CN114202539A - Hyperspectral image anomaly detection method based on end-to-end RX - Google Patents

Hyperspectral image anomaly detection method based on end-to-end RX Download PDF

Info

Publication number
CN114202539A
CN114202539A CN202111556995.6A CN202111556995A CN114202539A CN 114202539 A CN114202539 A CN 114202539A CN 202111556995 A CN202111556995 A CN 202111556995A CN 114202539 A CN114202539 A CN 114202539A
Authority
CN
China
Prior art keywords
hyperspectral
network
vae
hidden layer
anomaly detection
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202111556995.6A
Other languages
Chinese (zh)
Inventor
刘芳
刘嘉
肖亮
杨劲翔
张安迪
郜文菲
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Nanjing University of Science and Technology
Original Assignee
Nanjing University of Science and Technology
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Nanjing University of Science and Technology filed Critical Nanjing University of Science and Technology
Priority to CN202111556995.6A priority Critical patent/CN114202539A/en
Publication of CN114202539A publication Critical patent/CN114202539A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/0002—Inspection of images, e.g. flaw detection
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00—Machine learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T5/00—Image enhancement or restoration
    • G06T5/20—Image enhancement or restoration using local operators
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20024—Filtering details

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Software Systems (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Artificial Intelligence (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Medical Informatics (AREA)
  • Quality & Reliability (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Mathematical Physics (AREA)
  • Image Analysis (AREA)

Abstract

The invention discloses a hyperspectral image anomaly detection method based on end-to-end RX, which comprises the following steps: performing feature learning on each pixel point of the hyperspectral image by using variational self-coding; extracting hidden layer characteristics of the variational self-coding network as characteristic representation of hyperspectral pixels; estimating a local mean value and a local covariance matrix of characteristic representation of each pixel point in the hyperspectral image; constructing a differentiable RX anomaly detection algorithm to obtain an anomaly index; calculating a counter propagation mode of the differentiable RX structure; constructing a hyperspectral anomaly detection network structure based on end-to-end RX; fusing a variational self-coding network and an end-to-end RX loss function and training; and inputting the hyperspectral image into the trained network and outputting an abnormal detection result. The method has the capability of effectively fusing the feature space of the hidden layer of the VAE network and the RX abnormal feature information, and has excellent performance when being applied to a hyperspectral image abnormality detection task.

Description

Hyperspectral image anomaly detection method based on end-to-end RX
Technical Field
The invention relates to a hyperspectral image anomaly detection technology, in particular to a hyperspectral image anomaly detection method based on end-to-end RX.
Background
A hyperspectral image (HSI) is remote sensing data that contains both rich spectral band information and spatial information, and is usually represented as a three-dimensional cubic matrix. In view of the HSI's ability to detect anomalous ground objects (man-made small objects such as airplanes and buildings) pixel by pixel, it is rapidly developing in many fields such as geological surveying, environmental monitoring, urban planning, crop evaluation, etc. Conventional hyperspectral image anomaly detection algorithms include a detector (CRD) based on collaborative representation, a detector (AED) based on attribute and edge preservation, a detector (LRDM-MoG) based on low rank and sparse representation, and the like. With the rapid development of deep learning, the HIS anomaly detection based on the deep learning network has become one of the important research contents in this field.
An Auto Encoder (AE), as a typical self-supervision deep network learning model, can perform dimension reduction, denoising and the like on non-structural data such as images and the like, and extract the most representative information in original data while reducing the amount of input data, so that spectral band information and spatial information of hyperspectral data can be effectively extracted, and the model is widely paid attention to by researchers. However, when data is modeled by AE, the importance of different data often cannot be distinguished. In order to improve the robustness and expansibility of the model, researchers have proposed a variational self-encoder (VAE) for learning data distribution mapping, which limits the probability distribution of its implicit vectors to a standard normal distribution. The VAE is a generation model based on probability statistics, can capture structural changes of data such as images and can also mine abstract features of high-dimensional data, so that spectral band information and spatial information of the high-spectral data can be effectively extracted. In the hyperspectral anomaly detection algorithm based on AE and VAE, researchers usually use reconstruction errors as anomaly indicators of each pixel point. In Lu et al [ Lu X, Zhang W, Huang J.Exploiting Embedding Manual of Autoencoders for Hyperspectral analysis Detection [ J ]. IEEE Transactions on Geoscience and Remote Sensing,2020,58(3): 1527) 1537 ] the MC-AEN is designed by combining AE and Manifold learning, the internal structure of Hyperspectral data is preserved by adding Manifold constraint in the hidden layer, and the global reconstruction error and the local implicit characteristic error are fused to obtain the final Anomaly Detection result. However, this type of method only emphasizes the approximation of the data, and fails to fully utilize the abnormal target information in the hyperspectral data.
On the other hand, the anomaly detection algorithm of the conventional Reed-Xiaooli (RX) algorithm is a typical method based on a probability statistical model. Background pixels in a hyperspectral image are described through a Gaussian probability distribution model, and the Mahalanobis distance between each pixel and the background pixels is calculated, so that abnormal pixels and background pixels are distinguished. In the process of estimating and counting distribution parameters, a learner constructs a Global RX (GRX) anomaly detection algorithm by using global information of a hyperspectral image, and can effectively capture high-strength anomaly targets in the whole area; considering that a large number of local anomalies or small target anomalies exist in the hyperspectral image, local neighborhood information can be used for estimating statistical distribution parameters of background pixels, namely a Local RX (LRX) anomaly detection algorithm is adopted, the background pixels of a local area can be better described, and anomalous targets with small internal shapes of the area are excavated. In order to improve the computational efficiency of RX algorithm, ZHEN et al [ Y ZHEN, Y Li, Y Shi, et al. Acceleration scheme of RXD algorithm based on FPGAfor hyperspectral and target detection [ J ]. Beijing Hangkong antenna Daxue Xuebao/Journal of Beijing University of Aeronautics and Astronautics,2018,44(12): 2556) propose FPGA-based fast RX algorithm, which calculates covariance matrix and its inverse by block-parallel and QR decomposition. However, RX and its various derived algorithms do not have the ability to learn, and cannot effectively capture the spectral band and spatial information content of abnormal targets in different regions.
Disclosure of Invention
The invention discloses a hyperspectral image anomaly detection method based on end-to-end RX, which has the capability of extracting hyperspectral image anomalous pixels in the process of variational self-coding learning of high-level features, and has excellent performance when being applied to a hyperspectral image anomaly detection task.
The technical solution for realizing the purpose of the invention is as follows: a hyperspectral image next anomaly detection method based on end-to-end RX comprises the following steps:
firstly, performing feature learning on all pixel points in a hyperspectral image by using a variational self-coding network to remove redundant information, namely using high-dimensional hyperspectral data as training data of a VAE (virtual reality), and executing a VAE training process by using a random gradient descent method;
secondly, extracting independent same-distribution hidden layer data of each node as characteristic representation of the hyperspectral pixel points according to output data of the hyperspectral pixel points on the VAE hidden layer;
thirdly, estimating a local mean value and a local variance of the hidden layer feature representation data, namely performing mean filtering and square mean filtering on a hidden layer feature map of the hyperspectral image;
fourthly, constructing a differentiable RX anomaly detection algorithm to obtain an anomaly index, namely calculating the mahalanobis distance on a hidden layer feature map of the VAE, performing K & I threshold segmentation after normalization to obtain a threshold parameter, and further calculating the anomaly index of each pixel point;
fifthly, constructing a hyperspectral anomaly detection network structure based on end-to-end RX, namely constructing a network module for realizing differentiable RX, and embedding the module into a VAE network;
sixthly, fusing a preliminary abnormal detection result and an end-to-end RX loss function, namely taking the preliminary detection result as a weight coefficient of each pixel point in the hyperspectral image in the end-to-end RX network;
and seventhly, training the network by using a random gradient descent method and outputting a final abnormal detection result.
An electronic device includes a memory, a processor, and a computer program stored on the memory and executable on the processor, and the processor implements the end-to-end RX-based hyper-spectral map next anomaly detection method when executing the program.
A computer-readable storage medium, on which a computer program is stored, which, when executed by a processor, implements the end-to-end RX-based hyper-spectral map next anomaly detection method described above.
Compared with the prior art, the invention has the remarkable characteristics that: (1) learning the mapping relation of the hyperspectral image from a high-dimensional redundant data space to a low-dimensional hidden layer independent distribution space based on VAE; (2) establishing a differentiable RX anomaly detection module on the implicit characteristic diagram to obtain an anomaly index of each pixel point; (3) embedding a differentiable RX module into a VAE network to form an end-to-end RX-based abnormity detection network; (4) the network has the advantages of simple structure, small calculated amount, high training speed, high reasoning efficiency, high abnormality detection precision and the like.
The present invention is described in further detail below with reference to the attached drawing figures.
Drawings
FIG. 1 is a block diagram of the process of the present invention.
Fig. 2 is the classification results of different methods on the hydie dataset, (a) pseudo-color map, (b) group route, (c) GRX, (d) LRX, (e) PCA-LRX, (f) CRD, (g) AED.
Figure 3 is the classification results of different methods on the AVIRIS dataset. (a) False-color image (b) group route, (c) GRX, (d) LRX, (e) PCA-LRX, (f) CRD, (g) AED, (h) RX-VAE.
FIG. 4 is the classification result of different methods on ABU-urban data set. (a) False-color image (b) group route, (c) GRX, (d) LRX, (e) PCA-LRX, (f) CRD, (g) AED, (h) RX-VAE.
FIG. 5 is the classification result of different methods on the ABU-airport data set. (a) False-color image (b) group route, (c) GRX, (d) LRX, (e) PCA-LRX, (f) CRD, (g) AED, (h) RX-VAE.
FIG. 6 is the classification results of different methods on the ABU-beach dataset. (a) False-color image (b) group route, (c) GRX, (d) LRX, (e) PCA-LRX, (f) CRD, (g) AED, (h) RX-VAE.
FIG. 7 is a ROC curve on the ABU-airport data set.
Detailed Description
In view of the problems of the existing anomaly detection algorithm, the invention provides that the VAE is utilized to learn the implicit characteristics of the independent and same distribution of the hyperspectral data, a differentiable RX algorithm is designed to screen the anomaly information in the implicit characteristic diagram, further the VAE is guided to retain more anomaly information in training, and the anomaly detection performance of the hyperspectral image is improved. The method can effectively integrate the advantages of the VAE network and the traditional RX anomaly detection algorithm, and can keep anomaly information in an anomaly index weighting mode while mining the inherent characteristics of high-dimensional and high-redundancy hyperspectral data, so that the model has high robustness, strong expansibility and good anomaly detection effect, and can be widely applied to the field of related engineering. Experimental results show that the detection result graph and the detection accuracy of the method are superior to those of the traditional method in reference data sets Hyperspectral Digital image Collection Experimental (HYDICE), Airborne visual/extracted Imaging Spectrometers (AVIRIS) and Airport-Beach-Urban (ABU). The method has the capability of effectively fusing the feature space of the hidden layer of the VAE network and the RX abnormal feature information, and has excellent performance when being applied to a hyperspectral image abnormality detection task.
The following detailed description of the implementation of the present invention, with reference to fig. 1, includes the following steps:
firstly, performing feature learning on all pixel points in the hyperspectral image by using a variational self-coding network to remove redundant information, namely using high-dimensional hyperspectral data as training data of the variational self-coding network (VAE). Memory column vector
Figure BDA0003419122250000041
Is a pixel point in the hyperspectral image, a matrix
Figure BDA0003419122250000042
The method comprises the steps of collecting training samples of all pixel points, wherein B represents the wave band number of hyperspectral data, and M multiplied by N represents the size of a hyperspectral image, namely height and width. Firstly, the hyperspectral data is normalized to obtain
Figure BDA0003419122250000043
Then, the high spectrum pixel point matrix is processed
Figure BDA0003419122250000044
Into a VAE network, wherein the encoder comprises a circuitTwo 1 x 1 convolutional layers for capturing spectral band information at a 3 x 3 convolutional layer for capturing spatial features, where the output of each node in the hidden layer is in [ mu ] s1，μ2，…，μJ；σ1，σ2，…，σJ]In this representation, J corresponds to the dimension of the hidden layer feature representation. Finally, the output of the decoder is combined
Figure BDA0003419122250000045
As reconstruction data of the hyperspectral pixel point x, training is carried out through the following loss function, specifically
Figure BDA0003419122250000046
Wherein the first term and the second term correspond to the reconstruction error and the KL divergence, sigma, respectivelyjAnd mujRespectively, representing the corresponding outputs of the hidden layer nodes.
And secondly, extracting independent same-distribution hidden layer data of each node as characteristic representation of the hyperspectral pixel points according to output data of the hyperspectral pixel points on the VAE hidden layer. First, the encoder and decoder of the VAE use q, respectivelyφ(z | x) and pθ(x | z) where z is a characteristic representation of the hidden layer, the subscripts φ and
Figure BDA0003419122250000051
representing the network parameters of the encoder and decoder, respectively. The high-level feature representation z of the hyperspectral pixel points is obtained by the output of hidden layer nodes, specifically
z＝μ+σ⊙ε’
Wherein mu is [ mu ]1，μ2，…，μJ]，σ＝[σ1，σ2，…，σJ]For the output of the hidden layer nodes, ε N (0, I) obeys a standard normal distribution of the J dimension.
And thirdly, estimating a local mean value and a local variance of the hidden layer feature representation data, namely performing mean filtering and square mean filtering on a hidden layer feature map of the hyperspectral image. Firstly, the hyperspectral pixelHigh-level feature map representation of points is noted
Figure BDA0003419122250000052
Z2Representing an element-wise squaring operation of the matrix. Then, for Z and Z separately2A mean filtering operation with a window size of 7 x 7 is performed. In particular, for each element Z in any Z ∈ ZlE Z e Z, the corresponding representation of its neighborhood pixels respectively Zl(k) K is 1,2, K, and the local mean is denoted E zl]Representing local covariance
Figure BDA0003419122250000053
Local variance is given by var: (zl) is expressed by the specific calculation method
Figure BDA0003419122250000054
Figure BDA0003419122250000055
Figure BDA0003419122250000056
Where K denotes the number of neighborhood pixels, where K is 7 × 7 and 49, and the local variance is obtained by subtracting the square of the local mean from the local mean of the square.
And fourthly, constructing a differentiable RX anomaly detection algorithm to obtain an anomaly index, namely calculating the Mahalanobis distance on the hidden layer feature map, performing K & I threshold segmentation after normalization to obtain a threshold parameter, and further calculating the anomaly index of each pixel point. Applying an RX detection algorithm on a high-level feature map Z of the hyperspectral data to obtain a differentiable RX detection algorithm, specifically
Figure BDA0003419122250000057
Wherein d isRXWhich represents the mahalanobis distance,
Figure BDA0003419122250000058
local mean, C, of corresponding background pixelsbThe covariance matrix corresponding to the background pixels,
Figure BDA0003419122250000061
representing a zero mean feature vector.
And fifthly, constructing a hyperspectral anomaly detection network structure based on the end-to-end RX, namely constructing a network module for realizing differentiable RX, and embedding the module into the VAE network. First, let Mean (-) 7 be the Mean filter function with a window size of 7 × 7. The differentiable RX detection is then implemented with a filter function of a local mean and a squared local mean, in particular
Figure BDA0003419122250000062
Then, the mean filtering function is embedded into the VAE network according to the formula, the output layer of the RX detection algorithm is normalized and K is carried out&I, dividing the threshold value to obtain a threshold parameter theta, and then obtaining the abnormal index and d of each pixel pointRXIs in direct proportion, specifically
Figure BDA0003419122250000063
Wherein h isaEach pixel point is represented as an abnormal index, and the probability that each pixel point is an abnormal point is higher if the numerical value is larger. And finally, accessing the normalization layer and the abnormal index layer into the VAE network to form an end-to-end RX abnormal detection network.
And sixthly, fusing the preliminary abnormal detection result and the end-to-end RX loss function, namely taking the preliminary detection result as a weight coefficient of each pixel point in the hyperspectral image in the end-to-end RX network. Firstly, the abnormal index h of each pixel pointaAnd as a preliminary detection result, the weight coefficient is used as the weight coefficient of each pixel point in the VAE network training process. Then theAnd constructing an abnormal index weighted loss function, and guiding the VAE to retain more abnormal information.
In particular to
Figure BDA0003419122250000064
Wherein, the abnormal index of each hyperspectral pixel point x is corresponding haAnd W' are the network parameters of the encoder and decoder, respectively, in the VAE.
And seventhly, training the network by using a random gradient descent method and outputting a final abnormal detection result. Firstly, the encoder parameter and the decoder parameter in the trained network are respectively WFinal (a Chinese character of 'gan')And WFinal (a Chinese character of 'gan')', then the encoder function is f (·; W)Final (a Chinese character of 'gan')) The output of the encoder is muFinal (a Chinese character of 'gan')＝[μ1，μ2，…，μJ]And σ end ═ σ1，σ2，…，σJ]In cascade, i.e.
[μFinal (a Chinese character of 'gan')，σFinal (a Chinese character of 'gan')]＝[μ1，μ2，…，μJ，σ1，σ2，…，σJ]＝f(·；WFinal (a Chinese character of 'gan'))
Then, the high-level feature of the hyperspectral data in the hidden layer passes through a formula zFinal (a Chinese character of 'gan')＝μFinal (a Chinese character of 'gan')+σFinal (a Chinese character of 'gan')And e, sampling and inputting the epsilon into a differentiable RX layer to finally obtain the abnormal index h of each pixel point in the hyperspectral imageaForming a final abnormality detection result graph Ha＝[ha]M×N。
The effect of the invention can be further illustrated by the following simulation experiment: hyperspectral Digital image Collection Experimental (HYDICE), air Vision/incorporated Imaging Spectrometer (AVIRIS) and Airport-Beach-Urban (ABU).
Simulation conditions
The simulation experiment adopts three groups of real hyperspectral data: a HYDICE dataset, an AVIRIS dataset, and an ABU dataset.
The hyperspectral image in the HYDICE data set contains 175 wave bands, the spectral resolution and the spatial resolution are respectively 10nm and 1m, the image size is 80 multiplied by 100, and the covered ground objects are vegetable areas, building areas, roads, automobiles and the like. The AVIRIS data set remote sensing data collected by an airborne visible infrared imaging spectrometer in the San Diego experimental area of the United states comprises 2 hyperspectral images with the size of 100 multiplied by 100 and the wave band number of 189. The ABU data set comprises 13 hyperspectral images, the image size is 100 x 100 or 150 x 150, the wave band numbers are 191, 205, 102, 188, 191, 193, 204, 205 and 207 respectively, and the coverage areas are airports, beaches and urban areas respectively.
For the HYDICE, AVIRIS and ABU data sets, an abnormal detection result graph is mainly displayed in the experiment, and ROC curves of partial detection result graphs are listed. Wherein the ROC curve describes the false positive rate (FPR, P)d) True Positive Rate (TPR, P)f) And a threshold (τ). In the experiment, the method is marked as RX-VAE, and GRX, LRX, PCA-LRX, CRD and AED are selected by a comparison method. The simulation experiment is completed by adopting Python-3.6+ Tensorflow-1.12+ Pythrch-1.1 under the Windows 10 operating system. In addition, in the experiment of the invention, an encoder and a decoder of the VAE network respectively comprise two hidden layers, the number of nodes of an input layer of the decoder and the number of nodes of an output layer of the encoder correspond to the number of spectral segments of each hyperspectral image, the activation functions of all the hidden layers uniformly adopt Leaky ReLU, the network is trained by adopting a stochastic gradient descent method, the learning rates are respectively set to be 0.0005, 0.0002 and 0.0005 in three data sets, and the iteration times are respectively set to be 200, 500 and 600.
Analysis of simulation experiment results
Fig. 2 and fig. 3 are diagrams of abnormal detection results of simulation experiments performed on the hybrid and AVIRIS datasets by the method of the present invention and the comparison method thereof, and fig. 4 to fig. 6 are diagrams of abnormal detection results of three different types of scenes, urban areas, airports and beaches, respectively, in the ABU dataset. From experimental results, the method obtains better effect on three different data sets. As can be seen from the abnormal detection result graphs of different algorithms, the detection result graph of the method can show the abnormal target more clearly, and the background suppression is cleaner. One hyperspectral image is respectively selected from the three data sets, and ROC curves of different comparison methods and detection results of the method are displayed, as shown in FIG. 7. As can be seen from the ROC curve of FIG. 7, the quantitative index of the detection result of the method of the present invention is superior to that of each comparison algorithm. The effectiveness of the method is shown by the simulation experiment results of the three groups of different real hyperspectral image data sets on the comparison algorithm and the method.
The above examples are only intended to illustrate the technical solution of the present invention, but not to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, it will be understood by those of ordinary skill in the art that: the technical solutions described in the foregoing embodiments may still be modified, or some technical features may be equivalently replaced; and such modifications or substitutions do not depart from the spirit and scope of the corresponding technical solutions of the embodiments of the present invention.

Claims (10)

1. A hyperspectral image next anomaly detection method based on end-to-end RX is characterized by comprising the following steps:
firstly, performing feature learning on all pixel points in a hyperspectral image by using a variational self-coding network to remove redundant information, namely using high-dimensional hyperspectral data as training data of a VAE (virtual reality), and executing a VAE training process by using a random gradient descent method;
secondly, extracting independent same-distribution hidden layer data of each node as characteristic representation of the hyperspectral pixel points according to output data of the hyperspectral pixel points on the VAE hidden layer;
thirdly, estimating a local mean value and a local variance of the hidden layer feature representation data, namely performing mean filtering and square mean filtering on a hidden layer feature map of the hyperspectral image;
fourthly, constructing a differentiable RX anomaly detection algorithm to obtain an anomaly index, namely calculating the mahalanobis distance on a hidden layer feature map of the VAE, performing K & I threshold segmentation after normalization to obtain a threshold parameter, and further calculating the anomaly index of each pixel point;
fifthly, constructing a hyperspectral anomaly detection network structure based on end-to-end RX, namely constructing a network module for realizing differentiable RX, and embedding the module into a VAE network;
sixthly, fusing a preliminary abnormal detection result and an end-to-end RX loss function, namely taking the preliminary detection result as a weight coefficient of each pixel point in the hyperspectral image in the end-to-end RX network;
and seventhly, training the network by using a random gradient descent method and outputting a final abnormal detection result.
2. The method for detecting the next anomaly of the hyperspectral image based on end-to-end RX according to claim 1 is characterized in that in the first step, a variational self-coding network is utilized to perform feature learning on all pixel points in the hyperspectral image to remove redundant information, namely, high-dimensional hyperspectral data is used as training data of VAE; memory column vector
Figure FDA0003419122240000011
Is a pixel point in the hyperspectral image, a matrix
Figure FDA0003419122240000012
The method comprises the steps that a training sample set of all pixel points is formed, wherein B represents the wave band number of hyperspectral data, and M multiplied by N represents the size, namely height and width, of a hyperspectral image; firstly, the hyperspectral data is normalized to obtain
Figure FDA0003419122240000013
Then, the high spectrum pixel point matrix is processed
Figure FDA0003419122240000014
Input into a VAE network, wherein the encoder comprises a 3 × 3 convolutional layer for capturing spatial features, two 1 × 1 convolutional layers for capturing spectral band information, and the output of each node in the hidden layer is [ mu ] m1,μ2,…,μJ；σ1,σ2,…,σJ]Representing that J corresponds to the dimension of the hidden layer feature representation; finally, the output of the decoder is combined
Figure FDA0003419122240000015
As reconstruction data of the hyperspectral pixel point x, training is performed through the following loss function, specifically:
Figure FDA0003419122240000021
wherein the first term and the second term correspond to the reconstruction error and the KL divergence, sigma, respectivelyjAnd mujRespectively, representing the corresponding outputs of the hidden layer nodes.
3. The end-to-end RX-based hyperspectral image next anomaly detection method according to claim 1 is characterized in that in the second step, independent same-distribution hidden layer data of each node are extracted as the feature representation of a hyperspectral pixel according to the output data of the hyperspectral pixel in a VAE hidden layer; the method comprises the following specific steps:
encoder and decoder for VAE using q, respectivelyφ(z | x) and pθ(x | z) where z is a characteristic representation of the hidden layer, the subscripts φ and
Figure FDA0003419122240000022
network parameters representing an encoder and a decoder, respectively; the high-level feature representation z of the hyperspectral pixel points is obtained by the output of hidden layer nodes, specifically
z＝μ+σ⊙ε，
Wherein mu is [ mu ]1,μ2,…,μJ]，σ＝[σ1,σ2,…,σJ]For the output of the hidden layer nodes, ε N (0, I) obeys a standard normal distribution of the J dimension.
4. The end-to-end RX based next anomaly detection method for hyperspectral imagery according to claim 1, wherein in the third step, the local mean and local variance of the hidden layer feature representation data are estimated, i.e. mean filtering and square mean filtering are performed on the hidden layer feature map of the hyperspectral imagery; the method comprises the following specific steps:
high-rise characteristic map representation of hyperspectral pixel points is recorded as
Figure FDA0003419122240000023
Z2A element-wise squaring operation representing a matrix;
for Z and Z respectively2Performing a mean filtering operation with a window size of 7 × 7; for each element Z in any Z ∈ ZlE Z e Z, the corresponding representation of its neighborhood pixels respectively Zl(k) K is 1,2, K, and the local mean is denoted E zl]Representing local covariance
Figure FDA0003419122240000024
Var (z) for local variancel) The specific calculation method is shown as
Figure FDA0003419122240000025
Figure FDA0003419122240000026
Figure FDA0003419122240000027
Where K denotes the number of neighborhood pixels, where K is 7 × 7 and 49, and the local variance is obtained by subtracting the square of the local mean from the local mean of the square.
5. The end-to-end RX-based hyperspectral image next anomaly detection method according to claim 1 is characterized in that the fourth step is to construct a differentiable RX anomaly detection algorithm to obtain an anomaly index, namely to calculate Mahalanobis distance on the hidden layer feature map, to perform K & I threshold segmentation after normalization to obtain a threshold parameter, and further to calculate the anomaly index of each pixel point; applying an RX detection algorithm on a high-level feature map Z of hyperspectral data to obtain a differentiable RX detection algorithm, which specifically comprises the following steps:
Figure FDA0003419122240000031
wherein d isRXWhich represents the mahalanobis distance,
Figure FDA0003419122240000032
local mean, C, of corresponding background pixelsbThe covariance matrix corresponding to the background pixels,
Figure FDA0003419122240000033
representing a zero mean feature vector.
6. The end-to-end RX-based hyperspectral image next anomaly detection method according to claim 1 is characterized in that in the fifth step, an end-to-end RX-based hyperspectral anomaly detection network structure is constructed, that is, a network module for realizing differentiable RX is constructed, and the module is embedded into a VAE network, specifically, the following steps are performed:
let Mean (-) 7 be the Mean filter function with window size of 7 × 7; then, the differentiable RX detection algorithm is implemented with a filter function of local mean and squared local mean, in particular
Figure FDA0003419122240000034
Embedding the mean filtering function into VAE network according to the formula, normalizing the output layer of RX detection algorithm and performing K&I, dividing the threshold value to obtain a threshold parameter theta, and then obtaining the abnormal index and d of each pixel pointRXIs in direct proportion, specifically
Figure FDA0003419122240000035
Wherein h isaIndicating that each pixel point is an abnormal index;
and accessing the normalization layer and the anomaly index layer into the VAE network to form an end-to-end RX anomaly detection network.
7. The hyperspectral image next anomaly detection method based on end-to-end RX according to claim 1 is characterized in that the sixth step is to fuse the preliminary anomaly detection result and the loss function of the end-to-end RX, namely to use the preliminary anomaly detection result as the weight coefficient of each pixel point in the hyperspectral image in the end-to-end RX network; the method comprises the following specific steps:
anomaly index h of each pixelaAs a preliminary detection result, the weight coefficient is used as the weight coefficient of each pixel point in the VAE network training process;
constructing a loss function weighted by the abnormal index, and guiding the VAE to reserve more abnormal information; in particular to
Figure FDA0003419122240000041
Wherein, the abnormal index of each hyperspectral pixel point x is corresponding haAnd W' are network parameters of the encoder and decoder, respectively, in the VAE.
8. The hyperspectral image next anomaly detection method based on end-to-end RX according to claim 1 is characterized in that in the seventh step, a network is trained by using a stochastic gradient descent method and a final anomaly detection result is output; the method comprises the following specific steps:
the encoder parameter and the decoder parameter in the network after training are respectively WFinal (a Chinese character of 'gan')And WFinal (a Chinese character of 'gan')', then the encoder function is f (·; W)Final (a Chinese character of 'gan')) The output of the encoder is muFinal (a Chinese character of 'gan')＝[μ1,μ2,…,μJ]And σFinal (a Chinese character of 'gan')＝[σ1,σ2,…,σJ]In cascade, i.e.
[μFinal (a Chinese character of 'gan'),σFinal (a Chinese character of 'gan')]＝[μ1,μ2,…,μJ,σ1,σ2,…,σJ]＝f(·；WFinal (a Chinese character of 'gan'))
The high-level characteristic of the hyperspectral data in the hidden layer passes through a formula zFinal (a Chinese character of 'gan')＝μFinal (a Chinese character of 'gan')+σFinal (a Chinese character of 'gan')And e, sampling and inputting the epsilon into a differentiable RX layer to finally obtain the abnormal index h of each pixel point in the hyperspectral imageaForming a final abnormality detection result graph Ha＝[ha]M×N。
9. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor when executing the program implements an end-to-end RX-based hyperspectral next anomaly detection method according to any of claims 1-8.
10. A computer-readable storage medium, on which a computer program is stored, which, when being executed by a processor, carries out a method for end-to-end RX-based hyperspectral next anomaly detection according to any of claims 1 to 8.
CN202111556995.6A 2021-12-18 2021-12-18 Hyperspectral image anomaly detection method based on end-to-end RX Pending CN114202539A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202111556995.6A CN114202539A (en) 2021-12-18 2021-12-18 Hyperspectral image anomaly detection method based on end-to-end RX

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202111556995.6A CN114202539A (en) 2021-12-18 2021-12-18 Hyperspectral image anomaly detection method based on end-to-end RX

Publications (1)

Publication Number Publication Date
CN114202539A true CN114202539A (en) 2022-03-18

Family

ID=80655217

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202111556995.6A Pending CN114202539A (en) 2021-12-18 2021-12-18 Hyperspectral image anomaly detection method based on end-to-end RX

Country Status (1)

Country Link
CN (1) CN114202539A (en)

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114677574A (en) * 2022-05-26 2022-06-28 杭州宏景智驾科技有限公司 Method and system for diagnosing image fault for automatic driving
CN116612356A (en) * 2023-06-02 2023-08-18 北京航空航天大学 Hyperspectral anomaly detection method based on deep learning network
CN118135205A (en) * 2024-05-06 2024-06-04 南京信息工程大学 Hyperspectral image anomaly detection method
CN118135205B (en) * 2024-05-06 2024-07-16 南京信息工程大学 Hyperspectral image anomaly detection method

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114677574A (en) * 2022-05-26 2022-06-28 杭州宏景智驾科技有限公司 Method and system for diagnosing image fault for automatic driving
CN116612356A (en) * 2023-06-02 2023-08-18 北京航空航天大学 Hyperspectral anomaly detection method based on deep learning network
CN116612356B (en) * 2023-06-02 2023-11-03 北京航空航天大学 Hyperspectral anomaly detection method based on deep learning network
CN118135205A (en) * 2024-05-06 2024-06-04 南京信息工程大学 Hyperspectral image anomaly detection method
CN118135205B (en) * 2024-05-06 2024-07-16 南京信息工程大学 Hyperspectral image anomaly detection method

Similar Documents

Publication Publication Date Title
Wang et al. Auto-AD: Autonomous hyperspectral anomaly detection network based on fully convolutional autoencoder
Xu et al. Joint reconstruction and anomaly detection from compressive hyperspectral images using Mahalanobis distance-regularized tensor RPCA
CN110929607B (en) Remote sensing identification method and system for urban building construction progress
Wang et al. Urban impervious surface detection from remote sensing images: A review of the methods and challenges
CN103456018B (en) Remote sensing image change detection method based on fusion and PCA kernel fuzzy clustering
CN109766858A (en) Three-dimensional convolution neural network hyperspectral image classification method combined with bilateral filtering
CN105069468A (en) Hyper-spectral image classification method based on ridgelet and depth convolution network
CN103208011B (en) Based on average drifting and the hyperspectral image space-spectral domain classification method organizing sparse coding
CN114202539A (en) Hyperspectral image anomaly detection method based on end-to-end RX
CN112990313B (en) Hyperspectral image anomaly detection method and device, computer equipment and storage medium
CN111368691B (en) Unsupervised hyperspectral remote sensing image space spectrum feature extraction method
Feng et al. Local spatial constraint and total variation for hyperspectral anomaly detection
CN103413151A (en) Hyperspectral image classification method based on image regular low-rank expression dimensionality reduction
CN115236655B (en) Landslide identification method, system, equipment and medium based on fully-polarized SAR
CN111986162B (en) Hyperspectral abnormal point rapid detection method based on rough positioning and collaborative representation
Xiang et al. Visual attention and background subtraction with adaptive weight for hyperspectral anomaly detection
Qin et al. Spatiotemporal inferences for use in building detection using series of very-high-resolution space-borne stereo images
CN102540271B (en) Semi-supervised hyperspectral sub-pixel target detection method based on enhanced constraint sparse regression method
Ma et al. Hyperspectral anomaly detection based on low-rank representation with data-driven projection and dictionary construction
CN103116881A (en) Remote sensing image fusion method based on PCA (principal component analysis) and Shearlet conversion
CN104182985A (en) Remote sensing image change detection method
CN112381144B (en) Heterogeneous deep network method for non-European and Euclidean domain space spectrum feature learning
CN103258324A (en) Remote sensing image change detection method based on controllable kernel regression and superpixel segmentation
CN115144357A (en) Mineral identification method based on airborne mid-infrared hyperspectral remote sensing data of mining area
Meng et al. Improving the spatial resolution of hyperspectral image using panchromatic and multispectral images: An integrated method

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination