A small sample semi-supervised clustering method based on separation degree

Abstract

The application belongs to the technical field of computer simulation and method optimization, and discloses a small sample semi-supervised clustering method based on separation degree, which comprises the following steps: step 1, clustering center calculation; step 2, separation degree calculation; step 3, non-target clustering center determination; step 4, parameter calculation; step 5, class center migration; step 6, iterative calculation; and step 7, target separation. The application is aimed at the problem that traditional unsupervised clustering and semi-supervised clustering are poor in target classification and discrimination performance, and a small sample semi-supervised clustering method based on target separation degree is invented, the separation degree of the target and the non-target is set, and iterative operation is carried out on the basis of the known small sample based on the separation degree, so that the accurate separation of the target and the non-target with high similarity is realized. The application adopts a new algorithm framework to improve the accuracy of target classification and discrimination.

Description

Technical Field

[0001] The present invention belongs to the technical field of computer simulation and method optimization, and in particular relates to a small sample semi-supervised clustering method based on separation degree. Background Art

[0002] Target detection and target classification have always been the focus of research. When there is little prior knowledge of the target, low information consistency and serious interference information, learning-based methods are not applicable.

[0003] Under such conditions, when the target and non-target texture, structure and other features have a high similarity, traditional target classification and target identification methods cannot separate the target well, which leads to problems such as poor target detection performance.

[0004] In the absence of training, object detection usually adopts self-similar comparison methods or unsupervised / semi-supervised classification algorithms.

[0005] A negative exponential function based on Euclidean distance is used as a similarity metric to determine densely computed scale-invariant features, enabling detection of cars, faces, and general objects. Matrix cosine similarity is used to find similar matches in the image to be detected based on a single sample, enabling object detection.

[0006] Clustering algorithms are a classic example of unsupervised learning, aiming to group similar objects together. Common clustering algorithms include the distance-based Fuzzy C-Means (FCM) clustering algorithm, the Evidential C-Means (ECM) clustering algorithm, the Neutrosophic C-Means (NCM) clustering algorithm, and the density-based DBSCAN clustering algorithm. Clustering algorithms require no training and are technically simple. However, SAR images contain interference such as clutter that closely resembles target textures. Using clustering algorithms directly for automatic target classification results in poor performance.

[0007] Semi-supervised learning primarily relies on clustering models and manifold models to establish connections between unknown samples, targets, and known samples. The clustering hypothesis aligns with human intuition, leading to the emergence of numerous classification algorithms based on it. For example, the TSVM algorithm uses an iterative approach to find appropriate classification labels for unlabeled samples. The clustering hypothesis focuses primarily on overall characteristics, while the manifold hypothesis primarily considers the local characteristics of the model. For example, the Nuclear Norm-based Two-Dimensional Locality Preserving Projections (NN-2DLPP) method removes noise from the data by restoring the noisy data matrix through low-order learning. The clean data points are then projected onto a new subspace, maintaining the distance between data points of the same class as much as possible within this projected subspace.

[0008] Currently, most of the above methods are based on noise-free data and are directly used in feature-based target classification, resulting in low target detection performance. Summary of the Invention

[0009] The present invention aims to provide a small sample semi-supervised clustering method based on separation degree to solve the above technical problems.

[0010] In order to solve the above technical problems, the specific technical solution of the small sample semi-supervised clustering method based on separation degree of the present invention is as follows:

[0011] A small sample semi-supervised clustering method based on separation degree includes the following steps:

[0012] Step 1: Cluster center calculation: Use NCM algorithm to calculate the cluster center;

[0013] Step 2: Separation calculation: Adjust the cluster centers of the two categories during the NCM cluster calculation process to enhance the separability of targets and non-targets;

[0014] Step 3: Determine the non-target cluster center: Preliminarily calculate the sample cluster center according to step 1, and calculate the separation degree between the feature and the sample cluster center in the potential target, and use the feature with the largest separation degree as the initial non-target cluster center;

[0015] Step 4: Parameter calculation: Calculate the value of the [T, I, F] tuple and update the target and non-target cluster centers;

[0016] Step 5: Cluster center migration: During the parameter iteration process, the potential target cluster center is migrated to the sample cluster center, and the potential targets far away from the sample center are regarded as non-targets;

[0017] Step 6: Iterative calculation: perform iterative calculation for steps 3 to 5;

[0018] Step 7: Target separation: Based on the sample center, determine whether the potential target belongs to the target cluster and the ratio of the certainty F of the potential target not belonging to the target cluster to finally determine whether the potential target is the target. Further, the step 1 includes the following specific steps:

[0019] NCM uses a tuple [T, I, F] to describe the membership value, where T is the certainty of an object belonging to a cluster, F is the certainty of an object not belonging to a cluster, and I is the uncertainty of an object to a boundary cluster. The objective function expression of the NCM algorithm is:

[0020]

[0021] Where C={c j ,j=1,2,…,K} is the set of clusters, c j is the center of cluster j, d i,Kj Represents the object x i The distance to the center of the K cluster under consideration, ω i is the weighting coefficient, δ is the given distance for detecting non-targets, and only the two nearest deterministic clusters are selected to consider the maximum membership value, which is simplified to:

[0022]

[0023] in, is from the object x i The distance to the centers of two clusters, which are the closest and second closest clusters to the center, satisfies the above formula:

[0024]

[0025] Object x i The member tuple of [T ij ,I i ,F i ]Update and make J in the iteration NCM Minimize, and when the member values ​​reach convergence, the iteration will be interrupted,

[0026]

[0027] Among them, the three member values ​​[T, I, F] are iteratively updated as follows:

[0028]

[0029]

[0030] Furthermore, the step 2 includes the following specific steps:

[0031] Assume that the separation degree S of target X and non-target Y is xy , adjust the cluster centers of the two categories during the NCM clustering calculation process to enhance the separability of targets and non-targets, the separation degree S xy for:

[0032]

[0033] Among them, d xy is the between-class variance of class X and class Y; D(x) and D(y) are the within-class variances.

[0034] Furthermore, the step 4 includes the following specific steps:

[0035] Calculate the value of the [T, I, F] tuple and update the cluster centers of the target and non-target

[0036]

[0037]

[0038] Furthermore, the step 5 includes the following specific steps:

[0039] During the parameter iteration process, the potential target cluster center is migrated to the sample cluster center, and the potential target far away from the sample center is regarded as a non-target. The direct distance between the potential target and the sample cluster center is calculated by the separation degree. Discrete points are not calculated during the class center migration process, and their corresponding potential target areas are identified as non-targets.

[0040] Furthermore, step 6 includes the following specific steps:

[0041] Perform iterative calculations for steps 3 to 5. When [T, I, F] satisfies the following requirements, it is determined that the convergence point has been reached and the iteration stops.

[0042]

[0043] Furthermore, in the iterative process of step 6, a new variable f is introduced dp As the basis for judging discrete points, the cluster center of the new object set after eliminating discrete points is calculated in the iterative process. At this time, the objective function expression and constraint conditions can be expressed as:

[0044]

[0045]

[0046] The present invention provides a small sample semi-supervised clustering method based on separation degree, which has the following advantages: In view of the problem that traditional unsupervised clustering and semi-supervised clustering have poor target classification and identification performance, the present invention invents a small sample semi-supervised clustering method based on target separation degree, sets the separation degree between targets and non-targets, and performs iterative operations based on the separation degree based on known small samples, thereby achieving accurate separation of targets and non-targets with high similarity. The present invention adopts a new algorithm framework to improve the accuracy of target classification and identification. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 Schematic diagram of the NCM cluster center of the present invention;

[0048] Figure 2 Schematic diagram of cluster centers calculated after separating discrete points in the present invention;

[0049] FIG3( a ) is an original image of Example 1 of the present invention;

[0050] FIG3( b ) is a schematic diagram of a SAR potential target area in which potential targets are marked with circles in Example 1;

[0051] FIG4( a ) is a diagram showing the ECM classification results of Example 1;

[0052] FIG4( b ) is a diagram of the FCM classification results of Example 1;

[0053] FIG4( c ) is a diagram showing the NCM classification results of Example 1;

[0054] FIG4( d ) is a diagram showing the classification results of the method of the present invention in Example 1;

[0055] FIG5( a ) is an original image of Example 2 of the present invention;

[0056] FIG5( b ) is a schematic diagram of a SAR potential target area in which potential targets are marked with circles in Example 2;

[0057] FIG6( a ) is a diagram showing the ECM classification results of Example 2;

[0058] FIG6( b ) is a diagram of the FCM classification results of Example 2;

[0059] Figure 6(c) is the NCM classification result diagram of Example 2;

[0060] FIG6( d ) is a diagram showing the classification results of the method of the present invention in Example 2. FIG. DETAILED DESCRIPTION

[0061] In order to better understand the purpose, structure and function of the present invention, the following is a further detailed description of a small sample semi-supervised clustering method based on separation degree of the present invention in conjunction with the accompanying drawings.

[0062] This paper adopts a semi-supervised clustering algorithm based on separation to solve the problems of classifying and identifying targets and non-targets with high similarity. The steps of the solution are as follows:

[0063] Step 1: Cluster center calculation.

[0064] NCM uses a tuple [T, I, F] to describe the membership value, where T is the certainty of an object belonging to a cluster, F is the certainty of an object not belonging to a cluster, and I is the uncertainty of an object to the boundary cluster. The objective function of the NCM algorithm is expressed as:

[0065]

[0066] Where C={c j ,j=1,2,…,K} is the set of clusters, c j is the center of cluster j. i,Kj Represents the object x i The distance to the center of the K cluster under consideration. ω i is the weighting coefficient, and δ is the given distance for detecting non-targets. However, as the number of clusters increases, a large number of operations are required, which is very time-consuming. To solve this problem, it is recommended to only select the two closest deterministic clusters to consider the maximum membership value. Therefore, the above formula is simplified to:

[0067]

[0068] in, is from the object x i The distance to the centers of two clusters, which are the closest and second closest clusters to the center. The above formula satisfies:

[0069]

[0070] Object x i The member tuple of [T ij ,I i ,F i ]Update and make J in the iteration NCM Minimize. The iteration is interrupted when the member values ​​converge.

[0071]

[0072] Among them, the three member values ​​[T, I, F] are iteratively updated as follows:

[0073]

[0074]

[0075] Step 2: Calculation of separation.

[0076] Assume that the separation degree S of target X and non-target Y is xy , adjust the cluster centers of the two categories during the NCM clustering calculation process to enhance the separability of targets and non-targets. Separation degree S xy for:

[0077]

[0078] Among them, d xy is the between-class variance of class X and class Y; D(x) and D(y) are the within-class variances.

[0079] Step 3: Determine the non-target cluster center.

[0080] Preliminarily calculate the sample cluster centers according to step 1, and calculate the separation between the features and the sample cluster centers in the potential targets. The feature with the largest separation is used as the initial non-target cluster center.

[0081] Since some non-targets are not of the same type and may contain diverse clutter, the calculation process will make these two types of sample points not participate in the calculation of cluster centers. However, due to the diversity of clutter in SAR images, "clutter" has a greater impact on the clustering algorithm. If the "target-clutter" is directly classified into two categories, the cluster center deviation will be large. Figure 1 As shown in Figure 3, whether "target-clutter" is regarded as two categories or all objects are regarded as one category for calculation, there is a cluster center offset problem.

[0082] Step 4: Parameter calculation.

[0083] Calculate the value of the [T, I, F] tuple and update the cluster centers of the target and non-target

[0084]

[0085]

[0086] Step 5: Class center migration.

[0087] During parameter iteration, the potential target cluster center is relocated toward the sample cluster center, and potential targets that are far from the sample center are considered non-targets. The direct distance between a potential target and the sample cluster center is calculated using the separation factor. Discrete points are not counted during cluster center migration, and their corresponding potential target regions are considered non-targets. This method gradually migrates the cluster center toward the target and, during the clustering process, gradually eliminates non-targets such as clutter from potential targets.

[0088] Step 6: Iterate the calculation.

[0089] Perform iterative calculations for steps 3 to 5. When [T, I, F] satisfies the following requirements, it is determined that the convergence point has been reached and the iteration stops.

[0090]

[0091] It should be noted that during the iteration process, a new variable f is introduced dp As the basis for judging discrete points, the cluster center of the new object set after eliminating discrete points is calculated in the iterative process. Figure 2 It can be seen that the cluster centers calculated after separating the discrete points are more accurate than those of the NCM method.

[0092] At this point, the objective function expression and constraints can be expressed as:

[0093]

[0094]

[0095] Step 7: Target separation.

[0096] According to the sample center, the ratio of the degree of certainty that the potential target belongs to the target cluster to the degree of certainty that it does not belong to the target cluster is used to finally determine whether the potential target is the target.

[0097] Specific embodiments of the technical solution of the present invention

[0098] In order to verify the effectiveness of the semi-supervised clustering algorithm based on separation degree proposed in the present invention, simulation verification is carried out.

[0099] Example 1: Tank Target Identification

[0100] As shown in Figure 3(a), the experimental scene contains 13 tank targets and several non-targets with high grayscale values. First, the global CFAR algorithm is used to identify potential target regions with relatively high grayscale values ​​in the SAR image. After preliminary detection, the SAR potential target region, including a total of 26 potential targets, is circled, as shown in Figure 3(b).

[0101] After determining the feature points according to the CFAR detection results, the feature vectors of potential targets are classified using FCM, ECM, NCM and the algorithm of the present invention respectively. Figures 4(a)-4(d) The classification results of these clustering algorithms are shown, where the targets are marked with circles.

[0102] When the ECM algorithm is used to classify potential target features, there are missed targets and false alarm targets in the detection results (see Figure 4(a)); when the FCM algorithm is used to classify potential target features, there are non-targets detected as targets in the detection results (see Figure 4(b)); when the NCM algorithm is used for target detection, the parameter K is set to 1, that is, there is only one target in the image and the others are discrete points, but there are still missed targets and false alarm targets in the results (see Figure 4(c)); when the method of the present invention is used for classification, the detection results are better than those of the ECM, FCM and NCM clustering algorithms, and more accurate detection results can be obtained (see Figure 4(d)). Table 1 shows the specific comparative test results.

[0103] Table 1 Classification results of different methods

[0104]

[0105] Experiments show that the method of the present invention can effectively distinguish targets such as clutter from non-targets, and achieves good results in target detection in SAR images.

[0106] Example 2: Aircraft Target Identification

[0107] The experimental target is shown in Figure 5(a). The background of the original SAR image also contains many highlighted areas. The preliminary detection results contain 15 highlighted areas, using CFAR detection and size feature detection, as shown in Figure 5(b).

[0108] The algorithm of the present invention is used to cluster the features of potential SAR targets and automatically detect aircraft targets, such as Figures 6(a)-6(d) When using ECM, FCM and NCM algorithms to classify potential target features, there are missed targets and false alarm targets in the detection results (see Figure 6(a) 、 6(b) , 6(c)); when the NCM-1K method is used for classification, the detection results are better than those of the ECM, FCM, and NCM clustering algorithms, and more accurate detection results can be obtained (see Figure 6(d)). Experimental results show that the proposed method can effectively distinguish targets and non-targets in SAR image target detection.

[0109] It will be understood that the present invention is described by way of some embodiments, and it will be appreciated by those skilled in the art that various changes or equivalent substitutions may be made to these features and embodiments without departing from the spirit and scope of the present invention. In addition, under the teachings of the present invention, these features and embodiments may be modified to adapt to specific circumstances and materials without departing from the spirit and scope of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are intended to be protected by the present invention.

Claims

1. A small sample semi-supervised clustering method based on separation degree, which is used to process SAR images, characterized in that: The steps include: Step 1: Cluster center calculation: Use NCM algorithm to calculate the cluster center; NCM uses a tuple [T, I, F] to describe the membership value, where T is the certainty of an object belonging to a cluster, F is the certainty of an object not belonging to a cluster, and I is the uncertainty of an object to a boundary cluster. The objective function expression of the NCM algorithm is: Where C={c j ,j=1,2,…,K} is the set of clusters, c j is the center of cluster j, d i,Kj Represents the object x i The distance to the center of the K cluster under consideration, ω i is the weighting coefficient, δ is the given distance for detecting non-targets, and only the two nearest deterministic clusters are selected to consider the maximum membership value, which is simplified to: in, is from the object x i The square of the distance to the centers of the two clusters, which are the closest and second closest clusters to the center, satisfy the above formula: Object x i The member tuple of [T ij ,I i ,F i ]Update and make J in the iteration NCM Minimize, and when the member values ​​reach convergence, the iteration will be interrupted, Among them, the three member values ​​[T, I, F] are iteratively updated as follows: Step 2: Separation calculation: Adjust the cluster centers of the two categories during the NCM cluster calculation process to enhance the separability of targets and non-targets; Step 3: Determine the non-target cluster center: Preliminarily calculate the sample cluster center according to step 1, and calculate the separation degree between the feature and the sample cluster center in the potential target, and use the feature with the largest separation degree as the initial non-target cluster center; Step 4: Parameter calculation: Calculate the value of the [T, I, F] tuple and update the target and non-target cluster centers; Step 5: Cluster center migration: During the parameter iteration process, the potential target cluster center is migrated to the sample cluster center, and the potential targets far away from the sample center are regarded as non-targets; Step 6: Iterative calculation: perform iterative calculation for steps 3 to 5; Step 7: Target separation: Based on the sample center, determine whether the potential target is a target by comparing its degree of certainty of belonging to the target cluster to its degree of certainty F of not belonging to the target cluster.

2. The small sample semi-supervised clustering method based on separation degree according to claim 1, characterized in that The step 2 includes the following specific steps: Assume that the separation degree S of target X and non-target Y is xy , adjust the cluster centers of the two categories during the NCM clustering calculation process to enhance the separability of targets and non-targets, the separation degree S xy for: Among them, d xy is the between-class variance of class X and class Y; D(x) and D(y) are the within-class variances.

3. The small sample semi-supervised clustering method based on separation degree according to claim 1, characterized in that The step 4 includes the following specific steps: Calculate the value of the [T, I, F] tuple and update the cluster centers of the target and non-target 4. The small sample semi-supervised clustering method based on separation degree according to claim 1, characterized in that The step 5 includes the following specific steps: During the parameter iteration process, the potential target cluster center is migrated to the sample cluster center, and the potential target far away from the sample center is regarded as a non-target. The direct distance between the potential target and the sample cluster center is calculated by the separation degree. The discrete points are not calculated during the class center migration process, and the corresponding potential target area is identified as a non-target.

5. The small sample semi-supervised clustering method based on separation degree according to claim 1, characterized in that The step 6 includes the following specific steps: Perform iterative calculations for steps 3 to 5. When [T, I, F] satisfies the following requirements, it is determined that the convergence point has been reached and the iteration stops.

6. The small sample semi-supervised clustering method based on separation degree according to claim 5, characterized in that In the iterative process of step 6, a new variable f is introduced. dp As the basis for judging discrete points, the cluster center of the new object set after eliminating discrete points is calculated in the iterative process. At this point, the objective function expression and constraints can be expressed as:

Images