CN103268593B - The separation method of signal and noise in a kind of target in hyperspectral remotely sensed image - Google Patents

Abstract

The present invention relates to remote sensing technology field, particularly relate to the image treatment method signal in target in hyperspectral remotely sensed image and noise being easily separated. The separation method of signal and noise in this target in hyperspectral remotely sensed image, specifically includes: utilize low-pass filtering and homogeneity piecemeal, estimates signal and noise respectively; According to bayesian criterion, quantitative Analysis obtains the dimension of signal and the noise estimated; Build signal subspace and noise subspace, and by oblique subspace projection, separate the signal component in original target in hyperspectral remotely sensed image and noise component. The separation method of signal and noise in target in hyperspectral remotely sensed image provided by the invention, it is possible to the signal of quantitative Analysis estimation and the dimension of noise, it is ensured that the dimension obtained is not affected by subjective factors; The use of oblique subspace projection has fully taken into account the dependency between noise and signal, so that signal and noise can well separate.

Description

The separation method of signal and noise in a kind of target in hyperspectral remotely sensed image

Technical field

The present invention relates to remote sensing technology field, particularly relate to the image treatment method signal in target in hyperspectral remotely sensed image and noise being easily separated.

Background technology

The separation of signal and noise is the important technology that target in hyperspectral remotely sensed image processes. But existing Signal De-noising Method is based on various irrational hypothesis more: ignore noise multiformity, only consider additivity or multiplicative noise; Assuming that noise is separate with signal; Assuming that each row (OK) signal obeys same mathematical distribution; Assuming that whole scape image is homogenizing or approximate homogenizing, etc. And the method for existing acquisition signal dimension and noise dimension is subject to the impact of subjective factors, reduces the separation accuracy of signal and noise. Specifically can referring to document N.Acito, M.Diani, andG.Corsini, " Subspace-BasedStripingNoiseReductioninHyperspectralImage s, " IEEETransactionsonGeoscienceandRemoteSensing, vol.49, pp.1325-1342, Apr2011.

Summary of the invention

(1) to solve the technical problem that

It is an object of the invention to provide the separation method of signal and noise in a kind of target in hyperspectral remotely sensed image, to solve existing method is ignored the dependency of noise type multiformity and noise and signal, and the problem that dimension is excessively qualitative and subjective.

(2) technical scheme

In order to solve above-mentioned technical problem, the present invention provides the separation method of signal and noise in a kind of target in hyperspectral remotely sensed image, specifically includes:

Utilize low-pass filtering and homogeneity piecemeal, estimate signal and noise respectively;

According to bayesian criterion, quantitative Analysis obtains the dimension of signal and the noise estimated;

Build signal subspace and noise subspace, and by oblique subspace projection, separate the signal component in original target in hyperspectral remotely sensed image and noise component.

Preferably, data model is set up for target in hyperspectral remotely sensed image:

$X_l(i,j)=S_l(i,j)\⊕N_l(i,j)$

Wherein, X_kRepresent that wave band number is target in hyperspectral remotely sensed image data during k; S_k(i, j) represents signal; N_k(i, j) represents noise; K=1 ..., N_B, i=1 ..., N_L, j=1 ..., N_S.; SymbolRepresent direct sum;

Preferably, described estimation signal and noise specifically include:

The operator that low-pass filtering is selected is: $\left[\begin{array}{ccc}1/9& 1/9& 1/9\\ 1/9& 1/9& 1/9\\ 1/9& 1/9& 1/9\end{array}\right],$ Using low-frequency information as the estimation to signal;

Spatial continuity according to atural object and wave spectrum dependency, given less wave spectrum angle threshold value, image is carried out piecemeal;

The Noise Estimation of kth wave band the n-th piecemeal is C_nk(1-r_nk), wherein C_nkFor the pixel average of kth wave band the n-th piecemeal, r_nkMultiple correlation coefficient for kth wave band the n-th piecemeal with adjacent two wave band correspondence position piecemeals.

Preferably, pixel number contained by selected piecemeal must not less than 30.

Preferably, described obtain, according to bayesian criterion quantitative Analysis, the signal estimated and the dimension of noise specifically includes:

The bayesian criterion used is integrated with minimal error rate criterion and Neyman Pearson (N-P) criterion;

P is set_MRepresent false dismissal probability, P_FRepresenting false-alarm probability, �� represents a certain eigenvalue of covariance matrix, ��₁Represent noise component, ��₂Expression signal component, and p (�� | ��₁) and p (�� | ��₂) represent that �� belongs to approximate condition probability density corresponding when noise or signal, p (�� respectively₁| ��) and p (��₂| ��) represent that �� belongs to posterior probability density corresponding when noise or signal respectively;

Setting make p (�� | ��₁) and p (�� | ��₂) closest to eigenvalue ��₁For meeting the noise separation of minimal error rate criterion, now corresponding P_FIgnoring, the marginal value of N-P criterion is:

${\mathrm{\δ}}_{\mathrm{NP}}=\frac{p\left({\mathrm{\λ}}_1\right|{\mathrm{\ω}}_1)}{p\left({\mathrm{\λ}}_1\right|{\mathrm{\ω}}_2)}$

For arbitrary wave band, if $\frac{p\left(\mathrm{\&lambda;}\right|{\mathrm{\&omega;}}_1)}{p\left(\mathrm{\&lambda;}\right|{\mathrm{\&omega;}}_2)}\begin{array}{c}>\\ <\end{array}{\mathrm{\&delta;}}_{\mathrm{NP}},$ Then $\mathrm{\&lambda;}\&Element;\left\{\begin{array}{c}{\mathrm{\&omega;}}_1\\ {\mathrm{\&omega;}}_2\end{array}\right..$

Preferably, described set up signal subspace and noise subspace and separate the signal in original Hyperspectral imaging by oblique subspace projection and noise specifically includes:

The characteristic vector corresponding with the eigenvalue belonging to signal or noise constitutes the base S of corresponding signal subspace or noise subspace_basisOr N_basis, the orthogonal intersection cast shadow matrix of signal subspace is

S_orthproject=E-S_basis(S_basis ^HS_basis)^-1S_basis ^H

Wherein, E representation unit matrix, in like manner obtain the orthogonal intersection cast shadow matrix N of noise subspace_orthproject;

Parallel noise subspace to the oblique projection matrix of signal subspace is:

P_SN=S_basis(S_basis ^HN_orthprojectS_basis)^-1S_basis ^HN_orthproject

Signal is:

S=P_SNX;

Parallel signal subspace to the oblique projection matrix of noise subspace is

P_NS=N_basis(N_basis ^HS_orthprojectN_basis)^-1N_basis ^HS_orthproject

Noise is

N=P_NSX��

(3) beneficial effect

The separation method of signal and noise in target in hyperspectral remotely sensed image provided by the invention, it is possible to the signal of quantitative Analysis estimation and the dimension of noise, it is ensured that the dimension obtained is not affected by subjective factors; The use of oblique subspace projection has fully taken into account the dependency between noise and signal, so that signal and noise can well separate.

Accompanying drawing explanation

Fig. 1 is the flow chart of the separation method of signal and noise in target in hyperspectral remotely sensed image of the present invention;

First embodiment that Fig. 2 (a)-Fig. 2 (c) is the present invention: the result image of the original image of certain AVIRIS data the 218th wave band (centre wavelength: 2439.81nm) and the signal obtained after utilizing Fig. 1 method and noise;

Second embodiment that Fig. 3 (a)-Fig. 3 (c) is the present invention: the result image of the original image of certain Hyperion data the 57th wave band (centre wavelength: 925.41nm) and the signal obtained after utilizing Fig. 1 method and noise.

Detailed description of the invention

Following example are used for illustrating the present invention, but are not limited to the scope of the present invention.

Fig. 1 is the flow chart of the separation method of signal and noise in target in hyperspectral remotely sensed image of the present invention, comprises the following steps:

Step1, estimate signal and noise respectively from EO-1 hyperion raw video;

Step2, pass judgment on signal and the dimension of noise that criterion quantitative Analysis obtains estimating according to Bayes. In first embodiment, the dimension of Signal estimation is 9, and the dimension of Noise Estimation is 23; In second embodiment, the dimension of Signal estimation is 4, and the dimension of Noise Estimation is 149;

Step3, set up signal subspace and noise subspace, the oblique projection matrix of signal calculated subspace and noise subspace, by raw video oblique projection to different subspace, respectively obtain signal and noise. Being computed, in first embodiment, the subspace angle of signal subspace and noise subspace is 86.77 ��, and in second embodiment, the subspace angle of signal subspace and noise subspace is 72.26 ��, and signal and noise are also not completely independent. The result image of the raw video of Fig. 2 (a)-Fig. 2 (c) and Fig. 3 a (a)-Fig. 3 (c) respectively certain wave band of first embodiment and second embodiment and the signal obtained after utilizing Fig. 1 method and noise.

Introduce the separation method of signal and noise in this target in hyperspectral remotely sensed image in detail below, comprising:

Utilize low-pass filtering and homogeneity piecemeal, estimate signal and noise respectively;

According to bayesian criterion, quantitative Analysis obtains the dimension of signal and the noise estimated;

Build signal subspace and noise subspace, and by oblique subspace projection, separate the signal component in original target in hyperspectral remotely sensed image and noise component.

Wherein, data model is set up for target in hyperspectral remotely sensed image:

$X_l(i,j)=S_l(i,j)\&CirclePlus;N_l(i,j)$

Wherein, X_kRepresent that wave band number is target in hyperspectral remotely sensed image data during k; S_k(i, j) represents signal; N_k(i, j) represents noise; K=1 ..., N_B, i=1 ..., N_L, j=1 ..., N_S.; SymbolRepresent direct sum; N_BRepresent wave band number, N_LAnd N_SRepresent ranks number respectively.

Wherein, described estimation signal and noise specifically include:

The operator that low-pass filtering is selected is: $\left[\begin{array}{ccc}1/9& 1/9& 1/9\\ 1/9& 1/9& 1/9\\ 1/9& 1/9& 1/9\end{array}\right],$ Using low-frequency information as the estimation to signal;

Spatial continuity according to atural object and wave spectrum dependency, given less wave spectrum angle threshold value, image is carried out piecemeal;

The Noise Estimation of kth wave band the n-th piecemeal is C_nk(1-r_nk), wherein C_nkFor the pixel average of kth wave band the n-th piecemeal, r_nkMultiple correlation coefficient for kth wave band the n-th piecemeal with adjacent two wave band correspondence position piecemeals. Wherein, pixel number contained by selected piecemeal must not less than 30.

Wherein, described obtain, according to bayesian criterion quantitative Analysis, the signal estimated and the dimension of noise specifically includes:

The bayesian criterion used is integrated with minimal error rate criterion and Neyman Pearson (N-P) criterion;

P is set_MRepresent false dismissal probability, P_FRepresenting false-alarm probability, �� represents a certain eigenvalue of covariance matrix, ��₁Represent noise component, ��₂Expression signal component, and p (�� | ��₁) and p (�� | ��₂) represent that �� belongs to approximate condition probability density corresponding when noise or signal, p (�� respectively₁| ��) and p (��₂| ��) represent that �� belongs to posterior probability density corresponding when noise or signal respectively;

Setting make p (�� | ��₁) and p (�� | ��₂) closest to eigenvalue ��₁For meeting the noise separation of minimal error rate criterion, now corresponding P_FIgnoring, the marginal value of N-P criterion is:

${\mathrm{\&delta;}}_{\mathrm{NP}}=\frac{p\left({\mathrm{\&lambda;}}_1\right|{\mathrm{\&omega;}}_1)}{p\left({\mathrm{\&lambda;}}_1\right|{\mathrm{\&omega;}}_2)}$

For arbitrary wave band, if $\frac{p\left(\mathrm{\&lambda;}\right|{\mathrm{\&omega;}}_1)}{p\left(\mathrm{\&lambda;}\right|{\mathrm{\&omega;}}_2)}\begin{array}{c}>\\ <\end{array}{\mathrm{\&delta;}}_{\mathrm{NP}},$ Then $\mathrm{\&lambda;}\&Element;\left\{\begin{array}{c}{\mathrm{\&omega;}}_1\\ {\mathrm{\&omega;}}_2\end{array}\right..$

Wherein, described set up signal subspace and noise subspace and separate the signal in original Hyperspectral imaging by oblique subspace projection and noise specifically includes:

The characteristic vector corresponding with the eigenvalue belonging to signal or noise constitutes the base S of corresponding signal subspace or noise subspace_basisOr N_basis, the orthogonal intersection cast shadow matrix of signal subspace is

S_orthproject=E-S_basis(S_basis ^HS_basis)^-1S_basis ^H

Wherein, E representation unit matrix, in like manner obtain the orthogonal intersection cast shadow matrix N of noise subspace_orthproject;

Parallel noise subspace to the oblique projection matrix of signal subspace is:

P_SN=S_basis(S_basis ^HN_orthprojectS_basis)^-1S_basis ^HN_orthproject

Signal is:

S=P_SNX;

Parallel signal subspace to the oblique projection matrix of noise subspace is

P_NS=N_basis(N_basis ^HS_orthprojectN_basis)^-1N_basis ^HS_orthproject

Noise is

N=P_NSX, X represent target in hyperspectral remotely sensed image data.

The separation method of signal and noise in target in hyperspectral remotely sensed image provided by the invention, it is possible to the signal of quantitative Analysis estimation and the dimension of noise, it is ensured that the dimension obtained is not affected by subjective factors; The use of oblique subspace projection has fully taken into account the dependency between noise and signal, so that signal and noise can separate well.

The above is only the preferred embodiment of the present invention; it should be pointed out that, for those skilled in the art, under the premise without departing from the technology of the present invention principle; can also making some improvement and replacement, these improve and replace and also should be regarded as protection scope of the present invention.

Claims (3)

1. the separation method of signal and noise in a target in hyperspectral remotely sensed image, it is characterised in that: specifically include:

Utilize low-pass filtering and homogeneity piecemeal, estimate signal and noise respectively;

According to bayesian criterion, quantitative Analysis obtains the dimension of signal and the noise estimated;

Build signal subspace and noise subspace, and by oblique subspace projection, separate the signal component in original target in hyperspectral remotely sensed image and noise component;

Described estimation signal and noise specifically include:

The operator that low-pass filtering is selected is: $\left[\begin{array}{ccc}1/9& 1/9& 1/9\\ 1/9& 1/9& 1/9\\ 1/9& 1/9& 1/9\end{array}\right],$ Using low-frequency information as the estimation to signal;

Spatial continuity according to atural object and wave spectrum dependency, given less wave spectrum angle threshold value, image is carried out piecemeal;

The Noise Estimation of kth wave band the n-th piecemeal is C_nk(1-r_nk), wherein C_nkFor the pixel average of kth wave band the n-th piecemeal, r_nkMultiple correlation coefficient for kth wave band the n-th piecemeal with adjacent two wave band correspondence position piecemeals;

Described obtain, according to bayesian criterion quantitative Analysis, the signal estimated and the dimension of noise specifically includes:

The bayesian criterion used is integrated with minimal error rate criterion and Neyman Pearson (N-P) criterion;

P is set_MRepresent false dismissal probability, P_FRepresenting false-alarm probability, �� represents a certain eigenvalue of covariance matrix, ��₁Represent noise component, ��₂Expression signal component, and p (�� | ��₁) and p (�� | ��₂) represent that �� belongs to approximate condition probability density corresponding when noise or signal, p (�� respectively₁| ��) and p (��₂| ��) represent that �� belongs to posterior probability density corresponding when noise or signal respectively;

Setting make p (�� | ��₁) and p (�� | ��₂) closest to eigenvalue ��₁For meeting the noise separation of minimal error rate criterion, now corresponding P_FIgnoring, the marginal value of N-P criterion is:

${\mathrm{\&delta;}}_{NP}=\frac{p\left({\mathrm{\&lambda;}}_1\right|{\mathrm{\&omega;}}_1)}{p\left({\mathrm{\&lambda;}}_1\right|{\mathrm{\&omega;}}_2)}$

For arbitrary wave band, ifThen $\mathrm{\&lambda;}\&Element;\left\{\begin{array}{c}{\mathrm{\&omega;}}_1\\ {\mathrm{\&omega;}}_2\end{array}\right.;$

Described structure signal subspace and noise subspace also separate the signal component in original Hyperspectral imaging by oblique subspace projection and noise component specifically includes:

The characteristic vector corresponding with the eigenvalue belonging to signal or noise constitutes the base S of corresponding signal subspace or noise subspace_basisOr N_basis, the orthogonal intersection cast shadow matrix of signal subspace is

S_orthproject=E-S_basis(S_basis ^HS_basis)^-1S_basis ^H

Wherein, E representation unit matrix, in like manner obtain the orthogonal intersection cast shadow matrix N of noise subspace_orthproject;

Parallel noise subspace to the oblique projection matrix of signal subspace is:

P_SN=S_basis(S_basis ^HN_orthprojectS_basis)^-1S_basis ^HN_orthproject

Signal is:

S=P_SNX;

Parallel signal subspace to the oblique projection matrix of noise subspace is

P_NS=N_basis(N_basis ^HS_orthprojectN_basis)^-1N_basis ^HS_orthproject

Noise is

N=P_NSX, X represent target in hyperspectral remotely sensed image data.

2. the separation method of signal and noise in target in hyperspectral remotely sensed image as claimed in claim 1, it is characterised in that set up data model for target in hyperspectral remotely sensed image:

$X_k(i,j)=S_k(i,j)\&CirclePlus;N_k(i,j)$

Wherein, X_kRepresent that wave band number is target in hyperspectral remotely sensed image data during k;S_k(i, j) represents signal; N_k(i, j) represents noise; K=1 ..., N_B, i=1 ..., N_L, j=1 ..., N_S.; SymbolRepresent direct sum; N_BRepresent wave band number, N_LAnd N_SRepresent ranks number respectively.

3. the separation method of signal and noise in target in hyperspectral remotely sensed image as claimed in claim 1, it is characterised in that pixel number contained by selected piecemeal must not less than 30.