A target classification method and device based on evidence multi-view kernelization

Abstract

The application discloses a target classification method and device based on evidence multi-view kernelization, obtains feature information corresponding to different image attributes of a to-be-classified image set and the number of image categories corresponding to the to-be-classified image set, initializes the similarity of each to-be-classified image in different image attributes to each image category, and calculates the trust value of the to-be-classified image belonging to each image category in different image attributes based on the similarity, calculates the attribute weight of the image attribute according to the trust value, calculates the classification error of the to-be-classified image by using the attribute weight, the trust value and the similarity, and determines the category of the to-be-classified image in the to-be-classified image set based on the trust value when the change amount of the classification error is less than or equal to an error threshold. The application can improve the image classification precision by combining the influence of different image attributes on image classification.

Description

Technical Field

[0001] This invention belongs to the field of image classification technology, and in particular relates to a target classification method and apparatus based on evidence multi-view kernelization. Background Technology

[0002] With the rapid development of sensors, the internet, and storage devices, an image can be described by a series of different features, including color, histogram, texture, and so on. How to cluster these images with multiple perspective features has attracted considerable attention and is widely used in fields such as medical diagnosis and autonomous driving.

[0003] Several clustering methods based on evidential reasoning have emerged, termed evidential clustering. These methods apply evidential reasoning to image clustering, enabling the representation of clustering inaccuracies. In clustering problems, elements within the evidential reasoning identification framework are generally considered to fall into three categories: single-class, composite, and noisy. A set containing only one element is called a single-class, such as w1; conversely, a set containing more than one element is called a composite class, such as {w1, w2}. Notably, a composite class is also considered a separate class, which can be understood as a transitional class between these contained single-classes. (Empty set) It is generally used to represent noise classes. In generalized frame evidence clustering, images are allowed to be assigned to corresponding classes with different confidence values ​​(mass) to characterize uncertainty, where m(·) represents the confidence value of an image to a certain class. For each target, the confidence value is... A represents the image category. When an image is assigned to a composite class, it indicates that the image has local imprecision (indistinguishability), and based on the current information, the image cannot be accurately classified into a single class, but the target may belong to any single class contained in the composite class.

[0004] However, evidence-based clustering methods can only represent the impreciseness of clustering in single-view images and cannot perform cross-view information analysis and fusion in multi-view images. Therefore, these methods cannot be applied to multi-view clustering, or the accuracy of clustering results in multi-view data is very low. Summary of the Invention

[0005] The purpose of this invention is to provide a target classification method and apparatus based on evidence multi-view kernelization, so as to perform cross-attribute analysis and fusion on images with multiple attributes and improve the accuracy of multi-view image classification.

[0006] This invention adopts the following technical solution: a target classification method based on multi-view evidence kernelization, comprising the following steps:

[0007] Obtain feature information of different image attributes corresponding to the image set to be classified, as well as the number of image categories corresponding to the image set to be classified;

[0008] Initialize the similarity between each image to be classified and each image category in different image attributes, and calculate the trust value of the image to be classified belonging to each image category in different image attributes based on the similarity;

[0009] Attribute weights for image attributes are calculated based on trust values;

[0010] The classification error of the image to be classified is calculated using attribute weights, trust values, and similarity. When the change in classification error is less than or equal to the error threshold, the category of the image to be classified in the image set is determined based on the trust value.

[0011] Furthermore, when the change in classification error exceeds the error threshold, the similarity is updated based on the trust value, and the calculation continues.

[0012] Furthermore, the classification error is calculated as follows:

[0013]

[0014] Where Q represents the number of image attributes, q∈{1,2,...,Q}, a q Let represent the attribute weight of the q-th image attribute, N represent the number of images to be classified, i∈{1,2,...,N}, s∈{1,2,...,N}, k∈{1,2,...,N}, and C represent the number of image categories, j∈{1,2,...,C}. Let represent the trust value that the i-th image belongs to the j-th image category in the q-th image attribute, and β represent the correlation factor between the trust value and the similarity. Let θ represent the similarity between the i-th and j-th image categories in the q-th image attribute, and let θ represent the correlation factor between the image attribute and its weight. Let η represent the similarity between the s-th and k-th images in the q-th image attribute, and let η represent the difference factor of the attribute weights of different image attributes.

[0015] Furthermore, the trust value for an image to be classified belonging to each image category across different image attributes, based on similarity calculations, includes:

[0016]

[0017] Furthermore, determining the category of the images to be classified in the image set based on the trust value includes:

[0018] Select the highest trust value among the trust values ​​of the image to be classified in different image attributes;

[0019] When there is no such thing as When a trust value is established, the image to be classified is assigned to the image category corresponding to the highest trust value; where Belimax Bel represents the maximum trust value of the image i to be classified. ij Let represent the confidence value of image i in relation to the j-th single class of image categories. This represents the threshold for the trust value gap.

[0020] Furthermore, when there exists that makes When the trust value is established:

[0021] Reclassify the images to be classified.

[0022] Furthermore, reclassifying the images to be classified includes:

[0023] Choose to Establish a single image class and construct a set of reclassified image categories;

[0024] The images to be reclassified are classified based on the set of reclassified image categories.

[0025] Furthermore, classifying the image to be reclassified based on the set of reclassified image categories includes:

[0026] The second similarity between the image to be classified and each image category in the set of image categories for reclassification is calculated based on similarity.

[0027] Calculate the separation degree of each image category in the reclassified image category set based on similarity across different image attributes;

[0028] The second trust value of the reclassified image for different image categories is calculated based on the separation degree and the second similarity.

[0029] The image category of the reclassified image is determined based on the second trust value.

[0030] Furthermore, determining the image category of the reclassified image based on the second trust value includes:

[0031]

[0032] Where, μ ij′ The second trust value represents the category in which the i-th image belongs to the j′-th image category. This represents the second trust value indicating that the i-th image belongs to the j′-th image category in the q-th image attribute. This represents the normalized separation degree of the j′ image category in the q-th image attribute.

[0033] Another technical solution of the present invention: a target classification device based on evidence multi-view kernelization, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-described method.

[0034] The beneficial effects of this invention are as follows: This invention extracts feature information of an image in different image attributes, calculates the trust value and attribute weight of the image for each image category based on the feature information, and finally calculates the image classification error by combining the trust value and the corresponding attribute weight. When the classification error meets the preset requirements, the corresponding image category is determined. This invention can improve the image classification accuracy by combining the influence of different image attributes on image classification. Attached Figure Description

[0035] Figure 1 This is a schematic diagram illustrating that the image categories are distributed in a spherical shape.

[0036] Figure 2 This is a schematic diagram illustrating that the image categories are distributed in a non-spherical manner.

[0037] Figure 3 This is a schematic diagram of the clustering method in an embodiment of the present invention. Detailed Implementation

[0038] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0039] In addition to the problems mentioned above, traditional evidence clustering methods suffer from extremely high computational complexity due to their generalized frameworks, making them unsuitable for large-scale datasets. Finally, existing evidence clustering methods primarily focus on mining linear features of the data, thus only achieving good results with spherically distributed data (e.g., Figure 1 As shown), it cannot be applied to non-spherical distributed data (such as...). Figure 2 (As shown). Since the image attribute information of multi-view images has a certain correlation in each viewpoint, existing evidence clustering methods, which are only applied to single-view images, cannot take into account the relationship between different views. Therefore, even if existing evidence clustering methods are applied to multi-view images, it is difficult to improve the image classification accuracy.

[0040] In this invention, an image refers to an image with a single target object, such as a one-inch photograph of a person, an image within a target image frame that has already been identified, and so on.

[0041] This invention discloses a target classification method based on evidence multi-view kernelization, comprising the following steps: obtaining feature information of the image set to be classified corresponding to different image attributes, and the number of image categories corresponding to the image set to be classified; initializing the similarity between each image to be classified and each image category in different image attributes, and calculating the trust value of the image to be classified belonging to each image category in different image attributes based on the similarity; calculating the attribute weight of the image attribute based on the trust value; calculating the classification error of the image to be classified using the attribute weight, trust value and similarity; when the change in classification error is less than or equal to the error threshold, determining the category of the image to be classified in the image set based on the trust value.

[0042] This invention extracts feature information from images in different image attributes, calculates the trust value and attribute weight of the image for each image category based on the feature information, and finally calculates the image classification error by combining the trust value and the corresponding attribute weight. When the classification error meets the preset requirements, the corresponding image category is determined. This invention can improve image classification accuracy by combining the influence of different image attributes on image classification.

[0043] For a set of images, this invention uses their different attribute features to obtain different feature information for each image, and treats the same attribute feature of all images as a single viewpoint. Available attribute features include Histogram of Oriented Gradient (HOG), Local Binary Pattern (LBP), Scale-Invariant Feature Transform (SIFT), Local Texture Features, Soble operator, and Laplacian operator, etc.

[0044] Specifically, the Histogram of Oriented Gradients (HGP) divides the preprocessed image into cells / blocks, counts the gradient histogram (number of different gradients) of each block to form a feature (descriptor), and then concatenates the features of all blocks to obtain a row vector. That is, the feature information of the i-th image under the q-th image attribute (HOG viewpoint). The features of N images are combined into a matrix. Thus, the feature matrix of the image set under the HOG perspective was obtained.

[0045] Regarding Local Binary Pattern (LBP) features, the image is divided into 3×3 pixel blocks. Using the center pixel as a threshold, the grayscale values ​​of the surrounding eight pixels are compared. If a surrounding pixel's value is greater than the threshold, that pixel is marked as 1; otherwise, it's marked as 0. The eight points within the block can be arranged to generate an 8-bit binary number. The decimal representation of this binary number is the LBP value of the center pixel. The LBP values ​​of all pixels in the i-th image are then concatenated into a row vector. It represents the feature information under the LBP image attributes of the i-th image. The features of the N images are combined into a matrix. Thus, the feature matrix of the image set in the LBP perspective was obtained.

[0046] Regarding scale-invariant feature changes, the original image is sequentially subjected to scale-space extrema detection, keypoint localization, orientation assignment, and keypoint descriptor operations. These corresponding features are then concatenated into a row vector. This represents the feature information of the i-th image at the q-th viewpoint (SIFT viewpoint). The features of the N images are combined into a matrix. The feature matrix of the image set in the SIFT viewpoint is thus obtained.

[0047] In this invention, the value of Q is determined by the number of attribute features used. This invention is only applicable to cases where Q ≥ 1. In principle, the larger Q is, the longer it takes to obtain clustering results using this invention.

[0048] After feature extraction from the image, Q matrices are used. This represents a set of feature data containing Q viewpoints (i.e., image attributes). The next objective of this invention is to classify the images in the target image set into C image classes based on this feature information. However, some images may not be sufficiently classified based on the extracted features; in such cases, this invention will assign an imprecise result to that image.

[0049] In the clustering part, this invention adopts a dynamic multi-perspective evidence clustering based on kernel technology. It uses kernel technology to construct a multi-perspective high-dimensional feature space to mine the nonlinear features of the data, and designs a dynamic framework for multi-perspective evidence clustering within this space, thereby reducing the high computational complexity brought about by the generalized evidence clustering framework (and kernel technology).

[0050] This invention introduces kernel technology into multi-view evidence clustering to perform secondary mining of image features, constructing a new high-dimensional feature space and realizing nonlinear feature mining in evidence clustering. Compared to using Euclidean distance to describe the relationship of targets in the original feature space (HOG, LBP, SIFT), mapping feature information from each viewpoint from the original feature space to an unknown high-dimensional space is more conducive to mining nonlinear relationships. Using linear methods to collect and analyze features in the high-dimensional feature space can yield the corresponding nonlinear features in the original feature space.

[0051] This invention utilizes kernel technology to obtain the trust values ​​of images for different categories in a high-dimensional feature space from various perspectives, thereby achieving multi-view evidence clustering based on nonlinear features. Specifically, the kernel method constructs a nonlinear mapping function x→φ(x) to map data from the original feature space to a finite or infinite high-dimensional feature space. The specific form of φ(·) can be unknown, and the inner product in the high-dimensional feature space can be calculated using a kernel function K(x, y) = <φ(x), φ(y)> that conforms to Mercer's theorem. Specifically, in this high-dimensional feature space, the distance between the image and the class center is measured as follows:

[0052]

[0053] in, This represents the distance between the i-th image and the j-th image category in the q-th image attribute, i.e., the similarity. The larger the distance, the greater the difference between the two, indicating that the i-th image does not have the representative features of the j-th image category, i.e., it does not belong to the j-th image category. It is the feature of the i-th image in the q-th image attribute. The corresponding feature in the high-dimensional space φ(·) It is the class center of the j-th image category in the q-th image attribute. (This can also be understood as the representative feature of the j-th class) corresponding to the feature in the high-dimensional space φ(·).

[0054] In practical applications, the method for calculating φ(·) is unknown; only the method for calculating the kernel function K(·,·) is known. Therefore, the Euclidean distance... Convert to The distance value is calculated. Common kernel functions K(·,·) that can be used in this invention include: Gaussian kernel function, polynomial kernel function, etc.

[0055] In the embodiments of the present invention:

[0056]

[0057] The above formula can be used to derive the method for calculating the correlation kernel function in the high-dimensional feature space:

[0058]

[0059]

[0060] in, It is the class center of the m-th image category in the q-th image attribute. It is the class center of the nth image category in the qth image attribute. It is the confidence value used to determine whether the i-th image belongs to the m-th image category based on the q-th image attribute. It is the confidence value used to determine whether the i-th image belongs to the n-th image category based on the q-th image attribute. It is the confidence value used to determine whether the k-th image belongs to the j-th image category based on the q-th image attribute. It is the feature of the k-th image in the q-th image attribute.

[0061] In this embodiment of the invention, calculating the trust value of an image to be classified belonging to each image category in different image attributes based on similarity includes:

[0062]

[0063] in, Let represent the trust value of the i-th image to the j-th image category in the q-th image attribute, C represent the number of image categories, and β represent the correlation factor between the trust value and the similarity.

[0064] In this embodiment of the invention, the attribute weights of image attributes are calculated based on the trust value:

[0065]

[0066] Where N represents the number of images to be classified, s∈{1, 2, ..., N}, k∈{1, 2, ..., N}, Let η represent the similarity between the s-th and k-th images in the q-th image attribute, and let η represent the difference factor of the attribute weights of different image attributes.

[0067] In evidence clustering, performing clustering under a generalized framework for each image actually generates a lot of unnecessary computation because the number of potential categories a target could potentially belong to is often finite. For example, when classifying a group of portraits aged 0-60 into categories like 0-10, 11-20, 21-30, 31-40, 41-50, and 51-60, a single portrait could potentially belong to the categories {0-10, 11-20}, but not {0-10, ..., 41-50}. This invention uses a simple method to first perform a basic classification for all images, and then performs more detailed feature analysis on images that are difficult to distinguish to confirm their categories.

[0068] Specifically, this invention constructs a three-step multi-view evidence clustering dynamic framework: 1. Full image partitioning; 2. Image classification; 3. Partial image re-partitioning. Specifically, first, the support of all images for a single class is calculated; then, based on this support, the images are divided into precisely classified images and inaccurately classified images, and the latent classes of the inaccurately classified images are determined; finally, the inaccurately classified images are re-partitioned within their latent class ranges to obtain the final clustering result.

[0069] In this embodiment of the invention, firstly, a multi-view image classification error is designed for the entire image to obtain the multi-view confidence values ​​of the image that may belong to different single classes and the class centers of the single classes under each view. The corresponding objective function (i.e., classification error) is:

[0070]

[0071] Where Q represents the number of image attributes, q∈{1,2,...,Q}, a q Let represent the attribute weight of the q-th image attribute, i∈{1,2,...,N}, and C represent the number of image categories, j∈{1,2,...,C}. This represents the trust value that the i-th image belongs to the j-th image category in the q-th image attribute. Let θ represent the similarity between the i-th image and the j-th image category in the q-th image attribute, and let θ represent the correlation factor between the image attribute and the attribute weight.

[0072] Unlike single-view data analysis, this invention's method mines features from different aspects of an image through various image attribute features, resulting in varying data densities for different image attributes. In some image attributes, the image distribution is more concentrated (i.e., categories are easier to distinguish). In other image attributes, the image distribution is more uniform (i.e., category separability is poor). Therefore, the weights of image attributes should be different, assigning higher weights to image attributes with better cluster separability, and referring more to the confidence value provided by this feature when finally determining the image category.

[0073] In the above formula, the first term indicates that the conclusion that an image belongs to class j is highly reliable when the image i has representative features of class j, and vice versa. Therefore, The smaller the better. β controls the correlation between confidence and distance, i.e., whether the confidence value is sensitive to changes in distance (when β is large, even a slight increase in feature difference will significantly reduce the confidence value). q Control the weight of the q-th image attribute.

[0074] Second item A smaller value indicates that the image attributes are closer (more similar) in this high-dimensional space, meaning that images of the same type are close to each other. Therefore, the ability to classify images based on this attribute is stronger, and thus the weight 'a' of this viewpoint is higher. q It should be larger, that is The smaller the better, θ is used to control the correlation between the classification ability of view features and the weights.

[0075] Third item This refers to using the Shannon entropy concept to determine the viewpoint weight 'a'. q The larger the difference, the better. η is used to control the degree of difference in weights across multiple perspectives. The larger η is, the better. q The greater the gap, the better.

[0076] In addition, when the change in classification error exceeds the error threshold, the similarity is updated based on the trust value, and the calculation continues.

[0077] Next, as Figure 3 As shown, the image is divided into precisely divisible images and inprecisely divisible images; and a dynamic generalized subframe is determined for inprecisely divisible targets.

[0078] As a specific implementation, determining the category of an image in a set of images to be classified based on its trust value includes: selecting the image with the highest trust value among different image attributes. This invention constructs a multi-view evidence function. Bel i (w j ) represents the probability that the i-th image belongs to the single class of the j-th image. Bel imax This represents the maximum trust value of the image.

[0079] For different images, two situations may occur: 1. j does not exist. If (i.e., the largest trust value is much greater than others), then the image can be determined to belong to the j-th image class, which is called an image that can be precisely divided. These images do not need further clustering; 2. There exists j such that... If the differences between the largest trust values ​​are small, then the image cannot be assigned to a single image class at this stage. It is called an image that cannot be precisely divided. This part of the image will be further divided later to determine its possible single or composite class.

[0080] In other words, when there is no such thing as When a trust value is established, the image to be classified is assigned to the image category corresponding to the highest trust value; where Bel imax Bel represents the maximum trust value of the image i to be classified. ij This represents the trust value of image i in the single class of image category j. This represents the trust value gap threshold. When a threshold exists... When a trust value is established: the image to be classified is reclassified.

[0081] In real-world scenarios, the number of categories an image can potentially belong to is often finite; that is, the number of single classes it can belong to and the size of the number of composite classes it can belong to (|A|). j |) are often smaller. For example, when classifying a group of portraits aged 0-60 into categories such as 0-10, 11-20, 21-30, 31-40, 41-50, and 51-60, a portrait could potentially belong to the category {0-10, 11-20} or {21-30, 31-40, 41-50}. However, if its potential category is {0-10, 11-20, 21-30, 31-40, 41-50}, this is clearly unreasonable.

[0082] Therefore, this invention uses a simple method to initially eliminate many impossible classes, thereby reducing the computational cost of generalized frame evidence clustering (the number of classes in a generalized frame increases exponentially with the number of individual classes). Based on the global trust value of the image, this invention filters out the potentially classable targets for each imprecisely classifiable object. Furthermore, the present invention dynamically processes Ω i The corresponding generalized subframework, in the generalized subframework is Based on this, further limit the size of composite categories (|A j |), to obtain the dynamic generalized subframe:

[0083]

[0084] In other words, the choice makes Establish a single image class and construct a set of reclassified image categories; classify the images to be reclassified based on the set of reclassified image categories.

[0085] In one embodiment, if the first image potentially belongs to class Ω1 = {w1, w3, w4}, then its corresponding generalized subframe is If τ = 2, then its dynamic generalized subframework threshold It can be configured according to the needs of the scenario. The larger the value, the larger the range of suspicious classes of the target, and the higher the computational complexity, but the error rate of clustering will also decrease.

[0086] threshold And τ can be set according to the needs of the scenario. The larger τ is, the larger the dynamic generalized subframe of the target, and the higher the computational complexity, but the clustering error rate will also decrease.

[0087] In one embodiment, classifying an image to be reclassified based on a set of reclassified image categories includes: calculating a second similarity between the image to be classified and each image category in the set of reclassified image categories based on similarity; calculating a separation degree for each image category in the set of reclassified image categories based on similarity in different image attributes; calculating a second trust value for the reclassified image in different image attributes for different image categories based on the separation degree and the second similarity; and determining the image category of the reclassified image based on the second trust value.

[0088] The second similarity is calculated between the image to be classified and each image category in the set of image categories to be reclassified. The calculation method is as follows:

[0089]

[0090] Among them, γ can be used to manually control the influence of composite class centers and related single class center pairs. This embodiment uses... The objective function of multi-perspective evidence clustering was optimized by ρ, which reduced the influence of composite cluster centers on single cluster centers, and improved the accuracy of clustering results by mining local information from each perspective. Represents the features of the i-th reclassified image under the q-th image attribute. In the high-dimensional space φ(·), the class center of the j′-th image category Euclidean distance, class center of a single class As determined above, the class center of the composite class in the high-dimensional space φ(·) is calculated using the following formula:

[0091]

[0092] By using replace This invention reduces the possibility of images being misclassified into unreasonable composite classes (images will not be misclassified into composite classes of multiple classes that are far apart or cannot have overlapping areas). Similarly, when classifying a set of portraits aged 0-60, if a portrait could potentially belong to the category {21-30 years old, 31-40 years old, 41-50 years old}, its inaccurate classification result could be {21-30 years old, 31-40 years old} or {31-40 years old, 41-50 years old}. However, if its inaccurate classification range is defined as {21-30 years old, 41-50 years old}, this is obviously unreasonable. Formula 19 helps to eliminate the possibility of assigning images to unreasonable composite classes (inaccurate classification).

[0093] Represents the normalized separation degree of the j′-th image category in the q-th image attribute. The higher the separation degree of the j′-th image category in the q-th image attribute (the better the separability of the j-th class image in the q-th feature), the more reliable the trust value of the image in the j-th class of the q-th image attribute. Separation degree also reflects the importance of relevant information of a certain class from a certain perspective. Multi-view separation degree The calculation method is as follows:

[0094]

[0095]

[0096] Multi-view separation Each class under each perspective is evaluated separately to fully explore the local features of each perspective, so that the information from multiple perspectives can be fully utilized to obtain more accurate clustering results.

[0097] Therefore, the second trust values ​​for each image attribute are as follows:

[0098]

[0099] Furthermore, determining the image category of the reclassified image based on the second trust value includes:

[0100]

[0101] Where, μ ij′ The second trust value represents the category in which the i-th image belongs to the j′-th image category. This represents the second trust value indicating that the i-th image belongs to the j′-th image category in the q-th image attribute. This represents the normalized separation degree of the j′ image category in the q-th image attribute.

[0102] It is important to note that this invention does not require targeted training of the model. That is, depending on the set number of single-class images (C), this invention automatically analyzes the most obvious image categories that can divide the image set into three classes and clusters the images based on features. For example, for a set of face images of all ages and genders, if C=2, this invention may classify them according to male or female; if C=3, this invention will classify them according to age (teenager, middle-aged, elderly). When C=2, if images are difficult to distinguish in terms of gender, this invention will classify them as "indistinguishable between male and female," thereby reducing the high error rate that might result from forced classification. When C=3, in addition to the three precise categories of "teenager," "middle-aged," and "elderly," this technical solution will classify images with indistinct features into several imprecise categories such as "middle-aged or elderly," "teenager or middle-aged," "teenager or elderly," and "age cannot be determined."

[0103] In addition, to verify the effectiveness of the clustering method, it was tested on the Foresttype dataset, which contains training and testing data from a remote sensing study that used ASTER satellite imagery to map different forest types based on spectral characteristics from visible to near-infrared wavelengths. The output (forest type map) can be used to identify and quantify the ecosystem services (e.g., carbon storage, erosion protection) provided by forests.

[0104] During the testing process, metrics such as ACC (Accuracy), Purity, F-Score, Precision, and RI (Rand Index) were used. The meanings of each metric are as follows:

[0105] ACC represents the accuracy of each clustering method. It is a performance metric used to evaluate supervised learning models and measures the proportion of samples correctly predicted by the classification model. The value of ACC ranges from 0 to 1. The closer the value is to 1, the higher the consistency between the clustering results and the true labels, that is, the better the clustering effect.

[0106] Purity measures the degree to which samples within each cluster are correctly assigned to the same category in the clustering results. The value of Purity ranges from 0 to 1. The closer Purity is to 1, the better the clustering result is, that is, the more likely the samples within the cluster are to be correctly assigned to the same category.

[0107] F-Score is a metric that combines precision and recall to measure the accuracy of a classification model. The value of F-Score usually ranges from 0 to 1. The closer the F-Score is to 1, the better the clustering result.

[0108] Precision is a metric used in evaluating classification models. It measures the ratio between the number of samples correctly predicted as positive and the total number of samples classified as positive. Precision ranges from 0 to 1. If all samples in a cluster belong to the same class, then Precision will be 1; if some samples are incorrectly assigned to the cluster, Precision will decrease.

[0109] Rand Index (RI) is a metric used in clustering algorithms to evaluate clustering results. It compares the similarity between the clustering results and the true labels of the dataset. It examines whether the clustering algorithm groups similar samples into the same cluster, thus measuring the consistency of the clustering results. The RI value ranges from 0 to 1. An RI of 1 indicates perfect consistency between the clustering results and the true labels; a RI of 0 indicates no correlation between the clustering results and the true labels if the clustering results and the true labels are randomly assigned. The closer the RI value is to 1, the higher the consistency between the clustering results and the true labels.

[0110] Table 1

[0111]

[0112] As shown in Table 1, MvDKE represents the clustering method of the present invention. As can be seen from Table 1, the accuracy of the method of the present invention is greatly improved compared with many existing clustering methods.

[0113] In summary, this invention achieves multi-view image clustering and characterizes target imprecision in the clustering results. By characterizing imprecision, the error rate of the clustering results can be effectively reduced. A specific dynamic clustering framework is designed for multi-view data to eliminate interference from irrelevant composite classes, reduce redundant computation, and solve the high complexity problem brought about by multi-view evidence clustering and kernel techniques. This improves data processing speed and enables the application of evidence clustering in large-scale multi-view data. By utilizing kernel functions to map multi-view features in a high-dimensional nonlinear space, the nonlinear characteristics of the data can be deeply explored, enabling the application of evidence clustering in multi-view non-spherical distributed data.

[0114] Generalized multi-perspective evidence clustering still faces some practical challenges in real-world applications: First, the introduction of composite classes into the generalized framework significantly increases the computational complexity of existing methods, hindering its practical application in large-scale or multi-class datasets. Second, it struggles with special data structures commonly found in real-world scenarios, such as non-spherical distributions. This invention, while characterizing inaccuracies, can reduce the computational complexity of multi-perspective evidence clustering and improve its overall clustering performance and robustness in real-world data structures like non-spherical distributions, which has significant application value in real-world scenarios.

[0115] In summary, for the full-objective partitioning within the dynamic evidence clustering framework, multi-view fuzzy c-means clustering methods can be used to obtain confidence values ​​(referred to as support in fuzzy clustering) and single-class center points. Adopting this alternative strategy will not significantly impact the results, but may introduce new computational overhead. For initializing the objective function of the full-objective partitioning, other clustering techniques such as FCM, K-means, and spectral clustering can be used for each viewpoint to determine preliminary clustering results, thereby generating more reasonable initial single-class center points. This strategy can reduce the number of subsequent iterations, but will generate new computational overhead during the initialization phase. For the viewpoint weight adaptive variation method used in full image partitioning, a full-view equal-weight strategy can be used instead: w q = 1 / Q. If the above alternative strategy is adopted, the computational cost caused by weight adaptation during iteration can be reduced, but it cannot differentiate data from different perspectives. In scenarios with large differences in information content from different perspectives, this will lead to a decrease in the accuracy of clustering results. For the objective function of partial target re-partitioning in the dynamic evidence clustering framework, a perspective weight similar to that of full target partitioning can be used instead of multi-perspective category dispersion. Adopting this alternative strategy can slightly reduce computational complexity, but it is difficult to mine local features of different perspectives, leading to a decrease in the accuracy of clustering results.

[0116] The present invention also discloses a target classification device based on evidence multi-view kernelization, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-described method.

[0117] The present invention also discloses an embodiment that provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps in the above-described method embodiments.

[0118] The present invention also provides a computer program product that, when run on a data storage device, enables the data storage device to implement the steps in the above-described method embodiments.

[0119] If the integrated unit module is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of the present invention can be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include at least: any entity or device capable of carrying computer program code to a storage device, a recording medium, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium. Examples include USB flash drives, portable hard drives, magnetic disks, or optical disks.

[0120] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0121] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed in this invention can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0122] In the embodiments provided by this invention, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative. For instance, the division of modules or units is merely a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed.

[0123] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

Claims

1. A target classification method based on multi-perspective evidence kernelization, characterized in that, Includes the following steps: Obtain feature information of different image attributes corresponding to the image set to be classified, as well as the number of image categories corresponding to the image set to be classified; Initialize the similarity between each image to be classified and each image category in different image attributes, and calculate the trust value of the image to be classified belonging to each image category in different image attributes based on the similarity; The attribute weights of the image attributes are calculated based on the trust value; The classification error of the image to be classified is calculated using the attribute weight, trust value, and similarity. When the change in the classification error is less than or equal to the error threshold, the category of the image to be classified in the image set is determined based on the trust value. The method for calculating the classification error is as follows: ， Where Q represents the number of image attributes. , Let represent the attribute weight of the q-th image attribute, and N represent the number of images to be classified. , , C represents the number of image categories. , This represents the trust value that the i-th image belongs to the j-th image category in the q-th image attribute. This indicates the correlation factor between trust value and similarity. This represents the similarity between the i-th image and the j-th image category in the q-th image attribute. A factor representing the correlation between image attributes and attribute weights. This represents the similarity between the s-th image and the k-th image in the q-th image attribute. The difference factor representing the attribute weights of different image attributes; Based on the similarity, the trust value of the image to be classified belonging to each image category in different image attributes is calculated, including: ； Determining the category of the images to be classified in the image set based on the trust value includes: Select the highest trust value among the trust values ​​of the image to be classified in different image attributes; When there is no such thing as When a trust value is established, the image to be classified is assigned to the image category corresponding to the maximum trust value; wherein, This represents the maximum trust value of image i to be classified. Let represent the confidence value of image i in relation to the j-th single class of image categories. This represents the threshold for the trust value gap.

2. The target classification method based on multi-perspective evidence kernelization as described in claim 1, characterized in that, When the change in classification error exceeds the error threshold, the similarity is updated based on the trust value, and the calculation continues.

3. The target classification method based on multi-perspective evidence kernelization as described in claim 2, characterized in that, When existence makes When the trust value is established: Reclassify the images to be classified.

4. The target classification method based on multi-perspective evidence kernelization as described in claim 3, characterized in that, Reclassification of images to be classified includes: Choose to Establish a single image class and construct a set of reclassified image categories; The images to be reclassified are classified based on the set of reclassified image categories.

5. The target classification method based on multi-perspective evidence kernelization as described in claim 4, characterized in that, Classification of images to be reclassified based on the reclassified image category set includes: Calculate the second similarity between the image to be classified and each image category in the set of image categories for reclassification based on the similarity; The separation degree of each image category in the reclassified image category set is calculated based on the similarity among different image attributes; The second trust value of the reclassified image for different image categories is calculated based on the separation degree and the second similarity. The image category of the reclassified image is determined based on the second trust value.

6. The target classification method based on multi-perspective evidence kernelization as described in claim 5, characterized in that, The image category for reclassification based on the second trust value includes: ， in, This indicates that the i-th image belongs to the i-th image. The second trust value for each image category, This indicates that in the q-th image attribute, the i-th image belongs to the q-th image attribute. The second trust value for each image category, In the q-th image attribute, the first... Normalized separation of each image category.

7. A target classification device based on multi-view evidence kernelization, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method as described in any one of claims 1-6.

Images