Hyperspectral image band selection method for camouflage target identification

Abstract

The invention discloses a hyperspectral image waveband selection method for camouflage target identification. The method comprises the following steps: 1) obtaining a camouflage vehicle hyperspectral image data set; 2) preprocessing a hyperspectral image data set; 3) selection of spectral measurement parameters; 4) constructing a spectrum difference index model; 5) extracting an average spectrum vector and selecting a characteristic wave band; and 6) evaluating the wave band selection effect. According to the method, the spectral difference index model is established by combining spectral measurement parameters, and the hyperspectral image characteristic wave band is selected according to the model design algorithm; the problem that a characteristic wave band combination capable of effectively distinguishing a camouflage target and a similar background is difficult to select in the face of a camouflage target identification task by an existing wave band selection method is effectively improved.

Description

Technical Field

[0001] The present invention relates to the technical field of feature band selection of hyperspectral images, and in particular to a hyperspectral image band selection method for camouflage target recognition. Background Art

[0002] The development of remote sensing technology and imaging spectrometers has continuously improved the resolution of hyperspectral remote sensing images. The analysis of the spectral curves of various objects in hyperspectral images allows us to find the diagnostic spectral characteristics of various objects, which is the advantage of hyperspectral images over other image data. However, while the remote sensing detection capabilities have been improved, its huge amount of data will bring about the "dimensionality disaster" problem, which has brought great difficulties to the subsequent image analysis and processing. Therefore, it is very necessary to perform dimensionality reduction processing on hyperspectral images, which can eliminate redundant and irrelevant information in the image while retaining as much important discriminant information as possible.

[0003] Feature extraction and band selection are the two most commonly used methods for dimensionality reduction of hyperspectral images. Feature extraction generally uses different methods to establish a mapping between the original high-dimensional space and the low-dimensional subspace, which will cause the physical characteristics of the original hyperspectral data to be changed, which is very unfavorable for image interpretation. Commonly used feature extraction methods include principal component analysis and linear discriminant analysis. The purpose of band selection is to select a subset of bands with rich discriminant information and weak correlation from all bands while retaining the inherent characteristics of hyperspectral data as much as possible. According to the different levels of attention to the category information of the ground objects, band selection methods are often divided into two categories: information-based and category-based separability. Band selection methods based on information include the best index method, adaptive band selection method, etc., which often ignore the correlation between bands.

[0004] At present, most of the band selection methods based on category separability also focus on the overall separability of each band image. When facing the task of camouflaged target recognition, especially in the visible light band, due to the extremely high spectral similarity between the camouflaged target and the similar background, it is difficult to select a characteristic band combination that can effectively distinguish the camouflaged target from the similar background. Therefore, this paper proposes a hyperspectral image band selection method for camouflaged target recognition. Summary of the invention

[0005] Aiming at the problem that the existing band selection methods are difficult to select the characteristic band combination that can effectively distinguish the camouflaged target and the similar background when facing the camouflaged target recognition task, the present invention proposes a hyperspectral image band selection method for camouflaged target recognition. By combining spectral measurement parameters, a spectral difference index model is established to achieve accurate selection of hyperspectral image characteristic bands in the camouflaged target recognition task.

[0006] The technical solution adopted by the present invention to solve its technical problem is:

[0007] A hyperspectral image band selection method for camouflaged target recognition includes the following steps:

[0008] 1) Data collection: A hyperspectral video camera is used to obtain a hyperspectral image dataset of camouflaged vehicles; the data collection experimental scene consists of a background and a painted camouflaged vehicle;

[0009] 2) Data preprocessing: preprocess the collected hyperspectral image data set;

[0010] Due to the influence of atmospheric and lighting conditions on image quality, the acquired hyperspectral image dataset needs to be firstly atmospherically corrected and noise removed; then, the dataset is normalized to become a dimensionless dataset.

[0011] 3) Select spectral measurement parameters: Select spectral measurement parameters for building the model;

[0012] With the goal of maximizing the spectral difference between the camouflaged target and the similar background, three spectral measurement parameters, namely spectral gradient angle (SGA), normalized Fréchet distance (F) and Pearson correlation coefficient (r), were selected to construct a spectral difference index model.

[0013] 4) Constructing a spectral difference index model: Based on the correlation between the three selected spectral measurement parameters and the spectral difference, and taking into account the influence of scale, a spectral difference index model is constructed.

[0014] 5) Feature band selection: extract the average spectrum vector to select the feature band;

[0015] The average spectral vector of the camouflaged vehicle and the average spectral vector of the grass background are extracted from the data set respectively, and the two are divided into uniform subspaces. The spectral difference index SD of different bands in each subspace is calculated according to the spectral difference index model, and the band with the largest SD in each subspace is selected as the feature band.

[0016] 6) Band selection effect evaluation:

[0017] The data set is divided into a training set and a test set, and a support vector machine classifier (SVM) is used to carry out a target recognition experiment. The effect of the band selection method of the present invention is verified by comparing the overall accuracy OA and recognition result graph with other band selection methods.

[0018] Further, the process of step 3) is:

[0019] Step 3.1: To characterize the morphological difference between the spectral vectors of the camouflaged target and the similar background, the spectral gradient angle is introduced as the first component parameter of the model. It first takes the first-order derivative of the two spectral vectors to obtain the corresponding gradient vectors, and then calculates the generalized angle between the two gradient vectors, namely the spectral gradient angle. For two spectral vectors x and y with a dimension of n, their gradient vectors SG(x) and SG(y) are respectively:

[0020] SG(x)=(x2-x1,x3-x2,x4-x3,x5-x4,…,x n -x n-1 )

[0021] SG(y)=(y2-y1,y3-y2,y4-y3,y5-y4,…,y n -y n-1 )

[0022] In the formula, x i and i are the amplitude values ​​of the ith spectral dimension of the spectral vectors x and y respectively. The spectral gradient angle formed by the two gradient vectors is:

[0023]

[0024] Step 3.2: In order to characterize the amplitude difference between the spectral vectors of the camouflaged target and the similar background, the normalized Fréchet distance is introduced as the second component parameter of the model. The Fréchet distance between the spectral vectors is the discretized Fréchet distance. Assume that the hyperspectral image matrix with width and height of M and N is Y = [Y1,…,Y i ,…,Y j ,…,Y MN ], define two continuous mappings x:[1,n]→Y i and y:[1,n]→Y j , where n is the number of bands, x represents the spectral vector of the i-th pixel, and y represents the spectral vector of the j-th pixel. The Fréchet distance between two spectral vectors is expressed as:

[0025]

[0026] In the formula, d is the distance metric function, and here the Euclidean distance is used. In the formula, the maximum value of the Euclidean distance between the sampling points x(τ) and y(τ) between the two curves is first obtained, and then this series of maximum distances are normalized to find their minimum value, which is the normalized Fréchet distance.

[0027] Step 3.3: In order to measure the strength of the correlation between the spectral vectors of the camouflaged target and the similar background, and to strengthen the independence and difference of the selected bands, the Pearson correlation coefficient is introduced as the third component parameter of the model. The Pearson correlation coefficient between the spectral vectors x and y is expressed as:

[0028]

[0029] Where n is the spectral dimension of the spectral vectors x and y; x i and i are the amplitude values ​​of the i-th spectral dimension of the spectral vectors x and y respectively; x and y are the means of the two spectral vectors respectively; the value range of r(x,y) is (-1,1), the larger the value, the greater the correlation between the spectral vectors, and vice versa.

[0030] Further, the process of step 4) is:

[0031] Among the three selected spectral difference measurement parameters, the spectral gradient angle and the normalized Fréchet distance are both statistically positively correlated with the spectral difference index, but the degree of correlation between the two is different. Therefore, in the model, the two are combined by dot product and placed in the numerator; the Pearson correlation coefficient r is statistically negatively correlated with the spectral difference index. The Pearson correlation coefficient between the two spectral vectors is (-1,1). When its value is negative, it will cause the contribution of other model parameters to the model to drop sharply. Therefore, a constant b is added to the Pearson correlation coefficient to make the results all positive to eliminate this effect. At the same time, in order to maximize the contribution of the Pearson correlation coefficient to the model, the value of the constant b is taken as 1, so that the correlation coefficient is expanded to be measured by r+1∈(0,1) and is placed in the denominator.

[0032] Assuming that the two spectral vectors are x and y, the spectral difference index model is expressed as follows:

[0033]

[0034] The spectral difference index SD(x,y) can quantitatively characterize the difference between spectral vectors x and y. The larger the value of SD(x,y), the greater the morphological difference and amplitude difference between the two spectral vectors in the corresponding band, the smaller the correlation, and the higher the priority of being selected as the feature band.

[0035] Further, the process of step 6) is:

[0036] Step 6.1: Overall accuracy and recognition result comparison

[0037] The experiment uses the support vector machine classifier SVM to identify the camouflaged vehicle targets in the data set, and the radial basis function is used as the kernel function of the classifier. The overall accuracy OA (Overall Accuracy) is selected to evaluate the recognition effect of different band selection methods. 80% of the samples in the two data sets are randomly selected as training samples for each SVM, and the rest are used as test samples. All experiments are repeated 10 times, and the average of all results is taken as the final experimental result. The five methods involved in the comparative experiment are the optimal clustering framework method OCF, the end-to-end band selection network framework BS_Net, the genetic optimization algorithm GA, the band selection method based on spectral angle SAM and the band selection method based on Fréchet distance. The parameter settings in the comparative methods all follow their default settings.

[0038] The experiment compares and analyzes the performance of different methods by changing the number of selected bands, and selects bands in the range of 10-60 with a step of 10 to compare the recognition results;

[0039] Step 6.2: Computational efficiency comparison

[0040] The experiment compares the computational efficiency of the SD band selection method with five other methods.

[0041] The beneficial effects of the present invention are mainly manifested in that the present invention combines spectral measurement parameters, establishes a spectral difference index model, and selects characteristic bands of hyperspectral images according to a model design algorithm, thereby improving the problem that the existing band selection method is difficult to select a characteristic band combination that can effectively distinguish between a camouflaged target and a similar background when facing a camouflaged target recognition task. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 The real image map and ground truth map of the camouflaged vehicle data set used in the present invention;

[0043] Figure a is a real image of the camouflaged vehicle dataset used in the present invention; Figure b is a ground truth map of the camouflaged vehicle dataset used in the present invention;

[0044] Figure 2 It is a comparison chart of the overall recognition accuracy of different methods;

[0045] Figure 3 6 is a recognition result diagram of the band combination selected by the methods of the embodiment;

[0046] Figure a is the recognition result of the band combination selected by the SD method; Figure b is the recognition result of the band combination selected by the OCF method; Figure c is the recognition result of the band combination selected by the BS_Netd method; Figure e is the recognition result of the band combination selected by the GA method; and Figure f is the recognition result of the band combination selected by the SAM method. DETAILED DESCRIPTION

[0047] The present invention will be further described below in conjunction with the accompanying drawings.

[0048] Reference Figure 1 to Figure 3 , a hyperspectral image band selection method for camouflage target recognition, comprising the following steps:

[0049] Step 1) Data collection: Use Lingyunguang HV-125 series hyperspectral video cameras to obtain a hyperspectral image dataset of camouflaged vehicles;

[0050] The data collection experimental scene consists of green grass and green painted camouflaged vehicles. The image size is 1387 pixels × 1829 pixels. It has 128 spectral response bands in the wavelength range of 450-950 nanometers and a spectral resolution of 4 nanometers. The real image and ground truth image are shown in Figure 1 a and Figure 1 b.

[0051] Step 2) Data preprocessing: Preprocess the collected hyperspectral image dataset

[0052] Due to the influence of atmospheric and lighting conditions on image quality, the acquired hyperspectral image dataset needs to be firstly atmospherically corrected and noise removed; then, the dataset is normalized to become a dimensionless dataset.

[0053] Step 3) Select spectral metric parameters: Select the spectral metric parameters used to build the model

[0054] With the goal of maximizing the spectral difference between the camouflaged target and the similar background, three spectral measurement parameters, namely spectral gradient angle (SGA), normalized Fréchet distance (F) and Pearson correlation coefficient (r), were selected to construct a spectral difference index model.

[0055] Step 3.1: To characterize the morphological difference between the spectral vectors of the camouflaged target and the similar background, the spectral gradient angle is introduced as the first component parameter of the model. It first takes the first-order derivative of the two spectral vectors to obtain the corresponding gradient vectors, and then calculates the generalized angle between the two gradient vectors, namely the spectral gradient angle. For two spectral vectors x and y with a dimension of n, their gradient vectors SG(x) and SG(y) are respectively:

[0056] SG(x)=(x2-x1,x3-x2,x4-x3,x5-x4,…,x n -x n-1 )

[0057] SG(y)=(y2-y1,y3-y2,y4-y3,y5-y4,…,y n -yn-1 )

[0058] In the formula, x i and i are the amplitude values ​​of the ith spectral dimension of the spectral vectors x and y respectively. The spectral gradient angle formed by the two gradient vectors is:

[0059]

[0060] Step 3.2: In order to characterize the amplitude difference between the spectral vectors of the camouflaged target and the similar background, the normalized Fréchet distance is introduced as the second component parameter of the model. The Fréchet distance between the spectral vectors is the discretized Fréchet distance. Assume that the hyperspectral image matrix with width and height of M and N is Y = [Y1,…,Y i ,…,Y j ,…,Y MN ], define two continuous mappings x:[1,n]→Y i and y:[1,n]→Y j , where n is the number of bands, x represents the spectral vector of the i-th pixel, and y represents the spectral vector of the j-th pixel. The Fréchet distance between two spectral vectors is expressed as:

[0061]

[0062] In the formula, d is the distance metric function, and here the Euclidean distance is used. In the formula, the maximum value of the Euclidean distance between the sampling points x(τ) and y(τ) between the two curves is first obtained, and then this series of maximum distances are normalized to find their minimum value, which is the normalized Fréchet distance.

[0063] Step 3.3: In order to measure the strength of the correlation between the spectral vectors of the camouflaged target and the similar background, and to strengthen the independence and difference of the selected bands, the Pearson correlation coefficient is introduced as the third component parameter of the model. The Pearson correlation coefficient between the spectral vectors x and y is expressed as:

[0064]

[0065] Where n is the spectral dimension of the spectral vectors x and y; x i and i are the amplitude values ​​of the i-th spectral dimension of the spectral vectors x and y respectively; x and y are the means of the two spectral vectors respectively; the value range of r(x,y) is (-1,1), the larger the value, the greater the correlation between the spectral vectors, and vice versa.

[0066] Step 4) Construct spectral difference index model

[0067] According to the correlation between the three selected spectral measurement parameters and the spectral difference, and considering the influence of scale, a spectral difference index model is constructed.

[0068] Among the three selected spectral difference measurement parameters, the spectral gradient angle and the normalized Fréchet distance are both statistically positively correlated with the spectral difference index, but the degree of correlation between the two is different. Therefore, in the model, the two are combined by dot product and placed in the numerator; the Pearson correlation coefficient r is statistically negatively correlated with the spectral difference index. The Pearson correlation coefficient between the two spectral vectors is (-1,1). When its value is negative, it will cause the contribution of other model parameters to the model to drop sharply. Therefore, a constant b is added to the Pearson correlation coefficient to make the results all positive to eliminate this effect. At the same time, in order to maximize the contribution of the Pearson correlation coefficient to the model, the value of the constant b is taken as 1, so that the correlation coefficient is expanded to be measured by r+1∈(0,1) and is placed in the denominator.

[0069] Assuming that the two spectral vectors are x and y, the spectral difference index model is expressed as follows:

[0070]

[0071] The spectral difference index SD(x,y) can quantitatively characterize the difference between spectral vectors x and y. The larger the value of SD(x,y), the greater the morphological difference and amplitude difference between the two spectral vectors in the corresponding band, the smaller the correlation, and the higher the priority of being selected as the feature band.

[0072] Step 5) Feature band selection: Extract the average spectral vector to select the feature band

[0073] The average spectral vector of the camouflaged vehicle and the average spectral vector of the grass background are extracted from the data set respectively, and the two are divided into uniform subspaces. The spectral difference index SD of different bands in each subspace is calculated according to the spectral difference index model, and the band with the largest SD in each subspace is selected as the feature band.

[0074] Step 6) Band selection effect evaluation

[0075] The data set is divided into a training set and a test set, and a support vector machine classifier (SVM) is used to carry out a target recognition experiment. The effect of the band selection method of the present invention is verified by comparing the overall accuracy OA and recognition result graph with other band selection methods.

[0076] Step 6.1: Overall accuracy and recognition result comparison

[0077] The experiment uses the support vector machine classifier SVM to identify the camouflaged vehicle targets in the data set, and the radial basis function is used as the kernel function of the classifier. The overall accuracy OA (Overall Accuracy) is selected to evaluate the recognition effect of different band selection methods. 80% of the samples in the two data sets are randomly selected as training samples for each SVM, and the rest are used as test samples. All experiments are repeated 10 times, and the average of all results is taken as the final experimental result. The five methods involved in the comparative experiment are the optimal clustering framework method OCF, the end-to-end band selection network framework BS_Net, the genetic optimization algorithm GA, the band selection method based on spectral angle SAM and the band selection method based on Fréchet distance. The parameter settings in the comparative methods all follow their default settings.

[0078] The software environment for the proposed method and BS_Net, GA, SAM and F is Python 3.10, and the OCF method is run in Matlab 2021b. The hardware environment for all methods is: Intel Core i7-12700H 2.30GHz CPU and NVIDIA GeForce RTX 3060.

[0079] The experiment compares and analyzes the performance of different methods by changing the number of selected bands, and selects bands in the band range of 10-60 with a step size of 10 to compare the recognition results. Figure 2 It is the overall accuracy curve of different methods under different band subset sizes. Figure 3 The following figure shows the recognition results of 6 methods on this dataset when the band subset size is 60.

[0080] Figure 2 It shows that the SD band selection method proposed in the present invention outperforms other band selection methods. When the number of selected bands is small, the SD band selection method can achieve good recognition accuracy, and when the band subset size is 60, it exceeds the recognition accuracy of the full band, indicating that the SD band selection method can select representative feature band subsets. In addition, the BS_Net band selection method also achieves the recognition effect of the full band when the band subset size is 50 and 60. The recognition accuracy of the OCF band selection method decreases significantly with the decrease in the number of band subsets. The GA band selection method exhibits a "dimensionality disaster" phenomenon when the number of bands is 50, which is a very obvious decline. Figure 3 It can be seen that the identification result map obtained by the SD band selection method has better homogeneity and continuity.

[0081] Step 6.2: Computational efficiency comparison

[0082] The experiment compares the computational efficiency of the SD band selection method with the other five methods. Table 2 compares the time required for different methods to select 50 feature bands in the camouflaged vehicle dataset.

[0083] Table 2 Comparison of calculation time of different methods

[0084]

[0085] As can be seen from Table 2, the SD band selection method takes the least time in this data set; the OCF, SAM and F band selection methods also take less time; the BS_Net and GA methods have lower computational efficiency and both take the most time. At the same time, combined with the classification accuracy and classification result graph, in this experimental camouflaged vehicle data set, the SD band selection method has a better overall performance and can provide better support for the application of hyperspectral image camouflaged target recognition.

[0086] The method of the present invention combines spectral measurement parameters to establish a spectral difference index model, and selects the characteristic bands of the hyperspectral image according to the model design algorithm. It can improve the problem that the existing band selection method is difficult to select the characteristic band combination that can effectively distinguish the camouflaged target from the similar background when facing the camouflaged target recognition task, and has universality and generality.

[0087] The contents described in the embodiments of this specification are merely an enumeration of the implementation forms of the inventive concept. The protection scope of the present invention should not be regarded as limited to the specific forms described in the embodiments. The protection scope of the present invention also extends to equivalent technical means that can be conceived by those skilled in the art based on the inventive concept.

Claims

1. A hyperspectral image band selection method for camouflaged target recognition, characterized in that: The method comprises the following steps: Step 1) Data collection: The data collection experimental scene includes the background and the painted camouflaged vehicle, and a hyperspectral image dataset of the camouflaged vehicle is obtained using a hyperspectral video camera; Step 2) Data preprocessing: Perform atmospheric correction and noise removal on the collected hyperspectral image data set; perform data standardization on the data set to make it dimensionless; Step 3) Select spectral measurement parameters: To maximize the spectral difference between the camouflaged target and the similar background, three spectral measurement parameters, namely, spectral gradient angle SGA, normalized Fréchet distance F, and Pearson correlation coefficient r, are selected to construct a spectral difference index model; Step 4) constructing a spectral difference index model: constructing a spectral difference index model based on the correlation between the three selected spectral measurement parameters and the spectral difference and combining the influence of the scale; Step 5) Feature band selection: extract the average spectral vector of the camouflaged vehicle and the average spectral vector of the background from the data set, divide the two into uniform subspaces, calculate the spectral difference index SD of different bands in each subspace according to the spectral difference index model, and select the band with the largest SD in each subspace as the feature band; Step 6) Band selection effect evaluation: The data set is divided into a training set and a test set, and a target recognition experiment is carried out using a support vector machine classifier SVM. The effect of the band selection method of the present invention is verified by comparing the overall accuracy OA and the recognition result graph with other band selection methods.

2. The hyperspectral image band selection method for camouflaged target recognition according to claim 1, characterized in that: The process of step 3) is: Step 3.1: In order to characterize the morphological difference between the spectral vectors of the camouflaged target and the similar background, the spectral gradient angle is introduced as the first component parameter of the model. It first takes the first-order derivative of the two spectral vectors to obtain the corresponding gradient vectors, and then calculates the generalized angle between the two gradient vectors, namely the spectral gradient angle; for the two spectral vectors x and y with a dimension of n, their gradient vectors SG(x) and SG(y) are respectively: SG(x)=(x2-x1,x3-x2,x4-x3,x5-x4,…,x n -x n-1 ) <h2 style=";text-align:left;direction:ltr">SG(y) = (y2 - y1,y3 - y2,y4 - y3,y5 - y4,…,y<h2 style=";text-align:left;direction:ltr"> n <h2 style=";text-align:left;direction:ltr"> -y<h2 style=";text-align:left;direction:ltr"> n-1 <h2 style=";text-align:left;direction:ltr"> ) In the formula, x i and i are the amplitude values ​​of the i-th spectral dimension of the spectral vectors x and y respectively; the spectral gradient angle formed by the two gradient vectors is: Step 3.2: In order to characterize the amplitude difference between the spectral vectors of the camouflaged target and the similar background, the normalized Fréchet distance is introduced as the second component parameter of the model; the Fréchet distance between the spectral vectors is the discretized Fréchet distance. Assume that the hyperspectral image matrix with width and height of M and N is Y = [Y1,…,Y i ,…,Y j ,…,Y MN ], define two continuous mappings x:[1,n]→Y i and y:[1,n]→Y j , where n is the number of bands, x represents the spectral vector of the i-th pixel, and y represents the spectral vector of the j-th pixel; the Fréchet distance between two spectral vectors is expressed as: Where d is the distance metric function. First, the maximum value of the Euclidean distance between the sampling points x(τ) and y(τ) between the two curves is obtained. Then, this series of maximum distances are normalized to find their minimum value, which is the normalized Fréchet distance. Step 3.3: In order to measure the strength of the correlation between the spectral vectors of the camouflaged target and the similar background and to strengthen the independence and difference of the selected bands, the Pearson correlation coefficient is introduced as the third component parameter of the model; the Pearson correlation coefficient between the spectral vectors x and y is expressed as: Where n is the spectral dimension of the spectral vectors x and y; x i and i are the amplitude values ​​of the i-th spectral dimension of the spectral vectors x and y respectively; and are the means of the two spectral vectors respectively; the value range of r(x,y) is (-1,1), the larger the value is, the greater the correlation between the spectral vectors is, and vice versa.

3. The hyperspectral image band selection method for camouflaged target recognition according to claim 1, characterized in that: The process of step 4) is: The spectral gradient angle and the normalized Fréchet distance are both statistically positively correlated with the spectral difference index. In the model, the two are combined by dot product and placed in the numerator. The Pearson correlation coefficient r is statistically negatively correlated with the spectral difference index. The Pearson correlation coefficient between the two spectral vectors is (-1,1), and a constant b is added to the Pearson correlation coefficient. At the same time, in order to maximize the contribution of the Pearson correlation coefficient to the model, the constant b is taken as 1, so that the correlation coefficient is expanded to be measured by r+1∈(0,1) and placed in the denominator. Assuming that the two spectral vectors are x and y, the spectral difference index model is expressed as follows: The spectral difference index SD(x,y) can quantitatively characterize the difference between spectral vectors x and y. The larger the value of SD(x,y), the greater the morphological difference and amplitude difference between the two spectral vectors in the corresponding band, the smaller the correlation, and the higher the priority of being selected as the feature band.

4. The hyperspectral image band selection method for camouflaged target recognition according to claim 1, characterized in that: The process of step 6) is: Step 6.1: Overall accuracy and recognition result comparison: The experiment uses support vector machine classifier SVM to identify camouflaged vehicle targets in the data set, and radial basis function is used as the kernel function of the classifier; the overall accuracy OA is selected to evaluate the recognition effect of different band selection methods; 80% of the samples in the two data sets are randomly selected as training samples of each SVM, and the rest are used as test samples; all experiments are repeated 10 times, and the average of all results is taken as the final experimental result; The five methods involved in the comparative experiment are the optimal clustering framework method OCF, the end-to-end band selection network framework BS_Net, the genetic optimization algorithm GA, the spectral angle SAM and the Fréchet distance-based band selection method. The parameter settings in the comparative methods all follow their default settings. The experiment compares and analyzes the performance of different methods by changing the number of selected bands, and selects bands in the range of 10-60 with a step of 10 to compare the recognition results; Step 6.2: Computational efficiency comparison: The experiment compares the computational efficiency of the SD band selection method with five other methods.

Images