A hyperspectral image information classification method, device, equipment and medium

Abstract

The application discloses a hyperspectral image information classification method, device, equipment and medium, the hyperspectral image information classification method comprises the following steps: acquiring a hyperspectral image, and the hyperspectral image comprises a training image and a test image; acquiring an initial hyperspectral image information classification model through multiple analysis modules; taking the training image as an input variable, training and optimizing the initial hyperspectral image information classification model, and acquiring a target hyperspectral image information classification model; and processing the test image through the target hyperspectral image information classification model, and acquiring a hyperspectral image information classification result. Through the hyperspectral image information classification method disclosed by the application, the classification efficiency and accuracy of the hyperspectral image information are improved.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of image processing, and particularly relates to a hyperspectral image information classification method, device, equipment and medium. BACKGROUND

[0002] A hyperspectral image is a comprehensive carrier of multiple information composed of hundreds of spectral bands, and is an important research field and a hot research direction of remote sensing science. At present, hyperspectral remote sensing technology has been successfully applied to agricultural monitoring, environmental monitoring, marine remote sensing and military exploration, and has excellent performance in object classification and identification. When classifying hyperspectral image information, a hyperspectral image information classification model based on a convolutional neural network is prone to overfitting, large computation and insufficient feature extraction. Moreover, the classification module does not consider the interaction between the four dimensions of the channel, spectrum, horizontal direction and vertical direction of the hyperspectral image, which can easily lead to low classification efficiency and low accuracy. SUMMARY

[0003] The present application provides a hyperspectral image information classification method, device, equipment and medium, which effectively solves the problems of overfitting, large computation, insufficient feature extraction, lack of consideration of the interaction between multiple dimensions of the hyperspectral image, low classification efficiency and low accuracy of the existing hyperspectral image information classification model based on a convolutional neural network.

[0004] To solve the above technical problems, the present application is realized by the following technical scheme:

[0005] The present application provides a hyperspectral image information classification method, which comprises the following steps:

[0006] Obtain a hyperspectral image, and the hyperspectral image comprises a training image and a test image;

[0007] Obtain an initial hyperspectral image information classification model through multiple analysis modules;

[0008] Use the training image as an input variable of the initial hyperspectral image information classification model, train and optimize the initial hyperspectral image information classification model, and obtain a target hyperspectral image information classification model; and

[0009] Process the test image through the target hyperspectral image information classification model, and obtain a classification result of the hyperspectral image information.

[0010] In an embodiment of the present application, the step of obtaining the classification result of the hyperspectral image information comprises:

[0011] processing the hyperspectral image to obtain a plurality of low-dimensional hyperspectral image blocks;

[0012] performing multi-scale separable lightweight convolution processing on the plurality of low-dimensional hyperspectral image blocks to obtain a first output feature map; and

[0013] performing cross-dimensional dependency processing on the first output feature map to obtain a first attention feature map, wherein a calculation formula of the first attention feature map is:

[0014]

[0015] wherein f2 is the first attention feature map, r() represents a reshape operation, MLP() represents processing in a multi-layer perceptron, sigma() represents a Sigmoid function, represents dot product processing, and f represents the first output feature map.

[0016] In an embodiment of the present application, the step of obtaining the classification result of the hyperspectral image information further comprises:

[0017] performing spatial information fusion processing on the first attention feature map to obtain a second attention feature map, wherein a calculation formula of the second attention feature map is:

[0018]

[0019] wherein f3 is the second attention feature map, F1(), F2(), F3(), and F4() are convolution operations performed by a plurality of convolution kernels, H() is a batch normalization operation, L() is a linear correction operation, sigma() represents a Sigmoid function, and f2 is the first attention feature map;

[0020] performing dot product processing on the first output feature map and the second attention feature map to obtain a second output feature map; and

[0021] performing classification processing on feature information of the second output feature map to obtain a classification result of the hyperspectral image information.

[0022] In an embodiment of the present application, the step of obtaining a plurality of low-dimensional hyperspectral image blocks comprises:

[0023] performing dimensionality reduction processing on the hyperspectral image to obtain a low-dimensional hyperspectral image; and

[0024] performing segmentation on the low-dimensional hyperspectral image to obtain a plurality of low-dimensional hyperspectral image blocks.

[0025] In an embodiment of the present application, the step of obtaining the first output feature map comprises:

[0026] performing three-dimensional standard convolution processing on the plurality of low-dimensional hyperspectral image blocks to obtain a first feature map;

[0027] performing spectral feature and spatial feature extraction on the first feature map to obtain a second feature map;

[0028] performing connection processing on the first feature map and the second feature map to obtain a third feature map; and

[0029] performing multi-scale three-dimensional separable lightweight convolution processing on the third feature map to obtain the first output feature map.

[0030] In an embodiment of the present application, the step of obtaining the first output feature map comprises:

[0031] performing reshaping processing on the first output feature map to obtain a first reshaped feature map; and

[0032] performing processing on the first reshaped feature map on a multilayer perceptron to obtain a dimension-dependent feature map.

[0033] In an embodiment of the present application, the step of obtaining the first output feature map further comprises:

[0034] performing reshaping processing and normalization processing on the dimension-dependent feature map to obtain a first normalized feature map; and

[0035] performing dot product processing on the first output feature map and the first normalized feature map to obtain the first attention feature map.

[0036] The present application also provides a hyperspectral image information classification device, which comprises:

[0037] an image acquisition module configured to acquire a hyperspectral image, wherein the hyperspectral image comprises a training image and a test image;

[0038] an initial hyperspectral image information classification model establishment module configured to acquire an initial hyperspectral image information classification model through a plurality of analysis modules;

[0039] a training module configured to train and optimize the initial hyperspectral image information classification model by taking the training image as an input variable of the initial hyperspectral image information classification model to obtain a target hyperspectral image information classification model; and

[0040] a test module configured to process the test image through the target hyperspectral image information classification model to obtain a classification result of the hyperspectral image information.

[0041] An electronic device, comprising:

[0042] at least one processor;

[0043] a storage device for storing at least one program, which, when executed by the at least one processor, causes the electronic device to implement the hyperspectral image information classification method of any of the above.

[0044] A computer-readable storage medium having stored thereon a computer program, which, when executed by a processor of a computer, causes the computer to perform the hyperspectral image information classification method of any of the above.

[0045] The hyperspectral image information classification method, device, equipment and medium provided by the application, as described above, preprocess the hyperspectral image through the principal component analysis module, effectively suppress the influence of noise and reduce the redundancy of image data. Through the three-dimensional separable lightweight convolution module, the multiscale spatial features and spectral features of the preprocessed hyperspectral image are extracted, which reduces the parameter quantity and the amount of calculation, improves the feature expression ability and classification efficiency of the hyperspectral image, and further improves the dimension dependence and spatial information fusion ability of the spatial features and spectral features of the hyperspectral image through the global attention module. In addition, the pooling operation is removed in the global attention module, reducing information loss, and comprehensively considering the interaction between the four dimensions of channels, spectrum, spatial horizontal and vertical direction of the hyperspectral image, thereby improving the accuracy of classifying the hyperspectral image information.

[0046] Of course, implementing any product of the present application does not necessarily require all the advantages described above. BRIEF DESCRIPTION OF DRAWINGS

[0047] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the description of the embodiments. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0048] Figure 1 The flowchart of the hyperspectral image information classification method provided in the embodiments of the present application.

[0049] Figure 2 The structure diagram of the initial hyperspectral image information classification model provided in the embodiments of the present application.

[0050] Figure 3 The structure diagram of the three-dimensional separable lightweight convolution module provided in the embodiments of the present application.

[0051] Figure 4 A workflow diagram of the first separable convolution module provided in the embodiments of the present application.

[0052] Figure 5 A structural diagram of the global attention module provided in the embodiments of the present application.

[0053] Figure 6 A flowchart of the method for processing hyperspectral images by using the hyperspectral image information classification model provided in the embodiments of the present application.

[0054] Figure 7 A flowchart of the method for acquiring the first feature map provided in the embodiments of the present application.

[0055] Figure 8 A flowchart of the method for acquiring the first attention feature map provided in the embodiments of the present application.

[0056] Figure 9 A structural diagram of the hyperspectral image information classification device provided in the embodiments of the present application.

[0057] Figure 10 A structural diagram of a computer system of an electronic device suitable for implementing the embodiments of the present application is shown.

[0058] Figure 11 A comparison diagram of a first test image before and after processing provided in the embodiments of the present application.

[0059] Figure 12 A comparison diagram of a second test image before and after processing provided in the embodiments of the present application.

[0060] In the figure: 100, principal component analysis module; 200, three-dimensional separable lightweight convolution module; 210, first separable convolution module; 211, first complete three-dimensional convolution kernel; 212, first separated three-dimensional large convolution kernel; 213, second separated three-dimensional large convolution kernel; 214, first connection module; 220, second separable convolution module; 221, second complete three-dimensional convolution kernel; 222, first separated three-dimensional medium convolution kernel; 223, second separated three-dimensional medium convolution kernel; 224, second connection module; 230, third separable convolution module; 231, third complete three-dimensional convolution kernel; 232, first separated three-dimensional small convolution kernel; 233, second separated three-dimensional small convolution kernel; 234, third connection module; 300, global attention module; 310, channel attention module; 311, first reshaping module; 312, multilayer perceptron; 313, second reshaping module; 314, first normalization module; 315, first dot product module; 320, spatial attention module; 321, first convolution separation module; 322, first batch normalization module; 323, second convolution separation module; 324, second batch normalization module; 325, second normalization module; 326, second dot product module; 400, classification module; 410, fully connected layer; 420, activation function layer; 500, loss function layer; 600, low-dimensional hyperspectral image block; 610, first feature map; 620, second feature map; 630, third feature map; 710, image acquisition module; 720, initial hyperspectral image information classification model establishment module; 730, training module; 740, test module; 800, computer system; 801, central processing unit; 802, read-only memory; 803, random access memory; 804, bus; 805, I / O interface; 806, keyboard input part; 807, output part; 808, storage part; 809, communication part; 810, driver; 811, removable medium; 910, first test image; 911, first classification image; 920, second test image; 921, second classification image. DETAILED DESCRIPTION

[0061] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0062] The drawings provided in the following embodiments only schematically illustrate the basic concepts of the present application, and only show the components related to the present application in the drawings, not according to the number, shape and size of the components in actual implementation. The actual implementation of each component type, number and proportion can be a random change, and the component layout type can be more complex.

[0063] The classification method of hyperspectral image information disclosed by the present application can be applied to remote sensing identification and classification of crops. Crop classification and identification are the initial and key technologies of the crop condition detection technology system. Through classification and identification of crops, the accurate estimation of crop planting area, structure and spatial distribution can be realized, and the obtained crop planting area is the basis for studying regional food balance and predicting agricultural comprehensive productivity and population carrying capacity. However, because the distribution area of crops is wide, the planting structure is complex, and the regional difference is large, the design requirement of rapid and automatic identification and classification of crops has a great challenge. However, by applying the classification method of hyperspectral image information provided by the present application, the hyperspectral image of crops shot by remote sensing technology can be processed to obtain the accurate crop distribution feature map and classification result, so as to realize the accurate estimation of crop planting area, structure and spatial distribution.

[0064] Please refer to Figure 1 In an embodiment of the present application, the classification method of hyperspectral image information provided by the present application comprises the following steps:

[0065] In step S110, a hyperspectral image is obtained, and the hyperspectral image includes a training image and a test image.

[0066] Firstly, it is explained that in some embodiments, by setting the hyperspectral imaging device on the unmanned aerial vehicle or remote sensing satellite, the hyperspectral image of the ground object can be obtained by the unmanned aerial vehicle or remote sensing satellite, and the hyperspectral image is divided into training images and test images. In other embodiments, the obtained hyperspectral image can also use public hyperspectral image datasets, such as the Indian Pines dataset and the University of Pavia dataset. Among them, the Indian Pines dataset is a hyperspectral image with a pixel size of 120*120, with a spectral range of 0.4 microns to 2.5 microns and 224 discrete spectral bands. There are 16 different ground object categories in the Indian Pines dataset, including vegetation types such as brown land, shrubs and trees, buildings, roads, water areas and other non-natural objects, which are widely used in the fields of hyperspectral image classification, target detection and signal processing, etc. It is one of the standard datasets for performance evaluation of hyperspectral image processing models. The University of Pavia dataset is an image with a spectral range of 0.43 microns to 0.86 microns and 103 spectral bands. The University of Pavia dataset includes 9 different ground object categories, including lemon, shrub, wheat field, multi-color roof, desert ground, lawn, etc., which can also be used for performance evaluation of models for image processing such as target recognition and segmentation.

[0067] Step S120, obtaining a primary hyperspectral image information classification model through a plurality of analysis modules.

[0068] Referring to Figure 2 In an embodiment of the present application, a primary hyperspectral image information classification model is established through a plurality of analysis modules, for example, by connecting a principal component analysis module, a three-dimensional separable lightweight convolution module, a global attention module, a fully connected layer, a classification module and a loss function layer in sequence to obtain an initial hyperspectral image information classification model. The initial hyperspectral image information classification model includes a principal component analysis module 100, a three-dimensional separable lightweight convolution module 200, a global attention module 300, a classification module 400 and a loss function layer 500. The input end of the three-dimensional separable lightweight convolution module 200 is connected with the output end of the principal component analysis module 100, and the input end of the global attention module 300 is connected with the output end of the three-dimensional separable lightweight convolution module 200. The input end of the classification module 400 is connected with the output end of the global attention module 300, and the loss function 500 is connected with the output end of the classification module 400.

[0069] Referring to Figure 2As shown, in an embodiment of the present application, because the hyperspectral image has hundreds of spectral bands, in order to suppress the influence of noise and reduce the redundancy of data when processing the hyperspectral image, the method of principal component analysis (PCA) is used in the principal component analysis module 100 of the hyperspectral image information classification model to preprocess the hyperspectral image. Preprocessing the hyperspectral image using PCA is to project the high-dimensional spectral bands into a low-dimensional space containing a few principal components, reducing the spectral bands of the original hyperspectral image from A dimensions to b dimensions, A is greater than b. The A-dimensional hyperspectral image is represented as R W×H×A , and the b-dimensional hyperspectral image is represented as R W×H×b , where WxH represents the spatial dimension size of the hyperspectral image, and A and b represent the spectral dimension of the hyperspectral image. In the principal component analysis module 100, the reduced hyperspectral image also needs to be cropped to crop multiple hyperspectral image blocks, where the hyperspectral image block is represented as R w×h×b , and w x h represents the spatial dimension size of the hyperspectral image block.

[0070] Please refer to Figure 2 As shown, in an embodiment of the present application, in order to efficiently extract the features of the hyperspectral image, a three-dimensional separable lightweight convolution module 200 is proposed by combining a separable three-dimensional convolution with a Ghost module. The Ghost module is an enhancement module for convolutional neural networks, which aims to improve the ability of convolutional neural networks in feature expression and reduce the computational complexity and memory requirements of convolutional neural networks. The three-dimensional separable lightweight convolution module 200 includes a first separable convolution module 210, a second separable convolution module 220, and a third separable convolution module 230. The input end of the first separable convolution module 210 is connected to the output end of the principal component analysis layer 100. The input end of the second separable convolution module 220 is connected to the output end of the first separable convolution module 210. The input end of the third separable convolution module 230 is connected to the output end of the second separable convolution module 220. The output end of the third separable convolution module 230 is connected to the input end of the global attention module 300. Through multiple separable convolution modules, the input hyperspectral image block is sequentially extracted features, improving the feature extraction capability. Moreover, the three-dimensional separable lightweight convolution module 200 can also include a fourth separable convolution module and more separable convolution modules to achieve more detailed feature extraction of the input hyperspectral image block. The present application does not limit the specific number of separable convolution modules.

[0071] Please refer to Figure 2 and Figure 3As shown, in an embodiment of the present application, the first separable convolution module 210 includes a first full three-dimensional convolution kernel 211, a first separated three-dimensional large convolution kernel 212, a second separated three-dimensional large convolution kernel 213, and a first connection module 214. The second separable convolution module 220 includes a second full three-dimensional convolution kernel 221, a first separated three-dimensional medium convolution kernel 222, a second separated three-dimensional medium convolution kernel 223, and a second connection module 224. The third separable convolution module 230 includes a third full three-dimensional convolution kernel 231, a first separated three-dimensional small convolution kernel 232, a second separated three-dimensional small convolution kernel 233, and a third connection module 234. The input end of the first full three-dimensional convolution kernel 211 is connected with the output end of the principal component analysis module 100, the first separated three-dimensional large convolution kernel 212, the first full three-dimensional convolution kernel 211, and the second separated three-dimensional large convolution kernel 213 are connected in sequence, and the input end of the first connection module 214 is connected with the output end of the first full three-dimensional convolution kernel 211 and the output end of the second separated three-dimensional large convolution kernel 213. The input end of the second full three-dimensional convolution kernel 221 is connected with the output end of the first connection module 214, the second full three-dimensional convolution kernel 221, the first separated three-dimensional medium convolution kernel 222, and the second separated three-dimensional medium convolution kernel 223 are connected in sequence, and the input end of the second connection module 224 is connected with the output end of the second full three-dimensional convolution kernel 221 and the output end of the second separated three-dimensional medium convolution kernel 223. The input end of the third full three-dimensional convolution kernel 231 is connected with the output end of the second connection module 224, the third full three-dimensional convolution kernel 231, the first separated three-dimensional small convolution kernel 232, and the second separated three-dimensional small convolution kernel 233 are connected in sequence, and the input end of the third connection module 234 is connected with the output end of the third full three-dimensional convolution kernel 231 and the output end of the third separated three-dimensional small convolution kernel 233. Wherein, the number and size of the convolution kernel in each separable convolution module can be set according to specific classification requirements, and the present application does not limit the number and size of the convolution kernel in each separable convolution module.

[0072] In an embodiment of the present application, if a standard three-dimensional convolution is used to process a hyperspectral image block, wherein the hyperspectral image block input by any convolution layer is represented as Wherein, w x h represents the spatial dimension size of the hyperspectral image block, b is the spectral dimension number of the hyperspectral image block, C i is the input channel number of the hyperspectral image block. Wherein, the convolution kernel of the standard three-dimensional convolution layer is, for example, Wherein, k x k x k is the size of the convolution kernel, respectively the depth, height and width of the convolution kernel, C o is the output channel number of the hyperspectral image block, and C o =C i The feature map output after the input hyperspectral image block is filtered by the convolution kernel w′×h′ is the spatial dimension of the output feature map, and b′ is the spectral dimension of the output feature map. Then, the formula for calculating the number of parameters P in a standard 3D convolution is:

[0073] P=k×k×k×C i ×C o And the formula for calculating the floating-point operation quantity F is:

[0074] F=k×k×k×w′×h′×b′×C i ×C o .

[0075] Please see Figure 2 , Figure 3 and Figure 4 As shown, in one embodiment of the present invention, in the three-dimensional separable lightweight convolution module 200, the spectral and spatial features of the hyperspectral image patch are extracted by combining separable three-dimensional convolution and the Ghost module. First, the low-dimensional hyperspectral image patch x600 is processed by the first separable convolution module 210 through a three-dimensional convolutional layer with a first complete three-dimensional convolutional kernel 211 to obtain a first feature map 610. The first complete three-dimensional convolutional kernel 211 is, for example, represented as... The first feature map l610 is represented, for example, as follows: in, This represents the number of output channels for the hyperspectral image patch. In this case, the parameter P1 and the floating-point operation quantity F1 are calculated as follows:

[0076]

[0077]

[0078] Then, a separable 3D convolution is performed on the first feature map 610. The first feature map 610 is convolved by a first separable 3D large convolution kernel 212 and a second separable 3D large convolution kernel 213. The first separable 3D large convolution kernel 212 extracts the spatial features of each band on the first feature map 610, and the second separable 3D large convolution kernel 213 extracts the spectral features of the first feature map 610, thus obtaining the second feature map l1620. The first separable 3D large convolution kernel 212 is represented, for example, as: The second separation 3D large convolution kernel 213 is represented as follows: C o1 C represents the number of output channels of the first feature map 610. o The number of output channels for the second feature map 620, where the second feature map l1 is represented for example as... At this point, the parameters P2 and P3 are expressed as:

[0079] P2=k×k×1×Ci ×C o1

[0080] P3 = 1 x 1 x k x C i ×C o ,

[0081] The calculation formula of the floating-point operation F2 and F3 is:

[0082] F2 = k x k x 1 x w' x h' x b' x C i ×C o1

[0083] F3 = 1 x 1 x k x w' x h' x b' x C i ×C o ,

[0084] Therefore, the calculation formula of the total parameter quantity P and the floating-point operation F in the first separable convolution module 210 is:

[0085] P = P1 + P2 + P3

[0086] F = F1 + F2 + F3.

[0087] Compared with the parameter quantity P and the floating-point operation F of the standard three-dimensional convolution, the first separable convolution module 210 greatly reduces the parameter quantity and the calculation quantity. Moreover, in order to enhance the representation ability of the feature, the first connection module 214 is used to connect the obtained first feature map 610 and the second feature map 620 to obtain a third feature map 630. The third feature map is expressed as, for example: y1 ∈ R w′×h′×b′×C . The calculation formula of the third feature map 630 is:

[0088] y1 = Concat(G(x), H(G(x)))

[0089] wherein Concat() represents a connection function, G() represents processing by the first complete three-dimensional convolution kernel 211, and H() represents processing by the first separated three-dimensional large convolution kernel 212 and the second separated three-dimensional large convolution kernel 213. Furthermore, in order to further enhance the representation ability of the features, in some embodiments, the third feature map 630 can also be processed by the second separable convolution module 220 and the third separable convolution module 230 to obtain the output feature map of the three-dimensional separable lightweight convolution module 200, i.e., the first output feature map. The size of the convolution kernel in the second separable convolution module 220 and the third separable convolution module 230 is smaller than the size of the convolution kernel in the first separable convolution module 210, so as to extract more detailed features from the input feature map. Through the three-dimensional separable lightweight convolution module 200, the spatial characteristics and spectral characteristics of the hyperspectral image are effectively improved, the redundant features are effectively optimized, and the learning of the three-dimensional convolution features is promoted, and the parameter quantity and the calculation quantity of the three-dimensional convolution are reduced without changing the mapping size of the output feature map.

[0090] Referring to Figure 2 and Figure 5As shown, in an embodiment of the present application, considering that the hyperspectral image has the problems of same-spectrum different objects and different-spectrum same objects, only enhancing the features from a single dimension of space or spectrum will lead to the incompleteness of the significance of the features. Therefore, in order to extract more useful information of the hyperspectral image data, the cross-dimension interaction of the feature information can be enhanced from the two dimensions of spectrum and space, the output tensor contains the interaction of four dimensions, and the distinguishability of the features is enhanced to improve the classification performance of the hyperspectral image information classification model, and the present application proposes a global attention module 300. The global attention module 300 includes a channel attention module 310 and a spatial attention module 320, the input end of the channel attention module 310 is connected with the output end of the three-dimensional separable lightweight convolution module 200, and the input end of the spatial attention module 320 is connected with the output end of the three-dimensional separable lightweight convolution module 200 and the output end of the channel attention module 310. The channel attention module 310 includes a first reshaping module 311, a multilayer perceptron 312, a second reshaping module 313, a first normalization module 314 and a first dot product module 315, which are connected in sequence, and the input end of the first reshaping module 311 and the input end of the first dot product module 315 are connected with the output end of the three-dimensional separable lightweight convolution module 200, and the output end of the first dot product module 315 is connected with the input end of the spatial attention module 320. The spatial attention module 320 includes a first convolution separation module 321, a first batch normalization module 322, a second convolution separation module 323, a second batch normalization module 324, a second normalization module 325 and a second dot product module 326. The first convolution separation module 321, the first batch normalization module 322, the second convolution separation module 323, the second batch normalization module 324, the second normalization module 325 and the second dot product module 326 are connected in sequence, and the input end of the first convolution separation module 321 is connected with the output end of the first dot product module 330.

[0091] Referring to Figure 2 and Figure 5As shown, in an embodiment of the present application, the first output feature map first enters the channel attention module 310 for processing. The processing method of the channel attention module 310 in the present application is different from the conventional attention method, which uses the four-dimensional arrangement of the spectral dimension, the horizontal direction and the vertical direction of the channel number spatial dimension to retain the information of the four-dimensional feature map of the hyperspectral image. In order to adapt to the calculation requirement of the multilayer perceptron 312, the first reshaping module 311 first reshapes the dimension of the first output feature map to obtain a first reshaped feature map. Then the multilayer perceptron 312 is used to process the first reshaped feature map to obtain a dimension-dependent feature map. The multilayer perceptron 312 is, for example, a two-layer multilayer perceptron (MLP). Through the two-layer MLP, the size of the spatial dimension, the cross-dimension dependence of the spectral dimension and the channel number, and the nonlinear relationship between the channels can be enhanced, so as to better retain the information of the four-dimensional feature map of the hyperspectral image. Then the second reshaping module 313 reshapes the dimension-dependent feature map again to obtain a second reshaped feature map, and the first normalization module 314 is used to process the second reshaped feature map to obtain a first normalized feature map. In the first dot product module 315, the first output feature map and the first normalized feature map are dot product processed to obtain an attention feature map in the spectral direction, i.e. a first attention feature map, and the first attention feature map is output as the input of the spatial attention module 320. The calculation formula of the first attention feature map is represented as:

[0092]

[0093] wherein f2 is the first attention feature map, r() represents the reshape operation, MLP() represents the processing in the multilayer perceptron, and σ() represents the Sigmoid function, is represented as dot product processing, and f is represented as the first output feature map. The purpose of the Sigmoid function for normalizing the first feature map is to map the feature information between (0, 1), which is beneficial to improve the visualization, interpretability and anti-interference of the hyperspectral image information classification model, and increase the nonlinear transformation ability of the feature information, thereby enhancing the learning effect and generalization ability of the hyperspectral image information classification model. And through the dot product processing of the first dot product module 315, the feature correlation degree of the first output feature map and the first normalized feature map can be measured, thereby enhancing the feature extraction capability.

[0094] Please refer to Figure 2 and Figure 5As shown, in an embodiment of the present application, the first attention feature map f2 is taken as the input of the spatial attention module 320. In order to focus on the spatial information of the hyperspectral image, the nonlinear relationship of each pixel in space is established, and the spatial information fusion of the hyperspectral image is performed using the first convolution separation module 321 and the second convolution separation module 323 with different convolution kernel sizes, wherein the first convolution separation module 321 and the second convolution separation module 323 each include two three-dimensional convolution kernels. And in order to enhance the spatial features of the hyperspectral image, after the first attention feature map is processed by the first convolution separation module 321 and the second convolution separation module 323, the first batch normalization module 322 and the second batch normalization module 324 are added to perform batch normalization (Batch Normalization, BN) operation and rectified linear unit (Rectified Linear Unit, RELU) operation, and finally the attention feature map in the spatial direction, i.e. the second attention feature map, is obtained. The calculation formula of the second attention feature map can be expressed as:

[0095]

[0096] Wherein, f3 is the second attention feature map, F1() is the operation performed by the first convolution kernel in the first convolution separation module 321, F2() is the operation performed by the second convolution kernel in the first convolution separation module 321, F3() is the operation performed by the first convolution kernel in the second convolution separation module 323, F4() is the operation performed by the second convolution kernel in the second convolution separation module 323, H() is the BN operation, L() is the RELU operation, σ() represents the Sigmoid function, and f2 is the first attention feature map. And the size of the first convolution kernel and the second convolution kernel in the first convolution separation module 321 is, for example, 5x1x1 and 1x5x1, and the size of the first convolution kernel and the second convolution kernel in the second convolution separation module 323 is, for example, 7x1x1 and 1x7x1. And the result of the dot product processing of the second attention feature map and the first output feature map, i.e. the second output feature map As the input value of the classification module 400, the second output feature map The calculation formula is:

[0097]

[0098] Wherein, f represents the first output feature map, and f3 is the second attention feature map. By performing the dot product operation of the second attention feature map and the first output feature map, the feature expression of the hyperspectral image is enhanced, and the global attention module 300 removes the pooling operation, reducing the loss of characteristic information of the hyperspectral image.

[0099] Please refer to Figure 2As shown, in one embodiment of the present invention, the classification module 400 includes a fully connected layer 410 and an activation function layer 420. The input of the fully connected layer 410 is connected to the output of the global attention layer 300. The input of the activation function layer 420 is connected to the output of the fully connected layer 410, and the output of the activation function layer 420 is connected to the loss function layer 500. A second output feature map is applied to the fully connected layer 410. The data is mapped onto multiple class scores Y, and the activation function layer 420 transforms these scores into probability values, thereby enabling the mapping of the second output feature map. The classification of feature information. The formula for calculating the category score Y is:

[0100]

[0101] in, This is the second output feature map, and FC() is the fully connected layer operation.

[0102] Please see Figure 2 As shown, in one embodiment of the present invention, a loss function is calculated on the classification result output by the classification module 400 at the activation function layer 500. The loss function is, for example, the cross-entropy loss function, and the calculation formula for the loss function is, for example: Where Loss is the loss function, C is the total number of pixel categories in the hyperspectral image, T is the total number of pixels in the training image, and Y is the total number of pixels in the training image. t c It is the predicted output of the t-th pixel, that is, the true value Y indicating that the t-th pixel belongs to the c-th class. t c The classification model for hyperspectral image information can be optimized based on the loss function.

[0103] Step S130: Use the training image as the input variable of the initial hyperspectral image information classification model to train and optimize the initial hyperspectral image information classification model to obtain the target hyperspectral image information classification model.

[0104] Please see Figure 2As shown in the embodiment of the present application, the training image is taken as the input variable of the initial hyperspectral image information classification model, the initial hyperspectral image information classification model is trained and optimized, and the target hyperspectral image information classification model is obtained. First, the training image is preprocessed by the principal component analysis module 100 to obtain a preprocessed image. The feature information on the hyperspectral image is extracted and the cross-dimensional interaction of the hyperspectral image is enhanced by the three-dimensional separable lightweight convolution module 200 and the global attention module 300, and a second output feature map is generated. Then, the classification module 400 is used to classify the feature information on the second output feature map to obtain a classification result, and the loss function layer 500 is used to calculate the loss function of the classification result. The weight parameters of the initial hyperspectral image information classification model are trained through the classification result, and the weight parameters of the initial hyperspectral image information classification model are optimized according to the loss function, and the hyperspectral image information classification model with converged weights is obtained as the target hyperspectral image information classification model.

[0105] Step S140, processing the test image by the target hyperspectral image information classification model to obtain the classification result of the hyperspectral image information.

[0106] Referring to Figure 11 and Figure 12 As shown in an embodiment of the present application, the test image is taken as the input variable of the target hyperspectral image information classification model to obtain the classification result of the hyperspectral image information. The test image includes a first test image 910 and a second test image 920, wherein the first test image 910 is, for example, an Indian Pines data set 910, and the second test image is, for example, a University of Pavia data set 912. The classification result of the hyperspectral image information includes a first classification image 911 and a second classification image 912. As shown in Figure 11 and Figure 12 It can be seen that the feature information of the ground objects on the first classification image 911 and the second classification image 912 obtained after the first test image 910 and the second test image 920 are processed by the target hyperspectral image information classification model is very obvious, and the ground objects can be classified according to the spectral image, and the objects on the hyperspectral image are classified according to the hyperspectral image information.

[0107] Referring to Figure 6 As shown in an embodiment of the present application, the application of the target hyperspectral image information classification model to the method for processing the hyperspectral image includes the following steps:

[0108] Step S210, pre-processing the hyperspectral image to obtain a plurality of low-dimensional hyperspectral image blocks.

[0109] Referring to Figure 2As shown, in an embodiment of the present application, because the hyperspectral image has hundreds of spectral bands, in order to suppress the influence of noise and reduce the redundancy of data when processing the hyperspectral image, the method of principal component analysis (PCA) is used in the principal component analysis module 100 of the hyperspectral image information classification model to preprocess the hyperspectral image. The preprocessing of the hyperspectral image by PCA is to project the high-dimensional spectral band into a low-dimensional space containing a few main components, reducing the spectral band of the original hyperspectral image from A dimensions to b dimensions, A is greater than b. The A-dimensional hyperspectral image is represented as R W×H×A , and the b-dimensional hyperspectral image is represented as R W×H×b , where WxH represents the spatial dimension size of the hyperspectral image, and A and b represent the spectral dimension of the hyperspectral image. In the principal component analysis module 100, the reduced hyperspectral image also needs to be segmented and cropped to crop multiple low-dimensional hyperspectral image blocks, where the low-dimensional hyperspectral image block is represented as R w×h×b , and w x h represents the spatial dimension size of the low-dimensional hyperspectral image block.

[0110] Step S220, performing multi-scale separable lightweight convolution processing on the low-dimensional hyperspectral image block to obtain a first output feature map.

[0111] As shown in Figure 2 , in an embodiment of the present application, in order to efficiently extract the features of the hyperspectral image, a separable three-dimensional convolution combined with a Ghost module is adopted, and a three-dimensional separable lightweight convolution module 200 is proposed to perform multi-scale separable lightweight convolution processing on the low-dimensional hyperspectral image block to obtain a first output feature map. The Ghost module is an enhancement module for convolutional neural networks, which aims to improve the ability of convolutional neural networks in feature expression and reduce the computational complexity and memory requirements of convolutional neural networks.

[0112] Step S230, performing cross-dimensional dependency processing on the first output feature map to obtain a first attention feature map.

[0113] As shown in Figure 2 and Figure 5 , in an embodiment of the present application, the first output feature map will first enter the channel attention module 310 for cross-dimensional dependency processing to obtain a first attention feature map. The calculation formula of the first attention feature map is:

[0114]

[0115] Wherein, f2 is the attention feature map in the spectral direction, r() represents the reshape operation, MLP() represents the processing in the multi-layer perceptron 312, and sigma() represents the Sigmoid function. The point product processing is represented, and f represents the first output feature map. The processing method of the channel attention module 310 in the application is different from the traditional attention method, and the spectral dimension, the horizontal direction and the vertical direction of the channel number space dimension of the feature map are used to retain and enhance the information of the four-dimensional feature map of the hyperspectral image.

[0116] In step S240, the spatial information fusion processing is performed on the first attention feature map to obtain a second attention feature map, and the point product processing is performed on the first output feature map and the second attention feature map to obtain a second output feature map.

[0117] Referring to FIGS. 3 and 4, Figure 2 and Figure 5 In an embodiment of the application, the first attention feature map f2 is taken as the input of the spatial attention module 320, and a second attention feature map is output, wherein the calculation formula of the second attention feature map is represented as:

[0118]

[0119] Wherein, f3 is the second attention feature map, F1() is the operation performed by the first convolution kernel in the first convolution separation module 321, F2() is the operation performed by the second convolution kernel in the first convolution separation module 321, F3() is the operation performed by the first convolution kernel in the second convolution separation module 323, F4() is the operation performed by the second convolution kernel in the second convolution separation module 323, H() is the BN operation, L() is the RELU operation, sigma() represents the Sigmoid function, and f2 is the first attention feature map.

[0120] In step S250, the classification of the feature information of the second output feature map is performed to obtain the classification result of the hyperspectral image information.

[0121] Referring to FIGS. 3 and 4, Figure 7 In an embodiment of the application, the method for obtaining the first output feature map of the application comprises the following steps:

[0122] In step S310, the standard convolution processing is performed on the plurality of low-dimensional hyperspectral image blocks to obtain a first feature map.

[0123] Referring to FIGS. 3 and 4, Figure 2 , Figure 3 and Figure 4As shown, in one embodiment of the present invention, in the three-dimensional separable lightweight convolution module 200, the spectral and spatial features of a hyperspectral image patch can be extracted by combining separable three-dimensional convolution and a Ghost module. First, the hyperspectral image patch x is processed by the first separable convolution module 210 through a three-dimensional convolution layer with a first complete three-dimensional convolution kernel 211 using standard convolution processing to obtain a first feature map 610. The first complete three-dimensional convolution kernel 211 is, for example, represented as... The first feature map l610 is represented, for example, as follows: in, This represents the number of output channels for the hyperspectral image block.

[0124] Step S320: Extract spectral and spatial features from the first feature map to obtain the second feature map.

[0125] Please see Figure 2 , Figure 3 and Figure 4 As shown, in one embodiment of the present invention, a separable three-dimensional convolution is performed on the first feature map 610. The first feature map 610 undergoes multiple separable convolution processes through a first separable three-dimensional large convolution kernel 212 and a second separable three-dimensional large convolution kernel 213. The spatial features of each band on the first feature map 610 are extracted by the first separable three-dimensional large convolution kernel 212, and the spectral features of the first feature map 610 are extracted by the second separable three-dimensional large convolution kernel 213 to obtain the second feature map l1 620. The first separable three-dimensional large convolution kernel 212 is represented, for example, as: The second separation 3D large convolution kernel 213 is represented as follows: C o1 C represents the number of output channels of the first feature map 610. o The number of output channels for the second feature map 620, where the second feature map l1 is represented for example as...

[0126] Step S330: Connect the first feature map and the second feature map to obtain the third feature map.

[0127] Please see Figure 2 , Figure 3 and Figure 4 As shown, in one embodiment of the present invention, in order to enhance the representation capability of features, the first feature map 610 and the second feature map 620 are connected by the first connection module 214 to obtain a third feature map 630. The third feature map is represented, for example, as: y1∈R w′×h′×b′×C The formula for calculating the third feature map 630 is as follows:

[0128] y1 = Concat(G(x), H(G(x)))

[0129] wherein, Concat() represents a connection function, G() represents the first feature map 610 obtained after being processed by the first complete three-dimensional convolution kernel 211, and H() represents the second feature map 620 obtained after being processed by the first separated three-dimensional large convolution kernel 212 and the second separated three-dimensional large convolution kernel 213.

[0130] In step S340, the third feature map is sequentially subjected to multiple times of multi-scale separable lightweight convolution processing to obtain a first output feature map.

[0131] Referring to FIGS. 4 and 5, in an embodiment of the present application, the third feature map 630 can also be subjected to separable lightweight convolution processing by the second separable convolution module 220 and the third separable convolution module 230 to obtain the first output feature map. Figure 2 Figure 3 Figure 4 Referring to FIGS. 4 and 5, in an embodiment of the present application, the third feature map 630 can also be subjected to separable lightweight convolution processing by the second separable convolution module 220 and the third separable convolution module 230 to obtain the first output feature map.

[0132] Referring to FIGS. 4 and 5, in an embodiment of the present application, the method for obtaining the first attention feature map comprises the following steps: Figure 8 In step S410, a first output feature map is obtained.

[0133] In step S420, the first output feature map is subjected to reshaping processing to obtain a first reshaped feature map.

[0134] In step S430, the first reshaped feature map is processed on a multi-layer perceptron to obtain a dimension-dependent feature map.

[0135] In step S440, the dimension-dependent feature map is subjected to reshaping processing and normalization processing to obtain a first normalized feature map.

[0136] In step S450, the first output feature map and the first normalized feature map are subjected to dot product processing to obtain a first attention feature map.

[0137] Referring to FIGS. 4 and 5, in an embodiment of the present application, the calculation formula of the first attention feature map is represented as:

[0138] Figure 2 Figure 5

[0139] wherein, f2 is the first attention feature map, r() represents Reshape operation, MLP() represents processing in a multi-layer perceptron, σ() represents Sigmoid function,

[0140] wherein, f represents the first output feature map.

[0141] Referring to FIGS. 4 and 5, in an embodiment of the present application, the calculation formula of the first attention feature map is represented as:​​​​​Figure 9 As shown in the embodiment of the present application, the application provides a hyperspectral image information classification device, which comprises an image acquisition module 710, an initial hyperspectral image information classification model establishment module 720, a training module 730 and a test module 740. The image acquisition module 710 is configured to acquire a hyperspectral image, and the hyperspectral image comprises a training image and a test image. The initial hyperspectral image information classification model establishment module 720 is configured to acquire an initial hyperspectral image information classification model through a plurality of analysis modules. The training module 730 is configured to take the training image as an input variable of the initial hyperspectral image information classification model, train and optimize the initial hyperspectral image information classification model, and acquire a target hyperspectral image information classification model. The test module 740 is configured to process the test image through the target hyperspectral image information classification model, and acquire a classification result of the hyperspectral image information.

[0142] The embodiments of the application further provide an electronic device, which comprises at least one processor, and a storage device configured to store one or more programs, when the one or more programs are executed by the one or more processors, the electronic device is caused to implement the hyperspectral image information classification method provided in each of the above embodiments.

[0143] Figure 10 The structure of a computer system of the electronic device suitable for implementing the embodiments of the application is shown. It should be noted that, Figure 10 The computer system 800 of the electronic device shown is only an example, and should not limit the functions and use range of the embodiments of the application.

[0144] As Figure 10 shown, the computer system 800 comprises a central processing unit (CPU) 801, which can perform various appropriate actions and processes according to programs stored in a read-only memory (ROM) 802 or programs loaded from a storage portion 808 to a random access memory (RAM) 803, such as performing the methods described in the above embodiments. In the RAM 803, various programs and data required for system operation are also stored. The CPU 801, the ROM 802 and the RAM 803 are connected to each other through a bus 804. An input / output (I / O) interface 805 is also connected to the bus 804.

[0145] The following components are connected to the I / O interface 805: an input part 806 including a keyboard, a mouse, etc.; an output part 807 including a display such as a Cathode Ray Tube (CRT), a Liquid Crystal Display (LCD), etc., and a speaker, etc.; a storage part 808 including a hard disk, etc.; and a communication part 809 including a network interface card such as a LAN (Local Area Network) card, a modem, etc. The communication part 809 performs communication processing via a network such as the Internet. A drive 810 is also connected to the I / O interface 805 as necessary. A removable medium 811 such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, etc. is attached to the drive 810 as necessary, so that a computer program read out therefrom is installed in the storage part 808 as necessary.

[0146] In particular, the processes described above with reference to the flowcharts can be implemented as a computer software program according to embodiments of the present application. For example, embodiments of the present application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing a computer program for executing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network by the communication part 809, and / or installed from the removable medium 811. When the computer program is executed by the central processing unit (CPU) 901, various functions defined in the system of the present application are executed.

[0147] It should be noted that the computer-readable medium in the embodiments of the present application can be a computer-readable signal medium or a computer-readable storage medium or any combination of the two. The computer-readable storage medium may, for example, be an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or apparatus, or any combination of the above. More specific examples of the computer-readable storage medium can include, but are not limited to, an electrical connection having one or more wires, a portable computer diskette, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM), a flash memory, an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In this application, the computer-readable signal medium can include a data signal propagated in a baseband or as a carrier wave in a propagated data signal, in which the computer-readable computer program is carried. Such a propagated data signal can take many forms, including but not limited to an electromagnetic signal, an optical signal, or any suitable combination of the above. The computer-readable signal medium can also be any computer-readable medium other than the computer-readable storage medium, which can send, propagate, or transmit the program for use by or in connection with an instruction execution system, apparatus, or device. The computer program contained on the computer-readable medium can be transmitted in any suitable medium, including but not limited to wireless, wired, or the like, or any suitable combination of the above.

[0148] The flowcharts and block diagrams in the drawings illustrate the possible architectures, functions, and operations of systems, methods, and application structures in accordance with various embodiments of the present application. In this regard, each block in the flowcharts or block diagrams can represent a module, a segment, or a portion of code, which comprises one or more executable instructions for implementing the specified logical functions. It should also be noted that in some alternative implementations, the functions noted in the blocks can occur out of the order noted in the figures. For example, two blocks noted in succession can in fact be executed substantially concurrently or in the reverse order, depending on the functionality involved. It will also be noted that each block in the block diagrams or flowcharts, and combinations of blocks in the block diagrams or flowcharts, can be implemented by special-purpose hardware-based systems that perform the specified functions or operations, or can be implemented by a combination of special-purpose hardware and computer instructions.

[0149] The units described in the embodiments of the present application can be implemented in the form of software, or can be implemented in the form of hardware, and the units described can also be arranged in a processor. In some cases, the names of the units do not constitute a limitation on the units themselves.

[0150] Another aspect of the present application also provides a computer readable storage medium, which stores a computer program. When the computer program is executed by a processor of a computer, the computer executes the classification method of hyperspectral image information as described above. The computer readable storage medium can be included in the electronic device described in the above embodiments, or can exist separately without being assembled into the electronic device.

[0151] Another aspect of the present application also provides a computer program product or computer program, which includes computer instructions stored in a computer readable storage medium. A processor of a computer device reads the computer instructions from the computer readable storage medium, and the processor executes the computer instructions, so that the computer device executes the classification method of hyperspectral image information provided in each of the above embodiments.

[0152] To sum up, the classification method of hyperspectral image information provided by the present application effectively suppresses the influence of noise and reduces the redundancy of image data by preprocessing the hyperspectral image through the principal component analysis module. The three-dimensional separable lightweight convolution module extracts the multi-scale spatial features and spectral features of the preprocessed hyperspectral image, reduces the parameter quantity and the amount of calculation, improves the feature expression ability and the classification efficiency of the hyperspectral image, and further improves the dimension dependence and spatial information fusion ability of the spatial features and spectral features of the hyperspectral image through the global attention module. The pooling operation is removed in the global attention module, reducing information loss, comprehensively considering the interaction between the four dimensions of channels, spectrum, spatial horizontal and vertical directions of the hyperspectral image, and improving the accuracy of classifying the hyperspectral image information. And the fully connected layer and the activation function layer are used as the classification module, which effectively improves the generalization ability of the classification model of the hyperspectral image information and effectively prevents the classification model from overfitting, and improves the efficiency and accuracy of the classification of the hyperspectral image information.

[0153] In the description of the present specification, the description of the terms "the present embodiment", "example", "specific example" and the like means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present application. In the present specification, the illustrative description of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0154] The embodiments disclosed above are only used to illustrate the present application. The embodiments do not describe all of the details and do not limit the present application to the specific embodiments. Obviously, many modifications and variations can be made in light of the contents of the specification. The embodiments are chosen and described in order to best explain the principles of the present application and its practical application, so that others skilled in the art can better understand and utilize the present application. The present application is only limited to the scope of the claims and their full range of equivalents.

Claims

1. A method for classifying hyperspectral image information, characterized in that, Includes the following steps: Acquire hyperspectral images, wherein the hyperspectral images include training images and test images; An initial hyperspectral image information classification model is obtained through multiple analysis modules; The training images are used as input variables for the initial hyperspectral image information classification model to train and optimize the initial hyperspectral image information classification model, thereby obtaining the target hyperspectral image information classification model. The test image is processed using the target hyperspectral image information classification model to obtain the classification result of the hyperspectral image information; The steps for obtaining the classification results of the hyperspectral image information include: The hyperspectral image is processed to obtain multiple low-dimensional hyperspectral image patches; Multiple low-dimensional hyperspectral image patches are subjected to multi-scale separable lightweight convolution processing to obtain a first output feature map; The first output feature map is subjected to cross-dimensional dependency processing to obtain a first attention feature map, wherein the calculation formula of the first attention feature map is: Where f2 is the first attention feature map, r() represents the reshape operation, MLP() represents processing in a multilayer perceptron, and σ() represents the sigmoid function. This is represented as a dot product, and f represents the first output feature map; Spatial information fusion processing is performed on the first attention feature map to obtain a second attention feature map, wherein the calculation formula for the second attention feature map is: Where f3 is the second attention feature map, F1(), F2(), F3() and F4() are convolution operations performed by multiple convolution kernels, H() is batch normalization operation, L() is linear correction operation, σ() represents the Sigmoid function, and f2 is the first attention feature map. The first output feature map and the second attention feature map are multiplied by a dot product to obtain the second output feature map. The feature information of the second output feature map is classified to obtain the classification result of the hyperspectral image information.

2. The method for classifying hyperspectral image information according to claim 1, characterized in that, The steps for acquiring multiple low-dimensional hyperspectral image patches include: The hyperspectral image is subjected to dimensionality reduction processing to obtain a low-dimensional hyperspectral image; and The low-dimensional hyperspectral image is segmented to obtain multiple low-dimensional hyperspectral image blocks.

3. The method for classifying hyperspectral image information according to claim 1, characterized in that, The steps for obtaining the first output feature map include: Perform three-dimensional standard convolution processing on multiple low-dimensional hyperspectral image patches to obtain a first feature map; Spectral and spatial features are extracted from the first feature map to obtain the second feature map; The first feature map and the second feature map are concatenated to obtain the third feature map; and The third feature map is subjected to multiple multi-scale three-dimensional separable lightweight convolution processes to obtain the first output feature map.

4. The method for classifying hyperspectral image information according to claim 1, characterized in that, The steps for obtaining the first output feature map include: The first output feature map is reshaped to obtain a first reshaped feature map; and On the multilayer perceptron, the first reshaped feature map is processed to obtain a dimension-dependent feature map.

5. The method for classifying hyperspectral image information according to claim 4, characterized in that, The step of obtaining the first output feature map further includes: The dimension-dependent feature map is reshaped and normalized to obtain a first normalized feature map; and The first output feature map and the first normalized feature map are subjected to dot product processing to obtain the first attention feature map.

6. A training device for an occlusion image recognition model, characterized in that, The training device, which applies the classification method for hyperspectral image information as described in any one of claims 1 to 5, comprises: An image acquisition module is used to acquire hyperspectral images, wherein the hyperspectral images include training images and test images; The initial hyperspectral image information classification model establishment module is used to obtain the initial hyperspectral image information classification model through multiple analysis modules; The training module is used to train and optimize the initial hyperspectral image information classification model by using the training images as input variables to obtain the target hyperspectral image information classification model; and The testing module is used to process the test image using the target hyperspectral image information classification model to obtain the classification result of the hyperspectral image information.

7. An electronic device, characterized in that, The electronic device includes: At least one processor; A storage device for storing at least one program, which, when executed by the at least one processor, causes the electronic device to implement the classification method for hyperspectral image information according to any one of claims 1-5.

8. A computer-readable storage medium, characterized in that, It stores a computer program, which, when executed by the computer's processor, causes the computer to perform the classification method for hyperspectral image information as described in any one of claims 1-5.

Images