CN105738888A - Double-characteristic sea surface floating small-target detection method based on sea clutter suppression - Google Patents

Abstract

The invention discloses a double-characteristic sea surface floating small-target detection method based on sea clutter suppression, and mainly solve the problem of the low detection probability for a small sea surface target in short observation time in the prior art. The implementation process of the method comprises the following steps: 1, block whitening is performed on a training unit time sequence and a reference unit time sequence of pure clutter data; 2, relative time frequency double-characteristic vectors of training units are extracted; 3, a convex hull is formed by utilizing the extracted characteristic vectors, and a decision region is obtained by utilizing a convex hull learning algorithm; 4, relative time frequency double-characteristic vectors of to-be-detected units are extracted; 5, a detection statistical quantity is calculated according to the convex hull forming the decision region and the relative time frequency double-characteristic vectors of the to-be-detected units; and 6, whether a target exists is judged according to the detection statistical quantity, the target is judged to exist if the detection statistical quantity is greater than zero, otherwise, the target is judged to not exist. Detection for a small floating target under a sea clutter background in short observation time can be performed, and the method can used in identification and tracking for a small sea surface floating and low-speed target.

Description

Bicharacteristic offshore floating small target detecting method based on ocean clutter cancellation

Technical field

The invention belongs to signal processing technology field, be specifically related to a kind of object detection method, can be used for the recognition and tracking to offshore floating, at a slow speed Small object.

Background technology

Sea clutter is the radar return reflected from surface, sea that radar receives, and surface search radar, when sea is detected, particularly to when such as the floating Small object such as canoe, floating ice, floating thing detects on sea, can be subject to the impact of sea clutter unavoidably.The intensity of sea clutter can change with the difference of radar parameter, radar illumination direction, sea situation etc..Under high-resolution sea clutter background, clutter presents stronger non-Gaussian feature, and the existence of sea spike causes that a large amount of false-alarms occurs in the object detection method utilizing time domain energy to accumulate, and therefore, the method that floating Small object is difficult to by conventional energy is accumulated detects.

For solving this difficult problem, this is made that substantial amounts of research by a lot of scholars.Constantly perfect along with sea clutter statistical model, many self-adapting detecting methods are suggested, and sea clutter is modeled as complex Gaussian model by such method, non-stationary property during due to sea clutter empty, such method needs first sea clutter to be suppressed, and has certain limitation.Document " Hu; J.; Tung; W.W.andGao; J.B.:Detectionoflowobservabletargetswithinseaclutterbyst ructurefunctionbasedmultifractalanalysis; IEEETrans.AntennasPropag., 54 (1): 136-143,2006. " in propose based on the detection method of sea fractal characteristic; can effectively detect target when observation time is longer; but single ripple position generally cannot be carried out long resident observation by radar, therefore is difficult to be generalized in practical application based on the detector of fractal characteristic.

Detection to offshore floating Small object, a lot of methods meet certain statistical model for supposed premise with sea clutter, but existing statistical model is difficult to describe the complex characteristics of sea clutter, and this causes that testing result has certain limitation；Self-adapting detecting method is when sea situation is more complicated, namely when target cannot be distinguished by Doppler domain with clutter, it is impossible to offshore floating or low speed Small object are detected；Reaching good testing result when observation time is longer based on fractal object detection method, when observation time shortens, detection performance has and is decreased obviously, it is impossible to meet the requirement of seaward-looking radar.

Summary of the invention

Present invention aims to the deficiency of above-mentioned prior art, it is proposed to a kind of bicharacteristic offshore floating small target detecting method based on ocean clutter cancellation, to improve the detection performance to offshore floating Small object, meet the requirement of radar search at sea.

For achieving the above object, technical scheme includes as follows:

(1) from echo data, choose training unit, reference unit and unit to be detected:

Utilizing radar transmitter that sea is sent signal, utilize radar receiver to receive the echo data returned by sea surface reflection, this echo data is divided into pure clutter data and the echo data comprising target；

From pure clutter data, selected part distance unit is as one group of training unit, this training unit time series z is: z=[z (1), z (2), ..., z (N)], around training unit, choose Q close on unit as reference unit, this reference unit time series z_pFor: z_p=[z_p(1),z_p(2),…,z_p(N)], p=1,2 ..., Q, Q is number of reference, and N is seasonal effect in time series length；

From the echo data comprising target, selected part distance unit is as unit T to be detected；

(2) to the training unit time series z of pure clutter echo data and reference unit time sequence z_pCarry out block albefaction, obtain the training unit time series after albefactionWith the reference unit time series after albefaction

$\hat{z}=\[\hat{z}\left(1\right),\hat{z}\left(2\right),...,\hat{z}\left(N\right)\],{\hat{z}}_p=\[{\hat{z}}_p\left(1\right),{\hat{z}}_p\left(2\right),...,{\hat{z}}_p\left(N\right)\];$

(3) the training unit time series after albefaction is utilizedWith the reference unit time series after albefactionExtract the ridge energy η of the relative time-frequency distributions of training unit₁The ridge total variation η of (z) and the relative time-frequency distributions of training unit₂Z () both features, construct relative time-frequency bicharacteristic vector η: η of pure clutter data=[η 1 (z), η₂(z)]^T, []^TRepresent and matrix is carried out transposition；

(4) utilize the relative time-frequency bicharacteristic vector η of pure clutter data, two-dimensional feature space obtains two dimension convex closureAnd at given false-alarm probability P_FUnder, utilize greedy convex closure learning algorithm to two dimension convex closureShrink, and using the convex closure after contraction as detecting decision region Ω；

(5) unit T to be detected is extracted the ridge energy η of relative time-frequency distributions₁(T) with the ridge total variation η of relative time-frequency distributions₂(T) both features, construct the relative time-frequency bicharacteristic vector of unit to be detected: η_T=[η₁(T),η₂(T)]^T；

(6) characteristic time-frequency relative to the unit to be detected bicharacteristic vector η according to the convex closure constituting detection decision region Ω_T=[η₁(T),η₂(T)]^T, calculate detection statistic ω:

$\mathrm{\ω}=\underset{1\≤j\≤r}{min}\left\{\det \left(\left[\begin{array}{ccc}{\mathrm{\η}}_1\left(T\right)& {\mathrm{\η}}_1\left(T\right)& 1\\ x_j& y_j& 1\\ x_{j+1}& y_{j+1}& 1\end{array}\right]\right)\right\},$

Wherein, min{ } represent and take minima, det () represents and seeks matrix determinant, and r is the number on the summit of composition convex closure, x_jRepresent the ridge energy of the relative time-frequency distributions on jth convex closure summit, y_jThe ridge total variation of the relative time-frequency distributions on expression jth convex closure summit, j=1,2 ..., r；

(7) judge whether target exists according to the size of detection statistic ω: if detection statistic ω is more than zero, it was shown that the relative time-frequency bicharacteristic vector η of unit to be detected_TOutside detection decision region Ω, then judge that target exists, otherwise, it is determined that target is absent from.

The present invention has the advantage that compared with the prior art

1) present invention extracts time-frequency characteristics two kinds different from sea clutter sequence, and combine and utilize both time-frequency characteristics that the separating capacity of pure sea clutter data with echo data containing target is completed the detection to sea-surface target, compare the traditional detection method utilizing single features, it is possible in shorter observation time, obtain better Detection results；

2) present invention utilizes greedy convex closure learning algorithm, owing to amount of calculation is little, it is possible to obtain detection decision region rapidly, improve the training speed of detector, be more suitable for applying in practical application.

3) present invention utilizes block albefaction to carry out clutter reduction, strengthens target echo simultaneously, utilizes the average speckle covariance matrix covariance matrix to replace direct estimation to obtain, it is possible to effectively reduce the block effect of target energy in whitening process.

Accompanying drawing explanation

Fig. 1 is the flowchart of the present invention；

Fig. 2 is when observation time is 512ms, the detection Performance comparision figure under polarizing at four kinds with existing two kinds of detection methods by the present invention.

Fig. 3 is when observation time is 1024ms, the detection Performance comparision figure under polarizing at four kinds with existing two kinds of detection methods by the present invention.

Detailed description of the invention

With reference to Fig. 1, the present invention includes training and detection two parts, and it specifically comprises the following steps that

Step 1, obtains echo data, and chooses training unit, reference unit and unit to be detected from echo data.

Utilizing radar transmitter that sea is sent signal, utilize radar receiver to receive the echo data returned by sea surface reflection, this echo data is divided into pure clutter data and the echo data comprising target；

From pure clutter data, selected part distance unit is as one group of training unit, this training unit time series z is: z=[z (1), z (2), ..., z (N)], around training unit, choose Q close on unit as reference unit, this reference unit time series z_pFor: z_p=[z_p(1),z_p(2),…,z_p(N)], p=1,2 ..., Q, Q is number of reference, and N is seasonal effect in time series length；

From the echo data comprising target, selected part distance unit is as unit T to be detected.

Step 2, the training unit time series z and reference unit time sequence z to pure clutter echo data_pCarry out block albefaction.

Block albefaction refers to: time series is divided into the short gauge block of non-overlapping copies, utilizes the speckle covariance matrix of each short gauge block that reference unit time series estimates, each short gauge block is carried out albefaction, and its step is as follows:

(2a) by training unit time series z and reference unit time series z_pIt is divided into the short amount of the non-overlapping copies that length is M respectively, it may be assumed that

Z=[z₁,z₂,…,z_m,…,z_N/M],

z_p=[z_p,1,z_p,2,…,z_p,m,…,z_p,N/M], p=1,2 ..., Q

Wherein, z_mRepresent training unit seasonal effect in time series m-th short amount, z_p,mExpression reference unit seasonal effect in time series m-th short amount, m=1,2 ..., N/M；

(2b) above-mentioned each short amount is utilized, to training unit time series z and reference unit time sequence z_pCarry out block albefaction, obtain the training unit time series after albefactionWith the reference unit time series after albefaction

$\hat{z}=\[\hat{z}\left(1\right),\hat{z}\left(2\right),...,\hat{z}\left(N\right)\]=\[{\hat{z}}_1,{\hat{z}}_2,...,{\hat{z}}_m,...,{\hat{z}}_{N/M}\],$

${\hat{z}}_p=\[{\hat{z}}_p\left(1\right),{\hat{z}}_p\left(2\right),...,{\hat{z}}_p\left(N\right)\]=\[{\hat{z}}_{p,1},{\hat{z}}_{p,2},\mathrm{...},{\hat{z}}_{p,m},\mathrm{...},{\hat{z}}_{p,N/M}\],$

Wherein,Represent the training unit seasonal effect in time series m-th short amount after albefaction,Represent the reference unit seasonal effect in time series m-th short amount after albefaction,Represent reference unit seasonal effect in time series m-th short amount z_p,mSpeckle covariance matrix；

Described to reference unit seasonal effect in time series m-th short amount z_p,mSpeckle covariance matrix estimate, can by existing accomplished in many ways, for instance maximum likelihood estimate, the normalization sample covariance matrix estimation technique and the near-maximum-likelihood estimation technique, this example adopt the near-maximum-likelihood estimation technique obtain z_p,mSpeckle covariance matrix

Step 3, utilizes the training unit time series after above-mentioned albefactionWith the reference unit time series after albefactionExtract the ridge energy η of the relative time-frequency distributions of training unit₁The ridge total variation η of (z) and the relative time-frequency distributions of training unit₂Z () both features, construct the relative time-frequency bicharacteristic vector η of pure clutter data.

(3a) the smooth Eugene Wigner-Willie distribution of training unit is calculatedWith the smooth Eugene Wigner of reference unit-Willie distribution

$W(n,l,\hat{z})=\underset{m=-E}{\overset{E}{\Σ}}g\left(m\right)\underset{k=-F}{\overset{F}{\Σ}}h\left(k\right)\hat{z}(n+m+k){\hat{z}}^{*}(n+m-k)\exp (-4{\mathrm{j\πkl\Δf}}_d),$

$D(n,l,{\hat{z}}_p)=\underset{m=-E}{\overset{E}{\Σ}}g\left(m\right)\underset{k=-F}{\overset{F}{\Σ}}h\left(k\right){\hat{z}}_p(n+m+k){{\hat{z}}_p}^{*}(n+m-k)\exp (-4{\mathrm{j\πkl\Δf}}_d),$

Wherein, subscript * represents conjugation, and g (m) is time smoothing window, and h (k) is Frequency Smooth window, and the half of E express time smoothing windows length, F represents the half of Frequency Smooth window length, Δ f_dFor the sampling interval of normalization Doppler frequency, n=1,2 ..., N, l=1,2 ..., N；

Here time smoothing window and Frequency Smooth window adopt but are not limited to Hanning window, hamming window, Blackman window, triumphant plucked instrument window, this example adopt length be the triumphant plucked instrument window of 31 as time smoothing window, length is that the triumphant plucked instrument window of 63 is as Frequency Smooth window；

(3b) the smooth Eugene Wigner-Willie distribution of reference unit is utilizedEstimate training unit average time-frequency distributions A (n, l):

$A(n,l)=\frac{1}{9Q}\underset{p=1}{\overset{Q}{\Σ}}\underset{n^{\′}=-1}{\overset{1}{\Σ}}\underset{l^{\′}=-1}{\overset{1}{\Σ}}D(n+n^{\′},l+l^{\′},{\hat{z}}_p),$

Wherein, the value of n ' is-1,0,1, and the value of l ' is-1,0,1；

(3c) the smooth Eugene Wigner-Willie distribution of training unit is utilized(n l), calculates the relative time-frequency distributions of training unit with the average time-frequency distributions A of training unit

$B(n,l,\hat{z})=\frac{W(n,l,\hat{z})}{A(n,l)};$

(3d) the relative time-frequency distributions of training unit is calculatedTime-frequency ridge R (n):

$R\left(n\right)=\underset{l}{\arg \max }\left\{B(n,l,\hat{z})\right\},$

Wherein,RepresentTake the value of correspondence l during maximum；

(3e) the relative time-frequency distributions of training unit is utilizedRelative time-frequency distributions with training unitTime-frequency ridge R (n), calculate the ridge energy η of relative time-frequency distributions of training unit₁The ridge total variation η of (z) and the relative time-frequency distributions of training unit₂(z):

${\mathrm{\η}}_1\left(z\right)=\underset{n=1}{\overset{N}{\Σ}}B(n,R(n),\hat{z}),$

${\mathrm{\η}}_1\left(z\right)={\mathrm{\Δf}}_d\underset{n=2}{\overset{N}{\Σ}}|R(n+1)-R\left(n\right)|;$

(3f) the relative time-frequency distributions ridge energy η of training unit is utilized₁(z) time-frequency distributions ridge total variation η relative to training unit₂Z (), constructs the relative time-frequency bicharacteristic vector η of pure clutter data:

η=[η₁(z),η₂(z)]^T,

Wherein, []^TRepresent and matrix is carried out transposition.

Step 4, utilizes the relative time-frequency bicharacteristic vector η of pure clutter data, determines detection decision region Ω by convex closure learning algorithm.

Existing convex closure learning algorithm has: quickly convex closure learning algorithm, greedy convex closure learning algorithm, Graham scanning method, gift package pack, and this example adopts greedy convex closure learning algorithm to determine detection decision region Ω, and its step is as follows:

(4a) utilize the relative time-frequency bicharacteristic vector η of pure clutter data, form training sample set S:

$S=\{{\mathrm{\η}}_i={\[{\mathrm{\η}}_1^i\left(z\right),{\mathrm{\η}}_2^i\left(z\right)\]}^T:i=1,2,...,I\},$

Wherein, η i represents the relative time-frequency bicharacteristic vector of i-th training sample,Represent the ridge energy of the relative time-frequency distributions of i-th training sample,Representing the ridge total variation of the relative time-frequency distributions of i-th training sample, I is training sample number, and it is 20000 that this example takes number of training；

(4b) utilize training sample set S, obtain a set Ψ belonging to characteristic plane:

$\mathrm{\Ψ}=\{\mathrm{\η}={\[{\mathrm{\η}}_1,{\mathrm{\η}}_2\]}^T\∈R^2:{\mathrm{\η}}_1\≥{\stackrel{\‾}{\mathrm{\η}}}_1,{\mathrm{\η}}_2\≤{\stackrel{\‾}{\mathrm{\η}}}_2\},$

Wherein,Represent the meansigma methods of the ridge energy of the relative time-frequency distributions of training sample,Represent the meansigma methods of the ridge total variation of the relative time-frequency distributions of training sample；

(4c) convex closure of training sample set S composition is calculated

Wherein, polygon{ } represent by the convex polygon formed, v_jFor forming the jth summit of convex closure, j=1,2 ..., r, r is the number on the summit of composition convex closure；

(4d) calculate at convex closureIn remove a convex closure vertex v_jAfter convex closure amount of area reduction Δ (v_j):

Wherein, area represents the area of convex closure, and ∩ represents that taking of set ships calculation,Represent from convex closureMiddle deletion convex closure vertex v_jAfter set；

(4e) find out and make convex closure area reduction Δ (v_j) maximum vertex v_*:

$v_{*}=\arg \underset{j=1,2,...,r}{max}\left\{\mathrm{\Δ}\left(v_j\right)\right\},$

WhereinRepresent Δ (v_j) take the value of correspondence j during maximum；

(4f) by above-mentioned vertex v_*Remove from training sample set S, obtain new training sample set S'=S-{v_*}；

(4g) step 4a is repeated) to 4f), altogether remove C=[I × P_F] individual corresponding convex closure summit, finally give the convex closure after contraction and be and meet false-alarm probability P_FDetection decision region Ω, wherein, [I × P_F] represent take I × P_FInteger part.

Step 5, extracts the time-frequency bicharacteristic vector η of unit T to be detected_T。

In training link, after utilizing greedy convex closure learning algorithm to obtain decision region Ω, can detect, during detection, need first unit T to be detected to be extracted the ridge energy η of relative time-frequency distributions₁(T) with the ridge total variation η of relative time-frequency distributions₂(T) both features, construct the relative time-frequency bicharacteristic vector η of unit to be detected_T:

5a) time series of unit T to be detected is: T=[T (1), T (2) ..., T (N)], around unit T to be detected, to choose Q close on unit as reference unit, this reference unit time series is: T_p=[T_p(1),T_p(2),…,T_p(N)], p=1,2 ..., Q；N is seasonal effect in time series length；

5b) to the time series of unit T to be detected and reference unit time sequence T_pCarry out block albefaction, obtain the unit time series to be detected after albefaction and the reference unit time series after albefaction；

5c) utilize the unit time series to be detected after albefaction and the reference unit time series after albefaction, extract the ridge energy η of the relative time-frequency distributions of unit to be detected₁(T) with the ridge total variation η of the relative time-frequency distributions of unit to be detected₂(T) both features, obtain the relative time-frequency bicharacteristic vector of unit T to be detected: η_T=[η₁(T),η₂(T)]^T。

Step 6, characteristic time-frequency relative to the unit to be detected bicharacteristic vector η according to the convex closure constituting detection decision region Ω_TCalculate detection statistic ω.

(6a) the relative time-frequency bicharacteristic vector η of unit to be detected is utilized_TMatrix determinant is calculated with convex closure summit:

$\det \left(\left[\begin{array}{ccc}{\mathrm{\η}}_1\left(T\right)& {\mathrm{\η}}_1\left(T\right)& 1\\ x_j& y_j& 1\\ x_{j+1}& y_{j+1}& 1\end{array}\right]\right),$

Wherein, matrix determinant, x are asked in det () expression_jRepresent the ridge energy of the relative time-frequency distributions on jth convex closure summit, y_jThe ridge total variation of the relative time-frequency distributions on expression jth convex closure summit, j=1,2 ..., r；

(6b) take the minima of above-mentioned matrix determinant, obtain detection statistic ω:

$\mathrm{\ω}=\underset{1\≤k\≤r}{min}\left\{\det \left(\left[\begin{array}{ccc}{\mathrm{\η}}_1\left(T\right)& {\mathrm{\η}}_1\left(T\right)& 1\\ x_j& y_j& 1\\ x_{j+1}& y_{j+1}& 1\end{array}\right]\right)\right\},$

Wherein, min{ } represent take minima.

According to the size of detection statistic ω, step 7, judges whether target exists: if detection statistic ω is more than zero, it was shown that the time-frequency bicharacteristic vector η of unit to be detected_TOutside detection decision region Ω, then judge that target exists, otherwise, it is determined that target is absent from.

Based on step 1 to step 7, it is achieved that based on the detection of the bicharacteristic offshore floating Small object of clutter recognition.

One. experimental data

Data used by this example are the Observed sea clutter that 12 groups of IPIX radars obtain, and radar antenna height is 30m, and pulse recurrence frequency is 1000Hz, and range resolution ratio is 30m；Often group packet is containing four kinds of polarization data, and two of which is same polarization data HH, VV, and two kinds is cross polarization data HV, VH.Wherein having 10 groups of data is the sea clutter data gathered for 93 years, and every kind of polarization data includes 14 distance unit, and data length is 2¹⁷, target is diameter is the ball of 1 meter, and surface tinsel wraps up；Remaining 2 groups of data is the sea clutter data gathered for 98 years, and every kind of polarization data includes 28 distance unit, and data length is 60000, and target is a floating canoe.

Two. emulation experiment

Emulation 1, when observation time is 512ms, utilizes the present invention and based on fractal detection method with based on the detection method of three features, under four kinds of polarization data, Studies of Radar Detection performance is carried out simulation comparison, and result is as shown in Figure 2.Wherein Fig. 2 (a) is the Studies of Radar Detection performance comparison diagram under HH polarization data in the same direction；Fig. 2 (b) is the Studies of Radar Detection performance comparison diagram under VV polarization data in the same direction；Fig. 2 (c) is the Studies of Radar Detection performance comparison diagram under incorgruous HV polarization data；Fig. 2 (d) is the Studies of Radar Detection performance comparison diagram under incorgruous VH polarization data；

Figure it is seen that the detection performance of offshore floating Small object is better than the detection performance of existing two kinds of detection methods by the present invention.

Emulation 2, when observation time is 1024ms, utilizes the present invention and based on fractal detection method with based on the detection method of three features, under four kinds of polarization data, Studies of Radar Detection performance is carried out simulation comparison, and result is as shown in Figure 3.Wherein Fig. 3 (a) is the Studies of Radar Detection performance comparison diagram under HH polarization data in the same direction；Fig. 3 (b) is the Studies of Radar Detection performance comparison diagram under VV polarization data in the same direction；Fig. 3 (c) is the Studies of Radar Detection performance comparison diagram under incorgruous HV polarization data；Fig. 3 (d) is the Studies of Radar Detection performance comparison diagram under incorgruous VH polarization data；

From figure 3, it can be seen that the present invention is compared with existing two kinds of detection methods, the detection performance of offshore floating Small object is better.

Claims (4)

1., based on a bicharacteristic offshore floating small target detecting method for ocean clutter cancellation, comprise the steps:

(1) from echo data, choose training unit, reference unit and unit to be detected:

Utilizing radar transmitter that sea is sent signal, utilize radar receiver to receive the echo data returned by sea surface reflection, this echo data is divided into pure clutter data and the echo data comprising target；

From pure clutter data, selected part distance unit is as one group of training unit, this training unit time series z is: z=[z (1), z (2), ..., z (N)], around training unit, choose Q close on unit as reference unit, this reference unit time series z_pFor: z_p=[z_p(1),z_p(2),…,z_p(N)], p=1,2 ..., Q, Q is number of reference, and N is seasonal effect in time series length；

From the echo data comprising target, selected part distance unit is as unit T to be detected；

(2) to the training unit time series z of pure clutter echo data and reference unit time sequence z_pCarry out block albefaction, obtain the training unit time series after albefactionWith the reference unit time series after albefaction

$\begin{array}{cc}\hat{z}=\[\hat{z}\left(1\right),\hat{z}\left(2\right),...,\hat{z}\left(N\right)\],& {\hat{z}}_p=\[{\hat{z}}_p\left(1\right),{\hat{z}}_p\left(2\right),...,{\hat{z}}_p\left(N\right)\]\end{array};$

(3) the training unit time series after albefaction is utilizedWith the reference unit time series after albefactionExtract the ridge energy η of the relative time-frequency distributions of training unit₁The ridge total variation η of (z) and the relative time-frequency distributions of training unit₂Z () both features, construct relative time-frequency bicharacteristic vector η: the η=[η of pure clutter data₁(z),η₂(z)]^T, []^TRepresent and matrix is carried out transposition；

(4) utilize the relative time-frequency bicharacteristic vector η of pure clutter data, two-dimensional feature space obtains two dimension convex closureAnd at given false-alarm probability P_FUnder, utilize greedy convex closure learning algorithm to two dimension convex closureShrink, and using the convex closure after contraction as detecting decision region Ω；

(5) unit T to be detected is extracted the ridge energy η of relative time-frequency distributions₁(T) with the ridge total variation η of relative time-frequency distributions₂(T) both features, construct the relative time-frequency bicharacteristic vector of unit to be detected: η_T=[η₁(T),η₂(T)]^T；

(6) characteristic time-frequency relative to the unit to be detected bicharacteristic vector η according to the convex closure constituting detection decision region Ω_T=[η₁(T),η₂(T)]^T, calculate detection statistic ω:

$\mathrm{\ω}=\underset{1\≤j\≤r}{min}\left\{\det \left(\left[\begin{array}{ccc}{\mathrm{\η}}_1\left(T\right)& {\mathrm{\η}}_1\left(T\right)& 1\\ x_j& y_j& 1\\ x_{j+1}& y_{j+1}& 1\end{array}\right]\right)\right\},$

Wherein, min{ } represent and take minima, det () represents and seeks matrix determinant, and r is the number on the summit of composition convex closure, x_jRepresent the ridge energy of the relative time-frequency distributions on jth convex closure summit, y_jThe ridge total variation of the relative time-frequency distributions on expression jth convex closure summit, j=1,2 ..., r；

(7) judge whether target exists according to the size of detection statistic ω: if detection statistic ω is more than zero, it was shown that the relative time-frequency bicharacteristic vector η of unit to be detected_TOutside detection decision region Ω, then judge that target exists, otherwise, it is determined that target is absent from.

2. the bicharacteristic offshore floating small target detecting method based on ocean clutter cancellation as claimed in claim 1, it is characterised in that to the training unit time series z of pure clutter echo data and reference unit time sequence z in step (2)_pCarry out block albefaction, carry out as follows:

2a) by training unit time series z and reference unit time series z_pIt is divided into the short amount of the non-overlapping copies that length is M respectively, it may be assumed that

Z=[z₁,z₂,…,z_m,…,z_N/M],

z_p=[z_p,1,z_p,2,…,z_p,m,…,z_p,N/M], p=1,2 ..., Q,

Wherein, z_mRepresent training unit seasonal effect in time series m-th short amount, z_p,mExpression reference unit seasonal effect in time series m-th short amount, m=1,2 ..., N/M；

2b) utilize above-mentioned each short amount, to training unit time series z and reference unit time sequence z_pCarry out block albefaction, obtain the training unit time series after albefactionWith the reference unit time series after albefaction

$\hat{z}=\[\hat{z}\left(1\right),\hat{z}\left(2\right),...,\hat{z}\left(N\right)\]=\[{\hat{z}}_1,{\hat{z}}_2,\mathrm{...},{\hat{z}}_m,\mathrm{...},{\hat{z}}_{N/M}\],$

${\hat{z}}_p=\[{\hat{z}}_p\left(1\right),{\hat{z}}_p\left(2\right),...,{\hat{z}}_p\left(N\right)\]=\[{\hat{z}}_{p,2},{\hat{z}}_{p,2},\mathrm{...},{\hat{z}}_{p,m},\mathrm{...},{\hat{z}}_{p,N/M}\],$

Wherein,Represent the training unit seasonal effect in time series m-th short amount after albefaction,Represent the reference unit seasonal effect in time series m-th short amount after albefaction.

3. the bicharacteristic offshore floating small target detecting method based on ocean clutter cancellation as claimed in claim 1, it is characterised in that extract the ridge energy η of the relative time-frequency distributions of training unit in step (3)₁The ridge total variation η of (z) and the relative time-frequency distributions of training unit₂Z () both features, construct the relative time-frequency bicharacteristic vector η of pure clutter data, carry out as follows:

3a) calculate the smooth Eugene Wigner-Willie distribution of training unitWith the smooth Eugene Wigner of reference unit-Willie distribution

$W(n,l,\hat{z})=\underset{m=-E}{\overset{E}{\Σ}}g\left(m\right)\underset{k=-F}{\overset{F}{\Σ}}h\left(k\right)\hat{z}(n+m+k){\hat{z}}^{*}(n+m-k)\exp (-4{\mathrm{j\πkl\Δf}}_d),$

$D(n,l,{\hat{z}}_p)=\underset{m=-E}{\overset{E}{\Σ}}g\left(m\right)\underset{k=-F}{\overset{F}{\Σ}}h\left(k\right){\hat{z}}_p(n+m+k){{\hat{z}}_p}^{*}(n+m-k)\exp (-4{\mathrm{j\πkl\Δf}}_d),$

Wherein, subscript * represents conjugation, and g (m) is time smoothing window, and h (k) is Frequency Smooth window, and the half of E express time smoothing windows length, F represents the half of Frequency Smooth window length, Δ f_dFor the sampling interval of normalization Doppler frequency, n=1,2 ..., N, l=1,2 ..., N；

3b) utilize the smooth Eugene Wigner-Willie distribution of reference unitEstimate training unit average time-frequency distributions A (n, l):

$A(n,l)=\frac{1}{9Q}\underset{p=1}{\overset{Q}{\Σ}}\underset{n^{\′}=-1}{\overset{1}{\Σ}}\underset{l^{\′}=-1}{\overset{1}{\Σ}}D(n+n^{\′},l+l^{\′},{\hat{z}}_p),$

Wherein, the value of n ' is-1,0,1, and the value of l ' is-1,0,1；

3c) utilize the smooth Eugene Wigner-Willie distribution of training unit(n l), calculates the relative time-frequency distributions of training unit with the average time-frequency distributions A of training unit

$B(n,l,\hat{z})=\frac{W(n,l,\hat{z})}{A(n,l)};$

3d) calculate the relative time-frequency distributions of training unitTime-frequency ridge R (n):

$R\left(n\right)=\underset{l}{\mathrm{argmax}}\left\{B(n,l,\hat{z})\right\},$

Wherein,RepresentTake the value of correspondence l during maximum；

3e) utilize the relative time-frequency distributions of training unitRelative time-frequency distributions with training unitTime-frequency ridge R (n), calculate the ridge energy η of relative time-frequency distributions of training unit₁The ridge total variation η of (z) and the relative time-frequency distributions of training unit₂(z):

${\mathrm{\η}}_1\left(z\right)=\underset{n=1}{\overset{N}{\Σ}}B(n,R(n),\hat{z}),$

${\mathrm{\η}}_1\left(z\right)={\mathrm{\Δf}}_d\underset{n=2}{\overset{N}{\Σ}}|R(n+1)-R\left(n\right)|;$

3f) utilize the ridge energy η of the relative time-frequency distributions of training unit₁The ridge total variation η of (z) and the relative time-frequency distributions of training unit₂Z (), constructs the relative time-frequency bicharacteristic vector η of pure clutter data:

η=[η₁(z),η₂(z)]^T,

Wherein, []^TRepresent and matrix is carried out transposition.

4. the bicharacteristic offshore floating small target detecting method based on ocean clutter cancellation as claimed in claim 1, it is characterised in that at given false-alarm probability P in step (4)_FUnder, utilize greedy convex closure learning algorithm to two dimension convex closureShrink, carry out as follows:

4a) utilize the relative time-frequency bicharacteristic vector η of pure clutter data, form training sample set S:

$S=\{{\mathrm{\η}}_i={\[{\mathrm{\η}}_1^i\left(z\right),{\mathrm{\η}}_2^i\left(z\right)\]}^T:i=1,2,...,I\},$

Wherein, η_iRepresent the relative time-frequency bicharacteristic vector of i-th training sample,Represent the ridge energy of the relative time-frequency distributions of i-th training sample,Representing the ridge total variation of the relative time-frequency distributions of i-th training sample, I is training sample number；

4b) utilize training sample set S, obtain a set Ψ belonging to characteristic plane:

$\mathrm{\Ψ}=\{\mathrm{\η}={\[{\mathrm{\η}}_1,{\mathrm{\η}}_2\]}^T\∈R^2:{\mathrm{\η}}_1\≥\stackrel{\‾}{{\mathrm{\η}}_1},{\mathrm{\η}}_2\≤\stackrel{\‾}{{\mathrm{\η}}_2}\};$

Wherein,Represent the meansigma methods of the ridge energy of the relative time-frequency distributions of training sample,Represent the meansigma methods of the ridge total variation of the relative time-frequency distributions of training sample；

4c) calculate the convex closure of training sample set S composition

Wherein, polygon{ } represent by the convex polygon formed, v_jFor forming the jth summit of convex closure, j=1,2 ..., r, r is the number on the summit of composition convex closure；

4d) calculate at convex closureIn remove a convex closure vertex v_jAfter convex closure amount of area reduction Δ (v_j):

Wherein, area represents the area of convex closure, and ∩ represents that taking of set ships calculation,Represent from convex closureMiddle deletion convex closure vertex v_jAfter set；

4e) find out and make convex closure area reduction Δ (v_j) maximum vertex v_*:

$v_{*}=\arg \underset{j=1,2,...,r}{max}\left\{\mathrm{\Δ}\left(v_j\right)\right\},$

WhereinRepresent Δ (v_j) take the value of correspondence j during maximum；

4f) by above-mentioned vertex v_*Remove from training sample set S, obtain new training sample set S'=S-{v_*}；

4g) repeat step 4a) to 4f), altogether remove C=[I × P_F] individual corresponding convex closure summit, finally give the convex closure after contraction and be and meet false-alarm probability P_FDetection decision region Ω, wherein, [I × P_F] represent take I × P_FInteger part.