CN104657998A - DG-WMLE (degenerated filter-weighted maximum likelihood estimation) based radar image edge extraction method - Google Patents

Abstract

A DG-WMLE (degenerated filter-weighted maximum likelihood estimation) based radar image edge extraction method comprises steps as follows: a pixel point is selected randomly, two adjacent pixel of the pixel points are xi and xj, and noise-free true values of xi and xj are accurately estimated; dissimilarity of neighborhood pixels xi and xj in a current filter direction is calculated and taken as edge response of a pixel x in the direction; whether all filter directions are traversed in the position of the current pixel x is judged, if not, the steps are repeated, and the edge strength in a next filter direction is calculated; edge response values in all directions are traversed, the maximum value in the edge response values is taken as the current pixel edge strength, and the corresponding filter direction is the edge direction; whether all pixel points in an image are traversed is judged, if not, the steps are repeated, the edge strength of a next pixel point is continuously calculated, and if yes, the edge strength values of all pixels are output. With the adoption of the method, SAR (synthetic aperture radar) image edge extraction having advantages of image resolution guarantee and accurate edge positioning can be realized.

Description

Based on the SAR Image Edge Extraction of degeneration wave filter-weighting maximal possibility estimation

Technical field

The application relates to radar signal processing field, relate more specifically to a kind of SAR Image Edge Extraction based on degeneration wave filter-weighting maximal possibility estimation (degenerate filter-weighted maximum likelihoodestimation, DG-WMLE).

Background technology

The extraction at edge and analysis are the bases that synthetic-aperture radar (Synthetic Aperture Radar, SAR) image spatial information (si) is analyzed.

Wave filter based on regional compare is the classical way of SAR image rim detection, the calculating of outline map is each pixel in order one by one in access images, then by a series of, different towards wave filter be placed in this pixel place, analyze the otherness between pixel two side areas, for weighing the probability occurring edge at this pixel, and as the instruction of this pixel marginal information.Its schematic diagram is as follows:

This type of wave filter controls primarily of 4 parameters: length l _f, width w _f, the gap d between central pixel point two side areas _f, and the angle amplification Δ θ between two continuous directions _f.Therefore, at any one center pixel place, at l _f, w _fand d _fwhen three parameters are equal, total N _f=π/Δ θ _fthe wave filter of individual different angles.L _fand w _fthese two parameters define the yardstick of wave filter.Under this mechanism, in SAR image any one pixel edge strength by this pixel different towards the peak response of wave filter represent, the response of wave filter is then by filter center pixel two side areas R _iand R _jbetween the probability diversity based on region weigh determine.The method that diversity is weighed has a variety of, relatively more conventional has average ratio operator (Ratio), Generalized Likelihood Ratio operator (Generalized Likelihood Ratio Test, GLRT) and crosscorrelation operator (CrossCorrelation) etc.

This edge extracting method theory is complete, be easy to realize and efficient, but its prerequisite is sample in two regions meets independent identically distributed hypothesis.In the SAR image scene of reality, particularly in heterogeneous region, be difficult to ensure that independent identically distributed hypothesis is always set up.Therefore the restriction of this type of edge extracting method performance is mainly reflected in following two aspects:

Yardstick predicament: if the yardstick of wave filter increases, so in homogeney region, this wave filter can be estimated more accurately to distribution parameter, and the outline map in this case extracted in homogeney region will more " totally ", and false-alarm is less.But because yardstick is larger, the heterogeneous region on earth's surface, make independent same distribution suppose more to be difficult to meet, be difficult to extract marginal information accurately in heterogeneous region, false-alarm is more;

Deviations: in the calculating of reality, the direction of wave filter is always set as limited.When being positioned at the wave filter of center pixel towards when all misfitting with this pixel actual edge direction, edge extracting result there will be deviations, and filter scales is larger, and this deviations is more obvious.

Summary of the invention

In view of this, the present invention is directed to the defect that traditional SAR image edge extracting method based on regional compare exists in scale selection, edge local two, propose a kind of SAR Image Edge Extraction based on degeneration wave filter-weighting maximal possibility estimation, to realize ensureing image resolution ratio, edge local SAR image edge extracting accurately.

To achieve these goals, the invention provides a kind of SAR Image Edge Extraction based on degeneration wave filter-weighting maximal possibility estimation, comprise the following steps:

Estimate described radar image any point place without to make an uproar true value, the skirt response at this pixel place under calculating filter direction;

Traveled through all filter direction, get wherein maximal margin response as current pixel place edge strength;

Traveled through on described radar image institute a little, obtain the edge intensity value computing at all pixel places of described radar image, realize the edge extracting of described radar image thus.

And a kind of SAR Image Edge Extraction based on degeneration wave filter-weighting maximal possibility estimation, comprises the following steps:

Step 1, input radar image data;

Step 2, supposes that current pixel is x, selects a direction θ _dfbuild wave filter; Downward the party, centered by pixel x, two pixels making it adjacent are respectively x _iand x _j;

Step 3, builds search window respectively with

Step 4, in respective search window, estimates x respectively _iand x _jplace without to make an uproar true value;

Step 5, calculates neighborhood territory pixel x under present filter direction _iand x _jdiversity, as the skirt response at the downward pixel x place of the party;

Step 6, judges whether traveled through all filter direction at current pixel x place; If not, then return step 2, and calculate the edge strength in next filter direction; If so, then step 7 is continued;

Step 7, travels through the skirt response value under all directions, gets the edge strength of maximal value wherein as current pixel place, and corresponding filter direction is edge direction;

Step 8, judges whether to have traveled through pixels all in image; If not, then return step 2, and continue the edge strength calculating next pixel place; If so, then step 9 is continued;

Step 9, exports the edge intensity value computing at all pixel places and preserves.

Wherein, search window described in step 3 with size be 11 pixel * 5 pixels.

Wherein, x is estimated described in step 4 _iand x _jthe step without true value of making an uproar at place comprises:

${\widehat{\mathrm{\μ}}}^{\mathrm{WMIE}}\left(x_i\right)=\frac{{\mathrm{\Σ}}_{x^{\′}\∈R_i^{\mathrm{SW}}}\mathrm{\ω}(x_i,x^{\′})I\left(x^{\′}\right)}{{\mathrm{\Σ}}_{x^{\′}\∈R_i^{\mathrm{SW}}}\mathrm{\ω}(x_i,x^{\′})}---\left(1\right)$

In formula, the image intensity that I (x ') is x ' pixel place, for pixel x _ilocate the WMLE estimated value without true value of making an uproar, for the search window at pixel place estimated for WMLE, the region R before the size of described search window and scope and degeneration _iunanimously.

Wherein, described weights omega (x _i, x ') come from x _iand the image block P centered by x ' two pixels _xiand P _{x '}between probability diversity and through the conversion of index core, computing formula is as follows:

$\mathrm{\ω}(x_i,x^{\′})=\exp \{-\frac{D_{\mathrm{PPB}}(P_{x_i},P_{x^{\′}})}{h}\}---\left(2\right)$

Wherein, h is the parameter of index core, be the probability diversity measurement between two image blocks, calculate according to following formula:

$D_{\mathrm{PPB}}(P_{x_i},P_{x^{\′}})=\underset{\mathrm{\τ}\∈P}{\mathrm{\Σ}}\left\{2\ln \right[\frac{1}{2}(I(x_i+\mathrm{\τ})+I(x^{\′}+\mathrm{\τ}))]-\ln I(x_i+\mathrm{\τ})-\ln I(x^{\′}+\mathrm{\τ})\}---\left(3\right)$

Wherein, I (x _i+ τ) and I (x '+τ) represent that the intensity at corresponding pixel points place in two image blocks, τ are used for representing pixel x respectively _iwith x ' is at image block P _xiand P _{x '}in position consistency.

Wherein, neighborhood territory pixel x under present filter direction is calculated described in step 5 _iand x _jthe step of diversity comprise:

At center pixel x place, calculate the marginal information on certain direction with following formula:

$D_{\mathrm{DG}-\mathrm{WMLE}}(x_i,x_j)=2\ln \left(\frac{1}{2}({\widehat{\mathrm{\μ}}}^{\mathrm{WMLE}}\left(x_i\right)+{\widehat{\mathrm{\μ}}}^{\mathrm{WMLE}}\left(x_j\right))\right)-\ln {\widehat{\mathrm{\μ}}}^{\mathrm{WMLE}}\left(x_i\right)-\ln {\widehat{\mathrm{\μ}}}^{\mathrm{WMLE}}\left(x_j\right)---\left(4\right);$

Wherein, for pixel xi place is without the WMLE estimated value of true value of making an uproar, for pixel x _jplace is without the WMLE estimated value of true value of making an uproar;

Then, the edge strength of the filter response maximal value under all directions as this some place is got.

Wherein, the filter direction used in step 2 adds up to 4.

Known based on technique scheme, method of the present invention, compared with traditional SAR image edge extracting method, has the following advantages:

(1) there is not the difficulty on scale selection.Search window is larger, and pixel place is more accurate without the estimation of true value of making an uproar, but also can bring the efficiency reducing and calculate simultaneously.Therefore consider actual applicable cases, only need the search window (being generally 11 pixel * 5 pixels) selecting a certain size, ensure the quantity for the sample point estimated, just can obtain good edge extracting result.

(2) accurate positioning at edge.In this method, because the estimation without true value of making an uproar only is carried out at the neighborhood territory pixel place of center pixel, therefore the direction sum of wave filter is fixing.In addition, adopt non-local weighting maximum likelihood estimate can obtain each pixel place comparatively accurately without actual value of making an uproar, make the estimation of edge strength and edge local more accurate.

Accompanying drawing explanation

Fig. 1 is the schematic diagram of the SAR image edge extraction filter based on regional compare;

Fig. 2 is the process flow diagram of the SAR image edge extracting method based on regional compare;

Fig. 3 is the schematic diagram of SAR image edge extracting degeneration wave filter;

Fig. 4 is the schematic diagram utilizing the sample in search window and weighting maximal possibility estimation to carry out estimation of distribution parameters;

Fig. 5 is the process flow diagram of degeneration wave filter-weighting maximum likelihood estimate;

Fig. 6 is the edge extracting result of TerraSAR image and DG-WMLE method;

Fig. 7 is the edge extracting result of traditional GLRT method.

Embodiment

For making the object, technical solutions and advantages of the present invention clearly understand, below in conjunction with specific embodiment, and with reference to accompanying drawing, the present invention is described in further detail.

Degeneration wave filter-weighting maximum likelihood estimate is that one can ensure image resolution ratio, edge local SAR image edge extracting method accurately, and its core is a degeneration wave filter (degeneratefilter, DG).So-called degeneration wave filter refer to by traditional SAR edge extraction filter based on regional compare in spatial domain enterprising row degradation, by original two region R _iand R _jsharply be degenerated to two some x _iand x _j, along with filter direction θ _dfvertical line lay respectively at the both sides of center pixel x.The design diagram of this degeneration wave filter as shown in Figure 3.

Under the design of degeneration wave filter, center pixel x place exists the information at the probability at edge and edge can with two, its both sides pixel x _iand x _jbetween probability similarity weigh, the wave filter based on regional compare of classics is achieved maximum degeneration on spatial domain; In this case, independent identically distributed hypothesis certainly will meet.But the problem brought is, the calculating of marginal probability and information the probability similarity needed between the pixel of two, both sides is provided, therefore need the distribution parameter knowing these two points, this is the key weighing probability similarity between them.But under degeneration design of filter, the parameter of each probability distribution only has a sample to can be used for estimating, parameter estimation in this case certainly will exist great deviation, the performance of extreme influence edge extracting.For this problem, the present invention proposes a kind of edge extracting new method of SAR image: degeneration wave filter-weighting maximum likelihood estimate (degenerate filter-weighted maximum likelihood estimation, DG-WMLE), below degeneration wave filter-weighting maximum likelihood estimate of the present invention is set forth particularly.

When without loss of generality, below with x _ithis pixel is example, introduces the estimation procedure of its probability distribution parameters.According to the weighting maximal possibility estimation criterion that Deledalle etc. proposes, pixel x _ithe method of estimation of distribution parameter as follows:

${\widehat{\mathrm{\μ}}}^{\mathrm{WMIE}}\left(x_i\right)=\frac{{\mathrm{\Σ}}_{x^{\′}\∈R_i^{\mathrm{SW}}}\mathrm{\ω}(x_i,x^{\′})I\left(x^{\′}\right)}{{\mathrm{\Σ}}_{x^{\′}\∈R_i^{\mathrm{SW}}}\mathrm{\ω}(x_i,x^{\′})}---\left(1\right)$

The image intensity that in formula, I (x ') is x ' pixel place, for pixel x _ilocate the WMLE estimated value without true value of making an uproar, for the search window (searchwindow, SW) at pixel place estimated for WMLE, the region R before the size of this search window and scope and degeneration _iunanimously.And the weights omega (x in the middle of above formula _i, x ') come from x _iand the image block P centered by x ' two pixels _xiand P _{x '}between probability diversity and through the conversion of index core, as follows:

$\mathrm{\ω}(x_i,x^{\′})=\exp \{-\frac{D_{\mathrm{PPB}}(P_{x_i},P_{x^{\′}})}{h}\}---\left(2\right)$

Wherein, h is the parameter of index core, be that probability diversity between two image blocks weighs (probabilistic patch based, PPB), can by following formulae discovery according to the work of Deledalle:

$D_{\mathrm{PPB}}(P_{x_i},P_{x^{\′}})=\underset{\mathrm{\τ}\∈P}{\mathrm{\Σ}}\left\{2\ln \right[\frac{1}{2}(I(x_i+\mathrm{\τ})+I(x^{\′}+\mathrm{\τ}))]-\ln I(x_i+\mathrm{\τ})-\ln I(x^{\′}+\mathrm{\τ})\}---\left(3\right)$

Wherein I (x _i+ τ) and I (x '+τ) represent that the intensity at corresponding pixel points place in two image blocks, τ are used for representing pixel x respectively _iwith x ' is at image block P _xiand P _{x '}in position be consistent.

Mention above, the probability diversity derived based on Generalized Likelihood Ratio weighs the impact being subject to noise to a certain extent, more accurate in order to make weight estimate, the method that SAR image utilizes sliding window average by we carries out a simplest pre-filtering (pre-filtering, PF), then the probability diversity of image block is calculated and calculate in the SAR image through pre-filtering, but remain when utilizing weighting maximum likelihood method to carry out estimation of distribution parameters carry out in original SAR image data average weighted, thus keep the resolution of original image as much as possible.

With non local wave filter unlike, when edge extracting, we need done the marginate hypothesis of tool at central pixel point x place.On this hypothesis basis, we utilize pixel x _iand x _jdiversity between corresponding probability distribution is weighed as in the marginal probability of center pixel x and the instruction of information.Therefore, for pixel x _ithe estimation of distribution parameter is confined to search window in, can't pass through hypothesis by center pixel x, be oriented θ _dfedge arrive the search window of opposite side in, as shown in Figure 4.

In sum, at center pixel x place, the marginal information on certain direction can calculate with following formula:

$D_{\mathrm{DG}-\mathrm{WMLE}}(x_i,x_j)=2\ln \left(\frac{1}{2}({\widehat{\mathrm{\μ}}}^{\mathrm{WMLE}}\left(x_i\right)+{\widehat{\mathrm{\μ}}}^{\mathrm{WMLE}}\left(x_j\right))\right)-\ln {\widehat{\mathrm{\μ}}}^{\mathrm{WMLE}}\left(x_i\right)-\ln {\widehat{\mathrm{\μ}}}^{\mathrm{WMLE}}\left(x_j\right)---\left(4\right)$

Finally, the edge strength of the filter response maximal value under all directions as this some place is got.

The flow process of whole edge extracting method as shown in Figure 5.Below in conjunction with Fig. 5, the concrete steps of method of the present invention are further elaborated:

Step 1, input SAR image data.

Step 2, supposes that current pixel is x, selects a direction θ _dfbuild wave filter, such as 0 degree of direction.Downward the party, centered by pixel x, two pixels making it adjacent are respectively x _iand x _j.

Step 3, in order to accurately estimate x _iand x _jplace without to make an uproar true value, build search window respectively with window size is 11 pixel * 5 pixels.

Step 4, in respective search window, estimates x respectively according to formula (1), (2), (3) _iand x _jplace without to make an uproar true value.

Step 5, according to formula (4), calculates neighborhood territory pixel x under present filter direction _iand x _jdiversity, as the skirt response at the downward pixel x place of the party.

Step 6, judges whether traveled through all filter direction at current pixel x place.If not, then return step 2, calculate the edge strength in next filter direction; If so, then step 7 is continued.

Step 7, travels through the skirt response value under all directions, gets the edge strength of maximal value wherein as current pixel place, and corresponding filter direction is edge direction.

Step 8, judges whether to have traveled through pixels all in image.If not, then return step 2, continue the edge strength calculating next pixel place; If so, then step 9 is continued.

Step 9, exports the edge intensity value computing at all pixel places and preserves.

According to above step, edge extracting is carried out to a width TerraSAR satellite image, this image size is 608 pixel * 608 pixels, resolution is 1 meter, shooting area is certain port area of san francisco, usa, image and edge extracting result are as shown in Figure 6, wherein left figure is TerraSAR image, right figure is that left figure adopts DG-WMLE method edge extracting result, Fig. 7 is the result utilizing classical GLRT method this image to be carried out to edge extracting, in the method, the window size of edge extraction filter is also 11 pixel * 5 pixels, and the filter direction of use adds up to 4.The extraction result of contrast two kinds of methods, can find out under same filter direction number and filtering window size, the edge that method provided by the invention is extracted is more meticulous, the false-alarm extracted in result is less, for the edge contour of point target, line target and Ship Target in image, it is more complete careful to sketch the contours, and the location of marginal point is also more accurate simultaneously.

Above-described specific embodiment; object of the present invention, technical scheme and beneficial effect are further described; be understood that; the foregoing is only specific embodiments of the invention; be not limited to the present invention; within the spirit and principles in the present invention all, any amendment made, equivalent replacement, improvement etc., all should be included within protection scope of the present invention.

Claims (7)

1., based on a SAR Image Edge Extraction for degeneration wave filter-weighting maximal possibility estimation, comprise the following steps:

Estimate described radar image any point place without to make an uproar true value, the skirt response at this pixel place under calculating filter direction;

Traveled through all filter direction, get wherein maximal margin response as current pixel place edge strength;

Traveled through on described radar image institute a little, obtain the edge intensity value computing at all pixel places of described radar image, realize the edge extracting of described radar image thus.

2., based on a SAR Image Edge Extraction for degeneration wave filter-weighting maximal possibility estimation, comprise the following steps:

Step 1, input radar image data;

Step 2, supposes that current pixel is x, selects a direction θ _dfbuild wave filter; Downward the party, centered by pixel x, two pixels making it adjacent are respectively x _iand x _j;

Step 3, builds search window respectively with

Step 4, in respective search window, estimates x respectively _iand x _jplace without to make an uproar true value;

Step 5, calculates neighborhood territory pixel x under present filter direction _iand x _jdiversity, as the skirt response at the downward pixel x place of the party;

Step 6, judges whether traveled through all filter direction at current pixel x place; If not, then return step 2, and calculate the edge strength in next filter direction; If so, then step 7 is continued;

Step 7, travels through the skirt response value under all directions, gets the edge strength of maximal value wherein as current pixel place, and corresponding filter direction is edge direction;

Step 8, judges whether to have traveled through pixels all in image; If not, then return step 2, and continue the edge strength calculating next pixel place; If so, then step 9 is continued;

Step 9, exports the edge intensity value computing at all pixel places and preserves.

3. the SAR Image Edge Extraction based on degeneration wave filter-weighting maximal possibility estimation according to claim 1, wherein search window described in step 3 with size be 11 pixel * 5 pixels.

4. the SAR Image Edge Extraction based on degeneration wave filter-weighting maximal possibility estimation according to claim 1, wherein estimates x described in step 4 _iand x _jthe step without true value of making an uproar at place comprises:

${\widehat{\mathrm{\μ}}}^{\mathrm{WMLE}}\left(x_i\right)=\frac{{\mathrm{\Σ}}_{x^{\′}\∈R_i^{\mathrm{SW}}}\mathrm{\ω}(x_i,x^{\′})I\left(x^{\′}\right)}{{\mathrm{\Σ}}_{x^{\′}\∈R_i^{\mathrm{SW}}}\mathrm{\ω}(x_i,x^{\′})}---\left(1\right)$

In formula, the image intensity that I (x ') is x ' pixel place, for pixel x _ilocate the WMLE estimated value without true value of making an uproar, for the search window at pixel place estimated for WMLE, the region R before the size of described search window and scope and degeneration _iunanimously.

5. the SAR Image Edge Extraction based on degeneration wave filter-weighting maximal possibility estimation according to claim 3, wherein said weights omega (x _i, x ') come from x _iand the image block P centered by x ' two pixels _xiand P _{x '}between probability diversity and through the conversion of index core, computing formula is as follows:

$\mathrm{\ω}(x_i,x^{\′})=\exp \{-\frac{D_{\mathrm{PPB}}(P_{x_i},P_{x^{\′}})}{h}\}---\left(2\right)$

Wherein, h is the parameter of index core, be the probability diversity measurement between two image blocks, calculate according to following formula:

$D_{\mathrm{PPB}}(P_{x_i},P_{x^{\′}})=\underset{\mathrm{\τ}\∈P}{\mathrm{\Σ}}\left\{21n\right[\frac{1}{2}(I(x_i+\mathrm{\τ})+I(x^{\′}+\mathrm{\τ}))]-1\mathrm{nI}(x_i+\mathrm{\τ})-1\mathrm{nI}(x^{\′}+\mathrm{\τ})\}---\left(3\right)$

Wherein, I (x _i+ τ) and I (x '+τ) represent that the intensity at corresponding pixel points place in two image blocks, τ are used for representing pixel x respectively _iwith x ' is at image block P _xiand P _{x '}in position consistency.

6. the SAR Image Edge Extraction based on degeneration wave filter-weighting maximal possibility estimation according to claim 1, wherein calculates neighborhood territory pixel x under present filter direction described in step 5 _iand x _jthe step of diversity comprise:

At center pixel x place, calculate the marginal information on certain direction with following formula:

$D_{\mathrm{DG}-\mathrm{WMLE}}(x_i,x_j)=21n\left(\frac{1}{2}({\widehat{\mathrm{\μ}}}^{\mathrm{WMLE}}\left(x_i\right)+{\widehat{\mathrm{\μ}}}^{\mathrm{WMLE}}\left(x_j\right))\right)-1n{\widehat{\mathrm{\μ}}}^{\mathrm{WMLE}}\left(x_i\right)-1n{\widehat{\mathrm{\μ}}}^{\mathrm{WMLE}}\left(x_j\right)---\left(4\right);$

Wherein, for pixel x _ilocate the WMLE estimated value without true value of making an uproar, for pixel x _iplace is without the WMLE estimated value of true value of making an uproar;

Then, the edge strength of the filter response maximal value under all directions as this some place is got.

7. the SAR Image Edge Extraction based on degeneration wave filter-weighting maximal possibility estimation according to claim 1, the filter direction wherein used in step 2 adds up to 4.