Multi-spectral target recognition method based on double-branch multi-view spectral feature interaction

Abstract

The application discloses a multi-spectral target recognition method based on double-branch multi-view space spectrum feature interaction, comprising the following steps: performing principal component analysis operation on multi-spectral images to perform spectral band dimension reduction processing, and extracting three-dimensional image blocks as input data by using a pixel-level segmentation operation; capturing shallow spectral-spatial joint features of the input data; extracting local and global spectral-spatial joint features of the input data by using a CNN; extracting local and global correlation characteristics between spectral bands and spatial pixels by using a Transformer; early fusion of adaptive interaction of the local and global spectral-spatial joint features captured by the CNN and the Transformer to generate interdependent attention with more discriminative characteristics; and obtaining a multi-spectral target recognition result by using an output module through the interdependent attention.

Description

Technical Field

[0001] This invention belongs to the field of image processing and multispectral target recognition, specifically involving a multispectral target recognition method based on dual-branch multi-view spatial spectral feature interaction. Background Technology

[0002] Multispectral images organically combine high spatial and spectral resolution, macroscopic and microscopic remote sensing information, and spatial and spectral information. They are no longer traditional two-dimensional images, but rather three-dimensional data cubes. The number of spectral bands in a multispectral image can range from several to hundreds, covering the ultraviolet, visible, near-infrared, and mid-infrared electromagnetic spectrum regions, characterized by wide spectral coverage and a large number of bands. The fine spectral information of multispectral images can reflect the intrinsic physical and chemical properties of targets, as well as subtle differences between different targets, that are undetectable by the naked eye and ordinary optical sensors. The fine spatial information can reflect the target's physical location, structural relationships, and external quality characteristics such as shape proportions, exhibiting a "spatial-spectral integration" characteristic. This provides rich data support for target identification, making target information acquisition more comprehensive and beneficial to improving the accuracy of target identification.

[0003] Currently, there are two main types of multispectral target recognition methods: 1) Traditional multispectral target recognition methods utilize feature engineering to reduce the dimensionality of multispectral images and obtain features or spectral bands with discriminative properties. These features or spectral bands are then fed into a classifier for optimization training. However, traditional multispectral target recognition methods largely rely on handcrafted features or shallow descriptions. Most handcrafted features or shallow descriptions are typically designed for specific tasks, and the parameter setting stage requires experienced experts to design feature extraction methods, resulting in these methods being applicable only to certain specific scenarios and limiting their applicability to different scenarios; 2) Deep learning-based multispectral target recognition methods construct multiple hierarchical network structures to simulate the information processing structure of the human brain. Through training methods, they automatically learn from the original data to obtain the information features of this data. These methods utilize hierarchical structures to automatically learn important features from the original data and gradually abstract them into higher-level feature representations. They do not require manual feature engineering, possess strong generalization ability and universality, and are strong candidates for developing multispectral target recognition methods.

[0004] However, due to the fundamental differences in structural paradigms, there are insurmountable perceptual gaps in the spectral-spatial joint features captured by different networks. Furthermore, deep learning-based multispectral target recognition methods focus on improving recognition accuracy by increasing the depth or width of the network model, causing the discriminative spectral-spatial joint features to disappear during training as the model deepens or widens. Therefore, a specialized multispectral target recognition network model is needed to address these issues. Summary of the Invention

[0005] This invention addresses the shortcomings of existing deep learning-based multispectral target recognition methods by providing a method based on dual-branch, multi-view spatial-spectral feature interaction. This method fully leverages the complementary advantages of CNN and Transformer networks in their structural paradigms, employing an early fusion strategy to achieve adaptive interaction of local and global spectral-spatial joint features captured by both CNN and Transformer networks, thereby obtaining multispectral target recognition results.

[0006] The technical solution adopted in this invention is as follows: a multi-spectral target recognition method based on dual-branch multi-view spatial-spectral feature interaction, comprising the following six steps:

[0007] Step 1: Perform principal component analysis on the multi-spectral image to reduce the spectral dimension, and use pixel-level segmentation to extract three-dimensional image patches as input data;

[0008] Step 2: Capture shallow spectral-spatial joint features of the input data;

[0009] Step 3: Use CNN to extract local and global spectral-spatial joint features from the input data;

[0010] Step 4: Use Transformer to extract the local and global correlation characteristics between spectral bands and spatial pixels;

[0011] Step 5: Symmetric Multidimensional Consciousness Fusion Mechanism utilizes an early fusion strategy to adaptively fuse local and global spectral-spatial joint features captured by CNN and Transformer, generating more discriminative interdependent attention.

[0012] Step 6: Using the output module, obtain multi-spectral target recognition results through interdependent attention.

[0013] The advantages and beneficial effects of this invention compared with the prior art are as follows:

[0014] The CNN designed in this invention utilizes multiple spectral-spatial feature extraction modules and cross layers to interpret multi-spectral images locally and globally, effectively reducing the perceptual differences between CNN and Transformer.

[0015] The Transformer designed in this invention integrates the distortion invariance of CNN and utilizes multiple composite encoder modules and cross layers to transmit the correlation characteristics between spectral bands and spatial pixels locally and globally.

[0016] The symmetric multidimensional consciousness fusion mechanism designed in this invention obtains the interdependent attention of CNN and Transformer, effectively deepening the cross-dimensional interaction between the two networks, thereby avoiding the loss of important information. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the multi-spectral target recognition method based on dual-branch multi-view spatial spectral feature interaction of the present invention.

[0018] Figure 2 This is a structural diagram of the multi-spectral target recognition method based on dual-branch multi-view spatial spectral feature interaction of the present invention;

[0019] Figure 3 This is a structural diagram of the symmetrical multidimensional consciousness fusion mechanism of the present invention;

[0020] Figure 4 This is the experimental dataset used in this invention;

[0021] Figure 5 The figures show the experimental results of this invention, where (a) shows the ground truth diagram, (b) shows the experimental results of 2DCNN (Two-Dimensional Convolutional Neural Network), (c) shows the experimental results of 3DCNN (Three-Dimensional Convolutional Neural Network), (d) shows the experimental results of HybridSN (Hybrid Three-Dimensional-Two-Dimensional Convolutional Neural Network), (e) shows the experimental results of RSSAN (Residual Spectral-Spatial Attention Network), (f) shows the experimental results of DMCN (End-to-End Multilayer Hybrid Attention Network), (g) shows the experimental results of MSDAN (Multi-Scale Dense Connection Attention Network), (h) shows the experimental results of ViT (Visual Transformer), (i) shows the experimental results of Spectralformer, (j) shows the experimental results of SSFTT (Spectral-Spatial Feature Label Transformer), (k) shows the experimental results of GAHT (Group Consciousness Hierarchical Transformer), (l) shows the experimental results of CDCformer (Cascaded Convolution-Based Dense Connection Transformer), and (m) shows the experimental results of CDCformer (Cascaded Convolution-Based Dense Connection Transformer). The experimental results of the DLMFI of the present invention are shown. Detailed Implementation

[0022] The specific flowchart and structure diagram of the multi-spectral target recognition method based on dual-branch multi-view spatial-spectral feature interaction of this invention are as follows: Figure 1 and Figure 2 As shown in the accompanying drawings and specific embodiments, the present invention will now be described in further detail.

[0023] like Figure 1As shown, the multi-spectral target recognition method based on dual-branch multi-view spatial-spectral feature interaction achieves multi-spectral target recognition through the following six steps:

[0024] Step 1: Perform principal component analysis on the multi-spectral image to reduce the spectral dimension, and use pixel-level segmentation to extract three-dimensional image blocks (composed of target pixels and their surrounding neighboring pixels) as input data;

[0025] Step 2: Capture shallow spectral-spatial joint features of the input data;

[0026] Step 3: Use CNN to extract local and global spectral-spatial joint features from the input data;

[0027] Step 4: Use Transformer to extract the local and global correlation characteristics between the spectral bands and the spatial pixels of the 3D image patch;

[0028] Step 5: Utilize the symmetric multidimensional consciousness fusion mechanism to fuse the local and global correlation features captured by CNN and Transformer in the early stage to generate more discriminative interdependent attention;

[0029] Step 6: Using the output module, obtain multi-spectral target recognition results through interdependent attention.

[0030] In one embodiment, step one involves using principal component analysis to perform spectral band dimensionality reduction on the multi-spectral image, including: removing highly correlated spectral bands; and using pixel-level segmentation to extract target pixels and their neighboring pixels to form a three-dimensional image patch as the input data of this invention. .

[0031] In one embodiment, step two involves capturing shallow spectral-spatial joint features of the input data via an initialization module. This initialization module consists of two kernels with a size of [missing information]. and It consists of a two-dimensional convolutional layer, two batch normalization layers, and two ReLU activation functions, used to acquire input data. The shallow spectral-spatial joint characteristics are calculated as follows:

[0032] (1)

[0033] The symbols in formula (1) have the following meanings:

[0034] Input data;

[0035] : Intermediate layer output results;

[0036] : Initialize the module output results;

[0037] ReLU activation function;

[0038] : Batch normalization layer;

[0039] Kernel size is Two-dimensional convolutional layers;

[0040] Kernel size is Two-dimensional convolutional layers.

[0041] In one embodiment, the CNN described in step three consists of three Spectral-Spatial Feature Extraction Modules (SSFEMs) and cross-layer pass operations, used to mine local and global joint spectral-spatial features. The calculation process is as follows:

[0042] (2)

[0043] The symbols in formula (2) have the following meanings:

[0044] : Initialize the module output results;

[0045] : Dimensional reshaping operation;

[0046] : Spectral-spatial feature extraction module;

[0047] Output of the first spectral-spatial feature extraction module;

[0048] Cross-layer pass operation;

[0049] The first cross-layer passes the output of the operation;

[0050] Output of the second spectral-spatial feature extraction module;

[0051] The output result of the second cross layer is passed through.

[0052] Output of the third spectral-spatial feature extraction module;

[0053] The third cross layer passes the output of the operation;

[0054] : Output results of CNN branches;

[0055] ReLU activation function;

[0056] : Batch normalization layer;

[0057] The kernel is A two-dimensional convolutional layer of size.

[0058] In one implementation, the Transformer described in step four consists of four combined encoders (CEncoders) and cross-layer transfer operations, such as... Figure 3 As shown, the calculation process for capturing the local and global correlation characteristics between spectral bands and spatial pixels is as follows:

[0059] (3)

[0060] The symbols in formula (3) have the following meanings:

[0061] : Initialize the module output results;

[0062] : Dimensional reshaping operation;

[0063] Cross-layer pass operation;

[0064] Image segmentation operation;

[0065] Composite encoder;

[0066] : Output result of the first composite encoder;

[0067] The first cross layer passes the output result;

[0068] : Output result of the second composite encoder;

[0069] The second cross layer passes the output result;

[0070] Output of the third composite encoding module;

[0071] The third cross layer passes the output result;

[0072] Output of the fourth composite encoding module;

[0073] The fourth cross layer passes the output result;

[0074] ReLU activation function;

[0075] : Batch normalization layer;

[0076] : Output results of the Transformer branch.

[0077] In one embodiment, the symmetric multidimensional consciousness fusion mechanism described in step five utilizes an early fusion strategy to adaptively fuse the local and global spectral-spatial joint features captured by the CNN and Transformer dual networks, generating more discriminative interdependent attention. The calculation process is as follows:

[0078] Step a: Process the features output by the CNN branches Features of Transformer network branch outputs Global max pooling and global average pooling are performed along the row and column directions respectively to obtain the correlation in the height, width, and spectrum dimensions. The calculation formula is as follows:

[0079] (4)

[0080] The symbols in formula (4) have the following meanings:

[0081] : and of high;

[0082] : and The width;

[0083] : The output of performing global max pooling along the row direction;

[0084] : The output of performing global max pooling along the row direction;

[0085] : The output of performing global max pooling along the column direction;

[0086] : Output of global max pooling operation along the column direction

[0087] : The output of performing global average pooling along the row direction;

[0088] : The output of performing global average pooling along the row direction;

[0089] : The output of performing a global average pooling operation along the column direction;

[0090] : The output of the global average pooling operation performed along the column direction.

[0091] Step b: Aggregate the eight feature maps obtained in step a through a concatenation operation to generate global cross-dimensional awareness information. The calculation formula is as follows:

[0092] (5)

[0093] The symbols in formula (5) have the following meanings:

[0094] Cascading operation;

[0095] : The concatenated result of global average pooling and global max pooling output features along the row direction;

[0096] : The concatenated result of global average pooling and global max pooling output features along the row direction;

[0097] : and Cascaded results;

[0098] : The concatenated result of global average pooling and global max pooling output features along the column direction;

[0099] : The concatenated result of global average pooling and global max pooling output features along the column direction;

[0100] : and Cascaded results;

[0101] : and Cascaded results.

[0102] Step c: Consciousness information across global cross dimensions Perform customized conversion operations for further refinement. The calculation formula is as follows:

[0103] (6)

[0104] The symbols in formula (6) have the following meanings:

[0105] : Batch normalization layer;

[0106] Kernel size is Two-dimensional convolutional layers;

[0107] Customized conversion operations;

[0108] Element-wise addition;

[0109] : Refined global cross-dimensional awareness information.

[0110] Step d: Along the high and wide dimensions It is divided into two different feature maps, namely and ,right and Perform additional transformations and element-wise multiplication to obtain the correlation between high-dimensional and wide-dimensional features, calculated as follows:

[0111] (7)

[0112] The symbols in formula (7) have the following meanings:

[0113] Feature segmentation operation;

[0114] Additional conversion operations;

[0115] Element-wise multiplication;

[0116] The correlation between high-dimensional and wide-dimensional dimensions.

[0117] Step e: Utilize and Reweighting along high-dimensional, wide-dimensional, and spectral dimensions and Acquiring multidimensional consciousness characteristics and And use element-wise addition to , , and By aggregating these features, a more comprehensive and discriminative multidimensional consciousness fusion characteristic can be obtained. The calculation formula is as follows:

[0118] (8)

[0119] The symbols in formula (8) have the following meanings:

[0120] Element-wise multiplication;

[0121] Element-level addition.

[0122] In one embodiment, the step six involves obtaining multi-spectral target recognition results using the output module. The output module consists of a two-dimensional global average pooling layer, three fully connected layers, and a softmax function. The calculation formula is as follows:

[0123] (9)

[0124] The symbols in formula (9) have the following meanings:

[0125] Two-dimensional global average pooling;

[0126] Fully connected layer;

[0127] Multispectral target recognition results.

[0128] The aforementioned method comprises a dual-branch network of Convolutional Neural Network (CNN) and Transformer, along with an early fusion strategy (symmetric multidimensional awareness fusion mechanism). It fully leverages the complementary advantages of CNN and Transformer networks in their structural paradigms. The early fusion strategy enables adaptive interaction of local and global spectral-spatial joint features captured by both networks, thereby obtaining multispectral target recognition results. Specifically, the CNN utilizes multiple spectral-spatial feature extraction modules and cross-layers to locally and globally interpret multispectral images, effectively reducing the perceptual differences between CNN and Transformer. The Transformer integrates the distortion invariance of CNN, utilizing multiple composite encoder modules and cross-layers to locally and globally encode the correlation characteristics between spectral bands and spatial pixels. Furthermore, this invention employs a symmetric multidimensional awareness fusion mechanism to obtain interdependent attention between CNN and Transformer, effectively deepening the cross-dimensional interaction between the two networks and thus avoiding the loss of important information.

[0129] Example:

[0130] 1) Experimental conditions: The computer used in the experiment had an NVIDIA RTX 2080Ti GPU, 16.0 GB of memory, and Tensorflow 2.1.0 as the programming platform. The dataset size used in the experiment was... It has 200 spectral bands and 16 target categories.

[0131] 2) Experiment content:

[0132] The method proposed in this invention is compared with existing multi-spectral target recognition algorithms, among which commonly used algorithms mainly include 2DCNN (two-dimensional convolutional neural network), 3DCNN (three-dimensional convolutional neural network), HybridSN (hybrid three-dimensional-two-dimensional convolutional neural network), RSSAN (residual spectral-spatial attention network), DMCN (end-to-end multilayer hybrid attention network), MSDAN (multi-scale dense connection attention network), ViT (visual Transformer), Spectralformer (spectral Transformer), SSFTT (spectral-spatial feature labeling Transformer), GAHT (group consciousness hierarchical Transformer), and CDCformer (dense connection Transformer based on cascaded convolution).

[0133] Figure 4 The experimental dataset used in this invention is provided. Figure 5Visual results for 12 methods on experimental data are presented, including (a) the ground truth graph, (b) the experimental results for 2DCNN (Two-Dimensional Convolutional Neural Network), (c) the experimental results for 3DCNN (Three-Dimensional Convolutional Neural Network), (d) the experimental results for HybridSN (Hybrid Three-Dimensional-Two-Dimensional Convolutional Neural Network), (e) the experimental results for RSSAN (Residual Spectral-Spatial Attention Network), (f) the experimental results for DMCN (End-to-End Multilayer Hybrid Attention Network), (g) the experimental results for MSDAN (Multi-Scale Dense Connection Attention Network), (h) the experimental results for ViT (Visual Transformer), (i) the experimental results for Spectralformer, (j) the experimental results for SSFTT (Spectral-Spatial Feature Label Transformer), and (k) the experimental results for... Experimental results for GAHT (Group Consciousness Hierarchical Transformer) are shown in (l), experimental results for CDCformer (Densely Connected Transformer Based on Concatenated Convolutions) are shown in (m), and experimental results for the DLMFI method of this invention are shown in (m). The experimental results show that the visual output image of this invention is the clearest and cleanest, containing more detail and less noise. Visual output images from other methods contain salt-and-pepper noise (e.g., 2DCNN) or have smoother edges (e.g., HybridSN, MSDAN), demonstrating the superiority and robustness of this method.

[0134] This experiment uses overall accuracy (OA), average accuracy (AA), and Kappa coefficient as evaluation metrics to reflect the target recognition capability and accuracy of the algorithm. Table 1 shows the OA, AA, and Kappa of existing multi-spectral target recognition methods and the present invention (DLMFI shown in Table 1). Experimental results show that the present invention achieves the best OA, AA, and Kappa on the experimental dataset, indicating that the target recognition capability of the present invention is superior to existing multi-spectral target recognition methods.

[0135] Table 1

[0136]

[0137] The parts of this invention not described in detail belong to the well-known technology in this field. The above description is only a specific example of this invention and is not intended to limit this invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the protection scope of this invention.

Claims

1. A multi-spectral target recognition method based on dual-branch multi-view spatial-spectral feature interaction, characterized in that, include: Step 1: Perform principal component analysis on the multi-spectral image to reduce the spectral dimension, and use pixel-level segmentation to extract three-dimensional image patches as input data; Step 2: Capture shallow spectral-spatial joint features of the input data; Step 3: Use CNN to extract local and global spectral-spatial joint features from the input data; Step 4: Use Transformer to extract the local and global correlation characteristics between spectral bands and spatial pixels; Step 5: The symmetric multidimensional consciousness fusion mechanism utilizes an early fusion strategy to adaptively interact the local and global spectral-spatial joint features extracted by CNN and the local and global correlation features extracted by Transformer, generating more discriminative interdependent attention. The symmetric multidimensional consciousness fusion mechanism consists of a dual-branch network composed of CNN and Transformer, and an early fusion strategy. Step 6: Using the output module, obtain multi-spectral target recognition results through interdependent attention; The calculation process for step five is as follows: Step a: Process the features output by the CNN Features of Transformer output Global max pooling and global average pooling operations are performed along the row and column directions, respectively, to obtain the correlation in the height, width and spectral dimensions. Step b: Aggregate the eight feature maps obtained in step a through a cascade operation to generate global cross-dimensional awareness information. ; Step c: Consciousness information across global cross dimensions Perform customized conversion operations for further refinement. To obtain detailed global cross-dimensional consciousness information ; Step d: Integrate global cross-dimensional awareness information along the high and wide dimensions. It is divided into two different feature maps, namely and ,right and Perform additional transformation operations and element-wise multiplication operations to obtain correlations in high and wide dimensions; Step e: Utilize and Along the high dimension, wide dimension and spectral dimension and Reweighting yields multidimensional consciousness characteristics. and And use element-wise addition to , , and By aggregating these features, a more comprehensive and discriminative multidimensional consciousness fusion characteristic can be obtained. .

2. The multi-spectral target recognition method based on dual-branch multi-view spatial-spectral feature interaction according to claim 1, characterized in that: Step one involves performing principal component analysis on the multi-spectral image to reduce the spectral dimensionality, including: removing highly correlated spectral bands; and using pixel-level segmentation to extract target pixels and their neighboring pixels to form a three-dimensional image patch as input data. .

3. The multi-spectral target recognition method based on dual-branch multi-view spatial-spectral feature interaction according to claim 2, characterized in that: In step two, shallow spectral-spatial joint features of the input data are captured through an initialization module. This initialization module consists of two kernels with a size of [missing information]. and It consists of a two-dimensional convolutional layer, two batch normalization layers, and two ReLU activation functions, used to acquire input data. The shallow spectral-spatial joint characteristics are calculated as follows: (1) The symbols in formula (1) have the following meanings: Input data; : Intermediate layer output results; : Initialize the module output results; ReLU activation function; : Batch normalization layer; Kernel size is Two-dimensional convolutional layers; Kernel size is Two-dimensional convolutional layers.

4. The multi-spectral target recognition method based on dual-branch multi-view spatial-spectral feature interaction according to claim 1, characterized in that: The CNN described in step three consists of three spectral-spatial feature extraction modules and cross-layer pass operations, used to mine local and global spectral-spatial joint features. The calculation process is as follows: (2) The symbols in formula (2) have the following meanings: : Initialize the module output results; Dimensional reshaping operation; : Spectral-spatial feature extraction module; Output of the first spectral-spatial feature extraction module; Cross-layer pass operation; The first cross-layer passes the output of the operation. Output of the second spectral-spatial feature extraction module; The output result of the second cross layer is passed through. Output results of the third spectral-spatial feature extraction module; The third cross layer passes the output of the operation; : Output results of CNN branches; ReLU activation function; : Batch normalization layer; The kernel is A two-dimensional convolutional layer of size.

5. The multi-spectral target recognition method based on dual-branch multi-view spatial-spectral feature interaction according to claim 1, characterized in that: The Transformer described in step four consists of four composite encoders and cross-layer transfer operations, used to capture the local and global correlation characteristics between spectral bands and spatial pixels. The calculation process is as follows: (3) The symbols in formula (3) have the following meanings: : Initialize the module output results; Dimensional reshaping operation; Cross-layer pass operation; Image segmentation operation; Composite encoder; : Output result of the first composite encoder; The first cross layer passes the output result; : Output result of the second composite encoder; The second cross layer passes the output result; Output of the third composite encoding module; The third cross layer passes the output result; Output of the fourth composite encoding module; The fourth cross layer passes the output result; ReLU activation function; : Batch normalization layer; : Output results of the Transformer branch.

6. The multi-spectral target recognition method based on dual-branch multi-view spatial-spectral feature interaction according to claim 1, characterized in that: In step five: The calculation formula for step a is as follows: (4) The symbols in formula (4) have the following meanings: : sum High; : and The width; : The output of performing global average pooling along the row direction; : The output of performing global average pooling along the row direction; : The output of performing a global average pooling operation along the column direction; : The output of performing a global average pooling operation along the column direction; : The output of performing global max pooling along the row direction; : The output of performing global max pooling along the row direction; : The output of performing global max pooling along the column direction; : The output of performing global max pooling along the column direction; The calculation formula for step b is as follows: (5) The symbols in formula (5) have the following meanings: Cascading operation; : The concatenated result of global average pooling and global max pooling output features along the row direction; : The concatenated result of global average pooling and global max pooling output features along the row direction; : and Cascaded results; : The concatenated result of global average pooling and global max pooling output features along the column direction; : The concatenated result of global average pooling and global max pooling output features along the column direction; : and Cascaded results; : and Cascaded results; The calculation formula for step c is as follows: (6) The symbols in formula (6) have the following meanings: : Batch normalization layer; Kernel size is Two-dimensional convolutional layers; Customized conversion operations; Element-wise addition; : Detailed global cross-dimensional awareness information; The calculation formula for step d is as follows: (7) The symbols in formula (7) have the following meanings: Feature segmentation operation; Additional conversion operations; Element-wise multiplication; The correlation between high-dimensional and wide-dimensional dimensions; The calculation formula for step e is as follows: (8) The symbols in formula (8) have the following meanings: Element-wise multiplication; Element-level addition.

7. The multi-spectral target recognition method based on dual-branch multi-view spatial-spectral feature interaction according to claim 1, characterized in that: Step six describes obtaining multi-spectral target recognition results using the output module. The output module consists of a two-dimensional global average pooling layer, three fully connected layers, and a softmax function. The calculation formula is as follows: (9) The symbols in formula (9) have the following meanings: Two-dimensional global average pooling; Fully connected layer; Multispectral target recognition results.

Images