A hyperspectral image target detection method based on spectral mask driven background learning

Abstract

The application discloses a hyperspectral image target detection method based on spectral mask driving background learning, and belongs to the field of remote sensing image processing. The scheme is as follows: background sample screening is performed on an original hyperspectral image, and high-confidence background samples obtained are used as training samples of a network; a band mask is constructed for the obtained background spectrum as a mask training sample; two parallel automatic encoder networks are constructed, and the original background spectrum sample and the background spectrum sample subjected to the mask are used for training the two automatic encoder networks respectively; a constraint loss is constructed between the encoder outputs of the two automatic encoders during training; and the reconstruction error of a reconstructed hyperspectral image of the automatic encoder trained by using the original background spectrum is calculated and used as a detection result. The application can fully utilize background spectrum information, improve the target detection precision of the hyperspectral image, and can be used for target reconnaissance, mineral exploration and field search and rescue.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence and hyperspectral remote sensing image processing technology, and particularly relates to a hyperspectral image target detection method based on spectral mask driven background learning. BACKGROUND

[0002] The hyperspectral remote sensing imaging system can obtain an image technology containing two-dimensional spatial distribution information and one-dimensional continuous narrow-band spectral information of ground objects. The obtained image not only reveals the spatial geometric distribution of the observed ground objects, but also extracts the radiation intensity reflected by the ground objects through analysis of the continuous narrow-band spectral information. Due to the difference in chemical composition of different substances, they each have unique spectral characteristics, and the unique spectral information of different ground objects in the hyperspectral image can be used for identification and detection of these ground objects.

[0003] Traditional hyperspectral image target detection methods mainly focus on the analysis of spectral characteristics, and are divided into two categories: spectral matching algorithm and probability statistics algorithm. The spectral matching algorithm detects the target by designing various calculation methods to calculate the similarity between the target spectrum and the spectrum of all pixels in the image to be detected, which has the characteristics of simple calculation and fast recognition. However, in actual application, when the hyperspectral imager photographs the ground object, the imaging result will be distorted to different degrees due to factors such as clouds, light, scattering, and terrain shadows, resulting in different spectral curves of the target to be identified in different positions in the image, i.e. the "same object different spectrum" phenomenon, which makes it difficult to detect the target with spectral variation by simply matching the similarity. Spectral information difference (SID) and its derivative algorithms are typical representatives of deterministic algorithms. The probability statistics algorithm usually assumes that all pixels in the hyperspectral image follow a Gaussian distribution, and the target pixel to be detected is regarded as an abnormal point that does not conform to this distribution, and the target is detected by identifying these abnormal points. Among them, the constrained energy minimization (CEM) algorithm is one of the most representative algorithms in this category. Although these two methods have the advantages of simple implementation and rapid detection in target detection, they usually need to set more assumptions and constraints, and can only achieve better detection performance under ideal conditions, and ignore the rich background spectral information in the hyperspectral image.

[0004] In 2020, Xie et al. published a paper entitled "Background Learning Based on Target Suppression Constraint for Hyperspectral Target Detection" in IEEE J. Sel. Topics Appl. Earth Observ. Remote Sens., which disclosed a target suppression constraint-based hyperspectral target detection background learning (BLTSC) detection method that achieved advanced detection performance. The background learning-based solution thereof employs adversarial autoencoders and their various variants to pursue the best background spectrum simulation. The background spectrum distribution is learned by driving the autoencoder through adversarial learning, and then the hyperspectral image is reconstructed. The difference between the reconstructed image and the hyperspectral image to be detected is calculated to realize target detection. Such a method has less dependence on target prior spectrum and uses more abundant background samples as the training set. However, the adversarial-based background learning strategy is based on the assumption that the background spectrum distribution is consistent with the normal Gaussian distribution. This assumption is an idealized situation, and there are more uncertainties in real-world application scenarios. In addition, the pattern collapse problem existing in the adversarial learning process causes false detection in the detection result, making it difficult to deal with high-precision recognition in complex backgrounds. SUMMARY

[0005] The purpose of the present application is to propose a hyperspectral image target detection method based on spectrum mask-driven background learning to solve the problem of low detection accuracy caused by the dependence of existing background learning-based hyperspectral image detection algorithms on adversarial learning and the pattern collapse caused by adversarial learning.

[0006] The technical scheme of the present application is as follows: a double-window background selection strategy is used to select high-confidence background pixels as the original background training sample set, and a neighborhood spectrum mean mask operation is used to obtain a mask background training sample set. Two autoencoders are used to learn the original background spectrum reconstruction and the original background spectrum recovery, respectively. The consistency of the output of the middle encoder of the two autoencoders is constrained to promote the autoencoder learning the background spectrum reconstruction to have robust background reconstruction capability. The trained background spectrum reconstruction autoencoder is used to reconstruct the hyperspectral image to be detected, and target detection is realized by calculating the reconstruction error.

[0007] According to the above idea, the specific implementation steps of the present application include the following:

[0008] (1) A hyperspectral image X to be detected with a size of M×N×P and the true spectral vector d of the target to be detected are selected from the hyperspectral image library, where M and N represent the width and height of the image, respectively, and P represents the number of bands of the image. Both M and N are positive integers greater than 0, and P is a positive integer greater than or equal to 100.

[0009] (2) Screen the background samples of the hyperspectral image X and select C high-confidence background pixels from M×N pixels as training samples B;

[0010] (3) Performing a neighborhood spectrum mean masking operation on each band of the spectrum of each pixel in the training sample B to obtain a masked training sample B′;

[0011] (4) Construct a background learning network consisting of two autoencoder networks AE1 and AE2;

[0012] (5) Use B and B′ to perform joint iterative training on the autoencoder networks AE1 and AE2 respectively, and the loss function is the minimum mean square error function to obtain the trained autoencoder network AE1;

[0013] (6) Input the hyperspectral image to be detected X into the trained autoencoder AE1 to obtain the reconstructed hyperspectral image output X′, and calculate the pixel x at each corresponding position in X and X′. i and x′ i The spectral angle SAM(si) between , M×N spectral angle values ​​are obtained, where:

[0014]

[0015] Among them, cos -1 () represents the inverse cosine operation, T represents the transpose operation, and these M×N spectral angle values ​​are used as the detection results.

[0016] (7) According to the weak background detection characteristics, a mean square nonlinear background filter is constructed to convert the preliminary detection results Each element in is input into the mean square nonlinear background filter for background suppression, and the final detection result with good background suppression is obtained.

[0017] In step (2), C high-confidence background pixels are selected as training samples B, and a dual-window detector and spectral angle calculation are used to implement the method as follows:

[0018] (2a) In X, for each pixel x i As the center, construct two concentric square windows of different sizes, where the inner window size is w in , the outer window size is w out, take out the pixels between the two windows and record them as:

[0019]

[0020] Where L = win out ×win out -win in ×win in Indicates the number of pixels taken out, Indicates the pixel to be taken out;

[0021] (2b) Calculate the pixels taken out above The spectral angle A with the prior target spectral vector d l , calculated as follows:

[0022]

[0023] (2c) All A l Pixels > 0.3 Take out the high confidence background pixels in the pixel neighborhood and record these pixels as S BKG_i ;

[0024] (2d) Repeat (2a)-(2c) for all pixels in x to obtain the neighborhood background pixels S of all pixels. BKG =[S BKG_1 , S BKG_2 ,...,S BKG_H×W ];

[0025] (2e) For the above obtained S BKG =[S BKG_1 , S BKG_2 ,...,S BKG_H×W ]Use the deduplication function Unique(·) to remove S BKG The repeated background pixels in the image are used to obtain the non-repeated background pixel training sample set B.

[0026] In step (3), the mask training sample set B′ is obtained by performing a neighborhood spectrum mean mask operation, which is implemented as follows:

[0027] (3a) Take a background pixel sample from the training sample set B Any of the bands β p Replace it with the mean of the spectral bands in its three neighborhoods, and the calculation method is as follows:

[0028]

[0029] where β′ p Represents β pThe result after the neighborhood spectrum mean mask, mean() represents the mean operation;

[0030] (3b) Repeat the above operation (3a) for all pixels in the training sample set B to obtain the mask training sample set B′.

[0031] In the step (4), a background learning network consisting of two autoencoder networks AE1 and AE2 is constructed. AE1 and AE2 have the same structure and are both composed of an encoding network and a decoding network connected thereto, wherein the decoder network of AE1 and the decoder network of AE2 share weights with each other; the encoder network comprises an input layer, a fully connected layer, and an output layer, wherein the number of nodes in the input layer is equal to the number P of bands of the hyperspectral image to be measured, the number of nodes in the fully connected layer is equal to 256, and the number of nodes in the output layer is equal to 128; the decoder network comprises an input layer, a fully connected layer, and an output layer, and its structure is symmetrical to the encoder structure, wherein the dimension of the input layer is equal to 128, the dimension of the fully connected layer is equal to 256, and the dimension of the output layer is equal to P.

[0032] In step (5), B and B′ are used as inputs of the autoencoder networks AE1 and AE2 respectively for simultaneous iterative training, which is implemented as follows:

[0033] (5a) Set the number of rounds of current model training to e1, the total number of training rounds to E = 2000, and set e1 = 1;

[0034] (5b) Sample b in the background training sample set B i As the input of the autoencoder AE1, after passing through the encoder and decoder, the encoder output z is obtained i and decoder output At the same time, the sample b′ in the background training sample set B′ i As the input of the autoencoder AE2, after passing through the encoder and decoder, the encoder output z′ is obtained i and decoder output

[0035] (5c) Calculate the loss values ​​of autoencoders AE1 and AE2 in this round of training according to the following loss function calculation formula:

[0036]

[0037]

[0038]

[0039] Among them, L r1 is the reconstruction loss value of the autoencoder AE1, L r2 is the recovery loss value when the autoencoder AE2 restores the masked spectral sample to the pre-masked spectral sample, Lziz is the similarity loss between the encoder outputs in the two autoencoders AE1 and AE2, Indicates the two-norm operation;

[0040] (5d) Using the Adam optimizer and back-propagation algorithm, update the weights of the neuron nodes in the autoencoder AE1 and the autoencoder AE2 so that the total network loss L = L r1 +L r2 +L ziz Continuously decrease and approach 0;

[0041] (5e) Determine whether the current number of training rounds e1 is equal to the total number of training rounds E:

[0042] If yes, stop training, get the trained autoencoder AE1, and save it;

[0043] Otherwise, let e1=e1+1 and return to (5b).

[0044] The mean square nonlinear background filter f(x) constructed in step (7) is implemented as follows:

[0045]

[0046] Where λ is the adjustable parameter of the mean square nonlinear background filter, x is the preliminary detection result obtained in step (6) The preliminary detection results of each pixel in .

[0047] Compared with the prior art, the present invention has the following advantages:

[0048] First, the background spectrum sample selection method used in the present invention is implemented based on a local double-window strategy, which protects possible target pixels through the inner window and screens high-confidence background samples based on the spectral angles of the pixels between the inner window and the outer window and the prior spectrum, thereby ensuring the accuracy of subsequent background learning.

[0049] Second, the background learning strategy used in the present invention does not require the use of a discriminator for adversarial learning, avoiding the reliance of existing hyperspectral image detection algorithms based on background learning on adversarial learning, as well as the problem of low detection accuracy caused by pattern collapse caused by adversarial learning.

[0050] Third, the background filter constructed by the present invention effectively suppresses misdetected background pixels, increasing the discernibility of the detected target. This effectively avoids the problem of existing guided filters failing to suppress weakly detected background elements. Simulation results demonstrate that the present invention has higher detection accuracy and background suppression capabilities than existing detection algorithms. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0052] Figure 1 A schematic diagram of a method flow chart provided for an example of the present invention;

[0053] Figure 2 A pseudo-color image of a real hyperspectral image and a corresponding true value label map used in the embodiment of the present invention;

[0054] Figure 3 This is a target detection simulation image of a real hyperspectral image using the existing hyperspectral target detection method based on target suppression constraint-based hyperspectral target detection background learning BLTSC;

[0055] Figure 4 This is a target detection simulation image of the same real hyperspectral image using the method of the present invention. DETAILED DESCRIPTION

[0056] The present invention is further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention. After reading the present invention, modifications of various equivalent forms of the present invention made by those skilled in the art all fall within the scope defined by the claims attached to this application.

[0057] like Figure 1 As shown, the present invention provides a hyperspectral image target detection method based on spectral mask driven background learning, which includes the following steps.

[0058] Step 1: Obtain the hyperspectral image X to be detected and the true spectral vector d of the target to be detected. From the hyperspectral image library, select a hyperspectral image X to be detected of size M × N × P, and a true spectral vector d that is similar to the spectral curve of the target to be detected contained in the hyperspectral image X. M and N represent the width and height of the image, respectively, and P represents the number of image bands. Both M and N are positive integers greater than 0, and P is a positive integer greater than or equal to 100. In this example, the hyperspectral image X to be detected is a true hyperspectral image acquired by the airborne visible / infrared imaging spectrometer AVIRIS. It has 204 spectral bands, a size of 100 × 100, and a total of 67 target pixels to be detected.

[0059] Step 2: Screen the hyperspectral image X for background samples. C high-confidence background pixels are selected from the M×N pixels as training samples B. Traditional background spectrum sample selection methods include the constrained energy minimization (CEM) algorithm, the spectral angle mapping (SAM) algorithm, and the density-based noisy spatial clustering (DBSCAN) method. This example uses, but is not limited to, a dual-window detector and the spectral angle mapping (SAM) algorithm to screen high-confidence background samples from the hyperspectral image X. The specific implementation is as follows:

[0060] (2.1) Input hyperspectral image X and prior target spectrum vector d, in X, each pixel x i As the center, construct two concentric square windows of different sizes, where the inner window size is w in =9, the outer window size is w out =11, take out the pixels between the two windows and record them as:

[0061]

[0062] Where L = win out ×win out -win in ×win in =40 represents the number of pixels taken out, Indicates the pixel to be taken out;

[0063] (2.2) Calculate the pixels taken out above The spectral angle A with the prior target spectral vector d l , calculated as follows:

[0064]

[0065] (2.3) All A l Pixels > 0.3 Take out the high confidence background pixels in the pixel neighborhood and record these pixels as S BKG_i ;

[0066] (2.4) Repeat (2.1)-(2.3) for all pixels in x to obtain the neighborhood background pixels S of all pixels. BKG =[S BKG_1 , S BKG_2 ,...,S BKG_H×W ];

[0067] (2.5) For the above obtained S BKG =[S BKG_1 , S BKG_2 ,...,S BKG_H×W ]Use the deduplication function Unique(·) to remove S BKGThe repeated background pixels in the image are used to obtain the non-repeated background pixel training sample set B.

[0068] Step 3: Perform a neighborhood spectrum mean masking operation on each band of the spectrum of each pixel in the training sample set B to obtain a masked training sample set B′.

[0069] (3.1) Take a background pixel sample from the training sample set B Any of the bands β p Replace it with the mean of the spectral bands in its three neighborhoods, and the calculation method is as follows:

[0070]

[0071] Among them, β′ p Represents β p The result after the neighborhood spectrum mean mask, mean() represents the mean operation;

[0072] (3.2) Repeat the above operation (3a) for all pixels in the training sample set B to obtain the mask training sample set B′.

[0073] Step 4: Build and train autoencoders AE1 and AE2.

[0074] (4.1) The background learning network composed of the autoencoder networks AE1 and AE2 has the same structure, and is composed of an encoder network and a decoder network connected thereto, wherein the decoder network of AE1 and the decoder network of AE2 share weights with each other; the encoder network comprises an input layer, a fully connected layer, and an output layer, wherein the number of nodes in the input layer is equal to the number of bands P of the hyperspectral image to be measured, the number of nodes in the fully connected layer is equal to 256, and the number of nodes in the output layer is equal to 128; the decoder comprises an input layer, a fully connected layer, and an output layer, and its structure is symmetrical to that of the encoder, wherein the dimension of the input layer is equal to 128, the dimension of the fully connected layer is equal to 256, and the dimension of the output layer is equal to P, and in this example, P = 189;

[0075] (4.2) Iteratively train the autoencoder networks AE1 and AE2 to obtain the trained autoencoder network AE1;

[0076] (4.2a) Set the number of epochs for the current model training to e1 = 1, the total number of training epochs to E = 2000, and the learning rate lr = 0.0005;

[0077] (4.2b) Sample b in the background training sample set B i As the input of the autoencoder AE1, after passing through the encoder and decoder, the encoder output z is obtained i and decoder output At the same time, the sample b′ in the background training sample set B′i As the input of the autoencoder AE2, after passing through the encoder and decoder, the encoder output z′ is obtained i and decoder output

[0078] (4.2c) Calculate the loss values ​​of autoencoders AE1 and AE2 in this round of training according to the following loss function calculation formula:

[0079]

[0080]

[0081]

[0082] Among them, L r1 is the reconstruction loss value of the autoencoder AE1, L r2 is the recovery loss value when the autoencoder AE2 restores the masked spectral sample to the pre-masked spectral sample, L ziz is the similarity loss between the encoder outputs in the two autoencoders AE1 and AE2, Indicates the two-norm operation;

[0083] (4.2d) Using the Adam optimizer and back-propagation algorithm, update the weights of the neuron nodes in the autoencoder AE1 and the autoencoder AE2 so that the total network loss L = L r1 +L r2 +L ziz Continuously decrease and approach 0;

[0084] (4.2e) Determine whether the current number of training rounds e1 is equal to the total number of training rounds E:

[0085] If yes, stop training, get the trained autoencoder AE1, and save it;

[0086] Otherwise, let e1=e1+1 and return to (4.2b).

[0087] Step 5: Input the hyperspectral image to be detected X into the trained autoencoder AE1 to obtain the reconstructed hyperspectral image output X′, and calculate the pixel x at each corresponding position in X and X′. i and x′ i The spectral angle SAM(si) between , M×N spectral angle values ​​are obtained, where:

[0088]

[0089] Among them, cos -1 () represents the inverse cosine operation, T represents the transpose operation, and these M×N spectral angle values ​​are used as the preliminary detection results.

[0090] Step 6: Use the mean square nonlinear background filter to filter the preliminary detection results. Perform background suppression.

[0091] Construct a mean square nonlinear background filter f(x), which is implemented as follows:

[0092]

[0093] Among them, λ is the adjustable parameter of the mean square nonlinear background filter, and its value is selected according to the characteristics of the data set and engineering experience. In this example, λ = 10, and the preliminary detection results obtained in step 5 are The preliminary detection results of each pixel in are used as the input of f(x) to obtain the final detection results

[0094] The effects of the present invention will be further described below in conjunction with simulation experiments.

[0095] 1. Simulation conditions:

[0096] The simulation experiments of the present invention were conducted in a hardware environment with an Intel(R) Core(TM) i7-11800H CPU with a main frequency of 2.3 GHz, a memory of 16 GB, and a software environment including MATLAB;

[0097] The simulation experiment of the present invention uses a real hyperspectral image collected by the airborne visible light / infrared imaging spectrometer AVIRIS as the object of target detection. This image has 204 spectral bands and a size of 100 pixels × 100 pixels. There are 67 target pixels to be detected. The pseudo-color image of the real hyperspectral image is as follows: Figure 2 As shown in (a), the true value label image attached to the real hyperspectral image is as follows Figure 2 As shown in (b), Figure 2 The white marks in (b) are the real target pixels to be detected;

[0098] In order to effectively evaluate the performance of the proposed method, the simulation experiment of the present invention uses the AUC value of the detection result graph as the detection performance evaluation index of the hyperspectral image detection task. The AUC value is obtained by integrating the receiver operating characteristic curve ROC along the horizontal axis, that is, the area between the ROC curve and the horizontal axis. The ROC curve describes the false alarm probability P of different thresholds in the detection result graph. f and the corresponding detection probability P d The false alarm probability represents the probability of misdetecting a background sample as a target sample, and the detection probability represents the probability of successfully detecting a target sample.

[0099] 2. Simulation content and result analysis:

[0100] Simulation 1: Under the above simulation conditions, the existing hyperspectral target detection background learning BLTSC detection method based on target suppression constraint is used to detect the real hyperspectral image used in the simulation experiment. The detection result graph and the corresponding ROC curve are Figure 3 (a) and Figure 3 (b), where Figure 3 The white highlights in (a) represent the pixels identified as targets. Figure 3 The area under the curve of (b) is the AUC value of the detection result.

[0101] Simulation 2: Under the above simulation conditions, the method of the present invention is used to detect the real hyperspectral image used in the simulation experiment, and the detection result graph and the corresponding ROC curve are respectively Figure 4 (a) and Figure 4 (b), where Figure 4 The white highlights in (a) represent the pixels identified as targets. Figure 4 The area under the curve of (b) is the AUC value of the test result.

[0102] Depend on Figure 3 (a) and Figure 4 (a) By comparing the detection results, it can be found that the method of the present invention is more effective than existing background learning-based methods in target detection and background suppression in real hyperspectral images. This is because the present invention does not rely on the background assumption from the Gaussian distribution during background learning, thus avoiding the reconstruction of abnormal background in the detection stage and improving detection accuracy. The mean square nonlinear background filter increases the target-background separation.

[0103] Calculate separately Figure 3 (b) and Figure 4 (b) The area under the curve is used to obtain the target detection accuracy AUC of the prior art method in simulation 1 and the target detection accuracy AUC2 of the method of the present invention in simulation 2, as shown in Table 1.

[0104] Table 1 Comparison of target detection accuracy of the method of the present invention and the prior art method for real hyperspectral images

[0105]

[0106] As can be seen from Table 1, the AUC value of the method of the present invention is significantly improved compared with the existing method, indicating that the method of the present invention has stronger target detection ability and can achieve higher target detection accuracy.

[0107] In summary, the method of the present invention solves the problem of abnormal background reconstruction caused by the existing hyperspectral image target detection algorithm's reliance on adversarial learning by collaboratively mining the features of the original background samples and the masked background spectral samples: the background reconstruction framework proposed in the present invention can achieve more accurate modeling and reconstruction of the hyperspectral background distribution, and the target detection result can be obtained by calculating the reconstruction error of the input image, thereby achieving high-precision target detection.

Claims

1. A hyperspectral image target detection method based on spectral mask driven background learning, characterized in that: The steps include: (1) A size of Hyperspectral image to be detected And the real spectrum vector of the target to be detected ,in and Represents the width and height of the image respectively, Represents the number of bands of the image, and are all positive integers greater than 0, is a positive integer greater than or equal to 100; (2) Hyperspectral images Perform background sample screening, from Select from pixels High confidence background pixels are used as training samples ; (3) Training samples Each band of the spectrum of each pixel in the neighborhood spectrum is masked by the mean value of the spectrum, where each band contained in the pixel is masked by the mean value of the spectrum. Replace it with the mean of the spectral bands in its three neighborhoods, and the calculation method is as follows: ,in express The result after the neighborhood spectral mean mask is: Indicates the mean operation. Each pixel in the neighborhood spectrum is masked by the mean value mask operation to obtain the mask training sample ; (4) Construct a background learning network consisting of two autoencoder networks AE1 and AE2; (5) Use and The autoencoder networks AE1 and AE2 are jointly iteratively trained respectively, and the loss function is the minimum mean square error function to obtain the trained autoencoder network AE1; (6) The hyperspectral image to be detected Input to the trained autoencoder AE1 to obtain the reconstructed hyperspectral image output ,calculate and Each corresponding pixel in and The spectral angle between ,get spectral angle values, where: ， in, Indicates the inverse cosine operation, Represents the transpose operation, spectral angle values ​​as preliminary detection results ; (7) According to the weak background detection characteristics, a mean square nonlinear background filter is constructed to convert the preliminary detection results Each element in is input into the mean square nonlinear background filter for background suppression. The implementation is as follows: ， in, is the adjustable parameter of the mean square nonlinear background filter, Obtain preliminary test results for step (6) The preliminary detection results of each pixel in the image are obtained; the final detection results with good background suppression are finally obtained. .

2. The hyperspectral image target detection method based on spectral mask driven background learning according to claim 1 is characterized in that In step (2), select High confidence background pixels are used as training samples , using a dual-window detector and calculating the spectral angle, which is implemented as follows: (2a) In Each pixel As the center, construct two concentric square windows of different sizes, where the inner window size is , the outer window size is , take out the pixels between the two windows and record them as: , in, Indicates the number of pixels taken out, Indicates the pixel to be taken out; (2b) Calculate the pixels taken out above With the prior target spectrum vector spectral angle , calculated as follows: ； (2c) All Pixels Take out the high confidence background pixels in the pixel neighborhood and record these pixels as ; (2d) Yes Repeat (2a)-(2c) for all pixels in the image to obtain the neighborhood background pixels of all pixels. ; (2e) For the above obtained Use the deduplication function , remove The repeated background pixels in the image are used to obtain the non-repeated background pixel training sample set .

3. The hyperspectral image target detection method based on spectral mask driven background learning according to claim 1 is characterized in that In the step (4), a background learning network consisting of two autoencoder networks AE1 and AE2 is constructed. AE1 and AE2 have the same structure and are both composed of an encoding network and a decoding network connected thereto, wherein the decoder network of AE1 and the decoder network of AE2 share weights with each other; the encoder network comprises an input layer, a fully connected layer, and an output layer. The number of nodes in the input layer is equal to the number of bands of the hyperspectral image to be measured. The number of nodes in the fully connected layer is 256, and the number of nodes in the output layer is 128. The decoder network consists of an input layer, a fully connected layer, and an output layer. Its structure is symmetrical to that of the encoder. The dimension of the input layer is 128, the dimension of the fully connected layer is 256, and the dimension of the output layer is .

4. The hyperspectral image target detection method based on spectral mask driven background learning according to claim 1 is characterized in that In step (5), use and They are used as inputs of the autoencoder networks AE1 and AE2 for simultaneous iterative training, which is implemented as follows: (5a) Set the number of rounds of current model training to The total number of training rounds is , and set ; (5b) The background training sample set Samples in As the input of the autoencoder AE1, after passing through the encoder and decoder, the encoder output is obtained and decoder output ; At the same time, the background training sample set Samples in As the input of the autoencoder AE2, after passing through the encoder and decoder, the encoder output is obtained and decoder output ; (5c) Calculate the loss values ​​of autoencoders AE1 and AE2 in this round of training according to the following loss function calculation formula: ， ， ， in, is the reconstruction loss value of the autoencoder AE1, It is the recovery loss value when the autoencoder AE2 restores the masked spectral sample to the pre-masked spectral sample. is the similarity loss between the encoder outputs in the two autoencoders AE1 and AE2, Indicates the two-norm operation; (5d) Using the Adam optimizer and back-propagation algorithm, update the weights of the neuron nodes in the autoencoder AE1 and the autoencoder AE2 so that the total network loss Continuously decrease and approach 0; (5e) Determine the current number of training rounds Is it equal to the total number of training rounds? : If yes, stop training, get the trained autoencoder AE1, and save it; Otherwise, let , return to (5b).

Images