Hyperspectral anomaly detection method based on autoencoder and low-dimensional manifold modeling
Through the methods of autoencoder dimensionality reduction and low-dimensional manifold modeling, the problem of high background modeling complexity caused by redundancy of hyperspectral image information is solved, and efficient abnormal detection effect is achieved.
Patent Information
- Application Number
- CN202111122594.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-09-24
- Publication Date
- 2025-04-08
- Estimated Expiration
- 2041-09-24
AI Technical Summary
The information redundancy of hyperspectral images leads to high complexity of background modeling, affecting the abnormal detection effect.
The data dimensionality reduction of hyperspectral images is used to reduce the data through low-dimensional manifold modeling to reduce redundant calculations.
It effectively reduces the processing complexity of abnormal detection of hyperspectral images and improves the accuracy and speed of detection results.
Smart Images
Figure CN113902973B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of remote sensing image processing technology, and in particular relates to a hyperspectral anomaly detection method based on an autoencoder and low-dimensional manifold modeling. Background Art
[0002] Hyperspectral images add spectral dimensions to traditional remote sensing images, and are three-dimensional data structures that contain target spatial features and spectral information. Hyperspectral image detection technology can generally be divided into target detection and anomaly detection. Since the characteristics of abnormal targets can be detected without knowing the spectral information of the target object in advance, hyperspectral anomaly detection is widely used in many fields such as agriculture, meteorology, and military. The research on hyperspectral anomaly detection is becoming increasingly important.
[0003] Hyperspectral image anomaly detection mainly includes two steps: background modeling and anomaly detection. Background modeling establishes background metrics by finding an evaluation method for background signals in hyperspectral images, and anomaly detection evaluates whether each pixel belongs to the background based on the background metrics. Pixels that do not belong to the background are abnormal pixels. The quality of background modeling directly affects the final anomaly detection results. At the same time, hyperspectral images have spectral information of hundreds of bands, which leads to information redundancy of hyperspectral images and has a great impact on the construction of the background. Therefore, how to remove redundant information in the original data, reduce data dimensions, and reduce the processing complexity of anomaly detection is a focus in the field of hyperspectral anomaly detection. The present invention is characterized by making a series of improvements to the above-mentioned problems. Summary of the invention
[0004] In view of this, the main purpose of the present invention is to provide a hyperspectral anomaly detection method based on autoencoder and low-dimensional manifold modeling.
[0005] To achieve the above object, the technical solution of the present invention is achieved as follows:
[0006] An embodiment of the present invention provides a hyperspectral anomaly detection method using an autoencoder and low-dimensional manifold modeling, characterized in that the method is:
[0007] The stacked autoencoder is trained by using the original hyperspectral image to obtain a trained stacked autoencoder;
[0008] Extracting main features of the original hyperspectral image by using the trained stacked autoencoder to complete data dimensionality reduction and obtain a hyperspectral image Y after dimensionality reduction;
[0009] The hyperspectral image Y after dimension reduction is randomly sampled L times to obtain a sampled image l=1,2,…,L;
[0010] Each sampled image is obtained by low-dimensional manifold modeling. Reconstruct and obtain each feature sampling image The reconstructed background sub-image X l ; Reconstruct the background sub-image X l Find the average As the reconstructed background image of the original hyperspectral image
[0011] Determine the original hyperspectral image and reconstructed background image by l2 norm The residual r between them is taken as the final anomaly detection result.
[0012] In the above scheme, the stacked autoencoder is trained by the original hyperspectral image to obtain the trained stacked autoencoder, specifically:
[0013] Step 1.1, the size of the original hyperspectral image is M×N×B, where M represents the number of rows in the hyperspectral image, N represents the number of columns in the hyperspectral image, and B represents the number of bands;
[0014] Step 1.2, use stacked autoencoders to reduce the dimension of the hyperspectral image, use mean square error as the loss function during training, and use ReLU function as the activation function;
[0015] Step 1.3, train the first single-layer autoencoder, fix the number of encoding units of the first encoder, that is, the number of neurons in the first hidden layer, to 100, so that it can complete the reconstruction of the original input, and then freeze the parameters obtained during the training process of the first single-layer autoencoder, and use the output Hidden 1 of the hidden layer of the first single-layer autoencoder as the input of the second single-layer autoencoder;
[0016] Step 1.4, train the second single-layer autoencoder to complete the reconstruction of Hidden 1. The number of encoding units of the second encoder, i.e. the number of neurons in the second hidden layer, is fixed to 50;
[0017] Step 1.5, integrate the two trained autoencoders and train them as a whole. The number of encoding units of the third encoder of the autoencoder, that is, the number of neurons in the third hidden layer, is set to 25. The parameters of each layer are fine-tuned to obtain a stacked autoencoder. The parameters of each layer are fine-tuned to obtain a trained stacked autoencoder.
[0018] In the above scheme, the main features of the original hyperspectral image are extracted by the trained stacked autoencoder to complete data dimensionality reduction and obtain the hyperspectral image Y after dimensionality reduction, specifically: the original hyperspectral image Y∈R M×N×B As input, the main features of the original hyperspectral image are extracted through the trained stacked autoencoder to complete data dimensionality reduction and obtain the hyperspectral image Y∈R after dimensionality reductionM×N×n , n is the feature dimension after dimensionality reduction, n is equal to 25.
[0019] In the above scheme, each sampled image is obtained by low-dimensional manifold modeling. Reconstruct and obtain each feature sampling image The reconstructed background sub-image X l , specifically:
[0020] Step 4.1, initialize the reconstructed background sub-image X (0) ;
[0021] Step 4.2, from X (k) Extract the three-dimensional block set PX (k) , PX (k) Defined as X (k) A set of local three-dimensional blocks of size d1×d2×B in , where k represents the number of iterations, d1 and d2 represent the spatial dimensions of the block, d1 and d2 are 2, and the similarity matrix in the spatial domain is calculated as follows:
[0022]
[0023] Where σ(u) is the regularization factor, which represents the distance between the local 3D block u and its zth nearest neighbor local 3D block, and z is 10;
[0024] Step 4.3, assemble a new similarity matrix The symbol q {j} represents the jth component (in the spatial domain) in the three-dimensional block q, d s =d1×d2 is the spatial dimension, and the similarity calculation is truncated to 20 nearest neighbors, that is, only the 20 corresponding elements with the smallest adjacent distance are retained in the similarity matrix;
[0025] Step 4.4, in each spectral band t, use the generalized minimum residual method to update (X t ) (k+1) :
[0026]
[0027] Where μ = 1 / r-1, r is the sampling rate, which is 0.05, and λ is the regularization parameter, which is equal to 10 7 , represents a collection of hyperspectral image spatial blocks, is In this case, X t (q) The projection operator with projection 0, where Ω t A subset is obtained by randomly sampling from any band t∈[B] of the hyperspectral image;
[0028] Step 4.5, iteration number k = k + 1, make X = X (k+1) , X (k) =X (k+1) , repeat steps 4.2 to 4.4 until X converges, and then stop the loop to obtain the reconstructed background sub-image X l .
[0029] Compared with the prior art, the present invention firstly adopts a stacked autoencoder to perform data dimensionality reduction processing on the original hyperspectral image, and then uses a low-dimensional manifold modeling anomaly detector to perform anomaly detection on the hyperspectral image after dimensionality reduction; the anomaly detection method for hyperspectral images provided by the present invention aims at the problem that there are a large number of redundant calculations in the process of low-dimensional manifold modeling of hyperspectral images, and adopts a deep learning-based autoencoder dimensionality reduction method, which effectively reduces the redundant calculations in the hyperspectral image anomaly detection algorithm based on low-dimensional manifold modeling, improves the overall performance of the algorithm, and speeds up the algorithm. BRIEF DESCRIPTION OF THE DRAWINGS
[0030] The drawings described herein are used to disclose a further understanding of the present invention and constitute a part of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:
[0031] Figure 1 This is the structure diagram of the autoencoder network;
[0032] Figure 2 This is a diagram of the stacked autoencoder training process;
[0033] Figure 3 This is the structure diagram of the stacked autoencoder;
[0034] Figure 4 It is a flowchart of multiple random sampling background reconstruction in an example of the present invention;
[0035] Figure 5 Schematic diagram of a hyperspectral image target in an example of the present invention;
[0036] Figure 6 This is a diagram of anomaly detection results in an example of the present invention;
[0037] Figure 7 It is a flowchart of an example of the present invention. DETAILED DESCRIPTION
[0038] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.
[0039] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, article or device. In the absence of more restrictions, an element defined by the sentence "includes a ..." does not exclude the existence of other identical elements in the process, article or device including the element.
[0040] An embodiment of the present invention provides a hyperspectral anomaly detection method using an autoencoder and low-dimensional manifold modeling, which is implemented by the following steps:
[0041] Step 1, training the stacked autoencoder using the original hyperspectral image to obtain a trained stacked autoencoder;
[0042] Step 2, extracting the main features of the original hyperspectral image through the trained stacked autoencoder to complete data dimensionality reduction and obtain the hyperspectral image Y after dimensionality reduction;
[0043] Step 3: Perform L random sampling on the dimension-reduced hyperspectral image Y to obtain a sampled image l=1,2,…,L;
[0044] Step 4: Model each sampled image using low-dimensional manifold modeling Reconstruct and obtain each feature sampling image The reconstructed background sub-image X l ; Reconstruct the background sub-image X l Find the average As the reconstructed background image of the original hyperspectral image
[0045] Step 5: Determine the original hyperspectral image and reconstructed background image by L2 norm The residual r between them is taken as the final anomaly detection result.
[0046] Wherein, step 1 is specifically as follows:
[0047] Step 1.1, the size of the original hyperspectral image is M×N×B, where M represents the number of rows in the hyperspectral image, N represents the number of columns in the hyperspectral image, and B represents the number of bands;
[0048] Step 1.2, use stacked autoencoders to reduce the dimension of the hyperspectral image, use mean square error as the loss function during training, and use ReLU function as the activation function;
[0049] Step 1.3, train the first single-layer autoencoder, fix the number of encoding units of the first encoder, that is, the number of neurons in the first hidden layer, to 100, so that it can complete the reconstruction of the original input, and then freeze the parameters obtained during the training process of the first single-layer autoencoder, and use the output Hidden 1 of the hidden layer of the first single-layer autoencoder as the input of the second single-layer autoencoder;
[0050] Step 1.4, train the second single-layer autoencoder to complete the reconstruction of Hidden 1. The number of encoding units of the second encoder, i.e. the number of neurons in the second hidden layer, is fixed to 50;
[0051] Step 1.5, integrate the two trained autoencoders and train them as a whole. The number of encoding units of the third encoder of the autoencoder, that is, the number of neurons in the third hidden layer, is set to 25. The parameters of each layer are fine-tuned to obtain a stacked autoencoder. The parameters of each layer are fine-tuned to obtain a trained stacked autoencoder.
[0052] Among them, step 2 is specifically: the original hyperspectral image Y∈R M×N×B As input, the main features of the original hyperspectral image are extracted through the trained stacked autoencoder to complete data dimensionality reduction and obtain the hyperspectral image Y∈R after dimensionality reduction M×N×n , n is the feature dimension after dimensionality reduction, n is equal to 25.
[0053] Among them, step 4 is specifically as follows:
[0054] Step 4.1, initialize the reconstructed background sub-image X (0) ;
[0055] Step 4.2, from X (k) Extract the three-dimensional block set PX (k) , PX (k) Defined as X (k) A set of local three-dimensional blocks of size d1×d2×B in , where k represents the number of iterations, d1 and d2 represent the spatial dimensions of the block, d1 and d2 are 2, and the similarity matrix in the spatial domain is calculated as follows:
[0056]
[0057] Where σ(u) is the regularization factor, which represents the distance between the local 3D block u and its zth nearest neighbor local 3D block, and z is 10;
[0058] Step 4.3, assemble a new similarity matrix The symbol q {j} represents the jth component (in the spatial domain) in the three-dimensional block q, d s=d1×d2 is the spatial dimension, and the similarity calculation is truncated to 20 nearest neighbors, that is, only the 20 corresponding elements with the smallest adjacent distance are retained in the similarity matrix;
[0059] Step 4.4, in each spectral band t, use the generalized minimum residual method to update (X t ) (k+1) :
[0060]
[0061] Where μ = 1 / r-1, r is the sampling rate, which is 0.05, and λ is the regularization parameter, which is equal to 10. represents a collection of hyperspectral image spatial blocks, is In this case, X t (q) projection operator with projection equal to 0, where Ω t A subset is obtained by randomly sampling from any band t∈[B] of the hyperspectral image;
[0062] Step 4.5, iteration number k = k + 1, make X = X (k+1) , X (k) =X (k+1) , repeat steps 4.2 to 4.4 until X converges, and then stop the loop to obtain the reconstructed background sub-image X l ;
[0063] Step 4.6: Sample each feature image The reconstructed background sub-image X l Find the average As the reconstructed background image of the original hyperspectral image.
[0064] In order to demonstrate the effectiveness of the proposed method, real hyperspectral images are used for anomaly detection. The experiment uses a sub-image of 90×90 and 186 bands in the San Diego airport scene image collected by the Airborne Visible / Infrared Imaging Spectrometer (AVIRIS) sensor for anomaly detection, and three aircraft in the sub-image are taken as abnormal targets.
[0065] like Figure 7 As shown in the figure, the basic process of anomaly detection for SanDiego image data is as follows:
[0066] Step 101, the original hyperspectral image is a 90×90×186 three-dimensional matrix with a total of 90×90 spatial pixels, containing spectral data of ground targets in 186 bands. A one-dimensional vector of size 1×1×186 corresponding to each spatial pixel represents the corresponding spectral information.
[0067] Step 102, the autoencoder is a data preprocessing neural network that can reduce the dimension of data by constraining the number of encoding units to be smaller than the dimension of input data, and includes an input layer, a hidden layer, and an output layer.
[0068] Figure 1 As shown, the conversion process from the input layer to the hidden layer can be regarded as an encoding process. The relationship between the hidden layer vector h and the input vector x is shown in the following formula, where W h represents the weight matrix from the input layer to the hidden layer, b h represents the bias vector in the hidden layer, and s represents the activation function;
[0069] h=s(W h x+b h )
[0070] The process from the hidden layer to the output layer can be regarded as a decoding process. The relationship between the hidden layer vector h and the output vector y is shown in the following formula. y represents the weight matrix from the input layer to the hidden layer, b y represents the bias vector in the hidden layer, and s represents the activation function;
[0071] y=s(W y h+b y )
[0072] Step 103, similar to the idea of increasing the depth of the neural network to obtain a feature representation of a higher level of abstraction of the data, stacking multiple autoencoders can obtain a stacked autoencoder with a deeper network layer, such as Figure 3 As shown, the present invention uses a stacked autoencoder to complete the dimensionality reduction of the original hyperspectral image. The training process is as follows: Figure 2 As shown. During training, the mean square error is used as the loss function and the ReLU function is used as the activation function. In addition, the batch size, learning rate and epoch number are also parameters in the stacked autoencoder. In this invention, the batch size is set to 200 and the learning rate is set to 10. -3 , the number of training epochs for each autoencoder is 150.
[0073] Step 1.4, train the first single-layer autoencoder, fix the number of encoding units of the first encoder (that is, the number of neurons in the first hidden layer) to 100, so that it can complete the reconstruction of the original input, then freeze the parameters obtained during the training process in the first single-layer autoencoder, and use the output Hidden 1 of the hidden layer of the first single-layer autoencoder as the input of the second single-layer autoencoder.
[0074] Step 104 , training a second single-layer autoencoder to complete the reconstruction of Hidden 1, and the number of encoding units of the second encoder (ie, the number of neurons in the second hidden layer) is fixed to 50.
[0075] Step 105, integrating the two trained autoencoders and training them as a whole, setting the number of encoding units of the third encoder of the autoencoder, i.e., the number of neurons in the third hidden layer, to 25, and fine-tuning the parameters of each layer to obtain a stacked autoencoder;
[0076] Step 106: transform the original hyperspectral image Y∈R M×N×B As input, the trained stacked autoencoder obtained in step 1 is used to extract effective features from the original hyperspectral image to obtain the reduced-dimensional hyperspectral image Y∈R M×N×n , n is the feature dimension after dimensionality reduction, n is equal to 25;
[0077] Step 107, in order to reduce the influence of outliers on background reconstruction, the present invention utilizes the characteristics that outliers are sparsely distributed in the scene and have a small probability of being sampled, such as Figure 4 As shown in Figure 1, a multiple random sampling strategy is used to reduce the impact of outliers on background reconstruction. M×N×n Perform L random sampling to obtain the sampled image l=1,2,…,L;
[0078] Step 108: Initialize and reconstruct the background sub-image X (0) .
[0079] Step 109, from X (k) Extract the three-dimensional block set PX (k) , PX (k) Defined as X (k) A set of local three-dimensional blocks of size d1×d2×B, where k represents the number of iterations, d1 and d2 represent the spatial dimensions of the blocks, and d1 and d2 are set to 2 in the present invention, and the similarity matrix in the spatial domain is calculated according to the following formula.
[0080]
[0081] Wherein, σ(u) is a regularization factor, which represents the distance between the local three-dimensional block u and its zth nearest neighbor. In the present invention, z is set to 10.
[0082] Step 110, assemble a new similarity matrix The symbol q {j} represents the jth component (in the spatial domain) in the three-dimensional block q, d s =d1×d2 is the spatial dimension. At the same time, the similarity matrix It is an asymmetric matrix. In order to reduce the computational complexity, the similarity calculation is truncated to 20 nearest neighbors, that is, only the 20 corresponding elements with the smallest adjacent distance are retained in the similarity matrix.
[0083] Step 111: In each spectral band t, the generalized minimum residual method is used to update (X t ) (k+1) .
[0084]
[0085] Where μ = 1 / r-1, r is the sampling rate, which is 0.05, and λ is the regularization parameter, which is equal to 10. represents a collection of hyperspectral image spatial blocks, is In this case, X t (q) The projection operator with projection 0, where Ω t A subset is obtained by randomly sampling from any band t∈[B] of the hyperspectral image.
[0086] Step 112, the number of iterations k = k + 1, so that X = X (k+1) , X (k) =X (k+1) , repeat step 109 to step 111 until X converges and then stop the loop to obtain the reconstructed background image X l .
[0087] Step 113: each feature sampling image The reconstructed background sub-image X l Find the average As the reconstructed background image of the original hyperspectral image.
[0088] Step 114, using the l2 norm to calculate the original hyperspectral image and the reconstructed background image The residual r between them is taken as the final anomaly detection result.
[0089] Figure 6 The abnormal detection results of the GRX method, the RPCA-RX method and the method of the present invention can be seen. It can be seen that the existence of the detection target is difficult to see in the detection results of the GRX method and the RPCA-RX method, and the target is most obvious in the detection results of the method of the present invention. The method of the present invention can better detect abnormal target information in the hyperspectral image.
[0090] The above description is only a preferred embodiment of the present invention and is not intended to limit the protection scope of the present invention.
Claims
1. A hyperspectral anomaly detection method based on autoencoder and low-dimensional manifold modeling, characterized in that: The method is: The stacked autoencoder is trained by using the original hyperspectral image to obtain a trained stacked autoencoder; Extracting main features of the original hyperspectral image by using the trained stacked autoencoder to complete data dimensionality reduction and obtain a hyperspectral image Y after dimensionality reduction; The hyperspectral image Y after dimension reduction is randomly sampled L times to obtain a sampled image O l _, l = 1, 2, ..., L; Each sampled image O is obtained by low-dimensional manifold modeling. l _Reconstruct and obtain each feature sampling image O l Reconstructed background sub-image X of _ l ; Reconstruct the background sub-image X l Find the average As the reconstructed background image of the original hyperspectral image Determine the original hyperspectral image and reconstructed background image by l2 norm The residual r between them is used as the final anomaly detection result; The low-dimensional manifold modeling is used to obtain each sample image O l _Reconstruct and obtain each feature sampling image O l Reconstructed background sub-image X of _ l , specifically: Step 4.1, initialize the reconstructed background sub-image X (0) ; Step 4.2, from X (k) Extract the three-dimensional block set PX (k) , PX (k) Defined as X (k) A set of local three-dimensional blocks of size d1×d2×B in , where k represents the number of iterations, d1 and d2 represent the spatial dimensions of the block, d1 and d2 are 2, and the similarity matrix in the spatial domain is calculated as follows: Where σ(u) is the regularization factor, which represents the distance between the local 3D block u and its zth nearest neighbor local 3D block, and z is 10; Step 4.3, assemble a new similarity matrix The symbol q {j} represents the jth component (in the spatial domain) in the three-dimensional block q, d s =d1×d2 is the spatial dimension, and the similarity calculation is truncated to 20 nearest neighbors, that is, only the 20 corresponding elements with the smallest adjacent distance are retained in the similarity matrix; Step 4.4, in each spectral band t, use the generalized minimum residual method to update (X t ) (k+1) : Where μ = 1 / r-1, r is the sampling rate, which is 0.05, and λ is the regularization parameter, which is equal to 10 7 , represents a collection of hyperspectral image spatial blocks, is In this case, X t (q) projection operator with projection equal to 0, where Ω t A subset is obtained by randomly sampling from any band t∈[B] of the hyperspectral image; Step 4.5, iteration number k = k + 1, make X = X (k+1) , X (k) =X (k+1) , repeat steps 4.2 to 4.4 until X converges, and then stop the loop to obtain the reconstructed background sub-image X l .
2. The hyperspectral anomaly detection method based on autoencoder and low-dimensional manifold modeling according to claim 1, characterized in that: The stacked autoencoder is trained by the original hyperspectral image to obtain a trained stacked autoencoder, specifically: Step 1.1, the size of the original hyperspectral image is M×N×B, where M represents the number of rows in the hyperspectral image, N represents the number of columns in the hyperspectral image, and B represents the number of bands; Step 1.2, use stacked autoencoders to reduce the dimension of the hyperspectral image, use mean square error as the loss function during training, and use ReLU function as the activation function; Step 1.3, train the first single-layer autoencoder, fix the number of encoding units of the first encoder, that is, the number of neurons in the first hidden layer, to 100, so that it can complete the reconstruction of the original input, and then freeze the parameters obtained during the training process of the first single-layer autoencoder, and use the output Hidden 1 of the hidden layer of the first single-layer autoencoder as the input of the second single-layer autoencoder; Step 1.4, train the second single-layer autoencoder to complete the reconstruction of Hidden 1. The number of encoding units of the second encoder, i.e. the number of neurons in the second hidden layer, is fixed to 50; Step 1.5, integrate the two trained autoencoders and train them as a whole. The number of encoding units of the third encoder of the autoencoder, that is, the number of neurons in the third hidden layer, is set to 25. The parameters of each layer are fine-tuned to obtain a stacked autoencoder. The parameters of each layer are fine-tuned to obtain a trained stacked autoencoder.
3. The hyperspectral anomaly detection method based on autoencoder and low-dimensional manifold modeling according to claim 1 or 2, characterized in that: The main features of the original hyperspectral image are extracted by the trained stacked autoencoder to complete data dimensionality reduction and obtain the hyperspectral image Y after dimensionality reduction, specifically: the original hyperspectral image Y∈R M×N×B As input, the main features of the original hyperspectral image are extracted through the trained stacked autoencoder to complete data dimensionality reduction and obtain the hyperspectral image Y∈R after dimensionality reduction M×N×n , n is the feature dimension after dimensionality reduction, n is equal to 25.