CN104778706B - Method for detecting abnormality based on Non-negative Matrix Factorization and its device - Google Patents
Method for detecting abnormality based on Non-negative Matrix Factorization and its device Download PDFInfo
- Publication number
- CN104778706B CN104778706B CN201510191275.2A CN201510191275A CN104778706B CN 104778706 B CN104778706 B CN 104778706B CN 201510191275 A CN201510191275 A CN 201510191275A CN 104778706 B CN104778706 B CN 104778706B
- Authority
- CN
- China
- Prior art keywords
- matrix
- high spectrum
- image
- spectrum image
- iteration
- 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
Landscapes
- Image Processing (AREA)
Abstract
The invention discloses a kind of method for detecting abnormality based on Non-negative Matrix Factorization, the high spectrum image that pretreatment obtains cancelling noise is carried out to the high spectrum image of reading, the high spectrum image of the cancelling noise of the acquisition is converted into row vector and obtains two dimension initialization matrix V, linear decomposition is carried out to two dimension initialization matrix V again and generates random initializtion basic matrix W and coefficient matrix H, random initializtion basic matrix W and coefficient matrix H are iterated according to the multiplying property iterative rules of Non-negative Matrix Factorization and obtain the high spectrum image containing several wave bands, finally the high spectrum image containing a most wave band of exception information in the high spectrum image containing several wave bands is handled according to the adaptive Density Estimator operator of part, obtain the image for detecting abnormal object；The invention also discloses a kind of abnormal detectors based on Non-negative Matrix Factorization, can eliminate a large amount of redundancy wave band and noise information through the invention, to be effectively improved the efficiency of abnormality detection.
Description
Technical field
The invention belongs to Remote Sensing Data Processing technical fields, and in particular to a kind of abnormality detection based on Non-negative Matrix Factorization
Method and its device.
Background technology
Remote sensing technology is the twentieth century end mankind one of major technological breakthrough acquired in terms of earth observation, and bloom
Spectrum remote sensing technology is exactly the cutting edge technology of current remote sensing.High-spectrum seems a kind of 3-D view, except the two dimension on traditional images
There is spectrum dimension outside plane.Each wave band of high spectrum image can individually be regarded as an image, by the figure of all wave bands
As stacking up, an image cube can be formed.Each wave between theoretically spectrum is tieed up in verified hyperspectral datacube
The data of section have very strong correlation, and data are practically free of inside higher dimensional space, and most of data focus only on low
Dimension space, that is, focus only on wherein a small number of several wave bands, and contained by remaining most of wave band is all meaningless information or makes an uproar
Sound.Therefore how to EO-1 hyperion wave band dimensionality reduction, useful information is extracted from bulk redundancy wave band, becomes important for one
Proposition.
At present in the method for high-spectrum image dimensionality reduction, there are commonly principal component analysis (PCA) and independent component analysis
(ICA) two kinds, wherein：
Principal component analysis is a kind of feature extracting method projected based on linear subspaces, it is to study how to pass through minority
Several principal components explain the covariance structure of multivariable, by exporting several principal components, so that them is retained original as much as possible
The information of beginning variable, and it is uncorrelated to each other.PCA is also a kind of side that current high dimensional data dimensionality reduction aspect is most widely used
Method.In new subspace after PCA transformation, the correlation between image substantially eliminates, and can be represented with several principal components
Most of information in initial data, and the first factor therein contains the information of the overwhelming majority.Before only being taken due to us
One to two principal components, so this algorithm is simple, it is easy to implement, but other wave bands that do not get still contain perhaps
Mostly useful information, therefore the method efficiency is very low, acquired results assessment is poor.
Independent component analysis is that internal factor or a kind of statistical method of ingredient are found from multivariate statistics data, is obtained
The ingredient of statistical iteration and non-gaussian, final result are referred to as independent element (IC).Independent component analysis is in image procossing
It is widely applied, many important characteristic informations all have close relationship, ICA with the higher order statistical characteristic of image pixel
Exactly these order of information of method extraction, this is also consistent with the design feature of the high data dimension of EO-1 hyperion.The method efficiency is wanted
It is higher by much than PCA algorithm, but needs complicated preprocessing process, and the randomness of algorithm is very high, acquired results and institute
It takes initial value to have much relations, is not easy to realize.
Invention content
In view of this, the main purpose of the present invention is to provide a kind of method for detecting abnormality based on Non-negative Matrix Factorization and
Its device
In order to achieve the above objectives, the technical proposal of the invention is realized in this way：
The embodiment of the present invention provides a kind of method for detecting abnormality based on Non-negative Matrix Factorization, and this method is：To reading
High spectrum image carries out the high spectrum image that pretreatment obtains cancelling noise, by the high spectrum image of the cancelling noise of the acquisition
It is converted into row vector and obtains two dimension initialization matrix V, then linear decomposition is carried out to two dimension initialization matrix V and generates random initializtion
Basic matrix W and coefficient matrix H, according to the multiplying property iterative rules of Non-negative Matrix Factorization to random initializtion basic matrix W and coefficient square
Battle array H, which is iterated, obtains the high spectrum image containing several wave bands, is finally calculated according to the adaptive Density Estimator of part
Son to the high spectrum image containing a most wave band of exception information in the high spectrum image containing several wave bands into
Row processing, obtains the image for detecting abnormal object.
It is described that acquisition is iterated to random initializtion basic matrix W and coefficient matrix H according to rule of iteration in said program
High spectrum image containing each and every one several wave bands, specially：According to the multiplying property iterative rules of Non-negative Matrix Factorization respectively to random
Initialization basic matrix W and coefficient matrix H is iterated, and the multiplying property iterative rules of the Non-negative Matrix Factorization are respectively：
Wherein i, k, j respectively indicate line number and columns, matrix
WT、HTThe respectively transposed matrix of matrix W and T；
According to square of the Euclidean distance between two dimension initialization matrix V and random initializtion basic matrix W and coefficient matrix H
As object function come weigh decompose after result and original matrix degree of closeness, given threshold η, by after each iteration with
Machine initializes basic matrix W and coefficient matrix H and is sent into object function, obtains iteration result, if the iteration result is returned more than η
Continue iteration；If the iteration result is less than η, stop iteration, coefficient matrix H is changed into three-dimensional high-spectrum
As to get to the high spectrum image of several wave bands containing more exception information.
It is described that acquisition is iterated to random initializtion basic matrix W and coefficient matrix H according to rule of iteration in said program
High spectrum image containing several wave bands, later this method further include：It will contain several according to abnormality detection operator PASNR
Several band images arrive minispread by exception information number according to big in the high spectrum image of wave band, and therefrom choose 5 and contain
There is the band image of maximum exception information to be again transformed into two dimensional image.
It is described that several waves are finally contained to described according to the adaptive Density Estimator operator of part in said program
High spectrum image containing a most wave band of exception information in the high spectrum image of section is handled, and acquisition detects exception
The image of target, specially：Each section in the two dimensional image containing 5 wave bands is handled respectively according to LAKDE operators,
The whole image for detecting target is obtained after processing.
It is described that each section in the two dimensional image containing 5 wave bands is carried out respectively according to LAKDE operators in said program
Processing obtains the whole image for detecting target, specially after processing：The dualism hypothesis discriminant function of the LAKDE operators
For：
Wherein x is sample point to be detected, xiFor other sample points around, N is pixel number, and κ is kernel function, is used for certainly
Determine the shape in domain around the data sample in data space, h (k, x) is the bandwidth of kernel function, and η is discrimination threshold, is being assumed to be
H1In the case of, ΛLAKDE-AD(x) it is more than η, x is abnormal object；It is being assumed to be H0In the case of, ΛLAKDE-AD(x) it is less than η, x is
Background dot.
The embodiment of the present invention also provides a kind of abnormal detector based on Non-negative Matrix Factorization, which includes：Pre- place
Manage module, conversion and decomposing module, iteration module, detection module；
The preprocessing module carries out the high-spectrum that pretreatment obtains cancelling noise for the high spectrum image to reading
Picture；
The conversion and decomposing module, for obtaining the high spectrum image of the cancelling noise of the acquisition into row vector conversion
Two-dimentional initialization matrix V is obtained, then linear decomposition is carried out to two dimension initialization matrix V and generates random initializtion basic matrix W and coefficient
Matrix H；
The iteration module, for according to the multiplying property iterative rules of Non-negative Matrix Factorization to random initializtion basic matrix W and
Coefficient matrix H, which is iterated, obtains the high spectrum image containing several wave bands,
The detection module, for finally containing several to described according to the adaptive Density Estimator operator of part
High spectrum image containing a most wave band of exception information in the high spectrum image of wave band is handled, and acquisition detects different
The image of normal target.
In said program, the iteration module is distinguished particularly for according to the multiplying property iterative rules of Non-negative Matrix Factorization
Random initializtion basic matrix W and coefficient matrix H are iterated, the multiplying property iterative rules of the Non-negative Matrix Factorization are respectively：Wherein i, k, j respectively indicate line number and columns, matrix WT、HTRespectively
The transposed matrix of matrix W and T；
According to square of the Euclidean distance between two dimension initialization matrix V and random initializtion basic matrix W and coefficient matrix H
As object function come weigh decompose after result and original matrix degree of closeness, given threshold η, by after each iteration with
Machine initializes basic matrix W and coefficient matrix H and is sent into object function, obtains iteration result, if the iteration result is returned more than η
Continue iteration；If the iteration result is less than η, stopping is iterated, and coefficient matrix H is changed into three-dimensional bloom
Spectrogram picture is to get to the high spectrum image of several wave bands containing more exception information.
In said program, which further includes sorting module, for that will contain several according to abnormality detection operator PASNR
Several band images arrive minispread by exception information number according to big in the high spectrum image of wave band；
The conversion and decomposing module are additionally operable to therefrom to choose 5 band images containing maximum exception information and turn again
Become two dimensional image.
In said program, the conversion and decomposing module, specifically for according to LAKDE operators respectively to containing 5
Each section is handled in the two dimensional image of wave band, and the whole image for detecting target is obtained after processing.
In said program, the conversion and decomposing module are specifically used for the dualism hypothesis discriminant function of the LAKDE operators
For：
Wherein x is sample point to be detected, xiFor other sample points around, N is pixel number, is kernel function, is used for certainly
Determine the shape in domain around the data sample in data space, h (k, x) is the bandwidth of kernel function, and η is discrimination threshold, is being assumed to be
H1In the case of, ΛLAKDE-AD(x) it is more than η, x is abnormal object；It is being assumed to be H0In the case of, ΛLAKDE-AD(x) it is less than η, x is
Background dot.
Compared with prior art, beneficial effects of the present invention：
The present invention can improve the detection result to hyperspectral image abnormal target, make background by stronger inhibition, different
Normal target is more clear；It realizes easy feature, a large amount of redundancy wave band and noise information is eliminated, to be effectively improved
The efficiency of abnormality detection.
Description of the drawings
Fig. 1 is the implementation flow chart of the present invention；
Fig. 2 is the image that 5 wave bands after dimensionality reduction are obtained by NMF methods；
Fig. 3 is the image obtained after NMF-LAKDE methods are handled；
Fig. 4 is the ROC curve of RX, PCA-RX, ICA-RX and NMF-LAKDE；
The high spectrum image and the image after detection that Fig. 5 is embedded man-made target；
Fig. 6 is the Double-histogram of RX methods and NMF-LAKDE methods.
Specific implementation mode
The following describes the present invention in detail with reference to the accompanying drawings and specific embodiments.
The embodiment of the present invention provides a kind of method for detecting abnormality based on Non-negative Matrix Factorization, as shown in Figure 1, this method is logical
Cross following steps realization：
Step 101：The high spectrum image that pretreatment obtains cancelling noise is carried out to the high spectrum image of reading.
Specifically, obtaining high spectrum image, the spectrum wave of the high spectrum image by image collecting devices such as cameras
Contain 224 wave bands in segment table, the corresponding spectral image data of each pixel is extracted from the high spectrum image, preserves extraction
Obtained spectral image data；
One size of interception is the portion intercepts image of 192 × 90 pixels from whole high spectrum images, to the portion
Divide interception image to be pre-processed, that is, reject the wave band of strong noise and not exception information, after processing, obtains residue and contain 178
The high spectrum image of a wave band.
Step 102：The high spectrum image of the cancelling noise of the acquisition is converted into row vector and obtains two dimension initialization square
Battle array V, then linear decomposition is carried out to two dimension initialization matrix V and generates random initializtion basic matrix W and coefficient matrix H.
Specifically, three-dimensional is transformed into two dimension first, the two dimensional image of each wave band is c × p dimensions, and total n wave band turns to
It is tieed up for m × n after two-dimensional matrix, wherein m=c × p；Such as：Assuming that the image of each wave band is MI1 × MI2 dimensions, total MI3 wave
Section is tieed up after turning to two-dimensional matrix for Mn × MI3, wherein Mn=MI1 × MI2.
According to formula Vm×n≈Wm×rHr×nLinear decomposition is carried out, random initializtion basic matrix W and coefficient matrix H, institute are generated
It is all non-negative to state random initializtion basic matrix W and coefficient matrix H；Such as：W ties up for MI3 × r, and H ties up for r × Mn, and r is determined
Band image number after last obtained dimensionality reduction, it is maximum different without losing in order to reject the meaningless wave band of all redundancies
Normal information, compromise take r=5.
Step 103：According to the multiplying property iterative rules of Non-negative Matrix Factorization to random initializtion basic matrix W and coefficient matrix H
It is iterated and obtains the high spectrum image containing several wave bands.
Specifically, the random initializtion base generated respectively to step 102 according to the multiplying property iterative rules of Non-negative Matrix Factorization
Matrix W and coefficient matrix H are iterated, and the multiplying property iterative rules of the Non-negative Matrix Factorization are respectively：
Wherein i, k, j respectively indicate line number and columns, matrix WT、
HTThe respectively transposed matrix of matrix W and T；
According to square of the Euclidean distance between two dimension initialization matrix V and random initializtion basic matrix W and coefficient matrix H
As object function come weigh decompose after result and original matrix degree of closeness, given threshold η, by after each iteration with
Machine initializes basic matrix W and coefficient matrix H and is sent into object function, obtains iteration result, if the iteration result is returned more than η
Continue iteration；If the iteration result is less than η, stopping is iterated, and coefficient matrix H is changed into three-dimensional bloom
Spectrogram picture is to get to the high spectrum image of several wave bands containing more exception information.
Square of Euclidean distance of the object function between V and WH：
min||V-(WH)||2=Σi,j(Vi,j-(WH)ij)2
Using this optimization problem of gradient method solution, its most fast descent direction is found out first:
The step-length is taken to be respectively
W can be obtained, the rule of iteration of H is：
The object function can be the range formula based on K-L divergences,
Iterative rules are obtained using same method：
From iterative formula it will also be seen that, if WikInitial value will then cause the value of every step iteration big greatly, if Wik's
The small then result of initial value is opposite, so the initial value of W and H will make a big impact to the result of decomposition.The selection of initial value simultaneously is also
Will be to convergent speed, the effect etc. approached has an impact, so in the case of different, to choose suitable initial value.
Further include after step 103：It will be in the high spectrum image containing several wave bands according to abnormality detection operator PASNR
Several band images arrive minispread by exception information number according to big, and therefrom choose 5 waves containing maximum exception information
Section image is again transformed into two dimensional image, as shown in Figure 2；Such as：W is transformed back into again for the 3-D view of MI1 × MI2 × r,
Medium wave hop count is r.
Step 104：Finally according to the adaptive Density Estimator operator of part to the height containing several wave bands
High spectrum image containing a most wave band of exception information in spectrum picture is handled, and acquisition detects abnormal object
Image.
Specifically, being handled respectively each section in the two dimensional image containing 5 wave bands according to LAKDE operators, locate
The whole image for detecting target is obtained after reason.
The dualism hypothesis discriminant function of the LAKDE operators is：
Wherein x is sample point to be detected, xiFor other sample points around, N is pixel number, and κ is kernel function, is used for certainly
Determine the shape in domain around the data sample in data space, h (k, x) is the bandwidth of kernel function, and η is discrimination threshold, is being assumed to be
H1In the case of, ΛLAKDE-AD(x) it is more than η, x is abnormal object；It is being assumed to be H0In the case of, ΛLAKDE-AD(x) it is less than η, x is
Background dot.
The selection mode of the bandwidth of the kernel function is：
I.e. band width selection method is the information compression the local auto-adaptive extracted by KNN into an individual band
Width values, method are taken about M all image pixelsRk(x) intermediate value.
The effect of the present invention can be further illustrated by following simulation result：
Emulation 1：
The picture of 192 × 90 pixels on an airport is intercepted on the image of AVIRIS, totally 178 wave bands, in NMF methods
In, r=5 is set, the image of 5 wave bands shown in Fig. 2 is obtained after calculating.
Emulation 2：
In the image of finally obtained 5 wave bands of method, being chosen with PASNR exception informations evaluation operator should be containing abnormal letter
The most abundant wave band of breath, and LAKDE method processing is carried out to this piece image, it obtains background shown in Fig. 3 and is suppressed, only show
The image of final goal, it will be clear that six bright target points and a dark target point from figure, this be by
Caused by spectral characteristic difference between target.
Emulation 3：
As shown in figure 5, embedded 5 man-made targets in a panel height spectrum picture, Fig. 5 (a) is the original of one of wave band
Beginning image, due to containing noise wave band, therefore the wave band of target location by reference can be more clearly seen that by picking one
Image.Background is farmland, and the pixel value of 5 targets is different；Fig. 5 (b) be directly with RX methods treated image, can be with
See, the background of image is suppressed after treatment, but target is also inhibited simultaneously, becomes dim and very unintelligible, also
There are 3 obvious false-alarms；Fig. 5 (c) be after being handled with NMF methods containing exception information do more ingredients use again LAKDE into
Row treated image can be clear that each target in figure, and not have false-alarm, better.
Evaluation 1：
It is bent that following Fig. 4 depicts RX methods, PCA-RX methods, ICA-RX methods, the ROC of NMF-LAKDE methods respectively
Line chart；4 ROC curves are drawn under the same coordinate system, it can clearly be seen that the RX method effects under identical false alarm rate
Worst, the verification and measurement ratio of ICA-RX is higher than PCA-RX, and for NMF-LAKDE, its ROC curve substantially with ICA-RX phases
Together, lower in false alarm rate, verification and measurement ratio is slightly above ICA-RX methods, equally when false alarm rate is more than 10-2When, NMF-
The efficiency of LAKDE methods is slightly higher.Moreover, the result of NMF algorithms as ICA algorithm without having very strong randomness, LAKDE
Method for detecting abnormality is also more efficient than traditional RX method for detecting abnormality.
Evaluation 2：
Fig. 6 is respectively the Double-histogram of image after above-mentioned direct RX and NMF-LAKDE are handled.Fig. 6 (a) is at direct RX
Image after reason illustrates only 1 target, has 2 false-alarm points to be also revealed；Fig. 6 (b) is that treated by NMF-LAKDE
Image, 5 targets all show do not have false-alarm point, target and background point also to reach good separation.
The present invention is with traditional RX hyperspectral abnormity detection methods and with being carried out abnormality detection again after PCA or ICA dimensionality reductions
Method compare, have following remarkable advantage：
1) present invention provides a kind of new dimension reduction method using the method for Non-negative Matrix Factorization to high-spectrum image dimensionality reduction
Thinking, since negative value does not often have physical significance in practical problem in matrix, and NMF occurs without negative value in a matrix, because
This decomposition method realizes simplicity, and has specific physical significance.
2) memory space occupied by method of the invention is few, and arithmetic speed is fast, and the ability for obtaining exception information is high so that drop
That ties up is more efficient, and acquired results are more excellent.
3) for the present invention compared with traditional NMF methods, the present invention evaluates operator by the way that an exception information is added, and solves
Result generates the problem of randomness so that several dimensionality reduction wave bands after superposition are arranged from big to small by exception information, carry
The high reliability of method.
4) the LAKDE operators used in the present invention are fundamentally based on the expansion and extension of traditional RX methods, it is using part
Double window vocal imitation skill, while kernel function is added, based in the selection method of the bandwidth of kernel function in KNN arest neighbors methods,
Effect is more preferable than traditional RX methods.
The embodiment of the present invention also provides a kind of abnormal detector based on Non-negative Matrix Factorization, which includes：Pre- place
Manage module, conversion and decomposing module, iteration module, detection module；
The preprocessing module carries out the high-spectrum that pretreatment obtains cancelling noise for the high spectrum image to reading
Picture；
The conversion and decomposing module, for obtaining the high spectrum image of the cancelling noise of the acquisition into row vector conversion
Two-dimentional initialization matrix V is obtained, then linear decomposition is carried out to two dimension initialization matrix V and generates random initializtion basic matrix W and coefficient
Matrix H；
The iteration module, for random initializtion basic matrix W and coefficient matrix H to be iterated and obtained according to rule of iteration
The high spectrum image containing several wave bands is obtained,
The detection module, for finally containing several to described according to the adaptive Density Estimator operator of part
High spectrum image containing a most wave band of exception information in the high spectrum image of wave band is handled, and acquisition detects different
The image of normal target.
The iteration module, particularly for according to the multiplying property iterative rules of Non-negative Matrix Factorization respectively to random initializtion
Basic matrix W and coefficient matrix H are iterated, and the multiplying property iterative rules of the Non-negative Matrix Factorization are respectively：Wherein i, k, j respectively indicate line number and columns, matrix WT、HTRespectively
The transposed matrix of matrix W and T；
According to square of the Euclidean distance between two dimension initialization matrix V and random initializtion basic matrix W and coefficient matrix H
As object function come weigh decompose after result and original matrix degree of closeness, given threshold η, by after each iteration with
Machine initializes basic matrix W and coefficient matrix H and is sent into object function, obtains iteration result, if the iteration result is returned more than η
Continue iteration；If the iteration result is less than η, stopping is iterated, and coefficient matrix H is changed into three-dimensional bloom
Spectrogram picture is to get to the high spectrum image of several wave bands containing more exception information.
The invention also includes sorting module, for according to abnormality detection operator PASNR by the bloom containing several wave bands
Several band images arrive minispread by exception information number according to big in spectrogram picture；
The conversion and decomposing module are additionally operable to therefrom to choose 5 band images containing maximum exception information and turn again
Become two dimensional image.
The conversion and decomposing module are specifically used for according to LAKDE operators respectively to the X-Y scheme containing 5 wave bands
Each section is handled as in, and the whole image for detecting target is obtained after processing.
The conversion and decomposing module, the dualism hypothesis discriminant function specifically for the LAKDE operators are：
Wherein x is sample point to be detected, xiFor other sample points around, N is pixel number, and κ is kernel function, is used for certainly
Determine the shape in domain around the data sample in data space, h (k, x) is the bandwidth of kernel function, and η is discrimination threshold, is being assumed to be
H1In the case of, ΛLAKDE-AD(x) it is more than η, x is abnormal object；It is being assumed to be H0In the case of, ΛLAKDE-AD(x) it is less than η, x is
Background dot.
The foregoing is only a preferred embodiment of the present invention, is not intended to limit the scope of the present invention.
Claims (2)
1. a kind of method for detecting abnormality based on Non-negative Matrix Factorization, which is characterized in that this method is：To the high-spectrum of reading
The high spectrum image that cancelling noise is obtained as carrying out pretreatment, by the high spectrum image of the cancelling noise of the acquisition into row vector
Conversion obtains two dimension initialization matrix V, then carries out linear decomposition to two dimension initialization matrix V and generate random initializtion basic matrix W
With coefficient matrix H, random initializtion basic matrix W and coefficient matrix H are carried out according to the multiplying property iterative rules of Non-negative Matrix Factorization
Iteration obtains the high spectrum image containing several wave bands, finally according to the adaptive Density Estimator operator of part to described
High spectrum image containing a most wave band of exception information in high spectrum image containing several wave bands is handled, and is obtained
It must detect the image of abnormal object；It is described that random initializtion basic matrix W and coefficient matrix H are iterated according to rule of iteration
The high spectrum image containing each and every one several wave bands is obtained, specially：It is right respectively according to the multiplying property iterative rules of Non-negative Matrix Factorization
Random initializtion basic matrix W and coefficient matrix H are iterated, and the multiplying property iterative rules of the Non-negative Matrix Factorization are respectively：
Wherein i, k, j respectively indicate line number and columns, matrix WT、HTRespectively
The transposed matrix of matrix W and T；
According to square conduct of the Euclidean distance between two dimension initialization matrix V and random initializtion basic matrix W and coefficient matrix H
Object function come weigh decompose after result and original matrix degree of closeness, given threshold η, will after each iteration it is random just
Beginningization basic matrix W is sent into object function with coefficient matrix H, obtains iteration result, if the iteration result is more than η, return continues
It is iterated；If the iteration result is less than η, stop iteration, coefficient matrix H be changed into three-dimensional high spectrum image,
Obtain the high spectrum image of several wave bands containing more exception information；
It is described according to rule of iteration to random initializtion basic matrix W and coefficient matrix H be iterated acquisition contain several wave bands
High spectrum image, later this method further include：According to abnormality detection operator PASNR by the high-spectrum containing several wave bands
Several band images arrive minispread by exception information number according to big as in, and therefrom choose 5 containing maximum exception information
Band image be again transformed into two dimensional image；
It is described finally according to part adaptive Density Estimator operator to the high spectrum image containing several wave bands
In the high spectrum image containing a most wave band of exception information handled, obtain the image for detecting abnormal object, tool
Body is：Each section in the two dimensional image containing 5 wave bands is handled respectively according to LAKDE operators, is obtained after processing whole
The image for detecting target of body；
It is described that each section in the two dimensional image containing 5 wave bands is handled respectively according to LAKDE operators, it is obtained after processing
The whole image for detecting target, specially：The dualism hypothesis discriminant function of the LAKDE operators is：
Wherein x is sample point to be detected, xiFor other sample points around, N is pixel number, and κ is kernel function, for determining number
According to the shape in domain around the data sample in space, h (k, x) is the bandwidth of kernel function, and η is discrimination threshold, is being assumed to be H1's
In the case of, Λ LAKDE-AD (x) are more than η, and x is abnormal object；In the case where being assumed to be H0, Λ LAKDE-AD (x) are less than η, x
For background dot；
The selection mode of the bandwidth of the kernel function is：
That is band width selection method by KNN being extracted
For the information compression of local auto-adaptive at an individual bandwidth value, method is the r taken about M all image pixelsk(x)
Intermediate value.
2. a kind of abnormal detector based on Non-negative Matrix Factorization, which is characterized in that the device includes：Preprocessing module turns
It changes and decomposing module, iteration module, detection module；
The preprocessing module carries out the high spectrum image that pretreatment obtains cancelling noise for the high spectrum image to reading；
The conversion and decomposing module, for the high spectrum image of the cancelling noise of the acquisition to be obtained two into row vector conversion
Dimension initialization matrix V, then linear decomposition is carried out to two dimension initialization matrix V and generates random initializtion basic matrix W and coefficient matrix
H；
The iteration module, for random initializtion basic matrix W and coefficient matrix H to be iterated and contained according to rule of iteration
There are the high spectrum image of several wave bands, the detection module to be calculated for the adaptive Density Estimator finally according to part
Son to the high spectrum image containing a most wave band of exception information in the high spectrum image containing several wave bands into
Row processing, obtains the image for detecting abnormal object；
The iteration module, particularly for according to the multiplying property iterative rules of Non-negative Matrix Factorization respectively to random initializtion group moment
Battle array W and coefficient matrix H is iterated, and the multiplying property iterative rules of the Non-negative Matrix Factorization are respectively：
According to square conduct of the Euclidean distance between two dimension initialization matrix V and random initializtion basic matrix W and coefficient matrix H
Object function come weigh decompose after result and original matrix degree of closeness, given threshold η, will after each iteration it is random just
Beginningization basic matrix W is sent into object function with coefficient matrix H, obtains iteration result, if the iteration result is more than η, return continues
It is iterated；If the iteration result is less than η, stopping is iterated, and coefficient matrix H is changed into three-dimensional high-spectrum
As to get to the high spectrum image of several wave bands containing more exception information；
The device further includes sorting module, for according to abnormality detection operator PASNR by the high-spectrum containing several wave bands
Several band images arrive minispread by exception information number according to big as in；
The conversion and decomposing module are additionally operable to therefrom to choose 5 band images containing maximum exception information and are again transformed into
Two dimensional image；
The conversion and decomposing module are specifically used for according to LAKDE operators respectively in the two dimensional image containing 5 wave bands
Each section is handled, and the whole image for detecting target is obtained after processing；
The conversion and decomposing module, the dualism hypothesis discriminant function specifically for the LAKDE operators are：
Wherein x is sample point to be detected, xiFor other sample points around, N is pixel number, and κ is kernel function, for determining number
According to the shape in domain around the data sample in space, h (k, x) is the bandwidth of kernel function, and η is discrimination threshold, is being assumed to be H1's
In the case of, Λ LAKDE-AD (x) are more than η, and x is abnormal object；In the case where being assumed to be H0, Λ LAKDE-AD (x) are less than η, x
For background dot；
The selection mode of the bandwidth of the kernel function is：
I.e. band width selection method is that handle is carried by KNN
For the information compression of the local auto-adaptive taken out at an individual bandwidth value, method is taken about M all image slices
The r of elementk(x) intermediate value.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510191275.2A CN104778706B (en) | 2015-04-21 | 2015-04-21 | Method for detecting abnormality based on Non-negative Matrix Factorization and its device |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510191275.2A CN104778706B (en) | 2015-04-21 | 2015-04-21 | Method for detecting abnormality based on Non-negative Matrix Factorization and its device |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN104778706A CN104778706A (en) | 2015-07-15 |
| CN104778706B true CN104778706B (en) | 2018-10-30 |
Family
ID=53620152
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201510191275.2A Active CN104778706B (en) | 2015-04-21 | 2015-04-21 | Method for detecting abnormality based on Non-negative Matrix Factorization and its device |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN104778706B (en) |
Families Citing this family (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105427300B (en) * | 2015-12-21 | 2018-08-07 | 复旦大学 | A kind of hyperspectral image abnormal detection method based on low-rank representation and study dictionary |
| CN106156728B (en) * | 2016-06-24 | 2019-09-10 | 湖北久之洋红外***股份有限公司 | HYPERSPECTRAL IMAGERY dimension reduction method and system based on spectral space decomposition and noise contribution analysis |
| CN107067040A (en) * | 2017-05-11 | 2017-08-18 | 攀枝花学院 | Multispectral remote sensing image classification method based on tensor rarefaction representation and cluster |
| CN109815825B (en) * | 2018-12-27 | 2021-01-15 | 中国科学院西安光学精密机械研究所 | Hyperspectral abnormal target detection method based on similarity constrained convex-nonnegative matrix decomposition |
| CN109885029B (en) * | 2019-03-13 | 2021-07-06 | 东北大学 | Electric smelting magnesium oxide process fault detection method based on projection non-negative matrix decomposition |
| CN113553914B (en) * | 2021-06-30 | 2024-03-19 | 核工业北京地质研究院 | CASI hyperspectral data abnormal target detection method |
| CN114370820B (en) * | 2022-03-22 | 2022-07-01 | 武汉精立电子技术有限公司 | Peak extraction method, detection method and system of spectrum confocal displacement sensor |
Family Cites Families (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101221243B (en) * | 2007-11-01 | 2011-12-07 | 复旦大学 | Remote sensing image mixed pixels decomposition method based on nonnegative matrix factorization |
| US8670628B2 (en) * | 2011-08-16 | 2014-03-11 | Raytheon Company | Multiply adaptive spatial spectral exploitation |
| CN103870807A (en) * | 2014-02-25 | 2014-06-18 | 杭州电子科技大学 | High spectrum mixed nuclear RX anomaly detection method |
| CN104408705B (en) * | 2014-09-23 | 2017-10-20 | 西安电子科技大学 | A kind of method for detecting abnormality of high spectrum image |
-
2015
- 2015-04-21 CN CN201510191275.2A patent/CN104778706B/en active Active
Non-Patent Citations (3)
| Title |
|---|
| A Locally Adaptive Background Density Estimator An Evolution for RX-Based Anomaly Detectors;Stefania Matteoli et al.;《IEEE Geoscience and Remote Sensing Letters》;20140131;第11卷(第1期);第324页第II节 * |
| 基于光谱特性的高光谱图像异常目标检测算法研究;成宝芝;《中国博士学位论文全文数据库.信息科技辑》;20140415(第04期);第42页3.4节第2段，第90页5.3.1节第1段，第102页第5.4.3节第1-2段 * |
| 基于非负矩阵分解的高光谱图像解混技术研究;吕亚丽;《中国优秀硕士学位论文全文数据库.信息科技辑》;20110615(第06期);第11页第2章第1段、第2.1.2节第2段，第12页第2.1.2.1节第2段 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN104778706A (en) | 2015-07-15 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN104778706B (en) | Method for detecting abnormality based on Non-negative Matrix Factorization and its device | |
| Zhang et al. | Hyperspectral classification based on lightweight 3-D-CNN with transfer learning | |
| Liu et al. | Tensor matched subspace detector for hyperspectral target detection | |
| Gu et al. | A selective KPCA algorithm based on high-order statistics for anomaly detection in hyperspectral imagery | |
| Chang et al. | Linear spectral random mixture analysis for hyperspectral imagery | |
| Huang et al. | Robust face detection using Gabor filter features | |
| Jablonski et al. | Principal component reconstruction error for hyperspectral anomaly detection | |
| Prasad et al. | Information fusion in the redundant-wavelet-transform domain for noise-robust hyperspectral classification | |
| Geng et al. | CEM: more bands, better performance | |
| Zhang et al. | A nonlinear sparse representation-based binary hypothesis model for hyperspectral target detection | |
| CN106778680B (en) | A kind of hyperspectral image band selection method and device based on critical bands extraction | |
| CN101807301A (en) | High spectral image target detection method based on high order statistic | |
| CN102592280B (en) | Hyperspectral image anomaly detection method using multi-window feature analysis | |
| Zhang et al. | Regularization framework for target detection in hyperspectral imagery | |
| Du et al. | A signal-decomposed and interference-annihilated approach to hyperspectral target detection | |
| Zhang et al. | Infrared small target detection via interpatch correlation enhancement and joint local visual saliency prior | |
| Zhai et al. | A novel sense-through-foliage target recognition system based on sparse representation and improved particle swarm optimization-based support vector machine | |
| Yadav et al. | An improved illumination normalization and robust feature extraction technique for face recognition under varying illuminations | |
| Akhtar et al. | Adversarial examples detection using no-reference image quality features | |
| Matteoli et al. | Automatic target recognition within anomalous regions of interest in hyperspectral images | |
| Ling et al. | A constrained sparse-representation-based binary hypothesis model for target detection in hyperspectral imagery | |
| Jiao et al. | Triplet Spectral-Wise Transformer Network for Hyperspectral Target Detection | |
| CN105825512B (en) | The abnormal target in hyperspectral remotely sensed image object detection method returned based on steady background | |
| Abdul-Haleem et al. | Fingerprint recognition using haar wavelet transform and local ridge attributes only | |
| Bhairannawar et al. | FPGA implementation of face recognition system using efficient 5/3 2D-lifting scheme |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| EXSB | Decision made by sipo to initiate substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |