Hyperspectral center pixel double-branch image classification method under local security system

By adopting the central pixel dual-branch network and HiLo attention mechanism in hyperspectral image classification, the challenge of spectral and spatial feature extraction in hyperspectral image classification is solved, and the high accuracy and robust classification effect under finite sample conditions is achieved .

CN119091201APending Publication Date: 2024-12-06NORTHWESTERN POLYTECHNICAL UNIV
View PDF 0 Cites 3 Cited by

Patent Information

Application Number
CN202411129865.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-16
Publication Date
2024-12-06

AI Technical Summary

Technical Problem

Hyperspectral image classification task faces the large number of spectral bands, high redundancy, low spatial resolution, easy to affect weather and light, and high manual labeling costs. The existing methods are prone to overfitting under limited training samples. It is difficult to achieve efficient, reliable and precise classification.

Method used

A dual-branch image classification method for hyperspectral center pixels under the local security system is proposed. The global spatial feature extraction branch of ViT structure is extracted by the central spectral feature based on CNN, combined with the HiLo attention mechanism, the central pixel features and global features are separated to reduce the Independent information interferes with and enhances the extraction ability of spatial branches.

Benefits of technology

Under the condition of limited training samples, the accuracy and generalization performance of hyperspectral image classification are improved, the classification accuracy and global feature extraction ability are significantly improved, and the risk of overfitting is reduced.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119091201A_ABST
    Figure CN119091201A_ABST
Patent Text Reader

Abstract

A hyperspectral center pixel double-branch image classification method under a near-field security system comprises the following steps: data preprocessing: adopting a public hyperspectral image data set, and dividing training sets; constructing a double-branch network model based on a central pixel, wherein the double-branch network model comprises a central spectral feature extraction branch and a global spatial feature extraction branch; training the constructed network model by using the training set, and updating the network weight; and inputting the test set into the trained double-branch network model to carry out hyperspectral image classification and output a classification result graph. Center pixel region features are extracted by using a CNN structure-based spectral branch, the center pixel region features are separated from a global feature extraction process, the influence of irrelevant information in a neighborhood on a classification result is reduced, a HiLo attention mechanism is introduced into a ViT structure, the global feature extraction capability of a spatial branch is enhanced, and the classification accuracy is improved. The hyperspectral image classification model with better generalization performance is obtained under the condition of limited training samples.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of hyperspectral image classification, and in particular relates to a hyperspectral central pixel dual-branch image classification method in a local security system. Background Art

[0002] In today's increasingly complex global security environment, national and regional security are facing new challenges, and the concept of vicinagearth security (VS) has emerged. Vicinagearth security refers to a diversified, cross-domain, three-dimensional, collaborative, and intelligent technology system for defense, protection, production, safety, and rescue in vicinagearth space. Application scenarios include low-altitude security, underwater security, and cross-domain security, covering waters, ground, and airspace from 1,000 meters below sea level to 10,000 meters above sea level. In the vicinagearth security system, hyperspectral images are an important input data. Compared with natural images, they have higher spectral resolution and can provide more spectral feature information in the fine identification of various types of objects. Hyperspectral image classification technology uses the spectral feature information of objects to assign the correct material category label to each pixel in the hyperspectral image. It is not only commonly used to guide agricultural production, resource exploration, and urban planning, but also has broad application prospects in vicinityarth security tasks such as environmental monitoring, disaster warning, and reconnaissance. Since classification accuracy will affect the subsequent performance of the hyperspectral image analysis decision system, it is of great significance to study effective hyperspectral image classification methods.

[0003] Compared with image classification tasks in natural images, the classification task of hyperspectral images has the following characteristics: (1) the number of spectral bands is large, the correlation is strong, and the redundancy is high; (2) the spatial resolution of hyperspectral images is low, and there are some mixed pixels in the images. At the same time, due to the physical and technical characteristics of the sensor itself, the observation angle and the imaging mechanism, the imaging quality is often easily affected by various factors such as weather and light, and thus the phenomenon of "same object, different spectrum" and "same spectrum, different object" often occurs; (3) Since the pixel-level annotation of hyperspectral images requires certain expert knowledge in the field, manual annotation is time-consuming, labor-intensive and costly, so the number of available annotated samples in the existing hyperspectral image data is relatively small.

[0004] These complex characteristics make the hyperspectral image classification task challenging, so how to classify hyperspectral images efficiently, reliably and accurately is of great research significance. At present, the mainstream hyperspectral image classification method mainly relies on deep learning. According to the difference in feature extraction technology, the related work can be divided into the following three categories:

[0005] The first category: Deep learning methods are based on convolutional neural networks (CNNs). Hu et al. first used a one-dimensional convolutional neural network to extract spectral features in the paper "Hu W, Huang Y, Wei L, et al. Deep Convolutional Neural Networks for Hyperspectral Image Classification [J]. Journal of Sensors, 2015, 1-12.", achieving better performance than support vector machines and early deep learning-based methods. Zhang et al. divided the input spectral band into several non-overlapping continuous sub-bands in the paper "Zhang X, Shang S, Tang X, et al. Spectral partitioning residual network with spatial attention mechanism for hyperspectral image classification [J]. IEEE transactions on geoscience and remote sensing, 2021, 60: 1-14.", and used cascaded parallel improved residual blocks to extract spectral spatial features from these sub-bands respectively. Li et al. used a dual-branch CNN network in the paper "Li R, Zheng S, Duan C, et al. Classification of hyperspectral image based on double-branch dual-attention mechanism network[J]. Remote Sensing, 2020, 12(3): 582.", using spectral and spatial attention mechanisms to extract spectral and spatial features in the two branches, respectively, and then fused the features for classification. The deep learning method based on CNN has good local detail construction capabilities, but is limited by the size of the convolution window and needs to rely on stacking convolution layers to expand the receptive field.

[0006] The second category: Deep learning methods are based on visual Transformer (ViT). Hong et al. proposed a new Transformer-based backbone network SpectralFormer in the paper "Hong D, Han Z, Yao J, et al. Spectralformer: Rethinking Hyperspectral Image Classification With Transformers [J]. IEEE Transactions on Geoscience and Remote Sensing, 2021, 60: 1–15." to learn the spectral local sequence information in adjacent frequency bands, generate grouped spectral embeddings and use cross-layer jump connections to transfer the generated embeddings from shallow layers to deep layers, achieving impressive classification performance. Sun et al. proposed a spectral-spatial feature tokenization method in the paper "Sun L, Zhao G, Zheng Y, et al. Spectral–spatial feature tokenization transformer for hyperspectral image classification [J]. IEEE Transactions on Geoscience and Remote Sensing, 2022, 60: 1-14." to better adapt hyperspectral images to the ViT model and achieve better performance. The deep learning method built on ViT naturally has excellent long-distance dependency modeling capabilities and performs well in global feature extraction, but ViT has weak perception of spatial information and often requires larger-scale training data support.

[0007] The third category is the fusion of deep learning methods based on CNN and ViT structures. Xu et al. proposed a parallel structure of CNN and ViT in the paper "Xu R, Dong XM, Li W, et al. DBCTNet: Double Branch Convolution-Transformer Network for Hyperspectral Image Classification [J]. IEEE Transactions on Geoscience and Remote Sensing, 2024, 62: 3368141.", which used each other's characteristics to extract spectral features and spatial features respectively, and achieved better results than a single structure. Zhao et al. used group convolution and local window attention in the paper "Zhao Z, Xu X, LiS, et al. Hyperspectral Image Classification Using Groupwise SeparableConvolutional Vision Transformer Network [J]. IEEE Transactions on Geoscience and Remote Sensing, 2024, 62: 3377610." to reduce the number of parameters while ensuring model performance, making it more conducive to training and deployment. This type of hybrid structure of CNN and ViT can better utilize the model characteristics and achieve better results than a single structure, but this structure increases the complexity of the model and is prone to overfitting under conditions of limited training samples, making it more challenging in model design. Summary of the invention

[0008] In order to overcome the above problems, the present invention proposes a hyperspectral central pixel dual-branch image classification method under the local security system by combining the characteristics of hyperspectral image data. The central pixel area features are extracted by the spectral branch based on the CNN structure, which is separated from the global feature extraction process to reduce the influence of irrelevant information in the neighborhood on the classification results. At the same time, the HiLo attention mechanism is introduced into the ViT structure to enhance the ability of the spatial branch to extract global features, so as to achieve a hyperspectral image classification model with better generalization performance under the condition of limited training samples.

[0009] In order to achieve the above object, the technical solution adopted by the present invention is as follows:

[0010] A hyperspectral central pixel dual-branch image classification method for a local security system, comprising:

[0011] Step 1: Data preprocessing: using a public hyperspectral image dataset, the hyperspectral image is processed into several hyperspectral image patches and divided into training and test sets;

[0012] Step 2, constructing a dual-branch network model based on central pixels, wherein the dual branches are a central spectral feature extraction branch and a global spatial feature extraction branch;

[0013] Step 3, using the training set to train the dual-branch network model constructed in step 2, and updating the network weights;

[0014] Step 4: Input the test set into the trained dual-branch network model to classify the hyperspectral image, output the classification result graph, and verify the performance of the trained dual-branch network model.

[0015] When dividing the training set, several hyperspectral image patches and their corresponding central pixel labels are randomly selected from each category to generate the training set.

[0016] The central spectral feature extraction branch realizes multi-scale feature extraction through parallel group convolution, maps its own input feature dimension from coarse to fine to multi-scale feature expression, and splices spectral features of different scales in the channel dimension to obtain a multi-scale central spectral feature representation.

[0017] The global spatial feature extraction branch transforms its input features through a linear layer and then divides them into two paths. One path is calculated through HiLo attention and then passes through a feedforward network layer constructed by a fully connected layer and a deep separation convolution layer. The other path is mapped through an activation layer and the results of the two paths are subjected to a hadamard product operation to obtain a global spatial feature representation.

[0018] The HiLo attention in the global spatial feature extraction branch consists of a high-frequency attention part and a low-frequency attention part. The high-frequency attention is implemented by calculating the self-attention in the local window to characterize the image details. The low-frequency attention performs an average pooling operation on the features in the window and calculates the attention score between the reduced feature map and each image patch.

[0019] By adjusting the proportion of high-frequency attention and low-frequency attention, the global spatial feature extraction capability is enhanced.

[0020] The outputs of the central spectral feature extraction branch and the global spatial feature extraction branch are subjected to global average pooling and then concatenated, input into the fully connected layer, and output the classification result graph.

[0021] The training adopts the stochastic gradient descent method, the optimizer uses AdamW, the learning rate is set to 1e-3, the decay rate is set to 1e-4, and the loss function uses the cross entropy loss function.

[0022] The dual-branch network model selection accuracy AA, prediction accuracy OA and Kappa coefficient indicators measure the model accuracy.

[0023] A computer storable medium stores a computer program, and when the computer program is executed by a processor, a method for classifying a hyperspectral central pixel dual-branch image in a local security system is implemented.

[0024] Compared with the prior art, the present invention has the following beneficial effects:

[0025] 1. This application can effectively protect the central pixel features from being interfered by irrelevant information in the neighborhood by separating the key central pixel feature extraction process from the global feature extraction process, thereby improving the accuracy of classification.

[0026] 2. In the spatial extraction branch based on the ViT structure designed in this application, the HiLo attention mechanism is introduced to adjust the attention to the high- and low-frequency signal features in the image, which can smooth the intra-class spectral differences of the hyperspectral image while retaining the image detail information, thereby improving the global feature extraction capability.

[0027] 3. The center pixel-based dual-branch network designed in this application is concise and effective, and can achieve higher classification accuracy under limited training sample conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 A specific flow chart implemented for this application;

[0029] Figure 2 This is a structural diagram of a dual-branch network model of this application;

[0030] Figure 3 This is a structural diagram of the global spatial feature extraction branch model of this application;

[0031] Figure 4 This is the classification result diagram of this application on the Indian Pines dataset;

[0032] Figure 5 This is the classification result diagram of this application on the Pavia University dataset;

[0033] Figure 6 This is the classification result diagram of this application on the Houston2013 dataset. DETAILED DESCRIPTION

[0034] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0035] like Figure 1 As shown in FIG. 1 , a dual-branch image classification method for a hyperspectral central pixel in a local security system includes:

[0036] Step 1: Data preprocessing: using a public hyperspectral image dataset, the hyperspectral image is processed into several hyperspectral image patches and divided into training and test sets;

[0037] Step 2, constructing a dual-branch network model based on central pixels, wherein the dual branches are a central spectral feature extraction branch and a global spatial feature extraction branch;

[0038] Step 3, using the training set to train the dual-branch network model constructed in step 2, and updating the network weights;

[0039] Step 4: Input the test set into the trained dual-branch network model to classify the hyperspectral image, output the classification result graph, and verify the performance of the trained dual-branch network model.

[0040] When dividing the training set, several hyperspectral image patches and their corresponding central pixel labels are randomly selected from each category to generate the training set.

[0041] The features of the hyperspectral image are initially extracted, the central part of the overall feature map is used as the input of the central spectral feature extraction branch, and the overall feature map is patched as the input of the global spatial feature extraction branch.

[0042] The central spectral feature extraction branch realizes multi-scale feature extraction through parallel group convolution, maps its own input feature dimension from coarse to fine to multi-scale feature expression, and splices spectral features of different scales in the channel dimension to obtain a multi-scale central spectral feature representation.

[0043] The global spatial feature extraction branch transforms its input features through a linear layer and then divides them into two paths. One path is calculated through HiLo attention and then passes through a feedforward network layer constructed by a fully connected layer and a deep separation convolution layer. The other path is mapped through an activation layer and the results of the two paths are subjected to a hadamard product operation to obtain a global spatial feature representation.

[0044] The HiLo attention in the global spatial feature extraction branch consists of high-frequency attention and low-frequency attention. The high-frequency attention is implemented by calculating self-attention in the local window to characterize the image details. The low-frequency attention performs an average pooling operation on the features in the window and calculates the attention score between the reduced feature map and each image patch.

[0045] By adjusting the proportion of high-frequency attention and low-frequency attention, the global spatial feature extraction capability is enhanced.

[0046] The outputs of the central spectral feature extraction branch and the global spatial feature extraction branch are subjected to global average pooling and then concatenated, input into the fully connected layer, and output the classification result graph.

[0047] The training adopts the stochastic gradient descent method, the optimizer uses AdamW, the learning rate is set to 1e-3, the decay rate is set to 1e-4, and the loss function uses the cross entropy loss function.

[0048] The dual-branch network model selection accuracy AA, prediction accuracy OA and Kappa coefficient indicators measure the model accuracy.

[0049] A computer storable medium stores a computer program, which, when executed by a processor, implements a hyperspectral central pixel dual-branch image classification method in a local security system.

[0050] Example

[0051] A 27x27 image patch centered on the pixel to be classified is cropped from the hyperspectral image as the input of the network model. A small number of image patches and the corresponding center pixel labels in each category are randomly extracted for model training. A multi-scale spectral feature extraction branch based on the center pixel and a spatial feature extraction branch improved by the ViT structure and HiLo attention mechanism are designed and combined into the proposed dual-branch network. The constructed network model is supervised trained on the training samples, and the algorithm effect is verified on the test set.

[0052] 1. Prepare the dataset

[0053] Create the datasets required for training and testing:

[0054] (1) Indian Pine public dataset taken by AVIRIS sensor in Indiana: The size is 145*145, with 224 bands, including 200 effective bands, 16 crop categories, and a total of 10,249 labeled samples.

[0055] (2) The Pavia University public dataset taken by the ROSIS sensor: The size is 610*340, with 103 bands and 9 object categories, with a total of 42,776 labeled samples.

[0056] (3) Houston2013 public dataset provided by the Hyperspectral Image Analysis Group and the NSF-funded Center for Airborne Laser Mapping (NCALM) at the University of Houston: The dataset is 349×1905 in size and covers the urban area around the University of Houston. The images contain 144 spectral bands, 15 land feature categories, and a total of 15,029 annotated samples.

[0057] For all datasets, 20 samples from each class are selected for model training, and all the remaining samples are used for model testing.

[0058] 2. Design a dual-branch network based on center pixels

[0059] Design the central spectral feature extraction branch and the global spatial feature extraction branch to form a dual-branch network based on the central pixel. The specific network structure is as follows: Figure 2 As shown in the figure. The entire network consists of two feature extraction branches. The input data first passes through a 1x1 point convolution layer to map the number of spectral channels to the same dimension, and then passes through two group convolution layers of 1x1 and 3x3 sizes and 8 groups, a normalization layer, and an activation layer to achieve the initial extraction of spatial spectrum features. The central part of the obtained feature map will be input into the central spectral feature extraction branch, and the overall feature map will be patched and input into the global spatial feature extraction branch.

[0060] Central spectral feature extraction branch: The central spectral feature extraction branch extracts the most important central pixel discriminant features for classification from the central pixel area of ​​the hyperspectral image block. Due to the high dimensionality of hyperspectral images, the central spectral feature extraction branch uses parallel group convolution to achieve multi-scale feature extraction. Specifically, the three group convolutions are set to 4, 8, and 16 group numbers respectively, mapping the input feature dimensions from coarse to fine to multi-scale feature expressions, and then splicing the spectral features of different scales on the channel dimension to obtain the final multi-scale central spectral feature representation. The process is formalized as:

[0061]

[0062] Among them, F gconv1 ,F gconv2 ,F gconv3 They represent the features extracted when the number of groups is 4, 8, and 16, respectively. ′ Represents the embedding dimension of the feature, H ′ ×W ′ represents the size of the central pixel area, and [.] represents the concatenation operation. In the algorithm practice, the size of the final feature map is 96x3x3.

[0063] Global spatial feature extraction branch: The spatial feature extraction branch is mainly based on the ViT architecture to capture global spatial features. However, due to the existence of intra-class spectral differences, the ordinary self-attention mechanism may ignore the useful spectral information in the space. In order to overcome this problem, the spatial branch based on the ViT architecture is modified by introducing HiLo attention. The specific structure is as follows: Figure 3As shown. After the input is transformed by the linear layer, it is divided into two paths: one path is calculated by HiLo attention and then the feedforward network layer constructed by the fully connected layer and the deep separation convolution layer completes the extraction of global spatial features, and the other path uses the activation layer mapping as a gating mechanism to enhance the nonlinear expression of features. Among them, HiLo attention consists of high-frequency attention and low-frequency attention. High-frequency attention is realized by calculating self-attention in the local window to depict image details; the Q and K matrices of low-frequency attention are calculated on the feature map after average pooling, which pays more attention to the overall local part and can therefore represent low-frequency features; low-frequency attention performs average pooling operations on the features in the window, and then calculates the attention score between the reduced feature map and each image patch. The average pooling operation potentially smoothes the local spectral differences and can obtain a more general feature representation. By adjusting the proportion of high and low frequency attention α, the global feature extraction capability of the spatial branch is enhanced.

[0064] 3. Training a dual-branch network

[0065] The proposed dual-branch network of hyperspectral central pixels for on-site security system is trained. The training adopts stochastic gradient descent method, the optimizer uses AdamW, the learning rate is set to 1e-3, the decay rate is set to 1e-4, and the loss function uses the cross entropy loss function:

[0066]

[0067] Where N is the number of training samples, C is the number of categories, and y i,c is the true label (one-hot encoding) of sample i in category c. If the true category of sample i is c, then y i,c =1, otherwise y i,c =0, is the probability that the model predicts that sample i belongs to category c.

[0068] 4. Experimental Conditions

[0069] All experiments in this application were completed on a server equipped with an Intel(R) Xeon(R) Silver 4210R CPU and a 24GB NVIDIA Geforce RTX 4090 GPU for model training and testing. The system environment was Ubuntu 22.04.2. The algorithm was implemented in Python and relied on the Pytorch library. The Python version used was 3.9.18 and the Pytorch version used was 2.1.2.

[0070] The training and test data used in the experiment come from the public datasets IndianPine, Pavia University, and Houston2013.

[0071] 5. Experimental content

[0072] According to the steps given above, the two-branch network model is trained on the training set and the classification effect is evaluated on the test set. The results of the network model are shown in the following table.

[0073] Table 1 Performance of the present invention on the IndianPine dataset

[0074]

[0075]

[0076] Table 2 Performance of the present invention on the Pavia University dataset

[0077]

[0078] Table 3 Performance of the present invention on the Houston 2013 dataset

[0079]

[0080] Table 1, Table 2, and Table 3 are the results of comparing the method proposed in the present application with other methods on the data sets Indian Pine, Pavia University, and Houston 2013. The larger the accuracy (AA), prediction accuracy (OA), and Kappa coefficient, the better the classification effect. It can be seen from Table 1, Table 2, and Table 3 that compared with the prior art, the accuracy (AA), prediction accuracy (OA), and Kappa coefficient of the present application are the highest, and the classification effect is significantly better than other methods. Figure 4 , Figure 5 , Figure 6 The visualization results of the method proposed in this application are shown. Figure 4 (a) Figure 5 (a) and Figure 6 (a) is a pseudo-color image of the hyperspectral image. Figure 4 (b) Figure 5 (b) and Figure 6 (b) is the corresponding ground truth data. Figure 4 (c) Figure 5 (c) and Figure 6 (c) is the classification result diagram of the method proposed in this application. As can be seen from the figure, the classification result diagram obtained by the method of this application highly restores the real data and has a good classification effect.

[0081] In summary, the hyperspectral central pixel dual-branch image classification method for the local security system proposed in this application has achieved advanced level in hyperspectral image classification performance under limited sample conditions.

[0082] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A hyperspectral central pixel dual-branch image classification method for local security system. It is characterized in that include: Step 1: Data preprocessing: using a public hyperspectral image dataset, the hyperspectral image is processed into several hyperspectral image patches and divided into training and test sets; Step 2, constructing a dual-branch network model based on central pixels, wherein the dual branches are a central spectral feature extraction branch and a global spatial feature extraction branch; Step 3, using the training set to train the dual-branch network model constructed in step 2, and updating the network weights; Step 4: Input the test set into the trained dual-branch network model to classify the hyperspectral image, output the classification result graph, and verify the performance of the trained dual-branch network model.

2. According to the hyperspectral central pixel dual-branch image classification method for the local security system of claim 1, It is characterized in that When dividing the training set, several hyperspectral image patches and their corresponding central pixel labels are randomly selected from each category to generate the training set.

3. According to the hyperspectral central pixel dual-branch image classification method for the local security system of claim 1, It is characterized in that The central spectral feature extraction branch realizes multi-scale feature extraction through parallel group convolution, maps its own input feature dimension from coarse to fine to multi-scale feature expression, and splices spectral features of different scales in the channel dimension to obtain a multi-scale central spectral feature representation.

4. According to the hyperspectral central pixel dual-branch image classification method for the local security system of claim 1, It is characterized in that The global spatial feature extraction branch transforms its input features through a linear layer and then divides them into two paths. One path is calculated through HiLo attention and then passes through a feedforward network layer constructed by a fully connected layer and a deep separation convolution layer. The other path is mapped through an activation layer and the results of the two paths are subjected to a hadamard product operation to obtain a global spatial feature representation.

5. According to the method for classifying hyperspectral central pixel dual-branch images in a local security system according to claim 4, It is characterized in that The HiLo attention in the global spatial feature extraction branch consists of a high-frequency attention part and a low-frequency attention part. The high-frequency attention is implemented by calculating the self-attention in the local window to characterize the image details. The low-frequency attention performs an average pooling operation on the features in the window and calculates the attention score between the reduced feature map and each image patch.

6. The method for classifying hyperspectral central pixel dual-branch images in a local security system according to claim 5, It is characterized in that By adjusting the proportion of high-frequency attention and low-frequency attention, the global spatial feature extraction capability is enhanced.

7. The method for classifying hyperspectral central pixel dual-branch images in a local security system according to claim 1, It is characterized in that The outputs of the central spectral feature extraction branch and the global spatial feature extraction branch are subjected to global average pooling and then concatenated, input into the fully connected layer, and output the classification result graph.

8. The method for classifying hyperspectral central pixel dual-branch images in a local security system according to claim 1, It is characterized in that The training adopts the stochastic gradient descent method, the optimizer uses AdamW, the learning rate is set to 1e-3, the decay rate is set to 1e-4, and the loss function uses the cross entropy loss function.

9. The method for classifying hyperspectral central pixel dual-branch images in a local security system according to claim 1, It is characterized in that The dual-branch network model selection accuracy AA, prediction accuracy OA and Kappa coefficient indicators measure the model accuracy.

10. A computer storable medium storing a computer program, It is characterized in that When the computer program is executed by a processor, the method for classifying hyperspectral central pixel dual-branch images in a local security system as described in any one of claims 1 to 9 is implemented.

Citation Information

Cited By

  • Coral reef underwater hyperspectral benthic classification method based on improved Mama model

    CN121259554A

  • A method for classifying underwater hyperspectral benthic based on improved mamba model

    CN121259554B

  • Remote sensing hyperspectral image classification method, system, equipment and medium

    CN121280788A