CN111767806B - Ultra-narrow pulse radar ship target identification method based on Attribute - Google Patents

Ultra-narrow pulse radar ship target identification method based on Attribute Download PDF

Info

Publication number
CN111767806B
CN111767806B CN202010536147.8A CN202010536147A CN111767806B CN 111767806 B CN111767806 B CN 111767806B CN 202010536147 A CN202010536147 A CN 202010536147A CN 111767806 B CN111767806 B CN 111767806B
Authority
CN
China
Prior art keywords
attribute
image
ship
target
ship target
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.)
Active
Application number
CN202010536147.8A
Other languages
Chinese (zh)
Other versions
CN111767806A (en
Inventor
龙腾
李枫
李姗
李阳
姚迪
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.)
Beijing Institute of Technology BIT
Original Assignee
Beijing Institute of Technology BIT
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 Beijing Institute of Technology BIT filed Critical Beijing Institute of Technology BIT
Priority to CN202010536147.8A priority Critical patent/CN111767806B/en
Publication of CN111767806A publication Critical patent/CN111767806A/en
Application granted granted Critical
Publication of CN111767806B publication Critical patent/CN111767806B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/10—Terrestrial scenes
    • G06V20/13—Satellite images
    • 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
    • G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
    • 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
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
    • 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
    • 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
    • G06V10/26—Segmentation of patterns in the image field; Cutting or merging of image elements to establish the pattern region, e.g. clustering-based techniques; Detection of occlusion
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G06V10/50—Extraction of image or video features by performing operations within image blocks; by using histograms, e.g. histogram of oriented gradients [HoG]; by summing image-intensity values; Projection analysis
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
    • Y02A90/00—Technologies having an indirect contribution to adaptation to climate change
    • Y02A90/10—Information and communication technologies [ICT] supporting adaptation to climate change, e.g. for weather forecasting or climate simulation

Landscapes

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

Abstract

The invention discloses an extremely narrow pulse radar ship target identification method based on Attribute, which has controllable process, and the obtained middle layer features have stronger resolution, so that the robustness of ship target identification is further improved. And dividing the ship target in the ultra-narrow pulse radar image to obtain a divided image of the ship target as a training image. Low-level features of the training image are extracted. Constructing a mapping model of Attribute middle-layer features and low-layer features according to three properties of Attribute, namely reconfigurability, high discriminant and high predictability; and mapping the low-level features of the training image to attribute mid-level features by using the mapping model. And constructing an SVM classifier, taking the attribute middle layer characteristics of the training image as input, and training the SVM classifier to obtain the trained SVM classifier for identifying the target of the ultra-narrow pulse radar ship.

Description

Ultra-narrow pulse radar ship target identification method based on Attribute
Technical Field
The invention relates to the technical field of ship target identification, in particular to an ultra-narrow pulse radar ship target identification method based on Attribute.
Background
In recent years, the detection and identification of ships by using radar remote sensing images are highly valued in the field of marine remote sensing application. The radar can observe a large-range ocean area all the time and all the weather, and is one of effective means for identifying ships in a large-range sea area.
With the research and large-scale application of various radar systems, a series of intensive researches are carried out in China in the application field of ocean monitoring by utilizing radar images. For example, the national defense science and technology university, the noble and the like propose to identify radar image targets by utilizing the characteristics of target peak sequences and the characteristic extraction method thereof; the university of ocean in China, zhang Xi and the like, provides that primary characteristics such as length, width, peak value and the like of a ship and multi-polarization information are adopted to analyze scattering characteristics of the ship by considering structural characteristics of the radar ship, so that classification and identification of the ship are realized. In the field of ship identification, feature extraction mostly adopts various primary images such as appearance geometric structures, electromagnetic scattering and the like to describe features. In very narrow pulse radar images, the directional gradient histogram (Histogram of Oriented Gradient, HOG) features work well.
At present, two problems exist in ship identification, namely, when the self-posture and geometric form of a ship target and radar parameters are changed, the attribute of the ship target in a radar image is changed, and then the problem of large intra-class change is necessarily caused based on HOG characteristics extracted by an image layer. The robustness is not strong, and the identification accuracy is affected. Secondly, the radar ship image sample is few. To ensure accuracy of the identification, the identification system must obtain the target characteristics of the ship target under various attitudes, various geometric shapes and all radar parameters, which requires that radar ship images represent various imaging situations which can occur in practice, and the existing radar ship images are far from meeting the requirements.
Compared with the low-level features, for the extremely narrow pulse radar ship target, the middle-level features can describe the target stably, are not influenced by the outside as much as possible, and reduce the semantic gap between the low-level features and the high-level features. The Bag of Words (BoW) model is a simple and effective method for obtaining middle-layer semantic features of an image, but Bag of Words learning is an unsupervised learning method and cannot control the process of obtaining the middle-layer features.
Therefore, a process-controllable middle-layer semantic feature acquisition method is needed to improve the robustness of ship target identification.
Disclosure of Invention
In view of the above, the invention provides an extremely narrow pulse radar ship target identification method based on Attribute, the process is controllable, the obtained middle layer features have stronger resolution, and the robustness of ship target identification is further improved.
In order to achieve the above purpose, the technical scheme of the invention is as follows:
and dividing the ship target in the ultra-narrow pulse radar image to obtain a divided image of the ship target as a training image.
Low-level features of the training image are extracted.
Constructing a mapping model of Attribute middle-layer features and low-layer features according to three properties of Attribute, namely reconfigurability, high discriminant and high predictability; and mapping the low-level features of the training image to attribute mid-level features by using the mapping model.
And constructing an SVM classifier, taking the attribute middle layer characteristics of the training image as input, and training the SVM classifier to obtain the trained SVM classifier for identifying the target of the ultra-narrow pulse radar ship.
Further, the ship target in the ultra-narrow pulse radar image is segmented, and a segmented image of the ship target is obtained as a training image, specifically: and processing the ultra-narrow pulse radar image by using Radon transformation, determining a ship main shaft of the ship target, and dividing the ship target from the sea clutter background of the ultra-narrow pulse radar image to obtain a divided image of the ship target as a training image.
Further, the method comprises the steps of dividing the ship target in the ultra-narrow pulse radar image to obtain a divided image of the ship target as a training image, and specifically comprises the following steps:
step 101, preprocessing an extremely narrow pulse radar image, specifically: and sequentially carrying out gray scale expansion, wavelet denoising, otsu threshold segmentation, binary expansion, median filtering and binary corrosion on the extremely narrow pulse radar image to obtain a binary image, searching all connected domains in the binary image, and only reserving the largest connected domain to obtain a preprocessed image, wherein the largest connected domain is used as a ship target.
Step 102, radon transformation determines a vessel principal axis of the vessel target.
Radon transformation is carried out on the preprocessed image, and the line integral value of the ship target at each angle is calculated to be R (rho, theta) 0 ) ρ is the polar diameter of the polar coordinate system, θ 0 Is the polar angle of the polar coordinate system;
the projection direction of the ship main shaft is
Figure BDA0002537052390000031
Azimuth angle of ship target is θ=θ max -90°。
Step 103, determining the length range of the ship target by Radon transformation.
Rotating the preprocessed image clockwise by an angle theta to enable the ship spindle to be positioned in the horizontal direction, and performing the process again
The Radon transformation is carried out again on the rotated image, and all line integral values R of the 90-degree projection direction are taken * (ρ,90 °); t to be the maximum line integral value 2 The multiple is set to the second threshold value, and the line integral value smaller than the second threshold value is removed.
R * The coordinate range of the non-zero point in (ρ,90 DEG) is [ ρ 1 ，ρ 2 ]，ρ 1 Is the lower limit of the non-zero point coordinates, ρ 2 Is the upper limit of the non-zero coordinates; then [ ρ ] 1 ，ρ 2 ]I.e. the vertical direction division range of the ship in the image domain, and determining the length range [ x ] of the ship target in the binary image according to the pixel values 1 ，x 2 ],x 1 Is the lower limit of the target length of the ship, x 2 Is the upper limit of the target length of the ship.
t 2 Is a preset positive constant less than 1, and is empirically set.
And 104, dividing the ship target according to the length range of the ship target, and obtaining a divided image of the ship target as a training image.
Further, extracting low-level features of the training image, specifically:
and extracting the HOG characteristics of the directional gradient histogram of the training image.
Further, according to three properties of reconfigurability, high discriminant and high predictability of the Attribute Attribute, a mapping model of middle-layer characteristics and low-layer characteristics of the Attribute is constructed; the method comprises the following steps:
the optimized model of the constructed attribute middle-layer feature and low-layer feature mapping model is as follows:
Figure BDA0002537052390000041
Figure BDA0002537052390000042
Figure BDA0002537052390000043
Figure BDA0002537052390000044
Figure BDA0002537052390000045
wherein the method comprises the steps of
Figure BDA0002537052390000046
A reconfigurability model for Attribute Attribute; x is a low-level feature vector, and X dimension is Mxl, i.e. X ε R M×l L is the total number of training images; k (K) RBF Is a gaussian kernel function;
Figure BDA0002537052390000047
D i for the ith row of the dictionary D, D is a dictionary with the size of M multiplied by K, and M is the dimension of the low-level features; alpha is a middle layer feature of the attribute, the dimension of alpha is Kxl, and K is smaller than M; alpha j A j-th column of α; sigma > 0 is the bandwidth of the gaussian kernel.
J c (W, ζ, b, α) is a high discriminant model of the Attribute Attribute constructed using the SVM learning problem of class c, where for the ith image I i And category c, if y i =c, i.e. the I-th image I i Middle object class y c Class c, SVM classification result y ci Set to 1, otherwise y ci Set to-1.
Figure BDA0002537052390000048
Figure BDA0002537052390000049
ξ ci ≥0,i＝1,…,l
Wherein Q is 1 A constant value empirically set; w is a hyperplane set w= { W c } c C is the target class c, each class
Figure BDA00025370523900000410
Figure BDA00025370523900000411
For the total category, W c Is a hyperplane of category c for adding attribute vectors of category c and other categories +.>
Figure BDA0002537052390000051
Is divided by the attribute vector of (a); class c SVM classifier relaxation variable set ζ= { ζ c } c ， ξ c Relaxation variable, ζ, of class c SVM classifier ci An i-th term and a common l-th term in relaxation variables of the class c SVM classifier; class c SVM classifier displacement parameter set b= { b c } c ，b c Is a displacement parameter of the class c SVM classifier.
L k (β, α) is a high predictive model of the Attribute constructed using the dual problem of the support vector regression SVR formula, wherein:
Figure BDA0002537052390000052
Figure BDA0002537052390000053
Figure BDA0002537052390000054
Figure BDA0002537052390000055
wherein ε corresponds to the tolerance of SVR to prediction errors; q (Q) 2 To set constraint error experience value, Q 2 Is constant and empirically set; k is the dimension of the attribute alpha, and the range is 1-K; for each attribute k, a hyperplane w is learned using linear SVM classification k The method comprises the steps of carrying out a first treatment on the surface of the Beta is an attribute prediction related parameter,
Figure BDA0002537052390000056
β k predicting a correlation parameter beta for a first attribute k ＝{β ki } i＝1...l ，
Figure BDA0002537052390000057
Predicting the relevant parameter for the second property +.>
Figure BDA0002537052390000058
k is attribute dimension, and the value range is [1, K]An integer within the interval; x is x i For image I i The corresponding low-level features, i.e., row i in X; x is x j For attribute predictor f k (k.epsilon. {1, …, K }) according to x i Predicting the obtained test sample; kappa (x) i ,x j ) To be substituted into x i And x j Is a positive semi-finite kernel of (c); alpha ik Ith row and kth column for α; test sample x j The value of the kth attribute of (c) is:
Figure BDA0002537052390000059
and carrying out optimization solution on the constructed optimization model of the attribute middle layer feature and the low layer feature mapping model to obtain an optimal solution of the attribute middle layer feature and the low layer feature mapping model.
Further, the optimization solution is carried out on the constructed optimization model of the attribute middle layer feature and the low layer feature mapping model to obtain the optimal solution of the attribute middle layer feature and the low layer feature mapping model, and the method specifically comprises the following steps:
s1, reducing the dimension of a low-level feature vector from M to K through a Principal Component Analysis (PCA) method to obtain an initialization value alpha (0) 。
S2, attribute alpha takes on value alpha (t-1) Solving the optimization problem of the reconfigurability model of the Attribute Attribute by using a PSO algorithm, namely
Figure BDA0002537052390000061
Solution D of the t-th iteration of learning dictionary D (t) 。
t is the iterative optimization times, and the initial value of t is 1.
S3, solving an optimization problem of a high discriminant model of the Attribute Attribute:
Figure BDA0002537052390000062
learning parameters W of hyperplanes belonging to different classes (t) ，b (t) 。
S4, solving an optimization problem of a high-predictability model of the Attribute Attribute:
Figure BDA0002537052390000063
learning attribute prediction related parameter beta (t) 。
S5, predicting relevant parameters beta by utilizing attributes (t) Solving an attribute prediction function f k (t) 。
S6, solving:
Figure BDA0002537052390000064
Figure BDA0002537052390000065
Figure BDA0002537052390000066
Figure BDA0002537052390000067
learning to obtain the attribute alpha of the t-th iteration (t) 。
Wherein Q3 is a constant, empirically set
Figure BDA0002537052390000068
Is two approximation errors of the SVR predictor.
S7, judging whether t reaches the preset maximum iteration number, if so, outputting alpha (t) And (4) as the optimal solution of the attribute alpha, otherwise, t is increased by 1, and the method returns to S2.
Further, the trained SVM classifier is obtained for extremely narrow pulse radar ship target discrimination, and specifically comprises:
and dividing the ship targets in the ultra-narrow pulse radar image to be identified to obtain a test image.
Low-level features of the test image are extracted.
And mapping the low-level features of the test image to the attribute middle-level features of the test image by using the constructed attribute middle-level features and low-level features mapping model.
The middle layer characteristics in the attribute of the test image are input into a trained SVM classifier for extremely narrow pulse radar ship target identification.
The beneficial effects are that:
according to the ultra-narrow pulse radar ship target identification method based on the Attribute, provided by the invention, the bottom layer characteristics of the image can be mapped to the Attribute middle layer characteristics based on mathematical optimization, the process is controllable, the obtained middle layer characteristics are stronger in resolution, and the robustness of ship target identification is further improved.
Drawings
FIG. 1 is a main flow chart of an Attribute-based ultra-narrow pulse radar ship target identification method provided by the invention;
FIG. 2 is a detailed flow chart of an Attribute-based ultra-narrow pulse radar ship target identification method provided by an embodiment of the invention.
FIG. 3 is a diagram illustrating a process of preprocessing an ultra-narrow pulse radar image at step 101 in accordance with an embodiment of the present invention;
fig. 4 is a schematic diagram of a result of segmentation of a ship target in an extremely narrow pulse radar image in an embodiment of the present invention.
Detailed Description
The invention will now be described in detail by way of example with reference to the accompanying drawings.
The invention provides an extremely narrow pulse radar ship target identification method based on Attribute, which is shown in figure 1 and comprises the following steps:
dividing a ship target in an extremely narrow pulse radar image to obtain a divided image of the ship target as a training image;
and processing the ultra-narrow pulse radar image by using Radon transformation, determining a ship main shaft of the ship target, and dividing the ship target from the sea clutter background of the ultra-narrow pulse radar image to obtain a divided image of the ship target as a training image.
The method comprises the following steps:
step 101, preprocessing an extremely narrow pulse radar image, specifically: and sequentially carrying out gray scale expansion, wavelet denoising, otsu threshold segmentation, binary expansion, median filtering and binary corrosion on the extremely narrow pulse radar image to obtain a binary image, searching all connected domains in the binary image, and only reserving the largest connected domain to obtain a preprocessed image, wherein the largest connected domain is used as a ship target. Fig. 3 illustrates a process of preprocessing an extremely narrow pulse radar image in an embodiment of the present invention.
Step 102, radon transformation determines a vessel principal axis of the vessel target.
Carrying out Radon transformation on the preprocessed image to obtain an image of one Radon transformation, wherein all line integral values of the ship target along the main axis direction are all obtainedThe linear integral value of the ship target at each angle is calculated to be R (rho, theta 0 ) ρ is the polar diameter of the polar coordinate system, θ 0 Is the polar angle of the polar coordinate system.
The projection direction of the ship main shaft is
Figure BDA0002537052390000081
Azimuth angle of ship target is θ=θ max -90°。
Step 103, determining the length range of the ship target by Radon transformation.
The preprocessed image is rotated clockwise by an angle θ such that the ship's principal axis is in the horizontal direction.
The rotated image is subjected to Radon transformation again to obtain a secondary Radon transformed image, and all line integral values R in the 90-degree projection direction are taken * (ρ,90 °) the number of points of which is the width of the target; t to be the maximum line integral value 2 The multiple is set to the second threshold value, and the line integral value smaller than the second threshold value is removed.
R * The coordinate range of the non-zero point in (ρ,90 DEG) is [ ρ 1 ，ρ 2 ]，ρ 1 Is the lower limit of the non-zero point coordinates, ρ 2 Is the upper limit of the non-zero coordinates; then [ ρ ] 1 ，ρ 2 ]I.e. the vertical direction division range of the ship in the image domain, and determining the length range [ x ] of the ship target in the binary image according to the pixel values 1 ，x 2 ],x 1 Is the lower limit of the target length of the ship, x 2 Is the upper limit of the target length of the ship.
t 2 The positive constant smaller than 1 can be set empirically.
And 104, dividing the ship target according to the length range of the ship target, and obtaining a divided image of the ship target as a training image. Fig. 4 shows an effect diagram of segmentation of an extremely narrow pulse radar image using steps 101 to 104 in an embodiment of the present invention.
Step two, extracting low-level features of the training image; in the embodiment of the invention, the HOG characteristics of the directional gradient histogram of the training image are extracted.
Step three, constructing a mapping model of middle-layer features and low-layer features of the Attribute according to three properties of reconfigurability, high discriminant and high predictability of the Attribute Attribute; and mapping the low-level features of the training image to the attribute mid-level features using the mapping model.
The method comprises the following steps:
the optimized model of the constructed attribute middle-layer feature and low-layer feature mapping model is as follows:
Figure BDA0002537052390000091
Figure BDA0002537052390000092
Figure BDA0002537052390000093
Figure BDA0002537052390000094
Figure BDA0002537052390000095
the optimization model comprises three parts:
wherein the method comprises the steps of
Figure BDA0002537052390000096
A reconfigurability model for Attribute Attribute; x is a low-level feature vector, and X dimension is Mxl, i.e. X ε R M×l L is the total number of training images; k (K) RBF Is a gaussian kernel function;
Figure BDA0002537052390000101
D i for the ith row of the dictionary D, D is a dictionary with the size of M multiplied by K, and M is the dimension of the low-level features; alpha is the attribute middle layer feature, the dimension of alphaThe degree is Kxl, K is smaller than M; alpha j A j-th column of α; sigma > 0 is the bandwidth of the gaussian kernel.
J c (W, ζ, b, α) is a high discriminant model of the Attribute Attribute constructed using the SVM learning problem of class c, where for the ith image I i And category c, if y i =c, i.e. the I-th image I i Middle object class y c Class c, SVM classification result y ci Set to 1, otherwise y ci Set to-1.
Figure BDA0002537052390000102
Figure BDA0002537052390000103
ξ ci ≥0,i＝1,…,l
Wherein Q is 1 A constant value empirically set; w is a hyperplane set w= { W c } c C is the target class c, each class
Figure BDA0002537052390000104
Figure BDA0002537052390000105
For the total category, W c Is a hyperplane of category c for adding attribute vectors of category c and other categories +.>
Figure BDA0002537052390000106
Is divided by the attribute vector of (a); class c SVM classifier relaxation variable set ζ= { ζ c } c ， ξ c Relaxation variable, ζ, of class c SVM classifier ci An i-th term and a common l-th term in relaxation variables of the class c SVM classifier; class c SVM classifier displacement parameter set b= { b c } c ，b c Is a displacement parameter of the class c SVM classifier.
L k (beta, alpha) is a genus constructed for the dual problem using support vector regression SVR formulaA highly predictive model of Attribute, wherein
Figure BDA0002537052390000107
Figure BDA0002537052390000108
Figure BDA0002537052390000109
Figure BDA00025370523900001010
Wherein ε corresponds to the tolerance of SVR to prediction errors; q (Q) 2 To set constraint error experience value, Q 2 Is constant and empirically set; k is the dimension of the attribute alpha, and the range is 1-K; for each attribute k, a hyperplane w is learned using linear SVM classification k The method comprises the steps of carrying out a first treatment on the surface of the Beta is an attribute prediction related parameter,
Figure BDA0002537052390000111
β k predicting a correlation parameter beta for a first attribute k ＝{β ki } i＝1...■ ，
Figure BDA0002537052390000112
Predicting the relevant parameter for the second property +.>
Figure BDA0002537052390000113
k is attribute dimension, and the value range is [1, K]An integer within the interval; x is x i For image I i The corresponding low-level features, i.e., row i in X; x is x j For attribute predictor f k (k.epsilon. {1, …, K }) according to x i Predicting the obtained test sample; kappa (x) i ,x j ) To be substituted into x i And x j Is a positive semi-finite kernel of (c); alpha ik Alpha is the firsti row and k column; test sample x j The value of the kth attribute of (c) is:
Figure BDA0002537052390000114
And carrying out optimization solution on the constructed optimization model of the attribute middle layer feature and the low layer feature mapping model to obtain an optimal solution of the attribute middle layer feature and the low layer feature mapping model. In the embodiment of the invention, each part can be separately solved by optimizing the solution, and the method specifically comprises the following steps:
s1, reducing the dimension of a low-level feature vector from M to K through a Principal Component Analysis (PCA) method to obtain an initialization value alpha (0) 。
S2, attribute alpha takes on value alpha (t-1) Solving the optimization problem of the reconfigurability model of the Attribute Attribute by using a PSO algorithm, namely
Figure BDA0002537052390000115
Solution D of the t-th iteration of learning dictionary D (t) 。
t is the iterative optimization times, and the initial value of t is 1.
S3, solving an optimization problem of a high discriminant model of the Attribute Attribute:
Figure BDA0002537052390000116
learning parameters W of hyperplanes belonging to different classes (t) ，b (t) 。
S4, solving an optimization problem of a high-predictability model of the Attribute Attribute:
Figure BDA0002537052390000117
learning attribute prediction related parameter beta (t) 。
S5, predicting relevant parameters beta by utilizing attributes (t) Solving an attribute prediction function f k (t) 。
S6, solving:
Figure BDA0002537052390000121
Figure BDA0002537052390000122
Figure BDA0002537052390000123
Figure BDA0002537052390000124
learning to obtain the attribute alpha of the t-th iteration (t) 。
Wherein Q is 3 Is a constant, is empirically set,
Figure BDA0002537052390000125
is the two approximate errors of the SVR predictor.
S7, judging whether t reaches the preset maximum iteration number, if so, outputting alpha (t) And (4) as the optimal solution of the attribute alpha, otherwise, t is increased by 1, and the method returns to S2.
And fourthly, constructing an SVM classifier, and training the SVM classifier by taking the middle layer characteristics of the attribute of the training image as input to obtain the trained SVM classifier for identifying the target of the ultra-narrow pulse radar ship.
The training SVM classifier is used for identifying the target of the ultra-narrow pulse radar ship and comprises the following specific steps:
and dividing the ship targets in the ultra-narrow pulse radar image to be identified to obtain a test image.
Low-level features of the test image are extracted.
And mapping the low-level features of the test image to the attribute middle-level features of the test image by using the constructed mapping model of the attribute middle-level features and the test low-level features. In particular, low-level features of the test image can be used asIs x j Substitution attribute predictor f k (k.epsilon. {1, …, K }) to obtain test sample x j I.e. the middle layer characteristics of the test image.
The middle layer characteristics in the attribute of the test image are input into a trained SVM classifier for extremely narrow pulse radar ship target identification.
The effect of the invention is further illustrated by the experiments of the following measured data:
experimental scenario and parameters: the data used in the experiment are terra sar data set, 25 pictures of each of three types of bulk cargo ship, container ship and oil ship in the data set, 20 training sets and 5 test sets.
The radar image target cutting threshold is set to t 1 ＝0.8，t 2 =0.7, and the attribute represents dimension k=10.
Experimental content and results: randomly selecting training set and test set from 25 pictures, and obtaining discrimination statistics results of table 1
Figure BDA0002537052390000131
The comparison result of the average discrimination rate of the target discrimination of the terra sar experimental scene by the method and other several existing methods in the invention is shown in table 2:
table 2 comparison of discrimination rates for different methods
Method Discrimination (percent)
Combination of geometric features 80
HOG features 73.33
Attribute method 86.67
From the data presented in table 2, it can be found that the invention achieves 86,67% discrimination rate for terra sar data, which is higher than the HOG feature and geometry feature methods. The method has the advantages that the performance of the method is superior to that of the traditional radar target identification method, and the radar image target identification rate is remarkably improved.
In summary, the above embodiments are only preferred embodiments of the present invention, and are not intended to limit the scope of the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims (6)

1. The method for identifying the target of the ship by using the ultra-narrow pulse radar based on the Attribute is characterized by comprising the following steps of:
dividing a ship target in the extremely narrow pulse radar image to obtain a divided image of the ship target as a training image;
extracting low-level features of the training image;
constructing a mapping model of Attribute middle layer characteristics and the low layer characteristics according to three properties of Attribute, namely reconfigurability, high discriminant and high predictability; mapping the low-level features of the training image to attribute middle-level features by using the mapping model;
the optimized model of the constructed attribute middle layer feature and the low layer feature mapping model is as follows:
Figure FDA0003928482490000011
wherein the method comprises the steps of
Figure FDA0003928482490000012
A reconfigurability model for Attribute Attribute; x is a low-level feature vector, and X dimension is
Figure FDA0003928482490000013
I.e. < ->
Figure FDA0003928482490000014
Figure FDA0003928482490000015
The total number of training images; k (K) RBF Is a gaussian kernel function;
Figure FDA0003928482490000016
D i for the ith row of the dictionary D, D is a dictionary with the size of M multiplied by K, and M is the dimension of the low-level features; alpha is attribute middle layer feature, alpha dimension is +.>
Figure FDA0003928482490000017
K is less than M; alpha j A j-th column of α;
Figure FDA0003928482490000018
Bandwidth for gaussian kernel;
J c (W, ζ, b, α) is a high discriminant model of the Attribute Attribute constructed using the SVM learning problem of class c, where for the ith image I i And category c, if y i =c, i.e. the I-th image I i Middle object class y c Class c, SVM classification result y ci Set to 1, otherwise y ci Set to-1
Figure FDA0003928482490000021
Figure FDA0003928482490000022
Figure FDA0003928482490000023
Wherein Q is 1 A constant value empirically set; w is a hyperplane set w= { W c } c C is the target class c, each class
Figure FDA0003928482490000024
Figure FDA0003928482490000025
For the total category, W c Is a hyperplane of category c for adding attribute vectors of category c and other categories +.>
Figure FDA0003928482490000026
Is divided by the attribute vector of (a); class c SVM classifier relaxation variable set ζ= { ζ c } c ，ξ c Relaxation variable, ζ, of class c SVM classifier ci For the ith item in the relaxation variables of class c SVM classifier, co +.>
Figure FDA0003928482490000027
An item; class c SVM classifier displacement parameter set b= { b c } c ，b c The displacement parameters of the class c SVM classifier;
L k (β, α) is a high predictive model of the Attribute constructed using the dual problem of the support vector regression SVR formula, wherein:
Figure FDA0003928482490000028
Figure FDA0003928482490000029
Figure FDA00039284824900000210
Figure FDA00039284824900000211
wherein ε corresponds to the tolerance of SVR to prediction errors; q (Q) 2 To set constraint error experience value, Q 2 Is a constant; k is the dimension of the attribute alpha, and the range is 1-K; for each attribute k, a hyperplane w is learned using linear SVM classification k The method comprises the steps of carrying out a first treatment on the surface of the Beta is an attribute prediction related parameter,
Figure FDA00039284824900000212
β k predicting the relevant parameter for the first property +.>
Figure FDA00039284824900000213
Figure FDA00039284824900000214
Predicting the relevant parameter for the second property +.>
Figure FDA00039284824900000215
k is attribute dimension, and the value range is [1, K]An integer within the interval; x is x i For image I i The corresponding low-level features, i.e., row i in X; x is x j For attribute predictor f k (k.epsilon. {1, …, K }) according to x i Predicting the obtained test sample; kappa (x) i ,x j ) To be substituted into x i And x j Is a positive semi-finite kernel of (c); alpha ik Ith row and kth column for α; test sample x j The value of the kth attribute of (c) is:
Figure FDA00039284824900000216
carrying out optimization solution on the constructed attribute middle layer feature and the optimization model of the low-level feature mapping model to obtain an optimal solution of the attribute middle layer feature and the low-level feature mapping model;
and constructing an SVM classifier, taking the attribute middle layer characteristics of the training image as input, and training the SVM classifier to obtain the trained SVM classifier for identifying the target of the ultra-narrow pulse radar ship.
2. The method of claim 1, wherein the segmenting the ship target in the ultra-narrow pulse radar image obtains a segmented image of the ship target as a training image, specifically:
and processing the ultra-narrow pulse radar image by using Radon transformation, determining a ship main shaft of the ship target, and dividing the ship target from the sea clutter background of the ultra-narrow pulse radar image to obtain a divided image of the ship target as a training image.
3. The method of claim 1, wherein the segmenting the ship target in the ultra-narrow pulse radar image to obtain a segmented image of the ship target as the training image specifically comprises the steps of:
step 101, preprocessing the ultra-narrow pulse radar image, specifically: sequentially carrying out gray scale expansion, wavelet denoising, otsu threshold segmentation, binary expansion, median filtering and binary corrosion on the extremely narrow pulse radar image to obtain a binary image, searching all connected domains in the binary image, and only reserving the largest connected domain to obtain a preprocessed image, wherein the largest connected domain is used as a ship target;
102, determining a ship main shaft of a ship target through Radon transformation;
carrying out Radon transformation on the preprocessed image, and calculating the line integral value of the ship target at each angle as R (rho, theta) 0 ) ρ is the polar diameter of the polar coordinate system, θ 0 Is the polar angle of the polar coordinate system;
the projection direction of the ship main shaft is
Figure FDA0003928482490000031
Azimuth angle of ship target is θ=θ max -90°；
Step 103, determining the length range of the ship target through Radon transformation;
rotating the preprocessed image clockwise by an angle theta to enable the ship spindle to be positioned in the horizontal direction, carrying out Radon transformation on the rotated image again, and taking all line integral values R of the 90-degree projection direction * (ρ,90 °); t to be the maximum line integral value 2 Doubling to a second threshold value, and removing line integral values smaller than the second threshold value;
R * the coordinate range of the non-zero point in (ρ,90 DEG) is [ ρ 1 ，ρ 2 ]，ρ 1 Is the lower limit of the non-zero point coordinates, ρ 2 Is the upper limit of the non-zero coordinates; then [ ρ ] 1 ，ρ 2 ]I.e. the vertical direction division range of the ship in the image domain, and determining the length range [ x ] of the ship target in the binary image according to the pixel values 1 ，x 2 ],x 1 Is the lower limit of the target length of the ship, x 2 An upper limit for the target length of the watercraft;
t 2 a preset positive constant less than 1;
and 104, dividing the ship target according to the length range of the ship target, and obtaining a divided image of the ship target as a training image.
4. The method according to claim 1, wherein the extracting low-level features of the training image is in particular:
and extracting the HOG characteristics of the directional gradient histogram of the training image.
5. The method of claim 1, wherein the optimizing the constructed optimization model of the attribute middle layer feature and the low layer feature mapping model to obtain an optimal solution of the attribute middle layer feature and the low layer feature mapping model, specifically comprises the following steps:
s1, lead toReducing the dimension of the low-level feature vector from M to K by using PCA (principal component analysis) method to obtain an initialization value alpha (0) ；
S2, attribute alpha takes on value alpha (t-1) Solving the optimization problem of the reconfigurability model of the Attribute Attribute by using a PSO algorithm, namely
Figure FDA0003928482490000041
Solution D of the t-th iteration of learning dictionary D (t) ；
t is the iterative optimization times, and the initial value of t is 1;
s3, solving an optimization problem of a high discriminant model of the Attribute Attribute:
Figure FDA0003928482490000042
learning parameters W of hyperplanes belonging to different classes (t) ，b (t) ；
S4, solving an optimization problem of a high-predictability model of the Attribute Attribute:
Figure FDA0003928482490000051
learning attribute prediction related parameter beta (t) ；
S5, predicting relevant parameters beta by utilizing the attributes (t) Solving an attribute prediction function f k (t) ；
S6, solving:
Figure FDA0003928482490000052
learning to obtain the attribute alpha of the t-th iteration (t)
Where Q3 is a constant value and,
Figure FDA0003928482490000053
two approximation errors for the SVR predictor;
S7judging whether t reaches the preset maximum iteration number, if so, outputting alpha (t) And (4) as the optimal solution of the attribute alpha, otherwise, t is increased by 1, and the method returns to S2.
6. The method of claim 1, wherein the trained SVM classifier is used for very narrow pulse radar vessel target discrimination, comprising:
dividing ship targets in the ultra-narrow pulse radar image to be identified to obtain a test image;
extracting low-level features of the test image;
mapping the low-level features of the test image to the attribute middle-level features of the test image by using the constructed mapping model of the attribute middle-level features and the low-level features;
and inputting the middle-layer characteristics of the attribute of the test image into the trained SVM classifier for extremely narrow pulse radar ship target identification.
CN202010536147.8A 2020-06-12 2020-06-12 Ultra-narrow pulse radar ship target identification method based on Attribute Active CN111767806B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010536147.8A CN111767806B (en) 2020-06-12 2020-06-12 Ultra-narrow pulse radar ship target identification method based on Attribute

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010536147.8A CN111767806B (en) 2020-06-12 2020-06-12 Ultra-narrow pulse radar ship target identification method based on Attribute

Publications (2)

Publication Number Publication Date
CN111767806A CN111767806A (en) 2020-10-13
CN111767806B true CN111767806B (en) 2023-05-05

Family

ID=72720695

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010536147.8A Active CN111767806B (en) 2020-06-12 2020-06-12 Ultra-narrow pulse radar ship target identification method based on Attribute

Country Status (1)

Country Link
CN (1) CN111767806B (en)

Families Citing this family (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113378920A (en) * 2021-06-09 2021-09-10 东南大学 Ship classification and identification method for low-resolution radar
CN114415142B (en) * 2022-01-28 2022-08-16 北京海兰信数据科技股份有限公司 Rain clutter identification method and system based on navigation radar

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2003090879A (en) * 2001-09-18 2003-03-28 Toshiba Corp Category identification method of radar targets, computer program and device therefor
CN103544502A (en) * 2013-10-29 2014-01-29 上海市城市建设设计研究总院 High-resolution remote-sensing image ship extraction method based on SVM
CN107705313A (en) * 2017-06-23 2018-02-16 北京理工大学 A kind of remote sensing images Ship Target dividing method
CN108256471A (en) * 2018-01-16 2018-07-06 北方工业大学 High-resolution synthetic aperture radar ship identification method based on double-layer classification network

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2003090879A (en) * 2001-09-18 2003-03-28 Toshiba Corp Category identification method of radar targets, computer program and device therefor
CN103544502A (en) * 2013-10-29 2014-01-29 上海市城市建设设计研究总院 High-resolution remote-sensing image ship extraction method based on SVM
CN107705313A (en) * 2017-06-23 2018-02-16 北京理工大学 A kind of remote sensing images Ship Target dividing method
CN108256471A (en) * 2018-01-16 2018-07-06 北方工业大学 High-resolution synthetic aperture radar ship identification method based on double-layer classification network

Also Published As

Publication number Publication date
CN111767806A (en) 2020-10-13

Similar Documents

Publication Publication Date Title
An et al. DRBox-v2: An improved detector with rotatable boxes for target detection in SAR images
CN110414414B (en) SAR image ship target identification method based on multilevel feature depth fusion
Wang et al. New hierarchical saliency filtering for fast ship detection in high-resolution SAR images
CN105866775B (en) A kind of offshore spilled oil detection method based on polarimetric synthetic aperture radar image
CN108256471B (en) High-resolution synthetic aperture radar ship identification method based on double-layer classification network
CN111767806B (en) Ultra-narrow pulse radar ship target identification method based on Attribute
CN110991257B (en) Polarized SAR oil spill detection method based on feature fusion and SVM
Elnemr Feature selection for texture-based plant leaves classification
Xue et al. Rethinking automatic ship wake detection: state-of-the-art CNN-based wake detection via optical images
Chen et al. Contour refinement and EG-GHT-based inshore ship detection in optical remote sensing image
Feng et al. Rapid ship detection method on movable platform based on discriminative multi-size gradient features and multi-branch support vector machine
CN110298855B (en) Sea-sky-line detection method based on Gaussian mixture model and texture analysis
Xu et al. Intelligent ship recongnition from synthetic aperture radar images
Bi et al. A hierarchical salient-region based algorithm for ship detection in remote sensing images
Chen et al. A PointNet-Based CFAR Detection Method for Radar Target Detection in Sea Clutter
CN113869119A (en) Multi-temporal SAR ship target tracking method, system, equipment and medium
Li et al. Extraction of visual texture features of seabed sediments using an SVDD approach
Zhou et al. Oil spills identification in SAR image using mRMR and SVM model
CN105512682B (en) A kind of security level identification recognition methods based on Krawtchouk square and KNN-SMO classifier
Zhai et al. Segmentation-based ship detection in harbor for SAR images
Wu et al. Object-Enhanced YOLO Networks for Synthetic Aperture Radar Ship Detection
Kurmi et al. An algorithm for various crop diseases detection and classification using leaves images
Gao et al. Feature matching for multi-beam sonar image sequence using KD-Tree and KNN search
Ning et al. Automatic target recognition of ISAR object images based on neural network
Rao et al. An improved technique of plant leaf classificaion using hybrid feature modeling

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
GR01 Patent grant
GR01 Patent grant