CN106022280A - Typical ship target identification method based on graded invariance features - Google Patents

Abstract

The invention provides a typical ship target identification method based on graded invariance features. At first, the binary entropy and the normalized inertia moment of each image ship target are extracted as the primary feature; and then each image is subjected to wavelet decomposition to form four sub-images, and the weighted Hu moment, Zernike moment, and Fourier descriptor of the each sub-image ship target are extracted to be the secondary feature; the polar coordinate shape matrix of each image ship target is taken as the trinary feature; and all of the features are modified to have the properties of translation, rotation, and scaling invariance. The experimental result of a recognition classifier shows that the algorithm can describe the typical ship targets in satellite remote sensing images in details step by step, and the recognition accuracy is high. The method can be applied to the typical ship target recognition of the satellite remote sensing image database, and is an engineering method that is high in universality.

Description

A kind of typical ship seakeeping method based on classification Invariance feature

Technical field

The present invention relates to a kind of satellite remote sensing images typical case's ship seakeeping method, particularly a kind of based on The typical ship seakeeping method of classification Invariance feature, belongs to space remote sensing field.

Background technology

In the case of China's maritime safety interests increase rapidly, Optical remote satellite can to the earth on a large scale Region is observed, and energy accurately perception also obtains marine information, provides decision support in time, contributes to fast Speed solves ocean accident.By remote sensing satellite, ShipTargets is identified in-orbit, can be quickly Obtain the information such as the position of Ship Target, type, the application demand that targets in ocean is monitored by user can be met.

Shape is the principal character that typical case's Ship Target is detected and identified.Typical case's Ship Target is being defended Style characteristic in star chart picture is the most overall inseparable but also there is nuance.I.e. can not be partitioned into naval vessel Each composition primitive of target mates；But different types of target difference the most again is the least, Type identification only just correspond to the portion of playing a game have nuance target classify.The classical spy of this kind of target Levy extraction algorithm to have: Hu square, Zernike square and Fourier describe son etc..Hu square has translation, rotates With scaling invariance.But Hu square is global characteristics, and do not draw from Orthogonal Function Set, exist very Many redundancies, to local have nuance target be difficult to.The coefficient of Zernike square is orthogonal, Superior performance in terms of information redundancy and image reconstruction capabilities.But Zernike square remains global characteristics, base In such feature be not easy to local have nuance target be identified.Fourier describes son can Characterize the closed outline of target, but it be that the whole border to shape carries out distance measure that Fourier describes son, Therefore cause to local have nuance its identification ability of Ship Target the highest.

Making a general survey of existing classical feature extraction algorithm, advantage is to have solid theoretical basis, stable performance, But it is limited in one's ability to the shape characterization of target, it is impossible to reflect local detail exactly.In satellite image Typical case's Ship Target pixel count is little, particular type to be identified, in target shape, each details is to identifying Having contribution, these feature extraction algorithms can not meet the actual requirement of typical case's Ship Target type identification.

Therefore, each details in target shape during the typical Ship Target type identification in satellite image Very important, clarification of objective extraction algorithm must can describe the global property of target can take into account again it Local detail information.

Summary of the invention

Present invention solves the technical problem that and be: overcome the deficiencies in the prior art, it is provided that one is based on classification not The typical ship seakeeping method of Vertic features, solves satellite remote sensing images typical case's Ship Target in-orbit Identification problem, meets satellite remote sensing images typical case's Ship Target type identification in-orbit to a certain extent Demand.

The technical scheme is that a kind of typical ship seakeeping side based on classification Invariance feature Method, step is as follows:

(1) the every width satellite remote sensing images to input carries out adaptive-filtering, maximum between-cluster variance segmentation Normalization with correction skew processes；

(2) the two-value entropy of every width satellite remote sensing images Ship Target and normalization rotary inertia are extracted as one Level feature；

(3) bianry image of every width satellite remote sensing images being carried out one-level wavelet decomposition, image is broken down into Tetra-sub-band images of LL, LH, HL, HH, are respectively described image in level, vertical and diagonal side Details upwards；

(4) describe son with Hu square, Zernike square and Fourier and respectively all sub-band images are carried out spy Levy extraction, and be weighted processing to gained characteristic vector according to the descriptive power difference of each sub-band；By Hu square after wavelet decomposition and weighting process, Zernike square, Fourier describe son order and rearrange comprehensive Conjunction characteristic vector is as secondary characteristics, and ensures that it has translation, rotates and scale invariance；

(5) calculate three grades of features of acquisition, and represent with polar coordinate form matrix；

(6) utilize support vector machine and form matrix Similarity Measure that the feature extracted is identified Classification；Wherein, form matrix calculating formula of similarity is as follows:

$sim=1-\frac{dif}{tot}$

Wherein, sim is similar to the form matrix corresponding to " Ship Target to be identified " for " template Ship Target " Degree；Dif is the most right with the form matrix corresponding to " Ship Target to be identified " for " template Ship Target " Answer the absolute value sum of element difference；Tot is " template Ship Target " form matrix all elements value Summation；The identification to typical case's Ship Target is completed by form matrix similarity threshold T=0.8.

In step (2), the two-value entropy of every width satellite remote sensing images Ship Target is:

$H_b=log\frac{n}{MN}$

Wherein, n is the total pixel number that Ship Target comprises, and the size of every width satellite remote sensing images is M × N；

The normalization rotary inertia of every width satellite remote sensing images Ship Target is:

$NMI=\frac{\sqrt{\underset{y=1}{\overset{N}{\Σ}}\underset{x=1}{\overset{M}{\Σ}}\[{(x-C_x)}^2+{(y-C_y)}^2\]}}{MN};$

Wherein, the binary map of every width satellite remote sensing images be f (x, y), C_xAnd C_yBarycenter for bianry image.

Step (4) ensures have translation, rotate and scale invariance method particularly includes:

(4a) Hu moment preserving square itself has translation, rotates and scale invariance, it is not necessary to process；

(4b) following measures is taked to make Zernike square have translation, rotate and scale invariance: in calculating The origin translation of unit circle to Ship Target barycenter, calculate square with target centroid, to obtain translation not Degeneration；Each Zernike square is taken its amplitude to obtain rotational invariance；To each Zernike square It is normalized to obtain scaling invariance with following formula:

$\begin{array}{ccc}\left|\right|Z_{mn}|{|}^{\′}=\frac{\left|\right|Z_{mn}\left|\right|}{\left|\right|Z_{00}\left|\right|}& m=2,3,...& n=0,1,2,3,\end{array}...$

(4c) take following measures to make Fourier describe son to there is translation, rotate and scale invariance: go Fall Fourier and describe the Section 1 of son to obtain translation invariance；Owing to Ship Target profile is by the inverse time Pin is tracked, and Ship Target is non-zero region, and second order describes sub-z (1) ≠ 0, then make z (k)=z (k)/z (1) Not property is scaled to obtain, k=2,3,4 ..., K-1, K are equally distributed on Ship Target profile adopting Sampling point；Each item takes its amplitude to obtain rotational invariance；

In step (5), the construction algorithm of polar coordinate form matrix is as follows；

5a) every width satellite remote sensing images of input is carried out adaptive median filter, segmentation and binaryzation, Calculate the barycenter C=(C trying to achieve Ship Target_x, C_y)；

5b) ask for the Euclidean distance from naval vessel barycenter point furthest M to naval vessel barycenter C on Ship Target R (M, C), and to define r (M, C) be the greatest radius of Ship Target；Definition B is the form matrix of m × n；

5c) from the beginning of barycenter C, r (M, C) is divided into isometric n-1 section；

5d) with C as the center of circle, respectively with 0, r/n-1,2r/n-1 ..., (n-1) r/n-1 is that radius draws circle；

5e) start from r (M, C), along clockwise direction, each circle is divided into isometric m section arc, often The angle of section arc is 360 °/m；

If 5f) polar coordinate [ir/ (n-1)；(j 360 °)/m] point belong to Ship Target, then B (i, j)=1；Instead It, and B (i, j)=0；Wherein i, j are the element in matrix B；Finally give polar coordinate form matrix B.

Present invention advantage compared with prior art is:

(1) present invention be directly based upon shape satellite remote sensing images is carried out typical case Ship Target feature extraction, Whole method is simple and is prone to hardware realization；

(2) the graded features extraction algorithm that the present invention proposes, the descriptive power of algorithm improves step by step.One Level feature only carries out thick hierarchy description to target shape, can be used for the rough sort of typical case's Ship Target collection.Its Advantage is that calculating is simple, is prone to extract.Secondary characteristics descriptive power significantly improves, and can be used for satellite mapping The typical Ship Target less as style differences in storehouse is identified.Three grades of features have the strongest description energy Power, can be used as the typical case image registration of Ship Target, image rectification, target recognition etc..

(3) present invention introduces wavelet transformation as the micro-instrument of details, it is possible to the most micro-description typical case Ship Target image is in level, vertical and diagonally adjacent shape details；

(4) by choosing the imaging data of different satellite, to the remote sensing images acquired in polytypic satellite Carrying out experiment test, result shows: this method all can the allusion quotation in identification satellite remote sensing images rapidly and accurately Type Ship Target, is the engineering method that a kind of universality is the strongest.

Accompanying drawing explanation

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

Fig. 2 is that image wavelet decomposition data transmits block diagram；Fig. 3 is form matrix Similarity Measure block diagram.

Detailed description of the invention

Below in conjunction with accompanying drawing and implement example so that implementing of the present invention to be described:

The image in the satellite ship images storehouse of the present invention all derives from remote sensing satellite image storehouse.Image is all Jpeg format, gray level is 256.

Fig. 1 is the flow chart of the present invention, from fig. 1, it can be seen that the one that the present invention provides is constant based on classification The typical ship seakeeping method of property feature, it is characterised in that step is as follows:

(1) to input every width satellite remote sensing images carry out adaptive-filtering, maximum between-cluster variance segmentation and The normalization of correction skew processes.

(2) global characteristics that every width satellite remote sensing images carries out Ship Target extracts, by two-value entropy and returning One changes rotary inertia order rearranges multi-feature vector as one-level feature, and it is flat to ensure that it has Move, rotate and scale invariance.

(2a) entropy is the uncertainty measure of event probability of occurrence in theory of information, can effectively reflect event The information content comprised.Therefore, the entropy of piece image is exactly measuring of its comprised information content.Counting Local entropy is defined as follows by word image

Wherein, W × V is the image local window size comprising target, f (i, j) be in image local window (i, j) The grey scale pixel value at place, n is the total pixel number of target in image local window.Local entropy reflects a width The dispersion degree of pixel in image, local entropy is the biggest, illustrates that its pixel distribution is the most uniform.

Local entropy is introduced in bianry image, calculates the entropy of bianry image, referred to as two-value entropy by following two formulas.

$\begin{array}{cc}{p}_i=\left\{\begin{array}{cc}0,& f(i,j)=0\\ \frac{1}{n},& f(i,j)=1\end{array}\right.& H_t=-\underset{i=1}{\overset{n}{\Σ}}{p}_i\log p_i=-\underset{i=1}{\overset{n}{\Σ}}\frac{1}{n}log\frac{1}{n}=\log n\end{array}$

Wherein, n is the total pixel number that target area comprises, and image size is M × N.

Two-value entropy is translated, rotates and scales invariance analysis.It can be seen that two-value entropy only with The pixel count that target comprises is relevant, therefore has translation, rotational invariance.For obtaining scaling invariance, repair It it is being just following formula.

$H_b=log\frac{n}{MN}$

The span of two-value entropy is H_b∈ [-∞, 0], when in image without Ship Target in the presence of H_b=-∞, works as figure H during as being all Ship Target_b=0.

(2b) normalization rotary inertia (NMI, Normalized Moment of Inertia) is by two-dimensional digital ash Degree image regards the M on two dimensional surface × N number of pixel as, according to certain method, it is carried out two-value Change, obtain a width bianry image f (x, y).Then

The quality of bianry image is

$m=\underset{y=1}{\overset{N}{\Σ}}\underset{x=1}{\overset{M}{\Σ}}f(x,y)$

The barycenter of bianry image

$C_x=\frac{\underset{y=1}{\overset{N}{\Σ}}\underset{x=1}{\overset{M}{\Σ}}xf(x,y)}{\underset{y=1}{\overset{N}{\Σ}}\underset{x=1}{\overset{M}{\Σ}}f(x,y)}=\frac{\underset{x=1}{\overset{M}{\Σ}}x}{MN}$

$C_y=\frac{\underset{y=1}{\overset{N}{\Σ}}\underset{x=1}{\overset{M}{\Σ}}yf(x,y)}{\underset{y=1}{\overset{N}{\Σ}}\underset{x=1}{\overset{M}{\Σ}}f(x,y)}=\frac{\underset{y=1}{\overset{M}{\Σ}}y}{MN}$

The rotary inertia of bianry image

$J_{(C_x,C_y)}=\underset{y=1}{\overset{N}{\Σ}}\underset{x=1}{\overset{M}{\Σ}}\[{(x-C_x)}^2+{(y-C_y)}^2\]f(x,y)$

The normalization rotary inertia of the region bianry image obtained by gray scale threshold method

From its mechanism, normalization rotary inertia is translated, rotates and scales invariance analysis below.For flat Moving distortion, the coordinate of each pixel in target is x '=x+ △ x, y '=y+ △ y, but target after translation The center-of-mass coordinate of shape is C_x'=C_x+ △ x, C_y'=C_y+△y；So in target shape, each point arrives after Ping Yi The distance of barycenter does not change, and target area size M × N is also not changed in, therefore normalizing Change rotary inertia constant.The impact of its respective binary image is considered as first to mesh by the rotational distortion of target Mark region translates, and rotates further around its barycenter.In this case, target area does not occur Change, in target, each point does not the most change to the distance of barycenter, therefore normalization rotary inertia is constant. If the proportionality coefficient of scaling is k, then x '=kx, y '=ky；Image centroid C after scale transformation_x'=kC_x, C_y'=kC_y.Gray scale threshold method m '=k²As can be seen from the above equation, normalization rotary inertia is constant for m.

(3) image is carried out one-level wavelet decomposition, in resolution 2^jUnder (j≤0), image is projected to space V_jAnd O_jIn, it is as follows that image is broken down into four sub-band images.

${\mathrm{LL}}_{2^j}f=\{<f(x,y),g_{2^j}(x-2^{-j}n)g_{2^j}(y-2^{-j}m)>\}$

${\mathrm{LH}}_{2^j}f=\{<f(x,y),g_{2^j}(x-2^{-j}n)h_{2^j}(y-2^{-j}m)>\}$

${\mathrm{HL}}_{2^j}f=\{<f(x,y),h_{2^j}(x-2^{-j}n)g_{2^j}(y-2^{-j}m)>\}$

${\mathrm{HH}}_{2^j}f=\{<f(x,y),h_{2^j}(x-2^{-j}n)h_{2^j}(y-2^{-j}m)>\}$

N and m is integer,<..>represent scalar product computing,Yardstick Function g regards low pass filter as, and wavelet function h regards high pass filter as, then to every grade of wavelet decomposition,For the sub-band image of low pass filter output, it maintains the half small echo of wave filter input picture Coefficient, for input picture 2^jAn approximation under (j≤0) resolution；For high pass The output of wave filter, withSize is identical, for input picture 2^jDown-sampling figure under (j≤0) resolution Picture, but it is respectively described image in level, vertical and diagonally adjacent details.

(4) describe son with Hu square, Zernike square and Fourier and respectively all sub-band images are carried out spy Levy extraction, and be weighted processing to gained characteristic vector according to the descriptive power difference of each sub-band；By Hu square after wavelet decomposition and weighting process, Zernike square, Fourier describe son order and rearrange comprehensive Conjunction characteristic vector is as secondary characteristics, and ensures that it has translation, rotates and scale invariance.

(4a) seven Hu not bending moments itself are about the invariant translating, rotating and scale, as follows； Assume that R is the target represented with bianry image, then pth+q rank the central moment of R shape is following formula.

${\mathrm{\&mu;}}_{p+q}=\underset{(x,y)\&Element;R}{\&Sigma;}{(x-x_c)}^p{(y-y_c)}^q$

Wherein, (x_c, y_c) it is the barycenter of target.For obtaining the character unrelated to scaling, can be to this central moment It is standardized operation, such as following formula.

${\mathrm{\&eta;}}_{p+q}=\frac{{\mathrm{\&mu;}}_{p+q}}{{\mathrm{\&mu;}}^{{\mathrm{\&gamma;}}_{0,0}}};\mathrm{\&gamma;}=\frac{p+q}{2}+1;p+q=2,3,\mathrm{...}$

Based on these squares, there is translation, rotate and scale seven moment preserving formulas of invariance as follows

(4b) digital picture n rank Zernike square is defined as follows.

$Z_{nm}=\frac{n+1}{\mathrm{\&pi;}}\underset{y}{\&Sigma;}\underset{x}{\&Sigma;}f(x,y){\&lsqb;V_{nm}(x,y)\&rsqb;}^{*}$

Zernike square essence is a kind of mapping, is transformed to by image function on one group of orthogonal basis function.Image f (x, Y) Zernike square is that this image is at one group of orthogonal polynomial { V_nm(x, y) } on projection.Here what is called is just Hand over, refer to orthogonal polynomial { V_nm(x, y) } at unit circle { x²+y²Following condition is met on≤1}.

$V_{nm}(x,y)=R_{nm}(x,y)e^{jm\arctan \frac{y}{x}}$

In formula, n is positive integer or zero；M is just or negative integer, and meets n-m=even number, ∣m∣≤n；R_nm(x y) is radial polynomial, such as following formula

$R_{nm}(x,y)=\underset{s=0}{\overset{\left(n-\left|m\right|\right)/2}{\&Sigma;}}{(-1)}^s\frac{(n-s)!}{s!(\frac{n+\left|m\right|}{2}-s)!(\frac{n-\left|m\right|}{2}-s)!}{(x^2+y^2)}^{(\frac{n}{2}-s)}$

Following measures is taked to make Zernike square have translation, rotate and scale invariance: unit in calculating The origin translation of circle, to Ship Target barycenter, calculates square with target centroid, to obtain translation invariance； Each Zernike square is taken its amplitude to obtain rotational invariance；To each Zernike square following formula It is normalized to obtain scaling invariance:

$\begin{array}{ccc}\left|\right|Z_{mn}|{|}^{\&prime;}=\frac{\left|\right|Z_{mn}\left|\right|}{\left|\right|Z_{00}\left|\right|}& m=2,3,...& n=0,1,2,3\end{array},...$

(4c) border of the target shape in supposition 2-D cartesian coordinate system is S, then by side counterclockwise To being tracked and resampling, it is possible to obtain equally distributed K point.The coordinate of each point can be as follows Represent: (x₀, y₀), (x₁, y₁) ..., (x_k-1, y_k-1).These coordinates can be expressed as x (i)=x_i, Y (i)=y_iWherein, i=0,1,2 ..., k-1.In this case, border can be expressed as a series of Plural number: s (i)=x (i)+j y (i), wherein, i=0,1,2 ..., k-1.X-axis is made as real axis, y-axis For the imaginary axis.Then the Discrete Fourier Transform of sequence of complex numbers s (i) is shown below.Wherein, u=0,1,2 ..., K-1.Complex coefficient z (u) is referred to as the Fourier on border and describes son.

$z\left(u\right)=\frac{1}{K}\underset{i=0}{\overset{K-1}{\&Sigma;}}s\left(i\right)e^{-j2\mathrm{\&pi;}ui/K};$

Fourier describes son itself not to be had translation, rotates and scale invariance, in order to obtain to starting point, Translation, rotating and scale constant feature set, uses following normalization process: affecting the because translating One, Section 1 of dieing is to obtain translation invariance；Because profile is by being tracked counterclockwise, warship Ship target is non-zero region, and second order describes sub-a (1) ≠ 0, therefore with a (k)=a (k)/a (1) to obtain scaling not Property；Each item takes its amplitude a (k) to obtain rotational invariance.

(5) calculate three grades of features of acquisition, and represent with polar coordinate form matrix；Polar coordinate form matrix Construction algorithm is as follows；

5a) every width satellite remote sensing images of input is carried out adaptive median filter, segmentation and binaryzation, Calculate the barycenter C=(C trying to achieve Ship Target_x, C_y)；

5b) ask for the Euclidean distance from naval vessel barycenter point furthest M to naval vessel barycenter C on Ship Target R (M, C), and to define r (M, C) be the greatest radius of Ship Target；Definition B is the form matrix of m × n；

5c) from the beginning of barycenter C, r (M, C) is divided into isometric n-1 section；

5d) with C as the center of circle, respectively with 0, r/n-1,2r/n-1 ..., (n-1) r/n-1 is that radius draws circle；

5e) start from r (M, C), along clockwise direction, each circle is divided into isometric m section arc, often The angle of section arc is 360 °/m；

If 5f) polar coordinate [ir/ (n-1)；(j 360 °)/m] point belong to Ship Target, then B (i, j)=1；Instead It, and B (i, j)=0；Wherein i, j are the element in matrix B；Finally give polar coordinate form matrix B；

(6) utilize support vector machine and form matrix Similarity Measure that the feature extracted is identified Classification；Wherein, form matrix calculating formula of similarity is as follows:

$sim=1-\frac{dif}{tot};$

Wherein, sim is similar to the form matrix corresponding to " Ship Target to be identified " for " template Ship Target " Degree；Dif is the most right with the form matrix corresponding to " Ship Target to be identified " for " template Ship Target " Answer the absolute value sum of element difference；Tot is " template Ship Target " form matrix all elements value Summation；The identification of typical case's Ship Target is can be realized as by form matrix similarity threshold T=0.8.

The typical ship seakeeping that the inventive method is mainly used in satellite remote sensing images, it is possible to realize fast Speed accurately identification function.Can be applicable to the multinomial of high-resolution earth observation systems key special subjects subsequent satellites Engineering, has broad application prospects.The method that the present invention proposes passes in all satellite remote sensing images data Communication system can use.

The unspecified part of the present invention belongs to general knowledge as well known to those skilled in the art.

Claims (4)

1. a typical ship seakeeping method based on classification Invariance feature, it is characterised in that step is as follows:

(1) the every width satellite remote sensing images to input carries out the normalization process of adaptive-filtering, maximum between-cluster variance segmentation and correction skew；

(2) the two-value entropy of every width satellite remote sensing images Ship Target and normalization rotary inertia are extracted as one-level feature；

(3) bianry image of every width satellite remote sensing images being carried out one-level wavelet decomposition, image is broken down into tetra-sub-band images of LL, LH, HL, HH, is respectively described image in level, vertical and diagonally adjacent details；

(4) describe son with Hu square, Zernike square and Fourier and respectively all sub-band images are carried out feature extraction, and be weighted processing to gained characteristic vector according to the descriptive power difference of each sub-band；By the Hu square after wavelet decomposition and weighting process, Zernike square, the sub multi-feature vector that sequentially rearranges of Fourier description is as secondary characteristics, and ensures that it has translation, rotates and scale invariance；

(5) calculate three grades of features of acquisition, and represent with polar coordinate form matrix；

(6) utilize support vector machine and form matrix Similarity Measure that the feature extracted is identified classification；Wherein, form matrix calculating formula of similarity is as follows:

Wherein, sim is " template Ship Target " and the form matrix similarity corresponding to " Ship Target to be identified "；Dif is the absolute value sum of " template Ship Target " with the form matrix whole corresponding element difference corresponding to " Ship Target to be identified "；Tot is the summation of " template Ship Target " form matrix all elements value；The identification to typical case's Ship Target is completed by form matrix similarity threshold T=0.8.

A kind of typical ship seakeeping method based on classification Invariance feature the most according to claim 1, it is characterised in that: in step (2), the two-value entropy of every width satellite remote sensing images Ship Target is:

Wherein, n is the total pixel number that Ship Target comprises, and the size of every width satellite remote sensing images is M × N；

The normalization rotary inertia of every width satellite remote sensing images Ship Target is:

Wherein, the binary map of every width satellite remote sensing images be f (x, y), C_xAnd C_yBarycenter for bianry image.

A kind of typical ship seakeeping method based on classification Invariance feature the most according to claim 1, it is characterised in that: step (4) ensures have translation, rotate and scale invariance method particularly includes:

(4a) Hu moment preserving square itself has translation, rotates and scale invariance, it is not necessary to process；

(4b) following measures is taked to make Zernike square have translation, rotate and scale invariance: the origin translation of unit circle to Ship Target barycenter in calculating, to calculate square with target centroid, to obtain translation invariance；Each Zernike square is taken its amplitude to obtain rotational invariance；It is normalized each Zernike square following formula to obtain scaling invariance:

(4c) take following measures to make Fourier describe son to there is translation, rotate and scale invariance: remove Fourier and describe the Section 1 of son to obtain translation invariance；Owing to Ship Target profile is by being tracked counterclockwise, Ship Target is non-zero region, second order describes sub-z (1) ≠ 0, z (k)=z (k)/z (1) is then made to scale not property to obtain, k=2,3,4, ..., K-1, K are equally distributed sampled point on Ship Target profile；Each item takes its amplitude to obtain rotational invariance.

A kind of typical ship seakeeping method based on classification Invariance feature the most according to claim 1, it is characterised in that: in step (5), the construction algorithm of polar coordinate form matrix is as follows；

5a) the every width satellite remote sensing images to input carries out adaptive median filter, segmentation and binaryzation, calculates the barycenter C=(C trying to achieve Ship Target_x, C_y)；

5b) ask for Euclidean distance r (M, C) from naval vessel barycenter point furthest M to naval vessel barycenter C on Ship Target, and to define r (M, C) be the greatest radius of Ship Target；Definition B is the form matrix of m × n；

5c) from the beginning of barycenter C, r (M, C) is divided into isometric n-1 section；

5d) with C as the center of circle, respectively with 0, r/n-1,2r/n-1 ..., (n-1) r/n-1 is that radius draws circle；

5e) starting from r (M, C), along clockwise direction, each circle is divided into isometric m section arc, the angle of every section of arc is 360 °/m；

If 5f) polar coordinate [ir/ (n-1)；(j 360 °)/m] point belong to Ship Target, then B (i, j)=1；Otherwise, B (i, j)=0；Wherein i, j are the element in matrix B；Finally give polar coordinate form matrix B.