A hyperspectral anomaly detection method embedded in a robust subspace recovery layer

Abstract

The present invention discloses a hyperspectral anomaly detection method embedded with a robust subspace recovery layer, comprising the following steps: 1. constructing an autoencoder network and embedding the robust subspace recovery layer; 2. obtaining a total loss function during iterative anomaly detection of hyperspectral images; and 3. iteratively performing anomaly detection on the hyperspectral image to obtain a trained autoencoder network embedded with the robust subspace recovery layer. The method is simple and well-designed. By embedding the robust subspace recovery layer in the autoencoder network and iterating anomaly detection based on the robust subspace recovery layer and the total loss function of the autoencoder network, the method increases the ability to distinguish anomalies from background, thereby improving detection accuracy and effectiveness.

Description

Technical Field

[0001] The present invention belongs to the technical field of remote sensing image processing, and in particular relates to a hyperspectral anomaly detection method embedded in a robust subspace recovery layer. Background Art

[0002] Hyperspectral anomaly detection aims to classify all pixels in a hyperspectral image as anomalies or background. Objects in a hyperspectral image that are significantly different in spectrum from their surrounding pixels are considered anomalies. It is an important task in the field of hyperspectral image processing.

[0003] Hyperspectral anomaly detection methods mainly include (1) statistics-based, (2) representation-based, (3) subspace-based, (4) spatial-spectral-based, and (5) deep learning-based. Deep learning-based methods have gradually attracted the attention of researchers in recent years. Among the deep learning-based methods, the autoencoder network method has achieved good application in hyperspectral anomaly detection, but the hidden high-level semantic features represented by its hidden layer cannot directly distinguish between background and anomalies, thus affecting the accuracy of anomaly detection. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to address the deficiencies in the above-mentioned prior art and provide a hyperspectral anomaly detection method embedded with a robust subspace recovery layer. The method has simple steps and a reasonable design. By embedding the robust subspace recovery layer in the autoencoder network, anomaly detection iteration is performed according to the total loss function of the robust subspace recovery layer and the autoencoder network, thereby increasing the distinction between anomalies and background, thereby improving detection accuracy and effect.

[0005] To solve the above technical problems, the present invention adopts a technical solution: a hyperspectral anomaly detection method embedded in a robust subspace recovery layer, characterized in that the method comprises the following steps:

[0006] Step 1: Build an autoencoder network and embed a robust subspace recovery layer:

[0007] Step 101: Use a computer to construct an autoencoder network; wherein the autoencoder network includes an input layer, five hidden layers, and an output layer, and the five hidden layers from the input layer to the output layer are respectively a first hidden layer, a second hidden layer, a third hidden layer, a fourth hidden layer, and a fifth hidden layer;

[0008] Step 102: embedding a robust subspace recovery layer in the autoencoder network; wherein the robust subspace recovery layer sets a linear transformation matrix B, the robust subspace recovery layer includes a robust subspace recovery hidden layer and a robust subspace recovery output layer, and the third hidden layer of the autoencoder network is connected to the robust subspace recovery hidden layer;

[0009] Step 2: Obtaining the total loss function during the iteration of hyperspectral image anomaly detection:

[0010] Step 3: Iterate anomaly detection on the hyperspectral image to obtain the trained autoencoder network embedded in the robust subspace recovery layer;

[0011] Step 301: Use a computer to use the gradient descent method to calculate the total loss function L total Iterative anomaly detection on hyperspectral images;

[0012] Step 302: Repeat step 301 multiple times until the preset number of iterative training times is met, and a trained autoencoder network embedded with a robust subspace recovery layer is obtained.

[0013] The above-mentioned hyperspectral anomaly detection method embedded in a robust subspace recovery layer is characterized in that the number of neurons in the input layer and the output layer of the autoencoder network in step 101 is the same as the number of bands of the hyperspectral image.

[0014] The above-mentioned hyperspectral anomaly detection method embedded in a robust subspace recovery layer is characterized in that: in step 2, the total loss function is obtained during the iterative hyperspectral image anomaly detection, and the specific process is as follows:

[0015] Step 201: Set the size of the hyperspectral image A to m×n×I, and the input vector formed by the I bands at the j-th pixel in the hyperspectral image is recorded as A j ,and in, Represents the data corresponding to the i-th band at the j-th pixel, j is a positive integer, and the value range of j is 1 to mn; i and I are both positive integers, and 1≤i≤I, m represents the number of rows of the hyperspectral image, n represents the number of columns of the hyperspectral image, and I is the number of bands;

[0016] Step 202: Using a computer Get the reconstruction loss of the autoencoder network in, It represents the data corresponding to the i-th band at the j-th pixel after the hyperspectral image is reconstructed by the autoencoder network, D1(A j ) represents the hyperspectral image reconstructed by the autoencoder network and A j The corresponding output vector, ||·||2 represents the 2-norm;

[0017] Step 203: Using a computer Get the reconstruction loss of the robust subspace recovery layer Among them, among them, represents the data corresponding to the i-th band at the j-th pixel after the hyperspectral image is encoded by the autoencoder network and reconstructed by the robust subspace recovery layer, D2(Aj ) represents the hyperspectral image after encoding by the autoencoder network and reconstruction by the robust subspace recovery layer and A j The corresponding output vector, ||·||2 represents the 2-norm;

[0018] Step 204: Using a computer Get the linear transformation matrix loss L of the robust subspace recovery layer opt (B); where B T represents the transpose of B, ||·||2 represents the 2-norm, ||·|| F represents the F-norm, α represents the constraint coefficient, I D represents the identity matrix, z j Represents the hyperspectral image after being encoded by the hidden layer of the autoencoder network and A j The corresponding intermediate output vector;

[0019] Step 205: Using a computer The total loss function L of the autoencoder network embedded in the robust subspace recovery layer is obtained total .

[0020] The above-mentioned hyperspectral anomaly detection method embedded in the robust subspace recovery layer is characterized in that: the dimension of the linear transformation matrix B in step 204 is d×D, d represents the number of neurons in the robust subspace recovery hidden layer, D represents the number of neurons in the third hidden layer, the constraint coefficient α is 0<α<1, I D Represents the D×D identity matrix.

[0021] The above-mentioned hyperspectral anomaly detection method embedded with a robust subspace recovery layer is characterized in that: in step 301, the learning rate β in the gradient descent algorithm is set to 0<β<1;

[0022] The preset number of iterative training times in step 302 is 50 to 1000 times.

[0023] Compared with the prior art, the present invention has the following advantages:

[0024] 1. The method of the present invention has simple steps and reasonable design, and solves the problem that the potential representation of the autoencoder network currently used for hyperspectral anomaly detection cannot effectively distinguish between background and anomaly subspaces.

[0025] 2. The total loss function adopted by the present invention takes into account the reconstruction loss of the autoencoder network, the reconstruction loss of the robust subspace recovery layer, and the loss of the linear transformation matrix, thereby facilitating the use of the total loss function for subsequent anomaly detection iterations and effectively detecting anomalies in hyperspectral images.

[0026] 3. In the present invention, the hyperspectral data is input into the autoencoder network. The hidden layer encoding in the autoencoder network maps the hyperspectral data to the latent space. The robust subspace recovery layer extracts the underlying subspace in the latent space to eliminate outliers far away from the subspace during the anomaly detection training process, thereby realizing anomaly detection and effectively improving the discrimination between background and anomalies.

[0027] In summary, the method of the present invention has simple steps and reasonable design. By embedding a robust subspace recovery layer in the autoencoder network, anomaly detection iteration is performed according to the robust subspace recovery layer and the total loss function of the autoencoder network, the distinction between anomalies and background is increased, thereby improving the detection accuracy and effect.

[0028] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 It is a flowchart of the method of the present invention. DETAILED DESCRIPTION

[0030] like Figure 1 As shown, the hyperspectral anomaly detection method embedded in the robust subspace recovery layer of the present invention includes the following steps:

[0031] Step 1: Build an autoencoder network and embed a robust subspace recovery layer:

[0032] Step 101: Use a computer to construct an autoencoder network; wherein the autoencoder network includes an input layer, five hidden layers, and an output layer, and the five hidden layers from the input layer to the output layer are respectively a first hidden layer, a second hidden layer, a third hidden layer, a fourth hidden layer, and a fifth hidden layer;

[0033] Step 102: embedding a robust subspace recovery layer in the autoencoder network; wherein the robust subspace recovery layer sets a linear transformation matrix B, the robust subspace recovery layer includes a robust subspace recovery hidden layer and a robust subspace recovery output layer, and the third hidden layer of the autoencoder network is connected to the robust subspace recovery hidden layer;

[0034] Step 2: Obtaining the total loss function during the iteration of hyperspectral image anomaly detection:

[0035] Step 3: Iterate anomaly detection on the hyperspectral image to obtain the trained autoencoder network embedded in the robust subspace recovery layer;

[0036] Step 301: Use a computer to use the gradient descent method to calculate the total loss function L total Iterative anomaly detection on hyperspectral images;

[0037] Step 302: Repeat step 301 multiple times until the preset number of iterative training times is met, and a trained autoencoder network embedded with a robust subspace recovery layer is obtained.

[0038] In this embodiment, the number of neurons in the input layer and the output layer of the autoencoder network in step 101 is the same as the number of bands of the hyperspectral image.

[0039] In this embodiment, the specific process of obtaining the total loss function during the iteration of hyperspectral image anomaly detection in step 2 is as follows:

[0040] Step 201: Set the size of the hyperspectral image A to m×n×I, and the input vector formed by the I bands at the j-th pixel in the hyperspectral image is recorded as A j ,and in, Represents the data corresponding to the i-th band at the j-th pixel, j is a positive integer, and the value range of j is 1 to mn; i and I are both positive integers, and 1≤i≤I, m represents the number of rows of the hyperspectral image, n represents the number of columns of the hyperspectral image, and I is the number of bands;

[0041] Step 202: Using a computer Get the reconstruction loss of the autoencoder network in, It represents the data corresponding to the i-th band at the j-th pixel after the hyperspectral image is reconstructed by the autoencoder network, D1(A j ) represents the hyperspectral image reconstructed by the autoencoder network and A j The corresponding output vector, ||·||2 represents the 2-norm;

[0042] Step 203: Using a computer Get the reconstruction loss of the robust subspace recovery layer Among them, among them, represents the data corresponding to the i-th band at the j-th pixel after the hyperspectral image is encoded by the autoencoder network and reconstructed by the robust subspace recovery layer, D2(A j ) represents the hyperspectral image after encoding by the autoencoder network and reconstruction by the robust subspace recovery layer and A j The corresponding output vector, ||·||2 represents the 2-norm;

[0043] Step 204: Using a computer Get the linear transformation matrix loss L of the robust subspace recovery layer opt (B); where B T represents the transpose of B, ||·||2 represents the 2-norm, ||·|| F represents the F-norm, α represents the constraint coefficient, I Drepresents the identity matrix, z j Represents the hyperspectral image after being encoded by the hidden layer of the autoencoder network and A j The corresponding intermediate output vector;

[0044] Step 205: Using a computer The total loss function L of the autoencoder network embedded in the robust subspace recovery layer is obtained total .

[0045] In this embodiment, the dimension of the linear transformation matrix B in step 204 is d×D, where d represents the number of neurons in the robust subspace recovery hidden layer, D represents the number of neurons in the third hidden layer, and the constraint coefficient α is set to 0<α<1. D Represents the D×D identity matrix.

[0046] In this embodiment, the learning rate β in the gradient descent algorithm in step 301 is set to 0<β<1;

[0047] The preset number of iterative training times in step 302 is 50 to 1000 times.

[0048] In this embodiment, a computer is used to construct an autoencoder network and a robust subspace recovery layer using the Pytorch deep learning framework.

[0049] In this embodiment, the numbers of neurons in the first hidden layer, the second hidden layer, the third hidden layer, the fourth hidden layer, and the fifth hidden layer are 128, 64, 32, 64, and 128, respectively.

[0050] In this embodiment, the numbers of neurons in the robust subspace recovery hidden layer and the robust subspace recovery output layer are 10 and 189, respectively.

[0051] In this embodiment, the number of neurons in the robust subspace recovery output layer and the number of neurons in the input layer and output layer of the autoencoder network are the same as the number of bands of the hyperspectral image.

[0052] In this embodiment, in actual use, the number of neurons in the robust subspace recovery hidden layer is smaller than the number of neurons in the third hidden layer, so that the robust subspace recovery layer can extract the underlying subspace in the latent space, thereby eliminating outliers far away from the subspace during the anomaly detection training process.

[0053] In this embodiment, the number of neurons in the input layer and the output layer of the autoencoder network is 189.

[0054] In this example, the hyperspectral image was acquired from the San Diego dataset of the Naval Air Station in San Diego, California. A 100×100 pixel portion of the image was captured, containing data from 189 bands in the 370–2510 nm spectral range, with a spatial resolution of 3.5 m. Therefore, the hyperspectral image size is represented by m = 100, n = 100, and I = 189.

[0055] In this embodiment, the activation function in the autoencoder network is selected as the Sigmoid function.

[0056] In this embodiment, the preset number of iterative training times is 100, and the learning rate β is 0.001.

[0057] In this embodiment, it should be noted that after obtaining the trained autoencoder network embedded in the robust subspace recovery layer, the hyperspectral image is input into the autoencoder network of the present invention multiple times, and the hyperspectral image is input into the autoencoder network of the present invention multiple times. Get the total loss value at multiple j-th pixels, and average the total loss value at multiple j-th pixels as the abnormality score. represents the reconstruction loss of the autoencoder network at the jth pixel, represents the reconstruction loss of the robust subspace recovery layer at the j-th pixel, Represents the linear transformation matrix loss of the robust subspace recovery layer at the j-th pixel.

[0058] In this embodiment, the initial values ​​of the parameters of the autoencoder network, the parameters of the robust subspace recovery layer, and the linear transformation matrix are random numbers from a standard normal distribution; subsequently, a computer is used to optimize the parameters of the autoencoder network, the parameters of the robust subspace recovery layer, and the linear transformation matrix during the iterative process of hyperspectral image anomaly detection using the gradient descent method until a trained autoencoder network embedded in the robust subspace recovery layer is obtained.

[0059] In summary, the method of the present invention has simple steps and reasonable design. By embedding a robust subspace recovery layer in the autoencoder network, anomaly detection iteration is performed according to the robust subspace recovery layer and the total loss function of the autoencoder network, the distinction between anomalies and background is increased, thereby improving the detection accuracy and effect.

[0060] The above description is only a preferred embodiment of the present invention and does not limit the present invention in any way. Any simple modification, change and equivalent structural change made to the above embodiment based on the technical essence of the present invention shall still fall within the scope of protection of the technical solution of the present invention.

Claims

1. A hyperspectral anomaly detection method embedded in a robust subspace recovery layer, characterized in that: The method comprises the following steps: Step 1: Build an autoencoder network and embed a robust subspace recovery layer: Step 101: Use a computer to construct an autoencoder network; wherein the autoencoder network includes an input layer, five hidden layers, and an output layer, and the five hidden layers from the input layer to the output layer are respectively a first hidden layer, a second hidden layer, a third hidden layer, a fourth hidden layer, and a fifth hidden layer; Step 102: embedding a robust subspace recovery layer in the autoencoder network; wherein the robust subspace recovery layer sets a linear transformation matrix B, the robust subspace recovery layer includes a robust subspace recovery hidden layer and a robust subspace recovery output layer, and the third hidden layer of the autoencoder network is connected to the robust subspace recovery hidden layer; Step 2: Obtaining the total loss function during the iteration of hyperspectral image anomaly detection: Step 3: Iterate anomaly detection on the hyperspectral image to obtain the trained autoencoder network embedded in the robust subspace recovery layer; Step 301: Use a computer to use the gradient descent method to calculate the total loss function L total Iterative anomaly detection on hyperspectral images; Step 302: Repeat step 301 multiple times until a preset number of iterative training times is met, thereby obtaining a trained autoencoder network embedded with a robust subspace recovery layer; The total loss function is obtained during the iteration of hyperspectral image anomaly detection in step 2. The specific process is as follows: Step 201: Set the size of the hyperspectral image A to m×n×I, and the input vector formed by the I bands at the j-th pixel in the hyperspectral image is recorded as A j ,and in, Represents the data corresponding to the i-th band at the j-th pixel, j is a positive integer, and the value range of j is 1 to mn; i and I are both positive integers, and 1≤i≤I, m represents the number of rows of the hyperspectral image, n represents the number of columns of the hyperspectral image, and I is the number of bands; Step 202: Using a computer Get the reconstruction loss of the autoencoder network in, It represents the data corresponding to the i-th band at the j-th pixel after the hyperspectral image is reconstructed by the autoencoder network, D1(A j ) represents the hyperspectral image reconstructed by the autoencoder network and A j The corresponding output vector, ||·||2 represents the 2-norm; Step 203: Using a computer Get the reconstruction loss of the robust subspace recovery layer Among them, among them, represents the data corresponding to the i-th band at the j-th pixel after the hyperspectral image is encoded by the autoencoder network and reconstructed by the robust subspace recovery layer, D2(A j ) represents the hyperspectral image after encoding by the autoencoder network and reconstruction by the robust subspace recovery layer and A j The corresponding output vector, ||·||2 represents the 2-norm; Step 204: Using a computer Get the linear transformation matrix loss L of the robust subspace recovery layer opt (B); where B T represents the transpose of B, ||·||2 represents the 2-norm, ||·|| F represents the F-norm, α represents the constraint coefficient, I D represents the identity matrix, z j Represents the hyperspectral image after being encoded by the hidden layer of the autoencoder network and A j The corresponding intermediate output vector; Step 205: Using a computer The total loss function L of the autoencoder network embedded in the robust subspace recovery layer is obtained total .

2. The hyperspectral anomaly detection method embedded with a robust subspace recovery layer according to claim 1, characterized in that: In step 101 , the number of neurons in the input layer and the output layer of the autoencoder network is the same as the number of bands of the hyperspectral image.

3. The hyperspectral anomaly detection method embedded with a robust subspace recovery layer according to claim 1, characterized in that: The dimension of the linear transformation matrix B in step 204 is d×D, where d represents the number of neurons in the robust subspace recovery hidden layer, D represents the number of neurons in the third hidden layer, and the constraint coefficient α is 0<α<1. D Represents the D×D identity matrix.

4. The hyperspectral anomaly detection method embedded with a robust subspace recovery layer according to claim 1, characterized in that: In step 301, the learning rate β in the gradient descent algorithm is set to 0<β<1; The preset number of iterative training times in step 302 is 50 to 1000 times.

Images