CN113902973A - Hyperspectral anomaly detection method based on self-encoder and low-dimensional manifold modeling - Google Patents

Abstract

The invention discloses a self-encoder and a hyperspectral anomaly detection method of low-dimensional manifold modeling.A stack type self-encoder is trained through an original hyperspectral image to obtain a trained stack type self-encoder; performing main feature extraction on the original hyperspectral image through the trained stacked self-encoder to finish data dimension reduction and obtain a dimension-reduced hyperspectral image Y; carrying out L times of random sampling on the dimensionality-reduced hyperspectral image Y to obtain a sampled image

1,2, …, L; each of the resulting sampled images is modeled by a low-dimensional manifold

Reconstructing to obtain each characteristic sampling image

Reconstructed background sub-image X of_l(ii) a Reconstructing the background sub-image X_lCalculating the average value

Reconstructed background image as original hyperspectral image

(ii) a By a 1₂Norm determination of original hyperspectral image and reconstructed background image

And taking the residual r as the final abnormal detection result. The invention effectively reduces redundant calculation, improves the overall performance of the algorithm and accelerates the algorithm speed.

Description

Hyperspectral anomaly detection method based on self-encoder and low-dimensional manifold modeling

Technical Field

The invention belongs to the field of remote sensing image processing technology, and particularly relates to a hyperspectral anomaly detection method based on a self-encoder and low-dimensional manifold modeling.

Background

The hyperspectral image is added with a spectrum dimension on the basis of the traditional remote sensing image and is a three-dimensional data structure containing target space characteristics and spectrum information. The hyperspectral image detection technology can be generally divided into target detection and anomaly detection. The characteristic that the abnormal target can be detected without knowing the spectral information of the target object in advance is adopted, so that the hyperspectral abnormality detection is widely applied to various fields such as agriculture, meteorology, military and the like. The study of hyperspectral anomaly detection becomes increasingly important.

The hyperspectral image anomaly detection mainly comprises two steps of background modeling and anomaly detection. The background modeling establishes a background measurement standard by searching an evaluation mode of a background signal in a hyperspectral image, and if the abnormal image is detected, whether each pixel belongs to the background is evaluated according to the background measurement standard, and if the pixel does not belong to the background, the pixel is the abnormal pixel. The quality of background modeling directly affects the final anomaly detection result. Meanwhile, the hyperspectral image has spectrum information of hundreds of wave bands, so that the information redundancy of the hyperspectral image is caused, and the construction of the background is greatly influenced. Therefore, how to remove redundant information in original data, reduce data dimension and reduce processing complexity of anomaly detection is a key point in the field of hyperspectral anomaly detection.

Disclosure of Invention

In view of the above, the main objective of the present invention is to provide a hyper-spectral anomaly detection method based on an encoder and low-dimensional manifold modeling.

In order to achieve the purpose, the technical scheme of the invention is realized as follows:

the embodiment of the invention provides a hyperspectral anomaly detection method based on a self-encoder and low-dimensional manifold modeling, which is characterized by comprising the following steps:

training the stacked self-encoder through the original hyperspectral image to obtain a trained stacked self-encoder;

performing main feature extraction on the original hyperspectral image through the trained stacked self-encoder to finish data dimension reduction and obtain a dimension-reduced hyperspectral image Y;

carrying out L times of random sampling on the dimensionality-reduced hyperspectral image Y to obtain a sampled image

l＝1,2,…,L；

Each of the resulting sampled images is modeled by a low-dimensional manifold

Reconstructing to obtain each characteristic sampling image

Reconstructed background sub-image X of_l(ii) a Reconstructing the background sub-image X_lCalculating the average value

Reconstructed background image as original hyperspectral image

By a 1₂Norm determination of original hyperspectral image and reconstructed background image

And taking the residual r as the final abnormal detection result.

In the above scheme, the stacked self-encoder is trained through the original hyperspectral image, and the trained stacked self-encoder is obtained, specifically:

1.1, the size of an original hyperspectral image is MxNxB, wherein M represents the number of rows contained in the hyperspectral image, N represents the number of columns contained in the hyperspectral image, and B is the number of wavebands;

step 1.2, reducing the dimension of the hyperspectral image by adopting a stacked self-encoder, and using a mean square error as a loss function and a ReLU function as an activation function during training;

step 1.3, training a first single-layer self-encoder, fixing the number of coding units of the first encoder, namely the neuron number of a first Hidden layer, to 100, so that the reconstruction of the original input is completed, freezing parameters obtained in the training process of the first single-layer self-encoder, and taking output Hidden 1 of the Hidden layer of the first single-layer self-encoder as the input of a second single-layer self-encoder;

step 1.4, training a second single-layer self-encoder to complete reconstruction of Hidden 1, wherein the number of encoding units of the second encoder, namely the number of neurons of a second Hidden layer, is fixed to 50;

and step 1.5, integrating the two trained self-encoders, training the whole self-encoders, setting the number of encoding units of a third encoder of the self-encoders, namely the number of neurons of a third hidden layer to be 25, finely adjusting parameters of each layer to obtain a stacked self-encoder, and finely adjusting the parameters of each layer to obtain the trained stacked self-encoder.

In the scheme, the trained stacked self-encoder is used for extracting the main features of the original hyperspectral image to finish data dimension reduction and obtain a dimension-reduced hyperspectral image Y, and specifically, the original hyperspectral image Y belongs to R^M×N×BAs input, the main characteristics of the original hyperspectral image are extracted through a trained stacked self-encoder to finish dataReducing the dimension to obtain a dimension-reduced hyperspectral image Y epsilon R^M×N×nN is the dimensionality of the reduced feature, and n is equal to 25.

In the above scheme, each obtained sampling image is subjected to low-dimensional manifold modeling

Reconstructing to obtain each characteristic sampling image

Reconstructed background sub-image X of_lThe method specifically comprises the following steps:

step 4.1, initializing reconstructed background sub-image X⁽⁰⁾；

Step 4.2, from X^(k)Mid-decimation three-dimensional block set PX^(k)，PX^(k)Is defined as X^(k)Medium size is d₁×d₂Set of local three-dimensional blocks of x B, where k denotes the number of iterations, d₁And d₂Representing the spatial dimension of the block, d₁And d₂To 2, the similarity matrix in the spatial domain is calculated as follows:

where σ (u) is a regularization factor representing the distance between the local three-dimensional block u and its z-th nearest local three-dimensional block, z being 10;

step 4.3, assembling new similarity matrix

In which the symbol q_{j}Representing the jth component (in the spatial domain), d, in a three-dimensional block q_s＝d₁×d₂For the spatial dimension, the similarity calculation is truncated into 20 nearest neighbors, namely only 20 corresponding elements with the minimum adjacent distance are reserved in the similarity matrix;

step 4.4, at each spectral band t, updating (X) using the generalized minimum residual method according to the following formula^t)^(k+1):

Where μ 1/r-1, r is the sampling rate 0.05 and λ is the regularization parameter 10⁷，

A set of hyperspectral image space blocks is represented,

is that

In case of X^t(q) projection operator with projection 0, where Ω^tFrom hyperspectral image any wave band t ∈ [ B ]]The subset obtained by random sampling;

step 4.5, the iteration number k is k +1, so that X is X^(k+1)，X^(k)＝X^(k+1)And step 4.2-step 4.4 are circulated, and the circulation is stopped until X converges to obtain a reconstructed background sub-image X_l。

Compared with the prior art, the method comprises the steps of firstly, performing data dimension reduction processing on an original hyperspectral image by adopting a stacked self-encoder, and then performing anomaly detection on the dimension-reduced hyperspectral image by utilizing a low-dimensional manifold modeling anomaly detector; according to the anomaly detection method for the hyperspectral image, provided by the invention, aiming at the problem that a large amount of redundant calculation exists in the low-dimensional manifold modeling process of the hyperspectral image, a self-encoder dimension reduction method based on deep learning is adopted, so that the redundant calculation in the hyperspectral image anomaly detection algorithm based on the low-dimensional manifold modeling is effectively reduced, the overall performance of the algorithm is improved, and the algorithm speed is accelerated.

Drawings

The accompanying drawings, which are included to provide a further understanding of the invention and are incorporated in and constitute a part of this specification, illustrate embodiment(s) of the invention and together with the description serve to explain the invention without limiting the invention. In the drawings:

FIG. 1 is a diagram of a self-encoder network architecture;

FIG. 2 is a diagram of a stacked self-encoder training process;

FIG. 3 is a diagram of a stacked self-encoder;

FIG. 4 is a flow chart of background reconstruction with multiple random samples according to an embodiment of the present invention;

FIG. 5 is a schematic representation of a hyperspectral image object in an example of the invention;

FIG. 6 is a graph showing the results of anomaly detection in the example of the present invention;

fig. 7 is a block flow diagram in an example of the invention.

Detailed Description

In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is described in further detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

It should be noted that, in this document, the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, article, or apparatus that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, article, or apparatus. Without further limitation, an element defined by the phrase "comprising an … …" does not exclude the presence of other like elements in a process, article, or apparatus that comprises the element.

The embodiment of the invention provides a hyperspectral anomaly detection method based on a self-encoder and low-dimensional manifold modeling, which is realized by the following steps:

step 1, training a stacked self-encoder through an original hyperspectral image to obtain a trained stacked self-encoder;

step 2, extracting main features of the original hyperspectral image through the trained stacked self-encoder to finish data dimension reduction and obtain a dimension-reduced hyperspectral image Y;

step 3, carrying out random sampling on the hyperspectral image Y after dimension reduction for L times to obtain a sampled image

l＝1,2,…,L；

Step 4, each obtained sampling image is subjected to low-dimensional manifold modeling

Reconstructing to obtain each characteristic sampling image

Reconstructed background sub-image X of_l(ii) a Reconstructing the background sub-image X_lCalculating the average value

Reconstructed background image as original hyperspectral image

Step 5, passing through₂Norm determination of original hyperspectral image and reconstructed background image

And taking the residual r as the final abnormal detection result.

Wherein, the step 1 specifically comprises the following steps:

1.1, the size of an original hyperspectral image is MxNxB, wherein M represents the number of rows contained in the hyperspectral image, N represents the number of columns contained in the hyperspectral image, and B is the number of wavebands;

step 1.2, reducing the dimension of the hyperspectral image by adopting a stacked self-encoder, and using a mean square error as a loss function and a ReLU function as an activation function during training;

step 1.3, training a first single-layer self-encoder, fixing the number of coding units of the first encoder, namely the neuron number of a first Hidden layer, to 100, so that the reconstruction of the original input is completed, freezing parameters obtained in the training process of the first single-layer self-encoder, and taking output Hidden 1 of the Hidden layer of the first single-layer self-encoder as the input of a second single-layer self-encoder;

step 1.4, training a second single-layer self-encoder to complete reconstruction of Hidden 1, wherein the number of encoding units of the second encoder, namely the number of neurons of a second Hidden layer, is fixed to 50;

and step 1.5, integrating the two trained self-encoders, training the whole self-encoders, setting the number of encoding units of a third encoder of the self-encoders, namely the number of neurons of a third hidden layer to be 25, finely adjusting parameters of each layer to obtain a stacked self-encoder, and finely adjusting the parameters of each layer to obtain the trained stacked self-encoder.

Wherein the step 2 is to specifically determine that the original hyperspectral image Y belongs to R^M×N×BAs input, performing main feature extraction on the original hyperspectral image through a trained stacked self-encoder to finish data dimension reduction, and obtaining the dimension-reduced hyperspectral image Y belonging to R^M×N×nN is the dimensionality of the reduced feature, and n is equal to 25.

Wherein, the step 4 specifically comprises the following steps:

step 4.1, initializing reconstructed background sub-image X⁽⁰⁾；

Step 4.2, from X^(k)Mid-decimation three-dimensional block set PX^(k)，PX^(k)Is defined as X^(k)Medium size is d₁×d₂Set of local three-dimensional blocks of x B, where k denotes the number of iterations, d₁And d₂Representing the spatial dimension of the block, d₁And d₂To 2, the similarity matrix in the spatial domain is calculated as follows:

where σ (u) is a regularization factor representing the distance between the local three-dimensional block u and its z-th nearest local three-dimensional block, z being 10;

step 4.3, assembling new similarity matrix

In which the symbol q_{j}Representing the jth component (in the spatial domain), d, in a three-dimensional block q_s＝d₁×d₂For the spatial dimension, the similarity calculation is truncated into 20 nearest neighbors, namely only 20 corresponding elements with the minimum adjacent distance are reserved in the similarity matrix;

step 4.4, at each spectral band t, updating (X) using the generalized minimum residual method according to the following formula^t)^(k+1):

Where μ is 1/r-1, r is the sampling rate taken to be 0.05, λ is the regularization parameter equal to 10,

a set of hyperspectral image space blocks is represented,

is that

In case of X^t(q) projection operator with projection 0, where Ω^tFrom hyperspectral image any wave band t ∈ [ B ]]The subset obtained by random sampling;

step 4.5, the iteration number k is k +1, so that X is X^(k+1)，X^(k)＝X^(k+1)And step 4.2-step 4.4 are circulated, and the circulation is stopped until X converges to obtain a reconstructed background sub-image X_l；

Step 4.6, sampling each characteristic sample image

Reconstructed background sub-image X of_lCalculating the average value

As a sourceStarting a reconstructed background image of the hyperspectral image.

To demonstrate the effectiveness of the method of the invention, the anomaly detection was performed using real hyperspectral images. In the experiment, sub-images of 90 × 90 in size and 186 in wave band number in a scene image of san diego (san diego) of an airport, which is acquired by an onboard visible/infrared imaging spectrometer (AVIRIS) sensor, were used for abnormality detection, and three airplanes in the sub-images were used as abnormality targets.

As shown in fig. 7, the basic flow of anomaly detection for the sandiog image data is as follows:

step 101, an original hyperspectral image is a 90 × 90 × 186 three-dimensional matrix, and the total number of the spatial pixel points is 90 × 90, the spectral pixel points include spectral data of a ground target in 186 wave bands, and a one-dimensional vector corresponding to each spatial pixel and having the size of 1 × 1 × 186 represents corresponding spectral information.

In step 102, the self-encoder is a data preprocessing neural network, which can implement the dimensionality reduction of data by constraining the number of encoding units to be less than the dimensionality of input data, and comprises an input layer, a hidden layer and an output layer.

In FIG. 1, the conversion process from the input layer to the hidden layer can be regarded as an encoding process, and the relationship between the hidden layer vector h and the input vector x is shown as the following formula, where W is_hWeight matrix representing input layer to hidden layer, b_hRepresenting the bias vector in the hidden layer, s representing the activation function;

h＝s(W_hx+b_h)

the hidden layer to output layer can be considered as a decoding process, and the relationship between the hidden layer vector h and the output vector y is shown as the following formula. In the formula, W_yWeight matrix representing input layer to hidden layer, b_yRepresenting the bias vector in the hidden layer, s representing the activation function;

y＝s(W_yh+b_y)

step 103, similar to the idea of increasing the depth of the neural network to obtain a feature representation of a higher abstraction level of the data, stacking with multiple autoencoders can result in a stacked autoencoder with a deeper network hierarchy, as shown in fig. 3, so the present invention enablesThe dimension reduction of the original hyperspectral image is completed by using a stacked self-encoder, and the training process is shown in FIG. 2. The mean square error is used as a loss function and the ReLU function is used as an activation function in training. In addition, the batch size, the learning rate and the epoch number are parameters in the stack type self-encoder, and the batch size is set to be 200 and the learning rate is set to be 10 respectively in the invention^-3The number of epochs for each of the autocoders is 150.

Step 1.4, training the first single-layer self-encoder, fixing the number of coding units (namely the number of neurons in the first Hidden layer) of the first encoder to 100, so that the reconstruction of the original input is completed, freezing parameters obtained in the training process in the first single-layer self-encoder, and taking the output Hidden 1 of the Hidden layer of the first single-layer self-encoder as the input of the second single-layer self-encoder.

And step 104, training the second single-layer self-encoder to complete the reconstruction of Hidden 1, wherein the number of coding units (namely the number of second Hidden layer neurons) of the second encoder is fixed to 50.

Step 105, integrating the two trained self-encoders, training the whole self-encoders, setting the number of encoding units of a third encoder of the self-encoder, namely the number of neurons of a third hidden layer to be 25, and finely adjusting parameters of each layer to obtain a stacked self-encoder;

106, enabling the original hyperspectral image Y to be in the range of R^M×N×BAs input, extracting effective characteristics of the original hyperspectral image by using the trained stacked self-encoder obtained in the step 1 to obtain the dimension-reduced hyperspectral image Y belonging to R^M×N×nN is the characteristic dimension after dimensionality reduction, and n is equal to 25;

step 107, in order to reduce the influence of the abnormal point on the background reconstruction, the method utilizes the characteristics of sparse distribution and low sampling probability of the abnormal point in the scene, and adopts a multiple random sampling strategy to reduce the influence of the abnormal point on the background reconstruction as shown in fig. 4. For the hyperspectral image Y after dimensionality reduction to be in the middle of R^M×N×nPerforming random sampling for L times to obtain a sampled image

l＝1,2,…,L；

Step 108, initializing the reconstructed background sub-image X⁽⁰⁾。

Step 109, from X^(k)Mid-decimation three-dimensional block set PX^(k)，PX^(k)Is defined as X^(k)Medium size is d₁×d₂Set of local three-dimensional blocks of x B, where k denotes the number of iterations, d₁And d₂Representing the spatial dimension of the block, d is set in the present invention₁And d₂To 2, the similarity matrix in the spatial domain is calculated as follows.

Where σ (u) is a regularization factor, representing the distance between the local three-dimensional block u and its z-th nearest neighbor, and z is set to 10 in the present invention.

Step 110, assembling a new similarity matrix

In which the symbol q_{j}Representing the jth component (in the spatial domain), d, in a three-dimensional block q_s＝d₁×d₂Is the spatial dimension. At the same time, the similarity matrix

The similarity matrix is an asymmetric matrix, and in order to reduce the computational complexity, the similarity calculation is truncated into 20 nearest neighbors, namely, only 20 corresponding elements with the minimum adjacent distance are reserved in the similarity matrix.

Step 111, at each spectral band t, updating (X) using the generalized minimum residual method according to the following formula^t)^(k+1)。

Where μ is 1/r-1, r is the sampling rate taken to be 0.05, λ is the regularization parameter equal to 10,

a set of hyperspectral image space blocks is represented,

is that

In case of X^t(q) projection operator with projection 0, where Ω^tFrom hyperspectral image any wave band t ∈ [ B ]]To randomly sample the resulting subset.

Step 112, the iteration number k is k +1, so that X is X^(k+1)，X^(k)＝X^(k+1)And step 109-step 111 are circulated, and the circulation is stopped until X converges to obtain a reconstructed background image X_l。

Step 113, sampling each feature into an image

Reconstructed background sub-image X of_lCalculating the average value

As a reconstructed background image of the original hyperspectral image.

Step 114, use l₂Norm calculation of original hyperspectral image and reconstructed background image

And taking the residual error r as the final abnormal detection result.

FIG. 6 shows the GRX method, the RPCA-RX method and the abnormal detection results of the method of the present invention, it can be seen that the presence of the detection target is difficult to see in the detection results of the GRX method and the RPCA-RX method, the target in the detection results of the method of the present invention is most obvious, and the method of the present invention can better detect the abnormal target information in the hyperspectral image.

The above description is only a preferred embodiment of the present invention, and is not intended to limit the scope of the present invention.

Claims (4)

1. A hyperspectral anomaly detection method based on a self-encoder and low-dimensional manifold modeling is characterized by comprising the following steps:

training the stacked self-encoder through the original hyperspectral image to obtain a trained stacked self-encoder;

performing main feature extraction on the original hyperspectral image through the trained stacked self-encoder to finish data dimension reduction and obtain a dimension-reduced hyperspectral image Y;

carrying out L times of random sampling on the dimensionality reduced hyperspectral image Y to obtain a sampled image O _l ，l＝1,2,…,L；

Each of the resulting sampled images O is modeled by a low-dimensional manifold _l Reconstructing to obtain each characteristic sampling image O _l Reconstructed background sub-image X of_l(ii) a Reconstructing the background sub-image X_lCalculating the average value

Reconstructed background image as original hyperspectral image

By a 1₂Norm determination of original hyperspectral image and reconstructed background image

And taking the residual r as the final abnormal detection result.

2. The self-encoder and low-dimensional manifold modeling hyperspectral anomaly detection method according to claim 1, wherein the stacked self-encoder is trained through an original hyperspectral image to obtain a trained stacked self-encoder, specifically:

1.1, the size of an original hyperspectral image is MxNxB, wherein M represents the number of rows contained in the hyperspectral image, N represents the number of columns contained in the hyperspectral image, and B is the number of wavebands;

step 1.2, reducing the dimension of the hyperspectral image by adopting a stacked self-encoder, and using a mean square error as a loss function and a ReLU function as an activation function during training;

step 1.3, training a first single-layer self-encoder, fixing the number of coding units of the first encoder, namely the neuron number of a first Hidden layer, to 100, so that the reconstruction of the original input is completed, freezing parameters obtained in the training process of the first single-layer self-encoder, and taking output Hidden 1 of the Hidden layer of the first single-layer self-encoder as the input of a second single-layer self-encoder;

step 1.4, training a second single-layer self-encoder to complete reconstruction of Hidden 1, wherein the number of encoding units of the second encoder, namely the number of neurons of a second Hidden layer, is fixed to 50;

and step 1.5, integrating the two trained self-encoders, training the whole self-encoders, setting the number of encoding units of a third encoder of the self-encoders, namely the number of neurons of a third hidden layer to be 25, finely adjusting parameters of each layer to obtain a stacked self-encoder, and finely adjusting the parameters of each layer to obtain the trained stacked self-encoder.

3. The hyperspectral anomaly detection method by the self-encoder and the low-dimensional manifold modeling according to claim 1 or 2, characterized in that the trained stacked self-encoder performs main feature extraction on an original hyperspectral image to complete data dimension reduction and obtain a dimension-reduced hyperspectral image Y, specifically, the original hyperspectral image Y belongs to R^M×N×BAs input, performing main feature extraction on the original hyperspectral image through a trained stacked self-encoder to finish data dimension reduction, and obtaining the dimension-reduced hyperspectral image Y belonging to R^M×N×nN is the dimensionality of the reduced feature, and n is equal to 25.

4. The self-encoder and low-dimensional manifold-modeled hyperspectral anomaly detection method according to claim 3, wherein the low-dimensional manifold modeling is used for each sample obtainedImage O _l Reconstructing to obtain each characteristic sampling image O _l Reconstructed background sub-image X of_lThe method specifically comprises the following steps:

step 4.1, initializing reconstructed background sub-image X⁽⁰⁾；

Step 4.2, from X^(k)Mid-decimation three-dimensional block set PX^(k)，PX^(k)Is defined as X^(k)Medium size is d₁×d₂Set of local three-dimensional blocks of x B, where k denotes the number of iterations, d₁And d₂Representing the spatial dimension of the block, d₁And d₂To 2, the similarity matrix in the spatial domain is calculated as follows:

where σ (u) is a regularization factor representing the distance between the local three-dimensional block u and its z-th nearest local three-dimensional block, z being 10;

step 4.3, assembling new similarity matrix

In which the symbol q_{j}Representing the jth component (in the spatial domain), d, in a three-dimensional block q_s＝d₁×d₂For the spatial dimension, the similarity calculation is truncated into 20 nearest neighbors, namely only 20 corresponding elements with the minimum adjacent distance are reserved in the similarity matrix;

step 4.4, at each spectral band t, updating using the generalized minimum residual method according to the following formula

Where μ 1/r-1, r is the sampling rate 0.05 and λ is the regularization parameter 10⁷，

Representing a block of hyperspectral image spaceIn the collection of the images, the image data is collected,

is that

In case of X^t(q) projection operator with projection 0, where Ω^tFrom hyperspectral image any wave band t ∈ [ B ]]The subset obtained by random sampling;

step 4.5, the iteration number k is k +1, so that X is X^(k+1)，X^(k)＝X^(k+1)And step 4.2-step 4.4 are circulated, and the circulation is stopped until X converges to obtain a reconstructed background sub-image X_l。

Images