CN112580687A - Unsupervised classification method based on millimeter wave full-polarization SAR image - Google Patents

Unsupervised classification method based on millimeter wave full-polarization SAR image Download PDF

Info

Publication number
CN112580687A
CN112580687A CN202011319716.XA CN202011319716A CN112580687A CN 112580687 A CN112580687 A CN 112580687A CN 202011319716 A CN202011319716 A CN 202011319716A CN 112580687 A CN112580687 A CN 112580687A
Authority
CN
China
Prior art keywords
classification
polarization
unsupervised classification
scattering
unsupervised
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
CN202011319716.XA
Other languages
Chinese (zh)
Inventor
张帆
倪军
项徳良
韦立登
尹嫱
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 University of Chemical Technology
Original Assignee
Beijing University of Chemical Technology
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 University of Chemical Technology filed Critical Beijing University of Chemical Technology
Priority to CN202011319716.XA priority Critical patent/CN112580687A/en
Publication of CN112580687A publication Critical patent/CN112580687A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • 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
    • G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
    • G06F17/10—Complex mathematical operations
    • G06F17/16—Matrix or vector computation, e.g. matrix-matrix or matrix-vector multiplication, matrix factorization
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/23—Clustering techniques
    • G06F18/232—Non-hierarchical techniques
    • G06F18/2321—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
    • G06F18/23213—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions with fixed number of clusters, e.g. K-means clustering
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/10—Segmentation; Edge detection
    • G06T7/11—Region-based segmentation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/10—Segmentation; Edge detection
    • G06T7/13—Edge detection
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/10—Image acquisition modality
    • G06T2207/10032—Satellite or aerial image; Remote sensing
    • G06T2207/10044—Radar image

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • General Physics & Mathematics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Mathematical Physics (AREA)
  • General Engineering & Computer Science (AREA)
  • Pure & Applied Mathematics (AREA)
  • Artificial Intelligence (AREA)
  • Mathematical Analysis (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Mathematical Optimization (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Computational Mathematics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Evolutionary Computation (AREA)
  • Evolutionary Biology (AREA)
  • Computing Systems (AREA)
  • Probability & Statistics with Applications (AREA)
  • Algebra (AREA)
  • Databases & Information Systems (AREA)
  • Software Systems (AREA)
  • Radar Systems Or Details Thereof (AREA)
  • Image Analysis (AREA)

Abstract

The invention discloses an unsupervised classification method based on millimeter wave full polarization SAR images, which comprises the steps of firstly compressing data by using a linear compression method so that the uncalibrated data can be applied to a specific task; then, obtaining the category attribute of each pixel by utilizing a Wishart-H/A/alpha unsupervised classification algorithm; meanwhile, a self-adaptive polarization superpixel generation algorithm (Pol-ASLIC) is used for realizing a superpixel segmentation task so as to consider the spatial statistical property of the fully polarized SAR data; and finally, fusing spatial information obtained by the super pixels and unsupervised pixel label information to realize a final classification task.

Description

Unsupervised classification method based on millimeter wave full-polarization SAR image
Technical Field
The invention relates to an unsupervised classification method based on millimeter wave full-polarization SAR images, and belongs to the field of computer vision.
Background
Synthetic Aperture Radar (SAR) is a high-resolution active imaging Radar which obtains backscatter information including ground objects by using a sensor, and with the continuous improvement of the performance, the SAR of a low frequency band gradually cannot meet the demand. The millimeter wave SAR has the characteristics of small volume, high resolution, strong electronic countermeasure capability and the like, and gradually becomes an important direction for the development of radar imaging. Compared with single-polarized SAR data, the full-polarized SAR carries more ground feature scattering characteristics, and the analysis and classification of the full-polarized SAR data are important tasks for interpreting full-polarized SAR images, so that the research on the image classification of the millimeter wave full-polarized SAR is of great significance.
The classification method of the polarized SAR comprises a supervised and unsupervised classification algorithm, the supervised classification method needs actual labeled samples of ground features as a training set and combines a machine learning method to realize a supervised classification task, and ground truth cannot be easily obtained in various applications, so that the actual application value of the ground truth is limited. The unsupervised method is established based on a physical scattering mechanism and statistical distribution characteristics, so that the unsupervised method is not limited to specific surface feature classes and has greater research value and practical significance.
In the classification task, in order to take account of the spatial information of the image, the supervised classification algorithm usually considers the combination of the probability map obtained by the classifier and the image segmentation algorithm so as to take account of the spatial information and achieve a better classification effect. In the unsupervised classification task, because the probability map cannot be directly obtained through the classifier, the effective space segmentation strategy is the key for fusing the unsupervised classification label and the space information. The super-pixel segmentation algorithm can segment adjacent pixels with similar characteristics such as color, brightness, texture and the like into a plurality of super-pixels, and the super-pixels retain effective information for image segmentation and do not damage boundary information of objects in the image. Therefore, the method has great significance in obtaining the super-pixel segmentation model with pertinence by combining the statistical distribution characteristic of the polarized SAR with the traditional super-pixel segmentation algorithm.
Disclosure of Invention
The invention mainly aims to provide an unsupervised classification task realized by combining a traditional unsupervised classification algorithm and a superpixel segmentation algorithm for millimeter wave fully polarized SAR images which are not subjected to any calibration. Aiming at the statistical distribution characteristics of the polarized SAR data, an improved superpixel segmentation and fusion algorithm is used for the image segmentation task of the experiment, a method based on the combination of statistical characteristics and a physical scattering mechanism is used for realizing the basic unsupervised classification process, and meanwhile, the unsupervised classification label is fused with the image segmentation result to obtain the final unsupervised classification result.
The overall architecture of the unsupervised classification used by the present invention is shown in fig. 1. The classification task includes four processes in total: data compression and coherent T matrix extraction, superpixel segmentation, unsupervised classification and decision comprehensive classification. The treatment process comprises the following steps:
step 1: and linearly compressing the unscaled polarization data by using a data compression method, so that the full polarization SAR is in a calculable range.
Inputting millimeter wave full polarization SAR data which is not calibrated, and compressing the data to obtain the millimeter wave full polarization SAR data.
Step 2: extraction of T3And (4) matrix. Inputting the millimeter wave full polarization SAR data after data compression in the step 1, and enabling the millimeter wave full polarization SAR data after data compression to correspond to T3And (4) matrix.
And step 3: H/alpha/A classification, obtaining 16 classes of unsupervised class labels. Inputting T of step 23Matrix, unsupervised classification label of millimeter wave polarization SAR.
And 4, step 4: and carrying out statistical integration on the H/alpha/A classification result by utilizing wishart distance statistical information to obtain a more accurate polarization SAR unsupervised classification result. And inputting the unsupervised classification result of the step 3. And (5) unsupervised classification results.
And 5: obtaining an initial superpixel segmentation result by utilizing Pol-ASLIC superpixel classification algorithm, and inputting the initial superpixel segmentation result into T in the step 23And (4) matrix. And (3) outputting: initial superpixel segmentation results.
Step 6: and performing threshold fusion judgment by using a SIPV and HoM fusion method, and fusing similar initial superpixels to obtain a final superpixel segmentation result. Input of step 2T3And 5, obtaining a super pixel segmentation result by the matrix and the initial super pixel segmentation result in the step 5.
And 7: and fusing the unsupervised classification result of the fully-polarized SAR and the super-pixel segmentation result to obtain a final unsupervised classification result. And inputting the unsupervised classification result obtained in the step 4 and the super-pixel segmentation result obtained in the step 6 to obtain an unsupervised classification result.
Specifically, the implementation process of each step is as follows:
(1) data pre-processing
Electromagnetic wave signals transmitted by the radar system interact with a ground object target, so that the polarization characteristic of the electromagnetic wave is changed. The polarization scattering matrix S describes the conversion relationship between incident and scattered electromagnetic fields, and the radar incident wave and the radar scattered wave are assumed to be plane waves, and the corresponding Jones vectors are respectively EtAnd EsThen the target scattering process can be expressed as:
Figure BDA0002792483740000031
wherein G (r) is an electromagnetic wave propagation factor and reveals the change condition of amplitude and phase in the propagation process; for the scattering matrix S containing all polarization information:
Figure BDA0002792483740000032
each element SijThe complex backscattering coefficients for transmitting a signal in i-polarization and receiving a target in j-polarization are shown, H for horizontal polarization and V for vertical polarization. For the single-station backscatter case, the Jones vectors of the incident and scattered waves are described by the same set of orthogonal bases, since the transmit and receive antennas are located at the same position.
In order to fully utilize all polarization information of the multi-polarization SAR data, polarization calibration is generally applied to quantitatively analyze the correlation between different polarizations of different channels, and the multi-polarization SAR data after calibration is used for surface scattering characteristic analysis. However, the scaling process of the polarized SAR data is an extremely complex and high-cost process, and how to implement an effective image interpretation process for the unscaled polarized SAR data is a difficult problem.
The polar SAR data without internal and external scaling is usually not applicable due to the requirement of multiplication/division (inversion) of matrix in the target decomposition process of the polar data due to the large value. Therefore, before data processing, data compression is first required, using a linear compression process based on a polarization scattering matrix S:
Figure BDA0002792483740000033
in the formula, N is a compression ratio of the polarized SAR data. The compressed scattering matrix S' is expressed as a Pauli-based scattering vector
Figure BDA0002792483740000034
For the case of incoherent scattering, the scattering information is represented by the coherence matrix T
T＝<k·k*T> (5)
In the formula (I), the compound is shown in the specification,<·>representing statistical average, and obtaining a 3 x 3 coherent matrix T under a single-station polarization system3：
Figure BDA0002792483740000041
Coherence matrix T3As experimental basis data, is used for unsupervised classification and superpixel segmentation processes.
(2) Unsupervised classification process
Using the traditional wishart-H/α/a method as an unsupervised classification model, it comprises the following processes:
H/α/A unsupervised Classification based on Cloude decomposition
The Cloude decomposition decomposes the coherence matrix T into weights of three scattering mechanisms, dihedral, surface scattering and bulk scattering, with the advantage that the eigenvalues do not change with the change of the polarization basis. The decomposition features describe any scattering mechanism and thus, in conjunction with the cloud decomposition approach, achieve a finer classification of targets. The solution for H and α by Cloude decomposition is as follows:
first, the coherent matrix is decomposed into two unitary matrixes and a form of multiplication of diagonal matrixes formed by characteristic values
Figure BDA0002792483740000042
Figure BDA0002792483740000043
Wherein, U is unitary matrix, each column of which corresponds to orthogonal eigenvector of T, and λi≥0，i＝1,2,3。
From the characteristic value λiCalculating the scattering entropy H and the average scattering angle alpha:
the probability of occurrence of each scattering mechanism is defined as:
Figure BDA0002792483740000044
the polarization entropy H is defined as:
Figure BDA0002792483740000045
the scattering angle α is found from the normalized eigenvector, which is calculated as:
α＝arccos(|V(1,:)|)·P (15)
i.e. the first row vector | V (1,: | multiplied by the column vector P for the eigenvalue is the constant scattering angle α. The mean of the scattering angle α and the azimuth angle β is defined as:
Figure BDA0002792483740000046
Figure BDA0002792483740000047
the angle α is related to the scattering mechanism of the target, corresponding to surface scattering at α -0 °, volume scattering at α -45 ° and dihedral scattering at α -90 °.
The polarization entropy H (0 ≦ H ≦ 1) describes the randomness of the scatterers from isotropic (H ═ 0) to fully random scattering mechanisms (H ═ 1). When the depolarization effect of the target is weak, the eigenvector corresponding to the maximum eigenvalue of the scattering matrix is dominant, and other eigenvectors can be ignored. If the H value is large, the depolarization effect is strong. The object does not correspond to only one scattering matrix at this time, and therefore all its eigenvalues need to be taken into account. When H ═ 1, the target does not contain any polarization information, and the scattering mechanism is random noise.
Figure BDA0002792483740000051
Representing the mean scattering mechanism from dihedral to surface scattering. H and
Figure BDA0002792483740000052
the characteristics of the scattering of the object can be described effectively, and therefore for simplicity of description,
Figure BDA0002792483740000053
instead of alpha. The binary space consisting of H and a can be divided into 9 regions, and 8 active regions, ignoring the absence of scattering mechanisms.
Although polarization entropy H has some scalar characterization for describing the randomness of the scattering problem, it does not fully describe the eigenvalue ratio relationship. Thus, another characteristic parameter: the polarization anisotropy a is used for another description of the polarization characteristic value. The characteristic value is expressed as lambda1＞λ2＞λ3A polarization anisotropy A can be defined as
Figure BDA0002792483740000054
Since the eigenvalues are rotation invariant, the polarization anisotropy a is also a rotation invariant parameter. In addition to the polarization entropy H parameter, the polarization anisotropy a describes the relative magnitude of the second and third eigenvalues resulting from the eigen decomposition. With the introduction of the anisotropy parameter, it is possible to identify a second type of scattering target in a different scattering region.
Unsupervised classification based on Wishart metric criterion
The Wishart classifier is a maximum likelihood classification based on complex Wishart distributions. The probability density function of the coherence matrix T is represented as:
Figure BDA0002792483740000055
and K is a normalization factor V and is a correlation matrix of the clustering center. The expression of the normalization factor is
Figure BDA0002792483740000056
Wherein Gamma (·) is a Gamma function, and satisfies complex Wishart distribution with degree of freedom n. If the reciprocity condition is satisfied, q is 3. Defining the maximum likelihood distance between each target point and the ith class center as:
dm(<T>,Vi)＝n[ln|Vi|+Tr(Vi -1<T>)](20) wherein, ViIs the central scattering matrix of the i-th class label, namely:
Figure BDA0002792483740000061
in the formula, niIs the number of samples in the i-th class label. Pixels satisfying the following relationship are classified into the i-th class by the maximum likelihood criterion:
dm(<T>,Vi)＝dm(<T>,Vj) (22)
the Wishart classifier takes initial classification as training sample input and performs initial clustering on a ViThe basic entropy of (a) is continuously iterated and adjusted. Updating V according to the above formula and the new classification resultiAnd iteratively classifying the image again. And after the iteration times reach the requirements, stopping unsupervised classification.
In order to fully exert the advantages of the classification method based on the statistical distribution characteristic and the target classification method based on the physical scattering mechanism, Lee and the like propose an unsupervised classification method of the PolSAR target by combining the classification method and the target classification method. The method carries out primary classification by using cloud-Pottier decomposition or model decomposition, and then carries out iterative classification result by using a complex Wishart classifier. The Wishart-H/alpha/A method is used for carrying out unsupervised classification tasks on the polarization T matrix.
(3) Superpixel segmentation of millimeter wave polarization SAR
The super-pixel segmentation algorithm can reduce the speckle noise influence in the scattering image, is beneficial to storing the structural information of the image and accelerating the post-processing process of the polarized SAR image, and is widely combined with the processing processes of various polarized SAR images such as classification, change detection, target tracking and the like of the polarized SAR image.
Many optical superpixel segmentation methods have been applied to polarimetric SAR images, among them are traditional segmentation methods which generally produce compact regular superpixels in the heterogeneity and homogeneity regions of most polarimetric SAR images, but in superpixel generation, the edges of some images, especially the edges between different surface coverages, are not well preserved due to lack of edge constraints. In addition, the result of these super-pixel generation is image over-segmentation, which focuses mainly on the fine description of the local structure and content of the image. In order to fully exploit the superpixels and the initial super-segmentation results, some non-local clues of the image need to be considered, further grouping the superpixels, i.e. homogeneous small regions, into semantically meaningful large regions.
A new full polarization super pixel segmentation algorithm is used for realizing the image segmentation process of the millimeter wave polarization SAR, the algorithm firstly utilizes a directional span-driven adaptive (DSDA) window to realize the edge extraction of the polarization SAR, and the edge is used for optimizing the super pixel (Pol-ASLIC) obtained by the adaptive polarization super pixel generation algorithm. The Statistical Region Merging (SRM) image segmentation algorithm can capture the main structural information of an image, uses simple but effective statistical analysis, has high algorithm implementation efficiency, and has strong anti-noise capability. A superpixel segmentation result is initialized based on Pol-ASLIC, and a new superpixel dissimilarity measurement method based on edge penalty and polarization information is defined by using SRM and is used for obtaining a reasonable and accurate superpixel pair combination sequence. For the super-pixel fusion judgment, a polarization equilibrium measurement (HoM) is adopted to redefine a merging threshold value, so that the fusion judgment and the fusion threshold value are adaptive to the PolSAR image, and the problem of image complex parameters of the traditional SRM method is solved.
DSDA window based edge detection
The SIRV model defines the scattering vector k at Pauli basis as the product of the independent complex gaussian circular vector z and the square root of the positive random variable τ:
Figure BDA0002792483740000071
where z is the standard coherence matrix M ═ E { z · z }*TAnd the mean is 0, the variable τ is considered as spatial texture. The conditional Probability Density Function (PDF) of k is expressed as:
Figure BDA0002792483740000072
in the formula, N represents the number of independent observations of the data. Given M, it is possible to solve by maximum likelihood estimation
Figure BDA0002792483740000073
Estimation:
Figure BDA0002792483740000074
thus, the maximum likelihood estimate of M can be solved:
Figure BDA0002792483740000075
this estimate can be implemented in a fixed point recursive manner, which has a good convergence effect regardless of the initialization value of M. This parameter estimation is used in the SIRV parameter estimation and takes advantage of the DSDA window shape flexibility to accurately estimate the region centers on both sides of the center pixel.
Pol-ASLI superpixel generation optimization
In the SLIC algorithm, the original K-means clustering algorithm searches the whole image and clusters similar pixels, which takes a long time and cannot well store the local attributes of the image. And Pol-ASLIC can search limited local areas, reduce the calculation amount and simultaneously reserve the local information of the image. On the basis of Pol-ASLIC, a region search method with boundary constraint is used to redesign the segmentation model, as shown in FIG. 2.
In clustering within a limited 2S x 2S region, if edge pixels are searched, pixels outside the edge will not be traversed because they have different attributes from the center pixel due to the presence of the edge. The search area can not only keep edge information, but also accelerate the search speed clustering iteration process.
Fusion process of polarized SAR superpixels
To obtain an accurate superpixel pair merging order, a new inter-superpixel difference metric method is used for the superpixel fusion process, which utilizes both polarization information and image edge penalty terms. The SIPV product method can describe the statistical distribution category of the polarized SAR data, and the normalized coherent matrix and span data can be estimated under the condition that prior information of any texture is not set. Therefore, the SIRV distance metric algorithm can effectively implement edge detection and superpixel generation of polarized SAR in homogeneous or heterogeneous scenarios.
Suppose that
Figure BDA0002792483740000081
And
Figure BDA0002792483740000082
is a two-superpixel scene
Figure BDA0002792483740000083
And
Figure BDA0002792483740000084
a binary hypothesis test defined as:
Figure BDA0002792483740000085
the similarity of the two regions is achieved by the maximum likelihood ratio:
Figure BDA0002792483740000086
QMis equal to the minimum value of using SIRV to solve the following equation:
Figure BDA0002792483740000087
the distance measurement between the super pixels takes the polarization information of the adjacent pixels in the super pixels into consideration, and is beneficial to reducing multiplicative speckle influence. Since the formula has asymmetry, the use of a symmetric metric can be considered:
Figure BDA0002792483740000088
obtaining PolSAR by using edge detection methodAn edge intensity map of the image that can be used to compensate for the binning order of the superpixels. Suppose the edge intensity of pixel i is EiTwo super pixel scenes
Figure BDA0002792483740000089
And
Figure BDA00027924837400000810
the edge penalty term of (a) may be defined as:
Figure BDA00027924837400000811
where β is used to control the edge penalty strength, the dissimilarity of the two superpixel regions can be updated as:
Figure BDA0002792483740000091
by adding an edge penalty to the superpixel dissimilarity measure, adjacent superpixels are only entitled to preferential merging if the edge between them is weak. Therefore, the edge penalty term in the above equation can enhance the merging order of superpixels, avoiding overlap.
The blending decision of the superpixel can be realized by the following formula:
Figure BDA0002792483740000092
in the formula, ThmergeHoM fusion algorithm and asymmetry through polarization for fusion threshold
Figure BDA0002792483740000093
And (6) solving.
Compared with the traditional polarization SAR unsupervised classification process, the invention mainly aims at the unscaled full-polarization SAR millimeter wave data to realize the unsupervised classification process. And on the basis of unsupervised classification, the spatial information of the polarized SAR image is comprehensively considered. By utilizing a superpixel segmentation algorithm aiming at the fully polarized SAR, the image space information is rapidly and effectively segmented.
Drawings
Fig. 1 shows an unsupervised classification overall architecture of an uncalibrated millimeter wave fully polarized SAR image.
Fig. 2 searches for a limited area using an edge constraint optimization process.
Detailed Description
The invention relates to an unsupervised classification method based on millimeter wave full-polarization SAR images, which specifically comprises the following steps:
1) the uncalibrated polarimetric SAR data is first data compressed so that the uncalibrated data can be applied to a specific task.
2) And obtaining the class attribute of each pixel by utilizing a Wishart-H/A/alpha unsupervised classification algorithm.
3) And (3) realizing a superpixel segmentation task by using an adaptive polarization superpixel generation algorithm (Pol-ASLIC) so as to consider the spatial statistical property of the fully polarized SAR data.
4) And the spatial information obtained by the super pixels is fused with the unsupervised pixel label information to realize the final classification task.

Claims (5)

1. An unsupervised classification method based on millimeter wave full polarization SAR images is characterized in that: in the overall architecture of the unsupervised classification used, the classification task comprises four processes: data compression and coherent T matrix extraction, superpixel segmentation, unsupervised classification and decision comprehensive classification; the treatment process comprises the following steps:
step 1: linearly compressing the unscaled polarization data by a data compression method to ensure that the full polarization SAR is in a calculation range; inputting millimeter wave full polarization SAR data which is not calibrated, and compressing the data to obtain the millimeter wave full polarization SAR data;
step 2: extraction of T3A matrix; inputting the millimeter wave full polarization SAR data after data compression in the step 1, and enabling the millimeter wave full polarization SAR data after data compression to correspond to T3A matrix;
and step 3: H/alpha/A classification to obtain 16 classes of unsupervised class labels; inputting T of step 23A matrix, unsupervised classification label of millimeter wave polarization SAR;
and 4, step 4: carrying out statistical integration on the H/alpha/A classification result by utilizing wishart distance statistical information to obtain a polarized SAR unsupervised classification result;
and 5: obtaining an initial superpixel segmentation result by utilizing Pol-ASLIC superpixel classification algorithm, and inputting the initial superpixel segmentation result into T in the step 23A matrix outputting an initial superpixel segmentation result;
step 6: performing threshold fusion judgment by utilizing a SIPV and HoM fusion method, and fusing similar initial superpixels to obtain a final superpixel segmentation result; inputting T of step 23Obtaining a matrix and the initial superpixel segmentation result of the step 5 to obtain a superpixel segmentation result;
and 7: fusing an unsupervised classification result of the fully-polarized SAR and a super-pixel segmentation result to obtain a final unsupervised classification result; and inputting the unsupervised classification result obtained in the step 4 and the super-pixel segmentation result obtained in the step 6 to obtain an unsupervised classification result.
2. The unsupervised classification method based on the millimeter wave fully polarized SAR image according to claim 1 is characterized in that: the polarization scattering matrix S describes the conversion relationship between incident and scattered electromagnetic fields, and the radar incident wave and the radar scattered wave are assumed to be plane waves, and the corresponding Jones vectors are respectively EtAnd Es；
Before data processing, data compression is first required, using linear compression based on a polarization scattering matrix S.
3. The unsupervised classification method based on the millimeter wave fully polarized SAR image according to claim 2 is characterized in that: the unsupervised classification process comprises the following processes:
1) H/alpha/A unsupervised classification based on Cloude decomposition
Decomposing a coherent matrix T into weights of three scattering mechanisms of dihedral angle, surface scattering and bulk scattering by using the Cloude decomposition method, describing any one scattering mechanism by using decomposition characteristics, and realizing fine classification of targets by combining the Cloude decomposition method; the solution for H and α by Cloude decomposition is as follows:
decomposing a coherent matrix into a form of multiplying two unitary matrices by a diagonal matrix formed by eigenvalues;
from the characteristic value λiSolving scattering entropy H and an average scattering angle alpha;
because the characteristic value is rotation invariant, the polarization anisotropy A is also a rotation invariant parameter; as a complement to the polarization entropy H parameter, the polarization anisotropy a describes the relative magnitude of the second and third eigenvalues resulting from the eigen decomposition; after introducing the anisotropy parameters, identifying a second type of scattering target in different scattering areas;
2) unsupervised classification based on Wishart metric criterion
The Wishart classifier takes initial classification as training sample input and performs initial clustering on a ViContinuously iterating and adjusting the basic entropy; updating V according to classification resultiAnd iteratively classifying the images again; and after the iteration times reach the requirements, stopping unsupervised classification.
4. The unsupervised classification method based on the millimeter wave fully polarized SAR image according to claim 1 is characterized in that: to obtain the super-pixel pair merging order, a difference measurement method between super-pixels is adopted.
5. The unsupervised classification method based on the millimeter wave fully polarized SAR image according to claim 1 is characterized in that: obtaining an edge intensity image of the PolSAR image by using an edge detection method, wherein the edge intensity image is used for compensating the merging sequence of the superpixels; an edge penalty term is added to the superpixel dissimilarity measure.
CN202011319716.XA 2020-11-23 2020-11-23 Unsupervised classification method based on millimeter wave full-polarization SAR image Pending CN112580687A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202011319716.XA CN112580687A (en) 2020-11-23 2020-11-23 Unsupervised classification method based on millimeter wave full-polarization SAR image

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202011319716.XA CN112580687A (en) 2020-11-23 2020-11-23 Unsupervised classification method based on millimeter wave full-polarization SAR image

Publications (1)

Publication Number Publication Date
CN112580687A true CN112580687A (en) 2021-03-30

Family

ID=75124217

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202011319716.XA Pending CN112580687A (en) 2020-11-23 2020-11-23 Unsupervised classification method based on millimeter wave full-polarization SAR image

Country Status (1)

Country Link
CN (1) CN112580687A (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114898224A (en) * 2022-05-13 2022-08-12 北京科技大学 Change detection method based on physical scattering mechanism

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102982338A (en) * 2012-10-25 2013-03-20 西安电子科技大学 Polarization synthetic aperture radar (SAR) image classification method based on spectral clustering
CN105138970A (en) * 2015-08-03 2015-12-09 西安电子科技大学 Spatial information-based polarization SAR image classification method
CN106778821A (en) * 2016-11-25 2017-05-31 西安电子科技大学 Classification of Polarimetric SAR Image method based on SLIC and improved CNN
CN107491734A (en) * 2017-07-19 2017-12-19 苏州闻捷传感技术有限公司 Semi-supervised Classification of Polarimetric SAR Image method based on multi-core integration Yu space W ishart LapSVM
CN109145832A (en) * 2018-08-27 2019-01-04 大连理工大学 Polarimetric SAR image semisupervised classification method based on DSFNN Yu non local decision
US20200019817A1 (en) * 2018-07-11 2020-01-16 Harbin Institute Of Technology Superpixel classification method based on semi-supervised k-svd and multiscale sparse representation

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102982338A (en) * 2012-10-25 2013-03-20 西安电子科技大学 Polarization synthetic aperture radar (SAR) image classification method based on spectral clustering
CN105138970A (en) * 2015-08-03 2015-12-09 西安电子科技大学 Spatial information-based polarization SAR image classification method
CN106778821A (en) * 2016-11-25 2017-05-31 西安电子科技大学 Classification of Polarimetric SAR Image method based on SLIC and improved CNN
CN107491734A (en) * 2017-07-19 2017-12-19 苏州闻捷传感技术有限公司 Semi-supervised Classification of Polarimetric SAR Image method based on multi-core integration Yu space W ishart LapSVM
US20200019817A1 (en) * 2018-07-11 2020-01-16 Harbin Institute Of Technology Superpixel classification method based on semi-supervised k-svd and multiscale sparse representation
CN109145832A (en) * 2018-08-27 2019-01-04 大连理工大学 Polarimetric SAR image semisupervised classification method based on DSFNN Yu non local decision

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
韦立登 等: "毫米波全极化SAR 影像无监督分类方法", 第七届高分辨率对地观测学术年会论文集, pages 1 - 25 *

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114898224A (en) * 2022-05-13 2022-08-12 北京科技大学 Change detection method based on physical scattering mechanism

Similar Documents

Publication Publication Date Title
Yang et al. A novel CNN-based detector for ship detection based on rotatable bounding box in SAR images
CN108765458B (en) Sea surface target scale self-adaptive tracking method of high-sea-condition unmanned ship based on correlation filtering
Yang et al. CNN-based polarimetric decomposition feature selection for PolSAR image classification
Xiang et al. Adaptive superpixel generation for polarimetric SAR images with local iterative clustering and SIRV model
CN107102302B (en) Radar target fusion detection method based on non-Gaussian parameters
Xie et al. PolSAR image classification via D-KSVD and NSCT-domain features extraction
Zhang et al. Polarimetric HRRP recognition based on ConvLSTM with self-attention
CN106446935A (en) Kernel sparse representation and spatial constraint-based polarimetric SAR image classification method
Xue et al. Dynamic coarse‐to‐fine ISAR image blind denoising using active joint prior learning
Li et al. Object-oriented classification of polarimetric SAR imagery based on statistical region merging and support vector machine
CN104123563B (en) Cloude characteristic decomposition based polarimetric SAR (Synthetic Aperture Radar) image non-supervision classification method
Quan et al. Scattering feature-driven superpixel segmentation for polarimetric SAR images
CN112580687A (en) Unsupervised classification method based on millimeter wave full-polarization SAR image
Liu et al. Simultaneous diagonalization of Hermitian matrices and its application in PolSAR ship detection
Wei et al. Learning-based split unfolding framework for 3-D mmW radar sparse imaging
CN112215832B (en) SAR trail image quality evaluation and self-adaptive detection parameter adjustment method
Zhang et al. Learning scattering similarity and texture-based attention with convolutional neural networks for PolSAR image classification
Chen et al. Experimental feature-based SAR ATR performance evaluation under different operational conditions
Bi et al. l 2, p-norm sequential bilateral 2DPCA: a novel robust technology for underwater image classification and representation
Fernández-Michelli et al. Unsupervised classification algorithm based on EM method for polarimetric SAR images
CN110363105B (en) Method for inhibiting speckle of fully-polarized SAR image
Zhang et al. Region-based classification by combining MS segmentation and MRF for POLSAR images
Mishra et al. Bistatic sar atr
Yang et al. An information-geometric optimization method for ship detection in SAR images
Zhu et al. Target classification by conventional radar based on bispectrum and deep CNN

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