A pair strategy-based security image blur clustering method

Abstract

The present application relates to a kind of based on pair strategy's ground security image fuzzy clustering method, belong to image processing field.It includes: obtaining ground security image;The ground security image is spliced to obtain target data matrix;The target data matrix is normalized;Label matrix is randomly initialized;Distance matrix is calculated using Euclidean distance formula;Based on label matrix and distance matrix, construct the objective function of fuzzy clustering model, and the coordinate descent optimization method is used to optimize and solve target function;The label matrix after optimization is converted into hard label, and the target data matrix is clustered, so that the different features of target data matrix are associated, and then more accurate image category information is obtained.The present application aims to minimize the sum of intra-class distance of each class while effectively reducing the ratio in the re-clustering process of abnormal value, so as to improve the resistance of algorithm to abnormal value, optimize the clustering result of data, and better segment image samples.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology, specifically relating to a fuzzy clustering method for near-site security images based on a pairwise strategy, which is applied to the near-site security field. Background Technology

[0002] Clustering is the process of dividing a dataset into different groups according to specific criteria, maximizing the similarity between data points within the same cluster and maximizing the differences between data points in different clusters. Traditional hard clustering divides data into discrete groups, where each data point can belong to only one cluster center. In hard clustering, the elements in the cluster index matrix are either 0 or 1, while in soft clustering, the elements in the index matrix can be numbers between 0 and 1.

[0003] Multimodal cognitive computing is a crucial component of near-field security technology systems. Image cognition is an important modality within multimodal cognition, and clustering of different modalities, especially image data, is one of the most common application requirements. In the field of image cognition, clustering can play a significant role, helping us identify meaningful information and patterns within massive datasets, thereby improving work efficiency and effectiveness. It's worth noting that in real-world scenarios, an image often falls into multiple categories, indicating that each image has an appropriate proportion across all clusters and that a certain degree of membership is defined for each category. This means that many problems involve multiple choices rather than single-choice ones. It also implies that soft clustering, also known as fuzzy clustering, is more suitable for practical applications than hard clustering. Therefore, fuzzy clustering algorithms can be applied to image data clustering and recognition. By using these algorithms to cluster different image features and fusing the clustering results, the accuracy of image cognition and recognition can be improved.

[0004] Zadeh was the first to propose fuzzy sets, and Ruspini further developed fuzzy clustering algorithms based on this. Dunn combined K-means and fuzzy set theory to propose the most classic fuzzy clustering algorithm, Fuzzy C Means (FCM). FCM assigns each data point to each cluster center as a fuzzy membership metric, measuring the similarity of each data point to each cluster center, and solves the problem by iteratively iterating through the label matrix and cluster centers. Askari et al. proposed an improved algorithm based on FKM, which can adapt to data with uneven cluster sizes.

[0005] However, the above methods are generally ineffective in handling outliers. When FCM is applied to the multimodal domain, the weight of each dimension of the sample needs to be considered. In the process of clustering and segmentation, the goal of the algorithm is usually to minimize the sum of distances within the class. However, the occurrence of outliers often causes the data class center to deviate, resulting in a decrease in the data clustering effect. The images in the obtained local security data are unclear and not suitable for subsequent research on local security. Summary of the Invention

[0006] The technical problem to be solved by this invention is:

[0007] To avoid the problems of low accuracy and weak robustness in the clustering of near-site security images by existing technologies, this invention provides a fuzzy clustering method for near-site security images based on a pairwise strategy, which is used to obtain more accurate image category information.

[0008] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0009] A fuzzy clustering method for near-site security images based on a pairwise strategy is characterized by the following steps:

[0010] Step 1: Acquire local security images;

[0011] Step 2: Stitch the nearby security images to obtain the target data matrix X;

[0012] Step 3: Normalize the target data matrix;

[0013] Step 4: Randomly initialize the label matrix, wherein the label matrix F = [f1, ..., f n ] T ∈R n×c Where n is the number of samples, c is the number of categories, and f i Let F be the i-th row of the label matrix F, representing the membership degree of the i-th data to different categories. The sum of all elements in each row of F is 1.

[0014] Step 5: Calculate the distance matrix D using the Euclidean distance formula:

[0015]

[0016] Where, d ij Let x be the element in the i-th row and j-th column of the distance matrix D, representing the Euclidean distance between pixel j and cluster center i; i x j These represent the positions of pixel j and cluster center i, respectively.

[0017] Step 6: Construct the objective function of the fuzzy clustering model based on the label matrix and distance matrix, and use the coordinate descent optimization method to optimize and solve the objective function;

[0018] Step 7: Convert the optimized label matrix into hard labels, and cluster the target data matrix X to link different features of the target data matrix X, thereby obtaining more accurate image category information.

[0019] A further technical solution of the present invention: in step 1, a local security image is acquired via satellite.

[0020] A further technical solution of the present invention: the objective function is:

[0021]

[0022] Among them, f i Let F represent the i-th row, γ be the hyperparameter of the objective function, and α be the hyperparameter of the objective function. i =F r v, F r For F T The matrix formed by removing the column vector of the i-th column without changing the order of the other elements, and v represents the vector formed by removing the i-th element of the i-th column of the distance matrix without changing the order of the other elements.

[0023] A computer system is characterized by comprising: one or more processors, and a computer-readable storage medium for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement the method described above.

[0024] A computer-readable storage medium is characterized by storing computer-executable instructions, which, when executed, are used to implement the above-described method.

[0025] A computer program product is characterized by including computer-executable instructions, which, when executed, are used to implement the above-described method.

[0026] The beneficial effects of this invention are as follows:

[0027] This invention provides a fuzzy clustering method for near-site security data based on a pairwise strategy. It employs the idea of ​​optimizing the Euclidean squared distance between sample points within a cluster, transforming the problem of clustering samples into an optimization problem of minimizing the distance between a single sample and its fellow samples. A fuzzy algorithm is constructed and, leveraging this idea, a coordinate descent optimization strategy is used to efficiently solve the problem. The aim is to minimize the sum of intra-cluster distances within each class while effectively reducing the ratio of outliers during the clustering process, thereby improving the algorithm's resistance to outliers, optimizing the data clustering results, and better segmenting image samples. Attached Figure Description

[0028] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts.

[0029] Figure 1 This is a flowchart of the fuzzy clustering algorithm method based on pairwise strategies of the present invention.

[0030] Figure 2 These are the results of clustering the simulated dataset and the traditional fuzzy clustering algorithm using the present invention: (a) Clustering result of the FCM algorithm; (b) Clustering result of the algorithm of the present invention; (c) Clustering result of FCM on abnormal data; (d) Clustering result of the present invention on abnormal data.

[0031] Figure 3 These are the experimental results of this invention on real datasets with and without regularization terms in the algorithm; (a) with regularization terms; (b) without regularization terms.

[0032] Figure 4 These are the experimental results of this invention on real datasets before and after incorporating the K-nearest neighbor concept; (a) before incorporating the K-nearest neighbor concept; (b) after incorporating the K-nearest neighbor concept. Detailed Implementation

[0033] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0034] This invention provides a fuzzy clustering method for local security data based on pairwise strategies. First, the distance between data points is used to determine the membership degree of the data, and a regularization term is introduced to transform the hard clustering problem into a fuzzy clustering problem. Furthermore, the K-nearest neighbor idea is combined to further improve the robustness of the algorithm. Then, the coordinate descent method is used for optimization to minimize the distance between a single data point and data points belonging to the same class as that data point.

[0035] like Figure 1 As shown, it includes the following steps:

[0036] I. Constructing an Algorithm Model for Fuzzy Clustering

[0037] First, the algorithm requires that the greater the similarity between two sample points, the higher the probability that these two sample points belong to the same group. The objective function is as follows:

[0038] min F>0,F1＝1 Tr(F T DF)+γTr(F T F) (1)

[0039] Where Tr(F) T F) is a regularization term.

[0040] When k is between 1 and c, the probability of assigning the i-th sample and the j-th sample to the same cluster is relatively small.

[0041] The formula above can be easily derived into the following expression.

[0042]

[0043] II. Iterative Optimization Solution Algorithm Model

[0044] Step 1: Taking image data input as an example, X is the target data matrix, n is the number of samples, c is the pre-set number of clusters, therefore, d ij It represents the fuzzy membership degree of pixel j in the image with respect to cluster i. First, after labeling the data, the data is normalized.

[0045] Step 2: Each data point belongs to each class with a different probability. F is the label matrix, an n×c matrix where c represents the number of classes. F = [f1, ..., f2]. n ] T ∈R n×c , where f i Let be the i-th row of the label matrix F, representing the membership degree of the i-th data point to different categories. The sum of all elements in each row of F is 1.

[0046]

[0047] Update each row in F using this method.

[0048] Step 3: Optimize the objective function using the coordinate descent optimization method, specifically as follows:

[0049] Step 3.1: Randomly initialize the cluster label vector F according to the requirements, where k = max i＝1...n {f i} indicates that the i-th data point is assigned to the k-th cluster, i.e., x i ∈A k A k Let $k$ represent the set of the $k$-th cluster. Simultaneously, the distance matrix $D$ is calculated based on the Euclidean distance. The elements in the $i$-th row and $j$-th column are calculated as follows, representing the Euclidean distance between pixel $j$ and cluster center $i$:

[0050]

[0051] Step 3.2: Optimize the label matrix row by row according to the initialized cluster labels. We divide the label matrix into the first row and other rows, the second row and other rows, and iterate and update the first row and the second row of the matrix in turn. Taking the optimization of the i-th row as an example, the objective function (3) can be simplified to the following form.

[0052]

[0053] In the formula f i Let F represent the i-th row of F. r For the indicator matrix F T Let v be the distance matrix formed by removing the i-th column vector without changing the order of the other elements, and v be the distance vector formed by removing the i-th element from the i-th column without changing the order of the other elements. Calculate α. i The formula is as follows:

[0054] α i =F r v (6)

[0055] Step 3.3: Fix all rows in F except for the i-th row, optimize the i-th row based on the other rows, and solve the objective function after gradient descent using the Lagrange multiplier. For each row f... i Corresponding F r Both f and v can be represented, at which point f i * The calculation method is as follows:

[0056]

[0057] Step 3.4, at this point, the vector in the i-th row of the label matrix F is obtained by fi Updated to f i * The process involves updating the (i+1)th row of the updated label matrix, and then updating all n rows of the label matrix in sequence to form a preliminary label matrix. After the preliminary label matrix is ​​formed, the above process is repeated until the label matrix no longer changes, at which point the iteration terminates.

[0058] Step 4: After the iteration terminates, the cluster labels F are obtained. At this point, the label matrix is ​​a fuzzy matrix. We find the position of the maximum membership value k = max in each row. i＝1...n {f i The data in the i-th row has the largest value in the k-th column, so the i-th data is divided into the k-th class, thus realizing the transition from soft label to hard label and completing the final clustering result.

[0059] To enable those skilled in the art to better understand the present invention, the present invention will be described in detail below with reference to specific embodiments.

[0060] Example:

[0061] Step 1: Representation of Data Samples

[0062] For n images with resolution p×q, let d = p×q, and transform the n images into an n×d target data matrix X1.

[0063] Step 2: Algorithm Construction and Optimization

[0064] Suppose we need to divide n data points into c classes, where d represents the feature dimension of the data and n represents the number of data points. First, after labeling the data, we perform normalization preprocessing.

[0065] We employ random initialization and iterative updates of cluster centers, randomly initializing cluster labels F and calculating the distance matrix D′. Due to the high dimensionality and complex features of the data, it is more susceptible to interference from outliers. To better cluster the data, we use the K-nearest neighbor approach. When calculating the distance matrix, we use different values ​​for K, retaining only the distances to the K nearest samples and setting the distances to 0 for the other nK samples. This yields a new distance matrix D.

[0066] Based on the distance matrix D, fix all rows except the i-th row, and for each row f i Corresponding F r Both v and α can be expressed, and α can be calculated. i =F r v, update f line by line i * This method updates each row in F.

[0067] Step 3: Denote the updated label matrix as F. After updating each row, start updating again from the first row. Denote the updated label matrix as F′. If F′ = F, then exit the loop. If the condition F′ = F is not met, then continue updating in the loop until F′ = F and exit the loop.

[0068] Step 4: After the iteration terminates, the resulting cluster labels are fuzzy matrices. We search for the position of the maximum membership value in each column. The data in the i-th row has the largest value in the k-th column. Therefore, the i-th data is divided into the k-th class, thus realizing the transition from soft labels to hard labels and completing the final cluster division result.

[0069] from Figure 2 It can be seen that the clustering effect of the present invention is better than that of the traditional FCM. It is worth mentioning that the present invention also performs better than the traditional FCM algorithm when dealing with discrete points, thus proving that the present invention is more robust.

[0070] from Figure 3 It can be seen that the clustering performance of the present invention is better after adding the regularization term.

[0071] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the scope of the technology disclosed in the present invention, and such modifications or substitutions should all be covered within the scope of protection of the present invention.

Claims

1. A fuzzy clustering method for near-site security images based on pairwise strategies, characterized in that... The steps are as follows: Step 1: Acquire local security images; Step 2: Stitch the nearby security images to obtain the target data matrix. ; Step 3: Normalize the target data matrix; Step 4: Randomly initialize the label matrix. ,in For the sample size, For the number of categories, Label matrix The line, indicating the first The membership degree of each data point to different categories. The sum of all elements in each row is 1; Step 5: Calculate the distance matrix using the Euclidean distance formula. : in, Distance matrix No. line, number The elements of the column are used to represent pixels. and cluster center The Euclidean distance between them; , Representing pixels and cluster center Location; Step 6: Construct the objective function of the fuzzy clustering model based on the label matrix and distance matrix, and use the coordinate descent optimization method to optimize and solve the objective function; The specific steps for constructing the fuzzy clustering model are as follows: First, the algorithm requires that the greater the similarity between two sample points, the higher the probability that these two sample points belong to the same group; the objective function is as follows: in, It is a regular term; when k In 1 to c In between, the first i The first sample and the first j The probability of each sample being assigned to the same cluster is relatively small; a simple derivation of the above formula yields the following equation: The objective function is optimized using the coordinate descent optimization method, specifically as follows: Randomly initialize cluster label vectors ,in Indicates the first The data was divided into the first... Clusters, namely , Indicates the first A set of clusters; simultaneously, the distance matrix is ​​calculated based on the Euclidean distance. ; Based on the initialized clustering labels, the label matrix is ​​optimized row by row; the label matrix is ​​divided into the first row and other rows, and the second row and other rows, and the first row and second row of the matrix are updated iteratively in turn. The objective function is simplified to the following form: in, express The OK, Let be the hyperparameters of the objective function. , for Remove the first A matrix formed by concatenating column vectors without changing the order of other elements. The distance matrix represents the first... Remove the first column A vector formed by arranging elements without changing the order of other elements; fixed In addition to the first All rows other than row 1 are optimized based on the other rows. Line by line, using the Lagrange multiplier to solve the objective function after gradient descent, each line... corresponding and All of these can be represented at this time. The calculation method is as follows: The label matrix at this time The Middle Row vectors are Updated to Update the tag matrix again based on the updated tag matrix. Okay, update the label matrix in sequence. n The process is repeated to form a preliminary label matrix. After the preliminary label matrix is ​​formed, the above process is repeated until the label matrix no longer changes, at which point the iteration terminates. Step 7: Convert the optimized label matrix into hard labels for the target data matrix. Clustering is performed to make the target data matrix By linking different features together, more accurate image category information can be obtained.

2. The fuzzy clustering method for near-site security images based on pairwise strategies according to claim 1, characterized in that, In step 1, multimodal security data is acquired via satellite.

3. A computer system, characterized in that... include: One or more processors, a computer-readable storage medium, for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to perform the method of any one of claims 1-2.

4. A computer-readable storage medium, characterized in that... The device stores computer-executable instructions, which, when executed, are used to implement the method described in any one of claims 1-2.

5. A computer program product, characterized in that... It includes computer-executable instructions, which, when executed, are used to implement the method of any one of claims 1-2.

Images