CN105184753A - One-dimensional signal processing guided remote sensing image strip noise rapid filtering method - Google Patents

Abstract

The invention discloses a one-dimensional signal processing guided remote sensing image strip noise rapid filtering method. Starting from the directional characteristic of strip noise, the local statistical characteristics of a noise image are acquired firstly; then the normal statistical characteristics of pixels after removing of strips are estimated via a one-dimensional signal processing method by utilizing the relatively significant local statistical characteristic difference of the strips and non-strips in one-dimensional space; and finally the estimated characteristic values act as guidance information, correction coefficient is calculated through combination of the related technologies of moment matching and segmentation processing, and strip correction of the whole image can be completed. The corresponding characteristics of the noise-free image are estimated by utilizing the local statistical characteristics of the noise image, and the local characteristics of the image and local connection of the adjacent images are fully considered so that the processing results are enabled to be more stable and reliable. Besides, the method is high in calculation efficiency and can be suitable for different types and proportions of strip noise remote sensing images so that the method has relatively high practicality.

Description

The quick filtering method of remote sensing image Banded improvement under one-dimensional signal process guides

Technical field

The invention belongs to optical remote sensing image processing technology field, be specifically related to the quick filtering method of remote sensing image Banded improvement under one-dimensional signal process guiding.

Background technology

In the acquisition process of remote sensing image, the impact of camera subject performance and external environment, multidetector imager when imaging easily because there is the degradation problems such as Banded improvement in the response that its detector is inconsistent.Wherein, the appearance of Banded improvement can cover the true radiation signal distribution of atural object, directly reduces the quality of image, and affects follow-up decipher and the information extraction precision of data.Therefore, recovering in remote sensing image by the information that Banded improvement pollutes, is an important problem.

At present, following several class methods, comprising: the method based on transform domain, the method based on optimal model and the method based on spatial domain for the Banded improvement problem main development in remote sensing image.Based on the method for transform domain by image being transformed to a certain frequency space with the difference of band composition in outstanding image and non-band composition, then utilize a certain specific filter to reject the band composition in frequency field, finally result is gained image space by frequency space inversion.Although this class methods counting yield is high, be usually difficult to when band composition is rejected the concentrated frequency determining noise, thus easily occur non-banded zone information dropout or the problem such as destriping is not thorough.Method based on optimal model can take into full account each class feature of Banded improvement when modeling, and often result is more excellent for the denoising image that it solves, but is limited to the computation complexity of model, and these class methods are unsuitable for the remote sensing image bar tape handling of big data quantity.Method based on spatial domain is then how relevant with alignment technique to coupling, there is higher treatment effeciency, but the result of these class methods is subject to the restriction of reference information Selection Strategy, if not from the regarding feature of Banded improvement when formulating corresponding strategies, final result is probably unsatisfactory.

Although existing destriping method differs from one another, they are often difficult to take into account the requirement in processing accuracy, processing speed and method universality simultaneously.And along with the increase of various optical remote sensing image acquiring way, in the actual process of image, may faced by Banded improvement presentation also become increasingly complex, also will be more and more higher for the requirement in method stability and efficiency.In the face of the application demand of these reality, how to design a kind of to the blanket new method of Banded improvement all kinds of in remote sensing image, realize quick, the precise calibration of zones of different, different noise level band, there is very important Research Significance.

Summary of the invention

The object of the invention is to, limit to for existing technology, the quick filtering method of remote sensing image Banded improvement under providing a kind of one-dimensional signal process to guide, on the basis taking relation between image local feature and contiguous image local into account, the statistical nature of noise image is utilized to estimate without making an uproar the individual features of image, and with the filtering of guiding image strip, to obtain band result more fast and accurately.

Technical scheme of the present invention is: the quick filtering method of remote sensing image Banded improvement under one-dimensional signal process guides, is characterized in that, comprise the steps:

Step 1, based on image local, obtains the statistical nature curve of noise image;

Step 2, utilizes the statistical nature curve of noise image to estimate without making an uproar the statistical nature curve of image;

Step 3, according to noise and the character pair value estimated on characteristic curve, the method in conjunction with match by moment calculates the linear correction factor of each pixel;

Step 4, merges linear correction factor, optimizes;

Step 5, uses the linear correction factor after optimizing to carry out Pointwise filtering to band image;

Step 6, terminates.

Further, described step 1 comprises following sub-step:

Step 1.1, arranges two matrix M identical with raw video size and S, for recording the local feature value of raw video;

Step 1.2, with pixel (x _i, y _j) centered by, be radius with r, calculate the gray-scale value average μ of 2r+1 pixel on band bearing of trend _i,jand standard deviation sigma _i,j, and respectively by it stored in (i, j) element of Mean Matrix M and standard deviation matrix S.

Further, described step 2 comprises following sub-step:

Step 2.1, arranges two matrix M identical with raw video size ' and S ', for recording the local feature value estimated;

Step 2.2, along the bearing of trend perpendicular to band, respectively to M and the smoothing filtering of s-matrix, and charges to filter result in M ' and S '.

Further, described step 3 comprises following sub-step:

Step 3.1, arranges two matrix A identical with raw video size and B, for recording the linear correction factor of each pixel;

Step 3.2, with pixel (x _i, y _j) be target, extract the eigenwert being positioned at (i, j) place in M, S, M ' and S ' tetra-eigenmatrixes accordingly, and utilize the linear correction factor pair of match by moment formulae discovery object pixel, respectively by gain and deviation ratio stored in (i, j) element of A, B.

Further, in described step 3.2, adopt the linear correction factor of the pending image of match by moment technology node-by-node algorithm, comprise gain coefficient a _i,j, such as formula (1) and deviation ratio b _i,jsuch as formula (2);

$a_{i,j}=\frac{{\mathrm{\σ}}_{i,j}^{\′}}{{\mathrm{\σ}}_{i,j}}---\left(1\right)$

B _i,j=μ ' _i,j-a _i,jμ _i,j(2) wherein, μ _i,j, σ _i,jfor pixel (x _i, y _j) local original mean value and standard deviation, μ ' _i,jwith σ ' _i,jbe then filtered local mean value and standard deviation.

Further, described step 4 comprises following sub-step:

Step 4.1, arranges two matrix A identical with raw video size ' and B ', for record each pixel optimize after linear correction factor;

Step 4.2, with (i, the j) element in A for processing enter, is radius with r, calculates the mean value along the coefficient of 2r+1 in strip direction, and count in (i, j) element of A ' by this value;

Step 4.3, by the method described in step 4.2, merges the coefficient in B, acquisition optimized coefficients matrix B '.

Further, in described step 5, according to formula (3), Pointwise filtering is carried out to band image;

$D_{i,j}={\stackrel{\‾}{a}}_{i,j}{R}_{i,j}+{\stackrel{\‾}{b}}_{i,j}---\left(3\right)$

Wherein, R _i,jand D _i,jbe respectively original pending image and filter result image at pixel (x _i, y _j) the respective pixel value at place, with being then two optimization linear correction factor of this point, is (i, the j) element in A ', B '.

Preferably, the value of described r is 21.

The invention has the beneficial effects as follows: the quick filtering method of remote sensing image Banded improvement under one-dimensional signal process guides, utilize the statistical nature difference that in image, band and non-band pixel are comparatively outstanding in the one-dimensional space, estimated the normal statistics feature of pixel after band by the method for one-dimensional signal process, the correction of Banded improvement is quickly and reliably carried out under feature guides; Simultaneously, the local characteristics that the present invention fully takes image into account and the contact be close between image local, to the pixel of the different local of image, different band degree, according to the relation between itself and local guide features, carry out the filtering process of varying strength, efficiently solve the problem of the not in place and excessive correction of correction that complicated Banded improvement occurs in trimming process; The method that the present invention proposes can be effectively applied to the quick filtering of Banded improvement in optical remote sensing image, makes result while stress release treatment, at utmost retain the information characteristics of image.

Accompanying drawing explanation

Fig. 1 is overview flow chart of the present invention.

Embodiment

Understand for the ease of those of ordinary skill in the art and implement the present invention, below in conjunction with accompanying drawing and exemplifying embodiment, this technology is described in further detail, should be appreciated that exemplifying embodiment described herein is only for instruction and explanation of the present invention, is not intended to limit the present invention.

As shown in Figure 1, key step of the invention process can be divided into following two stages:

First stage: calculate correction coefficient;

Step 1, based on image local, obtains the statistical nature curve of noise image.

Step 1.1, arranges two matrix M identical with raw video size and S, for recording the local feature value of raw video.

Step 1.2, along on the bearing of trend of band, with pixel (x _i, y _j) centered by, r is radius, calculates the pixel fragment average μ comprising 2r+1 pixel _i,jand standard deviation sigma _i,j, and be designated as two local feature values of this point, respectively stored in (i, j) element of Mean Matrix M and standard deviation matrix S.It should be noted that the direction from extending perpendicular to band, eigenvalue matrix M and S actually contains many one-dimensional characteristic curves, and position and the strength information of band pixel have been reacted in the shake of these curves just.In the process calculating eigenwert, for making the statistical law in the eigenwert energy reasonable reaction region of calculating, the choosing value of r is unsuitable too small; Simultaneously for ensureing the locality of eigenwert, the choosing value of r is also unsuitable excessive.Be 21 in this reference value providing r, concrete setting can optionally become.

Step 2, utilizes the statistical nature curve of noise image to estimate without making an uproar the statistical nature curve of image.

Step 2.1, arranges two matrix M identical with raw video size ' and S ', for recording the local feature value estimated.

Step 2.2, along perpendicular on the bearing of trend of band, respectively to M and the smoothing filtering of s-matrix, and charges to filter result in M ' and S '.The one-dimensional characteristic curve that this process is equivalent to reacting band pixel characteristic in M and S carries out filtering process, and estimates out normal eigenvalues when image local does not comprise Banded improvement according to this.In conventional one-dimensional signal filtering method, SG filtering often has preferably filter effect and stability, therefore can be used as the default filter of this operation.Certainly, person skilled also can select other suitable one-dimensional signal filtering methods according to the actual conditions of pending image.

Step 3, according to noise and the character pair value estimated on characteristic curve, the method in conjunction with match by moment calculates the linear correction factor of each pixel.

Step 3.1, arranges two matrix A identical with raw video size and B, for recording the linear correction factor of each pixel.

Step 3.2, with pixel (x _i, y _j) be target, extract the eigenwert being positioned at (i, j) place in M, S, M ' and S ' tetra-eigenmatrixes accordingly, and utilize the linear correction factor pair of match by moment formulae discovery object pixel, respectively by gain and deviation ratio stored in (i, j) element of A, B.

Concrete, in conjunction with the primitive character value in eigenmatrix M, S and filtering (estimating) eigenwert in M ', S ', adopt the linear correction factor of the pending image of match by moment technology node-by-node algorithm, comprise gain coefficient a _i,jsuch as formula (1) and deviation ratio b _i,jsuch as formula (2), and respectively stored in (i, j) element of A, B matrix.

$a_{i,j}=\frac{{\mathrm{\σ}}_{i,j}^{\′}}{{\mathrm{\σ}}_{i,j}}---\left(1\right)$

B _i,j=μ ' _i,j-a _i,jμ _i,j(2) wherein, μ _i,j, σ _i,jfor pixel (x _i, y _j) local original mean value and standard deviation, μ ' _i,jwith σ ' _i,jbe then filtered local mean value and standard deviation.

Step 4, linear correction factor merges, optimization

Step 4.1, arranges two matrix A identical with raw video size ' and B ', for record each pixel optimize after linear correction factor.

Step 4.2, with (i, the j) element in A for processing enter, is radius with r, calculates the mean value along the coefficient of 2r+1 in strip direction, and count in (i, j) element of A ' by this value;

Step 4.3, by the method described in step 4.2, merges the coefficient in B, acquisition optimized coefficients matrix B '.

Concrete, when calculating image local feature value, certain any pixel value can be counted in a contiguous 2r+1 pixel fragment, and thus arbitrary pixel is actual has 2r+1 linear correction factor.For determining final correction coefficient, the simplest method is respectively to the coefficient in A, B, along on the direction of band, with (i, j) element for processing enter, r is radius, 2r+1 coefficient is averaged, obtains new optimized coefficients matrix A ' and B '.In addition, person skilled also can adopt other common technologies in piecemeal process, as average weighted method, and arranges weighting coefficient according to the distance relation at pixel to be asked and related pixel section center.

Subordinate phase: image Pointwise filtering

Step 5, uses the linear correction factor after optimizing to carry out Pointwise filtering to band image;

Concrete, the optimization linear correction factor that the A ' utilizing the first stage to obtain, B ' are interior, according to formula (3), carries out Pointwise filtering to band image.

$D_{i,j}={\stackrel{\‾}{a}}_{i,j}{R}_{i,j}+{\stackrel{\‾}{b}}_{i,j}---\left(3\right)$

Wherein, R _i,jand D _i,jbe respectively original pending image and filter result image at pixel (x _i, y _j) the respective pixel value at place, with be then two optimization linear correction factor of this point, (i, the j) element namely in A ', B '.

Step 6, terminates.

Should be understood that, the above-mentioned description for preferred embodiment is comparatively detailed, therefore can not think the restriction to scope of patent protection of the present invention.Those skilled in the art are under enlightenment of the present invention; not departing from the scope that the claims in the present invention protect; can also make distortion or replace, all fall within the scope of protection of the present invention, request protection domain of the present invention should be as the criterion with claims.

Claims (8)

1. the quick filtering method of remote sensing image Banded improvement under one-dimensional signal process guiding, is characterized in that, comprise the steps:

Step 1, based on image local, obtains the statistical nature curve of noise image;

Step 2, utilizes the statistical nature curve of noise image to estimate without making an uproar the statistical nature curve of image;

Step 3, according to noise and the character pair value estimated on characteristic curve, the method in conjunction with match by moment calculates the linear correction factor of each pixel;

Step 4, merges linear correction factor, optimizes;

Step 5, uses the linear correction factor after optimizing to carry out Pointwise filtering to band image;

Step 6, terminates.

2. the quick filtering method of remote sensing image Banded improvement under one-dimensional signal process guiding according to claim 1, it is characterized in that, described step 1 comprises following sub-step:

Step 1.1, arranges two matrix M identical with raw video size and S, for recording the local feature value of raw video;

Step 1.2, with pixel (x _i, y _j) centered by, be radius with r, calculate the gray-scale value average μ of 2r+1 pixel on band bearing of trend _i,jand standard deviation sigma _i,j, and respectively by it stored in (i, j) element of Mean Matrix M and standard deviation matrix S.

3. the quick filtering method of remote sensing image Banded improvement under one-dimensional signal process guiding according to claim 1, it is characterized in that, described step 2 comprises following sub-step:

Step 2.1, arranges two matrix M identical with raw video size ' and S ', for recording the local feature value estimated;

Step 2.2, along the bearing of trend perpendicular to band, respectively to M and the smoothing filtering of s-matrix, and charges to filter result in M ' and S '.

4. the quick filtering method of remote sensing image Banded improvement under one-dimensional signal process guiding according to claim 1, it is characterized in that, described step 3 comprises following sub-step:

Step 3.1, arranges two matrix A identical with raw video size and B, for recording the linear correction factor of each pixel;

Step 3.2, with pixel (x _i, y _j) be target, extract the eigenwert being positioned at (i, j) place in M, S, M ' and S ' tetra-eigenmatrixes accordingly, and utilize the linear correction factor pair of match by moment formulae discovery object pixel, respectively by gain and deviation ratio stored in (i, j) element of A, B.

5. the quick filtering method of remote sensing image Banded improvement under one-dimensional signal process guiding according to claim 4, is characterized in that: in described step 3.2, adopts the linear correction factor of the pending image of match by moment technology node-by-node algorithm, comprise gain coefficient a _i,j, such as formula (1) and deviation ratio b _i,j, such as formula (2);

$a_{i,j}=\frac{{\mathrm{\σ}}_{i,j}^{\′}}{{\mathrm{\σ}}_{i,j}}---\left(1\right)$

b _i,j＝μ′ _i,j-a _i,jμ _i,j(2)

Wherein, μ _i,j, σ _i,jfor pixel (x _i, y _j) local original mean value and standard deviation, μ ' _i,jwith σ ' _i,jbe then filtered local mean value and standard deviation.

6. the quick filtering method of remote sensing image Banded improvement under one-dimensional signal process guiding according to claim 1, it is characterized in that, described step 4 comprises following sub-step:

Step 4.1, arranges two matrix A identical with raw video size ' and B ', for record each pixel optimize after linear correction factor;

Step 4.2, with (i, the j) element in A for processing enter, is radius with r, calculates the mean value along the coefficient of 2r+1 in strip direction, and count in (i, j) element of A ' by this value;

Step 4.3, by the method described in step 4.2, merges the coefficient in B, acquisition optimized coefficients matrix B '.

7. the quick filtering method of remote sensing image Banded improvement under one-dimensional signal process guiding according to claim 1, is characterized in that: in described step 5, according to formula (3), carries out Pointwise filtering to band image;

$D_{i,j}={\stackrel{\‾}{a}}_{i,j}{R}_{i,j}+{\stackrel{\‾}{b}}_{i,j}---\left(3\right)$

Wherein, R _i,jand D _i,jbe respectively original pending image and filter result image at pixel (x _i, y _j) the respective pixel value at place, with being then two optimization linear correction factor of this point, is (i, the j) element in A ', B '.

8. the quick filtering method of remote sensing image Banded improvement under the one-dimensional signal process according to claim 2 or 6 guides, is characterized in that: the value of described r is 21.