Hyperspectral anomaly detection method based on double-line iterative convolutional neural network

Abstract

The invention discloses a hyperspectral anomaly detection method based on a double-line iterative convolutional neural network. The hyperspectral anomaly detection method comprises the following steps: S1, inputting and setting parameters; s2, abnormally enhancing a line; s3, a background reconstruction line; s4, judging an abnormal point; step S5, updating the hyperspectral data; and S6, generating a detection graph. According to the method, the background reconstruction line focuses on background reconstruction, the anomaly enhancement line is responsible for anomaly enhancement, and the background reconstruction line and the anomaly enhancement line complement and promote each other, so that more accurate anomaly detection is realized, and the problem that background reconstruction and anomaly detection are difficult to effectively combine in a traditional method is solved. According to the method, the superpixel technology is combined, and the anomaly detection strategy based on comparison with the surrounding pixel points is used, so that the local sensitivity is improved, the selection of the abnormal target is simpler, more convenient and more efficient, the detection efficiency is improved, and possible calculation redundancy and efficiency bottleneck are avoided.

Description

Technical Field

[0001] The present invention relates to the technical field of hyperspectral anomaly detection, and more specifically, to a hyperspectral anomaly detection method based on a dual-line iterative convolutional neural network. Background Art

[0002] In recent decades, researchers have proposed many effective hyperspectral anomaly detection algorithms, which can be divided into statistical metric-based, representation model-based and deep learning-based algorithms.

[0003] A. Anomaly Detection Based on Statistical Metrics

[0004] One of the most famous anomaly detection methods is the Reed-Xiaoli (RX) detector, which was proposed by Reed and Yu as a distance-based anomaly detection method and is widely used in anomaly detection tasks in hyperspectral images (HSI). The RX algorithm is based on the generalized likelihood ratio test, models the background as a multivariate Gaussian distribution, and uses the mean vector and covariance matrix of the background pixels to describe the statistical characteristics of the background. The degree of abnormality is determined by calculating the Mahalanobis distance between the pixel to be detected and the background mean. The RX detector is divided into global RX and local RX. The global RX detector assumes that the background in the data is statistically uniform and uses the entire data set to estimate the mean and covariance matrix of the background to construct a global background model; while the local RX is based on the local background information, selects the neighborhood window around the target pixel, estimates the mean and covariance matrix of the area as the background model, and can detect local anomalies more sensitively. Based on this idea of ​​the RX algorithm, subsequent studies have proposed some improved methods, such as kernel RX (KRX), which introduces kernel functions to process nonlinear data, projects the original image into a nonlinear high-dimensional feature space, and considers the high-order characteristics of different feature bands. The anomaly detection method based on the RX algorithm is limited by the Gaussian distribution assumption and the high computational complexity of the inversion of the high-dimensional covariance matrix. To solve the above problems, Schweizer et al. proposed an adaptive detection algorithm based on three-dimensional Gaussian Markov random fields, which uses spatial-spectral joint random fields to characterize the background statistical characteristics and combines spatial and spectral information to alleviate the computational cost of covariance inversion; Banerjee proposed a method based on support vector data description (SVDD) to detect target samples through nonlinear projection.

[0005] B. Anomaly Detection Based on Representation Model

[0006] The abnormal target detection method based on representation model does not rely on the probability distribution assumption of the background. Instead, it designs algorithm optimization criteria by analyzing the characteristics of local spectral distribution or applying multiple knowledge constraints, using a set of spectral vectors to describe the background and abnormal targets, and judging the abnormal degree of the sample by the residual value. Li et al. proposed a detector based on collaborative representation (CRD), which considers the information of neighborhood pixels to represent the central pixel, improving the accuracy and robustness of detection. In addition, in the abundance and dictionary based low-rank decomposition (ADLR) method, it is assumed that the unmixed background presents low-rank characteristics in the abundance subspace. Xu et al. proposed a low-rank and sparse representation (LRASR) method, which uses low-rank and sparse representation methods to model the background as a low-rank matrix and the abnormal target as sparse noise, enhancing the ability to separate abnormal points. At the same time, since the background can be approximately represented by a set of base background signals, low-rank priors are used for background estimation in hyperspectral anomaly detection. Low-rank sparse matrix decomposition (LRaSMD) uses Mahalanobis distance method (LSMAD) to decompose the data in hyperspectral images into low-rank background, sparse anomalies and noise parts. However, most methods only consider the spectral characteristics of background and anomaly components, but ignore their spatial distribution characteristics.

[0007] C. Anomaly Detection Based on Deep Learning

[0008] With the rapid development of deep learning technology, hyperspectral anomaly detection methods based on deep learning have gradually become a research hotspot. Some improved methods have begun to integrate multiple deep learning models to improve the accuracy and robustness of detection. For example, based on the characteristic that graph regularization can retain local spatial features in images, Fan et al. proposed a robust graph autoencoder (RGAE), adding graph regularization to AE to improve the robustness of the network to abnormal targets and noise. Wang et al. proposed a method based on a fully convolutional autoencoder, which uses the autoencoder to perform unsupervised learning on hyperspectral images to obtain anomaly probability maps, and then uses an adaptive weighted loss function to guide anomaly detection. This method can suppress abnormal reconstruction. Therefore, in the reconstruction error map, the anomaly has a high contrast with the background. The weight comes from the reconstruction error and can be updated adaptively. In addition, Generative Adversarial Networks (GANs) are also widely used in hyperspectral anomaly detection. Jiang et al. proposed an anomaly detection method based on GAN, using GAN to evaluate background statistical information and obtain spectral features in a new type of semi-supervised spectral learning, which can more accurately identify anomalies. GANs can learn complex distribution features from a large amount of normal data without relying on labels, and help detect abnormal pixels through generative networks. Zhong et al. used a rough search strategy based on morphological filtering to extract potential background subsets and used them as input for GAN training. Since the training samples are composed only of background elements, the abnormal samples cannot be accurately reconstructed in the test set, thereby improving the model's ability to distinguish between background and abnormal samples. Wu et al. used a pre-trained convolutional neural network to detect the input image data, which can extract deep information features with strong robustness and high recognition, and achieve separation of abnormal targets from the image background. Zhang et al. proposed an adaptive CNN network that uses tensor convolution and uses neighborhood tensor blocks as convolution kernels to evaluate the statistical characteristics of test pixels, thereby mining effective features. In order to reduce the impact of non-stationary noise on data, Zhang et al. proposed a tensor anomaly detection method that combines FrFT and CNN architectures. Through tensor transformation and PCA preprocessing strategies, the distinction between background and anomalies is enhanced in the frequency domain.

[0009] At present, the hyperspectral anomaly detection algorithm has the following defects:

[0010] (1) Noise sensitivity: Hyperspectral data may be affected by a variety of noise sources during the acquisition process, such as sensor noise, environmental interference, etc. The presence of noise may lead to false detection or missed detection, reducing the detection accuracy of the model;

[0011] (2) Limitations of background assumptions: Many hyperspectral anomaly detection methods usually rely on certain background assumptions; however, actual hyperspectral data are often complex and changing, and the background assumptions may not necessarily hold.

[0012] (3) High computational complexity: The processing of hyperspectral data often involves a large number of bands and pixels. If deep learning is used, the training time may be greatly increased.

[0013] (4) Poor generalization ability: Some hyperspectral anomaly detection methods perform well on some datasets, but their generalization ability is poor and they perform poorly on other datasets. Summary of the invention

[0014] The purpose of the present invention is to overcome the deficiencies of the above-mentioned prior art and provide a hyperspectral anomaly detection method based on a dual-line iterative convolutional neural network.

[0015] In order to achieve the above object, the present invention adopts the following technical solution:

[0016] A hyperspectral anomaly detection method based on a dual-line iterative convolutional neural network comprises the following steps:

[0017] Step S1, input and parameter setting, input hyperspectral data, put the hyperspectral data into the anomaly enhancement circuit and background reconstruction circuit, initialize the number of superpixels, convolution kernel size, learning rate and number of iterations;

[0018] Step S2: In the anomaly enhancement circuit, the hyperspectral data is input into the anomaly detector to obtain an anomaly probability map, and the connected region analysis is used to screen the abnormal region to generate an anomaly detection map, and the sliding average update strategy is adopted to obtain the anomaly probability detection map;

[0019] Step S3, in the background reconstruction circuit, the global robust CNN divides the hyperspectral data into superpixels and constructs a Laplacian matrix; the above operations are integrated into the convolutional neural network to reconstruct the hyperspectral data, and the error between the reconstructed data and the original data is used as an abnormal probability detection map;

[0020] Step S4, abnormal point determination, performing point multiplication on the abnormal probability detection maps obtained by the abnormal enhancement line and the background reconstruction line respectively to obtain the final abnormal probability map;

[0021] Step S5, updating the hyperspectral data, updating the hyperspectral data input by the anomaly enhancement line and the background reconstruction line in real time according to the final anomaly probability map;

[0022] Step S6, detection map generation, after the iteration is completed, the difference between the input data of the updated background reconstruction line and the original data is output as an anomaly detection map.

[0023] Furthermore, in step S3, the convolutional neural network adopts an encoding-decoding network structure, the encoding network is used to extract local features, and the Laplacian matrix is ​​added to its output before being input into the decoding network as the input of the decoding network.

[0024] Furthermore, the encoding network extracts local features including the following steps:

[0025] Step S31, the original hyperspectral data received by the network is represented as , m is the width of the image, n is the height of the image, and b is the number of bands of the image;

[0026] Step S32, in the encoding network, the hyperspectral data is input into the input layer, and in the first convolution layer, 64 convolution kernels of size 3×3×b are used to perform convolution operation, and the data is output after padding;

[0027] In step S33, the output data is passed to the second convolutional layer, and convolution is performed using 32 convolution kernels of size 3×3×64. The data is output after batch normalization and ReLU activation function.

[0028] Furthermore, the hyperspectral data is divided into M×M blocks, and the data of each block is represented as , and apply convolutional neural networks independently on each block.

[0029] Furthermore, in step S3, the hyperspectral data is divided into fewer pixel groups with consistent regions by introducing a superpixel segmentation algorithm, and the modified Laplacian matrix is ​​used to obtain the hyperspectral data. To speed up the construction of the Laplace matrix, the objective function of the proposed convolutional neural network is as follows:

[0030] (1)

[0031] In formula (1), is the reconstruction error, is the Laplace matrix regularization term, λ is the balance coefficient, N is the number of samples, and Z is the potential representation matrix of the hidden layer.

[0032] Furthermore, step S2 specifically includes the following steps:

[0033] Step S21, domain difference analysis, the hyperspectral data is put into the anomaly detector, the hyperspectral data is fused by band and divided into superpixels, the degree of abnormality of each pixel is judged according to the degree of difference with the surrounding pixels, and an abnormal probability map is obtained. The abnormal area is screened from the consistency image through connected region analysis, and the area with too large area is eliminated to generate an anomaly detection map;

[0034] Step S22, sliding average update, all anomaly detection maps are updated using the sliding average update strategy to obtain anomaly probability detection maps.

[0035] Furthermore, in step S2, the detection of the anomaly detector includes the following steps:

[0036] Step S211: Hyperspectral data Sum by band to generate grayscale image G;

[0037] Step S212, each pixel value of the grayscale image G is expressed as the sum of all bands:

[0038]

[0039] Normalize the grayscale image G to the interval [0, 1];

[0040] Step S213: divide the normalized grayscale image G into multiple superpixel regions using a superpixel segmentation algorithm, and assign a unique label to each superpixel. , represents the superpixel area to which the current pixel belongs;

[0041] Step S214, set the label of the current superpixel to , and set a distance parameter d, find the pixels surrounding the superpixel by distance field expansion, and set the current pixel value to , the pixel value of the area is , the calculation formula of the field is:

[0042] = ;

[0043] Step S215, calculate the current pixel value With the field pixel value The mean difference of :

[0044] ;

[0045] Step S216, square the difference value and record it in the abnormal probability map M, and perform standardization processing:

[0046] ;

[0047] Step S217, calculate The threshold T of the top 5% of pixel values ​​is:

[0048]

[0049] In the formula, Arrange all pixel values ​​in the abnormal probability map M in descending order and convert the first 5% of pixel values ​​to 1;

[0050] Step S218, perform connected region segmentation, calculate the number of pixels contained in each abnormal component, and record each pixel block as , i=1,2,3…; if If the number of pixels is greater than 50, the pixel block is considered not to be an outlier and is recorded as 0; otherwise, the original value is retained.

[0051] The beneficial effects of the present invention are:

[0052] 1. In the present invention, the background reconstruction circuit focuses on background reconstruction, and the anomaly enhancement circuit is responsible for anomaly enhancement. The two complement and promote each other, achieving more accurate anomaly detection and overcoming the problem that background reconstruction and anomaly detection are difficult to effectively combine in traditional methods.

[0053] 2. In the present invention, superpixel technology is combined with an anomaly detection strategy based on comparison with surrounding pixels to improve local sensitivity, make the selection of abnormal targets simpler and more efficient, improve detection efficiency, and avoid possible computational redundancy and efficiency bottlenecks.

[0054] 3. In the present invention, a convolutional neural network is introduced into hyperspectral anomaly detection, and combined with the global structural information of the Laplacian matrix, the extraction of multi-scale features is realized, which not only improves the model's ability to learn abnormal targets at different scales, but also effectively enhances the adaptability and accuracy of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 Schematic diagram of the overall architecture of the hyperspectral anomaly detection method based on a dual-line iterative convolutional neural network in this embodiment;

[0056] Figure 2 This is a flow chart of a hyperspectral anomaly detection method based on a dual-line iterative convolutional neural network in this embodiment. DETAILED DESCRIPTION

[0057] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0058] Embodiment: A hyperspectral anomaly detection method based on a dual-line iterative convolutional neural network, such as Figure 1 , Figure 2 As shown, the following steps are included:

[0059] Step S1, input and parameter setting, input hyperspectral data, put the hyperspectral data into the anomaly enhancement line and background reconstruction line, initialize the number of superpixels, convolution kernel size, learning rate, number of iterations and other parameters;

[0060] Step S2: In the anomaly enhancement circuit, the hyperspectral data is input into the anomaly detector to obtain an anomaly probability map, and the connected region analysis is used to screen the abnormal area to generate an anomaly detection map, and the sliding average update strategy is adopted to obtain the anomaly probability detection map; specifically:

[0061] Step S21, domain difference analysis, the hyperspectral data is put into the anomaly detector, the hyperspectral data is fused by band and divided into superpixels, the degree of abnormality of each pixel is judged according to the degree of difference with the surrounding pixels, and an abnormal probability map is obtained. The abnormal area is screened from the consistency image through connected region analysis, and the area with too large area is eliminated to generate an anomaly detection map;

[0062] Step S22, sliding average update, all anomaly detection maps are updated using the sliding average update strategy to obtain anomaly probability detection maps.

[0063] Among them, the purpose of the anomaly detector is to detect a small number of abnormal pixels from the image, which are usually significantly different from the statistical characteristics of the surrounding area. In hyperspectral data, anomaly detection not only needs to consider the differences in the spectral dimension, but also needs to combine the domain information in the spatial dimension. To this end, this embodiment proposes a hyperspectral anomaly detection method that combines grayscale, superpixel segmentation and domain difference analysis. For each pixel , calculate its difference with the domain pixel (from the superpixel area surrounding the current pixel), and the abnormality metric is measured by calculating the difference between the grayscale value of the current pixel and its domain mean. Specifically:

[0064] Step S211: Hyperspectral data Sum by band to generate grayscale image G, retaining the overall spectral information;

[0065] Step S212, each pixel value of the grayscale image G is expressed as the sum of all bands:

[0066]

[0067] Normalize the grayscale image G to the interval [0, 1];

[0068] Step S213, the normalized grayscale image G is divided into multiple superpixel regions by a superpixel segmentation algorithm (such as the SLIC algorithm). The goal of superpixel segmentation is to group pixels in the image so that each superpixel region contains similar pixel values, taking into account both the similarity of the grayscale values ​​of the pixels and their spatial proximity. Each superpixel is given a unique label. , represents the superpixel area to which the current pixel belongs;

[0069] Step S214, for each pixel , the difference between the pixel value and the mean of its surrounding area is taken as its abnormality degree; first, get the current pixel value , and then extract the domain pixel value from the superpixel where it is located; let the label of the current superpixel be , and set a distance parameter d to find the pixels surrounding the superpixel through distance domain expansion; set the domain pixel value to , that is, for each pixel The pixel value at belongs to the domain of the superpixel within the distance d. The domain calculation formula is:

[0070] = ;

[0071] Step S215, calculate the current pixel value With the field pixel value The mean difference of :

[0072] ;

[0073] Step S216, in order to highlight the abnormal features, the difference value is further squared and recorded in the abnormal probability map M, and then standardized:

[0074] ;

[0075] Step S217: Since the anomaly accounts for a very small part of the whole, threshold screening can be performed; calculate The threshold T of the top 5% of pixel values ​​is:

[0076]

[0077] In the formula, Arrange all pixel values ​​in the abnormal probability map M in descending order and convert the first 5% of pixel values ​​to 1;

[0078] Step S218, the above pixels are segmented into connected areas. Since the anomaly often occupies only a small part, in order to remove the pseudo anomaly, the number of pixels contained in each abnormal component is calculated, and each pixel block is recorded as , i=1,2,3…, you can set the pixel number threshold by yourself. In this embodiment, we believe that if it is greater than 50, it is a pseudo-anomaly, that is, if If the number of pixels is greater than 50, the pixel block is considered not to be an outlier and is recorded as 0; otherwise, the original value is retained.

[0079] Step S3, in the background reconstruction circuit, the global robust CNN divides the hyperspectral data into superpixels and constructs a Laplacian matrix; the above operations are integrated into the convolutional neural network to reconstruct the hyperspectral data, and the error between the reconstructed data and the original data is used as an abnormal probability detection map;

[0080] In step S3, the convolutional neural network adopts an encoding-decoding network structure, in which the encoding network is used to extract local features, and the Laplacian matrix is ​​added to its output before being input into the decoding network as the input of the decoding network. Specifically:

[0081] Step S31, the hyperspectral data received by the network is represented as , m is the width of the image, n is the height of the image, and b is the number of bands of the image;

[0082] Step S32, in the encoding network, the hyperspectral data is input into the input layer. In the first convolution layer, 64 convolution kernels of size 3×3×b are used for convolution operation, and padding is used to keep the spatial dimension unchanged, so the output size is still m×n, when the number of bands is reduced to 64;

[0083] Step S33, the output data is passed to the second convolutional layer. In this layer, convolution is performed using 32 convolution kernels of size 3×3×64, and batch normalization and ReLU activation function are also applied. After this layer of processing, the dimension of the output data becomes m×n×32.

[0084] The above stages are mainly to complete the extraction and expression of local features. At the same time, in order to make up for the insufficiency of convolutional neural network in modeling global features, the Laplace matrix L is introduced, and the processed data is input into the decoding network. By combining with the global features, the network further deconvolutions the data, restores it to the original dimension m×n×b, and realizes the reconstruction of hyperspectral data.

[0085] Considering that in practical applications, hyperspectral data usually has high spatial and spectral resolution, directly processing the entire data cube may result in excessive memory overhead or even inability to run. Therefore, this embodiment proposes a block processing strategy to divide the hyperspectral data into M×M blocks, and the data of each block is represented as The convolutional neural network is applied independently on each block. By processing in blocks, not only the memory requirement of a single calculation is reduced, but also the overall operation efficiency can be improved.

[0086] In order to accelerate the construction of the Laplacian matrix, the superpixel segmentation method is further introduced (corresponding to the PCA & SLIC method in ①). Superpixel technology can divide the hyperspectral data into fewer pixel groups with consistent areas to obtain a modified Laplacian matrix , thus significantly reducing the time complexity of calculating the Laplacian matrix. The objective function of the proposed convolutional neural network is as follows:

[0087] In formula (1), is the reconstruction error, is the Laplacian matrix regularization term, which retains the local geometric structure and spatial consistency, and encourages samples in local homogeneous regions to share similar representations. λ is the balancing coefficient, N is the number of samples, that is, the total number of pixels, which is m×n×b, and Z is the potential representation matrix of the hidden layer.

[0088] Step S4, abnormal point determination, performing point multiplication on the abnormal probability detection maps obtained by the abnormal enhancement line and the background reconstruction line respectively to obtain the final abnormal probability map;

[0089] Step S5, updating the hyperspectral data, updating the hyperspectral data input by the anomaly enhancement line and the background reconstruction line in real time according to the final anomaly probability map;

[0090] Step S6, detection map generation, after the iteration is completed, the difference between the input data of the updated background reconstruction line and the original data is output as an anomaly detection map.

[0091] For anomaly detection of hyperspectral data, anomaly enhancement and background reconstruction are combined to make the two complement and promote each other. Figure 1 As shown, in the initial state, Represents the original hyperspectral data, where m is the width of the image, n is the height of the image, and b is the number of bands in the image. , They represent the inputs of the background reconstruction circuit and the anomaly enhancement circuit at the kth iteration, respectively. , .

[0092] In the background reconstruction circuit (box ①), After reconstruction as input, the error between the two is ;

[0093] ;

[0094] In the abnormal enhancement line (box ②), The enhanced anomaly components are put into the anomaly detector based on the comparison with surrounding pixels and combined with superpixel technology. , and adopt a sliding average update strategy to enhance the robustness of the anomaly detector:

[0095] ;

[0096] Only when the anomaly detection images obtained by the background reconstruction circuit and the anomaly enhancement circuit consider the point to be an anomaly, the point is a real anomaly. Therefore, the anomaly weight of each pixel is finally obtained as:

[0097] ;

[0098] In order to make the hyperspectral data obtained by the background reconstruction line closer to the background, As the input of the next background reconstruction circuit; at the same time, in order to facilitate the enhancement of the abnormal part next time, As the input of the next abnormal enhancement circuit, since each iteration will result in The data of the outliers decreases rapidly, so the difference between the hyperspectral data obtained by the last iteration and the original hyperspectral data can be output in the end:

[0099] ;

[0100] Finally, D is spectrally fused according to the channel dimension to obtain

[0101] ;

[0102] Finally, D is normalized to obtain the final detection map D.

[0103] The above is only a preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, some improvements and modifications without departing from the principle of the present invention should also be regarded as the protection scope of the present invention.

Claims

1. A hyperspectral anomaly detection method based on a dual-line iterative convolutional neural network, characterized in that: The steps include: Step S1, input and parameter setting, input hyperspectral data, put the hyperspectral data into the anomaly enhancement circuit and background reconstruction circuit, initialize the number of superpixels, convolution kernel size, learning rate and number of iterations; Step S2: In the anomaly enhancement circuit, the hyperspectral data is input into the anomaly detector to obtain an anomaly probability map, and the connected region analysis is used to screen the abnormal region to generate an anomaly detection map, and the sliding average update strategy is adopted to obtain the anomaly probability detection map; Step S3, in the background reconstruction circuit, the global robust CNN divides the hyperspectral data into superpixels and constructs a Laplacian matrix; the above operations are integrated into the convolutional neural network to reconstruct the hyperspectral data, and the error between the reconstructed data and the original data is used as an abnormal probability detection map; Step S4, abnormal point determination, performing point multiplication on the abnormal probability detection maps obtained by the abnormal enhancement line and the background reconstruction line respectively to obtain the final abnormal probability map; Step S5, updating the hyperspectral data, updating the hyperspectral data input by the anomaly enhancement line and the background reconstruction line in real time according to the final anomaly probability map; Step S6, detection map generation, after the iteration is completed, the difference between the input data of the updated background reconstruction line and the original data is output as an anomaly detection map.

2. According to claim 1, a hyperspectral anomaly detection method based on dual-line iterative convolutional neural network is characterized in that: In step S3, the convolutional neural network adopts an encoding-decoding network structure. The encoding network is used to extract local features. After adding the Laplacian matrix to its output, it is input into the decoding network as the input of the decoding network.

3. According to claim 2, a method for detecting anomalies of a hyperspectral spectrum based on a dual-line iterative convolutional neural network is characterized in that: The encoding network extracts local features including the following steps: Step S31, the original hyperspectral data received by the network is represented as , m is the width of the image, n is the height of the image, and b is the number of bands of the image; Step S32, in the encoding network, the hyperspectral data is input into the input layer, and in the first convolution layer, 64 convolution kernels of size 3×3×b are used to perform convolution operation, and the data is output after padding; In step S33, the output data is passed to the second convolutional layer, and convolution is performed using 32 convolution kernels of size 3×3×64. The data is output after batch normalization and ReLU activation function.

4. According to claim 3, a method for hyperspectral anomaly detection based on dual-line iterative convolutional neural network is characterized in that: The hyperspectral data is divided into M×M blocks, and the data of each block is represented as , and apply convolutional neural networks independently on each block.

5. According to claim 1, a method for hyperspectral anomaly detection based on dual-line iterative convolutional neural network is characterized in that: In step S3, the hyperspectral data is divided into fewer pixel groups with consistent regions by introducing a superpixel segmentation algorithm, and the modified Laplacian matrix is ​​used to To speed up the construction of the Laplace matrix, the objective function of the proposed convolutional neural network is as follows: (1) In formula (1), is the reconstruction error, is the Laplace matrix regularization term, λ is the balance coefficient, N is the number of samples, and Z is the potential representation matrix of the hidden layer.

6. According to claim 1, a method for hyperspectral anomaly detection based on dual-line iterative convolutional neural network is characterized in that: Step S2 specifically includes the following steps: Step S21, domain difference analysis, the hyperspectral data is put into the anomaly detector, the hyperspectral data is fused by band and divided into superpixels, the degree of abnormality of each pixel is judged according to the degree of difference with the surrounding pixels, and an abnormal probability map is obtained. The abnormal area is screened from the consistency image through connected region analysis, and the area with too large area is eliminated to generate an anomaly detection map; Step S22, sliding average update, all anomaly detection maps are updated using the sliding average update strategy to obtain anomaly probability detection maps.

7. The method for hyperspectral anomaly detection based on dual-line iterative convolutional neural network according to claim 1, characterized in that: In step S2, the detection of the anomaly detector includes the following steps: Step S211: Hyperspectral data Sum by band to generate grayscale image G; Step S212, each pixel value of the grayscale image G is expressed as the sum of all bands: Normalize the grayscale image G to the interval [0, 1]; Step S213: divide the normalized grayscale image G into multiple superpixel regions using a superpixel segmentation algorithm, and assign a unique label to each superpixel. , represents the superpixel area to which the current pixel belongs; Step S214, set the label of the current superpixel to , and set a distance parameter d, find the pixels surrounding the superpixel by distance field expansion, and set the current pixel value to , the pixel value of the area is , the calculation formula of the field is: = ； Step S215, calculate the current pixel value With the field pixel value The mean difference of : ； Step S216, square the difference value and record it in the abnormal probability map M, and perform standardization processing: ； Step S217, calculate The threshold T of the top 5% of pixel values ​​is: In the formula, Arrange all pixel values ​​in the abnormal probability map M in descending order and convert the first 5% of pixel values ​​to 1; Step S218, perform connected region segmentation, calculate the number of pixels contained in each abnormal component, and record each pixel block as , i=1,2,3…; if If the number of pixels is greater than 50, the pixel block is considered not to be an outlier and is recorded as 0; otherwise, the original value is retained.

Images