CN107833208B - Hyperspectral anomaly detection method based on dynamic weight depth self-encoding - Google Patents
Hyperspectral anomaly detection method based on dynamic weight depth self-encoding Download PDFInfo
- Publication number
- CN107833208B CN107833208B CN201711027488.7A CN201711027488A CN107833208B CN 107833208 B CN107833208 B CN 107833208B CN 201711027488 A CN201711027488 A CN 201711027488A CN 107833208 B CN107833208 B CN 107833208B
- Authority
- CN
- China
- Prior art keywords
- pixel
- image
- local
- detected
- reconstruction error
- 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.)
- Active
Links
- 238000001514 detection method Methods 0.000 title claims abstract description 82
- 238000000034 method Methods 0.000 claims abstract description 47
- 230000002159 abnormal effect Effects 0.000 claims abstract description 45
- 230000008569 process Effects 0.000 claims abstract description 24
- 238000004364 calculation method Methods 0.000 claims description 27
- 238000012549 training Methods 0.000 claims description 15
- 210000002569 neuron Anatomy 0.000 claims description 14
- 238000004422 calculation algorithm Methods 0.000 claims description 12
- 238000012360 testing method Methods 0.000 claims description 8
- 239000010410 layer Substances 0.000 description 57
- 230000006870 function Effects 0.000 description 9
- 230000003595 spectral effect Effects 0.000 description 7
- 238000010586 diagram Methods 0.000 description 6
- 230000004913 activation Effects 0.000 description 5
- 230000002547 anomalous effect Effects 0.000 description 5
- 230000000694 effects Effects 0.000 description 3
- 238000005516 engineering process Methods 0.000 description 3
- 238000011156 evaluation Methods 0.000 description 3
- XEEYBQQBJWHFJM-UHFFFAOYSA-N Iron Chemical compound [Fe] XEEYBQQBJWHFJM-UHFFFAOYSA-N 0.000 description 2
- 230000003044 adaptive effect Effects 0.000 description 2
- 238000013528 artificial neural network Methods 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 2
- 210000004027 cell Anatomy 0.000 description 2
- 238000013135 deep learning Methods 0.000 description 2
- 238000003384 imaging method Methods 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 238000012795 verification Methods 0.000 description 2
- 238000010521 absorption reaction Methods 0.000 description 1
- 150000001875 compounds Chemical class 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 238000012937 correction Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000004069 differentiation Effects 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 238000011835 investigation Methods 0.000 description 1
- 229910052742 iron Inorganic materials 0.000 description 1
- 238000002372 labelling Methods 0.000 description 1
- 239000011159 matrix material Substances 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000003287 optical effect Effects 0.000 description 1
- 238000003909 pattern recognition Methods 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
- 239000002356 single layer Substances 0.000 description 1
- 238000001228 spectrum Methods 0.000 description 1
- 239000000126 substance Substances 0.000 description 1
- 230000002194 synthesizing effect Effects 0.000 description 1
- XLYOFNOQVPJJNP-UHFFFAOYSA-N water Chemical compound O XLYOFNOQVPJJNP-UHFFFAOYSA-N 0.000 description 1
- 238000005303 weighing Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/194—Segmentation; Edge detection involving foreground-background segmentation
-
- 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
- G06T2207/10—Image acquisition modality
- G06T2207/10032—Satellite or aerial image; Remote sensing
-
- 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
- G06T2207/20081—Training; Learning
Landscapes
- Engineering & Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The invention discloses a hyperspectral anomaly detection method based on dynamic weight depth self-coding, and relates to a hyperspectral anomaly detection method. The invention aims to solve the problem that an abnormal target pollutes a local model in the existing hyperspectral anomaly detection method, so that the detection precision is low. The process is as follows: firstly, obtaining an optimized DBN model; secondly, obtaining a coded image and a reconstruction error image; thirdly, obtaining a local coding image; executing the fifth step; fourthly, obtaining a local reconstruction error set; executing step six; fifthly, obtaining a local distance factor; executing step seven; sixthly, obtaining all dynamic weights of the local distance; executing step seven; seventhly, obtaining an abnormal detection operator value, setting a threshold value, and when the abnormal detection operator value is greater than or equal to the threshold value, taking the detected pixel as an abnormal target; otherwise, it is a background pixel; and taking the next pixel in the detected image as the detected pixel, and executing three to seven again until all the pixels in the detected image are judged. The hyperspectral image anomaly detection method is used for the field of hyperspectral anomaly detection.
Description
Technical Field
The invention relates to a hyperspectral anomaly detection method.
Background
With the continuous development and progress of remote sensing imaging technology, the hyperspectral image plays an increasingly important role in the fields of precision agriculture, urban planning, military investigation and the like, and the research and application of the hyperspectral remote sensing image are the key points focused by related scientific researchers for a long time. Compared with the visible light or infrared remote sensing imaging technology, the hyperspectral image can not only acquire ground object space distribution information, but also collect spectral information corresponding to tens of to hundreds of continuous narrow bands of ground objects on each pixel point, and the data of the hyperspectral image has the characteristic of map unification, so that ground object substance information can be distinguished through the spectral information. However, in practical application, due to the lack of prior spectrum information of most ground objects and high cost of labeling hyperspectral image data, the unsupervised abnormal target detection method becomes an extremely important means in the practical application of hyperspectral images.
At present, the academic community has no unified definition for the abnormal target, and the ground object target which has a significant difference from the background information is generally called the abnormal target. Among the classical anomaly detection algorithms, the detection method based on the multivariate normal distribution model is most widely applied, and comprises an RXD algorithm, a balanced target detection (UTD) algorithm and a Low Probability Target Detection (LPTD) algorithm. The most classical of them is the RXD algorithm[1]([1]Reed I S,Yu X.Adaptive multiple-band CFAR detection of an optical pattern with unknown spectral distribution[J].IEEE Transactions on Acoustics Speech&Signal Processing,1990,38(10): 1760-. A plurality of researchers improve and research on the problems of data nonlinearity, sub-pixel detection precision and the like of an RXD algorithm, and the problems that the dimension of data is high and nonlinear characteristics have certain influence on the detection precision commonly existing in hyperspectral image data.
In addition, the Banerjee aims at the problem that the false alarm rate of the classical algorithm is high because the real hyperspectral image cannot meet the distribution assumption[2]([2]Banerjee A,Burlina P,Diehl C.A support vector method for anomaly detection in hyperspectral imagery[J]IEEE Transactions on Geoscience and Remote Sensing,2006,44(8):2282-[3]([3]Khazai S,Homayouni S,Safari A,et al.Anomaly detection in hyperspectral images based on an adaptive support vector method[J]IEEE Geoscience and Remote Sensing Letters,2011,8(4):646-And (5) detecting the precision of the mark.
In addition, the abnormal target detection aiming at the hyperspectral image also comprises a method based on the sparse theory, Zongze Yuan[4]([4]Yuan Z,Sun H,Ji K,et al.Local sparsity divergence for hyperspectral anomaly detection[J]IEEE Geoscience and Remote Sensing Letters,2014,11(10): 1697-.
In recent years, the deep learning method is greatly developed in the field of image recognition, and a new idea is provided for hyperspectral image abnormal target detection. 2016, Sarah M.Erfani[5]([5]Erfani S M,Rajasegarar S,Karunasekera S,et al.High-dimensional and large-scale anomaly detection using a linear one-class SVM with deep learning[J]Pattern Recognition,2016,58: 121-.
YihuiXiong[6]([6]Xiong Y,Zuo R.Recognition of geochemical anomalies using a deep autoencoder network[J].Computers&Geosciences,2016,86:75-82.) utilizes a DBN network, the small-probability sample points contribute little to model construction during training, reconstruction errors are relatively large, abnormal detection of remote sensing data is achieved, and Gaussian noise is added to a visible layer and a hidden layer during network training to improve the anti-noise capability of the model, so that iron ore distribution in southwest areas of Fujian is detected. The reconstruction error of the anomalous target is higher than the background sample. Accordingly, Ma proposes a hyperspectral anomalous target detector based on DBN reconstruction error, which obtains better accuracy than the conventional RX anomalous detector, while the model is easily contaminated due to anomalous pixels used for training, resulting in low detection accuracy.
Disclosure of Invention
The invention aims to solve the problem that an abnormal target pollutes a local model in the existing hyperspectral anomaly detection method and the detection precision is low, and provides a hyperspectral anomaly detection method based on dynamic weight deep self-coding.
A hyperspectral anomaly detection method based on dynamic weight depth self-coding comprises the following specific processes:
inputting original hyperspectral image data into a DBN model, and training parameters of the DBN model to obtain an optimized DBN model;
inputting the detected image into the optimized DBN model, and coding the detected image to obtain a coded image of the detected image and a corresponding reconstruction error image;
the measured image is an original hyperspectral image;
inputting the coded image obtained in the step two into a local pixel coding selection block, taking one pixel in the detected image as a detected pixel, and aiming at the detected pixel, obtaining a local coded image of the detected pixel; executing the step five;
inputting the reconstruction error image obtained in the step two into a local pixel reconstruction error selection module, and aiming at the measured pixel selected in the step three, obtaining a local reconstruction error set of the measured pixel; executing the step six;
inputting the local coded image of the pixel to be detected obtained in the step three into a local neighborhood distance calculation module to obtain a local distance factor of the pixel to be detected; executing the step seven;
inputting the local reconstruction error set of the measured pixel obtained in the fourth step into a dynamic weight generation module, calculating the mean value and the variance of the local reconstruction error set of the measured pixel, checking each reconstruction error in the local reconstruction error set of the measured pixel and calculating a dynamic weight to obtain all dynamic weights of the local distance of the measured pixel; executing the step seven;
step seven, inputting the local distance factor of the pixel to be detected obtained in the step five and all the dynamic weights of the local distance of the pixel to be detected obtained in the step six into an abnormal operator calculation module to obtain an abnormal detection operator value of the pixel to be detected, setting a threshold value, and when the abnormal detection operator value of the pixel to be detected is greater than or equal to the threshold value, the pixel to be detected is an abnormal target; when the detected pixel abnormity detection operator value is smaller than the threshold value, the detected pixel is a background pixel;
and taking the next pixel in the detected image as the detected pixel, and re-executing the third step to the seventh step until all the pixels in the detected image are judged.
The invention has the beneficial effects that:
the invention provides a hyperspectral anomaly detection method based on dynamic weight self-coding. The effect of the possible abnormal target in the detection process is reduced, and the effect of the possible normal target in the detection process is increased, so that the pollution problem of the abnormal target to the local model is solved, and the detection precision is improved.
By combining the figure 8 and the table 1, the AUC value of the existing Global Reed-xiaolio Detector method is 0.690, the AUC value of the existing Local Reed-xiaolio Detector method is 0.776, the AUC value of the existing SVDD Detector method is 0.754, the AUC value of the method using only reconstruction errors is 0.870, and the AUC value of the dynamic weighting method provided by the invention is 0.935, so that the result shows that the provided dynamic weighting method is superior to the traditional detection method.
Drawings
FIG. 1 is a diagram of a hyperspectral abnormal target detection method according to the invention;
FIG. 2 is a schematic view of a single layer RBM model;
FIG. 3 is a diagram of a DBN network structure, where a First layer is a First layer, an Input layer is an Input layer, a Hidden layer is a Hidden layer, a Code layer is an encoding layer, an Output layer is an Output layer, Neurons are Neurons, and convention weights are connection weight coefficients;
FIG. 4 is a schematic diagram of a partial pixel selection;
FIG. 5 is a pseudo-color view of airport data from san Diego;
FIG. 6 is a schematic diagram of the real ground position of an anomalous target from a hyperspectral image of an airport in san Diego;
FIG. 7 is a diagram of the detection results obtained by the dynamic weight self-coding anomaly detection technique proposed by the present invention;
FIG. 8 is a ROC curve diagram of the detection result obtained by the dynamic weight self-coding anomaly detection technology provided by the present invention, where DBN rerAD is a ROC curve result for directly using DBN reconstruction to perform anomaly detection, DBN Rer local AD is a ROC curve result for performing local anomaly detection on DBN reconstruction errors, pro field weight base is a ROC curve result for anomaly detection achieved by the method provided by the present invention, Global RX AD is a ROC curve result for anomaly detection of a Global reed-xiaoli method, Collaboratory ReservationAIAD is a ROC curve result for anomaly detection based on joint representation, False positive Rate is False alarm probability, and True positive Rate is discovery probability.
Detailed Description
The first embodiment is as follows: the embodiment is described with reference to fig. 1, and a specific process of the hyperspectral anomaly detection method based on dynamic weight depth self-encoding of the embodiment is as follows:
inputting original hyperspectral image data into a DBN model, and training parameters of the DBN model to obtain an optimized DBN model;
inputting the detected image into the optimized DBN model, and coding the detected image to obtain a coded image of the detected image and a corresponding reconstruction error image;
the measured image is an original hyperspectral image;
inputting the coded image obtained in the step two into a local pixel coding selection block, taking one pixel in the detected image as a detected pixel, and aiming at the detected pixel, obtaining a local coded image of the detected pixel; executing the step five;
inputting the reconstruction error image obtained in the step two into a local pixel reconstruction error selection module, and aiming at the measured pixel selected in the step three, obtaining a local reconstruction error set of the measured pixel; executing the step six;
inputting the local coded image of the pixel to be detected obtained in the step three into a local neighborhood distance calculation module to obtain a local distance factor of the pixel to be detected; executing the step seven;
inputting the local reconstruction error set of the measured pixel obtained in the fourth step into a dynamic weight generation module, calculating the mean value and the variance of the local reconstruction error set of the measured pixel, checking each reconstruction error in the local reconstruction error set of the measured pixel and calculating a dynamic weight to obtain all dynamic weights of the local distance of the measured pixel; executing the step seven;
step seven, inputting the local distance factor of the pixel to be detected obtained in the step five and all the dynamic weights of the local distance of the pixel to be detected obtained in the step six into an abnormal operator calculation module to obtain an abnormal detection operator value of the pixel to be detected, setting a threshold value, and when the abnormal detection operator value of the pixel to be detected is greater than or equal to the threshold value, the pixel to be detected is an abnormal target; when the detected pixel abnormity detection operator value is smaller than the threshold value, the detected pixel is a background pixel;
and taking the next pixel in the detected image as the detected pixel, and re-executing the third step to the seventh step until all the pixels in the detected image are judged.
The second embodiment is as follows: the first difference between the present embodiment and the specific embodiment is: inputting original hyperspectral image data into a DBN model in the first step, and training parameters of the DBN model to obtain an optimized DBN model; the specific process is as follows:
step one, constructing a DBN model, and pre-training the DBN model to obtain a preliminary estimation value of parameters of the DBN model;
the specific process is as follows:
the DBN model (DBN neural network) is composed of a multi-layered RBM model, and a single-layered RBM model is shown in fig. 2. The system comprises n visible layers, m hidden layers and a weight coefficient w connecting the visible layers and the hidden layers;
the visible layer v is input by the RBM and is an n multiplied by 1 column vector, and each component corresponds to each spectral wave band of the hyperspectral image;
the hidden layer h is the output of the RBM and is a column vector of m multiplied by 1;
j,vj∈{0,1},hi∈{0,1},vjj is more than or equal to 1 and less than or equal to n; h isiI is more than or equal to 1 and less than or equal to m; n and m are positive integers;
a matrix with a weight coefficient w of n x m connecting the visible layer and the hidden layer;
the RBM model fits the input hyperspectral image data (data of unknown distribution) by an energy function E (v, h) as follows:
wherein, ciFor biasing of the ith hidden layer unit, wijA weight coefficient connecting the jth visible layer and the ith hidden layer; bjA bias for the jth visible layer cell;
the joint probability density of the simultaneous occurrence of a specific visible layer v and hidden layer h is:
therefore, the probability distribution of v is found by the edge distribution,
similarly, the probability distribution of h is found by the following edge distribution,
combining the above equations (2) - (4), it is proved that the probability that h takes a value of 1 and the probability that v takes a value of 1 are the following conditional probability equations:
therefore, fitting the hyperspectral image data to build a DBN network model, i.e. to the parameter ci、wij、bj(ii) an estimate of (d);
let q (x) be the true distribution of the visible layer v in the sample space Ω, the estimation of each parameter in the above formula should make the difference between p (v) and q (x) as small as possible, i.e. the difference between Kullback-Leibler as small as possible, and the difference between p (v) and q (x) is expressed by KL distance as follows,
get Σx=ΩA maximum value of lnp (x) to minimize KL;
p (v) represents the probability distribution of v; p (x) represents the edge distribution of v; x is sample input, the first layer is image input, and the other layers are output of the previous layer; since q (x) is definite at the time of sample determination and is greater than zero, if KL is minimum, only Σx=Ωlnp (x) is the maximum. And (4) synthesizing the formulas (2) to (7) to derive the DBN model parameter ci、wij、bjThe gradient of (c) is as follows:
to accelerate the gradient convergence process, the MCD (minimum structural differentiation) algorithm is proposed according to Hinton in combination with equations (8) to (10), and parameter ci、wij、bjThe iterative update gradient is performed as follows:
wherein eta iswIs a parameter wijLearning rate (predetermined constant), ηbIs a parameter bjLearning rate (predetermined constant), ηcIs a parameter ciLearning rate of (a predetermined constant), Δ bjIs b isjIteratively updating the gradient, Δ wijIs wijIteratively updating the gradient, Δ cjIs cjIteratively updating the gradient, cjFor the bias of the jth hidden layer unit, hjIs the output of the jth hidden layer unit,respectively representing the values of visible layer units and hidden layer units obtained after one-time reverse reconstruction and forward calculation,representing the value of the hidden layer unit obtained after one forward calculationIn the mode, traversing training and parameter updating are carried out on a hyperspectral image to be detected to obtain a layer of preliminary estimation value of RBM parameters;
because the DBN is formed by stacking a plurality of RBMs, each RBM layer is independently subjected to parameter estimation during training, and the output of the previous RBM layer is used as the input of the next RBM layer to obtain the initial estimation value of the whole DBN parameter;
and step two, performing parameter fine adjustment on the initial estimation values of the DBN model parameters obtained in the step one by one on the basis of a back propagation algorithm to obtain an optimized DBN model.
The specific process is as follows;
the idea is consistent with the parameter training of the BP neural network. And the model is closer to the sample space through fine adjustment of the parameters. For hyperspectral image abnormal target detection, a cost function is defined for a sample input x as follows:
wherein h isw,b(x) Representing the forward propagation output value of the output layer of the DBN model, and calculating the residual error of the ith unit of the output layer of the DBN model according to the following formula,
wherein x isiFor inputting the spectral information of the ith waveband of a pixel of the hyperspectral image,for the output value of the ith cell of the output layer of the DBN model,for the activation input value of the DBN model output layer, nl represents the output layer;
for the middle layer residuals of the DBN model, the following formula,
wherein Sl+1The number of the l +1 layer neurons,inputting values for an activation function in layer I neurons (the activation function is a function in the neurons, the function of the activation function is equivalent to the neurons, and the function of the neurons is embodied by the activation function), wherein l is 2 to nl-1; because the network is a multilayer structure, l represents the layer I, and nl represents the number of neurons in the layer I;
thus, the compounds are obtained according to the formulae (15) to (17)And ci,And ciThe update method of (2) is as follows,
through iteration, parameters of the DBN network model can be finely adjusted, and the optimized DBN model is obtained. After the DBN network training is completed, the hyperspectral image is input into the DBN network pixel by pixel, the reconstruction error of the hyperspectral image is calculated, and then the detection of each pixel of the hyperspectral image can be completed.
Other steps and parameters are the same as those in the first embodiment.
The third concrete implementation mode: the present embodiment differs from the first or second embodiment in that: inputting the detected image into the optimized DBN model in the second step, and coding the detected image to obtain a coded image of the detected image and a corresponding reconstruction error image; the specific process is as follows:
DBN inference image coding
The method mainly obtains the image code of the image after the image passes through the DBN network. And inputting the original hyperspectral image into a DBN reasoning and coding module, wherein the module is a standard DBN model, and parameters of the module are generated in a DBN model training stage. The model structure is shown in FIG. 3
Inputting a detected image into an optimized DBN model, wherein the number of neurons in the middle layer of the DBN model is generally lower than that of nodes in an input and output layer, the whole DBN model is of a symmetrical structure, the output of the neurons in the middle layer is used as a coding result of the detected image, all detected pixels are independently coded for one time, so that a coded image of the detected image is obtained, and the number of wave bands of each detected image after being coded is lower than that of the original input detected image; the coded image is sent to the local pixel coding selection module in the third step.
Decoding the obtained coded image by using the optimized DBN model so as to obtain a decoded image, wherein the decoded image and the input measured image have the same number of wave bands, and calculating a reconstruction error image corresponding to the measured image by the following formula:
wherein z is the pixel of the image to be detected, hw,b(z) represents the DBN model output layer forward propagation output value.
After the reconstruction error image calculation is finished, the reconstruction error image is sent to the local pixel reconstruction error selection module in the step four
Other steps and parameters are the same as those in the first or second embodiment.
The fourth concrete implementation mode: the difference between this embodiment mode and one of the first to third embodiment modes is: inputting the coded image obtained in the second step into a local pixel coding selection block in the third step, taking a pixel in the detected image as a detected pixel, and obtaining a local coded image of the detected pixel aiming at the detected pixel; the specific process is as follows:
and after the DBN reasoning image coding module finishes the coding of the image, the coded image is input to a local pixel coding selection module and is used for selecting a local pixel set evaluated by the detected target.
A pixel in a detected image is taken as a detected pixel, a square pixel region (determined by artificial experience) is constructed on a coded image by taking the detected pixel coordinate as the center aiming at the detected pixel, the size of the square pixel region is far smaller than that of the coded image, a square window is nested in the square pixel region (the size is determined artificially in advance), the nested square window region is an adjacent pixel region, the pixel of the region between two squares is a selected pixel region, and the coded image of the selected pixel region is a local coded image.
The coded image of the adjacent pixel to be detected (excluding pixel area in the figure) of the square pixel area is excluded, and the coded image of other part of pixels (selected pixel area) in the square area is a local coded image; and sending the local coded image to a local neighborhood distance calculation module for calculating a local distance factor. As shown in fig. 4.
Other steps and parameters are the same as those in one of the first to third embodiments.
The fifth concrete implementation mode: the difference between this embodiment and one of the first to fourth embodiments is: inputting the reconstruction error image obtained in the second step into a local pixel reconstruction error selection module in the fourth step, and obtaining a local reconstruction error set of the measured pixel aiming at the measured pixel selected in the third step; the specific process is as follows:
after the DBN inference image coding module completes the reconstruction error calculation of the image, the coded image is input to the local pixel coding selection module, and meanwhile, the reconstruction error image is input to the local pixel reconstruction error selection module and used for selecting the reconstruction error value used for the local pixel set of the detected target. When a coded pixel of a certain coordinate value is selected for detection, the corresponding local reconstruction error is performed according to the local pixel selection of the following graph, the process of the local reconstruction error is consistent with the execution process of the local pixel coding selection module, and the method comprises the following steps:
aiming at the measured pixel selected in the step three, a square area is constructed on the reconstruction error image by taking the measured pixel coordinate as the center, and the size of the square pixel area is consistent with that constructed in the local pixel coding selection module; a square window is nested in the square pixel region (the size is artificially determined in advance), the nested square window region is an adjacent pixel region, the region pixels between the two squares are local reconstruction errors, and all the local reconstruction errors of the measured pixels form a local reconstruction error set.
The 1 pixel under test has only one local reconstruction error set, which is composed of all the reconstruction errors of the pixels between the two square windows outside the pixel under test.
The pixels (excluding pixel areas in the figure) adjacent to the detected pixels in the square pixel area are excluded, and other parts of pixels (selected pixel areas) in the square area are local reconstruction errors and are sent to a dynamic weight generation module for generating dynamic weights;
other steps and parameters are the same as in one of the first to fourth embodiments.
The sixth specific implementation mode: the difference between this embodiment and one of the first to fifth embodiments is: inputting the local coded image of the pixel to be detected obtained in the third step into a local neighborhood distance calculation module to obtain a local distance factor of the pixel to be detected; the specific process is as follows:
inputting the local coding image of the detected pixel obtained in the step three into a local neighborhood distance calculation module, and completing the calculation of a local distance factor in the local neighborhood distance calculation module according to the following formula:
neighborhood distance Dist between detected pixel y in coded image and jth pixel of local coded imagejThe following formula:
wherein D represents the code length, yiIndicating the i-th symbol value, z, of the pixel y being measured after it has been encodedjiAn ith symbol value representing a jth pixel of the partially encoded image;
all neighborhood distances of the pixel under test constitute a local distance factor of the pixel under test.
Other steps and parameters are the same as those in one of the first to fifth embodiments.
The seventh embodiment: the difference between this embodiment and one of the first to sixth embodiments is: inputting the local reconstruction error set of the measured pixel obtained in the fourth step into a dynamic weight generation module, calculating the mean and variance of the local reconstruction error set of the measured pixel, checking each reconstruction error in the local reconstruction error set of the measured pixel and calculating a dynamic weight to obtain all dynamic weights of the local distance of the measured pixel; the method comprises the following steps:
subtracting the mean value of the local reconstruction error set from each reconstruction error in the local reconstruction error set of the measured pixel obtained in the step four, and if the result is more than 3 times of the variance of the reconstruction error set, then the dynamic weight wt corresponding to the local reconstruction errorjAs calculated by the following formula,
Rerjthe reconstruction error corresponding to the jth pixel of the hyperspectral image is that j is more than or equal to 1 and less than or equal to K,k is the number of the selected local pixel points, namely the size of the reconstruction error set;
otherwise, the dynamic weight wt corresponding to the local reconstruction errorjThe following calculation is carried out in accordance with the formula,
wherein, Pf is more than or equal to 0 and less than or equal to 1, and Pf is a constant parameter (predetermined for weighing the magnitude of the weight coefficient correction);
all dynamic weights of the local distance of the pixel to be detected are obtained by detecting all reconstruction errors in the local reconstruction error set, and the dynamic weights are sent to an abnormal operator calculation module for calculating an abnormal detection operator.
Other steps and parameters are the same as those in one of the first to sixth embodiments.
The specific implementation mode is eight: the present embodiment differs from one of the first to seventh embodiments in that: and Pf is 0.
Through verification, the effect of Pf taking 0 is optimal.
Other steps and parameters are the same as those in one of the first to seventh embodiments.
The specific implementation method nine: the present embodiment differs from the first to eighth embodiments in that: in the seventh step, all the dynamic weights of the local distance factor of the pixel to be detected obtained in the fifth step and the local distance of the pixel to be detected obtained in the sixth step are input into an abnormal operator calculation module to obtain an abnormal detection operator value of the pixel to be detected; the specific process is as follows:
calculating an anomaly detection operator value delta for a pixel under testDBN-ρThe calculation method is carried out according to the following formula:
calculating abnormal operators of all pixels of each pixel in the image through the local pixel coding selection module, the local pixel reconstruction error selection module, the dynamic weight generation module, the local neighborhood distance calculation module and the abnormal operator calculation module, so as to set a threshold value, judging whether the pixel to be detected is an abnormal target or not according to the abnormal detection operator value and the threshold value, and when the abnormal detection operator value is greater than or equal to the threshold value, the pixel point is the abnormal target; when the abnormal detection operator value is smaller than the threshold value, the pixel point is a target;
when the abnormal operator value is larger, the probability that the pixel point is an abnormal target is larger.
Other steps and parameters are the same as those in one to eight of the embodiments.
The following examples were used to demonstrate the beneficial effects of the present invention:
the first embodiment is as follows:
the hyperspectral anomaly detection method based on dynamic weight depth self-coding is specifically prepared according to the following steps:
the detection method provided by the text is verified by adopting a public standard hyperspectral data set for NASA acquisition, adopting a part of image data of the san Diego airport acquired by AVIRIS as a real data set, wherein the size of the image is 100 multiplied by 100, the number of spectral wave bands is 126 after water vapor absorption and interference wave bands are removed, 38 airplanes in the image are taken as abnormal targets, and a data pseudo-color image is shown in FIG. 5. The real ground position of the anomaly target in the hyperspectral image of the san Diego airport is shown in FIG. 6.
Evaluation method and experimental environment
The hyperspectral anomaly detection classical algorithms RXD and SVDD are used as comparison algorithms, and a receiver characteristic curve (ROC curve) and an area under the ROC curve (AUC area) which are commonly used in hyperspectral anomaly detection are used as evaluation indexes. The ROC curve can obtain different false alarm rates and detection rates by changing an abnormal discrimination threshold value, so that the performance of the detector is evaluated, when the ROC curve is not obvious enough in comparison, the area under the ROC curve (AUC area) can be used as another common evaluation index, the hardware environment for experiments is a Dell T7910 workstation, the software environment is Matlab 2015b, and when the DBN network is used for detecting abnormal targets, the number of nodes of input visible layers and output layers is the same as the number of spectral bands.
Verification result
The detection result obtained by the proposed dynamic weight self-coding anomaly detection technique is shown in fig. 7;
the ROC curves are shown in fig. 8, and the AUC values of each curve are compared with other methods as shown in the following table:
TABLE 1
Name of comparison method | AUC value |
Global Reed-Xiaoli Detector | 0.690 |
Local Reed-Xiaoli Detector | 0.776 |
SVDD Detector | 0.754 |
Method using only reconstruction errors | 0.870 |
Proposed dynamic weighting method | 0.935 |
The results show that the proposed dynamic weight method is superior to the conventional detection method.
The present invention is capable of other embodiments and its several details are capable of modifications in various obvious respects, all without departing from the spirit and scope of the present invention.
Claims (7)
1. A hyperspectral anomaly detection method based on dynamic weight depth self-coding is characterized by comprising the following steps: the method comprises the following specific processes:
inputting original hyperspectral image data into a DBN model, and training parameters of the DBN model to obtain an optimized DBN model;
inputting the detected image into the optimized DBN model, and coding the detected image to obtain a coded image of the detected image and a corresponding reconstruction error image;
the measured image is an original hyperspectral image;
inputting the coded image obtained in the step two into a local pixel coding selection block, taking one pixel in the detected image as a detected pixel, and aiming at the detected pixel, obtaining a local coded image of the detected pixel; executing the step five;
inputting the reconstruction error image obtained in the step two into a local pixel reconstruction error selection module, and aiming at the measured pixel selected in the step three, obtaining a local reconstruction error set of the measured pixel; executing the step six;
inputting the local coded image of the pixel to be detected obtained in the step three into a local neighborhood distance calculation module to obtain a local distance factor of the pixel to be detected; executing the step seven;
inputting the local reconstruction error set of the measured pixel obtained in the fourth step into a dynamic weight generation module, calculating the mean value and the variance of the local reconstruction error set of the measured pixel, checking each reconstruction error in the local reconstruction error set of the measured pixel and calculating a dynamic weight to obtain all dynamic weights of the local distance of the measured pixel; executing the step seven;
step seven, inputting the local distance factor of the pixel to be detected obtained in the step five and all the dynamic weights of the local distance of the pixel to be detected obtained in the step six into an abnormal operator calculation module to obtain an abnormal detection operator value of the pixel to be detected, setting a threshold value, and when the abnormal detection operator value of the pixel to be detected is greater than or equal to the threshold value, the pixel to be detected is an abnormal target; when the detected pixel abnormity detection operator value is smaller than the threshold value, the detected pixel is a background pixel;
taking the next pixel in the detected image as the detected pixel, and re-executing the third step to the seventh step until all pixels in the detected image are judged;
inputting the coded image obtained in the second step into a local pixel coding selection block in the third step, taking a pixel in the detected image as a detected pixel, and obtaining a local coded image of the detected pixel aiming at the detected pixel; the specific process is as follows:
taking a pixel in a detected image as a detected pixel, constructing a square pixel area on a coded image by taking the detected pixel coordinate as a center aiming at the detected pixel, wherein the size of the square pixel area is smaller than that of the coded image, a square window is nested in the square pixel area, the nested square window area is an adjacent pixel area, the area pixel between two squares is a selected pixel area, and the coded image of the selected pixel area is a local coded image;
inputting the reconstruction error image obtained in the second step into a local pixel reconstruction error selection module in the fourth step, and obtaining a local reconstruction error set of the measured pixel aiming at the measured pixel selected in the third step; the specific process is as follows:
aiming at the measured pixel selected in the step three, a square area is constructed on the reconstruction error image by taking the measured pixel coordinate as the center, and the size of the square pixel area is consistent with that constructed in the local pixel coding selection module; a square window is nested in the square pixel region, the nested square window region is an adjacent pixel region, the region pixel between two squares is a local reconstruction error, and all the local reconstruction errors of the measured pixel form a local reconstruction error set.
2. The hyperspectral anomaly detection method based on dynamic weight depth self-coding according to claim 1, characterized in that: inputting original hyperspectral image data into a DBN model in the first step, and training parameters of the DBN model to obtain an optimized DBN model; the specific process is as follows:
step one, constructing a DBN model, and pre-training the DBN model to obtain a preliminary estimation value of parameters of the DBN model;
and step two, performing parameter fine adjustment on the initial estimation values of the DBN model parameters obtained in the step one by one on the basis of a back propagation algorithm to obtain an optimized DBN model.
3. The hyperspectral anomaly detection method based on dynamic weight depth self-coding according to claim 2 is characterized in that: inputting the detected image into the optimized DBN model in the second step, and coding the detected image to obtain a coded image of the detected image and a corresponding reconstruction error image; the specific process is as follows:
inputting a measured image into an optimized DBN model, wherein the number of neurons in the middle layer of the DBN model is lower than the number of nodes in an input and output layer, the whole DBN model is of a symmetrical structure, the output of the neurons in the middle layer is used as a coding result of the measured image, all measured pixels are independently coded for one time, so that a coded image of the measured image is obtained, and the number of wave bands of each measured image after being coded is lower than the number of wave bands of the pixels of the originally input measured image;
decoding the obtained coded image by using the optimized DBN model so as to obtain a decoded image, wherein the decoded image and the input measured image have the same number of wave bands, and calculating a reconstruction error image corresponding to the measured image by the following formula:
wherein z is the pixel of the image to be detected, hw,b(z) represents the DBN model output layer forward propagation output value.
4. The hyperspectral anomaly detection method based on dynamic weight depth self-coding according to claim 3, characterized in that: inputting the local coded image of the pixel to be detected obtained in the third step into a local neighborhood distance calculation module to obtain a local distance factor of the pixel to be detected; the specific process is as follows:
inputting the local coding image of the detected pixel obtained in the step three into a local neighborhood distance calculation module, and completing the calculation of a local distance factor in the local neighborhood distance calculation module according to the following formula:
neighborhood distance Dist between detected pixel y in coded image and jth pixel of local coded imagejThe following formula:
wherein D represents the code length, yiIndicating the i-th symbol value, z, of the pixel y being measured after it has been encodedjiAn ith symbol value representing a jth pixel of the partially encoded image;
all neighborhood distances of the pixel under test constitute a local distance factor of the pixel under test.
5. The hyperspectral anomaly detection method based on dynamic weight depth self-coding according to claim 4, characterized in that: inputting the local reconstruction error set of the measured pixel obtained in the fourth step into a dynamic weight generation module, calculating the mean and variance of the local reconstruction error set of the measured pixel, checking each reconstruction error in the local reconstruction error set of the measured pixel and calculating a dynamic weight to obtain all dynamic weights of the local distance of the measured pixel; the specific process is as follows:
subtracting the mean value of the local reconstruction error set from each reconstruction error in the local reconstruction error set of the measured pixel obtained in the step four, and if the result is more than 3 times of the variance of the reconstruction error set, then the dynamic weight wt corresponding to the local reconstruction errorjAs calculated by the following formula,
Rerjj is more than or equal to 1 and less than or equal to K, wherein K is the number of selected local pixel points, namely the size of a reconstruction error set;
otherwise, the dynamic weight wt corresponding to the local reconstruction errorjThe following calculation is carried out in accordance with the formula,
wherein Pf is more than or equal to 0 and less than or equal to 1, and is a constant parameter;
and obtaining all dynamic weights of the local distance of the pixel to be measured by detecting all reconstruction errors in the local reconstruction error set.
6. The hyperspectral anomaly detection method based on dynamic weight depth self-coding according to claim 5, characterized in that: and Pf is 0.
7. The hyperspectral anomaly detection method based on dynamic weight depth self-coding according to claim 6, characterized in that: in the seventh step, all the dynamic weights of the local distance factor of the pixel to be detected obtained in the fifth step and the local distance of the pixel to be detected obtained in the sixth step are input into an abnormal operator calculation module to obtain an abnormal detection operator value of the pixel to be detected;
the specific process is as follows:
calculating an anomaly detection operator value delta for a pixel under testDBN-ρThe calculation method is carried out according to the following formula:
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201711027488.7A CN107833208B (en) | 2017-10-27 | 2017-10-27 | Hyperspectral anomaly detection method based on dynamic weight depth self-encoding |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201711027488.7A CN107833208B (en) | 2017-10-27 | 2017-10-27 | Hyperspectral anomaly detection method based on dynamic weight depth self-encoding |
Publications (2)
Publication Number | Publication Date |
---|---|
CN107833208A CN107833208A (en) | 2018-03-23 |
CN107833208B true CN107833208B (en) | 2021-07-13 |
Family
ID=61649862
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201711027488.7A Active CN107833208B (en) | 2017-10-27 | 2017-10-27 | Hyperspectral anomaly detection method based on dynamic weight depth self-encoding |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN107833208B (en) |
Families Citing this family (12)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108985330B (en) * | 2018-06-13 | 2021-03-26 | 华中科技大学 | Self-coding network and training method thereof, and abnormal power utilization detection method and system |
CN108830320B (en) * | 2018-06-15 | 2021-11-30 | 南京农业大学 | Hyperspectral image classification method based on identification and robust multi-feature extraction |
CN108960303B (en) * | 2018-06-20 | 2021-05-07 | 哈尔滨工业大学 | Unmanned aerial vehicle flight data anomaly detection method based on LSTM |
CN109493338B (en) * | 2018-11-16 | 2020-08-04 | 西安电子科技大学 | Hyperspectral image anomaly detection method based on combined extraction of spatial spectral features |
CN109799245A (en) * | 2019-03-29 | 2019-05-24 | 云南电网有限责任公司电力科学研究院 | A kind of insulator contamination degree non-contact detection method and device |
CN110781832B (en) * | 2019-10-28 | 2023-04-18 | 大庆师范学院 | Hyperspectral image abnormal target detection method based on combined depth confidence network |
CN112819769B (en) * | 2021-01-26 | 2022-05-20 | 复旦大学 | Nonlinear hyperspectral image anomaly detection algorithm based on kernel function and joint dictionary |
CN113011321B (en) * | 2021-03-17 | 2022-05-06 | 中南大学 | Spectral signal denoising method, system, terminal and readable storage medium based on joint dictionary |
CN113222972B (en) * | 2021-05-31 | 2024-03-19 | 辽宁工程技术大学 | Image anomaly detection method based on variation self-encoder algorithm |
CN113378683B (en) * | 2021-06-03 | 2024-06-21 | 大连海事大学 | Residual error driven anomaly detection band selection method |
CN114112992B (en) * | 2022-01-20 | 2022-04-12 | 武汉工程大学 | Detection method and device for blue pigment of cream and storage medium |
CN117132504B (en) * | 2023-10-20 | 2024-01-30 | 武汉怡特环保科技有限公司 | Method for removing interference based on single photon spectrum counting technology imaging |
Citations (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US7263226B1 (en) * | 2002-07-09 | 2007-08-28 | United States Of America As Represented By The Secretary Of The Navy | Method and system for detecting anomalies in multispectral and hyperspectral imagery employing the normal compositional model |
CN103024349A (en) * | 2012-11-13 | 2013-04-03 | 大连民族学院 | MLE (Maximum Likelihood Estimation) video target tracking method based on sparse limitation |
CN105335703A (en) * | 2015-10-13 | 2016-02-17 | 西北工业大学 | Traffic scene abnormity detection method based on movement reconstruction technology |
CN105989597A (en) * | 2015-02-13 | 2016-10-05 | 中国科学院西安光学精密机械研究所 | Hyperspectral image abnormal target detection method based on pixel selection process |
CN106023218A (en) * | 2016-05-27 | 2016-10-12 | 哈尔滨工程大学 | Hyperspectral abnormity detection method based on spatial spectrum combined background co-sparse representation |
CN106503734A (en) * | 2016-10-14 | 2017-03-15 | 哈尔滨工程大学 | Based on trilateral filter and the image classification method of the sparse autocoder of storehouse |
CN106919952A (en) * | 2017-02-23 | 2017-07-04 | 西北工业大学 | EO-1 hyperion Anomaly target detection method based on structure rarefaction representation and internal cluster filter |
CN107203750A (en) * | 2017-05-24 | 2017-09-26 | 中国科学院西安光学精密机械研究所 | Hyperspectral target detection method based on combination of sparse expression and discriminant analysis |
CN107274432A (en) * | 2017-06-10 | 2017-10-20 | 北京航空航天大学 | A kind of common scene intelligent video monitoring method of view-based access control model conspicuousness and depth own coding |
-
2017
- 2017-10-27 CN CN201711027488.7A patent/CN107833208B/en active Active
Patent Citations (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US7263226B1 (en) * | 2002-07-09 | 2007-08-28 | United States Of America As Represented By The Secretary Of The Navy | Method and system for detecting anomalies in multispectral and hyperspectral imagery employing the normal compositional model |
CN103024349A (en) * | 2012-11-13 | 2013-04-03 | 大连民族学院 | MLE (Maximum Likelihood Estimation) video target tracking method based on sparse limitation |
CN105989597A (en) * | 2015-02-13 | 2016-10-05 | 中国科学院西安光学精密机械研究所 | Hyperspectral image abnormal target detection method based on pixel selection process |
CN105335703A (en) * | 2015-10-13 | 2016-02-17 | 西北工业大学 | Traffic scene abnormity detection method based on movement reconstruction technology |
CN106023218A (en) * | 2016-05-27 | 2016-10-12 | 哈尔滨工程大学 | Hyperspectral abnormity detection method based on spatial spectrum combined background co-sparse representation |
CN106503734A (en) * | 2016-10-14 | 2017-03-15 | 哈尔滨工程大学 | Based on trilateral filter and the image classification method of the sparse autocoder of storehouse |
CN106919952A (en) * | 2017-02-23 | 2017-07-04 | 西北工业大学 | EO-1 hyperion Anomaly target detection method based on structure rarefaction representation and internal cluster filter |
CN107203750A (en) * | 2017-05-24 | 2017-09-26 | 中国科学院西安光学精密机械研究所 | Hyperspectral target detection method based on combination of sparse expression and discriminant analysis |
CN107274432A (en) * | 2017-06-10 | 2017-10-20 | 北京航空航天大学 | A kind of common scene intelligent video monitoring method of view-based access control model conspicuousness and depth own coding |
Non-Patent Citations (1)
Title |
---|
A DBN based anomaly targets detector for HSI;Ning Ma等;《AOPC 2017: 3D Measurement Technology for Intelligent Manufacturing》;20171024;1-6 * |
Also Published As
Publication number | Publication date |
---|---|
CN107833208A (en) | 2018-03-23 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN107833208B (en) | Hyperspectral anomaly detection method based on dynamic weight depth self-encoding | |
Liu et al. | Detection of multiclass objects in optical remote sensing images | |
Li et al. | Structure-aware collaborative representation for hyperspectral image classification | |
CN110929080B (en) | Optical remote sensing image retrieval method based on attention and generation countermeasure network | |
CN111986193B (en) | Remote sensing image change detection method, electronic equipment and storage medium | |
Tian et al. | Joint learning model for underwater acoustic target recognition | |
CN112905894B (en) | Collaborative filtering recommendation method based on enhanced graph learning | |
CN111783884A (en) | Unsupervised hyperspectral image classification method based on deep learning | |
CN113435253A (en) | Multi-source image combined urban area ground surface coverage classification method | |
CN114266982A (en) | SAR image change detection method based on network parameter and structure collaborative optimization | |
CN111462196A (en) | Remote sensing image matching method based on cuckoo search and Krawtchouk moment invariant | |
CN110956601B (en) | Infrared image fusion method and device based on multi-sensor mode coefficients and computer readable storage medium | |
CN114048810A (en) | Hyperspectral image classification method based on multilevel feature extraction network | |
Aires et al. | Convolutional neural networks for satellite remote sensing at coarse resolution. Application for the SST retrieval using IASI | |
CN116152206A (en) | Photovoltaic output power prediction method, terminal equipment and storage medium | |
CN113470036B (en) | Hyperspectral image unsupervised waveband selection method and system based on knowledge distillation | |
CN117524353B (en) | Molecular large model based on multidimensional molecular information, construction method and application | |
CN112446256A (en) | Vegetation type identification method based on deep ISA data fusion | |
CN111062888B (en) | Hyperspectral image denoising method based on multi-target low-rank sparsity and spatial-spectral total variation | |
Sun et al. | Tensor decomposition-inspired convolutional autoencoders for hyperspectral anomaly detection | |
Lv et al. | Novel Distribution Distance based on Inconsistent Adaptive Region for Change Detection Using Hyperspectral Remote Sensing Images | |
Celik et al. | Change detection without difference image computation based on multiobjective cost function optimization | |
Kong et al. | Regularized multiple sparse Bayesian learning for hyperspectral target detection | |
Elsisi et al. | Robust indoor positioning of automated guided vehicles in internet of things networks with deep convolution neural network considering adversarial attacks | |
CN115631341A (en) | Point cloud registration method and system based on multi-scale feature voting |
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 | ||
GR01 | Patent grant | ||
GR01 | Patent grant |