Attribute missing target fusion identification method and system based on covariance bidirectional alignment

Abstract

The invention provides an attribute missing target fusion identification method and system based on covariance bidirectional alignment. The system comprises an initialization module, a feature learning module, a soft classification module and an evidence fusion module. The initialization module is used for inputting data and preliminarily filling missing values in test data based on a mean vector of training data; the feature learning module obtains new feature representation of the training data and the filled test data through the feature conversion matrix; the soft classification module is used for training two classifiers to obtain two soft classification results of the test data under the original feature representation and the new feature representation; and the evidence fusion module obtains the optimal weight by minimizing the objective function, and fuses the discounted soft classification result based on the evidence theory to obtain the final category of the test data. According to the method, the alignment process of data distribution is fused into learning of missing values, so that the negative influence of data distribution offset on a downstream target identification task is effectively reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of target recognition, and in particular to a method and system for fusion recognition of attribute-missing targets based on covariance bidirectional alignment. Background Art

[0002] In the field of target recognition, data-driven classification models generally assume that the attribute values in the training and test data are complete. However, in actual scenarios such as industrial inspection, environmental monitoring, and energy system management, the problem of missing attributes is extremely common. Affected by sensor failures, transmission errors, or data collection limitations, missing data values will seriously damage the statistical characteristics of the data, causing the performance of traditional classification models to decline sharply. If samples with missing values are directly ignored, once the number of complete data is too small, a high-quality model cannot be trained, which will also cause a decrease in classification performance. Therefore, how to achieve reliable classification and recognition under incomplete data conditions has become a key challenge in practical engineering.

[0003] The current mainstream solution is mainly the missing value filling technology. Traditional methods include simple filling based on statistics (such as mean, median), and complex filling based on generative models (such as generative adversarial networks and variational autoencoders). For example, mean filling (MI) quickly fills missing values with the average value of training data, but this will cause the distribution of the filled test data to deviate from the actual situation (such as Figures 2 to 4 As shown in the figure); although deep learning methods such as Generative Adversarial Infill Network (GAIN) and Missing Data Importance Weighted Autoencoder (MIWAE) can generate more realistic infill values, their models are complex and computationally expensive, making them difficult to apply in scenarios with high missing rates.

[0004] In addition, these methods generally have the following defects:

[0005] Distribution shift: Padding changes the statistical properties of the training data (e.g., mean and variance), disrupting the original distribution of the test data and widening the discrepancy in the covariance structure between the training and test data. Even if the local padding values appear reasonable, the global distribution inconsistency can still weaken the classifier's generalization ability.

[0006] Limitations of feature alignment: Existing methods often focus on unidirectional feature mapping (e.g., aligning test data to the training data distribution), neglecting the modeling of bidirectional covariance relationships. Domain adaptation methods such as Maximum Mean Divergence (MMD) only align global distributions and fail to optimize for local covariance shifts caused by missing attributes.

[0007] Static fusion strategies: Most multi-classifier fusion methods use fixed weights or empirical rules and cannot dynamically adjust to the uncertainty of different missing patterns. For example, combination methods based on evidence theory can handle uncertainty, but the weight determination lacks a clear connection with distribution differences, limiting fusion effectiveness when missing rates are high.

[0008] A search of patent documents revealed an invention patent with publication number CN117456291A, which discloses a defect classification method and device, electronic device, and storage medium. The method includes: inputting the image to be tested and the first position prompt information into the feature extraction network of the defect classification model to obtain image features and position identification features; inputting the image features and position identification features into the fully connected neural network of the defect classification model to obtain the defect classification results; the defect classification model is obtained by training a target sample data set, the target sample data set includes labeled data and unlabeled data, the labeled data includes multiple first sample images and corresponding second position prompt information and annotation labels, the unlabeled data includes multiple second sample images and corresponding third position prompt information, and the second position prompt information and / or third position prompt information are obtained by positionally annotating the defect area in the corresponding sample image through a semi-automatic annotation network. This patent focuses on image defect classification and is only applicable to the image field. The scenarios are limited, the technical versatility is lacking, and it does not address the problems of missing similar attributes and data fusion, making it difficult to cope with complex data conditions.

[0009] In summary, in response to the above-mentioned problems of the existing technology, studying a method and system for fusion recognition of attribute-missing targets based on covariance bidirectional alignment has become a key task that needs to be solved urgently. Summary of the Invention

[0010] In view of the defects in the prior art, the purpose of the present invention is to provide a method and system for fusion recognition of attribute-missing targets based on covariance bidirectional alignment.

[0011] According to the present invention, a method for fusion recognition of attribute-missing targets based on covariance bidirectional alignment is provided, comprising the following steps:

[0012] Step S1: input training data and test data. The training data includes complete attribute values and corresponding labels, and the test data includes incomplete attribute values. The missing values in the test data are preliminarily filled based on the mean vector of the training data to obtain the filled test data.

[0013] Step S2, calculate the covariance matrix C of the training data and the filled test data tr and By minimizing the difference in covariance matrix, we obtain the feature transformation matrices A and B of the training data and the padded test data, respectively, and generate new feature representations Z of the training data and the padded test data.tr and Z te , and retain the original feature representation of the training data and the filled test data;

[0014] Step S3, new feature representation Z based on training data tr and original feature representation, respectively train the first classifier C A and the second classifier C B ; Use the first classifier and the second classifier to classify the padded test data under the original feature representation and the new feature representation Z te Classify the padded test data under , and obtain the first soft classification result and the second soft classification result;

[0015] Step S4: performing weighted fusion on the first soft classification result and the second soft classification result based on evidence theory to obtain the final category of the test data.

[0016] Preferably, step S1 includes the following sub-steps:

[0017] Step S1.1, calculate the mean vector and covariance matrix of the training data: the training data is set to The test data is set to The pattern sets of training data and test data are represented as and The label space of training data and test data is Ω={ω1,ω2,...,ω C},in, is the attribute value of the i-th training sample, is the label of the i-th training sample, is the attribute value of the jth test sample; Represents the sample set of training data; Represents the sample set of test data; K is the feature dimension, N tr and N te Represent the number of training data and test data, ω C represents the Cth category,

[0018] Mean vector of training data and covariance matrix Calculated using the following formula:

[0019]

[0020] in It contains N te A column vector of 1s;

[0021] In step S1.2, the missing values of the test data are filled based on the mean vector and covariance matrix.

[0022] Preferably, step S1.2 is based on the training data X tr The mean vector of the test data X is filled te The missing values in are filled and the test data is obtained The following sub-steps are included:

[0023] Step S1.2.1, set is the test data X te The elements in the pth row and qth column of the matrix are initialized with 0 to fill the missing values, and the result is X te* , and calculate X te* The mean vector of

[0024] Step S1.2.2, calculate the final estimate of missing values

[0025] in, is the mean vector of the training data; is the covariance matrix of the training data; It contains N te A column vector of 1s; Fill missing values in the test data with the mean vector of the training data; To fill the test data The jth vector of ; For X te The element in the pth row and qth column of the matrix; For test data, fill missing values with 0; For X te* The mean vector of ; and They are and The pth element of ; and are the number of missing and non-missing values in the p-th row, respectively.

[0026] Preferably, step S2 includes:

[0027] Step S2.1, calculate the mean vector of the padded test data, and construct the centering matrix of the training data and the padded test data, as well as the covariance matrix C of the training data and the padded test data. tr and

[0028] Step S2.2, calculate the new feature representation Z of the training data tr And the covariance matrix corresponding to the new feature representation of the training data, the feature transformation matrix A of the training data is obtained by minimizing the difference of the covariance matrix;

[0029] Step S2.3, calculate the new feature representation Z of the filled test data te And the covariance matrix corresponding to the new feature representation of the filled test data, the feature transformation matrix B of the filled test data is obtained by minimizing the difference of the covariance matrix.

[0030] Preferably, in step S2.1, the mean vector of the padded test data is Where, is the attribute value of the jth filled test sample, the training data X tr and filled test data ) is the centralization matrix of

[0031] Where, X tr and are training data and filled test data, respectively. and are the mean vectors of the training data and the padded test data, respectively, It contains N tr A column vector of 1s, It contains N te A column vector of 1s;

[0032] The covariance matrices of the training data and the imputed test data are and

[0033] in, is the mean vector of the test data after filling;

[0034] is the centering matrix of the training data;

[0035] To fill the central matrix of the post-test data;

[0036] C tr is the covariance matrix of the training data;

[0037] is the covariance matrix of the test data after filling;

[0038] The superscript T indicates the transpose of the matrix.

[0039] Preferably, in step S2.2, a new feature representation of the training data is constructed And the covariance matrix corresponding to the new feature representation of the training data Minimize the F-norm difference to obtain the feature transformation matrix

[0040] Wherein, the superscript T indicates transposition;

[0041] Z tr New feature representation for training data;

[0042] Represent the corresponding covariance matrix for the new features of the training data;

[0043] A is the feature transformation matrix corresponding to the new feature representation of the training data;

[0044] represents the identity matrix;

[0045] arg min represents the value when minimizing the expression;

[0046] ‖‖ F represents the F-norm.

[0047] Preferably, in step S2.3, a new feature representation of the padded test data is constructed And the covariance matrix corresponding to the new feature representation of the filled test data Minimize the F-norm difference to obtain the feature transformation matrix

[0048] Among them, Z te New feature representation for the test data after filling;

[0049] The covariance matrix corresponding to the new feature representation of the test data after filling;

[0050] B is the feature transformation matrix corresponding to the new feature representation of the test data after filling.

[0051] Preferably, step S3 includes the following sub-steps:

[0052] Step S3.1, first classifier C A New feature representation of the training data and its actual label training, using the first classifier C A Classify the padded test data under the original feature representation to obtain the first soft classification result

[0053] Among them, C A is a classifier trained by new feature representations of the training data;

[0054] is the jth vector representing the new feature of the training data;

[0055] is the first classifier CA Test samples of the original feature representation of the padded test data Perform classification to obtain a first soft classification result;

[0056] is the first classifier C A After classifying the filled test data, c class probability;

[0057] Step S3.2, second classifier C B Original feature representation of the training data and its actual label training, using the second classifier C B Classify the padded test data under the new feature representation to obtain the second soft classification result

[0058] Among them, C B It is a classifier trained by the original feature representation of the training data;

[0059] is the second classifier C B Test samples of the new feature representation of the padded test data Perform classification to obtain a second soft classification result;

[0060] is the second classifier C B After classifying the filled test data, c The probability of the class.

[0061] Preferably, step S4 includes the following sub-steps:

[0062] Step S4.1, discount the fusion soft classification results, the first soft classification results and the second soft classification result The weights are β A ∈[0,1] and β B ∈[0,1], the discount result is and

[0063] Among them, β A and β B express and The weight of A and β B express and After β A and β B weighted; is the first classifier C AAfter classifying the filled test data, c class probability; is the second classifier C B After classifying the filled test data, c class probability;

[0064] Using the Dempster rule in evidence theory to combine the discounted soft classification results, the weighted evidence fusion result is expressed as The DS combination operator is calculated as

[0065] in,

[0066]

[0067] Symbol ⊕ DS Represents the DS combination operator; and Indicates the probability that cannot be clearly assigned to a specific category;

[0068] β A and β B express and The weight of

[0069] Superscript β A and β B express and After β A and β B weighted; is the first classifier C A After classifying the filled test data, c class probability; is the second classifier C B After classifying the filled test data, c class probability;

[0070] Step S4.2: Construct the objective function on the labeled training data set up and The training samples are represented in the original feature and new feature representation The soft classification results under the y-axis are: The mean square error between the combined results and the true labels on the labeled training data is calculated using the following formula:

[0071] The constraint condition is: 0≤β A ≤1,0≤β B ≤1,

[0072] The optimal weight is obtained by minimizing the objective function Using the formula and weights in step S4.1 Get the weighted soft classification results, and combine the two weighted soft classification results to get

[0073] in, is the objective function;

[0074] ‖‖ represents the L2 norm;

[0075] t i is a training sample The actual category One-hot vector of ;

[0076] Step S4.3, convert the combination result into a probability value, the formula is as follows: If p j ({ω c}) is the maximum value, then the test sample Classified as category {ω c}, where p j ({ω c}) represents the test sample Classified as category {ω c}probability.

[0077] The present invention also provides an attribute-missing target fusion recognition system based on covariance bidirectional alignment, comprising:

[0078] Initialization module: input training data and test data. The training data includes complete attribute values and corresponding labels, and the test data includes incomplete attribute values. The missing values in the test data are preliminarily filled based on the mean vector of the training data to obtain the filled test data.

[0079] Feature learning module, calculates the covariance matrix C of training data and filled test data tr and By minimizing the difference in covariance matrix, we obtain the feature transformation matrices A and B of the training data and the padded test data, respectively, and generate new feature representations Z of the training data and the padded test data. tr and Z te , and retain the original feature representation of the training data and the filled test data;

[0080] Soft classification module, new feature representation Z based on training data tr and original feature representation, respectively train the first classifier C A and the second classifier C B; Use the first classifier and the second classifier to classify the padded test data under the original feature representation and the new feature representation Z te Classify the padded test data under , and obtain the first soft classification result and the second soft classification result;

[0081] The evidence fusion module performs weighted fusion on the first soft classification result and the second soft classification result based on evidence theory to obtain the final category of the test data.

[0082] Compared with the prior art, the present invention has the following beneficial effects:

[0083] 1. The present invention adopts a covariance alignment strategy including mean alignment, and uses the alignment of first-order and second-order statistical distributions to reduce the distribution difference between the padded data and the real data, successfully solving the problem of limited classification performance due to data distribution differences in the incomplete data classification process.

[0084] 2. The present invention generates two complementary soft classification results through the covariance bidirectional alignment method. These complementary results can be effectively combined to improve classification performance.

[0085] 3. The present invention obtains the optimal weight by minimizing the mean square error between the combination result and the true label. This method can dynamically adapt to the uncertainty under different missing patterns, is more reasonable and reliable, and effectively reduces the negative impact of distribution differences on fusion.

[0086] 4. The present invention uses a weighted evidence fusion model to comprehensively process the complementary soft classification results generated by covariance bidirectional alignment, fusing complementary information in different directions to improve the classification accuracy of incomplete data. BRIEF DESCRIPTION OF THE DRAWINGS

[0087] Other features, objects and advantages of the present invention will become more apparent upon reading the detailed description of non-limiting embodiments with reference to the following drawings:

[0088] Figure 1 Schematic diagram of the flow of the attribute missing target fusion recognition method based on covariance bidirectional alignment in an embodiment of the present invention;

[0089] Figure 2 Schematic diagram of training data in an embodiment of the present invention;

[0090] Figure 3 Schematic diagram of test data in an embodiment of the present invention;

[0091] Figure 4 Schematic diagram of filling data in an embodiment of the present invention;

[0092] Figure 5Schematic diagram of the attribute-missing target fusion recognition system based on covariance bidirectional alignment in the comparative example of the present invention;

[0093] Figure 6 Schematic diagram of basic information of a UCI target recognition dataset in an embodiment of the present invention. DETAILED DESCRIPTION

[0094] The present invention will be described in detail below with reference to specific embodiments. The following examples will help those skilled in the art to further understand the present invention, but are not intended to limit the present invention in any form. It should be noted that, for those skilled in the art, several changes and improvements can be made without departing from the scope of the present invention. These all fall within the scope of protection of the present invention.

[0095] The present invention provides a method and system for fusion recognition of attribute-missing targets based on covariance bidirectional alignment. The system includes an initialization module, a feature learning module, a soft classification module, and an evidence fusion module. The initialization module is used to input data and preliminarily fill in missing values in the test data based on the mean vector of the training data; the feature learning module learns to obtain two feature conversion matrices, and then obtains new feature representations of the training data and the filled test data through the feature conversion matrices; the soft classification module uses the training data under the new feature representation and the original feature representation to train two classifiers respectively, and obtains two soft classification results of the test data under the original feature representation and the new feature representation; the evidence fusion module obtains the optimal weight by minimizing the objective function, and then discounts the soft classification results according to the optimal weight. Based on evidence theory, the discounted soft classification results are fused to obtain the final category of the test data. The present invention integrates the alignment process of data distribution into the learning of missing values, effectively reducing the negative impact of data distribution offset on downstream target recognition tasks.

[0096] When dealing with the classification problem of incomplete data of type C, this paper integrates the data distribution alignment process into the learning of missing values, effectively reducing the negative impact of data distribution offset. At the same time, it fuses the complementary information of different classifiers with the help of evidence theory, and still obtains accurate recognition results even when there are a large number of missing attribute values, thereby improving the classification accuracy of test data samples and successfully overcoming the problem of limited classification performance due to data distribution offset caused by filling in the missing attribute values in the fusion recognition of target objects.

[0097] Example 1:

[0098] Figure 1 The figure is a flow chart of a method for fusion recognition of attribute-missing targets based on covariance bidirectional alignment in an embodiment of the present invention.

[0099] like Figure 1As shown, this embodiment provides a method for fusion recognition of attribute-missing targets based on covariance bidirectional alignment, including the following steps:

[0100] Step S1: input training data and test data. The training data includes complete attribute values and corresponding labels, and the test data includes incomplete attribute values. The missing values in the test data are preliminarily filled based on the mean vector of the training data to obtain the filled test data.

[0101] Figure 2 Schematic diagram of training data in an embodiment of the present invention; Figure 3 Schematic diagram of test data in an embodiment of the present invention; Figure 4 Schematic diagram of padding data in an embodiment of the present invention.

[0102] like Figure 2-4 As shown, step S1 includes the following sub-steps:

[0103] Step S1.1, calculate the mean vector and covariance matrix of the training data: the training data is set to The test data is set to The pattern sets of training data and test data are represented as and The label space of training data and test data is Ω={ω1,ω2,...,ω C},in, is the attribute value of the i-th training sample, is the label of the i-th training sample, is the attribute value of the jth test sample; Represents the sample set of training data; Represents the sample set of test data; K is the feature dimension, N tr and N te Represent the number of training data and test data, ω C represents the Cth category,

[0104] Mean vector of training data and covariance matrix Calculated using the following formula:

[0105]

[0106] in is a containing N te A column vector of 1s.

[0107] In the standard data classification framework, it is usually assumed that the training data and the test data come from the same distribution. Therefore, the mean vector and covariance matrix (i.e., the first-order and second-order statistical distribution) of the training and test data are close or even the same. However, the filling process of missing attribute values in the test data causes their mean vector (i.e., and ) and the covariance matrix (ie C tr and C te ) difference increases. In this case, the filled missing values cannot truly reflect the properties of the original data, and the distribution shift problem will occur, affecting the classification effect of incomplete data.

[0108] In step S1.2, the missing values of the test data are filled based on the mean vector and covariance matrix.

[0109] Furthermore, in step S1.2, in order to improve the accuracy of classification, it is ensured that the distribution of the training data and the filled test data is as close as possible. tr The mean vector of the test data X is filled te The missing values in are filled and the test data is obtained The following sub-steps are included:

[0110] Step S1.2.1, set is the test data X te The elements in the pth row and qth column of the matrix are initialized with 0 to fill the missing values, and the result is X te* , and calculate X te* The mean vector of

[0111] Step S1.2.2, calculate the final estimate of missing values

[0112]

[0113] in, is the mean vector of the training data; is the covariance matrix of the training data; is a containing N te A column vector of 1s; Fill missing values in the test data with the mean vector of the training data; To fill the test data The jth vector of ; For X te The element in the pth row and qth column of the matrix; X te* For test data, fill missing values with 0; For X te* The mean vector of ; and They are and The pth element of ; and are the number of missing and non-missing values in the p-th row, respectively.

[0114] After this step, the mean vectors of the training data and the padded test data are the same, but their covariance matrices are different. To address the covariance difference, a better strategy is to map the training data and the padded test data using a feature transformation matrix so that their covariance matrices are closer to the covariance matrices of the padded test data and the training data.

[0115] Step S2, calculate the covariance matrix C of the training data and the filled test data tr and By minimizing the difference in covariance matrix, we obtain the feature transformation matrices A and B of the training data and the padded test data, respectively, and generate new feature representations Z of the training data and the padded test data. tr and Z te , and retain the original feature representations of the training data and the padded test data.

[0116] Specifically, step S2 includes:

[0117] Step S2.1, calculate the mean vector of the padded test data, and construct the centering matrix of the training data and the padded test data, as well as the covariance matrix C of the training data and the padded test data. tr and

[0118] Furthermore, in step S2.1, the mean vector of the padded test data is Where, is the attribute value of the jth filled test sample, the training data X tr and filled test data The centralization matrix is Where, X tr and are training data and filled test data, respectively. and are the mean vectors of the training data and the padded test data, respectively, is a containing N tr A column vector of 1s, is a containing N te The covariance matrices of the training data and the filled test data are and in, is the mean vector of the test data after filling; is the centering matrix of the training data; is the central matrix of the filled test data; C tr is the covariance matrix of the training data; is the covariance matrix of the filled test data; the superscript T indicates the transpose of the matrix.

[0119] Step S2.2, calculate the new feature representation Z of the training data tr And the covariance matrix corresponding to the new feature representation of the training data, the feature transformation matrix A of the training data is obtained by minimizing the difference of the covariance matrix.

[0120] Specifically, in step S2.2, a new feature representation of the training data is constructed And the covariance matrix corresponding to the new feature representation of the training data Minimize the F-norm difference to obtain the feature transformation matrix

[0121] Wherein, the superscript T indicates transposition;

[0122] Z tr New feature representation for training data;

[0123] Represent the corresponding covariance matrix for the new features of the training data;

[0124] A is the feature transformation matrix corresponding to the new feature representation of the training data;

[0125] represents the identity matrix;

[0126] argmin represents the value when minimizing the expression;

[0127] ‖‖ F represents the F-norm.

[0128] In this embodiment, during the conversion process of training data, the formula Get the new feature representation of the training data, where A is the feature transformation matrix, is the centralization matrix of the training data, and the covariance matrix corresponding to the new feature representation is Among them C tr is the covariance matrix of the training data. To learn the feature transformation matrix A, minimize the covariance difference between the training data under the new feature representation and the padded test data under the original feature representation, that is, minimize and The difference in F-norm between:

[0129]

[0130] Its closed-form solution is in, The covariance matrix corresponding to the new feature representation of the training data, To fill the covariance matrix corresponding to the test data, represents the identity matrix, argmin represents the value when minimizing the expression, ‖‖ F represents the F-norm.

[0131] In this case, the new features of the training data are represented as because and The covariance difference between and the classifier C trained with its actual labels A Can classify the padded test data well

[0132] Step S2.3, calculate the new feature representation Z of the filled test data te And the covariance matrix corresponding to the new feature representation of the filled test data, the feature transformation matrix B of the filled test data is obtained by minimizing the difference of the covariance matrix.

[0133] Specifically, in step S2.3, a new feature representation of the padded test data is constructed And the covariance matrix corresponding to the new feature representation of the filled test data Minimize the F-norm difference to obtain the feature transformation matrix

[0134] Among them, Z te New feature representation for the test data after filling;

[0135] The covariance matrix corresponding to the new feature representation of the test data after filling;

[0136] B is the feature transformation matrix corresponding to the new feature representation of the test data after filling.

[0137] In this embodiment, during the conversion process of the filled test data, the formula Get the new feature representation of the filled test data, where B is the feature transformation matrix, To fill the central matrix of the test data, the covariance matrix corresponding to the new feature representation is in is the covariance matrix corresponding to the padded test data. To learn the feature transformation matrix B, minimize the covariance difference between the padded test data under the new feature representation and the training data under the original feature representation, that is, minimize and C tr The difference in F-norm between:

[0138]

[0139] Its closed-form solution is in, To fill in the covariance matrix corresponding to the new feature representation of the test data, C tr is the covariance matrix corresponding to the training data, represents the identity matrix, argmin represents the value when minimizing the expression, ‖‖ F represents the F-norm.

[0140] In this case, the new features of the imputed test data are represented as because and The covariance difference between and the classifier C trained with its actual labels B New feature representations that can classify the imputed test data well

[0141] Step S3: Based on the new feature representation and the original feature representation of the training data, the first classifier C is trained. A and the second classifier C B ; Use the first classifier and the second classifier to classify the padded test data under the original feature representation and the new feature representation Z te The padded test data is classified to obtain the first soft classification result and the second soft classification result.

[0142] Specifically, step S3 includes the following sub-steps:

[0143] Step S3.1, first classifier C A New feature representation of the training data and its actual label training, using the first classifier C A Classify the padded test data under the original feature representation to obtain the first soft classification result

[0144] Among them, C A is a classifier trained by new feature representations of the training data;

[0145] is the jth vector representing the new feature of the training data;

[0146] is the first classifier C A Test samples of the original feature representation of the padded test data Perform classification to obtain a first soft classification result;

[0147] is the first classifier C A After classifying the filled test data, c class probability;

[0148] Step S3.2, second classifier C B Original feature representation of the training data and its actual label training, using the second classifier C B Classify the padded test data under the new feature representation to obtain the second soft classification result

[0149] Among them, C B It is a classifier trained by the original feature representation of the training data;

[0150] is the second classifier C B Test samples of the new feature representation of the padded test data Perform classification to obtain a second soft classification result;

[0151] is the second classifier C B After classifying the filled test data, c The probability of the class.

[0152] In this embodiment, the first classifier is a classifier trained based on the new feature representation of the training data, and the second classifier is a classifier trained based on the original feature representation of the training data. A and the second classifier C B They are usually complementary and can be effectively combined at the decision level to improve classification performance.

[0153] Step S4: performing weighted fusion on the first soft classification result and the second soft classification result based on evidence theory to obtain the final category of the test data.

[0154] In practical applications, the importance of the two classifiers is different, and the two soft classification results cannot be treated equally when fusion. In order to obtain a good combination result, the two soft classification results must be considered. and The weight of .

[0155] Specifically, step S4 includes the following sub-steps:

[0156] Step S4.1, discount the fusion soft classification results, the first soft classification results and the second soft classification result The weights are β A ∈[0,1] and β B ∈[0,1], the discount result is and

[0157] Among them, β A and β B express and The weight of A and β B express and After β A and β B weighted; is the first classifier C A After classifying the filled test data, c class probability; is the second classifier C B After classifying the filled test data, c The probability of the class.

[0158] In this case, the probabilities of different categories will be partially redistributed to the element Ω, reducing the importance of the soft classification results in the downstream combination tasks. and represents the probability that cannot be clearly assigned to a specific category. or When is 0, it indicates that the corresponding soft classification result is completely reliable and plays an important role in the combination. or When , the soft classification result will be completely unreliable, and it will play a neutral role in the combination. Specifically, the larger the weight, the higher its importance in the fusion.

[0159] The two discounted soft classification results are combined using the Dempster (DS) rule in evidence theory. The weighted evidence fusion result is expressed as The DS combination operator is calculated as

[0160] in,

[0161]

[0162] Symbol ⊕ DS Represents the DS combination operator; and Indicates the probability that cannot be clearly assigned to a specific category;

[0163] β A and β B express and The weight of

[0164] Superscript β A and β B express and After β A and β B weighted; is the first classifier C A After classifying the filled test data, c class probability; is the second classifier C B After classifying the filled test data, c class probability;

[0165] Step S4.2: To automatically learn the optimal weights, construct the objective function on the labeled training data. set up and The training samples are represented in the original feature and new feature representation The soft classification results under . The mean square error (MSE) between the combined results and the true labels on the labeled training data is calculated using the following formula:

[0166]

[0167] The constraint condition is: 0≤β A ≤1,0≤β B ≤1,

[0168] The optimal weight is obtained by minimizing the objective function Using the formula and weights in step S4.1 Get the weighted soft classification results, and combine the two weighted soft classification results to get

[0169] in, is the objective function;

[0170] ‖‖ represents the L2 norm;

[0171] t i is a training sample The actual category One-hot vector of ;

[0172] In step S4.3, in order to make the final category decision, the combined result is converted into a probability value, as follows: If p j ({ω c}) is the maximum value, then the test sample Classified as category {ω c}.

[0173] Among them, p j ({ω c}) represents the test sample Classified as category {ω c}probability.

[0174] Through the above steps, the method of this embodiment fills the missing attribute values of the test data based on the mean vector of the complete training data, reducing the first-order distribution difference. By minimizing the difference in the covariance matrix of the training data and the filled test data, two feature transformation matrices are learned. New feature representations are obtained by mapping the complete training data and incomplete test data to these matrices using two feature transformation matrices, and using the two classifiers learned from the training data under the original feature representation and the new feature representation. The optimal weights of the two classifiers are determined by minimizing the mean square error (MSE) between the combined result and the true value of the labeled training data. The complementary information of the two classifiers is fused through evidence theory to classify the filled test data and obtain more accurate target recognition results.

[0175] A simple numerical example is used to illustrate step S4. The label space of the data is Ω = {ω1, ω2, ω3, ω4}. Assume that the classifier C A and C B The soft classification results of the two parts are given respectively.

[0176]

[0177] and

[0178]

[0179] For the covariance hypothesis, the optimal weights learned by minimizing the mean squared error loss are In this case, the discounted soft classification results calculated according to formulas (5) and (6) will be as follows.

[0180]

[0181] and

[0182]

[0183] The calculated evidence combination result is:

[0184] m j :m j ({ω1})=0.1029,m j ({ω2})=0.2468,

[0185] m j ({ω3})=0.1640,m j ({ω4})=0.4059,

[0186] m j (Ω)=0.0804;

[0187] The probability value is calculated as follows:

[0188]

[0189] Because p j The value of ({ω4}) is the largest, and the test data pattern will be classified as {ω4}.

[0190] Comparative Example 1:

[0191] Figure 5 Schematic diagram of the attribute-missing target fusion recognition system based on covariance bidirectional alignment in the comparative example of the present invention.

[0192] To verify the effectiveness of the method of the present invention, a number of real-valued benchmark datasets were extracted from the machine learning repository of the University of California, Irvine (UCI), such as Figure 5 As shown in the figure, 13 datasets are included: Alforestfires, Cancer, Diabetes, Heartstatlog, Ilpd, Ionosphere, Vehicle, WBC, WBDC, Wine, RedWine, WhiteWine, and Seeds. Different datasets have different attribute values, number of categories, and sample sizes. Each dataset is evenly divided into two parts, serving as training data and test data, to construct the attribute-missing object recognition task.

[0193] The example samples and basic information of the target recognition dataset are as follows Figure 5 shown.

[0194] Mean imputation (MI), median imputation (MedI), K-nearest neighbor imputation (KNNI), fuzzy C-means imputation (FCMI), fuzzy-based information decomposition (FID), local linear approximation (LLA), data-driven incremental imputation model (DIM), imputation based on cluster linear regression (ICLR), improved K-nearest neighbor for interval-valued imputation (IKNN), fuzzy K-top matching (FKTM) are used as benchmark algorithms for attribute missing target recognition and for comparison. Support vector machine (SVM) with a linear kernel is used as the base classifier. For each test sample, the number of missing attributes is denoted as N (N < K), and is set to {4, 5, 6, 7} to evaluate covariance. The overall accuracy (OA) of the test data is used to measure the classification performance of the method of the present invention and other related methods, and the calculation formula is: Where is an indicator function; and are the predicted label and the actual label of the test sample respectively; N te is the number of test data.

[0195] For fair comparison, the method of the present invention and the benchmark algorithm use the same basic classification model to calculate the overall accuracy. The test results are shown in Tables 1-4.

[0196] Table 1 (N = 4)

[0197]

[0198] Table 2 (N = 5)

[0199]

[0200] Table 3 (N = 6)

[0201]

[0202] Table 4 (N = 7)

[0203]

[0204] The experimental results show that the method of the present invention reduces the negative impact of missing values through the distribution alignment strategy, improves the target recognition performance, and is particularly significant when the number of missing attributes is large. Compared with other algorithms that do not consider covariance alignment, the highest recognition accuracy can be achieved in most datasets. Therefore, using two-way covariance alignment to reduce the decline in classification performance caused by data imputation is an effective means to achieve accurate recognition of attribute missing targets and has important practical application value.

[0205] Example 2:

[0206] The present invention also provides an attribute-missing target fusion recognition system based on covariance bidirectional alignment. The attribute-missing target fusion recognition system based on covariance bidirectional alignment can be implemented by executing the process steps of the attribute-missing target fusion recognition method based on covariance bidirectional alignment, that is, those skilled in the art can understand the attribute-missing target fusion recognition method based on covariance bidirectional alignment as a preferred implementation of the attribute-missing target fusion recognition system based on covariance bidirectional alignment.

[0207] Figure 6 Schematic diagram of basic information of a UCI target recognition dataset in an embodiment of the present invention.

[0208] like Figure 6 As shown, the present invention provides an attribute-missing target fusion recognition system based on covariance bidirectional alignment, comprising:

[0209] Initialization module: input training data and test data. The training data includes complete attribute values and corresponding labels, and the test data includes incomplete attribute values. The missing values in the test data are preliminarily filled based on the mean vector of the training data to obtain the filled test data.

[0210] Feature learning module, calculates the covariance matrix C of training data and filled test data tr and By minimizing the difference in covariance matrix, we obtain the feature transformation matrices A and B of the training data and the padded test data, respectively, and generate new feature representations Z of the training data and the padded test data. tr and Z te , and retain the original feature representation of the training data and the filled test data;

[0211] Soft classification module, new feature representation Z based on training data tr and original feature representation, respectively train the first classifier C A and the second classifier C B ; Use the first classifier and the second classifier to classify the padded test data under the original feature representation and the new feature representation Z te Classify the padded test data under , and obtain the first soft classification result and the second soft classification result;

[0212] The evidence fusion module performs weighted fusion on the first soft classification result and the second soft classification result based on evidence theory to obtain the final category of the test data.

[0213] The training data includes a plurality of training samples, each of which includes attribute values and labels corresponding to the attribute values; the test data includes a plurality of test samples, each of which includes attribute values with missing values;

[0214] Feature learning module, including:

[0215] Matrix calculation unit, used to calculate the centering matrix and covariance matrix of the training data and the padded test data under the new feature representation and the original feature representation;

[0216] Construct a difference unit to find the optimal solution for the difference term and obtain the feature transformation matrix;

[0217] Evidence fusion module, including:

[0218] The weight calculation unit obtains the optimal weight by minimizing the mean square error between the combination result and the true value of the labeled training data;

[0219] The weighted fusion unit is used to weight and combine the two soft classification results according to the calculated optimal weight, and convert the combined result into a probability value.

[0220] In more preferred embodiments, the initialization module inputs training data and test data with missing values, and preliminarily fills the missing values in the test data based on the average value of the training data;

[0221] The feature learning module learns two feature transformation matrices by minimizing the difference in the covariance matrix between the training data and the padded test data;

[0222] The soft classification module trains two classifiers using the new feature representation and labels of the training data and the training data and labels, and applies them to the padded test data and the new feature representation of the padded test data, respectively, to obtain two soft classification results;

[0223] The evidence fusion module obtains the optimal weight by minimizing the objective function, discounts the fused soft classification results according to the optimal weight, and performs weighted evidence fusion to make a category decision to obtain the category of the target image to be identified.

[0224] Those skilled in the art will appreciate that, in addition to implementing the system and its various devices, modules, and units provided by the present invention in purely computer-readable program code, it is entirely possible to implement the same functions of the system and its various devices, modules, and units provided by the present invention in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers by logically programming the method steps. Therefore, the system and its various devices, modules, and units provided by the present invention can be considered a hardware component, and the devices, modules, and units included therein for implementing various functions can also be considered as structures within the hardware component; the devices, modules, and units for implementing various functions can also be considered as both software modules implementing the method and structures within the hardware component.

[0225] The above describes specific embodiments of the present invention. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art may make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. The embodiments of this application and the features in the embodiments may be combined with each other in any manner unless there is a conflict.

Claims

1. A method for fusion recognition of attribute-missing targets based on covariance bidirectional alignment, characterized in that: The steps include: Step S1: input training data and test data, wherein the training data includes complete attribute values and corresponding labels, and the test data includes incomplete attribute values. The missing values in the test data are preliminarily filled based on the mean vector of the training data to obtain the filled test data. Step S2: Calculate the covariance matrix C of the training data and the padded test data. tr and By minimizing the covariance matrix difference, the feature transformation matrices A and B of the training data and the padded test data are obtained respectively, and new feature representations Z of the training data and the padded test data are generated. tr and Z te , and retain the original feature representations of the training data and the padded test data; Step S3, new feature representation Z based on the training data tr and original feature representation, respectively train the first classifier C A and the second classifier C B ; Use the first classifier and the second classifier to classify the padded test data under the original feature representation and the new feature representation Z respectively te Classify the padded test data under , and obtain the first soft classification result and the second soft classification result; Step S4: performing weighted fusion on the first soft classification result and the second soft classification result based on evidence theory to obtain the final category of the test data.

2. The attribute-missing target fusion recognition method based on covariance bidirectional alignment according to claim 1 is characterized in that: The step S1 includes the following sub-steps: Step S1.1, calculate the mean vector and covariance matrix of the training data: the training data is set to The test data is set as The pattern sets of the training data and the test data are respectively expressed as and The label space of the training data and the test data is Ω={ω1,ω2,...,ω C },in, is the attribute value of the i-th training sample, is the label of the i-th training sample, is the attribute value of the jth test sample; Represents the sample set of training data; Represents the sample set of test data; K is the feature dimension, N tr and N te Represent the number of training data and test data, ω C represents the Cth category, The mean vector of the training data and covariance matrix Calculated using the following formula: in It contains N te A column vector of 1s; Step S1.2: Fill in the missing values of the test data based on the mean vector and the covariance matrix.

3. The attribute-missing target fusion recognition method based on covariance bidirectional alignment according to claim 2 is characterized in that: The step S1.2 is based on the training data X tr The mean vector fills the test data X te The missing values in are filled and the test data is obtained The following sub-steps are included: Step S1.2.1, set is the test data X te The elements in the pth row and qth column of the matrix are initialized with 0 to fill the missing values, and we get and calculate The mean vector of Step S1.2.2, calculate the final estimate of missing values in, is the mean vector of the training data; is the covariance matrix of the training data; It contains N te A column vector of 1s; Fill missing values in the test data with the mean vector of the training data; To fill the test data The jth vector of ; For X te The element in the pth row and qth column of the matrix; For test data, fill missing values with 0; for The mean vector of ; and They are and The pth element of ; and are the number of missing and non-missing values in the p-th row, respectively.

4. The attribute-missing target fusion recognition method based on covariance bidirectional alignment according to claim 1 is characterized in that: The step S2 comprises: Step S2.1, calculate the mean vector of the padded test data, and construct the centering matrix of the training data and the padded test data, as well as the covariance matrix C of the training data and the padded test data. tr and Step S2.2, calculate the new feature representation Z of the training data tr and the covariance matrix corresponding to the new feature representation of the training data, and obtaining the feature transformation matrix A of the training data by minimizing the difference of the covariance matrix; Step S2.3, calculate the new feature representation Z of the padded test data te And the covariance matrix corresponding to the new feature representation of the padded test data, the feature transformation matrix B of the padded test data is obtained by minimizing the covariance matrix difference.

5. The attribute-missing target fusion recognition method based on covariance bidirectional alignment according to claim 4 is characterized in that: In step S2.1, the mean vector of the padded test data is Where, is the attribute value of the jth filled test sample, the training data X tr and the padded test data The centralization matrix is Where, X tr and are training data and filled test data, respectively. and are the mean vectors of the training data and the padded test data, respectively, It contains N tr A column vector of 1s, It contains N te A column vector of 1s; The covariance matrices of the training data and the filled test data are and in, is the mean vector of the test data after filling; is the centering matrix of the training data; To fill the central matrix of the post-test data; C tr is the covariance matrix of the training data; is the covariance matrix of the test data after filling; The superscript T indicates the transpose of the matrix.

6. The attribute-missing target fusion recognition method based on covariance bidirectional alignment according to claim 5 is characterized in that: In step S2.2, a new feature representation of the training data is constructed. And the covariance matrix corresponding to the new feature representation of the training data Minimize the F-norm difference to obtain the feature transformation matrix Wherein, the superscript T indicates transposition; Z tr New feature representation for training data; Represent the corresponding covariance matrix for the new features of the training data; A is the feature transformation matrix corresponding to the new feature representation of the training data; represents the identity matrix; argmin represents the value when minimizing the expression; ‖‖ F represents the F-norm.

7. The attribute-missing target fusion recognition method based on covariance bidirectional alignment according to claim 6 is characterized in that: In step S2.3, a new feature representation of the padded test data is constructed. And the covariance matrix corresponding to the new feature representation of the filled test data Minimize the F-norm difference to obtain the feature transformation matrix Among them, Z te New feature representation for the test data after filling; The covariance matrix corresponding to the new feature representation of the test data after filling; B is the feature transformation matrix corresponding to the new feature representation of the test data after filling.

8. The attribute-missing target fusion recognition method based on covariance bidirectional alignment according to claim 7 is characterized in that: The step S3 includes the following sub-steps: Step S3.1, first classifier C A New feature representation of the training data and its actual label training, using the first classifier C A Classify the padded test data under the original feature representation to obtain the first soft classification result Among them, C A is a classifier trained by new feature representations of the training data; is the jth vector representing the new feature of the training data; is the first classifier C A Test samples of the original feature representation of the padded test data Perform classification to obtain a first soft classification result; is the first classifier C A After classifying the filled test data, c class probability; Step S3.2, second classifier C B Original feature representation of the training data and its actual label training, using the second classifier C B Classify the padded test data under the new feature representation to obtain the second soft classification result Among them, C B It is a classifier trained by the original feature representation of the training data; is the second classifier C B Test samples of the new feature representation of the padded test data Perform classification to obtain a second soft classification result; is the second classifier C B After classifying the filled test data, c The probability of the class.

9. The attribute-missing target fusion recognition method based on covariance bidirectional alignment according to claim 8 is characterized in that: The step S4 includes the following sub-steps: Step S4.1, discount the fusion soft classification results, the first soft classification results and the second soft classification result The weights are β A ∈[0,1] and β B ∈[0,1], the discount result is and Among them, β A and β B express and The weight of A and β B express and After β A and β B weighted; is the first classifier C A After classifying the filled test data, c class probability; is the second classifier C B After classifying the filled test data, c class probability; Using the Dempster rule in evidence theory to combine the discounted soft classification results, the weighted evidence fusion result is expressed as The DS combination operator is calculated as in, Symbol ⊕ DS Represents the DS combination operator; and Indicates the probability that cannot be clearly assigned to a specific category; β A and β B express and The weight of Superscript β A and β B express and After β A and β B weighted; is the first classifier C A The probability of the padded test data being in the ωcth class after classification; is the second classifier C B After classifying the filled test data, c class probability; Step S4.2: Construct the objective function on the labeled training data set up and The training samples are represented in the original feature and new feature representation The soft classification results under the y-axis are: The mean square error between the combined results and the true labels on the labeled training data is calculated using the following formula: The constraint condition is: 0≤β A ≤1,0≤β B ≤1, The optimal weight is obtained by minimizing the objective function Using the formula and weights in step S4.1 Get the weighted soft classification results, and combine the two weighted soft classification results to get in, is the objective function; ‖‖ represents the L2 norm; t i is a training sample The actual category One-hot vector of ; Step S4.3, convert the combination result into a probability value, the formula is as follows: If p j ({ω c }) is the maximum value, then the test sample Classified as category {ω c }, where p j ({ω c }) represents the test sample Classified as category {ω c }probability.

10. A system for fusion recognition of attribute-missing objects based on covariance bidirectional alignment, characterized in that: include: Initialization module, inputting training data and test data, wherein the training data includes complete attribute values and corresponding labels, and the test data includes incomplete attribute values, preliminarily filling the missing values in the test data based on the mean vector of the training data to obtain the filled test data; Feature learning module, calculates the covariance matrix C of the training data and the filled test data tr and By minimizing the covariance matrix difference, the feature transformation matrices A and B of the training data and the padded test data are obtained respectively, and new feature representations Z of the training data and the padded test data are generated. tr and Z te , and retain the original feature representations of the training data and the padded test data; Soft classification module, based on the new feature representation Z of the training data tr and original feature representation, respectively train the first classifier C A and the second classifier C B ; Use the first classifier and the second classifier to classify the padded test data under the original feature representation and the new feature representation Z respectively te Classify the padded test data under , and obtain the first soft classification result and the second soft classification result; The evidence fusion module performs weighted fusion on the first soft classification result and the second soft classification result based on evidence theory to obtain a final category of the test data.

Images