A hyperspectral target detection method and device based on target-driven autoencoder

Abstract

The application discloses a hyperspectral target detection method and device based on a target-driven autoencoder. The method comprises the following steps: 1) setting the hidden layer dimension of the target-driven autoencoder; 2) designing a loss function of the target-driven autoencoder; 3) training the target-driven autoencoder to obtain a reconstructed hyperspectral image; 4) calculating a CEM detection result of reconstructed data output by the target-driven autoencoder and prior knowledge spectrum, and performing dimension expansion on the detection result; 5) calculating a TI value based on the CEM detection result obtained in the step 4), and performing iteration to finally obtain a hyperspectral target detection result T, and realize the hyperspectral target detection. The hyperspectral target detection model based on the target-driven autoencoder designed in the application can effectively eliminate noise and interference, suppress a background, enhance a target and improve the performance of the hyperspectral target detection.

Description

Technical Field

[0001] This invention relates to the field of hyperspectral image processing, and in particular to a hyperspectral target detection method and apparatus based on a target-driven autoencoder. Background Technology

[0002] Hyperspectral target detection (HTD) is an important aspect of hyperspectral image (HSI) processing, aiming to distinguish targets from background (BKG) using fully or partially known prior knowledge. Traditional HTD methods typically rely on distance or similarity measurements and statistical techniques to identify discriminative features, such as matched filters (MF) and constrained energy minimization (CEM). These methods are known for their simplicity and intuitiveness, but are often limited in detection accuracy. Furthermore, component representation-based methods have become a hot topic in HTD research. For example, the sparse representation-based binary hypothesis (SRBBH) model and the dual-dictionary nonlinear representation model (DDNRTD) for hyperspectral subpixel target detection have improved detection capabilities and gained popularity, but are sensitive to inter-class similarity and intra-class variability.

[0003] Recently emerging machine learning and deep learning algorithms have become a research hotspot in HTD. For example, 3-D macro-micro convolutional residual autoencoders (3-D-MMRAE) are known for their remarkable feature extraction capabilities, but they are usually complex in structure, have low time efficiency, and are sensitive to hyperparameters.

[0004] Autoencoders (AEs), as an unsupervised model in deep learning, excel in feature generation and dimensionality reduction, and possess significant development potential. In HSI (High-Speed ​​Interaction), AEs are commonly used for tasks such as denoising, anomaly detection, and dimensionality reduction. In HTD (High-Level Distributed ... Summary of the Invention

[0005] To address the shortcomings of existing technologies, the present invention aims to provide a hyperspectral target detection method and apparatus based on a target-driven autoencoder. This method fully utilizes the characteristics of the target itself to drive the autoencoder to improve data quality, and uses an iterative structure to further enhance the target and suppress other uninteresting signals. This method can efficiently achieve hyperspectral image target detection, and the target detection capability, background suppression capability, and robustness to noise of this algorithm have been effectively verified.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] This invention provides a hyperspectral target detection method based on a target-driven autoencoder, comprising the following steps:

[0008] Step 1): Decompose the hyperspectral image into a linear combination of four components: target, background, interference, and noise; use the NWHFC method to estimate the sum of the virtual dimensions of the target and background in the hyperspectral image, and use the sum of the virtual dimensions of the target and background to set the hidden layer dimension of the target-driven autoencoder.

[0009] Step 2): Calculate the truncated KL divergence based on the hyperspectral image X and the prior spectrum, and design the loss function of the target-driven autoencoder in combination with the root mean square error;

[0010] Step 3): Construct and train the target-driven autoencoder based on the hidden layer dimension designed in Step 1) and the loss function designed in Step 2); input the hyperspectral image X into the trained target-driven autoencoder to obtain the reconstructed hyperspectral image.

[0011] Step 4): Based on the reconstructed hyperspectral image The CEM detection results are calculated using prior spectra, and the CEM detection results are dimensionally expanded to obtain a matrix. The obtained matrix is ​​then weighted and summed with the hyperspectral image X to obtain a new hyperspectral image X'.

[0012] Step 5): Calculate the TI value based on the CEM detection results obtained in Step 4). If the TI value is less than the set threshold, replace the hyperspectral image X with X' and repeat the operation of Steps 2) to 5). If the TI value is greater than the set threshold, the iteration ends. The CEM detection result obtained in the last round is the final target detection result.

[0013] The present invention also provides a hyperspectral target detection device based on a target-driven autoencoder, comprising:

[0014] The hidden layer dimension design module is used to decompose a hyperspectral image into a linear combination of four components: target, background, interference, and noise. It also estimates the sum of the virtual dimensions of the target and background in the hyperspectral image and uses the sum of the virtual dimensions to set the hidden layer dimension of the target-driven autoencoder.

[0015] The loss function design module is used to calculate the truncated KL divergence of the hyperspectral image and the prior knowledge spectrum, and to design the loss function of the target-driven autoencoder in combination with the root mean square error.

[0016] The target-driven autoencoder training module is used to construct a target-driven autoencoder based on the designed hidden layer dimensions and train it using a loss function.

[0017] The hyperspectral image reconstruction module is used to input the hyperspectral image X into the trained target-driven autoencoder to obtain the reconstructed hyperspectral image.

[0018] The CEM detection module is used to calculate the CEM detection results based on the reconstructed hyperspectral image and the prior spectrum.

[0019] The hyperspectral image update module is used to expand the dimensions of the CEM detection results to obtain a matrix, and then weighted summation of the obtained matrix with the hyperspectral image X to obtain a new hyperspectral image X'.

[0020] The iteration stop determination module is used to determine whether the iteration stop condition is met.

[0021] The target detection result output module is used to output the final hyperspectral target detection result image.

[0022] Compared with the prior art, the present invention has the following beneficial effects:

[0023] 1) The target-driven autoencoder designed in this invention can effectively achieve adaptive quality improvement of hyperspectral images. By adaptively determining the intermediate layer dimension of the target-driven autoencoder, the network is driven to learn the salient features of the target and the background in a biased manner, thereby achieving the suppression and elimination of noise and interference.

[0024] 2) This invention proposes a truncated KL divergence, which optimizes the distribution of the original KL divergence, making it more suitable for the field of hyperspectral target detection. The loss function of the target-driven autoencoder is designed using the truncated KL divergence. The loss function is the root mean square error between the input data and the reconstructed data weighted by the truncated KL divergence, which drives the network to further reconstruct the target rather than the background. Combined with the constraints on the intermediate layer, the emphasis on the target component and the suppression of other uninteresting components are achieved.

[0025] 3) This invention utilizes the results of target-driven autoencoder reconstruction for coarse detection, and uses the results of coarse detection to further improve the quality of the image, thereby iteratively improving the accuracy of target detection results layer by layer. Attached Figure Description

[0026] Figure 1 This is a flowchart illustrating the hyperspectral target detection based on a target-driven autoencoder according to the present invention;

[0027] Figure 2 This is a schematic diagram of the hyperspectral target detection device based on a target-driven autoencoder according to the present invention;

[0028] Figure 3 This is a schematic diagram of a hyperspectral image;

[0029] Figure 4This is a location diagram of the target;

[0030] Figure 5 This is a target detection result image obtained from a hyperspectral image using the method proposed in this invention.

[0031] Figure 6 This is a comparison chart of target detection results obtained from hyperspectral images using the present invention, CEM, E-CEM, DM-BDL, BA-OSP, BLTSC, and DSC.

[0032] Figure 7 The target detection result image obtained by this invention is obtained after adding noise with SNR=5dB and SNR=10dB to the hyperspectral image. Detailed Implementation

[0033] To make the objectives, technical solutions, and advantages of this invention clearer, the invention is described in detail below with reference to specific embodiments. Specific embodiments are described below to simplify the invention. However, it should be understood that the invention is not limited to the described embodiments, and various modifications are possible without departing from the basic principles; these equivalent forms also fall within the scope defined by the appended claims.

[0034] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0035] like Figure 1 The diagram shown is a flowchart illustrating the basic steps of the hyperspectral target detection method and apparatus based on a target-driven autoencoder according to this invention in this embodiment, mainly including:

[0036] Step 1: Decompose the hyperspectral image into a linear combination of four components: target, background, interference, and noise; use the NWHFC method to estimate the sum of the virtual dimensions of the target and background in the hyperspectral image, and use the sum of the virtual dimensions of the target and background to set the hidden layer dimension of the target-driven autoencoder.

[0037] Specifically, hyperspectral image data is decomposed into a linear combination of target, background, interference, and noise, represented as:

[0038] r=Sα+Lβ+Iγ+N (1)

[0039] in, This represents the spectral vector of a pixel in a hyperspectral image. Represents the target spectral matrix. Represents the background spectral matrix. The spectral matrix representing the interference. Let L represent the noise matrix, and let n represent the spectral dimension of the hyperspectral data. t n represents the target dimension of interest in hyperspectral data. b n represents the background dimension of the hyperspectral data. i Indicates the interference dimension of hyperspectral data. These represent the abundance of the target, background, and interference, respectively.

[0040] The purpose of decomposing hyperspectral data into a linear combination of four components is to leverage the statistical properties of different components in hyperspectral data to enhance the ability of autoencoders in unsupervised deep learning methods to perceive salient features.

[0041] The sum of the dimensions of the target and background components in the hyperspectral data is estimated using the NWHFC method, thereby determining the hidden layer dimension of the target-driven autoencoder.

[0042] The NWHFC method includes noise whitening and virtual dimension (VD) estimation:

[0043] A) Noise whitening:

[0044] First, the covariance matrix of the noise is calculated, and then the original data is whitened. For a spectral matrix... Calculate the cross-correlation matrix R and the covariance matrix K, where L represents the number of spectral bands and N represents the total number of pixels in the image.

[0045]

[0046]

[0047] in, It is the average spectrum of the hyperspectral image data, calculated using the following formula: X(:,i) represents the i-th column vector of the input spectral matrix; the covariance matrix K is a diagonal matrix; the inverse of the diagonal matrix of the inverse of the covariance matrix K is the inverse of the diagonal matrix of the covariance matrix K. n To estimate noise characteristics:

[0048] K n =(diag(K) -1 )) -1 (4)

[0049] Where, diag(·) represents the diagonal matrix operator, (·) -1 Represents the inverse matrix operator;

[0050] The data is whitened using the estimated noise covariance matrix:

[0051]

[0052] Among them, X w This represents hyperspectral image data after whitening.

[0053] B) Virtual Dimension (VD) estimation:

[0054] Virtual dimension estimation is performed using whitened hyperspectral image data. First, the correlation matrix R of the noisy whitened data is calculated according to formulas (2) and (3). w Covariance matrix K w And obtain the two sets of eigenvalues ​​corresponding to the two matrices, namely {γ1≥γ2≥…≥γ L} and {ν1≥ν2≥...≥ν L};

[0055] Calculate R w and K w The difference between the sets of eigenvalues ​​yields the difference matrix D:

[0056]

[0057] Based on the theory of multiple regression analysis, the threshold τ of band l is calculated by setting the false alarm rate ρ. l :

[0058]

[0059] Where erfinv(·) is the inverse function of the error function, used to calculate the cumulative distribution function of the normal distribution;

[0060] The virtual dimension (VD) is estimated as follows:

[0061]

[0062] Where p is the sum of the virtual dimensions of the target and background in the hyperspectral image, [D i,j >0] indicates that when D i,j When >0, the value is 1, otherwise it is 0; set the dimension of the intermediate layer of the target-driven autoencoder to the p value estimated in (8).

[0063] Step 2: Calculate the truncated KL divergence between the hyperspectral image and the prior knowledge spectrum, and design the loss function of the target-driven autoencoder in combination with the root mean square error;

[0064] Specifically, firstly, the spectral vectors and prior spectra of each pixel in the hyperspectral image are normalized and preprocessed to ensure that the spectral vector of each pixel conforms to the property that the sum of the probability distributions is 1:

[0065]

[0066] Where i = 1, 2, ..., L, j = 1, 2, ..., N, p i,j q represents the i-th element of the j-th spectral vector after preprocessing. i Let e ​​represent the i-th element of the preprocessed prior spectrum. (·) It represents the exponent with base e;

[0067] The KL divergence D of the j-th column vector of the spectral matrix X KL The formula for calculating (j) is:

[0068]

[0069] Among them, D KL (j) represents the distribution distance vector The smaller the value, the more similar the pixel is to the prior spectrum, and the higher the probability that it is the target. To make the value proportional to the similarity, the value is negative and the similarity vector is normalized. Outliers are identified using truncation techniques. Data quality is evaluated using the standard deviation and the mean. Data points falling below three times the standard deviation of the mean are identified as outliers and truncated to the margin.

[0070]

[0071] Set values ​​greater than μ-3σ to a specific value μ-3σ:

[0072]

[0073] Normalize the adjusted vector:

[0074]

[0075] Among them, s j It is the j-th element in the truncated vector S of length N. To guide the target-driven autoencoder to prioritize the reconstruction of pixels more similar to the target, the loss function is designed as the Hadamard product of MSE and truncated KL divergence K, where truncated KL divergence K is:

[0076] K = U × S (14)

[0077] in, K is an L×N weighting matrix, representing the weighted hyperspectral image X and the reconstructed hyperspectral image. The MSE between the two values ​​is used as the loss function, and its calculation formula is as follows:

[0078]

[0079] Among them, "⊙" represents the Hadamarda accumulation.

[0080] Step 3: Train the target-driven autoencoder based on the structure designed in Step 1 and the loss function designed in Step 2 to obtain the reconstructed hyperspectral image;

[0081] A target-driven autoencoder is constructed, comprising an encoder, intermediate layers, and a decoder. The encoder consists of a fully connected layer and a sigmoid activation function. The dimension of the intermediate layer is the same as the hidden layer dimension calculated in step 1). The decoder consists of a fully connected layer and a sigmoid activation function. The learning rate, batch size, and number of training epochs are set in the constructed target-driven autoencoder, and the autoencoder is trained based on the loss function obtained in step 2) until the maximum number of training epochs is reached. The hyperspectral image X is input into the trained target-driven autoencoder to obtain the reconstructed hyperspectral image.

[0082] Step 4: Based on the reconstructed hyperspectral image The CEM detection results are calculated using prior spectra, and the CEM detection results are dimensionally expanded to obtain a matrix. The obtained matrix is ​​then weighted and summed with the input of the target-driven autoencoder, and the new matrix is ​​used as the input matrix of the new target-driven autoencoder.

[0083] Specifically, in order to further improve the diversity between the target and the background and suppress noise and interference, and to make full use of prior and posterior information, an iterative framework is proposed based on the target-driven autoencoder to improve detection performance layer by layer through iteration.

[0084] Hyperspectral target detection using a CEM detector, utilizing the reconstructed hyperspectral image The target detection result calculated from the prior spectrum d is as follows:

[0085]

[0086] Where C is The autocorrelation matrix, To align the object detection results with the input size of the object-driven autoencoder for iterative data quality enhancement, dimensional expansion is performed, resulting in the following matrix:

[0087]

[0088] Thus, the target detection result image is obtained using the hyperspectral image X and the prior spectrum iteration; for the next iteration, the hyperspectral image X is updated to X', where X' is the weighted sum of X and M from the previous round:

[0089] X=X+λM (18)

[0090] The trade-off parameter is λ.

[0091] Step 5: Using the updated X', calculate the Tanimoto index (TI). If the TI value is less than the set threshold, replace the hyperspectral image X with X' and repeat steps 2) to 5) for data reconstruction and target detection. If the TI value is greater than the set threshold, the iteration ends. The TI value for the nth iteration is calculated as follows:

[0092]

[0093] Where the superscript of M represents the number of iterations; the final CEM detection result obtained after stopping the iteration is the final target detection result.

[0094] Corresponding to the aforementioned embodiment of a hyperspectral target detection method based on an iterative autoencoder, the present invention also provides an embodiment of a hyperspectral target detection device based on an iterative autoencoder.

[0095] Figure 2 This is a block diagram illustrating a hyperspectral target detection device according to an exemplary embodiment, such as... Figure 2 As shown, the device includes: a hidden layer dimension design module for decomposing a hyperspectral image into a linear combination of four components: target, background, interference, and noise, estimating the sum of virtual dimensions of the target and background in the hyperspectral image, and using the sum of virtual dimensions to set the hidden layer dimension of the target-driven autoencoder in the iterative autoencoder; a loss function design module for calculating the truncated KL divergence between the hyperspectral image and the prior knowledge spectrum, and designing the loss function of the target-driven autoencoder in combination with the root mean square error; a target-driven autoencoder training module for constructing the target-driven autoencoder according to the designed hidden layer dimensions and training it through the loss function; and a hyperspectral image X input into the trained target-driven autoencoder to obtain the reconstructed hyperspectral image. The system includes: a hyperspectral image reconstruction module; a CEM detection module for calculating CEM detection results based on the reconstructed hyperspectral image and prior spectra; a hyperspectral image update module for expanding the dimensions of the CEM detection results to obtain a matrix, and then weighted summing the obtained matrix with the hyperspectral image X to obtain a new hyperspectral image X'; an iteration stop determination module for determining whether the iteration stop condition is met; and a target detection result output module for outputting the final hyperspectral target detection result map.

[0096] Regarding the apparatus in the above embodiments, the specific manner in which each module performs its operation has been described in detail in the embodiments related to the method, and will not be elaborated upon here.

[0097] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to the description of the method embodiments. The device embodiments described above are merely illustrative; the various modules in the device represent a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules may be combined or integrated into another unit. Furthermore, the connections between the displayed or discussed modules may be communication connections through some interfaces, which may be electrical or other forms. Some or all of the modules can be selected to achieve the purpose of this application according to actual needs. Those skilled in the art can understand and implement this without creative effort. The following uses publicly available real hyperspectral target detection data as an example to illustrate specific implementation methods, including hyperspectral image data and prior spectral information, to demonstrate the technical effects of the present invention. Specific steps in the embodiments will not be repeated.

[0098] Example 1

[0099] In this embodiment, the effectiveness of a hyperspectral target detection method based on an iterative autoencoder will be verified using publicly available hyperspectral image data. Qualitatively, the method is evaluated using a visual result image of the target detection from the hyperspectral image data and its corresponding 3D visualization grayscale image. Quantitatively, the algorithm's detection performance is evaluated using the 3D-ROC evaluation index system. The 3D-ROC evaluation index system includes target detection capability (TD), background suppression capability (BS), and a comprehensive effectiveness evaluation index for the detector. The specific indices are described in the table below:

[0100] Table 1 3D-ROC Evaluation Index System

[0101]

[0102]

[0103] Given a publicly available, real hyperspectral image (80×100×162 pixels), depicting a complex urban road background with bright buildings and significant noise clutter, the target is a man-made vehicle. Figure 3 The image shown is a schematic diagram of the hyperspectral image in this embodiment. Figure 4 The image shown is a schematic diagram of the target's location in the hyperspectral image of this embodiment. Figure 5As shown, this is the target detection result image corresponding to the hyperspectral image in this embodiment. The detection result image shows that background, noise, and interference are completely suppressed, and the target is significantly enhanced. To quantitatively evaluate the effectiveness of the proposed algorithm, Table 2 presents the 3D-ROC evaluation system metrics for hyperspectral target detection results using CEM, E-CEM, DM-BDL, BA-OSP, BLTSC, DSC, and the proposed method, where AUC... (D,F) The optimal AUC value is 1.0000; the suboptimal AUC value is 0.9999; AUC (D,τ) The optimal AUC value is 1.0000; the suboptimal AUC value is 0.9153; AUC (F,τ) The optimal AUC value is 0.0002; the suboptimal AUC value is 0.0008; AUC TD The optimal AUC value is 2.0000; the suboptimal AUC value is 1.9153; AUC BS The optimal AUC value is 0.9998; the suboptimal AUC value is 0.9992; AUC SNPR The optimal AUC value is 0.9998; the second-best AUC value is 0.8364; AUC TDBS The optimal AUC value is 4581.9694; the second-best AUC value is 619.6162; AUC ODP The optimal AUC value is 1.9998; the suboptimal AUC value is 1.8364.

[0104] Table 2

[0105] method <![CDATA[AUC (D,F) ]]> <![CDATA[AUC (D,τ) ]]> <![CDATA[AUC (F,τ) ]]> <![CDATA[AUC TD ]]> <![CDATA[AUC BS ]]> <![CDATA[AUC SNPR ]]> <![CDATA[AUC TDBS ]]> <![CDATA[AUC ODP ]]> This invention 1.0000 1.0000 0.0002 2.0000 0.9998 0.9998 4581.9694 1.9998 CEM 1.0000 0.6663 0.1093 1.6662 0.8907 0.5570 6.0950 1.5569 E-CEM 1.0000 <![CDATA[ 0.9153 ]]> 0.0789 <![CDATA[ 1.9153 ]]> 0.9211 <![CDATA[ 0.8364 ]]> 11.6001 <![CDATA[ 1.8364 ]]> DM-BDL 0.9996 0.6426 0.0252 1.6422 0.9744 0.6174 25.5239 1.6170 BA-OSP 0.9986 0.7516 0.2061 1.7503 0.7925 0.5455 3.6466 1.5442 BLSTC <![CDATA[ 0.9999 ]]> 0.4674 <![CDATA[ 0.0008 ]]> 1.4673 <![CDATA[ 0.9992 ]]> 0.4666 <![CDATA[ 619.6162 ]]> 1.4666 DSC 0.9998 0.5132 0.0025 1.5130 0.9973 0.5107 206.0357 1.5105

[0106] like Figure 6 The image shown is a comparison of target detection results obtained from hyperspectral images using the present invention, CEM, E-CEM, DM-BDL, BA-OSP, BLTSC, and DSC. Figure 7 The images show the target detection results obtained by applying this invention to hyperspectral images with SNR=5dB and SNR=10dB, respectively. Qualitative results show that some of the selected contrast algorithms achieve good background suppression, but the target is not salient enough, such as BLTSC. Some contrast algorithms can detect the target, but the background suppression effect is poor, such as BA-OSP. The proposed detection algorithm, however, can completely suppress the background and enhance target saliency. The quantitative results shown in Table 2 also reflect that the background suppression and target detection capabilities of the method proposed in this invention are superior to all contrast methods. Specifically, the AUC of the method proposed in this invention... (F,τ) Reaching 0.0002, AUC (D,τ) Reaching 1, AUC SNPRThe result reached 4581.9694, indicating excellent background suppression capability. Based on both qualitative and quantitative analysis, the hyperspectral target detection method proposed in this invention possesses superior target detection capability, background suppression capability, and overall effectiveness.

[0107] The accompanying drawings illustrating the embodiments of the present invention will make the objectives, technical solutions, and advantages of the present invention clearer. It should be noted that the specific embodiments described herein are merely illustrative and not intended to limit the scope of the invention. All equivalent substitutions and improvements made within the framework of the methods and principles provided by the present invention should be included within the protection scope of the present invention.

Claims

1. A method for hyperspectral target detection based on target-driven autoencoder, characterized in that, Comprising the following steps: Step 1): decompose the hyperspectral image into a linear combination of target, background, interference and noise four components; estimate the sum of virtual dimensions of target and background of the hyperspectral image by using the NWHFC method, and set the hidden layer dimension of the target-driven autoencoder by using the sum of virtual dimensions of target and background; the NWHFC method comprises noise whitening and virtual dimension VD estimation; Step 2): calculate the truncated KL divergence based on the hyperspectral image X and the prior spectrum, and design the loss function of the target-driven autoencoder in combination with the root mean square error; The KL divergence D of the jth column vector of hyperspectral image X KL The calculation formula of (j) is: where D KL (j) represents the distance vector of distribution The value of D KL (j) is smaller, the more similar the pixel is to the prior spectrum, and the higher the likelihood that it is the target. To make the size of D KL (j) proportional to the similarity, D KL (j) is taken as a negative value and normalized. Outliers are identified using a truncation technique, and the data quality is assessed using the standard deviation and mean. Data points falling below the mean minus three times the standard deviation are identified as outliers and truncated to the edge value: Set the value greater than μ-3σ to a specific value μ-3σ: Normalize the adjusted vector: where s j is the jth element in the truncated vector S of length N; to guide the target-driven autoencoder to preferentially reconstruct pixels that are more similar to the target, the loss function is designed as the Hadamard product of MSE and the truncated KL divergence K, where the truncated KL divergence K is: K=U×S (14) wherein, K is a L x N weighted matrix, weighted hyperspectral image X and reconstructed hyperspectral image MSE between them as loss function, its calculation formula is: Where "⊙" is Hadamard product; Step 3): Construct and train the target-driven autoencoder according to the hidden layer dimension designed in step 1) and the loss function designed in step 2); input the hyperspectral image X into the trained target-driven autoencoder to obtain the reconstructed hyperspectral image Step 4): based on the reconstructed hyperspectral image The CEM detection result is calculated based on the prior spectrum, and the CEM detection result is dimensionally expanded to obtain a matrix. The obtained matrix is weighted and summed with the hyperspectral image X to obtain a new hyperspectral image X'. Step 5): obtain the TI value based on the CEM detection result obtained in step 4), if the TI value is less than the set threshold, replace the hyperspectral image X with X', repeat the operations of steps 2)-5), if the TI value is greater than the set threshold, the iteration is ended, and the CEM detection result obtained in the last round is the final target detection result.

2. The target-driven autoencoder based hyperspectral target detection method according to claim 1, wherein, The step 1) is specifically: Decompose the hyperspectral image data into a linear combination of target, background, interference and noise, represented as r=Sα+Lβ+Iγ+N (1) wherein, represents a spectral vector of a pixel in a hyperspectral image, represents a target spectral matrix, represents a background spectral matrix, represents a spectral matrix of an interference, represents a noise matrix, L represents a spectral dimension of the hyperspectral data, n t represents a target dimension of interest of the hyperspectral data, n b represents a background dimension of the hyperspectral data, n i represents an interference dimension of the hyperspectral data, respectively represent the abundance of the target, background and interference. Estimate the sum of dimensions of target and background components in the hyperspectral image according to the NWHFC method, so as to determine the hidden layer dimension of the target-driven autoencoder; The NWHFC method comprises noise whitening and virtual dimension VD estimation: A) Noise whitening: For a spectral matrix The cross-correlation matrix R and the covariance matrix K are calculated, where L represents the number of spectral bands, and N represents the total number of pixels of the image, wherein is the average spectrum of the hyperspectral image data, calculated as X(:,i) denotes the i-th column vector of the input spectral matrix; the covariance matrix K is a diagonal matrix; the inverse of the diagonal of the inverse of the covariance matrix K n estimates the noise characteristics: K n = (diag(K -1 )) -1 (4) where diag(·) denotes the diagonal matrix operator, (·) -1 denotes the inverse matrix operator; Use the estimated noise covariance matrix to whiten the data: wherein X w represents hyperspectral data after whitening processing; B) Virtual dimension VD estimation: The virtual dimension estimation is performed using the hyperspectral data after the whitening processing. First, the correlation matrix R of the noise whitened data is calculated according to the formula (2) and (3) w and the covariance matrix K w , and two sets of eigenvalues corresponding to the two matrices are obtained, respectively, {γ1≥γ2≥...≥γ L} and {ν1≥ν2≥...≥ν L}. Compute R w and K w The difference between the sets of eigenvalues, resulting in a difference matrix D: Based on the theory of multiple regression analysis, the threshold τ of the wave band l is calculated by presetting the false alarm rate ρ l : Where erfinv(·) is the inverse function of the error function, used to calculate the cumulative distribution function of the normal distribution; The virtual dimension VD estimation is: where p is the sum of the virtual dimensions of the target and background of the hyperspectral image, [D i,j > 0] is 1 if D i,j > 0, otherwise 0; the dimension of the intermediate layer of the target-driven autoencoder is set to the estimated value of p in (8).

3. The target-driven autoencoder based hyperspectral target detection method according to claim 1, wherein, The step 2) before calculating the truncated KL divergence, First, normalize the spectral vector of each pixel in the hyperspectral image and the prior spectrum for pretreatment, to ensure that the spectral vector of each pixel meets the characteristic that the sum of probability distribution is 1: where i = 1, 2,..., L, j = 1, 2,..., N, p i,j denotes the i-th element of the j-th pre-processed spectral vector, q i denotes the i-th element of the pre-processed prior spectrum, e (·) denotes the exponential value of (·) with base e.

4. The target-driven autoencoder based hyperspectral target detection method of claim 1, wherein, The step 3) is specifically: Build a target-driven autoencoder, which comprises an encoder part, an intermediate layer and a decoder part, wherein the encoder part is a fully connected layer and a Sigmoid activation function, the dimension of the intermediate layer is the hidden layer dimension calculated in step 1), and the decoder part is a fully connected layer and a Sigmoid activation function; Setting learning rate, batch size and training rounds in the constructed target-driven autoencoder, and based on the loss function obtained in step 2), training the target-driven autoencoder until the maximum training rounds are reached; inputting the hyperspectral image X into the trained target-driven autoencoder to obtain the reconstructed hyperspectral image 5. The target-driven autoencoder based hyperspectral target detection method of claim 1, wherein, The step 4) is specifically: In order to further improve the diversity between target and background, and suppress noise and interference, make full use of prior information and posterior information, and propose an iterative framework based on the target-driven autoencoder, which improves the detection performance layer by layer through iteration; The CEM detector is used for hyperspectral target detection, and the reconstructed hyperspectral image is used The target detection result calculated by the prior spectrum d is: where C is autocorrelation matrix of To align the target detection results and the input size of the target-driven autoencoder for iterative data quality enhancement, dimension expansion is performed to obtain the following matrix: Thus, the target detection result image using the hyperspectral image X and the prior spectrum is obtained; for the next iteration, the hyperspectral image X is updated to X', X' is the weighted sum of X and M in the last round: X'=X+λM (18) Where the weighting parameter is λ.

6. The target-driven autoencoder based hyperspectral target detection method of claim 1, wherein, The step 5) is specifically: Using the updated X', a Tanimoto index (TI) is calculated, if the TI value is less than a set threshold, the hyperspectral image X is replaced by X', and the data reconstruction and target detection are repeated by steps 2) to 5), if the TI value is greater than the set threshold, the iteration is ended; the TI value of the nth iteration is calculated as follows: Wherein, the superscript of M is the number of iterations; the last round of CEM detection result obtained after stopping iteration is the final target detection result.

7. A hyperspectral target detection device based on target-driven autoencoder, the hyperspectral target detection device is used to implement the hyperspectral target detection method based on target-driven autoencoder of claim 1, characterized in that, The hyperspectral target detection device comprises: A hidden layer dimension design module is configured to decompose the hyperspectral image into a linear combination of four components of target, background, interference and noise, and estimate the sum of virtual dimensions of the target and the background of the hyperspectral image, and set the hidden layer dimension of the target-driven autoencoder by using the sum of virtual dimensions; A loss function design module is configured to calculate the truncated KL divergence of the hyperspectral image and the prior knowledge spectrum, and design the loss function of the target-driven autoencoder by combining the root mean square error; A target-driven autoencoder training module is configured to construct the target-driven autoencoder according to the designed hidden layer dimension and train the target-driven autoencoder by using the loss function; a hyperspectral image reconstruction module configured to input the hyperspectral image X into the trained target-driven autoencoder to obtain a reconstructed hyperspectral image A CEM detection module is configured to calculate the CEM detection result according to the reconstructed hyperspectral image and the prior spectrum; A hyperspectral image updating module is configured to perform dimension expansion on the CEM detection result to obtain a matrix, and perform weighted summation on the obtained matrix and the hyperspectral image X to obtain a new hyperspectral image X'; An iteration stop judgment module is configured to judge whether the iteration stop condition is met; A target detection result output module is configured to output the final hyperspectral target detection result image.

Images