A hyperspectral image clustering method, system, device, medium and product

Abstract

The present application discloses a hyperspectral image clustering method, system, device, medium and product, which relates to the field of image data processing. The method includes: obtaining a hyperspectral image; using a superpixel segmentation method to convert the hyperspectral image into a superpixel-level graph; the superpixel-level graph is a regional-level graph data; obtaining a clustering model; inputting the superpixel-level graph into the clustering model to obtain a clustering result. Among them, the clustering model includes an adaptive filter graph encoder, a graph embedding clustering self-training decoder and an isomorphic enhancement structure learning module; the adaptive filter graph encoder is used to extract low-frequency features and high-frequency features; the graph embedding clustering self-training decoder is used for self-training and obtains clustering results using the k-means algorithm; the isomorphic enhancement structure learning module is used to dynamically adjust the input graph structure according to different clustering tasks. The present application can improve the clustering accuracy of hyperspectral images.

Description

Technical Field

[0001] The present application relates to the field of image data processing, and in particular to a hyperspectral image clustering method, system, device, medium and product. Background Art

[0002] Hyperspectral images (HSI) have the unique advantage of obtaining continuous spectral information of target objects. Hyperspectral images (HSI) have been widely used in many fields, such as earth observation, environmental monitoring, military detection and resource exploration. HSI contains rich feature recognition and classification potential, so HSI interpretation has important research significance. In the known technology, the development of graph clustering methods makes it possible for deep clustering methods to effectively encode spatial structural information, which makes deep clustering methods transform from only processing Euclidean data to non-Euclidean geometric graph structure data. Based on this, in the known technology, many structural graph clustering networks have been developed, and they have shown significant effectiveness in learning structural information through graph embedding. Although significant progress has been made in graph-based HSI clustering, the filters in existing graph networks are invariant, and more capabilities are required to process the signal characteristics of different nodes, resulting in poor clustering accuracy of large-scale HSI. In addition, mainstream graph networks cannot update the original graph according to different clustering tasks. That is, during the network training process, the graph is fixed, which makes it challenging to correct the incorrect edge connections in the original graph, which seriously limits the improvement of HSI clustering accuracy. Summary of the invention

[0003] The purpose of this application is to provide a hyperspectral image clustering method, system, device, medium and product that can improve the clustering accuracy of HSI.

[0004] To achieve the above objectives, this application provides the following solutions:

[0005] In a first aspect, the present application provides a hyperspectral image clustering method, comprising:

[0006] Acquire a hyperspectral image; the hyperspectral image is pixel-level image data;

[0007] The hyperspectral image is converted into a super-pixel level image by using a super-pixel segmentation method; the super-pixel level image is a regional level image data;

[0008] Obtain a clustering model; the clustering model is a network model trained using a sample data set; the clustering model includes an adaptive filter graph encoder, a graph embedding clustering self-training decoder, and an isomorphic enhancement structure learning module; the adaptive filter graph encoder is used to extract low-frequency features and high-frequency features; the graph embedding clustering self-training decoder is used to perform self-training and obtain clustering results using a k-means algorithm; the isomorphic enhancement structure learning module is used to dynamically adjust the input graph structure according to different clustering tasks;

[0009] The super-pixel level graph is input into the clustering model to obtain a clustering result.

[0010] Optionally, a super-pixel segmentation method is used to convert the hyperspectral image into a super-pixel level image, specifically including:

[0011] Performing dimensionality reduction processing on the hyperspectral image by using linear discriminant analysis to obtain a dimensionality reduction image;

[0012] Using a simple linear iterative clustering method to divide the reduced-dimensional image into a plurality of spatially connected superpixels;

[0013] Each superpixel is regarded as a graph node, and a nearest neighbor relationship is established between superpixels to generate the superpixel level graph.

[0014] Optionally, a simple linear iterative clustering method is used to divide the dimension-reduced image into a plurality of spatially connected superpixels, specifically comprising:

[0015] Set the correlation matrix for the simple linear iterative clustering method;

[0016] The correlation matrix is ​​used to perform an image back-projection operation on the reduced-dimensional image, and data features in the reduced-dimensional image are converted from pixels to super-pixels to obtain a plurality of spatially connected super-pixels.

[0017] Optionally, the t-layer stack of the adaptive filter graph encoder is:

[0018] ;

[0019] In the formula, represents the input graph node spectrum features, is the representation of the adaptive filter graph encoder after stacking t layers, is the graph node representation after t layers of convolution, represents the adaptive graph filter, represents the adaptive filter graph encoder layer.

[0020] Optionally, a linear transformation weight matrix is ​​used in the adaptive filter graph encoder to encode the superpixel level graph, which is expressed as:

[0021] ;

[0022] In the formula, represents the output of the adaptive filter graph encoder, represents the output of layer 0 of the adaptive filter graph encoder, Represents the linear transformation weight matrix.

[0023] Optionally, the graph is embedded in a clustered self-training decoder, using t distribution to measure the Cluster Centers and embed The similarity between nodes is calculated, and the cluster distribution is optimized by minimizing the KL divergence; represents the output of the i-th layer of the adaptive filter graph encoder.

[0024] In a second aspect, the present application provides a hyperspectral image clustering system, which is applied to the hyperspectral image clustering method provided above; the hyperspectral image clustering system comprises:

[0025] A data acquisition module, used to acquire a hyperspectral image; the hyperspectral image is pixel-level image data;

[0026] An initial graph construction module is used to convert the hyperspectral image into a super-pixel level graph by using a super-pixel segmentation method; the super-pixel level graph is a regional level graph data;

[0027] A model acquisition module is used to acquire a clustering model; the clustering model is a network model trained using a sample data set; the clustering model includes an adaptive filter graph encoder, a graph embedding clustering self-training decoder, and an isomorphic enhancement structure learning module; the adaptive filter graph encoder is used to extract low-frequency features and high-frequency features; the graph embedding clustering self-training decoder is used to perform self-training and obtain clustering results using a k-means algorithm; the isomorphic enhancement structure learning module is used to dynamically adjust the input graph structure according to different clustering tasks;

[0028] The clustering module is used to input the superpixel level graph into the clustering model to obtain a clustering result.

[0029] In a third aspect, the present application provides a computer device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of any of the hyperspectral image clustering methods described above.

[0030] In a fourth aspect, the present application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any one of the hyperspectral image clustering methods described above.

[0031] In a fifth aspect, the present application provides a computer program product, including a computer program, which, when executed by a processor, implements the steps of any one of the hyperspectral image clustering methods described above.

[0032] According to the specific embodiments provided in this application, this application has the following technical effects:

[0033] The present application provides a hyperspectral image clustering method, system, device, medium and product. By obtaining a clustering model including an adaptive filter graph encoder, a graph embedding clustering self-training decoder and an isomorphic enhanced structure learning module, the clustering result can be accurately obtained based on the superpixel level graph. In addition, by designing an adaptive filter graph encoder, the present application can process the signal characteristics of different nodes and improve the clustering accuracy of HSI. By designing an isomorphic enhanced structure learning module, the graph structure can be updated according to different clustering tasks, thereby improving the clustering accuracy of HSI. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0035] Figure 1 This is an application environment diagram of a hyperspectral image clustering method in one embodiment of the present application;

[0036] Figure 2 A schematic diagram of a process of a hyperspectral image clustering method provided in one embodiment of the present application;

[0037] Figure 3 An implementation architecture diagram of a hyperspectral image clustering method provided in one embodiment of the present application;

[0038] Figure 4 A schematic diagram of the structure of a computer device provided in one embodiment of the present application. DETAILED DESCRIPTION

[0039] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0040] In order to make the above-mentioned objects, features and advantages of the present application more obvious and easy to understand, the present application is further described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0041] The hyperspectral image clustering method provided in the embodiment of the present application can be applied to Figure 1 In the application environment shown. Among them, the terminal 102 communicates with the server 104 through the network. The data storage system can store the data that the server 104 needs to process. The data storage system can be set up separately, integrated on the server 104, or placed on the cloud or other servers. The terminal 102 can send the hyperspectral image to the server 104. After the server 104 receives the hyperspectral image, it uses the superpixel segmentation method to convert the hyperspectral image into a superpixel level image. A clustering model including an adaptive filter graph encoder, a graph embedding clustering self-training decoder, and an isomorphic enhanced structure learning module is obtained. The superpixel level image is input into the clustering model to obtain a clustering result. Among them, the adaptive filter graph encoder is used to extract low-frequency features and high-frequency features. The graph embedding clustering self-training decoder is used for self-training, and the clustering result is obtained using the k-means algorithm. The isomorphic enhanced structure learning module is used to dynamically adjust the input graph structure according to different clustering tasks. The server 104 can feedback the obtained clustering results to the terminal 102. In addition, in some embodiments, the hyperspectral image clustering method can also be implemented independently by the server 104 or the terminal 102. For example, the terminal 102 can directly perform clustering processing on the hyperspectral image, or the server 104 can obtain the hyperspectral image from the data storage system and perform clustering processing on the hyperspectral image.

[0042] The terminal 102 may be, but is not limited to, various desktop computers, laptop computers, smart phones, tablet computers, IoT devices, and portable wearable devices. The IoT devices may be smart speakers, smart TVs, smart air conditioners, smart vehicle-mounted devices, etc. The portable wearable devices may be smart watches, smart bracelets, head-mounted devices, etc. The server 104 may be implemented as an independent server or a server cluster consisting of multiple servers, or may be a cloud server.

[0043] In an exemplary embodiment, Figure 2As shown, a hyperspectral image clustering method is provided. The method is executed by a computer device, and can be executed by a computer device such as a terminal or a server alone, or by a terminal and a server together. In the embodiment of the present application, the method is applied to Figure 1 The server 104 in the example is used as an example to illustrate, including the following steps 200 to 203. Among them:

[0044] Step 200: Acquire a hyperspectral image. A hyperspectral image is pixel-level image data.

[0045] Step 201: Using a superpixel segmentation method, the hyperspectral image is converted into a superpixel level map. The superpixel level map is a regional level map data.

[0046] Step 202: Obtain a clustering model. The clustering model is a network model trained using a sample data set. The clustering model includes an adaptive filter graph encoder, a graph embedding clustering self-training decoder, and an isomorphic enhanced structure learning module. The adaptive filter graph encoder is used to extract low-frequency features and high-frequency features. The graph embedding clustering self-training decoder is used for self-training and obtains clustering results using the k-means algorithm. The isomorphic enhanced structure learning module is used to dynamically adjust the input graph structure according to different clustering tasks.

[0047] Step 203: Input the superpixel level graph into the clustering model to obtain a clustering result.

[0048] Implementing the above steps 200 to 203 can improve the clustering accuracy of HSI. In addition, by designing an adaptive filter graph encoder, the present application can solve the problem that the filter in the existing graph network remains unchanged and more capabilities are required to process the signal characteristics of different nodes, resulting in poor clustering accuracy of large-scale HSI. By designing an isomorphic enhanced structure learning module, the problem that the mainstream graph network cannot update the original graph according to different clustering tasks can be solved.

[0049] In another exemplary embodiment of the present application, the hyperspectral image contains many pixels. If each pixel is used as the input of AHSGC, it requires a huge amount of calculation. In order to solve this problem, this embodiment proposes a superpixel segmentation method to generate uniform regions. This superpixel segmentation method converts HSI from pixels to regions while retaining local spatial spectral structure information and reducing the number of nodes. Based on this, the implementation process of the above step 201 can be described as:

[0050] (1) Linear discriminant analysis (LDA) is used to reduce the dimension of the hyperspectral image to obtain a reduced-dimensional image. This step uses LDA to reduce the dimension of the hyperspectral image, mainly to facilitate subsequent processing.

[0051] (2) The simple linear iterative clustering (SLIC) method is used to divide the reduced-dimensional image into multiple spatially connected superpixels.

[0052] (3) Each superpixel is regarded as a graph node, and the nearest neighbor relationship is established between superpixels to generate a superpixel-level graph.

[0053] For example, given a hyperspectral image , the hyperspectral image can be converted into an undirected graph (i.e., superpixel-level graph) by treating each superpixel as a node of the graph and establishing the nearest neighbor relationship between superpixels, where , and Respectively represent the height, width and number of bands of the hyperspectral image, , represents the number of superpixels. In addition, Represents the spectral feature of the input graph node, which is calculated as the average spectral feature of the pixels contained in the superpixel. Connection relationship between superpixels It is expressed as:

[0054] (1)

[0055] In the formula, Represents superpixel With superpixels The connection relationship between It is superpixels, express Contains pixels, express The number of pixels contained, Indicates Superpixels.

[0056] In another exemplary embodiment of the present application, in order to explore the connection between pixels and superpixels, an image back-projection operation can be designed to convert data features from pixels to superpixels. Based on this, the implementation process of the above step (2) can be described as:

[0057] (2-1) Set the correlation matrix of the simple linear iterative clustering method. For example, is the correlation matrix introduced by SLIC, defined as:

[0058] (2)

[0059] In the formula, Represents superpixel With superpixels The correlation matrix between .

[0060] (2-2) The image back-projection operation is performed on the reduced-dimensional image using the correlation matrix to convert the data features in the reduced-dimensional image from pixels to superpixels, and obtain multiple spatially connected superpixels. Among them, graph projection can encode the reduced-dimensional image into superpixel-level graph node features through matrix multiplication, expressed as:

[0061] (3)

[0062] In the formula, represents the super-pixel-level graph node features, represents matrix multiplication, It is column-normalized ,Right now , Represents column-wise normalization , Represents superpixel With superpixels The correlation matrix between . Represents the expansion of hyperspectral imagery according to spatial dimensions.

[0063] In another exemplary embodiment of the present application, the superpixel level graph is the input of the clustering model. In order to extract and filter the low-frequency and high-frequency features on the superpixel level graph for clustering, an adaptive filter graph encoder is designed by using an adaptive graph filter to aggregate neighbor information. The adaptive graph filter is associated with a Laplacian matrix, an affinity matrix and its variants, respectively. Low-pass filter and high pass filter It can be expressed as:

[0064] (4)

[0065] In the formula, is the normalized affine matrix of the random walk, expressed as , is the corresponding Laplacian matrix, expressed as . Represents the degree matrix of the graph.

[0066] Given a convolution kernel and image signal , the convolution operation can be expressed as:

[0067] (5)

[0068] In the formula, ,for Representation in the spectral domain. ⊙ is the matrix dot product. represents the eigenvector matrix.

[0069] if and is regarded as a convolution kernel, as shown in formula (4), then for the image signal The filtering can be expressed as:

[0070] (6)

[0071] In the formula, and yes and Representation in the spectral domain. If and , is the generated image in the spectral domain, then equation (6) can be written as:

[0072] (7)

[0073] In formula (7), if the eigenvector ,but The difference between the projection is amplified in the projection, and then a non-smoothed signal is obtained. If λ<1, the projection middle The difference between them is reduced and a smooth signal is obtained.

[0074] After analysis, it is noticed that low-pass filters can capture similarity information, while high-pass filters do the opposite. Therefore, both types of filters play a vital role in graph networks.

[0075] In order to capture the complete node information, an adaptive graph filter is designed , that is:

[0076] (8)

[0077] In the formula, the parameters represents a learnable parameter to balance the high-pass and low-pass information. is with The associated normalized Laplacian matrix. represents the eigenvalue, represents the feature vector, express The components in the frequency domain, is the order of the filter.

[0078] Based on the above description, in order to capture the high-pass and low-pass representations of the construction graph, the adaptive filter graph encoder in this embodiment The layers are stacked as:

[0079] (9)

[0080] In the formula, is the adaptive filter graph encoder stack The representation after the layer, yes The graph node representation after layer convolution, represents the adaptive graph filter, represents the adaptive filter graph encoder layer.

[0081] In another exemplary embodiment of the present application, the graph embedding clustering self-training decoder can use auxiliary distribution to self-train the clustering model, and then use the k-means algorithm to implement clustering. Based on this, the data processing process of the graph embedding clustering self-training decoder can be described as:

[0082] 1) Graph Embedding: Based on the above description, the adaptive filter graph encoder can extract high-level structural semantic information and simultaneously learn high-pass and low-pass representations of the graph.

[0083] In order to achieve an effective fusion representation of structural information and spectral features, this embodiment develops a linear transformation mechanism to encode the graph structure, as shown below:

[0084] (10)

[0085] In the formula, represents the output of the adaptive filter graph encoder, represents the output of layer 0 of the adaptive filter graph encoder, , Represents the linear transformation weight matrix.

[0086] 2) Self-training clustering: As mentioned above, high-pass and low-pass graph representations can be learned using adaptive filters. However, high-level structural semantic information cannot be used for clustering, and the lack of labels is also an obstacle to graph clustering.

[0087] To address these issues, this step designs a self-training clustering mechanism using auxiliary distributions. Specifically, the t distribution is used to measure the Cluster Centers and embedded adaptive filter graph encoder Output of the layer The similarity between nodes is as follows:

[0088] (11)

[0089] In the formula, Indicates Cluster centers, It is super pixel The clustering distribution of To cluster center Then the cluster distribution can be optimized by minimizing the KL divergence, that is:

[0090] (12)

[0091] (13)

[0092] In the formula, is the auxiliary distribution, as shown in formula (13), Will Raised to the second power, will be a sharp distribution and avoid the collapse of clusters into a single cluster. represents the clustering loss function, Also represents auxiliary distribution. It is super pixel The cluster assignment distribution of .

[0093] In another exemplary embodiment of the present application, in order to obtain sparse edges, this embodiment develops inter-cluster edge removal and intra-cluster edge recovery to dynamically adjust the graph structure according to different clustering tasks. Based on this, the data processing flow of the isomorphic enhanced structure learning module can be described as:

[0094] 1) Directed correlation estimation: Given an intermediate clustering result, the inner product is usually used to measure the pairwise correlation between nodes. However, in practice, it is found that this is a suboptimal method because the clustering results contain considerable noise during the network training process.

[0095] To solve this problem, this embodiment designs a hierarchical correlation estimation mechanism. Specifically, node similarity is estimated by combining the guidance of latent space and clustering in a hierarchical manner. First, the pseudo-label is calculated. The calculation formula of the pseudo-label is expressed as:

[0096] (14)

[0097] In the formula, represents a pseudo label.

[0098] According to formula (14), the pseudo labels of all nodes are obtained, and then in each cluster with the highest assignment probability Select the front Therefore, a node subset is obtained, namely:

[0099] (15)

[0100] In the formula, is the highest confidence node subset, Representation Node In cluster The cluster assignment ranking probability in , because one cluster corresponds to one cluster center, so the number of clusters is also adopted express.

[0101] By using Equation (15), the noise effect of node clustering can be reduced. However, if the edge weights are still determined by the cluster space, confident but incorrect fragments will still destroy the accuracy of the similarity matrix. To overcome these limitations, the latent space is used instead of the cluster space to estimate node similarity.

[0102] In formula (10), the node embedding matrix is ​​calculated , based on this, the similarity between all nodes can be estimated by dot product, expressed as:

[0103] (16)

[0104] In the formula, is the node similarity. Different from equation (16), this algorithm is based on Internal calculation , the corresponding node embedding similarity It is expressed as:

[0105] (17)

[0106] In the formula, Represents superpixel The embedding matrix of the nodes in Represents a collection, express The number of nodes in .

[0107] Formula (17) comprehensively utilizes the embedding similarity matrix and soft assignment to identify incorrect or missing edges, thereby significantly improving the accuracy of edge connections, thereby achieving more reliable and accurate graph structure learning.

[0108] 2) Graph edge sparsification: Most known methods keep the graph unchanged during network training, which means that the edges in the graph remain unchanged. Based on this, the known methods have two disadvantages, for example, incorrect edge connections in the graph cannot be corrected and bring unnecessary difficulties to clustering. To this end, this embodiment proposes a graph edge sparsification method, which includes inter-cluster edge removal and intra-cluster edge restoration, so that the nodes within the cluster are more closely connected and the nodes between clusters are more sparse.

[0109] Among them, 2-1) Intra-cluster edge recovery: In order to increase the connection within the cluster, add edges to the cluster Inclusive of the previous percentage The maximum similarity The node pairs are:

[0110] (18)

[0111] In the formula, represents the restored edge, express The number of nodes in is the edge set on the current graph learning, Represents a node pair The similarity is Arranged in order, represents the j-th pseudo label.

[0112] It should be noted that existing edges are not considered in the ranking process. Based on this, the complete set of intra-cluster edges is recovered It is expressed as:

[0113] (19)

[0114] 2-2) Inter-cluster edge removal: This strategy aims to remove unnecessary edge connections between clusters, weaken inter-cluster connections, and promote clustering. The expression is as follows:

[0115] (20)

[0116] in, is the percentage of edges to be retained. represents the removed edge, Represents the similarity between the i-th node and the j-th node, Indicates that the node similarities are arranged in order.

[0117] 3) Homogeneous Enhanced Structure Learning: Equations (19) and (20) show the restored and removed edge sets. The edge set in the graph learning process can be expressed as , . Adjacency Matrix It can be calculated as:

[0118] (twenty one)

[0119] In the formula, and yes and The adjacency matrix of .

[0120] Then, the graph is reconstructed by updating the adjacency matrix, that is:

[0121] (twenty two)

[0122] In the formula, is the image reconstruction loss.

[0123] In another exemplary embodiment of the present application, in order to improve the structural consistency and graph edge sparsity of the graph and train the proposed network, this embodiment proposes a new joint objective loss, namely, a joint network optimization loss function.

[0124] In this embodiment, the self-training clustering loss is shown in equation (12), and the structure graph reconstruction loss is shown in equation (22). Then the overall goal of the clustering model provided by this application is It can be expressed as:

[0125] (twenty three)

[0126] In the clustering model, The hyperparameters are considered as “true values” by minimizing the overall objective conduct Iteration update. Node The clustering results It is expressed as:

[0127] (twenty four)

[0128] Therefore, the adaptive filter graph encoder set in this application is mainly used to capture the potential features of the graph and generate a graph embedding representation Then, based on the initialized cluster centers , using k-means algorithm Perform clustering and obtain clustering results.

[0129] Based on the above description, combined with Figure 3 The implementation architecture of the hyperspectral image clustering method provided in this application is shown in Table 1.

[0130] Table 1 Description of the implementation process of hyperspectral image clustering method

[0131]

[0132] In Table 1, AHSGC represents the hyperspectral image clustering method provided by this application.

[0133] Based on the same inventive concept, the embodiment of the present application also provides a hyperspectral image clustering system for implementing the hyperspectral image clustering method involved above. The implementation scheme for solving the problem provided by the system is similar to the implementation scheme recorded in the above method, so the specific limitations in one or more hyperspectral image clustering system embodiments provided below can refer to the limitations on the hyperspectral image clustering method above, and will not be repeated here.

[0134] In an exemplary embodiment, a hyperspectral image clustering system is provided, the system comprising:

[0135] The data acquisition module is used to acquire hyperspectral images. Hyperspectral images are pixel-level image data.

[0136] The initial graph construction module is used to convert the hyperspectral image into a super-pixel level graph using the super-pixel segmentation method. The super-pixel level graph is the regional level graph data.

[0137] The model acquisition module is used to obtain the clustering model. The clustering model is a network model trained using a sample data set. The clustering model includes an adaptive filter graph encoder, a graph embedding clustering self-training decoder, and an isomorphic enhanced structure learning module. The adaptive filter graph encoder is used to extract low-frequency features and high-frequency features. The graph embedding clustering self-training decoder is used for self-training and obtains clustering results using the k-means algorithm. The isomorphic enhanced structure learning module is used to dynamically adjust the input graph structure according to different clustering tasks.

[0138] The clustering module is used to input the superpixel level graph into the clustering model to obtain the clustering result.

[0139] In an exemplary embodiment, the hyperspectral image clustering system provided may be composed of five modules, namely, an adaptive filter graph encoder, a graph embedding clustering self-training decoder, an isomorphic enhanced structure learning module, a joint network optimization module, and an HSI preprocessing and initial graph construction module. The main functions of the five modules are as follows:

[0140] (1) HSI preprocessing and initial graph construction module: A superpixel segmentation method is established to convert HSI from pixels to regions while preserving the local spatial spectral structure information and reducing the number of nodes.

[0141] (2) Adaptive filter graph encoder: Taking the superpixel level graph as its input, an adaptive filter graph encoder is designed by aggregating neighbor information using an adaptive graph filter.

[0142] (3) Graph embedding clustering self-training decoder: The clustering model is self-trained using auxiliary distribution, and then clustering is achieved using the k-means algorithm.

[0143] (4) Isomorphic enhanced structure learning module: In order to obtain sparse edges, inter-cluster edge removal and intra-cluster edge recovery are developed, which can dynamically adjust the graph structure according to different clustering tasks.

[0144] (5) Joint network optimization module: This module proposes a new joint objective loss to improve the structural consistency and edge sparsity of the graph and train the proposed network. Finally, the k-means algorithm is used to represent the latent features.

[0145] In actual application, the five modules can be connected in an end-to-end network for HSI clustering, and each part interacts with each other.

[0146] In an exemplary embodiment, a computer device is provided. The computer device may be a server or a terminal. The internal structure diagram thereof may be as follows: Figure 4 As shown. The computer device includes a processor, a memory, an input / output interface (Input / Output, referred to as I / O) and a communication interface. The processor, the memory and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store clustering processing data. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a hyperspectral image clustering method is implemented.

[0147] Those skilled in the art will understand that Figure 4 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components. In an exemplary embodiment, a computer device is provided, including a memory and a processor, wherein a computer program is stored in the memory, and the processor implements the steps in the above-mentioned method embodiments when executing the computer program.

[0148] In an exemplary embodiment, a computer-readable storage medium is provided, storing a computer program, and when the computer program is executed by a processor, the steps in the above method embodiments are implemented.

[0149] In an exemplary embodiment, a computer program product is provided, including a computer program, and when the computer program is executed by a processor, the steps in the above method embodiments are implemented.

[0150] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant regulations.

[0151] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to the memory, database or other medium used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM may be in various forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0152] The database involved in each embodiment provided in this application may include at least one of a relational database and a non-relational database. The non-relational database may include a distributed database based on blockchain, etc., but is not limited thereto. The processor involved in each embodiment provided in this application may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., but is not limited thereto.

[0153] The technical features of the above embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0154] This article uses specific examples to illustrate the principles and implementation methods of this application. The description of the above embodiments is only used to help understand the method and core ideas of this application. At the same time, for those skilled in the art, according to the ideas of this application, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting this application.

Claims

1. A hyperspectral image clustering method, characterized in that: The hyperspectral image clustering method comprises: Acquire a hyperspectral image; the hyperspectral image is pixel-level image data; The hyperspectral image is converted into a super-pixel level image by using a super-pixel segmentation method; the super-pixel level image is a regional level image data; Obtain a clustering model; the clustering model is a network model trained using a sample data set; the clustering model includes an adaptive filter graph encoder, a graph embedding clustering self-training decoder, and an isomorphic enhancement structure learning module; the adaptive filter graph encoder is used to extract low-frequency features and high-frequency features; the graph embedding clustering self-training decoder is used to perform self-training and obtain clustering results using a k-means algorithm; the isomorphic enhancement structure learning module is used to dynamically adjust the input graph structure according to different clustering tasks; Inputting the superpixel level graph into the clustering model to obtain a clustering result; The t-layer stack of the adaptive filter graph encoder is: ； In the formula, represents the input graph node spectrum features, is the representation of the adaptive filter graph encoder after stacking t layers, It is the graph node representation after t layers of convolution, represents the adaptive graph filter, represents the adaptive filter graph encoder layer; In the adaptive filter graph encoder, a linear transformation weight matrix is ​​used to encode the super-pixel level graph, which is expressed as: ； In the formula, represents the output of the adaptive filter graph encoder, represents the output of layer 0 of the adaptive filter graph encoder, Represents the linear transformation weight matrix.

2. The hyperspectral image clustering method according to claim 1, characterized in that: The hyperspectral image is converted into a super-pixel level image using a super-pixel segmentation method, specifically including: Performing dimensionality reduction processing on the hyperspectral image by using linear discriminant analysis to obtain a dimensionality reduction image; Using a simple linear iterative clustering method to divide the reduced-dimensional image into a plurality of spatially connected superpixels; Each superpixel is regarded as a graph node, and a nearest neighbor relationship is established between superpixels to generate the superpixel level graph.

3. The hyperspectral image clustering method according to claim 2, characterized in that: The reduced-dimensional image is divided into a plurality of spatially connected superpixels using a simple linear iterative clustering method, specifically comprising: Set the correlation matrix for the simple linear iterative clustering method; The correlation matrix is ​​used to perform an image back-projection operation on the reduced-dimensional image, and data features in the reduced-dimensional image are converted from pixels to super-pixels to obtain a plurality of spatially connected super-pixels.

4. The hyperspectral image clustering method according to claim 1, characterized in that: The graph is embedded in the clustered self-training decoder, and the t distribution is used to measure the Cluster Centers and embed The similarity between nodes is calculated, and the cluster distribution is optimized by minimizing the KL divergence; represents the output of the i-th layer of the adaptive filter graph encoder.

5. A hyperspectral image clustering system, characterized in that: The hyperspectral image clustering system is applied to the hyperspectral image clustering method according to any one of claims 1 to 4; the hyperspectral image clustering system comprises: A data acquisition module, used to acquire a hyperspectral image; the hyperspectral image is pixel-level image data; An initial graph construction module is used to convert the hyperspectral image into a super-pixel level graph by using a super-pixel segmentation method; the super-pixel level graph is a regional level graph data; A model acquisition module is used to acquire a clustering model; the clustering model is a network model trained using a sample data set; the clustering model includes an adaptive filter graph encoder, a graph embedding clustering self-training decoder, and an isomorphic enhancement structure learning module; the adaptive filter graph encoder is used to extract low-frequency features and high-frequency features; the graph embedding clustering self-training decoder is used to perform self-training and obtain clustering results using a k-means algorithm; the isomorphic enhancement structure learning module is used to dynamically adjust the input graph structure according to different clustering tasks; The clustering module is used to input the superpixel level graph into the clustering model to obtain a clustering result.

6. A computer device comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the hyperspectral image clustering method according to any one of claims 1 to 4.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the hyperspectral image clustering method described in any one of claims 1 to 4 is implemented.

8. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the hyperspectral image clustering method described in any one of claims 1 to 4 is implemented.

Images