Hyperspectral deep learning network reconstruction method and device based on graph structure and heat map

Abstract

The application discloses a hyperspectral deep learning network reconstruction method and device based on a graph structure and a heat map, and comprises the following steps: acquiring multi-target and background data of a hyperspectral image in a complex environment; performing feature extraction on the hyperspectral data through a deep learning network, performing importance calculation on traditional features and deep features of the hyperspectral image, and acquiring a hyperspectral image correlation feature weight density measurement standard; performing correlation mapping on the traditional features and the deep features, and acquiring a feature corresponding dictionary; introducing a hidden variable constraint function into the network by using the feature corresponding dictionary, and acquiring a reconstructed graph structure hyperspectral deep learning network; visualizing a network feature parameter and a space spectrum dimension by using image features acquired through feature extraction, and acquiring a reconstructed full cycle period deep learning network. The method and device can reconstruct a graph structure and a heat map deep learning network for a hyperspectral image, and have high accuracy.

Description

Technical Field

[0001] This invention relates to the field of hyperspectral image data processing, and in particular to a method and apparatus for reconstructing hyperspectral deep learning networks based on graph structures and visualized heatmaps in complex environments. Background Technology

[0002] Currently, in the process of data processing using deep learning networks, the application of extracted physical meaning of features has always been a challenge in this field. Existing feature extraction techniques can be mainly divided into traditional feature extraction methods and deep learning-based feature extraction methods. Traditional feature extraction methods use the image's own feature information as the extraction target, such as grayscale values, edges, and textures. Deep learning-based feature extraction methods extract image information features through convolutional network models.

[0003] These methods demonstrate good performance in target detection, rapid identification, and accurate localization. Traditional feature extraction methods such as LBP and SIFT offer some adaptability to changes in brightness, angle, and noise. However, due to differences in networks, different methods exhibit significant variations in computation time and results. For instance, when extracting features from images, HOG requires image segmentation, extracting local features. Furthermore, the varying resolutions and image sizes of different hyperspectral images result in substantial differences in computation speed. Currently, there is no method that can combine global and target information and accelerate extraction by pruning target features. Summary of the Invention

[0004] This invention provides a method and apparatus for reconstructing hyperspectral deep learning networks based on graph structures and heatmaps, in order to solve the technical problem that existing technologies are difficult to apply in a targeted manner to extract the physical meaning of features.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0006] In a first aspect, embodiments of this application propose a method for reconstructing a hyperspectral deep learning network based on graph structure and heatmap, comprising the following steps:

[0007] Acquire multi-target and background data from hyperspectral images in complex environments;

[0008] Feature extraction of hyperspectral data is performed using deep learning networks. The importance of traditional and deep features of hyperspectral images is calculated to obtain a dense metric standard for the weights of associated features in hyperspectral images.

[0009] Perform an association mapping between traditional features and deep features to obtain a feature-correspondence dictionary;

[0010] By utilizing a feature-correspondence dictionary, a latent parameter constraint function is introduced into the network to obtain a reconstructed graph-structured hyperspectral deep learning network.

[0011] By utilizing the image features obtained through feature extraction, the network feature parameters and spatial spectral dimension are represented by a visual heatmap to obtain the reconstructed full-cycle deep learning network.

[0012] The image is a hyperspectral image under complex conditions, and the target includes multiple targets and the background. Preferably, the hyperspectral image includes naturally captured images and images synthesized in post-processing, and the multiple targets include single-target and multi-target information.

[0013] Furthermore, the feature extraction of hyperspectral data, including the importance calculation of traditional features and depth features of the hyperspectral image, includes:

[0014] By analyzing the target characteristics, background characteristics, and multi-target distribution characteristics of hyperspectral images under complex backgrounds, target-to-target correlation matrices and target-to-background correlation matrices are constructed.

[0015] Feature extraction is performed on the spatial texture and morphological features, spatial neighborhood features, and spectral features of the target and background. The feature importance of the target and background is calculated, and a feature importance matrix of the target and background is constructed.

[0016] Furthermore, through the target-to-target association matrix, the target-to-background association matrix, and the feature importance matrices of the target and background, a dense metric for the weighted association features of hyperspectral images is obtained, including:

[0017] Through the target-target correlation matrix R TT With the target feature importance matrix I T The dense matrix D of target-to-target correlation feature weights is calculated. TT ;

[0018] Through the target-background correlation matrix R TB With background feature importance matrix I B The dense matrix D of the target-background association feature weights is calculated. TB .

[0019] Furthermore, the step of associating and mapping traditional features with deep features to obtain a feature-corresponding dictionary includes:

[0020] By aligning across feature vectors, two feature vectors will be used. s and l t The traditional feature vector distribution X∈R obtained by training a single feature library d×n With depth feature vector distribution Y∈R d×wMapped to a common vector space, where l s Indicating traditional characteristics, l t Let d represent the dimension of the feature vector, and n and w represent the number of feature vectors in the two vector spaces, respectively, such that the distance between feature vectors with the same characteristics in this shared vector space is minimized. Feature-to-feature translation is performed using aligned feature vectors with the same characteristics to obtain a feature-to-feature dictionary.

[0021] Furthermore, the step of using a feature-correspondence dictionary to introduce the latent parameter constraint function into the network to obtain the reconstructed graph-structured hyperspectral deep learning network includes:

[0022] Based on the feature mapping dictionary, by applying soft constraints to multiple hyperparameters in the network, a hidden parameter constraint function C is introduced to constrain the entire network model by a linear function.

[0023] The feature map structure is iterated continuously and visualized in the form of a graph structure to obtain the reconstructed graph structure hyperspectral deep learning network.

[0024] Furthermore, the iterative process of the feature map structure, visualized through a graph structure, includes:

[0025] The target element logical graph structure unit treats each node in the graph data as a feature of a target or environment, and the link edges between nodes represent the dependency relationship between two features.

[0026] By using iterative theory to calculate the state of nodes in the logical graph structure of the target elements, and through continuous iteration, the attribute characteristics of the graph nodes eventually converge and reach an equilibrium state.

[0027] Furthermore, the step of using the image features obtained through feature extraction to visualize the network feature parameters and spatial spectral dimension using a heatmap to obtain the reconstructed fully recurrent deep learning network includes:

[0028] By extracting high-level features from hyperspectral data, the feature parameters in the network are represented by a visual heatmap to obtain the feature types of the images extracted by the network, the features of the images, and the regions most sensitive to the network.

[0029] By training the spatial location and spectral information in hyperspectral images with saliency weights, effectively extracting bands in the spectral dimension, and pruning features in the spatial dimension, a reconstructed full-cycle deep learning network is obtained after importance analysis and weight allocation.

[0030] Secondly, embodiments of this application also propose a hyperspectral deep learning network reconstruction device based on graph structure and heatmap, used to implement the method described in any embodiment of the first aspect of this application, including: an acquisition module, a feature module, and a learning module; the learning module further includes a dictionary unit, a first reconstruction unit, and a second reconstruction unit, and the device is configured to implement the steps of the method described in any embodiment of the first aspect of this application.

[0031] This application also proposes a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in any embodiment of the first aspect of this application.

[0032] This application also proposes an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the method described in any embodiment of the first aspect of this application.

[0033] The beneficial effects of the technical solution provided by this invention include at least the following:

[0034] The present invention provides a method and apparatus for reconstructing hyperspectral deep learning networks based on graph structures and visualized heatmaps. First, it utilizes traditional and deep features to obtain a dense metric standard for the weights of hyperspectral image-related features, constructing a feature-correspondence dictionary. Then, it reconstructs the network using a constraint function with hidden parameters. Finally, it prunes and reconstructs the network using saliency weights based on spatial spectral information, achieving high-precision and rapid deep learning network reconstruction. The reconstruction method provided by this invention can achieve network reconstruction that highlights target features for complex environments with multiple hyperspectral targets and background data. Attached Figure Description

[0035] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0036] Figure 1 This is a flowchart illustrating the hyperspectral deep learning network reconstruction method based on graph structure and heatmap provided in this embodiment of the invention.

[0037] Figure 2 This is a schematic flowchart of a method for obtaining a dense metric standard for the weighted association features of hyperspectral images, provided in an embodiment of the present invention.

[0038] Figure 3 This is a flowchart illustrating the method for obtaining a feature-corresponding dictionary provided in an embodiment of the present invention;

[0039] Figure 4 This is a schematic diagram of the method for obtaining a reconstructed graph structure deep learning network provided in an embodiment of the present invention;

[0040] Figure 5 This is a schematic diagram of the method for obtaining the reconstructed full-cycle deep learning network provided in an embodiment of the present invention;

[0041] Figure 6 This is an embodiment of the hyperspectral deep learning network reconstruction device based on graph structure and heat map of the present invention;

[0042] Figure 7 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0043] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0044] Embodiments of the present invention provide a method for reconstructing a hyperspectral deep learning network based on graph structure and heatmap, such as... Figure 1 As shown, the method includes the following steps:

[0045] Step 110: Acquire multi-target and background data from hyperspectral images under complex environments;

[0046] Step 120: Extract features from hyperspectral data using a deep learning network, calculate the importance of traditional and deep features of hyperspectral images, and obtain a dense metric standard for the weights of hyperspectral image association features.

[0047] Step 130: Obtain the feature-corresponding dictionary by performing an association mapping between traditional features and deep features;

[0048] Step 140: Using the feature correspondence dictionary, the hidden parameter constraint function is introduced into the network to obtain the reconstructed graph structure hyperspectral deep learning network;

[0049] Step 150: Using the image features obtained from feature extraction, visualize the network feature parameters and spatial spectral dimension using a heatmap to obtain the reconstructed full-cycle deep learning network.

[0050] This invention addresses the challenges of diverse features extracted from multiple targets and backgrounds in hyperspectral data under complex backgrounds. It proposes a densely distributed weighting metric for hyperspectral image association features to achieve optimal feature category identification. To address the issue of unclear physical meaning in features extracted by deep learning networks, it proposes an association mapping of feature vector space and constructs a feature-correspondence dictionary to represent the physical meaning of similar features. Furthermore, to address the difficulty in associating multiple targets and backgrounds in complex backgrounds, it introduces a latent parameter constraint function into the network to obtain graph-structured feature association representations. Finally, considering the large volume, complex features, and dispersed sensitive regions of hyperspectral image data, it proposes visualizing network feature parameters and spatial-spectral dimensions using heatmaps, pruning the network, and obtaining full-lifecycle sensitive heatmap representations. This invention can reconstruct deep learning networks for target hyperspectral images under complex backgrounds, accurately representing sensitive regions in a focused manner.

[0051] Furthermore, such as Figure 2 As shown, step 120 involves extracting features from hyperspectral data using a deep learning network, calculating the importance of traditional and deep features of the hyperspectral image, obtaining a dense metric for the weights of associated hyperspectral image features, and achieving optimal feature category recognition. Specifically, this includes:

[0052] By analyzing the relationships between different objectives, an association matrix R between objectives is constructed. TT The matrix is ​​as follows:

[0053]

[0054] In this context, the diagonal lines of the correlation matrix represent the relationships between identical targets, and therefore all are 1; x 1n and x n1 Both represent the relationship between target 1 and target n, therefore x 1n =x n1 ;

[0055] By analyzing the relationship between different objectives and backgrounds, an association matrix R between objectives and backgrounds is constructed. TB The matrix is ​​as follows:

[0056]

[0057] Where, x nm This represents the relationship between the target n and the background m.

[0058] By calculating the feature importance of each target and background in the image, a target feature importance matrix and a background feature importance matrix are constructed. Target Feature Importance Matrix I T =[i t1 i t2 i t3… i tn ], where i tn This represents the feature importance of the nth target. Calculate the dense matrix D of target-target association weights. TT =I T ·R TT , where D TT Given a 1×n matrix. By detecting D... TT The validity of a feature is determined by whether any element in the matrix is ​​zero. Background Feature Importance Matrix I B =[i b1 i b2 i b3 … i bm ], where i bn This represents the feature importance of the m-th background element. Calculate the dense matrix of feature weights associated with the target and background. Among them, D TB Given a 1×n matrix. By detecting D... TB The validity of a feature is determined by whether the elements in the array contain zeros.

[0059] Furthermore, such as Figure 3 As shown, step 130, which involves associating traditional features with deep features to obtain a feature-corresponding dictionary, specifically includes:

[0060] The model uses a generative adversarial network to achieve feature mapping by employing a cross-feature vector alignment method.

[0061] Cross-feature vector alignment refers to mapping feature vectors from different features to a common vector space, where feature vectors with similar characteristics are also close in distance. In deep learning, aligned feature vectors with similar characteristics are used for feature-to-feature translation.

[0062] Specifically, assume X∈R d×n and Y∈R d×w These represent the use of two features l s and l t The feature vector space distribution obtained by training a single feature library, where l s Indicating traditional characteristics, l t Let represent depth features, d represent the dimension of the feature vector, and n and w represent the number of feature vectors in the two vector spaces, respectively. Then a parameter matrix W∈R is needed. d×d To map X to the vector space containing Y, and to minimize the distance between feature vectors of features with the same characteristics: Among them, ||·|| F This represents the F-norm, used for distance calculations.

[0063] Furthermore, a GAN (Generative Adversarial Network) is first trained to learn an initial W. The discriminator D in the GAN aims to distinguish, as far as possible, whether vectors in the same vector space after mapping belong to feature l. s or feature l t Assuming The discriminator predicts the feature vector e based on feature l. s The probability of the vector obtained by mapping the feature vector in the matrix, where θ D Let represent the trainable parameters of the discriminator. Then, the objective function for training the discriminator based on the cross-entropy loss function is:

[0064] Where x i and y i Both represent a vector in the vector space. In contrast, the generator in a GAN is trained by mapping matrix W, making it impossible for the discriminator to predict which feature a feature vector belongs to. The objective function for training the generator is:

[0065]

[0066] The most frequent feature pairs are selected from the dictionary as anchor points, and W is fine-tuned to further align the distribution spaces of the two feature vectors. For example... Figure 3 As shown, the dashed irregular shape represents the depth feature vector space, and the solid irregular shape represents the traditional feature vector space. In the same aligned vector space, the shapes of the depth and traditional vector spaces roughly overlap. Circles represent the positions of certain depth features in the vector space, and triangles represent the positions of certain traditional features in the vector space.

[0067] Furthermore, such as Figure 4 As shown, step 150 is implemented by using a feature correspondence dictionary to introduce the latent parameter constraint function into the network to obtain the reconstructed graph-structured hyperspectral deep learning network, specifically including:

[0068] By using a mapping model between deep features and traditional features, target elements (such as target spectral features, background spectral features, etc.) are extracted from hyperspectral images to construct target element logical graph units. By making full use of the correlation between targets and between targets and the background, it is possible to determine whether a target is a real target.

[0069] In this context, the target element logical graph structure unit treats each node in the graph data as a feature of a target or background. The connecting edges between nodes represent the dependency relationship between two features. Therefore, the state vector of each node is composed of its own state information and the state information of its neighboring nodes.

[0070] The target element logical graph structure uses iterative theory to calculate the state of the nodes in the graph. Through continuous iteration, the attribute features of the graph nodes are recursively propagated until convergence, and finally the whole graph reaches a balanced state.

[0071] Furthermore, a latent parameter constraint function is constructed to apply soft constraints to the changes in latent parameters in the deep learning network, thereby reconstructing the network. To make the network insensitive to input perturbations and improve its generalization performance, ||f| needs to be very small. w (x1)-f w (x2)|| should also be as small as possible, where f w (x) represents the simplified result of the model. To minimize the above expression, a constraint with controllable variance is constructed, i.e., a function C is constructed such that ||f w (x1)-f w (x2)||≤C(w)·||x1-x2|| always holds true.

[0072] By constraining and controlling multiple hyperparameters contained in the network, the reconstruction of graph-structured hyperspectral deep learning networks can be achieved.

[0073] Furthermore, such as Figure 5 As shown, step 160 involves using the image features obtained through feature extraction to visualize the network feature parameters and spatial spectral dimension using a heatmap, thereby obtaining the reconstructed full-cycle deep learning network. This specifically includes:

[0074] By visualizing the convolutional kernels of the deep learning feature extraction network, we can understand the types of image features extracted by the network; by visualizing intermediate activations, we can obtain the extracted image features; and finally, by visualizing the activation response categories using heatmaps, we can easily find the regions most sensitive to the network, thus achieving a visual representation of the entire lifecycle of the deep learning network.

[0075] Furthermore, based on the visualization of features of the full-cycle deep learning network, for different recognition target tasks, the importance analysis of target image elements is carried out by combining the attention prior knowledge of the spectrum and space of hyperspectral target recognition. Band extraction and feature pruning are performed in the spatial and spectral dimensions respectively. After target importance analysis and weight allocation, the deep learning model is updated and iterated to obtain a deep learning network with optimized recognition performance.

[0076] Specifically, spatial dimension image feature pruning is performed. By utilizing spatial dimension attention conclusions, the target and its surroundings are located during the network input stage, and feature pruning is performed to avoid irrelevant feature input, reduce network computation and running time, and effectively improve recognition accuracy.

[0077] Spectral bands are extracted in the spectral dimension. The corresponding feature band weights are obtained through feature visualization by deep learning network. Based on this prior knowledge, feature bands are extracted to reduce the hyperspectral dimension, making the network lighter and improving recognition efficiency.

[0078] To implement the method of any embodiment of the first aspect of this application, embodiments of this application also propose a hyperspectral deep learning network reconstruction device based on graph structure and heatmap, such as... Figure 6 As shown, including

[0079] The acquisition module 510 is used to acquire multi-target and background data of hyperspectral images under complex environments, as in step 110;

[0080] Feature module 520 is used to extract features from hyperspectral data through a deep learning network, calculate the importance of traditional features and deep features of hyperspectral images, and obtain a dense metric standard for the weights of hyperspectral image association features, as in step 120.

[0081] Learning module 500 further includes dictionary unit 530, first reconstruction unit 540, and second reconstruction unit 550.

[0082] The dictionary unit is used to perform association mapping between traditional features and deep features to obtain the feature-corresponding dictionary, as in step 130;

[0083] The first reconstruction unit is used to introduce the hidden parameter constraint function into the network using the feature correspondence dictionary to obtain the reconstructed graph structure hyperspectral deep learning network, as in step 140.

[0084] The second key unit utilizes the image features obtained through feature extraction to visualize the network feature parameters and spatial spectral dimension using a heatmap, thereby obtaining the reconstructed full-cycle deep learning network, as in step 150.

[0085] Furthermore, embodiments of further optimizations for implementing steps 110-150 in each module of the device of this application are provided in the appendix. Figures 2-5 The relevant explanations will not be repeated here.

[0086] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0087] Therefore, this application also proposes a computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the methods described in any embodiment of this application.

[0088] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0089] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0090] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0091] Figure 7 This is a schematic diagram of the structure of an electronic device (or computing device) provided in an embodiment of this application. In a typical configuration, the computing device includes one or more processors (CPUs), input / output interfaces, network interfaces, and memory. Memory may include non-persistent storage in computer-readable media, random access memory (RAM), and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0092] The electronic device 600 shown is merely an example and should not be construed as limiting the functionality and scope of use of the embodiments of this application. It includes: one or more processors 620; and a storage device 610 for storing one or more programs, which, when executed by the one or more processors 620, cause the one or more processors 620 to implement the method described in any embodiment of the first aspect of this application.

[0093] The electronic device 600 also includes an input device 630 and an output device 640; the processor 620, storage device 610, input device 630 and output device 640 in the electronic device can be connected by a bus or other means, as shown in the figure, which is connected by a bus 650.

[0094] Storage device 610, as a computer-readable storage medium, can be used to store software programs, computer-executable programs, and module units, such as the program instructions corresponding to the cloud bottom height determination method in the embodiments of this application. Storage device 610 may mainly include a program storage area and a data storage area. The program storage area may store the operating system and at least one application program required for a function; the data storage area may store data created based on terminal usage. Furthermore, storage device 610 may include high-speed random access memory and non-volatile memory, such as at least one disk storage device, flash memory device, or other non-volatile solid-state storage device. In some instances, storage device 610 may further include memory remotely located relative to processor 620, and these remote memories can be connected via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0095] Input device 630 can be used to receive input digital, character, or voice information, and to generate key signal inputs related to user settings and function control of the electronic device. Output device 640 may include electronic devices such as a display screen and a speaker.

[0096] It should also be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal device. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal device that includes said element.

[0097] Finally, it should be noted that the above description represents a preferred embodiment of the present invention. It should be pointed out that although preferred embodiments have been described, those skilled in the art, once they understand the basic inventive concept of the present invention, can make various improvements and modifications without departing from the principles described herein. These improvements and modifications should also be considered within the scope of protection of the present invention. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the embodiments of the present invention.

Claims

1. A hyperspectral deep learning network reconstruction method based on graph structure and heat map, characterized in that, The method comprises the following steps: Obtaining multi-target and background data of hyperspectral images in a complex environment; Performing feature extraction on the hyperspectral data, calculating the importance of the traditional features and the deep features of the hyperspectral images, and obtaining a hyperspectral image correlation feature weight density measurement standard; Performing correlation mapping on the traditional features and the deep features, and obtaining a feature corresponding dictionary; Introducing a hidden variable constraint function into the network by using the feature corresponding dictionary, and obtaining a reconstructed graph structure hyperspectral deep learning network; Visualizing the network feature parameters and the spatial and spectral dimensions by using the image features obtained by feature extraction, and obtaining a reconstructed full cycle period deep learning network.

2. The hyperspectral deep learning network reconstruction method based on graph structure and heat map according to claim 1, wherein, The feature extraction on the hyperspectral data and the calculation of the importance of the traditional features and the deep features of the hyperspectral images comprise: Constructing a target-target correlation matrix and a target-background correlation matrix based on the target characteristics, the background characteristics and the multi-target distribution characteristics of the hyperspectral images in a complex background; Performing feature extraction on the spatial texture and morphological features, the spatial neighborhood features and the spectral features of the target and the background, calculating the feature importance of the target and the background, and constructing a feature importance matrix of the target and the background.

3. The hyperspectral deep learning network reconstruction method based on graph structure and heat map according to claim 2, wherein, The hyperspectral image correlation feature weight density measurement standard is obtained based on the target-target correlation matrix, the target-background correlation matrix and the feature importance matrix of the target and the background, and comprises: By target and target association matrix With target feature importance matrix Calculate target and target association feature weight density matrix ; By target and background association matrix With background feature importance matrix The target and background association feature weight density matrix is calculated . 4.The hyperspectral deep learning network reconstruction method based on graph structure and heat map of claim 1, wherein, The correlation mapping on the traditional features and the deep features and the obtaining of the feature corresponding dictionary comprise: By aligning across feature vectors, two features will be used. and Traditional feature vector distribution obtained by training a single feature library With depth feature vector distribution Mapped into a common vector space, where Indicating traditional characteristics, Let d represent the depth feature, n and w represent the number of feature vectors in the two vector spaces, and let w be the number of feature vectors in the two vector spaces. The goal is to minimize the distance between feature vectors with the same characteristics in the common vector space. The aligned feature vectors with the same characteristics are used to translate features into features and obtain the feature-to-feature dictionary.

5. The hyperspectral deep learning network reconstruction method based on graph structure and heat map according to claim 1, wherein, The introduction of the hidden variable constraint function into the network by using the feature corresponding dictionary and the obtaining of the reconstructed graph structure hyperspectral deep learning network comprise: Based on the feature mapping dictionary, the multiple hyperparameters in the network are controlled by soft constraint, the hidden variable constraint function C is introduced, and the entire network model is constrained by a linear function; The feature graph structure is iterated constantly, visualized in the form of a graph structure, and the reconstructed graph structure hyperspectral deep learning network is obtained.

6. The hyperspectral deep learning network reconstruction method based on graph structure and heat map according to claim 5, wherein, The constant iteration of the feature graph structure and the visualization in the form of a graph structure comprise: The target element logical graph structure unit regards each node in the graph data as a target or environmental feature, and the link between the nodes represents the dependency between the two features. The state of the node in the graph is calculated by using the iterative theory on the target element logical graph structure, the attribute features of the graph nodes are finally converged, and a balance state is reached.

7. The hyperspectral deep learning network reconstruction method based on graph structure and heat map of claim 1, wherein, The visualization of the network feature parameters and the spatial and spectral dimensions by using the image features obtained by feature extraction, and the obtaining of the reconstructed full cycle period deep learning network comprise: The feature parameters in the network are visualized by using the high-level features extracted from the hyperspectral data, the feature types of the network extracted images, the features of the images and the most sensitive regions of the network are obtained, and the reconstructed full cycle period deep learning network is obtained. The spatial position and the spectral information in the hyperspectral images are subjected to significant weight training, the spectral dimension is subjected to effective band extraction, the spatial dimension is subjected to feature pruning, importance analysis and weight distribution are performed, and the reconstructed full cycle period deep learning network is obtained.

8. A hyperspectral deep learning network reconstruction device based on graph structure and heat map, for implementing the method of any one of claims 1-7, characterized in that, The method comprises the following steps: An acquisition module is configured to acquire multi-target and background data of a hyperspectral image in a complex environment; A feature module is configured to perform feature extraction on the hyperspectral data through a deep learning network, to perform importance calculation on traditional features and deep features of the hyperspectral image, and to acquire a hyperspectral image correlation feature weight density metric standard; A learning module includes a dictionary unit, a first reconstruction unit and a second reconstruction unit; The dictionary unit is configured to perform correlation mapping on the traditional features and the deep features, and to acquire a feature corresponding dictionary; The first reconstruction unit is configured to introduce an implicit variable constraint function into the network by using the feature corresponding dictionary, and to acquire a reconstructed graph structure hyperspectral deep learning network; The second reconstruction unit is configured to perform visualization thermograph representation on network feature parameters and spatial spectral dimensions by using image features acquired through feature extraction, and to acquire a reconstructed full cycle period deep learning network.

9. A computer readable storage medium having stored thereon a computer program, characterized in that, The program is executed by the processor to implement the method of any one of claims 1-7.

10. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the computer program to implement the method of any one of claims 1-7.

Images