CN108681732A - A kind of gunz optimization SAR radar airbound target identifying systems - Google Patents

A kind of gunz optimization SAR radar airbound target identifying systems Download PDF

Info

Publication number
CN108681732A
CN108681732A CN201810457315.7A CN201810457315A CN108681732A CN 108681732 A CN108681732 A CN 108681732A CN 201810457315 A CN201810457315 A CN 201810457315A CN 108681732 A CN108681732 A CN 108681732A
Authority
CN
China
Prior art keywords
feature
module
target
solution
sar
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN201810457315.7A
Other languages
Chinese (zh)
Inventor
刘兴高
吴俊�
孙元萌
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Zhejiang University ZJU
Original Assignee
Zhejiang University ZJU
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Zhejiang University ZJU filed Critical Zhejiang University ZJU
Priority to CN201810457315.7A priority Critical patent/CN108681732A/en
Publication of CN108681732A publication Critical patent/CN108681732A/en
Pending legal-status Critical Current

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/20—Image preprocessing
    • G06V10/255—Detecting or recognising potential candidate objects based on visual cues, e.g. shapes
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • General Physics & Mathematics (AREA)
  • Artificial Intelligence (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Biology (AREA)
  • Evolutionary Computation (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • General Engineering & Computer Science (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Multimedia (AREA)
  • Radar Systems Or Details Thereof (AREA)

Abstract

The invention discloses a kind of gunzs to optimize SAR radar airbound target identifying systems, including SAR radars, database and host computer；SAR radars, database, host computer are sequentially connected, the SAR radars to being monitored in real time in the air, and in the image data storage to the database for obtaining SAR radars, the host computer includes image pre-processing module, characteristic extracting module, feature selection module, classifier training module, gunz optimization module and result display module.A kind of realization online recognition of present invention offer, airflight target identification system with high accuracy.

Description

A kind of gunz optimization SAR radar airbound target identifying systems
Technical field
The present invention relates to radar data process fields, particularly, are related to a kind of gunz optimization SAR radar airbound targets Identifying system.
Background technology
Airflight target identification technology has shown powerful vitality, occurs various theoretical methods over 30 years, The algorithm of the comparative maturity of research has classical statistics judgement, Subjective Bayes deduction, D-S evidence theory method, neural network and support Vector machine etc..Because the difference of the algorithm of target identification is possible to that the inconsistent of expression recognition result can be caused, system is It is convenient for handling and comparing that the expression-form of all kinds of recognition results is unified for subjective probability.In addition to this, some grind When studying carefully personnel and studying target identification, rough set theory and data mining are introduced.Also someone by data mining with it is coarse Collection is theoretical to be combined, and is studied target identification system.Currently, inductive learning process is widely used in data mining technology, number Potential rule can be extracted from mass data according to excavating.Wherein airflight target is monitored using SAR image and Identification is a current international forward position and hot spot, can be identified by carrying out the monitoring of airflight target to SAR image, Obtain the important informations parameter such as type, position and the course of airbound target.For obtain airflight target initiative, It is vital to ensure that the success of airflight goal activity plays the role of.
Invention content
In order to overcome the shortcomings of currently based on not high, the of the invention mesh of the airflight target identification accuracy rate of SAR image Be provide it is a kind of realize analyze in real time gunz optimization SAR radar airbound target identifying systems.
The technical solution adopted by the present invention to solve the technical problems is：A kind of gunz optimization SAR radar flight mesh Identifying system, including SAR radars, database and host computer are marked, SAR radars, database and host computer are sequentially connected, described SAR radars to being monitored in real time in the air, and the image data that SAR radars are obtained is stored into the database, described Host computer include：
Image pre-processing module is completed to carry out SAR radar image data pretreatments using following process：
1) the SAR image gray level transmitted from database is L, f (x0,y0) it is pixel (x0,y0) at gray value, g (x0,y0) it is pixel (x0,y0) N × N neighborhoods in pixel average value, wherein x0,y0The abscissa of pixel is indicated respectively And ordinate；
2) the number of pixels h (m, n) for meeting f=m and g=n by calculating, obtains joint probability density pmn：
pmn=p (m, n)=h (m, n)/M
Wherein, M indicates the total number of image pixel；
3) the mean vector μ of two-dimensional histogram is calculated：
4) probability P that target and background occurs in image is calculated separately0,1With mean vector μ0,1：
Wherein, t, s, subscript 0, subscript 1 indicate f segmentation thresholds, g segmentation thresholds, target area, background area respectively；
5) inter-class variance BCV is calculated：
BCV=P0(μ0-μ)(μ0-μ)′+P1(μ1-μ)(μ1-μ)′；
Wherein, μ indicates mean vector, the transposition of subscript ' representing matrix.
6) optimal threshold is the Two Dimensional Thresholding vector [s so that when BCV is maximum value0,t0]：
Characteristic extracting module is completed to carry out the extraction of airbound target characteristic feature using following process：
1) the SAR image slice I (m, n) only comprising an airbound target transmitted from image pre-processing module, wherein only Including the binary map of target area is B (m, n), then only include the image T (m, n) of target：
T (m, n)=I (m, n) × B (m, n)
Wherein, × indicate that respective pixel is multiplied；
2) minimum for acquiring airbound target body region according to the major axes orientation of airbound target individual in B (m, n) is external Rectangle, then the long side length Length of the rectangle is the length of airbound target individual, and the bond length Width of rectangle is to fly The width of row target individual；
3) geometry feature is calculated, including perimeter, area, length-width ratio, shape complexity, target centroid position It sets and rotary inertia：
PerimeterAreaLength-width ratio R=Length/ Width；Shape complexity C=Length2/4 π S；The centroid position of target area
Rotary inertiaIn formula, r represents the distance between target pixel points and barycenter,
4) gray-scale statistical characteristics are calculated, including quality, mean value, coefficient of variation, standard deviation, fractal dimension, add Weigh packing ratio：
QualityMean valueCoefficient of variation Standard deviationIn formula Gray scale logarithm and gray scale logarithm quadratic sum are indicated respectively；Fractal dimension H=(log10N1-log10N2)/(log10d1-log10d2), The computational methods of this feature are：With one K for remaining target area of SAR image slice structure after segmentation (K=is taken here 50) the binary map B of a brightest pixel point2One size is first d by (m, n)1×d2Window it is continuous in this binary map Sliding writes down the window sum comprising bright spot in window and is denoted as N1, with a size it is again then d2×d2Window this two It is continuously slipping in value figure, it writes down the window sum comprising bright spot in the window and is denoted as N2；Weight packing ratio
Feature selection module is completed to select optimal feature subset using following process：
1) inter- object distance is calculatedBetween class distanceAnd class spacing J in classi:
Wherein, i indicates that feature label, ω indicate the label of airbound target classification, | | Fi (ω)||2Indicate feature vector Fi (ω) 2 norms,Indicate the population mean of training set sample, NωIndicate the number of ω class airbound targets Amount, N indicate that airbound target sum in training set, E indicate it is expected, subscript W, subscript B are indicated respectively in class, between class.
2) normalization coefficient of variation ρ is calculatedi (ω)：
ρi (ω)=E [| | Fi (ω)||2 2]-E2[||Fi (ω)||2]/E[||Fi (ω)||2 2]
Wherein, i indicates that feature label, ω indicate the label of airbound target classification, | | Fi (ω)||2Indicate feature vector Fi (ω) 2 norms, E [| | Fi (ω)||2 2] and E2[||Fi (ω)||2] mean value of square of feature and square of mean value are indicated respectively.Feature Coefficient of variation ρi (ω)It is smaller, show that the stability of the target signature is better；
3) correlation coefficient r is calculatedi,j：
Wherein, i, j indicate feature label, | | Fi||2Indicate feature Fi2 norms,WithF is indicated respectivelyiAnd FjIt is equal Value, σi,iAnd σj,jF is indicated respectivelyiAnd FjStandard deviation.By the property of related coefficient it is found that 0≤ri,j≤1；If two features Complete uncorrelated, ri,j=0；If two features are perfectly correlated, ri,j=1；If the correlation between two features is very low, i.e., Information redundancy between feature is considerably less, then ri,jIt will be closer to 0；, whereas if the correlation between two features is very Height, i.e., the information redundancy between feature is very more, then ri,jIt will be closer to 1；
4) optimal feature subset is filtered out by class spacing, normalization coefficient of variation, related coefficient in class obtained above, Construct optimal input feature value；
Classifier training module is completed to carry out classifier training using following process：
5) N number of SAR radar images x is acquired from feature selection moduleiAs training sample, i=1,2 ..., N；
6) training sample is normalized, obtains normalization sample
7) training sample after normalization is reconstructed, respectively obtains input matrix X and corresponding output matrix Y:
Wherein, D indicates reconstruct dimension, and D is natural number, and D<The value range of N, D are 50-70；
8) classifier training set is obtainedWherein xiRepresent the input feature vector of n × 1 to Amount, tiRepresent the object vector of m × 1.Give the node number of an activation primitive g (x) and hidden layerSo ELM points Class device is：
Wherein, ωiRepresent the weight vectors between i-th of hidden layer node and input layer, biRepresent i-th of hidden layer section The biasing of point, βiRepresent the weight vectors between i-th of hidden layer node and output layer, ojRepresent the target of j-th of input data Output.In addition, ωi·xjRepresent ωiAnd xjInner product.
The output of the network can be ad infinitum close to N number of sample of input, i.e.,：
It can obtain：
Above formula can be expressed as matrix form：H β=T
Wherein, H indicates that the output matrix of hidden layer, the i-th row of H indicate i-th of node of hidden layer corresponding to N number of respectively Input x1,x2,…,xNOutput valve.The input weights of Single hidden layer feedforward neural networks (SLFNs) and the deviation of hidden layer are in net It need not be adjusted during network training, it can be any given.Based on above-mentioned theory, output weight can be by calculating H β=T Least square solutionIt acquires：
Non trivial solution can be quickly acquired using linear method, as shown in formula：
Wherein,The Moore-Penrose generalized inverse matrix of H are represented,LS solution of the least norm is represented, it is just It is the solution of Norm minimum in least square solution well.Compared to many existing classifier systems, extreme learning machine passes through this The solution of Moore-Penrose generalized inverses can reach good training effect at a very rapid rate.
Gunz optimization module, to use the optimization module based on swarm intelligence algorithm to the nuclear parameter θ of grader and punishment Factor gamma optimizes, and is completed using following process：
1) algorithm initialization constructs initial disaggregation S=(s according to grader form to be optimized1,s2,…,sn), really Determine the size m of ant colony, the threshold value MaxGen of ant optimization algorithm iteration number is set and initializes the iterations of ant optimization Serial number gen=0；
2) the corresponding fitness value Fit of disaggregation S are calculatedi(i=1,2 ..., n), fitness value is bigger to represent Xie Yuehao； The probability P that solution concentrates each solution to be fetched into the initial solution as ant optimizing is determined further according to following formulai(i=1,2 ..., n)
Initialization executes the ant number i=0 of optimizing algorithm；
3) ant i chooses initial solution of the solution as optimizing in S, and selection rule is to do wheel disc choosing according to P；
4) ant i carries out optimizing on the basis of the initial solution of selection, finds preferably solution si′；
If 5) i<M, then i=i+1, return to step 3)；Otherwise continue to execute step 6) downwards；
If 6) gen<MaxGen, then gen=gen+1, replaces S using the best solution that all ants in step 4) obtain In homographic solution, return to step 2)；Otherwise step 7) is executed downwards；
7) the corresponding fitness value Fit of disaggregation S are calculatedi(i=1,2 ..., n) chooses the maximum solution conduct of fitness value The optimal solution of algorithm terminates algorithm and returns.
Result display module shows the type for inputting airbound target in SAR image the display of result is identified Show in screen.
The present invention technical concept be：The present invention is directed to SAR radars round-the-clock, all weather operations and the spy penetrated by force Property, image preprocessing is carried out to the aerial image that SAR radars monitor, then carries out the extraction of feature and the selection of feature, Airflight Model of Target Recognition is established finally by the training process of grader, to realize SAR radar airbound targets Identification.
Beneficial effects of the present invention are mainly manifested in：1, airflight target can be identified in real time；2, recognition methods used Only need less training sample；3, intelligent, small by interference from human factor.
Description of the drawings
Fig. 1 is the overall structure figure of system proposed by the invention；
Fig. 2 is the functional block diagram of host computer proposed by the invention.
Specific implementation mode
The present invention is illustrated below according to attached drawing.Above-described embodiment is used for illustrating the present invention, rather than to this hair It is bright to be limited, in the protection domain of spirit and claims of the present invention, to any modifications and changes for making of the present invention, Both fall within protection scope of the present invention.
Embodiment
Referring to Fig.1, Fig. 2, a kind of gunz optimization SAR radar airbound target identifying systems, including SAR radars 1, data Library 2 and host computer 3, SAR radars 1, database 2 and host computer 3 are sequentially connected, and the SAR radars 1 shine monitored marine site It penetrates, and by SAR radar images storage to the database 2, the host computer 3 includes：
Image pre-processing module 4 is completed to carry out SAR radar image data pretreatments using following process：
1) the SAR image gray level transmitted from database is L, f (x0,y0) it is pixel (x0,y0) at gray value, g (x0,y0) it is pixel (x0,y0) N × N neighborhoods in pixel average value, wherein x0,y0The abscissa of pixel is indicated respectively And ordinate；
2) the number of pixels h (m, n) for meeting f=m and g=n by calculating, obtains joint probability density pmn：
pmn=p (m, n)=h (m, n)/M
Wherein, M indicates the total number of image pixel；
3) the mean vector μ of two-dimensional histogram is calculated：
4) probability P that target and background occurs in image is calculated separately0,1With mean vector μ0,1：
Wherein, t, s, subscript 0, subscript 1 indicate f segmentation thresholds, g segmentation thresholds, target area, background area respectively；
5) inter-class variance BCV is calculated：
BCV=P0(μ0-μ)(μ0-μ)′+P1(μ1-μ)(μ1-μ)′；
Wherein, μ indicates mean vector, the transposition of subscript ' representing matrix.
6) optimal threshold is the Two Dimensional Thresholding vector [s so that when BCV is maximum value0,t0]：
Characteristic extracting module 5 is completed to carry out the extraction of airbound target characteristic feature using following process：
1) the SAR image slice I (m, n) only comprising an airbound target transmitted from image pre-processing module, wherein only Including the binary map of target area is B (m, n), then only include the image T (m, n) of target：
T (m, n)=I (m, n) × B (m, n)
Wherein, × indicate that respective pixel is multiplied；
2) minimum for acquiring airbound target body region according to the major axes orientation of airbound target individual in B (m, n) is external Rectangle, then the long side length Length of the rectangle is the length of airbound target individual, and the bond length Width of rectangle is to fly The width of row target individual；
3) geometry feature is calculated, including perimeter, area, length-width ratio, shape complexity, target centroid position It sets and rotary inertia：
PerimeterAreaLength-width ratio R=Length/ Width；Shape complexity C=Length2/4πS；The centroid position of target area
Rotary inertiaIn formula, r represents the distance between target pixel points and barycenter,
4) gray-scale statistical characteristics are calculated, including quality, mean value, coefficient of variation, standard deviation, fractal dimension, add Weigh packing ratio：
QualityMean valueCoefficient of variation Standard deviationIn formula Gray scale logarithm and gray scale logarithm quadratic sum are indicated respectively；Fractal dimension H=(log10N1-log10N2)/(log10d1-log10d2), The computational methods of this feature are：With one K for remaining target area of SAR image slice structure after segmentation (K=is taken here 50) the binary map B of a brightest pixel point2One size is first d by (m, n)1×d2Window it is continuous in this binary map Sliding writes down the window sum comprising bright spot in window and is denoted as N1, with a size it is again then d2×d2Window this two It is continuously slipping in value figure, it writes down the window sum comprising bright spot in the window and is denoted as N2；Weight packing ratio
Feature selection module 6 is completed to select optimal feature subset using following process：
1) inter- object distance is calculatedBetween class distanceAnd class spacing J in classi:
Wherein, i indicates that feature label, ω indicate the label of airbound target classification, | | Fi (ω)||2Indicate feature vector Fi (ω) 2 norms,Indicate the population mean of training set sample, NωIndicate the number of ω class airbound targets Amount, N indicate that airbound target sum in training set, E indicate it is expected, subscript W, subscript B are indicated respectively in class, between class.
2) normalization coefficient of variation ρ is calculatedi (ω)：
ρi (ω)=E [| | Fi (ω)||2 2]-E2[||Fi (ω)||2]/E[||Fi (ω)||2 2]
Wherein, i indicates that feature label, ω indicate the label of airbound target classification, | | Fi (ω)||2Indicate feature vector Fi (ω) 2 norms, E [| | Fi (ω)||2 2] and E2[||Fi (ω)||2] mean value of square of feature and square of mean value are indicated respectively.Feature Coefficient of variation ρi (ω)It is smaller, show that the stability of the target signature is better；
3) correlation coefficient r is calculatedi,j：
Wherein, i, j indicate feature label, | | Fi||2Indicate feature Fi2 norms,WithF is indicated respectivelyiAnd FjIt is equal Value, σi,iAnd σj,jF is indicated respectivelyiAnd FjStandard deviation.By the property of related coefficient it is found that 0≤ri,j≤1；If two features Complete uncorrelated, ri,j=0；If two features are perfectly correlated, ri,j=1；If the correlation between two features is very low, i.e., Information redundancy between feature is considerably less, then ri,jIt will be closer to 0；, whereas if the correlation between two features is very Height, i.e., the information redundancy between feature is very more, then ri,jIt will be closer to 1；
4) optimal feature subset is filtered out by class spacing, normalization coefficient of variation, related coefficient in class obtained above, Construct optimal input feature value；
Classifier training module 7 is completed to carry out classifier training using following process：
1) N number of SAR radar images x is acquired from feature selection moduleiAs training sample, i=1,2 ..., N；
2) training sample is normalized, obtains normalization sample
3) training sample after normalization is reconstructed, respectively obtains input matrix X and corresponding output matrix Y:
Wherein, D indicates reconstruct dimension, and D is natural number, and D<The value range of N, D are 50-70；
4) classifier training set is obtainedWherein xiRepresent the input feature vector of n × 1 to Amount, tiRepresent the object vector of m × 1.Give the node number of an activation primitive g (x) and hidden layerSo ELM points Class device is：
Wherein, ωiRepresent the weight vectors between i-th of hidden layer node and input layer, biRepresent i-th of hidden layer section The biasing of point, βiRepresent the weight vectors between i-th of hidden layer node and output layer, ojRepresent the target of j-th of input data Output.In addition, ωi·xjRepresent ωiAnd xjInner product.
The output of the network can be ad infinitum close to N number of sample of input, i.e.,：
It can obtain：
Above formula can be expressed as matrix form：H β=T
Wherein, H indicates that the output matrix of hidden layer, the i-th row of H indicate i-th of node of hidden layer corresponding to N number of respectively Input x1,x2,…,xNOutput valve.The input weights of Single hidden layer feedforward neural networks (SLFNs) and the deviation of hidden layer are in net It need not be adjusted during network training, it can be any given.Based on above-mentioned theory, output weight can be by calculating H β=T Least square solutionIt acquires：
Non trivial solution can be quickly acquired using linear method, as shown in formula：
Wherein,The Moore-Penrose generalized inverse matrix of H are represented,LS solution of the least norm is represented, it is just It is the solution of Norm minimum in least square solution well.Compared to many existing classifier systems, extreme learning machine passes through this The solution of Moore-Penrose generalized inverses can reach good training effect at a very rapid rate.
Gunz optimization module 9, to use the optimization module based on swarm intelligence algorithm to the nuclear parameter θ of grader and punishment Factor gamma optimizes, and is completed using following process：
1) algorithm initialization constructs initial disaggregation S=(s according to grader form to be optimized1,s2,…,sn), really Determine the size m of ant colony, the threshold value MaxGen of ant optimization algorithm iteration number is set and initializes the iterations of ant optimization Serial number gen=0；
2) the corresponding fitness value Fit of disaggregation S are calculatedi(i=1,2 ..., n), fitness value is bigger to represent Xie Yuehao； The probability P that solution concentrates each solution to be fetched into the initial solution as ant optimizing is determined further according to following formulai(i=1,2 ..., n)
Initialization executes the ant number i=0 of optimizing algorithm；
3) ant i chooses initial solution of the solution as optimizing in S, and selection rule is to do wheel disc choosing according to P；
4) ant i carries out optimizing on the basis of the initial solution of selection, finds preferably solution si′；
If 5) i<M, then i=i+1, return to step 3)；Otherwise continue to execute step 6) downwards；
If 6) gen<MaxGen, then gen=gen+1, replaces S using the best solution that all ants in step 4) obtain In homographic solution, return to step 2)；Otherwise step 7) is executed downwards；
7) the corresponding fitness value Fit of disaggregation S are calculatedi(i=1,2 ..., n) chooses the maximum solution conduct of fitness value The optimal solution of algorithm terminates algorithm and returns.
Result display module 8 will input the type of airbound target in SAR image the display of result is identified It is shown in host computer.
The hardware components of the host computer 3 include：I/O elements, the transmission of acquisition and information for data；Data store Device, the required data sample of storage running and operating parameter etc.；The software journey of function module is realized in program storage, storage Sequence；Arithmetic unit executes program, realizes specified function；Display module shows the parameter and recognition result of setting.
Above-described embodiment is used for illustrating the present invention, rather than limits the invention, the present invention spirit and In scope of the claims, to any modifications and changes that the present invention makes, protection scope of the present invention is both fallen within.

Claims (5)

1. a kind of gunz optimizes SAR radar airbound target identifying systems, it is characterised in that：Including SAR radars, database with And host computer, SAR radars, database and host computer are sequentially connected, the SAR radars to being monitored in real time in the air, and by SAR To in the database, the host computer includes that image pre-processing module, feature carry for the image data storage that radar obtains Modulus block, classifier training module, classifier training module, gunz optimization module and result display module, described image are located in advance Reason module, characteristic extracting module, classifier training module, classifier training module and result display module are sequentially connected, and are classified Device training module is connected with gunz optimization module.
2. gunz optimizes SAR radar airbound target identifying systems according to claim 1, it is characterised in that：The figure As preprocessing module is to carry out SAR radar image data pretreatments, completed using following process：
1) the SAR image gray level transmitted from database is L, f (x0,y0) it is pixel (x0,y0) at gray value, g (x0, y0) it is pixel (x0,y0) N × N neighborhoods in pixel average value, wherein x0,y0The abscissa of pixel is indicated respectively and is indulged Coordinate；
2) the number of pixels h (m, n) for meeting f=m and g=n by calculating, obtains joint probability density pmn：
pmn=p (m, n)=h (m, n)/M
Wherein, M indicates the total number of image pixel；
3) the mean vector μ of two-dimensional histogram is calculated：
4) probability P that target and background occurs in image is calculated separately0,1With mean vector μ0,1：
Wherein, t, s, subscript 0, subscript 1 indicate f segmentation thresholds, g segmentation thresholds, target area, background area respectively；
5) inter-class variance BCV is calculated：
BCV=P0(μ0-μ)(μ0-μ)′+P1(μ1-μ)(μ1-μ)′；
Wherein, μ indicates mean vector, the transposition of subscript ' representing matrix.
6) optimal threshold is the Two Dimensional Thresholding vector [s so that when BCV is maximum value0,t0]：
3. gunz optimizes SAR radar airbound target identifying systems according to claim 1, it is characterised in that：The spy Extraction module is levied to carry out the extraction of airflight target characteristic feature, is completed using following process：
1) the SAR image slice I (m, n) only comprising an airbound target transmitted from image pre-processing module, wherein including only The binary map of target area is B (m, n), then only includes the image T (m, n) of target：
T (m, n)=I (m, n) × B (m, n)
Wherein, × indicate that respective pixel is multiplied；
2) minimum enclosed rectangle of airbound target body region is acquired according to the major axes orientation of airbound target individual in B (m, n), Then the long side length Length of the rectangle is the length of airbound target individual, and the bond length Width of rectangle is flight mesh Mark the width of individual；
3) geometry feature is calculated, including perimeter, area, length-width ratio, shape complexity, target centroid position with And rotary inertia：
PerimeterAreaLength-width ratio R=Length/ Width；Shape complexity C=Length2/4πS；The centroid position of target area
Rotary inertiaIn formula, r represents the distance between target pixel points and barycenter,
4) gray-scale statistical characteristics are calculated, are filled out including quality, mean value, coefficient of variation, standard deviation, fractal dimension, weighting Fill ratio：
QualityMean valueCoefficient of variation Standard deviationIn formula Gray scale logarithm and gray scale logarithm quadratic sum are indicated respectively；Fractal dimension H=(log10N1-log10N2)/(log10d1-log10d2), The computational methods of this feature are：With one K for remaining target area of SAR image slice structure after segmentation (K=is taken here 50) the binary map B of a brightest pixel point2One size is first d by (m, n)1×d2Window it is continuous in this binary map Sliding writes down the window sum comprising bright spot in window and is denoted as N1, with a size it is again then d2×d2Window this two It is continuously slipping in value figure, it writes down the window sum comprising bright spot in the window and is denoted as N2；Weight packing ratio
4. gunz optimizes SAR radar airbound target identifying systems according to claim 1, it is characterised in that：The spy Selecting module is levied to select optimal feature subset, is completed using following process：
1) inter- object distance is calculatedBetween class distanceAnd class spacing J in classi:
Wherein, i indicates that feature label, ω indicate the label of airbound target classification, | | Fi (ω)||2Indicate feature vector Fi (ω)2 models Number,Indicate the population mean of training set sample, NωIndicate the quantity of ω class airbound targets, N tables Show that airbound target sum in training set, E indicate it is expected, subscript W, subscript B are indicated respectively in class, between class.
2) normalization coefficient of variation ρ is calculatedi (ω)：
Wherein, i indicates that feature label, ω indicate the label of airbound target classification, | | Fi (ω)||2Indicate feature vector Fi (ω)2 models Number, E [| | Fi (ω)||2 2] and E2[||Fi (ω)||2] mean value of square of feature and square of mean value are indicated respectively.The variance of feature Coefficient ρi (ω)It is smaller, show that the stability of the target signature is better；
3) correlation coefficient r is calculatedi,j：
Wherein, i, j indicate feature label, | | Fi||2Indicate feature Fi2 norms,WithF is indicated respectivelyiAnd FjMean value, σi,iAnd σj,jF is indicated respectivelyiAnd FjStandard deviation.By the property of related coefficient it is found that 0≤ri,j≤1；If two features are complete It is uncorrelated, ri,j=0；If two features are perfectly correlated, ri,j=1；If the correlation between two features is very low, i.e. feature Between information redundancy it is considerably less, then ri,jIt will be closer to 0；, whereas if the correlation between two features is very high, i.e., Information redundancy between feature is very more, then ri,jIt will be closer to 1；
4) optimal feature subset is filtered out by class spacing, normalization coefficient of variation, related coefficient in class obtained above, constructed Optimal input feature value.
5. gunz optimizes SAR radar airbound target identifying systems according to claim 1, it is characterised in that：Described point Class device training module is completed to carry out classifier training using following process：
1) N number of SAR radar images x is acquired from feature selection moduleiAs training sample, i=1,2 ..., N；
2) training sample is normalized, obtains normalization sample
3) training sample after normalization is reconstructed, respectively obtains input matrix X and corresponding output matrix Y:
Wherein, D indicates reconstruct dimension, and D is natural number, and D<The value range of N, D are 50-70；
4) classifier training set is obtainedWherein xiRepresent the input feature value of n × 1, ti Represent the object vector of m × 1.Give the node number of an activation primitive g (x) and hidden layerSo ELM graders For：
Wherein, ωiRepresent the weight vectors between i-th of hidden layer node and input layer, biRepresent i-th hidden layer node Biasing, βiRepresent the weight vectors between i-th of hidden layer node and output layer, ojThe target for representing j-th of input data is defeated Go out.In addition, ωi·xjRepresent ωiAnd xjInner product.
The output of the network can be ad infinitum close to N number of sample of input, i.e.,：
It can obtain：
Above formula can be expressed as matrix form：H β=T
Wherein, H indicates that the output matrix of hidden layer, the i-th row of H indicate that i-th of node of hidden layer corresponds to N number of input respectively x1,x2,…,xNOutput valve.The input weights of Single hidden layer feedforward neural networks (SLFNs) and the deviation of hidden layer are instructed in network It need not be adjusted in experienced process, it can be any given.Based on above-mentioned theory, output weight can be by calculating H β=T most Small two multiply solutionIt acquires：
Non trivial solution can be quickly acquired using linear method, as shown in formula：
Wherein,The Moore-Penrose generalized inverse matrix of H are represented,LS solution of the least norm is represented, it is exactly most Small two multiply the solution of Norm minimum in solution.Compared to many existing classifier systems, extreme learning machine passes through this Moore- The solution of Penrose generalized inverses can reach good training effect at a very rapid rate.
The gunz optimization module to using the optimization module based on swarm intelligence algorithm to the nuclear parameter θ of grader and punishment because Sub- γ is optimized, and is completed using following process：
1) algorithm initialization constructs initial disaggregation S=(s according to grader form to be optimized1,s2,…,sn), determine ant The size m of group is arranged the threshold value MaxGen of ant optimization algorithm iteration number and initializes the iterations serial number of ant optimization Gen=0；
2) the corresponding fitness value Fit of disaggregation S are calculatedi(i=1,2 ..., n), fitness value is bigger to represent Xie Yuehao；Further according to Following formula determines the probability P that solution concentrates each solution to be fetched into the initial solution as ant optimizingi(i=1,2 ..., n)
Initialization executes the ant number i=0 of optimizing algorithm；
3) ant i chooses initial solution of the solution as optimizing in S, and selection rule is to do wheel disc choosing according to P；
4) ant i carries out optimizing on the basis of the initial solution of selection, finds preferably solution si′；
If 5) i<M, then i=i+1, return to step 3)；Otherwise continue to execute step 6) downwards；
If 6) gen<MaxGen, then gen=gen+1, is replaced using the best solution that all ants in step 4) obtain in S Homographic solution, return to step 2)；Otherwise step 7) is executed downwards；
7) the corresponding fitness value Fit of disaggregation S are calculatedi(i=1,2 ..., n) chooses the maximum solution of fitness value and is used as algorithm Optimal solution, terminate algorithm simultaneously return.
The result display module shows the type for inputting airbound target in SAR image the display of result is identified Show in screen.
CN201810457315.7A 2018-05-14 2018-05-14 A kind of gunz optimization SAR radar airbound target identifying systems Pending CN108681732A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201810457315.7A CN108681732A (en) 2018-05-14 2018-05-14 A kind of gunz optimization SAR radar airbound target identifying systems

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201810457315.7A CN108681732A (en) 2018-05-14 2018-05-14 A kind of gunz optimization SAR radar airbound target identifying systems

Publications (1)

Publication Number Publication Date
CN108681732A true CN108681732A (en) 2018-10-19

Family

ID=63806249

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201810457315.7A Pending CN108681732A (en) 2018-05-14 2018-05-14 A kind of gunz optimization SAR radar airbound target identifying systems

Country Status (1)

Country Link
CN (1) CN108681732A (en)

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102183749A (en) * 2011-03-03 2011-09-14 浙江大学 Sea target detecting system of adaptive radar and method thereof
CN103824121A (en) * 2013-12-09 2014-05-28 浙江大学 Propylene polymerization production process optimal prediction system based on multimode crowd-sourcing and method
CN103823369A (en) * 2013-12-09 2014-05-28 浙江大学 Propylene polymerization production process prediction system based on BP multimode network and method

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102183749A (en) * 2011-03-03 2011-09-14 浙江大学 Sea target detecting system of adaptive radar and method thereof
CN103824121A (en) * 2013-12-09 2014-05-28 浙江大学 Propylene polymerization production process optimal prediction system based on multimode crowd-sourcing and method
CN103823369A (en) * 2013-12-09 2014-05-28 浙江大学 Propylene polymerization production process prediction system based on BP multimode network and method

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
JUN WU 等: "A novel ship classification approach for high resolution SAR images based on the BDA-KELM classification model", 《INTERNATIONAL JOURNAL OF REMOTE SENSING》 *
陈文婷: "SAR图像舰船目标特征提取与分类识别方法研究", 《中国优秀硕士学位论文全文数据库 基础科学辑》 *

Similar Documents

Publication Publication Date Title
CN108596156A (en) A kind of intelligence SAR radar airbound target identifying systems
CN103489005B (en) A kind of Classification of High Resolution Satellite Images method based on multiple Classifiers Combination
CN110135267A (en) A kind of subtle object detection method of large scene SAR image
CN111160268B (en) Multi-angle SAR target recognition method based on multi-task learning
CN110060233A (en) A kind of corn ear damage testing method
CN105160623B (en) Unsupervised high-spectral data dimension reduction method based on chunking low-rank tensor model
CN108490415A (en) Ship seakeeping system in a kind of intelligence SAR Radar Seas
CN106529563B (en) EO-1 hyperion band selection method based on the sparse Non-negative Matrix Factorization of digraph
CN111914728A (en) Hyperspectral remote sensing image semi-supervised classification method and device and storage medium
CN108932455B (en) Remote sensing image scene recognition method and device
CN113344045B (en) Method for improving SAR ship classification precision by combining HOG characteristics
CN114399685B (en) Remote sensing monitoring and evaluating method and device for forest pest and disease damage
CN103246894A (en) Ground nephogram identifying method solving problem of insensitiveness in illumination
CN108629382A (en) Ship seakeeping system in a kind of gunz optimization SAR Radar Seas
Sabri et al. Nutrient deficiency detection in maize (Zea mays L.) leaves using image processing
Chakrabarti et al. Disaggregation of remotely sensed soil moisture in heterogeneous landscapes using holistic structure-based models
CN108596155A (en) A kind of SAR radars airbound target identifying system
CN116824585A (en) Aviation laser point cloud semantic segmentation method and device based on multistage context feature fusion network
Hsieh Evolution of machine learning in environmental science—A perspective
CN114862596A (en) Crop bank loan data risk analysis method and device
CN107563420B (en) The Classification of Polarimetric SAR Image method encoded certainly based on scattering energy and storehouse
CN108875578A (en) A kind of adaptive colony intelligence optimization SAR radar airbound target identifying system
CN108647653A (en) A kind of intelligence SAR radars land tank target identifying system
CN115205693B (en) Method for extracting enteromorpha in multi-feature integrated learning dual-polarization SAR image
CN108681732A (en) A kind of gunz optimization SAR radar airbound target identifying systems

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
RJ01 Rejection of invention patent application after publication

Application publication date: 20181019

RJ01 Rejection of invention patent application after publication