Image noise marker feature selection methods, systems, storage media, and computers

Abstract

This invention provides a method, system, storage medium, and computer for selecting image noise-labeled features. The method includes: acquiring an image noise-labeled dataset to be processed; embedding a sample set from the image noise-labeled dataset into a multi-granularity fuzzy cluster to construct a dynamic fuzzy membership evaluation matrix; dynamically evolving a multi-level high-precision sub-granular sphere cluster; obtaining a highly discriminative label distribution; constructing a coarse-perceptual feature evaluation framework based on sphere topology-driven methods, combining upper and lower approximations of rough sets and extended positive domain models to extract decision equivalence classes; based on a dependency quantification model, fusing multi-sphere decision boundary information to determine the contribution of each feature to the decision system; introducing a sphere structure consistency verification mechanism, and evaluating the importance of features at multiple levels through dependency and consistency. This invention obtains an optimal feature subset with strong noise resistance and high discriminative ability, providing stable and efficient input support for subsequent image noise-labeled learning models.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, in particular to an image noise label feature selection method, system, storage medium and computer. BACKGROUND

[0002] With the rapid development of large-scale multimedia data and computer vision technology, automatic annotation of image data has become a key task in practical applications. Due to the dependence on crowd-sourced annotation, online scraping or weakly supervised information to generate labels, there are often problems such as omission, error or redundancy in the labels of image data, forming a significant noise label phenomenon. Especially in the task of bird image recognition and classification, due to the large number of bird species, high similarity in appearance and complex habitat, the bird image labels obtained through the crowd-sourcing platform often have noise problems such as species mislabeling, subspecies confusion and juvenile bird label errors. These noise labels can interfere with the training process of classification, retrieval and other models, and affect the accuracy of applications such as bird protection, ecological monitoring and intelligent bird watching. Therefore, how to extract accurate and effective label information from noisy bird image data is a technical problem that needs to be solved in the field of weakly supervised learning.

[0003] To reduce the negative impact of noise labels on bird image learning tasks, one of the key technical problems is to perform effective feature selection on noisy bird image data. By extracting high-dimensional features from the original bird image data and then selecting the most relevant discriminative features to the true semantics using a feature selection method, the performance of subsequent visual tasks such as classification, retrieval or target detection can be significantly improved. However, existing bird image feature selection methods often rely on a single correlation measure or assume that the labels are completely accurate, making it difficult to adapt to the feature-label dependence distortion and label confusion problems caused by label noise.

[0004] At the same time, image data itself usually has complex nonlinear distribution characteristics and multi-level local structure characteristics. Traditional feature selection methods often ignore the local consistency and structure information between samples, resulting in insufficient resistance of the selected features to noise labels, thereby affecting the generalization ability of the model. Therefore, how to combine multi-level granular structure evolution and fuzzy label representation, fully utilize the topological properties and local stability of image data, and perform robust feature selection on noisy bird image samples has become a core technical difficulty that needs to be broken through in the field of image noise label learning. SUMMARY

[0005] Therefore, the purpose of the present application is to provide an image noise label feature selection method, system, storage medium and computer to at least solve the deficiencies in the above technical field.

[0006] The present application provides an image noise label feature selection method, which comprises:

[0007] S1: Obtain the image noise labeled dataset to be processed, wherein the image noise labeled dataset contains both feature space and label space;

[0008] S2: An adaptive multi-core fusion and dynamic parameter optimization mechanism is adopted to embed the sample set in the image noise labeling dataset into a multi-granularity fuzzy cluster to construct a dynamic fuzzy membership evaluation matrix and determine the multi-dimensional fuzzy mapping relationship between quantified sample-cluster and sample-class.

[0009] S3: Based on the global spectrum feature mapping mechanism, a macroscopic coarse-grained particle skeleton is constructed. For particles with insufficient label isomorphism or structural entropy exceeding the threshold, a quality measurement model is integrated for joint evaluation. Through a progressive self-organizing fission algorithm, multi-level high-precision sub-particle clusters are dynamically evolved.

[0010] S4: Within each fine sphere, weight-driven sphere calculation and fuzzy inference are integrated. Weights are automatically assigned based on the sample-sphere center distance, and a category-cluster coupling association matrix is ​​constructed. Through fuzzy synthesis operation, the cluster membership is optimized into category membership, and finally a highly recognizable label distribution is obtained.

[0011] S5: Construct a rough perception feature evaluation framework based on particle topology-driven, define adaptive neighborhoods with particle geometric characteristics, dynamically divide cohesive and non-overlapping local neighborhoods, combine rough set upper and lower approximations and extended positive domain models to extract decision equivalence classes, and based on the dependency metric model, fuse multi-particle decision boundary information to determine the contribution of each feature to the decision system.

[0012] S6: Introduces a particle-sphere structure consistency verification mechanism to measure the expression stability of features in the local domain of particles from the dimension of structure preservation, and to evaluate the importance of features at multiple levels through dependency and consistency.

[0013] Furthermore, step S1 includes:

[0014] Using ResNet to extract features from noise-labeled image data, a dataset is obtained. ,in, For the first The feature vector of each sample For the first A set of candidate labels for each sample, for each candidate label element ,when When, it indicates a mark. For the sample Candidate tags; when When, it indicates a mark. With sample Irrelevant.

[0015] Furthermore, in step S2, a combination of random perturbation initialization and nondeterministic exploration strategies is used to initially partition the sample set in the image noise-labeled dataset. Step S2 includes:

[0016] S21: In a multi-kernel framework, each kernel function is assigned dynamic weights, and an entropy regularization term is introduced into the objective function to constrain the weight distribution.

[0017] ;

[0018] in, Let represent the overall objective function of multi-kernel fuzzy clustering, used to simultaneously optimize clustering quality and the rationality of kernel weight allocation, where Represents the membership matrix. Represents the cluster center matrix. Represents the kernel weight allocation vector; This represents the preset number of clusters; This represents the total number of samples in the dataset. This represents the total number of kernel functions in a multi-kernel framework; in the fuzzy clustering process, Represents the original sample For the The fuzzy membership degree of each cluster, which is between 0 and 1 and satisfies the normalization constraint. , The fuzziness index, typically greater than 1, is used to control the degree of fuzziness of cluster boundaries; in multi-kernel mapping mechanisms, Representing the The dynamic weight coefficients of each kernel function, the weights satisfying and Constraints; Indicates the first Each kernel function will convert the original sample The feature vector obtained after mapping to a high-dimensional feature space; Indicates the first The clustering in the th case is at the th case. Cluster centers in the feature space induced by each kernel function; distance term Calculate the squared Euclidean distance between a sample and its corresponding cluster center in a specific kernel space to measure the similarity between the sample and the cluster center; The regularization parameter controls the uniformity of the kernel weight distribution; the entropy term... As a constraint, the more uniform the distribution of kernel weights, the larger the negative value of this term.

[0019] S22: By dynamically adjusting the degree to which samples belong to each cluster, the rationality of fuzzy partitioning is ensured.

[0020] ;

[0021] in, Indicates the first The clustering in the th case is at the th case. Cluster centers in the feature space induced by each kernel function; S23: weights The assignment is based on the contribution of the current kernel function to the clustering, and the regularization parameter Controlling the uniformity of weight distribution:

[0022] ;

[0023] in, Indicates the first The clustering in the th case is at the th case. S24: Update the cluster centers in the feature space induced by the kernel function; update the cluster centers using the mean of the membership-weighted kernel mapping data to ensure the feasibility of the algorithm in the feature space.

[0024] ;

[0025] S25: After the update is completed, calculate the deviation between the result of this iteration and the result of the previous iteration. If the deviation is less than a preset threshold... If the condition is met, the algorithm terminates; otherwise, return to step S21 and repeat until the stopping condition is met.

[0026] or ;

[0027] in, and Indicates the first The set of all cluster centers generated in the next iteration and Indicates the first The set of all cluster centers generated in the next iteration The pre-set convergence threshold.

[0028] Furthermore, step S3 rapidly generates initial macroscopic particles based on full data spectral information. For particles with complex distributions or purity constraints, a dual-mode evaluation mechanism is embedded for real-time screening. Then, a progressive self-organizing fission algorithm is used to decompose them into finer sub-particle clusters. Specific steps include:

[0029] S31: Transfer the dataset As the initial macroscopic particles, in the dataset The binary clustering algorithm is executed to obtain two initial spheres. and , where each sphere The formula for calculating the center is:

[0030] ;

[0031] wherein, is the number of samples in the granule;

[0032] S32: Each granule The calculation formula of the radius is:

[0033] ;

[0034] S33: Check whether the above granule reaches the threshold Requirements:

[0035] ;

[0036] wherein, wherein, denotes the number of samples in the granule , is a candidate marker set, denotes the marker set of the sample ; denotes the information entropy of the granule , which is used to measure the uncertainty of the internal markers of the granule, wherein, denotes the probability of the marker appearing in the granule ; the parameter adjusts the weight ratio of purity and entropy, and when the value is lower than the preset threshold, it indicates that the marker consistency of the granule is insufficient or the topological structure is too complex; If the purity of the granule is lower than the preset threshold, perform binary clustering to split it into two sub-granules with higher consistency, recursively until the purity of all granules meets the standard, and finally complete the construction of the granule.

[0037] Further, the step S4, inside each hierarchical granule, according to the geometric situation of the granule, fuses the distance measurement strategy triggered by multiple source weights, constructs a cluster-class tensor mapping framework, realizes the adaptive mapping fusion of sample-cluster membership to sample-class membership, and finally outputs a highly recognized semantic disambiguation distribution. The specific steps include:

[0038] S41: Construct a hybrid distance measurement algorithm that fuses local geometric structure and dynamic feature importance to obtain the weight information of the sample:

[0039] S41: Construct a hybrid distance measurement algorithm that fuses local geometric structure and dynamic feature importance to obtain the weight information of the sample:

[0040] ;

[0041] wherein, is a dynamic feature weight function, is the weight of the i-th feature​ the mean value of the feature, is the sensitivity coefficient, denotes the sample In the first the value of the feature on the dimension, is the reference center value of the feature in the dimension, is the standard deviation of the feature in the dimension within the granule, eliminating the influence of dimensional difference; is the hyperbolic tangent function, whose value range is [0, 1], and the parameter is an adjustment parameter, used to control the smoothness of the density value to weight mapping, is a local density estimation based on the nearest neighbor; S42: On the basis of having obtained the fuzzy membership matrix and the sample weight

[0042] , a class-cluster association matrix is constructed:

[0043] ;

[0044] The element in the matrix represents the association strength of the first class and the first cluster. By using the fuzzy logic reasoning method, the association matrix is subjected to fuzzy composition operation with the membership of the sample to each cluster , so as to realize the conversion of the membership of the sample to the cluster to the membership to the class. Further, the step S5 constructs a rough perception feature evaluation framework based on granular topology driving. The marked distribution space is adaptively divided into decision classes by granular computation, and the feature dependency is measured by granular neighborhood rough set. The specific steps include:

[0045] S51: For any sample

[0046] , define its adaptive neighborhood under the feature subset as:

[0047] ;

[0048] wherein and are the center and radius of the of the granule, is the distance under the feature subset .​​​

[0049] S52: Constructing dynamic marked granular ball group set, through adaptive granular ball splitting algorithm, performing multi-scale nonlinear spatial reconstruction on real marked global space:

[0050] ;

[0051] Each represents a marked granular ball, which collectively covers the entire marked space, and each granular ball reaches the optimal purity and coverage;

[0052] S53: In a given decision system , the decision attribute divides the sample into equivalence classes, and the granular ball set generated by the feature subset defines the upper and lower approximation sets:

[0053] , ;

[0054] where, , through the non-empty intersection relationship between granular ball and equivalence class , determine the coverage boundary of the decision class;

[0055] ;

[0056] Through the full inclusion relationship between granular ball and equivalence class , extract the core support domain of the decision class;

[0057] S54: Constructing an extended positive region dependence quantification model for image noise marked decision system:

[0058] ;

[0059] where, is the sample that belongs to both and , represents the sample set in but not in , is the distance from the sample to the center of the granular ball, is the distance from the sample to the center of the granular ball;

[0060] S55: Through dynamic multi-granularity collaborative feature importance quantification algorithm, integrating rough set theory and granular ball calculation framework, realize the accurate evaluation of features: ​

[0061] .

[0062] Further, the step S6 is robust in feature selection from the dual-modal perspective, and the specific steps include:

[0063] S61: measure the similarity of granule samples based on a similarity quantization algorithm of sample structure consistency within the granule:

[0064] ;

[0065] Wherein, The cardinality of the granule set;

[0066] S62: based on the additional feature importance evaluation mechanism, the discriminant contribution of the feature is quantified by measuring the degree of change of the granule structure before and after the feature is removed:

[0067] ;

[0068] S63: by constructing the feature dependency matrix and the sample consistency matrix within the granule, the method of jointly evaluating the feature weight by multi-source dual-modal dependency-consistency is adopted to realize the fine quantization and sorting of the key features in high-dimensional data:

[0069] .

[0070] The application also provides an image noise label feature selection system, comprising:

[0071] A data acquisition and processing module is configured to acquire an image noise label data set to be processed, wherein the image noise label data set contains a feature space and a label space;

[0072] A fuzzy clustering dynamic division module is configured to embed a sample set in the image noise label data set into a multi-granularity fuzzy cluster group by using an adaptive multi-kernel fusion and dynamic parameter optimization mechanism, to construct a dynamic fuzzy membership evaluation matrix, and to determine a multi-dimensional fuzzy mapping relationship between the sample-cluster and the sample-class;

[0073] A multi-scale granule hierarchical construction module is configured to construct a macroscopic coarse-granularity granule skeleton based on a global spectral feature mapping mechanism, to fuse a quality measurement model into a granule with insufficient label isomorphism or structural entropy exceeding a threshold value for joint evaluation, and to dynamically evolve a multi-level high-precision sub-granule cluster by using a progressive self-organizing fission algorithm;

[0074] The coupled granule label disambiguation module is used for fusing weight-driven granule calculation and fuzzy reasoning in each fine granule, automatically assigning weights according to sample-granule center distance, and constructing a class-cluster coupled correlation matrix, and finally obtaining a high-resolution label distribution through fuzzy synthesis operation.

[0075] The granule-rough set collaborative feature evaluation module is used for constructing a rough set perception feature evaluation framework based on granule topology driving, defining an adaptive neighborhood based on granule geometric characteristics, dynamically dividing local neighborhoods that are cohesive and do not overlap, extracting decision equivalence classes in combination with rough set upper and lower approximation and extended positive region model, determining the contribution of each feature to the decision system based on a dependence quantification model and fusing multi-granule decision boundary information.

[0076] The multi-granularity feature stability verification module is used for introducing a granule structure consistency verification mechanism, measuring the expression stability of the feature in the local domain of the granule from the structure preservation dimension, and performing multi-level evaluation on the importance of the feature through dependence and consistency.

[0077] The application further provides a storage medium having a computer program stored thereon, and the computer program is executed by a processor to implement the image noise label feature selection method.

[0078] The application further provides a computer comprising a memory, a processor and a computer program stored on the memory and executable on the processor, and the processor implements the image noise label feature selection method when executing the computer program.

[0079] The image noise label feature selection method, system, storage medium and computer in the application adopt a fuzzy membership degree driven sample soft partition mechanism to initialize fuzzy clusters, and on this basis, introduce a quality driven granule evolution strategy to construct a multi-granularity granule system from coarse to fine, realize dynamic disambiguation and noise interference suppression of the label space through progressive self-organizing fission, fuse granule geometric topology and rough set approximation theory to define an adaptive neighborhood range, use the upper and lower approximation set and the extended positive region model to depict the decision equivalence class boundary, and based on the dependence measurement model of multi-granule boundary fusion, realize high-precision quantitative evaluation of feature importance, construct a granule structure stability verification mechanism, evaluate the expression consistency and generalization ability of the selected features at different granularity levels from the two dimensions of sample similarity and feature disturbance response in the granule, and improve the robustness and reliability of feature selection, design a hierarchical weight integration strategy to weight and fuse the feature ranking results at different granularities, and finally obtain an optimal feature subset with strong noise resistance and high discrimination ability, providing stable and efficient input support for subsequent image noise label learning models. BRIEF DESCRIPTION OF DRAWINGS

[0080] Figure 1 Flow chart of image noise label feature selection method in first embodiment of the present application;

[0081] Figure 2 Frame diagram of image noise label feature selection method in first embodiment of the present application;

[0082] Figure 3 Structure block diagram of image noise label feature selection system in second embodiment of the present application;

[0083] Figure 4 Structure block diagram of computer in third embodiment of the present application.

[0084] The following detailed description will further illustrate the present application in conjunction with the above-mentioned drawings. DETAILED DESCRIPTION

[0085] In order to facilitate the understanding of the present application, the present application will be described more fully below with reference to the accompanying drawings. The drawings show several embodiments of the present application. However, the present application can be implemented in many different forms and is not limited to the embodiments described herein. On the contrary, the purpose of providing these embodiments is to make the disclosure of the present application more thorough and comprehensive.

[0086] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the present application belongs. The terminology used in the description of the present application herein is for the purpose of describing the specific embodiments only and is not intended to limit the present application. The term "and / or" used herein includes any and all combinations of one or more of the associated listed items.

[0087] Embodiment one

[0088] Please refer to Figure 1 , which shows the image noise label feature selection method in the first embodiment of the present application, which specifically includes steps S1 to S6:

[0089] S1: obtaining the image noise label data set to be processed, wherein the image noise label data set contains both feature space and label space;

[0090] In specific implementation, the frame diagram of the bird image noise label feature selection method provided in the present embodiment is as shown in Figure 2 The pre-trained ResNet network is used to extract deep features of the bird image, and the image data is converted into a structured multi-dimensional feature vector, forming a data set that covers both image feature space and candidate label space;

[0091] Specifically, let Examples of a sample set in a feature space, let denote a label space containing possible labels. For a given noisy-labeled bird image dataset where is the embedding vector of the th sample in the feature space, is the candidate label set associated with the sample. The candidate label set can be mixed with true labels and noise labels, typically characterized by label uncertainty. For any candidate label , if , it indicates that the label is a candidate label for the sample ; if , it indicates that the label is irrelevant to the sample .

[0092] S2: Embed the sample set in the image noisy-labeled dataset into multi-granularity fuzzy cluster groups using an adaptive multi-kernel fusion and dynamic parameter optimization mechanism to construct a dynamic fuzzy membership evaluation matrix and determine the multi-dimensional fuzzy mapping relationship between the quantized sample-cluster and sample-class;

[0093] In specific implementation, this step faces the distribution characteristics of bird image features (such as feather texture details, feather color texture details, beak contour, background environment, etc.), and uses an adaptive multi-kernel fusion and dynamic parameter optimization mechanism to perform prior subspace division on the sample set, thereby constructing multiple original semantic cluster centers. This mechanism can automatically adjust the kernel function parameters according to the nonlinear distribution of bird images in the feature space, effectively capturing feature differences caused by changes in lighting, differences in posture, occlusion, and background diversity. Subsequently, through the fusion of fuzzy set theory, a soft attachment relationship between samples and cluster centers is realized, so that the gradual aggregation of true species semantics can still be maintained in the presence of label noise. The step S2 specifically includes:

[0094] Step S21: In the multi-kernel framework, a dynamic weight is given to each kernel function, and an entropy regularization term is introduced into the objective function to constrain the weight distribution:

[0095] ;

[0096] wherein, denotes the overall objective function of multi-kernel fuzzy clustering, which is used to simultaneously optimize the clustering quality and the rationality of kernel weight distribution, wherein, denotes the membership matrix, denotes the cluster center matrix, denotes the kernel weight distribution vector; denotes the preset number of clustering clusters; denotes the total number of samples in the data set; denotes the total number of kernel functions in the multi-kernel framework; in the fuzzy clustering process, denotes the original sample (bird image sample) the fuzzy membership of the first cluster, the value is between 0 and 1, and satisfies the normalization constraint condition , is a fuzzy index, usually greater than 1, used to control the fuzziness of the cluster boundary;

[0097] In the multi-kernel mapping mechanism, denotes the dynamic weight coefficient of the first kernel function, and the weight satisfies and constraint condition; denotes the feature vector obtained after the first kernel function maps the original sample (bird image sample) to a high-dimensional feature space; denotes the cluster center of the first cluster in the feature space induced by the first kernel function; the distance term calculates the Euclidean distance square of the sample in the specific kernel space and the corresponding cluster center, which is used to measure the similarity between the sample and the cluster center;

[0098] In the bird image with noisy labels, different kernel functions are designed to capture different types of visual features: color kernel extracts feather color distribution, eye circle color, etc. Features are used to distinguish species with significant color differences; texture kernel captures feather texture, wing spot details, etc. Information should be used to deal with the situation where the feather color is similar but the texture is obviously different; shape kernel analyzes the beak shape geometry, body contour and wing shape, etc. Used to distinguish species with large differences in body characteristics; multi-scale kernel captures detailed features under different shooting distances, poses and scales.

[0099] The second part of the objective function introduces an entropy regularization mechanism, where is a regularization parameter used to control the uniformity of the kernel weight distribution; the entropy term as a constraint term, the greater the negative value of this term when the kernel weight distribution is more uniform, thereby preventing the algorithm from degenerating into a single kernel function, ensuring the effectiveness of multi-kernel fusion. The whole objective function realizes the optimization of sample clustering in multi-kernel space and the adaptive adjustment of kernel weight by minimizing the weighted clustering error and maximizing the entropy constraint term;

[0100] Step S22: By dynamically adjusting the degree of membership of each cluster to the sample, the rationality of fuzzy division is ensured:

[0101] ;

[0102] wherein, denotes the cluster center of the i-th cluster in the feature space induced by the j-th kernel function;

[0103] Step S23: The kernel weight update formula dynamically allocates the importance of different kernel functions through the entropy regularization term. The allocation of the weight is based on the contribution of the current kernel function in the cluster, and the regularization parameter controls the uniformity of the weight distribution:

[0104] ;

[0105] wherein, denotes the cluster center of the i-th cluster in the feature space induced by the j-th kernel function;

[0106] Step S24: In the kernel space, the cluster center cannot be explicitly expressed and needs to be implicitly calculated through the kernel matrix. The membership degree weighted kernel mapping data mean is used to update the cluster center, ensuring the feasibility of the algorithm in the feature space:

[0107] ;

[0108] This process can make the cluster centers of bird images in multiple feature modalities closer to the visual feature distribution of real species;

[0109] Step S25: After updating, the new membership matrix and the deviation between the previous membership matrix and the new center and the previous center are calculated in turn. If the deviation is less than the preset threshold , the algorithm terminates; otherwise, return to step S21 and repeat execution until the stop condition is met:

[0110] or ;

[0111] wherein, and denote the set of all cluster centers generated in the i-th iteration, and denote the set of all cluster centers generated in the i-th iteration, ​​​​​​The preset convergence threshold is set according to the complexity of the bird image feature distribution (such as species diversity and background interference intensity).

[0112] S3: Based on the global spectrum feature mapping mechanism, a macro coarse-grained granule skeleton is constructed, the granules with insufficient marking isomorphism or structural entropy exceeding the threshold are fused into a quality measurement model for joint evaluation, and a multi-level high-precision sub-granule cluster is dynamically evolved through a progressive self-organizing fission algorithm.

[0113] In specific implementation, this step is based on the distribution characteristics of bird image features in high-dimensional space (for example, the appearance difference of the same species under different postures, light, and habitat environment) to quickly generate an initial macro granule based on full data spectrum information, and a double-mode evaluation mechanism is embedded for real-time screening of granules containing high-noise labels or mixed species. Through a progressive self-organizing fission algorithm, complex granules are gradually decomposed into high-purity sub-granules that can distinguish fine-grained species characteristics, thereby improving the ability to distinguish subspecies differences and gender differences. The specific steps include:

[0114] Step S31: The data set is divided into two initial granules and by performing a binary clustering algorithm on the data set Step S32: The radius of each granule is calculated according to the following formula: wherein, the calculation formula of the center of each granule is as follows:

[0115] ;

[0116] wherein, n is the number of samples in the granule, and

[0117] Step S32: The radius of each granule is calculated according to the following formula:

[0118] ;

[0119] Step S33: Check whether the above granules meet the threshold value requirement:

[0120] ;

[0121] wherein, n is the number of samples in the granule, and ​​​​​​​​​​The purity value is a set of markers. The formula measures the concentration of the marker distribution by calculating the maximum frequency of each marker in the sample within the sphere. The higher the purity value, the more consistent the marker distribution of the samples within the sphere. For example, when most images in the same sphere are labeled as "light-vented bulbul" and have consistent feather features, the purity is high. Indicates granules The information entropy is used to measure the uncertainty of the markings inside the granule, where... Indicates a marker In granules The probability of occurrence within a granule is calculated using this formula, which measures the degree of uncertainty in the distribution of markers within the granule. For example, when different species with similar plumage (Grey-headed Bulbul and White-headed Bulbul) are mixed in the same granule, the entropy value is higher. A higher entropy value indicates a more complex and chaotic distribution of markers within the granule. The parameter adjusts the weight ratio of purity and entropy, when When the value is lower than the preset threshold, it indicates that the labeling consistency of the granule is insufficient or the topology is too complex, and further splitting is required to improve the quality of the granule.

[0122] If the purity of a particle is lower than a preset threshold, binary clustering is performed to split it into two more consistent sub-particles. This process is recursively repeated until the purity of all particles meets the threshold, thus completing particle construction. This process is iterated until the purity of all particles meets the threshold requirement, at which point the splitting process terminates, completing adaptive particle construction.

[0123] S4: Within each fine sphere, weight-driven sphere calculation and fuzzy inference are integrated. Weights are automatically assigned based on the sample-sphere center distance, and a category-cluster coupling association matrix is ​​constructed. Through fuzzy synthesis operation, the cluster membership is optimized into category membership, and finally a highly recognizable label distribution is obtained.

[0124] In specific implementation, within each level of the granular sphere, based on the granular sphere geometric posture of bird image features, a multi-source weight-triggered distance metric strategy (such as color difference, texture difference, and shape difference) is fused to construct a cluster-class tensor mapping framework. This achieves adaptive mapping and fusion of sample-cluster membership to sample-class membership, thereby optimizing label disambiguation in species identification. Step S4 specifically includes:

[0125] Step S41: Based on the core idea of ​​multi-source weight triggering, and combined with the ideas of manifold learning and adaptive feature weighting, a hybrid distance metric algorithm that integrates local geometric structure and dynamic feature importance is constructed to obtain the weight information of samples. By introducing a third-order nonlinear control mechanism, the scene adaptability of the distance metric is significantly improved.

[0126] ;

[0127] in, , is the dynamic feature weight function. is the dimensional mean value of the dimensional feature, is the sensitivity coefficient, represents the sample value on the dimensional feature, is the reference center value of the dimensional feature in the granule, is the standard deviation of the dimensional feature in the granule, eliminating the influence of dimensional difference; is the hyperbolic tangent function, whose value range is [0, 1], and the parameter is the adjustment parameter, used to control the smoothness of the density value to weight mapping, is the local density estimation based on the nearest neighbor;

[0128] Step S42: On the basis of having obtained the fuzzy membership matrix and the sample weight , a category-cluster association matrix is constructed:

[0129] ;

[0130] wherein the initial value of the matrix A is a zero matrix, and the dimension is , is the number of categories, is the number of clusters, and the element in the matrix represents the association strength of the th category and the th cluster. By using the fuzzy logic reasoning method, the association matrix is subjected to fuzzy composition operation with the sample membership of each cluster, so as to realize the conversion of the sample membership of the cluster to the membership of the category, thereby reducing the marking confusion between the species with similar colors. S5: A rough perception feature evaluation framework based on granule topology driving is constructed, the adaptive neighborhood is defined according to the granule geometric characteristics, the local neighborhood which is cohesive and non-overlapping is dynamically divided, the decision equivalent class is extracted by combining the rough set upper and lower approximation and the extended positive region model, the contribution degree of each feature to the decision system is determined by fusing the multi-granule decision boundary information based on the dependence quantification model;

[0131]

[0132] In specific implementation, a coarse perception feature evaluation framework based on particle topology was constructed. Relying on the geometric topological information of the particles, an adaptive neighborhood for each sample was dynamically defined, achieving multi-dimensional partitioning of the label distribution space. Using extended upper and lower approximate boundaries and a positive domain extension model, a refined characterization of decision equivalence classes was completed. Based on a dynamic dependency measurement model, the dependency of each feature in different decision classes was calculated in real time, thereby accurately identifying and ranking discriminative features. Step S5 specifically includes:

[0133] Step S51: Let the complete sample set be The complete set of features is Select a subset of attributes as .right The particle sphere calculation is applied to obtain the set of particles in the characteristic space. For any sample Define it in the feature subset The adaptive neighborhood is as follows:

[0134] ;

[0135] in, and It is a pellet. The center and radius, For feature subset The distance below, defined by the center and radius of the sphere, dynamically divides the sample into “cohesive” and “non-overlapping” local neighborhoods, avoiding the rigidity of fixed radius neighborhoods and ensuring that samples within the neighborhood maintain a high degree of consistency in species characteristics;

[0136] Step S52: Construct a dynamic labeled spheroid group set, and use an adaptive spheroid splitting algorithm to split the real labeled global domain. Multi-scale nonlinear spatial reconstruction is performed to form an optimal granular topological partition with optimal coverage of the entire domain:

[0137] ;

[0138] Each This represents a single marked sphere that collectively covers the entire marked space, with each sphere achieving optimal purity and coverage.

[0139] Step S53: In the given decision system In the middle, decision attributes Divide the samples into equivalence classes, using feature subsets The generated set of spheres Define upper and lower approximate sets:

[0140] , ;

[0141] wherein, , by granular ball and the non-empty intersection relation of equivalence class , determine the coverage boundary of the decision class;

[0142] ;

[0143] by granular ball and the total inclusion relation of equivalence class , extract the core support domain of the decision class. In the bird application, the upper approximation set usually contains typical samples with completely consistent feather color, texture and morphological characteristics, such as high-definition photos of adult male birds; the lower approximation set contains boundary samples that are easy to be confused with other species, such as juvenile birds, intermediate individuals with feather color, or photos with partial occlusion;

[0144] Step S54: Because the label space of the multi-label data set itself has inherent ambiguity, it is difficult to clearly segment the decision boundary, and the lower approximation set alone cannot fully reflect the contribution of boundary samples to the decision. Therefore, samples outside the lower approximation set but still have "weak association" or "partial overlap" with a certain decision class are also included in the measurement of the positive domain, so as to more comprehensively characterize the multi-label decision boundary. An extended positive region dependence quantification model for image noise label decision system is constructed:

[0145] ;

[0146] wherein, is the sample that belongs to and , and represents the sample set in but not in , is the distance from the sample to the center of the granular ball, is the distance from the sample to the center of the granular ball;

[0147] Step S55: Through the dynamic multi-granularity collaborative feature importance quantification algorithm, the rough set theory and the granular ball calculation framework are fused to realize the accurate evaluation of the features:

[0148] .

[0149] Specifically, the multi-granularity decision boundary information is fused to output a ranking of the contribution of each feature to the classification of birds. In actual deployment, the ranking result can be used to prioritize the features with the highest contribution to species discrimination and the least sensitivity to environmental changes for model training and updating of an automatic bird recognition system in the wild. S6: A particle ball structure consistency verification mechanism is introduced to measure the expression stability of the features in the local domain of the particle ball from the perspective of structure preservation, and the importance of the features is evaluated at multiple levels through dependency and consistency.

[0150] In specific implementation, the conditional attribute subset from the training sample set The generated particle ball set A similarity quantification method based on the consistency of the internal sample structure of the particle ball is introduced to measure the similarity of the particle ball samples, specifically as follows:

[0151] Step S61: A similarity quantification algorithm based on the consistency of the internal sample structure of the particle ball is used to measure the similarity of the particle ball samples:

[0152] ;

[0153] wherein, indicates the cardinality of the particle ball set, The greater the value, the more consistent the feature distribution of the particle ball is, the higher the similarity between them is, and the clearer the boundary between different particle balls is, proving that the current feature subset has strong discrimination ability and expression effectiveness;

[0154] Step S62: Based on the additional feature importance evaluation mechanism, the discrimination contribution of the features is quantified by measuring the degree of change in the particle ball structure before and after the features are removed:

[0155] ;

[0156] Step S63: By constructing a feature dependency matrix and a particle ball internal sample consistency matrix, a multi-source dual-modal dependency-consistency joint evaluation method is used to evaluate the feature weight, realizing the fine quantification and ranking of the key features in high-dimensional data:

[0157] .

[0158] In summary, the image noise label feature selection method in the above embodiments of the present application adopts a sample soft partition mechanism driven by fuzzy membership, initializes fuzzy clusters, and on this basis introduces a quality-driven granular ball evolution strategy to construct a multi-granularity granular ball system from coarse to fine, and realizes dynamic disambiguation and noise interference suppression of the label space through progressive self-organizing fission; fusion of granular ball geometric topology and rough set approximation theory, define adaptive neighborhood range, use upper and lower approximation set and extended positive domain model to depict the boundary of decision equivalence class, and based on the dependence measurement model of multi-granular ball boundary fusion, realize high-precision quantitative evaluation of feature importance; construct a granular ball structure stability verification mechanism, evaluate the expression consistency and generalization ability of the selected features at different granularity levels from the two dimensions of sample similarity and feature disturbance response in the granular ball, and improve the robustness and reliability of feature selection; design a hierarchical weight integration strategy to weight and fuse the feature ranking results at multiple granularities, and finally obtain an optimal feature subset with strong noise resistance and high discrimination ability, providing stable and efficient input support for subsequent image noise label learning model.

[0159] Embodiment two

[0160] Another aspect of the present application also provides an image noise label feature selection system, please refer to Figure 3 , which is an image noise label feature selection system in the second embodiment of the present application, the system comprises:

[0161] Data acquisition and processing module 201 is used for acquiring image noise label data set to be processed, wherein the image noise label data set contains feature space and label space at the same time;

[0162] Fuzzy clustering dynamic partition module 202 is used for embedding the sample set in the image noise label data set into multi-granularity fuzzy clusters by using adaptive multi-core fusion and dynamic parameter optimization mechanism, so as to construct a dynamic fuzzy membership evaluation matrix and determine the multi-dimensional fuzzy mapping relationship between sample-cluster and sample-class;

[0163] Multi-scale granular ball level construction module 203 is used for constructing a macro coarse-grained granular ball skeleton based on a global spectral feature mapping mechanism, jointly evaluating the granular ball with insufficient label isomorphism or structure entropy exceeding the threshold value, and dynamically evolving multi-level high-precision sub-granular ball clusters through a progressive self-organizing fission algorithm;

[0164] Coupled granular ball label disambiguation module 204 is used for fusing weight-driven granular ball calculation and fuzzy reasoning in each fine granular ball, automatically assigning weights according to sample-granular ball center distance, constructing a class-cluster coupling correlation matrix, optimizing cluster membership to class membership through fuzzy synthesis operation, and finally obtaining a high-resolution label distribution;

[0165] The granular-rough set collaborative feature evaluation module 205 is configured to construct a granular topology driven rough set aware feature evaluation framework, define adaptive neighborhoods based on granular geometric properties, dynamically divide local neighborhoods that are cohesive and non-overlapping, extract decision equivalence classes in combination with rough set upper and lower approximation and extended positive region models, determine the contribution of each feature to the decision system based on a dependency quantification model, and fuse multi-granular decision boundary information.

[0166] The multi-granularity feature stability verification module 206 is configured to introduce a granular structure consistency verification mechanism, measure the expression stability of a feature in a granular local domain from the perspective of structure preservation, and perform multi-level evaluation of the importance of the feature based on dependency and consistency.

[0167] The functions or operation steps realized when the above modules and units are executed are substantially the same as those of the above method embodiments, and thus will not be described here again.

[0168] The image noise marker feature selection system provided in the embodiments of the present application has the same implementation principle and technical effects as the above method embodiments, and for brevity of description, the parts not mentioned in the system embodiment part can be referred to the corresponding contents in the above method embodiments.

[0169] Embodiment Three

[0170] The present application also provides a computer, please refer to Figure 4 , which is a computer in the third embodiment of the present application, comprising a memory 10, a processor 20, and a computer program 30 stored in the memory 10 and executable on the processor 20, wherein the processor 20 implements the above-mentioned image noise marker feature selection method when executing the computer program 30.

[0171] The memory 10 includes at least one type of storage medium, such as flash memory, hard disk, multimedia card, card-type memory (e.g., SD or DX memory, etc.), magnetic memory, disk, optical disk, etc. In some embodiments, the memory 10 can be an internal storage unit of the computer, such as a hard disk of the computer. In other embodiments, the memory 10 can also be an external storage device, such as a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, etc. Further, the memory 10 can include both an internal storage unit and an external storage device of the computer. The memory 10 can be used not only to store application software and various data installed in the computer, but also to temporarily store data that has been output or will be output.

[0172] The processor 20 may, in some embodiments, be an electronic control unit (ECU), a central processing unit (CPU), a controller, a microcontroller, a microprocessor, or other data processing chip, for running program codes or processing data stored in the memory 10, such as executing the access restriction program.

[0173] It should be noted that, Figure 4 The illustrated structure does not constitute a limitation on the computer, which may, in other embodiments, include fewer or more components than shown, or combine certain components, or have different arrangements of components.

[0174] The embodiments of the present application also propose a storage medium having a computer program stored thereon, which, when executed by a processor, implements the image noise marking feature selection method as described above.

[0175] Those skilled in the art can understand that the logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a list of executable instructions for implementing the logic function, which can be specifically implemented in any computer readable medium for use by or in conjunction with an instruction execution system, device or apparatus, such as a computer-based system, a system including a processor or other system that can fetch and execute instructions from the instruction execution system, device or apparatus. For the present specification, the "computer readable medium" can be any device that can contain, store, communicate, propagate or transport programs for use by or in conjunction with the instruction execution system, device or apparatus, or in conjunction with these instruction execution systems, devices or apparatus.

[0176] More specific examples (a non-exhaustive list) of the computer readable medium include the following: an electrical connection having one or more wires (electrical devices), a portable computer diskette (magnetic devices), a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disc read-only memory (CDROM). In addition, the computer readable medium can even be paper or other suitable medium on which the program can be printed, as the program can be electronically obtained, for example, by optical scanning of the paper or other medium, followed by editing, interpreting or otherwise processing the program as necessary, and then storing it in a computer memory.

[0177] It should be understood that portions of the present application can be implemented with hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented with software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, any of the following technologies, known in the art, or a combination thereof, can be used: discrete logic circuitry having logic gates for implementing logic functions upon an application of data signals, application specific integrated circuits having appropriate combinational logic gates, programmable gate arrays (PGA), field programmable gate arrays (FPGA), and the like.

[0178] The technical features of the above-described embodiments can be combined in any manner. For brevity, not all possible combinations of the technical features in the above-described embodiments are described, but it should be understood that any combination of the technical features is within the scope of the present disclosure as long as the combination does not result in a contradiction.

[0179] The above-described embodiments only express several implementation manners of the present application, and the description is specific and detailed, but it should not be understood as a limitation on the scope of the patent. It should be noted that for those skilled in the art, without departing from the concept of the present application, a number of modifications and improvements can be made, which are all within the scope of the present application. Therefore, the scope of the patent of the present application should be subject to the appended claims.

Claims

1. An image noise marker feature selection method characterized by, Comprise: S1: obtain the image noise mark data set to be processed, wherein the image noise mark data set contains feature space and mark space at the same time; S2: adopt adaptive multi-core fusion and dynamic parameter optimization mechanism, embed sample set in multi-granularity fuzzy cluster group in the image noise mark data set, construct dynamic fuzzy membership evaluation matrix, determine multi-dimensional fuzzy mapping relationship between quantitative sample-cluster and sample-class; S3: based on global spectrum feature mapping mechanism, construct macroscopic coarse-grained granular ball skeleton, for granular ball with insufficient mark isomorphism or structure entropy exceeding threshold, embed quality measurement model for joint evaluation, and through progressive self-organizing fission algorithm, dynamically evolve multi-level high-precision sub-granular ball cluster; S4: in each fine granular ball, fuse granular ball calculation and fuzzy reasoning driven by weight, automatically assign weight according to sample-granular ball center distance, and construct class-cluster coupling correlation matrix, through fuzzy synthesis operation, optimize cluster group membership to class membership, and finally obtain high-identification mark distribution; S5: construct rough perception feature evaluation framework based on granular ball topology driving, define adaptive neighborhood according to granular ball geometric characteristics, dynamically divide local neighborhood which is cohesive and does not overlap, combine rough set upper and lower approximation and extended positive region model, extract decision equivalence class, based on dependence measurement model, fuse multi-granular ball decision boundary information, and determine the contribution degree of each feature to decision system; S6: introduce granular ball structure consistency verification mechanism, measure the expression stability of features in local domain of granular ball from structure preservation dimension, and evaluate the importance of features through dependence and consistency in multiple levels.

2. The image noise marker feature selection method of claim 1, wherein, The step S1 comprises: Using ResNet to extract features from given image noise-labeled data, we can obtain features with A dataset of samples ,in, Indicates the first The feature vector of each sample Indicates the first The set of candidate labels for each sample, for Each candidate tag element in ,when When, it indicates the first The first label is the Candidate labels for each sample; when When, it means the first The first tag and the first The individual samples are irrelevant.

3. The image noise marker feature selection method of claim 2, wherein, In step S2, the sample set in the image noise mark data set is preliminarily partitioned by combining random disturbance initialization with non-deterministic exploration strategy, and the step S2 comprises: S21: in the multi-core framework, dynamic weight is given to each kernel function, and entropy regularization term is introduced into the objective function to constrain weight allocation: ； wherein, denotes the overall objective function of multi-kernel fuzzy clustering, which is used to optimize the clustering quality and the rationality of kernel weight distribution simultaneously, wherein, denotes the membership matrix, denotes the cluster center matrix, denotes the kernel weight distribution vector; represents the preset number of clustering clusters; is the total number of samples in the data set; denotes the total number of kernel functions in the multi-kernel framework; in the fuzzy clustering process, denotes the fuzzy membership of the th sample to the th cluster, is between 0 and 1, and satisfies the normalization constraint condition , is a fuzzification index, which is greater than 1, and is used to control the fuzzy degree of the clustering boundary; in the multi-kernel mapping mechanism, represents the dynamic weight coefficient of the th kernel function, and the weight satisfies the constraint condition and ; denotes the feature vector obtained after the th kernel function maps to a high-dimensional feature space; denotes the cluster center of the th cluster in the feature space induced by the th kernel function; the distance term calculates the Euclidean distance square of the sample in the specific kernel space and the corresponding cluster center, which is used to measure the similarity between the sample and the cluster center; is a regularization parameter, which is used to control the uniformity degree of the kernel weight distribution; the entropy term is a constraint term, and the greater the negative value of this term is, the more uniform the kernel weight distribution is; S22: by dynamically adjusting the attribution degree of samples to each cluster, the rationality of fuzzy division is ensured: ； wherein, represents the i-th cluster center in the feature space induced by the i-th kernel function; represents the i-th cluster center in the feature space induced by the i-th kernel function; represents the i-th cluster center in the feature space induced by the i-th kernel function; S23: the weights The allocation of the weights is based on the contribution of the current kernel function in the cluster, the regularization parameter Controlling the uniformity of the weight distribution: ； wherein, represents the cluster center of the th cluster in the feature space induced by the th kernel function; S24: the cluster center is updated by using the mean value of the kernel mapping data weighted by membership, so as to ensure the feasibility of the algorithm in the feature space: ； S25: After the update is completed, the degree of deviation between the result of this iteration and the last iteration is calculated. If the deviation is less than a preset threshold , the algorithm terminates; otherwise, return to step S21 for repeated execution until the stop condition is met: or ; wherein, and respectively represent the membership matrix set and the cluster center matrix set generated in the th iteration, and respectively represent the membership matrix set and the cluster center matrix set generated in the th iteration, is a pre-set convergence threshold.

4. The image noise marker feature selection method according to claim 3, characterized in that, The step S3 generates initial macroscopic granular ball based on full data spectrum information, for granular ball with complex distribution or purity constraint violation, embeds double-mode evaluation mechanism for real-time screening, and through progressive self-organizing fission algorithm, decomposes it into more fine sub-granular ball cluster, and the specific steps comprise: S31: obtaining a data set As initial macro-particle balls, a bisection clustering algorithm is performed on the data set to obtain two initial particle balls and wherein, the calculation formula of the center of each particle ball is: ； wherein, is the number of granulocytic samples; S32: each granulocyte The formula for the radius is: ； wherein, is to the center of the granulocyte Euclidean distance; S33: Check if the above granulocyte reaches a threshold request, wherein the threshold of the first granulocyte is defined as: ​ ； In the formula, ,in, Indicates granules The number of samples in For the candidate tag set, Indicates the first A set of candidate labels for each sample; Indicates granules The information entropy is used to measure the uncertainty of the markings inside the granule, where... Indicates a marker In granules The probability of occurrence in; through The parameter adjusts the weight ratio of purity and entropy, when When the value is lower than the preset threshold, it indicates that the labeling consistency of the particle is insufficient or the topology is too complex. If the purity of the granular ball is lower than the preset threshold, the granular ball is divided into two sub-granular balls with higher consistency by performing binary clustering, and the granular ball construction is finally completed by recursion until the purity of all granular balls meets the standard.

5. The image noise marker feature selection method of claim 4, wherein, In the step S4, according to the geometric situation of the granular ball, the distance measurement strategy triggered by multi-source weight is fused to construct cluster-class tensor mapping framework, realize adaptive mapping and fusion of sample-cluster membership to sample-class membership, and finally output highly recognized semantic disambiguation distribution, and the specific steps comprise: S41: Construct a hybrid distance metric algorithm that fuses local geometric structure and dynamic feature importance, and obtain the weight information of the sample: ； in, , is the dynamic feature weight function. For the first dimensional feature mean, This is the sensitivity coefficient. Indicates the first The sample at the th Values ​​on the dimensional features, For the first Reference center value of dimensional feature For the first granule Standard deviation of dimensional characteristics to eliminate the influence of dimensional differences; The function is the hyperbolic tangent function, and its range is [0,1]. The adjustable parameters control the smoothness of the mapping from density values ​​to weights. Based on Local density estimation of nearest neighbors; S42: On the basis of the obtained fuzzy membership matrix and sample weights , a category-cluster association matrix is constructed: ； matrix elements in Indicates the first The first tag and the first The association strength of each cluster is determined using fuzzy logic reasoning, which correlates the association matrix with... Membership degree of each cluster Perform fuzzy synthesis operation Thus achieving The membership degree of a cluster is transformed into the membership degree of a category.

6. The image noise marker feature selection method of claim 5, wherein, The step S5 constructs a rough perception feature evaluation framework based on the granular ball topology driving, performs adaptive decision class division on the label distribution space through granular ball calculation, and measures feature dependency degree through granular ball neighborhood rough set. The specific steps include: S51: For any , define its adaptive neighborhood under the feature subset as: ； wherein, and is the center and radius of a sphere of is the distance under the feature subset ​​ S52: Constructing dynamic labeled particle group set, through adaptive particle splitting algorithm, to the real labeled global Performing multi-scale nonlinear spatial reconstruction: ； each represents a marked particle, collectively covering the entire marked space, and each particle is optimal in purity and coverage; S53: In a given decision system decision attributes are partitioned into decision classes, and a feature subset is generated , define upper and lower approximation sets: ， ； Wherein, , through the granulocyte With the non-empty overlap relationship of decision class Determine the coverage boundary of the decision class; ; By granules With the full inclusion relationship of decision class Extract the core support domain of the decision class; S54: Construct an extended positive region dependency quantization model for the image noise label decision system: ； wherein, is a sample belonging to both is a sample belonging to both is a sample belonging to both denotes the set of samples in but not in is the distance of the th sample to the center of the granulocyte, is the distance of the th sample to the center of the granulocyte;​ S55: Through a dynamic multi-granularity collaborative feature importance quantization algorithm, the rough set theory and the granular ball calculation framework are fused to realize accurate evaluation of the features: 。 7. The image noise marker feature selection method of claim 6, wherein, The step S6 evaluates the robustness of feature selection from the perspective of dual modalities. The specific steps include: S61: Measure the similarity of granular ball samples based on the similarity quantization algorithm of the internal sample structure consistency of the granular ball: ； wherein, represents the cardinality of the granulocyte collection; S62: Based on the additional feature importance evaluation mechanism, the discriminant contribution of the feature is quantified by measuring the degree of change of the granular ball structure before and after the feature is removed: ； S63: By constructing a feature dependency matrix and a granular ball internal sample consistency matrix, a multi-source dual-modality dependency-consistency joint evaluation method is used to evaluate the feature weight, so as to realize fine quantization and sorting of key features in high-dimensional data: 。 8. An image noise marker feature selection system characterized by, It includes: A data acquisition and processing module is configured to acquire an image noise label data set to be processed, wherein the image noise label data set contains a feature space and a label space; A fuzzy clustering dynamic division module is configured to embed a sample set in the image noise label data set into a multi-granularity fuzzy cluster group by using an adaptive multi-kernel fusion and dynamic parameter optimization mechanism, to construct a dynamic fuzzy membership evaluation matrix, and to determine a multi-dimensional fuzzy mapping relationship between the quantized sample-cluster and sample-class; A multi-scale granular ball level construction module is configured to construct a macro coarse-grained granular ball skeleton based on a global spectral feature mapping mechanism, to jointly evaluate the granular balls with insufficient label isomorphism or structural entropy exceeding a threshold value by integrating a quality measurement model, and to dynamically evolve a multi-level high-precision sub-granular ball cluster by using a progressive self-organizing fission algorithm; A coupled granular ball label disambiguation module is configured to fuse granular ball calculation and fuzzy reasoning driven by weight within each fine granular ball, to automatically assign weights according to sample-granular ball center distance, to construct a class-cluster coupling correlation matrix, and to optimize cluster group membership to class membership by fuzzy synthesis operation, so as to finally obtain a high-resolution label distribution; A granular ball-rough set collaborative feature evaluation module is configured to construct a rough perception feature evaluation framework based on granular ball topology driving, to define an adaptive neighborhood based on granular ball geometric characteristics, to dynamically divide local neighborhoods that are cohesive and do not overlap, to extract decision equivalence classes by combining rough set upper and lower approximation and extended positive region model, to determine the contribution of each feature to the decision system by fusing multi-granular decision boundary information based on dependency quantization model; A multi-granularity feature stability verification module is configured to introduce a granular ball structure consistency verification mechanism, to measure the expression stability of the feature in the local domain of the granular ball from the structural preservation dimension, and to perform multi-level evaluation on the importance of the feature through dependency and consistency.

9. A storage medium having stored thereon a computer program, characterized in that The program, when executed by the processor, implements the image noise marker feature selection method according to any one of claims 1 to 7.

10. A computer comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor implements the image noise marker feature selection method according to any one of claims 1 to 7 when executing the computer program.

Images