CN104091335A - Polarization SAR image ship target detection method - Google Patents

Abstract

The invention discloses a polarization SAR image target detection method. The method mainly solves the problems that an existing detection method depends on energy, and the false alarm rate is high. The method comprises the implementation steps that two types of training samples including targets and clutter are selected, and feature extraction and energy normalization are carried out; a polarization dictionary learning target function is defined, and supervised learning is carried out on a polarization dictionary which can differentiate the clutter and the targets; polarization feature extraction and energy normalization are carried out on a tested image; sparse coding is carried out on a tested feature vector through the polarization dictionary; the reconstruction error proportion of the feature vector under a clutter sub-dictionary serves as test statistics; the test statistics is compared with a threshold, pixels larger than that threshold are determined as the targets, other pixels are determined as the clutter, and target detection is completed. According to the polarization SAR image ship target detection method, target detection is carried out only through polarization information, and the method has the advantages that robustness is good, the false alarm rate is low, and weak targets can be detected, and the method is applicable to polarization SAR image target detection when the contrast ratio of the targets to the clutter is low.

Description

Polarimetric SAR Image Ship Target Detection method

Technical field

The invention belongs to Radar Targets'Detection technical field, be particularly related to the Polarimetric SAR Image Ship Target Detection method of Weak target, the method is suitable for region to be detected may there is the Polarimetric SAR Image Ship Target Detection of weak Ship Target of radar back scattering.

Background technology

Naval vessel detect marine monitoring is significant, due to synthetic-aperture radar SAR have round-the-clock, round-the-clock work characteristics, SAR image has its unique advantage to marine monitoring compared to optical imagery.One group of full polarimetric SAR I comprises four POLARIZATION CHANNEL, i.e. level-horizontal polarization I conventionally _hH, level-vertical utmost point I _hV, vertical-horizontal polarization I _vH, vertical-vertical polarization I _vVthe identical image of four width sizes, is the measured value to four of the same area kinds of polarization modes, has level-vertical polarization I in the time meeting polarization reciprocity theorem _hVwith vertical-horizontal polarization I _vHapproximately equal.Polarization SAR can obtain the polarization characteristic of target, and polarization characteristic is to object construction, degree of roughness, and the reflection of the information such as symmetry, is different from other parameters of radar, and it has portrayed the key property of target, thereby in target detection, has clear superiority.

Existing polarization synthetic aperture radar image PolSAR Ship Target Detection, adopts CFAR detection method CFAR conventionally.These class methods first merge POLARIZATION CHANNEL, and the image after merging is carried out to statistical modeling, then under selected statistical model, set false alarm rate, thereby determine detection threshold, using Ship Target pixel as outlier detection out.These class methods all based on Ship Target echo amplitude or intensity far above clutter, there is the implicit hypothesis of the long-pending RCS of larger radar cross section.Although CFAR is in Ship Target Detection for CFAR detection method, has reached good accuracy of detection.But due under actual application environment, naval vessel may adopt the design with the long-pending RCS of lower radar cross section, in order to reduce the detection probability of Ship Target, now, the application that depends on the CFAR detection of echo amplitude or intensity will be subject to great restriction.

Adopt the Polarimetric detection method of Polarization scattering mechanism to be in recent years also studied much, wherein part object detection method PTD, polarization Symmetry Detection method RSD has obtained testing result relatively preferably.PTD method is utilized the polarization characteristic vector of clutter sample extraction, using this vectorial orthogonal subspaces as target subspace, construct the target subspace of a disturbance simultaneously, for test sample book, in this two sub spaces, carry out respectively rectangular projection and using the related coefficient of two projection vectors as test statistics.RSD is according to polarization symmetric theory, a large amount of sub-stochastic distribution of scattering in overwhelming majority natural feature on a map scattering unit, there is Polarization scattering symmetry, but man-made target is owing to existing very strong structural being often difficult to meet, accordingly from polarization coherence matrix, extract polarization asymmetry part as test statistics, to this test statistics modeling, complete the detection of target.These two kinds of detection methods that method is all the scattering mechanisms based on single, due to the difference of sea situation and the existence of coherent spot, these class methods can not always well be distinguished clutter and target resolution element, and robustness is lower, there will be higher false-alarm.

Summary of the invention

The object of the invention is the deficiency for above-mentioned prior art, and a kind of Ship Target Detection method of Polarimetric SAR Image is provided, and with in the situation that not relying on energy, realizes the detection of weak target and has lower false alarm rate.

The technical thought that realizes the object of the invention is: choose the training sample of Ship Target and clutter, learn a structuring polarization dictionary being made up of target polaron dictionary and clutter polaron dictionary.By design learning criterion, reduce the expression ability of sub-dictionary to cross sample, reduce the sub-dictionary of target to clutter sample, the sub-dictionary of clutter represents ability to target sample; At test phase, adopt two class hypothesis, wherein null hypothesis: sample is clutter sample; Alternative hypothesis: sample is target sample; Test sample book is carried out to Its Sparse Decomposition according to null hypothesis on polarization dictionary; By the coefficient decomposing, the reconstructed error ratio of calculating the sub-dictionary of clutter obtains the test statistics of sample; Test statistics and predetermined threshold value are compared, higher than threshold value be judged to target, otherwise be judged to clutter, the detection of realize target.Its concrete steps comprise as follows:

A. training step:

(A1) from Polarimetric SAR Image scene, choose M clutter pixel as the set of clutter training sample, pixel i=1 in pair set ..., M, adopts the interior pixel of 3 × 3 neighborhood windows of pixel i to calculate clutter polarization covariance matrix C _i;

(A2) according to clutter polarization covariance matrix C _i, calculate clutter polarization characteristic vector $\stackrel{\‾}{x_i}={[C_i\left(\mathrm{1,1}\right),C_i\left(\mathrm{2,2}\right),C_i\left(\mathrm{3,3}\right),R\left(C_i\left(\mathrm{1,2}\right)\right),R\left(C_i\left(\mathrm{1,2}\right)\right),R\left(C_i\left(\mathrm{1,3}\right)\right),I\left(C_i\left(\mathrm{1,3}\right)\right),I\left(C_i\left(\mathrm{2,3}\right)\right),I\left(C_i\left(\mathrm{2,3}\right)\right)]}^T$ Wherein () ^trepresenting matrix transposition; C in formula _i(a, b) represents polarization covariance matrix C _ithe element of the capable b row of a, a≤b, a=1,2,3, b=1,2,3; R (), I () represents respectively to get real part and imaginary-part operator;

(A3) to clutter polarization characteristic vector carry out energy normalizing, obtain normalized clutter polarization characteristic vector: wherein || || ₂represent 2 norm operators;

(A4) by normalization clutter proper vector x _ibe spliced into clutter eigenmatrix X ₁=[x ₁..., x _i..., x _m];

(A5) from Polarimetric SAR Image scene, choose N object pixel as the set of target training sample, pixel j=1 in pair set, ..., N, adopts and step (A1), (A2), (A3) identical operation, obtains object pixel j normalization polarization characteristic vector x _j, normalization target polarization characteristic vector is spliced into target signature matrix X ₂=[x ₁..., x _j..., x _n];

(A6) by clutter eigenmatrix X ₁with target signature matrix X ₂splice, obtain eigenmatrix: X=[X ₁, X ₂];

(A7) the definition clutter dictionary D that encodes ₁=[d ₁..., d _p..., d _k], each row d of dictionary _pbe called a wherein p=1 of atom ..., K, K is atom number; The objective definition dictionary D that encodes ₂=[d ₁..., d _q..., d _k], q=1 ..., K; Clutter encoder dictionary and target code dictionary are spliced, obtain encoder dictionary D=[D ₁, D ₂];

(A8) utilize eigenmatrix X, the method that adopts sparse coding and dictionary updating to replace iteration is learnt encoder dictionary D, obtains the encoder dictionary D after output is optimized ^*=D ^l+1, L is dictionary updating iterations.

B. testing procedure:

(B1), for one group of full polarimetric SAR I to be detected, define one and level-horizontal polarization image I _hHthe indicating image B that size is identical, in order to carry out mark to the position of object pixel in testing result;

(B2) the pixel t in Polarimetric SAR Image I to be detected is adopted to the operation identical with training stage step (A1), calculate test polarization covariance matrix C _t;

(B3) according to test polarization covariance matrix C _t, carry out the operation identical with training stage step (A2), extract test polarization proper vector

(B4) to test polarization proper vector carry out the operation identical with training stage step (A3), obtain normalization test polarization proper vector x _t;

(B5) suppose that test pixel point t is clutter, the encoder dictionary D that study obtains according to training stage step (A8) ^*, to test polarization proper vector x _tcarry out sparse coding, obtain test polarization proper vector x _tsparse coding coefficient $z_t=\left[\begin{array}{c}{z}_{t1}\\ z_{t2}\end{array}\right],$ Wherein z _t1, z _t2be respectively test polarization proper vector x _tat clutter encoder dictionary D ₁with target code dictionary D ₂code coefficient;

(B6) according to sparse coding coefficient z _t, build test statistics:

(B7) according to false alarm rate, detection threshold T=0.42 is set, by test statistics l (x _t) compare with this threshold value, differentiate for target if be greater than detection threshold T, in indicating image B, pixel t position is labeled as to 1, otherwise is labeled as 0;

(B8) treat each pixel of detected image I, all carry out the operation identical with step (B2)～(B7), complete the assignment to indicating image B, indicating image B is the testing result corresponding to image I to be detected.

The present invention compared with prior art has the following advantages:

1. can detect weak target

The present invention adopts polarization characteristic to build polarization characteristic vector, and to polarization characteristic vector energy normalizing, remove energy information completely, only depending on polarization information detects, need to not amass the hypothesis of RCS far above the long-pending RCS of extra large clutter normalized radar backscatter cross section by target radar backscattering cross, therefore can realize the detection to weak target.

2. detect and there is good robustness

The present invention adopts polarization dictionary to portray Polarization scattering mechanism, by dictionary learning method, study obtains polarization dictionary clutter polarization characteristic vector sum target feature vector to rarefaction representation ability, the defect that scattering mechanism is portrayed to scarce capacity that has overcome the method existence of the Polarization scattering mechanism based on single, has improved the robustness detecting.

3. reduced false alarm rate

The present invention, by the expression ability of siding stopping dictionary to cross sample, limits clutter dictionary D ₁to target feature vector X ₂re-configurability and target dictionary D ₂to clutter eigenmatrix X ₁re-configurability, improved the separating capacity of polarization dictionary to clutter and object pixel, thereby effectively reduced false-alarm, realized under lower false-alarm and target having been detected high probability.

4. without target sizes priori

Than traditional CFAR detection method CFAR, do not need CFAR detection window is set, thereby do not need according to target sizes, CFAR detection window size to be arranged, make to detect operation and become simple.

Brief description of the drawings

Fig. 1 is realization flow figure of the present invention;

Fig. 2 is that in the present invention, training sample is chosen schematic diagram;

Fig. 3 is the artificial image schematic diagram that emulation experiment of the present invention is used;

Fig. 4 is the test statistics comparison diagram obtaining in artificial image by the present invention and existing two kinds of Polarimetric detection methods;

Fig. 5 be the present invention and existing two kinds of Polarimetric detection methods in artificial image, receiver operating characteristic curve comparison diagram;

Fig. 6 is as span compared with large scene regional polarization energygram;

Fig. 7 is the inventive method test statistics output map;

Fig. 8 is the output of part object detection method PTD test statistics;

Fig. 9 is that false alarm rate is set to 5 × 10 ^-3time, the testing result of the inventive method;

Figure 10 is that false alarm rate is set to 5 × 10 ^-3time, the result that part object detection method PTD detects.

Embodiment

With reference to Fig. 1, recognition methods of the present invention comprises two stages of training and testing, and concrete steps are as follows:

One. the training stage

Step 1, from Fig. 2, label is to choose M=800 clutter pixel as the set of clutter training sample in 1,2,3 rectangle frame, the pixel i=1 in pair set ..., M, adopts 3 × 3 neighborhood territory pixel set of pixel i to calculate clutters polarization covariance matrix C _i:

$C_i={<k}_{\mathrm{si}}*k_{\mathrm{si}}^H>=\frac{1}{9}\underset{\mathrm{si}}{\mathrm{\&Sigma;}}{k}_{\mathrm{si}}*k_{\mathrm{si}}^H$

Wherein si represents 3 × 3 neighborhood territory pixel set with pixel i, k _sifor Polarization scattering vector corresponding to the pixel si adjacent with clutter pixel i, () ^hfor matrix complex conjugation operator, <> represents to get arithmetic mean.

Step 2, according to polarization covariance matrix C _i, calculate clutter polarization characteristic vector

Extract polarization covariance matrix C _i9 real numbers comprising of upper triangle element form clutter polarization characteristic vectors, its concrete operations are carried out according to following formula:

$\stackrel{\&OverBar;}{x_i}={[C_i\left(\mathrm{1,1}\right),C_i\left(\mathrm{2,2}\right),C_i\left(\mathrm{3,3}\right),R\left(C_i\left(\mathrm{1,2}\right)\right),R\left(C_i\left(\mathrm{1,2}\right)\right),R\left(C_i\left(\mathrm{1,3}\right)\right),I\left(C_i\left(\mathrm{1,3}\right)\right),I\left(C_i\left(\mathrm{2,3}\right)\right),I\left(C_i\left(\mathrm{2,3}\right)\right)]}^T,$

Wherein, C _i(a, b) represents polarization covariance matrix C _ithe element of the capable b row of a, a≤b, a=1,2,3, b=1,2,3; R (), I () represents respectively to get real part and imaginary-part operator; () ^trepresenting matrix transposition.

Step 3, to clutter polarization characteristic vector carry out energy normalizing, obtain normalized polarization characteristic vector x _i.

In order to eliminate the dependence of detection method to energy, this example carries out energy normalizing to polarization characteristic vector, obtains normalized polarization characteristic vector x _i:

$x_i=\stackrel{\&OverBar;}{x_i}/{\left|\right|\stackrel{\&OverBar;}{x_i}\left|\right|}_2.$

Wherein || || ₂represent 2 norm operators.

Step 4, by normalization clutter proper vector x _ibe spliced into clutter eigenmatrix X ₁=[x ₁..., x _i..., x _m], i=1 ..., M, wherein clutter training sample number M=800.

Step 5, calculates target signature matrix X ₂.

From Polarimetric SAR Image scene, choose N object pixel as the set of target training sample, pixel j=1 in pair set ..., N, adopts the operation identical with training step 1,2,3, obtains object pixel j normalization target polarization characteristic vector x _j, normalization target polarization characteristic vector is spliced into target signature matrix X ₂=[x ₁..., x _j..., x _n], j=1 ..., N, wherein target training sample number N=730.

Step 6, by clutter eigenmatrix X ₁with target signature matrix X ₂splice, obtain training characteristics matrix X=[X ₁, X ₂].

Step 7, definition encoder dictionary D, each row of dictionary D are called an atom.

The definition clutter dictionary D that encodes ₁=[d ₁..., d _p..., d _k], wherein p=1 ..., K, K=10 is sub-dictionary atom number; The objective definition dictionary D that encodes ₂=[d ₁..., d _q..., d _k], q=1 ..., K; Clutter encoder dictionary and target code dictionary are spliced, obtain encoder dictionary D=[D ₁, D ₂], using each column vector of D as an atom.

Step 8, utilizes training characteristics matrix X, adopts sparse coding and dictionary updating to replace the method for iteration, and to encoder dictionary, D learns.

8a), in order to improve the separating capacity of dictionary to clutter and target, be defined as follows objective function dictionary learnt:

$\underset{D,Z}{\arg \min }\{\frac{1}{2}{\left|\right|X-\mathrm{DZ}\left|\right|}_F^2+\mathrm{\&lambda;}{\left|\right|Z\left|\right|}_1\}+\{\frac{1}{2}{\left|\right|D_1{Z}_{\mathrm{1,2}}\left|\right|}_F^2+\frac{1}{2}{\left|\right|D_2{Z}_{\mathrm{2,1}}\left|\right|}_F^2\}+\frac{\mathrm{\&gamma;}}{2}{\left|\right|D_1^T{D}_2\left|\right|}_F^2$

s.t. i＝1,...,2K

Wherein, sparse coding coefficient $Z=\left[\begin{array}{cc}{Z}_{\mathrm{1,1}}& Z_{\mathrm{1,2}}\\ Z_{\mathrm{2,1}}& Z_{\mathrm{2,2}}\end{array}\right],$ Z _i,jrepresentation feature matrix X _jcorresponding to sub-dictionary D _icode coefficient, i=1,2, j=1,2; In objective function sparse reconstruct is retrained, || || ₁representing matrix l ₁norm, || || _ffor matrix F norm operator; for retraining the rarefaction representation ability of sub-dictionary to cross sample; representing the correlativity between sub-dictionary, is equally also in order to reduce sub-dictionary cross means ability; Two balance parameters λ, γ value is respectively λ=0.01, γ=0.005;

8b) to encoder dictionary initialization, obtain initialization dictionary

Respectively from clutter eigenmatrix X ₁with target signature matrix X ₂in random select K=10 clutter proper vector, to clutter encoder dictionary D ₁with target code dictionary D ₂carry out initialization, obtain initialization codes dictionary wherein r=0 ..., L is that iterative steps r=0 is primary iteration step number, L=30 is greatest iteration step number;

8c), for r step iteration, utilize encoder dictionary according to optimization aim function, adopt proximal-point method to carry out sparse coding to eigenmatrix X, obtain sparse coding matrix of coefficients $Z^r=\left[\begin{array}{cc}{Z}_{\mathrm{1,1}}^r& Z_{\mathrm{1,2}}^r\\ Z_{\mathrm{2,1}}^r& Z_{\mathrm{2,2}}^r\end{array}\right];$

8d) utilize sparse coding matrix of coefficients Z ^r, solve and obtain encoder dictionary D according to optimization aim function ^r+1;

8e) upgrade iterations r=r+1, if iterations r is <=L, return to step 8b), otherwise termination of iterations, the encoder dictionary D after output is optimized ^*=D ^l+1.

Two. test phase

Step 9, for one group of full polarimetric SAR I to be detected, defines one and level-horizontal polarization image I _hHthe indicating image B that size is identical, in order to carry out mark to the position of object pixel in testing result.

Step 10, adopts the operation identical with training stage step 1 to the pixel t in Polarimetric SAR Image I to be detected, calculates test polarization covariance matrix C _t.

Step 11, according to test polarization covariance matrix C _t, carry out the operation identical with training stage step 2, extract test polarization proper vector

Step 12, to test polarization proper vector carry out the operation identical with training stage step 3, obtain normalization test polarization proper vector x _t.

Step 13, supposes that test pixel point t is clutter, according to training stage step 8e) the encoder dictionary D that obtains ^*, to test polarization proper vector x _tadopt and training stage step 8c) identical sparse coding operation, obtain test polarization proper vector x _tsparse coding coefficient $z_t=\left[\begin{array}{c}{z}_{t1}\\ z_{t2}\end{array}\right],$ Wherein z _t1, z _t2be respectively test polarization proper vector xt at clutter encoder dictionary D ₁with target code dictionary D ₂on code coefficient.

Step 14, according to sparse coding coefficient z _t, build test statistics: wherein || x _t-D ₁z _t1|| ₂for sparse coding is at the sub-dictionary D of clutter ₁the reconstructed error of lower acquisition, clutter sample is at the sub-dictionary D of clutter ₁under there is less reconstructed error, and for target sample, at the sub-dictionary D of clutter ₁next have a larger reconstructed error.

Step 15, arranges detection threshold T=0.42 according to false alarm rate, by test statistics l (x _t) compare with detection threshold T, if l is (x _t) be greater than detection threshold T, differentiate for target, and in indicating image B, pixel t position is labeled as to 1, otherwise differentiate for clutter, in indicating image B, pixel t position is labeled as to 0.

Step 15, treats each pixel of detected image I, all carries out with test phase step 10, to the identical operation of step 15, completing the assignment to indicating image B, and indicating image B is the testing result corresponding to image I to be detected.

Effect of the present invention can describe by following emulation experiment and the checking of large scene test experience

1. experiment condition

Testing data used is C-band RADARSAT-2 data set.These data are certain harbour complete polarization SAR data on August 4th, 2010.Azimuth-range resolution is about 8 × 12m, and test zone polarization energy image span as shown in Figure 2.

Training sample is chosen: adopt the testing result of energy threshold detection method as priori, obtain target and clutter resolution cell.In Fig. 2, choose in rectangle frame 1～3 region totally 800 pixels, as extra large clutter training sample.Choose 730 object pixels that obtain in 7 targets in oval marks region as target training sample.

2. experiment content

Experiment 1: generate artificial image and artificial image is detected:

In order to obtain the real information of target pixel location, the result of detection method is carried out to qualitative assessment, from Fig. 2, choose target and extra large clutter pixel generates artificial image, generate target pixel location accurately known artificial image as Fig. 3, wherein:

Fig. 3 (a): level-horizontal polarization I _hHmagnitude image;

Fig. 3 (b): level-vertical polarization I _hVmagnitude image;

Fig. 3 (c): level-horizontal polarization I _vVmagnitude image;

Fig. 3 (d): object pixel actual position marking image.

The training sample that utilization the is chosen dictionary learning that polarizes, obtains polarization dictionary D ^*, utilize polarization dictionary D ^*image shown in Fig. 3 is detected, detect respectively by the inventive method and existing two kinds of methods, calculate three kinds of test statistics images that method is corresponding, result is as Fig. 4, wherein:

The test statistics that Fig. 4 (a) obtains for the inventive method;

Fig. 4 (b) is the test statistics that existing part object detection method PTD obtains;

The test statistics that Fig. 4 (c) obtains for polarization Symmetry Detection method RSD.

By Fig. 4 (a) and Fig. 4 (b), Fig. 4 (c) compares, can find out that the inventive method can detect three targets in artificial image graph 3 effectively, and the test statistics of clutter and target is distinguished obviously, can better suppress the test statistics of clutter output.

According to the test statistics of three kinds of methods, ask and be receiver operating characteristic curve ROC as Fig. 5, wherein horizontal ordinate represents false alarm rate, ordinate is verification and measurement ratio; Receiver operating characteristic curve ROC area under a curve AUC size is commonly used to the quality of qualitative assessment detection method, and area under curve AUC is larger, and detection method is better.

As can be seen from Figure 5: the inventive method, part object detection method PTD, the receiver operating characteristic curve ROC area under curve AUC of polarization Symmetry Detection method RSD is respectively 0.9957,0.9942,0.8665.Receiver operating characteristic curve ROC area under a curve AUC value of the present invention is slightly larger than the area under curve AUC that part object detection method PTD obtains, and the area under curve AUC obtaining much larger than polarization symmetry method RSD.Show the inventive method realize target detection better.

Experiment 2: the test experience on measuring image: in order to verify that the present invention is in the validity compared with under large scene, apply the present invention to Fig. 2 entire image and carry out test experience.By the inventive method and existing detection method, region shown in Fig. 2 is detected respectively, testing result is as Fig. 6-Figure 10, wherein:

Fig. 6 is polarization energy image span, and Ship Target marks by rectangular box, and weak target adopts ellipse to carry out mark;

Fig. 7 is the inventive method test statistics output map;

Fig. 8 is the output of part object detection method PTD test statistics;

Fig. 9 is that false alarm rate is set to 5 × 10 ^-3time, corresponding threshold value is 0.42, the testing result of the inventive method, and oval marks region is corresponding to the weak target in Fig. 6;

Figure 10 is that false alarm rate is set to 5 × 10 ^-3time, corresponding threshold value is 0.76, the result that part object detection method PTD detects, and what white rectangle collimation mark was remembered is the false dismissal target in detecting.

Comparison diagram 7, Fig. 8 can find, and the inventive method is in the region, the upper left corner of image, and clutter inhibition is better, and clutter test statistic output entirety is lower.

Comparison diagram 9, the testing result in Figure 10 can be found:

1) the inventive method can all detect target-marking, and part object detection method PTD exists false dismissal, and false dismissal target is remembered by white rectangle collimation mark in Figure 10;

2) it is better that the inventive method detects goal succession, and this is under identical false alarm rate, and the verification and measurement ratio of the inventive method is the reflection in Pixel-level higher than part object detection method PTD.

In Fig. 9, oval marks region is corresponding one by one with elliptic region in Fig. 6, and these elliptic region correspondences Weak target, shows that the inventive method can successfully detect these targets.

Comprehensive Experiment 1, experiment 2 can draw: the inventive method is only utilized polarization information, realizes polarimetric synthetic aperture radar PolSAR Ship Target Detection, and compared to existing Polarimetric detection method, can under higher verification and measurement ratio, obtain lower false-alarm.And because the present invention does not rely on energy information, can detect weak realization of goal.

Claims (4)

1. a Ship Target Detection method for Polarimetric SAR Image, comprising:

A. training step:

(A1) from Polarimetric SAR Image scene, choose M clutter pixel as the set of clutter training sample, pixel i=1 in pair set ..., M, adopts the interior pixel of 3 × 3 neighborhood windows of pixel i to calculate clutter polarization covariance matrix C _i;

(A2) according to clutter polarization covariance matrix C _i, calculate clutter polarization characteristic vector $\stackrel{\&OverBar;}{x_i}={[C_i\left(\mathrm{1,1}\right),C_i\left(\mathrm{2,2}\right),C_i\left(\mathrm{3,3}\right),R\left(C_i\left(\mathrm{1,2}\right)\right),R\left(C_i\left(\mathrm{1,2}\right)\right),R\left(C_i\left(\mathrm{1,3}\right)\right),I\left(C_i\left(\mathrm{1,3}\right)\right),I\left(C_i\left(\mathrm{2,3}\right)\right),I\left(C_i\left(\mathrm{2,3}\right)\right)]}^T$ Wherein () ^trepresenting matrix transposition; C in formula _i(a, b) represents polarization covariance matrix C _ithe element of the capable b row of a, a≤b, a=1,2,3, b=1,2,3; R (), I () represents respectively to get real part and imaginary-part operator;

(A3) to clutter polarization characteristic vector carry out energy normalizing, obtain normalized clutter polarization characteristic vector: wherein || || ₂represent 2 norm operators;

(A4) by normalization clutter proper vector x _ibe spliced into clutter eigenmatrix X ₁=[x ₁..., x _i..., x _m];

(A5) from Polarimetric SAR Image scene, choose N object pixel as the set of target training sample, pixel j=1 in pair set, ..., N, adopts and step (A1), (A2), (A3) identical operation, obtains object pixel j normalization polarization characteristic vector x _j, normalization target polarization characteristic vector is spliced into target signature matrix X ₂=[x ₁..., x _j..., x _n];

(A6) by clutter eigenmatrix X ₁with target signature matrix X ₂splice, obtain eigenmatrix: X=[X ₁, X ₂];

(A7) the definition clutter dictionary D that encodes ₁=[d ₁..., d _p..., d _k], each row d of dictionary _pbe called a wherein p=1 of atom ..., K, K is atom number; The objective definition dictionary D that encodes ₂=[d ₁..., d _q..., d _k], q=1 ..., K; Clutter encoder dictionary and target code dictionary are spliced, obtain encoder dictionary D=[D ₁, D ₂];

(A8) utilize eigenmatrix X, the method that adopts sparse coding and dictionary updating to replace iteration is learnt encoder dictionary D, obtains the encoder dictionary D after output is optimized ^*=D ^l+1, L is dictionary updating iterations:

B. testing procedure:

(B1), for one group of full polarimetric SAR I to be detected, define one and the flat ?horizontal polarization of water image I _hHthe indicating image B that size is identical, in order to carry out mark to the position of object pixel in testing result;

(B2) the pixel t in Polarimetric SAR Image I to be detected is adopted to the operation identical with training stage step (A1), calculate test polarization covariance matrix C _t;

(B3) according to test polarization covariance matrix C _t, carry out the operation identical with training stage step (A2), extract test polarization proper vector

(B4) to test polarization proper vector carry out the operation identical with training stage step (A3), obtain normalization test polarization proper vector x _t;

(B5) suppose that test pixel point t is clutter, the encoder dictionary D that study obtains according to training stage step (A8) ^*, to test polarization proper vector x _tcarry out sparse coding, obtain test polarization proper vector x _tsparse coding coefficient $z_t=\left[\begin{array}{c}{z}_{t1}\\ z_{t2}\end{array}\right],$ Wherein z _t1, z _t2be respectively test polarization proper vector x _tat clutter encoder dictionary D ₁with target code dictionary D ₂code coefficient;

(B6) according to sparse coding coefficient z _t, build test statistics:

(B7) according to false alarm rate, detection threshold T=0.42 is set, by test statistics l (x _t) compare with this threshold value, differentiate for target if be greater than detection threshold T, in indicating image B, pixel t position is labeled as to 1, otherwise is labeled as 0;

(B8) treat each pixel of detected image I, all carry out the operation identical with step (B2)～(B7), complete the assignment to indicating image B, indicating image B is the testing result corresponding to image I to be detected.

2. according to the method described in claims 1, in wherein said step (A1), calculate clutter polarization covariance matrix C _i, adopt following formula to calculate:

$C_i={<k}_{\mathrm{si}}*k_{\mathrm{si}}^H>=\frac{1}{9}\underset{\mathrm{si}}{\mathrm{\&Sigma;}}{k}_{\mathrm{si}}*k_{\mathrm{si}}^H,$

Wherein si represents the interior pixel set of 3 × 3 neighborhood windows of pixel i, k _sifor Polarization scattering vector corresponding to the pixel si adjacent with clutter pixel i, () ^hfor matrix complex conjugation operator, <> represents to get arithmetic mean.

3. according to the method described in claims 1, the method that wherein the described employing sparse coding of step (A8) and dictionary updating replace iteration is learnt encoder dictionary D, carries out as follows:

(A8.1) definition dictionary learning objective function:

$\underset{D,Z}{\arg \min }\{\frac{1}{2}{\left|\right|X-\mathrm{DZ}\left|\right|}_F^2+\mathrm{\&lambda;}{\left|\right|Z\left|\right|}_1\}+\{\frac{1}{2}{\left|\right|D_1{Z}_{\mathrm{1,2}}\left|\right|}_F^2+\frac{1}{2}{\left|\right|D_2{Z}_{\mathrm{2,1}}\left|\right|}_F^2\}+\frac{\mathrm{\&gamma;}}{2}{\left|\right|D_1^T{D}_2\left|\right|}_F^2$

s.t. i＝1,...,2K

Wherein, sparse coding coefficient $Z=\left[\begin{array}{cc}{Z}_{\mathrm{1,1}}& Z_{\mathrm{1,2}}\\ Z_{\mathrm{2,1}}& Z_{\mathrm{2,2}}\end{array}\right],$ Z _i,jrepresentation feature matrix X _jcorresponding to sub-dictionary D _icode coefficient, i=1,2, j=1,2; In formula || || ₁, || || _fbe respectively matrix l ₁norm operator, F norm operator, λ, γ is balance parameters;

(A8.2) respectively from clutter eigenmatrix X ₁with target signature matrix X ₂in random select K clutter proper vector, to clutter encoder dictionary D ₁with target code dictionary D ₂carry out initialization, obtain initialization codes dictionary wherein r=0 ..., L is iterative steps, L is greatest iteration step number, establishes primary iteration step number r=0;

(A8.3) for r step iteration, utilize encoder dictionary according to optimization aim function, adopt proximal ?point method eigenmatrix X is carried out to sparse coding, obtain sparse coding matrix of coefficients $Z^r=\left[\begin{array}{cc}{Z}_{\mathrm{1,1}}^r& Z_{\mathrm{1,2}}^r\\ Z_{\mathrm{2,1}}^r& Z_{\mathrm{2,2}}^r\end{array}\right];$

(A8.4) utilize sparse coding matrix of coefficients Z ^r, solve and obtain encoder dictionary D according to optimization aim function ^r+1;

(A8.5) upgrade iterations r=r+1, if iterations r is <=L, return to step (A8.2), otherwise termination of iterations, the encoder dictionary D after output is optimized ^*=D ^l+1.

4. according to the method described in claims, a kind of method that detection threshold T is set according to false alarm rate of wherein said step (B7), is to adopt Beta to distribute to test statistics l (x _t) carry out statistical modeling, obtain statistical distribution g (l (x _t)); An artificial selected false alarm rate P _f, adopt following formula to calculate detection threshold T:

$1-P_f={\&Integral;}_0^{T}g\left(l\left(x_t\right)\right)\mathrm{dl}\left(x_t\right)$

According to statistical model g (l (x _t)) can unique true upper limit of integral T, upper limit of integral is detection threshold T.