the recognition of intelligent systems. For specific tasks, accurate and appropriate measurement of the similarity between related objects can often significantly improve the performance of many machine learning algorithms such as k-nearest neighbors and k-means. As a solution to the problem of how to measure the similarity between objects, metric learning adaptively learns the similarity measure between objects by mining the association and comparison relationship between data. The similarity measure learned by metric learning can align the data distribution with its essential semantics, that is, similar objects of the same type are closer, and dissimilar objects of different types are farther away. Thanks to the powerful similarity characterization ability of metric learning, it has been widely used in real-world application fields such as face recognition, pedestrian re-identification, information retrieval, and recommendation systems in recent years.
[0003] Although metric learning has achieved great success, most methods only focus on a single semantic scene and aim to characterize the similarity of a single semantic object. When faced with polysemantic objects, which exist in more common forms, these monosemantic object metric learning methods will no longer be applicable due to the complexity of the semantic form of polysemantic objects. Therefore, how to measure the similarity of polysemantic objects has become a new hot topic in metric learning research in recent years, and many polysemantic object metric learning methods have been proposed to characterize the similarity between polysemantic objects.
[0004] It is worth noting that the label information of the training samples of polysemous objects is clear, that is, for a specific polysemous object example, each label in the label space is either associated with the example or not associated with the example. Therefore, the current research on polysemous object metric learning inevitably adopts the assumption of equal label importance during the learning process, that is, each associated label of the example has the same label importance. However, for polysemous objects in the real world, each associated label should essentially have a different label importance. For example, elements such as mountains, water, and sky in a landscape photo often convey different scenery emphasis information due to different presentation areas. Therefore, the basic assumptions adopted by the existing polysemous object metric learning are contrary to the facts, which leads to its ability to severely limit the similarity of polysemous objects. Summary of the invention
[0005] The present invention is aimed at the problems existing in the prior art, and provides a method and system for learning a metric of a polysemous object based on implicit label importance recovery, which includes at least three stages: implicit label importance recovery, discriminant metric learning and similarity characterization. In the implicit label importance recovery stage, the implicit label importance of the polysemous object is restored by locally linearly reconstructing the feature space to model its potential manifold structure, and the manifold structure is transferred to the label space. In the discriminant metric learning stage, a discriminant metric learning objective function is constructed to constrain the predicted output of the discriminant model to be aligned with the restored implicit label importance, and the discriminant model coefficients and discriminant metrics are solved through an iterative optimization strategy, so that samples with similar label importance are implicitly brought closer to each other, and samples with large differences in label importance are pushed away from each other. In the similarity characterization stage, the discriminant model coefficients and discriminant metrics obtained by the solution are used to explicitly characterize the similarity of polysemous objects. The method of the present invention further improves the ability of metric learning to characterize the similarity of polysemous objects.
[0006] In order to achieve the above object, the technical solution adopted by the present invention is: a polysemous object metric learning method based on implicit label importance recovery and utilization, comprising at least the following steps:
[0007] S1, implicit tag importance recovery: by locally linearly reconstructing the feature space to model its potential manifold structure, the manifold structure is transferred to the tag space to recover the implicit tag importance of the polysemous object;
[0008] S2, discriminant metric learning: construct a discriminant metric learning objective function, constrain the prediction output of the discriminant model to be aligned with the recovered implicit label importance, solve the discriminant model coefficients and discriminant metrics through an iterative optimization strategy, bring samples with similar label importance closer to each other, and push samples with different importance away from each other;
[0009] S3, similarity characterization: the similarity of the polysemous objects is explicitly characterized according to the discriminant model coefficients and discriminant metrics obtained in step S2.
[0010] As an improvement of the present invention, the step S1 specifically includes the following steps:
[0011] S11: Perform local linear reconstruction on the feature space of the polysemous object training set to model its potential manifold structure; the reconstruction parameters are specifically:
[0012]
[0013] in, represents the reconstruction parameter of the i-th example in the training set D; G i =D i D i ∈R k×k , x i represents the features of the i-th example in the training set D, Represents the features of the jth neighbor of the i-th example in the training set D; 1 k is a k-dimensional column vector whose values are all 1;
[0014] S12: transferring the manifold structure obtained in step S11 to the tag space to restore the implicit tag importance of the ambiguous object; the transferring of the manifold structure of the feature space to the tag space is specifically as follows:
[0015]
[0016] The first term in the above formula is used to ensure that the label space and the feature space have similar popular structures, and the second term is used to constrain the recovered implicit label importance F to be similar to the original binary logic label Y;
[0017] Among them, ||·||2 represents the 2-norm of the vector; ||·|| F represents the Frobenius norm of the matrix; n represents the number of samples in the training set D; x i represents the features of the i-th example in the training set D; N k (x i )={i r |1≤r≤k} refers to x i In {x1,x2,K,x n}\x i The subscript set of the k nearest neighbors in the set; m is the trade-off hyperparameter; f i represents the implicitly labeled importance vector to be recovered corresponding to the i-th example in the training set D; F = [f1,f2,K,f n ]∈R n×q represents the implicit label importance matrix to be restored; Y = [y1, y2, K, y n ]∈R n×q Represents the original binary logical label matrix of the training set D, y i represents the binary label of the ith example in the training set D.
[0018] As an improvement of the present invention, the implicit marking importance F in step S12 is specifically:
[0019]
[0020] Where F = [f1,f2,K,f n ]∈R n×q represents the implicit label importance matrix to be restored, f irepresents the implicitly labeled importance vector to be recovered corresponding to the i-th example in the training set D; n represents the number of samples in the training set D; I n represents the n×n unit matrix; S = [s1, s2, K, s n ]∈R n×n ,s i =[s i1 ,s i2 ,K,s in ], if j∈N k (x i ), then s ij Obtained by step S11, and if Then ij =0,x i represents the features of the i-th example in the training set D, N k (x i )={i r |1≤r≤k} refers to x i In {x1,x2,K,x n}\x i The subscript set of the k nearest neighbor examples in the set; m is the trade-off hyperparameter.
[0021] As another improvement of the present invention, the step S2 specifically includes the following steps:
[0022] S21: construct a discriminant metric learning objective function to constrain the prediction output of the discriminant model to be aligned with the restored implicit tag importance; the discriminant model is specifically:
[0023]
[0024] The first constraint W in the above formula is and its corresponding Close to each other, the second constraint W Stay away from non The implicit label importance vector of Stay away from non-W The feature encoding of , the fourth term is used to limit the complexity of M to avoid overfitting;
[0025] Where M is the metric matrix to be learned; n represents the number of samples in the training set D; and are the feature vector and implicit label importance vector of the ith example after centralization; f(·) is a nonlinear mapping determined by the kernel function k:X×X→R, which maps the original d-dimensional data to the d′-dimensional Hilbert space, i.e., f(x i )∈R d′ ; W = [w1,w2,K,w q ]∈Rd′×q is the prediction model coefficient; represent exist The subscript set of the k nearest neighbors in the set; represent exist The subscript set of the k nearest neighbors in the set; g is a trade-off hyperparameter; I q represents the q×q unit matrix, is the symmetric LogDet divergence;
[0026] S22: Solving the discriminant model coefficients and the discriminant metric through an iterative optimization strategy, bringing samples with similar label importance closer to each other, and pushing samples with large differences in label importance farther away from each other; the iterative optimization strategy is: fixing one of the prediction model coefficients and the discriminant metric, and optimizing the other until convergence.
[0027] As another improvement of the present invention, in step S22, the discriminant model coefficient and the discriminant metric are solved by an iterative optimization strategy. When the discriminant metric M is fixed and the prediction model coefficient W is solved, the prediction model coefficient W can be equivalently expressed as ΦΘ, where refers to the centered example feature matrix after nonlinear mapping, f(·) is the nonlinear mapping determined by the kernel function k:X×X→R, is the feature vector of the ith example after centering; Θ = [q1, q2, K, q q ]∈R n×q is the combination coefficient to be solved, solve
[0028]
[0029] Where n represents the number of samples in the training set D; h is a trade-off hyperparameter; K = ΦΦ∈R n×d′ represents the kernel matrix, where the element in the i-th row and j-th column is represented as is the implicit label importance matrix after centralization, is the implicitly labeled importance vector of the ith example after centering; M is the metric matrix to be learned. Θ can be solved by the optimization solver;
[0030] When the prediction model coefficient W is fixed and the discriminant metric M is solved,
[0031] M=(U+gI q ) -1 # 1 / 2 (V+gI q ).
[0032] Among them, M is the metric matrix to be learned; A# 1 / 2 B=A 1 / 2(A 1 / 2 BA -1 / 2 ) 1 / 2 A 1 / 2 ;
[0033]
[0034]
[0035] Where n is the number of samples in the training set D; k is the number of nearest neighbors; I q represents the identity matrix of q×q; g is the trade-off hyperparameter; represents the centered example feature matrix after nonlinear mapping, f(·) is the nonlinear mapping determined by the kernel function k:X×X→R, is the feature vector of the ith example after centering, Θ = [q1, q2, K, q q ]∈R n×q To solve the combined coefficient; K = ΦΦ∈R n×d′ represents the kernel matrix, where the element in the i-th row and j-th column is represented as k:X×X→R represents the nonlinear mapping determined by the kernel function; is the implicitly labeled importance vector of the ith example after centering; represent exist The subscript set of the k nearest neighbors in the set; represent exist The subscript set of the k nearest neighbors in the set; W is the prediction model coefficient; K r =Φ r Φ r ,in
[0036] As another improvement of the present invention, in the similarity characterization of step S3, based on the discriminant model coefficient W and the discriminant metric M obtained by optimizing step S2, the polysemous object example pair (x i ,x j ) is characterized by the Mahalanobis distance shown below:
[0037]
[0038] Among them, Dis(x i ,x j ) represents the example x i and x j The Mahalanobis distance between W is the prediction model coefficient; M is the discriminant metric; represents the centered example feature matrix after nonlinear mapping, f(·) is the nonlinear mapping determined by the kernel function k:X×X→R, is the feature vector of the ith example after centering, Θ = [q1, q2, K, q q ]∈R n×q K is the combined coefficient obtained by solving i ∈R n And the rth element n represents the number of samples in the training set D.
[0039] In order to achieve the above-mentioned purpose, the present invention also adopts a technical solution: a polysemous object metric learning system based on implicit label importance recovery, including a computer program, which implements the steps of any of the above-mentioned methods when executed by a processor.
[0040] Compared with the prior art, the technical advantages and effects of the present invention are as follows: the present invention discloses a polysemous object metric learning method and system based on implicit label importance recovery and utilization. Different from the prior polysemous object metric learning that adopts the equal label importance assumption, the present invention takes into account the essential characteristic that the associated labels of polysemous objects have different label importances in the process of polysemous object metric learning, and effectively recovers the implicit label importance of the polysemous object by modeling the manifold structure of the feature space and transferring it to the label space; then, by introducing a discriminative metric learning framework, the recovered implicit label importance information is fully utilized to constrain samples with similar label importance to be close to each other, and samples with large label importance differences to be away from each other. Through this process, the present invention significantly improves the ability of metric learning to characterize the similarity of polysemous objects. BRIEF DESCRIPTION OF THE DRAWINGS
[0041] Figure 1 It is a flowchart of the steps of the polysemous object metric learning method based on implicit label importance recovery and utilization according to the present invention;
[0042] Figure 2 is a flow chart of step S11 of this embodiment;
[0043] Figure 3 It is a flowchart of step S22 of this embodiment. DETAILED DESCRIPTION
[0044] The present invention will be further explained below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention.
[0045] Example 1
[0046] The metric learning method for polysemous objects based on implicit label importance recovery is as follows: Figure 1 As shown, at least the following steps are included:
[0047] Step S1, implicit tag importance recovery: by locally linearly reconstructing the feature space to model its potential manifold structure, the manifold structure is transferred to the tag space to recover the implicit tag importance of the ambiguous object;
[0048] Step S11: Perform local linear reconstruction on the feature space of the polysemous object training set to model its potential manifold structure. The specific implementation method is as follows:
[0049] Let X = R d represents the d-dimensional feature space, represents a label space with q labels. A polysemous object sample is represented as (x,Y), where x∈X represents its feature vector, Represents the corresponding set of related tags. q ]∈{0,1} q Used to represent Y, where y p = 1 if and only if l p ∈Y,y p = 0 if and only if D={(x i ,y i )|1≤i≤n} represents a polysemous object training sample set containing n samples. For each example x in the training set D i (1≤i≤n), which can be reconstructed by linear combination of the k nearest neighboring examples. The reconstruction parameters can be solved by the optimization problem shown in formula (1):
[0050]
[0051] Among them, N k (x i )={i r |1≤r≤k} refers to x i The set of subscripts of the k nearest neighbor examples of . Let Then formula (1) can be equivalently converted to formula (2):
[0052]
[0053] Among them, G i =D i D i ∈R k×k , 1 k is a k-dimensional column vector whose values are all 1.
[0054] To solve formula (2), construct the Lagrangian function as shown in formula (3):
[0055]
[0056] Then let them be about The partial derivative of and l is 0, so we have
[0057]
[0058] Substituting formula (4) into formula (5), we can get
[0059]
[0060] Substituting formula (6) into formula (4), we can obtain the closed-form solution of the optimization problem shown in formula (2):
[0061]
[0062] Figure 2 This is a flow chart of step S11 of this embodiment. By traversing each example x in the entire training set D, i (1≤i≤n), the reconstruction parameters corresponding to each example can be obtained by calculating formula (7) This enables modeling of the manifold structure of the feature space.
[0063] Step S12: transferring the manifold structure obtained in step S11 to the tag space to restore the implicit tag importance of the ambiguous object, which is specifically implemented as follows:
[0064] Let Y = [y1,y2,K,y n ]∈R n×q Represents the original binary logical label matrix of the training set D, F = [f1,f2,K,f n ]∈R n×q represents the implicit tag importance matrix to be restored. Based on the modeling of the feature space manifold structure in step 1, F = [f1,f2,K,f n ]∈R n×q It can be obtained by solving the optimization problem shown in formula (9), thereby transferring the manifold structure of the feature space to the label space.
[0065]
[0066] Among them, ||·|| F represents the Frobenius norm of the matrix, and m is a trade-off hyperparameter. The first term of formula (8) is used to ensure that the tag space and the feature space have similar popular structures, and the second term is used to constrain the recovered implicit tag importance to be similar to the original binary logic tag Y.
[0067] To facilitate the solution, formula (8) can be equivalently converted to formula (9):
[0068]
[0069] In formula (9), tr(·) is the matrix trace operator, I n represents the n×n unit matrix. In addition, S=[s1,s2,K,s n ]∈R n×n ,s i =[s i1 ,s i2 ,K,s in ], where if j∈N k (x i ), then s ij By solving formula (1), we can get Then ij =0.
[0070] Let G(F) denote the objective function of the optimization problem shown in formula (9). The first-order partial derivative of G(F) with respect to F is
[0071]
[0072] By setting it to 0, we can obtain the closed-form solution of the implicit tag importance F:
[0073]
[0074] Step S2, discriminant metric learning: construct a discriminant metric learning objective function, constrain the prediction output of the discriminant model to be aligned with the recovered implicit label importance, solve the discriminant model coefficients and discriminant metrics through an iterative optimization strategy, bring samples with similar label importance closer to each other, and push samples with different importance away from each other.
[0075] Step S21: construct a discriminative metric learning objective function to constrain the prediction output of the discriminative model to be aligned with the restored implicit tag importance. The specific construction method is as follows:
[0076] First, based on the implicit tag importance F recovered in step S12, the ridge regression model is used to realize the discriminant prediction from the example features to the corresponding implicit tag importance:
[0077]
[0078] Where h is a trade-off hyperparameter, and f(·) is a nonlinear mapping determined by the kernel function k:X×X→R, which maps the original d-dimensional data to the d′-dimensional Hilbert space, that is, f(x i )∈R d′ W=[w1,w2,K,wq ]∈R d′×q is the prediction model coefficient, b=[b1,b2,K,b q ]∈R q is the intercept. Let X=[x1,x2,K,x n ]∈R n×d Representing the example feature matrix of the training set D, formula (12) can be equivalently transformed into formula (13) by centering X and F, thereby omitting the intercept b.
[0079]
[0080] in, and are the ith example and implicit label importance after centering, respectively. Let is the centered example feature matrix, is the implicit label importance matrix after centralization.
[0081] Considering that formula (13) essentially processes q tags independently and ignores the relationship between tags, the present invention introduces the Mahalanobis metric M on the basis of formula (13) to measure W and The distance between them is calculated to take the marker correlation into account in the modeling of the prediction model:
[0082]
[0083] in, M can be viewed as a discriminant metric that reduces the i The feature encoding W and the corresponding implicitly labeled importance The distance between.
[0084] In order to further enhance the discriminative performance of the metric M, the present invention further penalizes the feature encodings and implicit labeling importance that are not aligned with each other on the basis of formula (14):
[0085]
[0086] in, represent exist The subscript set of the k nearest neighbors in the set, represent exist The subscript set of the k nearest neighbors in the set, g is the trade-off hyperparameter, is the symmetric LogDet divergence, I q represents the q×q unit matrix. The first constraint W in formula (15) and its corresponding Close to each other, the second constraint W Stay away from non The implicit label importance vector of Stay away from non-W The fourth term is used to limit the complexity of M to avoid overfitting.
[0087] Step S22: Solve the discriminant model coefficients and discriminant metrics through an iterative optimization strategy, so as to implicitly bring samples with similar label importance closer to each other and push samples with large label importance differences further away from each other. The specific implementation method is as follows:
[0088] The present invention adopts an iterative optimization strategy to solve the prediction model coefficient W and the discriminant metric M, that is, fix one of them and optimize the other until convergence:
[0089] Fix M and solve for W
[0090] make Refers to the centralized example feature matrix after nonlinear mapping. According to the representation theorem, the prediction model coefficients can be obtained by linear combination of all training examples, so Based on this, the prediction model coefficient W can be equivalently expressed as ΦΘ, where Θ = [q1, q2, K, q q ]∈R n×q is the combination coefficient to be solved. Let H(W) refer to the objective function of formula (14), and substitute W = ΦΘ into H(W), we can get
[0091]
[0092] Let K = Φ ∈ R n×d′ represents the kernel matrix, where the element in the i-th row and j-th column is represented as Therefore, the first-order partial derivative of H(Θ) with respect to Θ is
[0093]
[0094] Let formula (17) equal to 0, we can get
[0095] nh(KK) -1 Kθ+θM=(KK) -1 KFM. (18)
[0096] Formula (18) is the classic Sylvester equation, which can be solved for Θ using an off-the-shelf optimization solver.
[0097] Fix W and solve for M
[0098] The optimization problem shown in formula (15) can be equivalently transformed into:
[0099]
[0100] in,
[0101]
[0102] In formula (21), K r =Φ r Φ r ,
[0103]
[0104] Let T(M) refer to the objective function of formula (19), and the first-order partial derivative of T(M) with respect to M is:
[0105]
[0106] Let formula (22) be 0, we can get
[0107] M(U+gI q )M=(V+gI q ). (twenty three)
[0108] Formula (23) is the classic Riccati equation, which has a closed-form solution:
[0109] M=(V+gI q ) -1 # 1 / 2 (V+gI q ). (twenty four)
[0110] Among them, A# 1 / 2 B=A 1 / 2 (A 1 / 2 BA -1 / 2 ) 1 / 2 A 1 / 2 .
[0111] Figure 3 The flowchart of step S22 of this embodiment is as follows: first, M is initialized to a q×q dimensional unit matrix I q , and then the discriminant model coefficient W and the discriminant metric M are updated alternately by calculating formula (18) and formula (24) until convergence, thereby implicitly bringing samples with similar label importance closer to each other and pushing samples with large label importance differences further away from each other.
[0112] Step S3, similarity characterization: the similarity of the polysemous objects is explicitly characterized according to the discriminant model coefficients and discriminant metrics obtained in step S2. The specific implementation method is as follows:
[0113] Based on the discriminant model coefficient W and the discriminant metric M obtained by optimizing step S22, the present invention explicitly converts the polysemous object example pair (x i ,x j ) is characterized by the Mahalanobis distance shown in formula (25):
[0114]
[0115] Among them, K i ∈R n And the rth element Polysemous objects with smaller Mahalanobis distances have higher similarity, while polysemous objects with larger Mahalanobis distances have lower similarity.
[0116] In summary, the method and system of the present invention aim to restore the implicit label importance that essentially exists in polysemous objects and utilize it in polysemous object metric learning, thereby further improving the ability of metric learning to characterize the similarity between polysemous objects.
[0117] It should be noted that the above content merely illustrates the technical idea of the present invention and cannot be used to limit the protection scope of the present invention. For ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications all fall within the protection scope of the claims of the present invention.
Claims
1. A polysemous object metric learning method based on implicit label importance recovery, characterized by , including at least the following steps: S1, implicit tag importance recovery: by locally linearly reconstructing the feature space to model its potential manifold structure, the manifold structure is transferred to the tag space to recover the implicit tag importance of the polysemous object; S2, discriminant metric learning: construct a discriminant metric learning objective function, constrain the prediction output of the discriminant model to be aligned with the recovered implicit label importance, solve the discriminant model coefficients and discriminant metrics through an iterative optimization strategy, bring samples with similar label importance closer to each other, and push samples with different importance away from each other; S3, similarity characterization: the similarity of the polysemous objects is explicitly characterized according to the discriminant model coefficients and discriminant metrics obtained in step S2.
2. The ambiguous object metric learning method based on implicit label importance recovery and utilization as claimed in claim 1, characterized in that: The step S1 specifically includes the following steps: S11: Perform local linear reconstruction on the feature space of the polysemous object training set to model its potential manifold structure; the reconstruction parameters are specifically: in, represents the reconstruction parameter of the i-th example in the training set D; G i =D i D i ∈R k×k , x i Represents the features of the i-th example in the training set D; 1 k is a k-dimensional column vector whose values are all 1; S12: transferring the manifold structure obtained in step S11 to the tag space to restore the implicit tag importance of the ambiguous object; the transferring of the manifold structure of the feature space to the tag space is specifically as follows: The first term in the above formula is used to ensure that the label space and the feature space have similar popular structures, and the second term is used to constrain the recovered implicit label importance F to be similar to the original binary logic label Y; Among them, ||·||2 represents the 2-norm of the vector; ||·|| F represents the Frobenius norm of the matrix; n represents the number of samples in the training set D; N k (x i )={i r |1≤r≤k} refers to x i In {x1,x2,K,x n }\x i The subscript set of the k nearest neighbors in the set; m is the trade-off hyperparameter; f i represents the implicitly labeled importance vector to be recovered corresponding to the i-th example in the training set D; F = [f1,f2,K,f n ]∈R n×q represents the implicit label importance matrix to be restored; Y = [y1, y2, K, y n ]∈R n×q Represents the original binary logical label matrix of the training set D, y i represents the binary label of the ith example in the training set D.
3. The ambiguous object metric learning method based on implicit label importance recovery as claimed in claim 2, characterized in that: The implicit tag importance F in step S12 is specifically: Where F = [f1,f2,K,f n ]∈R n×q represents the implicit label importance matrix to be restored; n represents the number of samples in the training set D; I n represents the n×n unit matrix; S = [s1, s2, K, s n ]∈R n×n ,s i =[s i1 ,s i2 ,K,s in ].
4. The ambiguous object metric learning method based on implicit label importance recovery as claimed in claim 1, characterized in that: The step S2 specifically includes the following steps: S21: construct a discriminant metric learning objective function to constrain the prediction output of the discriminant model to be aligned with the restored implicit tag importance; the discriminant model is specifically: The first constraint in the above formula is and its corresponding Close to each other, the second constraint Stay away from non The implicit label importance vector of Stay away from non The feature encoding of , the fourth term is used to limit the complexity of M to avoid overfitting; Where M is the metric matrix to be learned; n represents the number of samples in the training set D; and are the feature vector and implicit label importance vector of the ith example after centralization; f(·) is a nonlinear mapping determined by the kernel function k:X×X→R, which maps the original d-dimensional data to the d′-dimensional Hilbert space, i.e., f(x i )∈R d′ ; W = [w1,w2,K,w q ]∈R d′×q is the prediction model coefficient; represent exist The subscript set of the k nearest neighbors in the set; represent exist The subscript set of the k nearest neighbors in the set; g is a trade-off hyperparameter; I q represents the q×q unit matrix, is the symmetric LogDet divergence; S22: Solving the discriminant model coefficients and the discriminant metric through an iterative optimization strategy, bringing samples with similar label importance closer to each other, and pushing samples with large differences in label importance farther away from each other; the iterative optimization strategy is: fixing one of the prediction model coefficients and the discriminant metric, and optimizing the other until convergence.
5. The ambiguous object metric learning method based on implicit label importance recovery and utilization as claimed in claim 4, characterized in that: In step S22, the discriminant model coefficient and the discriminant metric are solved by an iterative optimization strategy. When the discriminant metric M is fixed and the prediction model coefficient W is solved, the prediction model coefficient W is expressed as ΦΘ, where refers to the centered example feature matrix after nonlinear mapping, f(·) is the nonlinear mapping determined by the kernel function k:X×X→R, is the feature vector of the ith example after centering; Θ = [q1, q2, K, q q ]∈R n×q is the combination coefficient to be solved, solve Where n represents the number of samples in the training set D; h is a trade-off hyperparameter; K = ΦΦ∈R n×d′ represents the kernel matrix, where the element in the i-th row and j-th column is represented as is the implicit label importance matrix after centralization, is the implicitly labeled importance vector of the ith example after centering; M is the metric matrix to be learned; Θ can be solved by the optimization solver; When the prediction model coefficient W is fixed and the discriminant metric M is solved, M=(U+gI q ) -1 # 1 / 2 (V+gI q ). Among them, M is the metric matrix to be learned; A# 1 / 2 B=A 1 / 2 (A 1 / 2 BA -1 / 2 ) 1 / 2 A 1 / 2 ; Where n is the number of samples in the training set D; k is the number of nearest neighbors; I q represents the identity matrix of q×q; g is the trade-off hyperparameter; represents the centered example feature matrix after nonlinear mapping, f(·) is the nonlinear mapping determined by the kernel function k:X×X→R; k:X×X→R represents the nonlinear mapping determined by the kernel function; represent exist The subscript set of the k nearest neighbors in the set; represent exist The subscript set of the k nearest neighbors in the set; W is the prediction model coefficient; K r =Φ r Φ r ,in 6. The ambiguous object metric learning method based on implicit label importance recovery as claimed in claim 1, characterized in that: In the similarity characterization of step S3, based on the discriminant model coefficient W and the discriminant metric M obtained by optimizing step S2, the polysemous object example pair (x i ,x j ) is shown by the following formula The Mahalanobis distance is used to characterize: Among them, Dis(x i ,x j ) represents the example x i and x j The Mahalanobis distance between W is the prediction model coefficient; M is the discriminant metric; represents the centered example feature matrix after nonlinear mapping, f(·) is the nonlinear mapping determined by the kernel function k:X×X→R, is the feature vector of the ith example after centering, Θ = [q1, q2, K, q q ]∈R n×q K is the combined coefficient obtained by solving i ∈R n And the rth element n represents the number of samples in the training set D.
7. A system for learning polysemous object metrics based on implicitly labeled importance recovery, comprising a computer program, characterized in that: When the computer program is executed by a processor, the steps of any of the above methods are implemented.