CN103177443A - SAR (synthetic aperture radar) target attitude angle estimation method based on randomized hough transformations - Google Patents

Abstract

The invention provides an SAR (synthetic aperture radar) target attitude angle estimation method based on randomized hough transformations. The method includes: firstly, analyzing invariant features of SAR images changing along with azimuth on the basis of analyzing a randomized hough transformation method and SAR target features; secondly, segmenting the SAR target through the Otsu's method according to the characteristic that double edges of one side of the SAR target facing the radar are sharp; thirdly, calculating threshold of the two sharp edges through the randomized hough transformation method so as to complete accurate attitude angle estimation. By the aid of the SAR target attitude angle estimation method, the problem of imaging target self-occlusion of the SAR is resolved, and accurate attitude angle estimation is realized. According to experimental results of measured data of MSTAR, the SAR target attitude angle estimation method has the advantages of high accuracy and short calculation time.

Description

SAR object attitude angle method of estimation based on random hough conversion

Technical field

The invention belongs to the signal processing technology field, especially relate to the SAR object attitude angle method of estimation based on random hough conversion.

Background technology

Synthetic-aperture radar (Synthetic Aperture Radar, SAR) imaging is very responsive to object attitude angle, when the attitude of SAR target changed, the scattering center of target also can change, and caused the target under different attitude angle that obvious difference is arranged.At the SAR target classification with in detecting, accurate pose estimation can reduce the object matching number and detect error.

The SAR object attitude angle refers to target main shaft and the angle of radar line of sight between the projection of ground, military terrain object in the MSTAR data, generally comprise the chassis of a rectangle and stretch out parts hinged with target outside the chassis, as antenna and gun turret, the tank model schematic diagram of Fig. 1 for simplifying comprises length, width and attitude angle information.For the MSTAR data, the attitude angle of target is exactly that the main shaft of target and SAR image distance are to the angle of negative sense.

Present pose estimation method mainly contains target envelope box method, target method of principal axis, Radon method and Hough converter technique.Due to the singularity of SAR image imaging, these methods accuracy when characterizing the SAR object attitude angle is not high enough.

Summary of the invention

Technical matters to be solved by this invention is to overcome the deficiencies in the prior art, has proposed the SAR object attitude angle method of estimation based on random hough conversion.Accurate SAR object attitude angle estimates it is one of important step of SAR target identification, characteristics due to SAR image imaging itself, a side of radar due to the blocking of target itself, is difficult to split dorsad, makes traditional pose estimation method all be difficult to accurate estimation.The present invention is based on the SAR target towards bilateral sharp-edged characteristics of a side of radar, utilize between quick maximum kind poor method Otsu to be partitioned into the SAR target, then calculate clear bilateral threshold value with the random Hough transformation method, obtain pose estimation accurately.By the MSATR data verification, the present invention can solve the problem that the SAR imageable target self is blocked, and realizes accurate SAR object attitude angle estimation.

For solving the problems of the technologies described above, the technical solution adopted in the present invention is:

Based on the SAR object attitude angle method of estimation of random hough conversion, at first described method carries out the partial differential denoising to the SAR image, then the SAR image after denoising is carried out Otsu and cuts apart, and obtains cutting apart meticulous SAR target; At last the segmentation result that obtains is adopted pose estimation algorithm based on random Hough transformation, realize that accurate SAR object attitude angle estimates; Its concrete steps are as follows:

Steps A is carried out denoising with partial differential to the SAR image, makes image improve cohesion under the prerequisite of preserving edge;

Step B utilizes the Otsu method to cut apart the SAR target

Choose optimal separation gray-scale value threshold value t', make between the target and background of SAR image separation property best; Obtaining the probability of happening in each zone to be split based on histogram, according to setting threshold, is SAR target and background two classes with image segmentation; Optimal threshold t' satisfies following formula:

${\mathrm{\&sigma;}}_B^2\left(t^{\&prime;}\right)=\underset{1\&le;t<L}{\max }\left\{{\mathrm{\&sigma;}}_B^2\left(t\right)\right\}$

Wherein

μ _TBe the entire image average, ω (t) is the ratio of separation gray-scale value 1 to the t part of image, and μ (t) is the average of separation gray-scale value 1 to the t part of image, and t is for separating gray-scale value, 1≤t＜L; L is gradation of image progression, and L is natural number;

Step C, based on the SAR object attitude angle algorithm for estimating of random Hough transformation: its implementation procedure is:

Step C-1, point set D of all boundary dot generation of the image after cutting apart, setup parameter space F;

Step C-2 chooses a pair of point (d from point set D _i, d _j), setpoint distance condition d if this does not satisfy pre-determined distance condition to point, chooses another to this step of continuation; I, j are natural numbers, are the labels of point set D mid point;

Step C-3 utilizes straight-line equation to (d _i, d _j) find the solution, obtain parameter space parameter (ρ, θ); If there has been a f in parameter space F _{(ρ, θ)}The point that perhaps closes on it adds 1 with this accumulated value, otherwise (ρ, θ) inserted parameter space, forms new accumulated value point f _{(ρ, θ)}Wherein, ρ, θ are respectively utmost point footpath and the polar angle under the polarization coordinate system; Cumulative complete after, in point set D with this to a removing, step C-2 is returned in the initiation parameter space;

Step C-4 is for the f in parameter space _{(ρ, θ)}Set minimum value m, the polar angle θ in parameter space is separated by 90 °, and single threshold value all surpasses two some accumulated values addition of minimum value m, obtain two points of respective value maximum, two limits of these two some correspondences are testing result.

The invention has the beneficial effects as follows: the present invention proposes the SAR object attitude angle method of estimation based on random hough conversion, described method is on the basis of analyzing random Hough transformation method and SAR target characteristic, analyze the SAR image with the invariant features of azimuthal variation, based on the bilateral sharp-edged characteristics of SAR target towards a side of radar, utilize between quick maximum kind poor method Otsu to be partitioned into the SAR target, then calculate clear bilateral threshold value with the random Hough transformation method, obtain pose estimation accurately.The present invention can solve the problem that the SAR imageable target self is blocked, and realizes that accurate SAR object attitude angle estimation shows based on the measured data experimental result of MSTAR, and the present invention has advantages of that at estimation SAR object attitude angle estimated accuracy is high, computing time is short.

Description of drawings

Fig. 1 is the object attitude angle schematic diagram.

Fig. 2 is 0～180 degree attitude angle experimental result picture of the SNS7 of T72.

Embodiment

Below in conjunction with accompanying drawing, the SAR object attitude angle method of estimation based on random hough conversion that the present invention is proposed is elaborated:

Based on the SAR object attitude angle method of estimation of random hough conversion, described method adopts quick Otsu partitioning algorithm, at first the SAR image is carried out the partial differential denoising, then the SAR image after denoising is carried out Otsu and cuts apart, and obtains cutting apart meticulous SAR target; At last the segmentation result that obtains is adopted pose estimation algorithm based on random Hough transformation; Its concrete steps are as follows:

Steps A, the partial differential denoising,

The partial differential denoising model hold concurrently to be removed the advantage that noise and edge keep two aspects simultaneously, as invasin, different gradients is got different diffuseness values with the inverse of gradient corresponding to picture point, carry out in edge weak level and smooth, with the protection marginal information; Become large in non-marginarium diffusion, larger smooth interaction is arranged; Model must lack edge is level and smooth automatically, and it is much level and smooth at smooth place, can identify boundary position to a certain extent, therefore can solve preferably the contradiction of denoising and Edge preserving, and can repair the edge that disconnects due to noise pollution and the cohesion of improving intra-zone

Step B, the Otsu method is cut apart

The method take image 1 the dimension histogram as foundation, be criterion to the maximum with variance between target and background, even thereby do not have obvious trough crest can obtain good threshold value between background and target yet;

If gradation of image progression be L (1,2 ..., L), selected threshold t' makes between inhomogeneity separation property best; At first obtain the probability of happening in each zone to be split based on histogram, take Threshold segmentation as two classes, obtain in each class class and inter-class variance, choose the ratio that makes class internal variance and inter-class variance

Maximum t' is as optimal threshold;

Gray-scale value separately is that the objective function of t' can be following three:

$\mathrm{\&lambda;}={\mathrm{\&sigma;}}_B^2/{\mathrm{\&sigma;}}_W^2,$ $k={\mathrm{\&sigma;}}_T^2/{\mathrm{\&sigma;}}_W^2,$ $\mathrm{\&eta;}={\mathrm{\&sigma;}}_B^2/{\mathrm{\&sigma;}}_T^2$

Wherein

For poor between class,

Be the class internal variance,

Be population variance, due to

Can get k=λ+1, η=λ/(λ+1), therefore top three equivalences; Here with asking η to replace getting λ, again because

Be not subjected to the separately impact of gray-scale value t', as long as optimal threshold t' satisfies following formula:

${\mathrm{\&sigma;}}_B^2\left(t^{\&prime;}\right)=\underset{1\&le;t<L}{\max }\left\{{\mathrm{\&sigma;}}_B^2\left(t\right)\right\}$

Wherein

μ _TBe the entire image average, ω (t) be image (1 ..., t) part ratio, μ (t) be (1 ..., t) part average

Step C, based on the SAR object attitude angle algorithm for estimating of random Hough transformation:

The random Hough transformation method

Random Hough transformation, adopted random sampling, convergence mapping (namely being calculated the parameter group of corresponding primitive by minimum point set) and dynamic link table, these three kinds of mechanism have replaced respectively the exhaustive and one-to-many mapping of Hough conversion, thereby the main difficulty that the Hough conversion that makes random Hough transformation overcome standard faces, make arithmetic speed accelerate, the peak value of totalizer array is more obvious, and computing time and storage space all greatly reduce;

The Hough principle is substantially as follows, and a point (x, y) can be mapped as straight line b=xk+y by (k, b), and each point can be mapped as straight line, finally can find the longest straight line by the accumulative total intersection point;

Based on the SAR object attitude angle algorithm for estimating of random Hough transformation, its implementation procedure is:

Step C-1, point set D of all boundary dot generation of the image after cutting apart, setup parameter space F;

Step C-2 chooses a pair of point (d from point set D _i, d _j), setpoint distance condition d if this does not satisfy pre-determined distance condition to point, chooses another to this step of continuation;

Step C-3 utilizes straight-line equation to (d _i, d _j) find the solution, obtain parameter space parameter (ρ, θ); If there is this f in parameter space F _{(ρ, θ)}The point that perhaps closes on it, this accumulated value adds 1, otherwise (ρ, θ) inserted parameter space, forms new accumulated value point f _{(ρ, θ)}Wherein, ρ, θ are respectively utmost point footpath and the polar angle under the polarization coordinate system.Cumulative complete after, will remove by corresponding point in point set D, step C-2 is returned in the initiation parameter space;

Step C-4 is for the f in parameter space _{(ρ, θ)}Set minimum value m, be separated by 90 ° of left and right and single threshold value of the θ in parameter space all surpassed two some accumulated values addition of m, obtain two points of respective value maximum, two limits of these two some correspondences are testing result.

according to said method, the SNS7 data of the T72 of MSTAR are tested, at first target is cut apart, then carry out pose estimation, can find out and adopt more accurate partitioning algorithm, can obtain edge more clearly, therefore Hough conversion estimation is more accurate, and because the reason of blocking fails all to be partitioned into image, therefore estimate that for Radon and enclosure rectangle the rule error ratio is larger, and the algorithm that we propose is due to the advantage that has absorbed Hough conversion and EPC conversion, only two sharp edges being carried out random Hough transformation estimates, therefore the attitude estimated result is more accurate.

Fig. 2 is 0～180 degree attitude angle experimental result of the SNS7 of T72, can find out except 50 degree left and right owing to cutting apart, be mistaken for the error ratio of 0 degree left and right large outside, remaining error is substantially all in 10 degree.

Claims (1)

1. based on the SAR object attitude angle method of estimation of random hough conversion, it is characterized in that, at first described method carries out the partial differential denoising to the SAR image, then the SAR image after denoising is carried out Otsu and cuts apart, and obtains cutting apart meticulous SAR target; At last the segmentation result that obtains is adopted pose estimation algorithm based on random Hough transformation, realize that accurate SAR object attitude angle estimates; Its concrete steps are as follows:

Steps A is carried out denoising with partial differential to the SAR image, makes image improve cohesion under the prerequisite of preserving edge;

Step B utilizes the Otsu method to cut apart the SAR target

Choose optimal separation gray-scale value threshold value t', make between the target and background of SAR image separation property best; Obtaining the probability of happening in each zone to be split based on histogram, according to setting threshold, is SAR target and background two classes with image segmentation; Optimal threshold t' satisfies following formula:

${\mathrm{\&sigma;}}_B^2\left(t^{\&prime;}\right)=\underset{1\&le;t<L}{\max }\left\{{\mathrm{\&sigma;}}_B^2\left(t\right)\right\}$

Wherein

μ _TBe the entire image average, ω (t) is the ratio of separation gray-scale value 1 to the t part of image, and μ (t) is the average of separation gray-scale value 1 to the t part of image, and t is for separating gray-scale value, 1≤t＜L; L is gradation of image progression, and L is natural number;

Step C, based on the SAR object attitude angle algorithm for estimating of random Hough transformation: its implementation procedure is:

Step C-1, point set D of all boundary dot generation of the image after cutting apart, setup parameter space F;

Step C-2 chooses a pair of point (d from point set D _i, d _j), setpoint distance condition d if this does not satisfy pre-determined distance condition to point, chooses another to this step of continuation; I, j are natural numbers, are the labels of point set D mid point;

Step C-3 utilizes straight-line equation to (d _i, d _j) find the solution, obtain parameter space parameter (ρ, θ); If there has been a f in parameter space F _{(ρ, θ)}The point that perhaps closes on it adds 1 with this accumulated value, otherwise (ρ, θ) inserted parameter space, forms new accumulated value point f _{(ρ, θ)}Wherein, ρ, θ are respectively utmost point footpath and the polar angle under the polarization coordinate system; Cumulative complete after, in point set D with this to a removing, step C-2 is returned in the initiation parameter space;

Step C-4 is for the f in parameter space _{(ρ, θ)}Set minimum value m, the polar angle θ in parameter space is separated by 90 °, and single threshold value all surpasses two some accumulated values addition of minimum value m, obtain two points of respective value maximum, two limits of these two some correspondences are testing result.

Images