CN109345537B - SAR image segmentation method based on high-order multi-scale CRF semi-supervision - Google Patents

SAR image segmentation method based on high-order multi-scale CRF semi-supervision Download PDF

Info

Publication number
CN109345537B
CN109345537B CN201810943115.2A CN201810943115A CN109345537B CN 109345537 B CN109345537 B CN 109345537B CN 201810943115 A CN201810943115 A CN 201810943115A CN 109345537 B CN109345537 B CN 109345537B
Authority
CN
China
Prior art keywords
scale
synthetic aperture
aperture radar
sar image
pixel point
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
CN201810943115.2A
Other languages
Chinese (zh)
Other versions
CN109345537A (en
Inventor
张鹏
江银银
李明
宋婉莹
谭啸峰
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.)
Xidian University
Original Assignee
Xidian University
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 Xidian University filed Critical Xidian University
Priority to CN201810943115.2A priority Critical patent/CN109345537B/en
Publication of CN109345537A publication Critical patent/CN109345537A/en
Application granted granted Critical
Publication of CN109345537B publication Critical patent/CN109345537B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/10—Segmentation; Edge detection
    • 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
    • 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
    • G06T2207/20—Special algorithmic details
    • G06T2207/20048—Transform domain processing
    • G06T2207/20064—Wavelet transform [DWT]

Landscapes

  • Engineering & Computer Science (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Radar Systems Or Details Thereof (AREA)

Abstract

A Synthetic Aperture Radar (SAR) image segmentation method based on semi-supervision of a high-order multi-scale Conditional Random Field (CRF) is disclosed. The method comprises the following implementation steps: (1) inputting a Synthetic Aperture Radar (SAR) image; (2) performing wavelet transformation on the image; (3) obtaining a feature vector of a high-order multi-scale conditional random field CRF; (4) obtaining the local class conditional probability of the high-order multi-scale conditional random field CRF; (5) initially segmenting a Synthetic Aperture Radar (SAR) image; (6) calculating the edge probability of each pixel point; (7) calculating the combined posterior marginal probability of each pixel point; (8) calculating the posterior marginal probability of each pixel point; (9) segmenting a Synthetic Aperture Radar (SAR) image; (10) and finishing the division. According to the invention, the high-order potential energy and the inter-scale potential energy of each pixel point are considered when the posterior marginal probability is calculated, the spatial context structure information is fully utilized, and the accuracy of the segmentation result is improved.

Description

SAR image segmentation method based on high-order multi-scale CRF semi-supervision
Technical Field
The invention belongs to the technical field of image processing, and further relates to a Synthetic Aperture Radar (SAR) (synthetic Aperture radar) image segmentation method based on semi-supervision of a high-order multi-scale Conditional Random Field (CRF) (conditional Random fields) in the technical field of radar image processing. The method can be used for segmenting the SAR image.
Background
Synthetic aperture radar SAR is a high resolution imaging radar. The synthetic aperture radar SAR image interpretation technology is required to be used for support in civil and military, the synthetic aperture radar SAR image segmentation is an important link of the synthetic aperture radar SAR image interpretation technology, and the synthetic aperture radar SAR image segmentation can provide integral structure information of the synthetic aperture radar SAR image, so that the application of the synthetic aperture radar SAR system in many fields, such as geological exploration, environment monitoring and the like, is promoted. The random field model is regarded as an important means for processing the segmentation problem of the synthetic aperture radar SAR image, and has the advantage that the spatial correlation among pixels can be considered in the image classification process.
In the patent document "SAR image segmentation method based on semantic conditional Random field model" (application number: 201611237232.4, application publication number: CN106683109A) applied by the university of Western's electronics science and technology, a method for segmenting images by using a semantic conditional Random field CRF (conditional Random fields) model is proposed. According to the method, according to a regional diagram of an SAR image, the SAR image is divided into a mixed aggregation structure ground object subspace, a structure region subspace and a homogeneous region subspace. And extracting features of the feature subspace of the mixed aggregation structure by adopting a bag-of-words model, and segmenting by using an AP clustering method. And then constructing a semantic conditional random field model to segment the structural region subspace and the homogeneous region subspace. And merging the segmentation results of the ground object subspace, the structural region subspace and the homogeneous region subspace of the mixed aggregation structure to obtain the segmentation result of the SAR image. The method has the disadvantage that the accuracy of the segmentation of the SAR image is not high.
The patent technology owned by the university of electronic technology of Xian "polarized SAR image segmentation method based on deconvolution network and sparse classification" (application number: 201510953343.4, publication number: CN105608692B) proposes image segmentation by using deconvolution network and a dilution classification method. The patent technology trains the samples of the aggregation area by using a deconvolution network, and constructs a similarity matrix by using a filter obtained by training. And then, segmenting the aggregation region by utilizing the obtained similarity matrix, and finally merging segmentation results of the aggregation region, the homogeneous region and the structural region to obtain the segmented SAR image. The method has the disadvantages that a large amount of training data is needed to learn the parameters of the deconvolution network, the method is only limited to processing the segmentation problem of the supervised synthetic aperture radar image, and the method is difficult to deal with the situation that only a small amount of training data exists in practical application.
Disclosure of Invention
The invention aims to provide an SAR image segmentation method based on high-order multi-scale conditional random field semi-supervision aiming at the defects of the prior art.
The specific idea for realizing the purpose of the invention is as follows: firstly, feature extraction is carried out on a synthetic aperture radar SAR image to be segmented, and then the image is initially segmented. And then, the posterior marginal probability of the SAR image is calculated step by utilizing the initially segmented SAR image. And then, parameters are estimated by using an iterative condition estimation method ICE, the posterior marginal probability is maximized to obtain a segmentation result of the image, and then multiple iterations are carried out. And in each iteration, the current segmentation result is used as the basis of the next iteration segmentation, so that the segmentation result after multiple iterations approaches the optimal segmentation of the synthetic aperture radar image SAR.
The steps of the invention comprise:
(1) inputting a Synthetic Aperture Radar (SAR) image;
(2) performing wavelet transformation on an input Synthetic Aperture Radar (SAR) image;
(3) obtaining a feature vector of a high-order multi-scale conditional random field CRF:
(3a) calculating the edge strength between each pixel point in each scale and each pixel point in the adjacent domain system of the image after wavelet transformation by using an exponential weighted average ratio operator;
(3b) sliding a window with the radius of 5 pixel points at intervals in the wavelet-transformed image, performing sliding window operation on the wavelet-transformed image, and calculating histogram features of all pixel values in each sliding window;
(3c) sliding a window with the radius of 7 pixel points at intervals in the wavelet-transformed image, performing sliding window operation on the wavelet-transformed image, and calculating all gray level co-occurrence matrix texture features in each sliding window;
(3d) forming local feature vectors by using all histogram features and gray level co-occurrence matrix texture features of the wavelet-transformed image;
(3e) forming a feature vector of a high-order multi-scale conditional random field CRF by using all local feature vectors and edge intensities of the wavelet-transformed image;
(4) obtaining the local class conditional probability of the high-order multi-scale conditional random field CRF:
inputting the feature vector of the high-order multi-scale conditional random field CRF into a multi-class Support Vector Machine (SVM) classifier, and outputting the classifier as the local class conditional probability of the high-order multi-scale conditional random field CRF;
(5) initially segmenting a Synthetic Aperture Radar (SAR) image:
initially segmenting the SAR image by utilizing a method of maximizing the local class conditional probability;
(6) calculating the edge probability of each pixel:
calculating the edge probability of each pixel point high-order multi-scale conditional random field CRF in the current segmented synthetic aperture radar SAR image by using a mean field estimation method;
(7) calculating the combined posterior marginal probability of each pixel point:
calculating the joint posterior marginal probability of each pixel point high-order multi-scale conditional random field CRF in the current segmented synthetic aperture radar SAR image by using the marginal probability of each pixel point by adopting a bottom-up recursion method;
(8) calculating the posterior marginal probability of each pixel point:
calculating the posterior marginal probability of each pixel point of the currently segmented synthetic aperture radar SAR image in a high-order multi-scale conditional random field CRF by using a top-down recursion method and the combined posterior marginal probability of each pixel point;
(9) segmenting the SAR image:
performing parameter estimation by using an Iterative Condition Estimation (ICE) method, and segmenting a Synthetic Aperture Radar (SAR) image;
(10) judging whether the current iteration number is 10, if so, executing the step (11); otherwise, adding 1 to the current iteration number and then executing the step (6);
(11) and (5) finishing the segmentation:
and finishing the segmentation of the SAR image to obtain a segmentation result of the SAR image semi-supervised by the high-order multi-scale conditional random field CRF.
Compared with the prior art, the invention has the following advantages:
firstly, the method for calculating the posterior marginal probability of the high-order multi-scale conditional random field CRF of each pixel point is adopted, so that the problem of low accuracy in the process of segmenting the SAR image in the prior art is solved, the high-order potential energy and the inter-scale potential energy of each pixel point are considered in the process of calculating the posterior marginal probability, the spatial context structure information is fully utilized, and the accuracy of the segmentation result is improved.
Secondly, because the invention estimates the model parameter by using the iterative condition estimation ICE method, and adopts the method of multiple iterations to obtain the final segmentation result, the problem that the prior art lacks training data when segmenting the SAR image is overcome, so that the invention uses the current segmentation result as the basis of the next iteration segmentation when segmenting, and the segmentation result after multiple iterations approaches the optimal segmentation of the SAR image, thereby reducing the requirement on the quantity of training data.
Drawings
FIG. 1 is a flow chart of the present invention;
FIG. 2 is a simulation diagram of a synthetic aperture radar SAR image combined by three homogeneous areas by using the method of the present invention;
FIG. 3 is a simulation diagram of synthetic aperture radar SAR images of a river channel respectively by adopting the method of the invention and the generalized conditional random field CRF method;
FIG. 4 is a diagram of simulation results of ESAR images of a Parvowalen area using the method of the present invention and a generalized conditional random field CRF method.
Detailed Description
The invention is further described below with reference to the accompanying drawings.
The steps of the present invention will be further described with reference to fig. 1.
Step 1, inputting a synthetic aperture radar SAR image.
And 2, performing wavelet transformation on the input synthetic aperture radar SAR image.
And 3, obtaining a feature vector of the high-order multi-scale conditional random field CRF.
And calculating the edge strength between each pixel point in each scale and each pixel point in the adjacent domain system of the image after wavelet transformation by using an exponential weighted average ratio operator.
The formula of the exponential weighted average ratio operator is as follows:
Figure BDA0001769551240000041
wherein the content of the first and second substances,
Figure BDA0001769551240000042
the method is characterized in that the method represents the edge strength between the s-th pixel point of the multi-scale synthetic aperture radar SAR image in the nth scale and the t-th pixel point of the multi-scale synthetic aperture radar SAR image in a neighborhood system, the value range of n is {0,1,2}, exp (·) represents exponential operation with a natural constant as a base,
Figure BDA0001769551240000043
representing the pixel value of the s-th pixel point of the multi-scale synthetic aperture radar SAR image in the nth scale,
Figure BDA0001769551240000044
and expressing the pixel value of the t-th pixel point of the multi-scale synthetic aperture radar SAR image in the nth scale.
And sliding the wavelet-transformed image by taking one pixel point as an interval in the wavelet-transformed image by using a window with the radius of 5 pixel points, and calculating the histogram characteristics of all pixel values in each sliding window.
The specific steps for calculating the histogram features of all pixel values in each sliding window are as follows:
step 1, calculating the average value of all pixel values in each sliding window according to the following formula:
Figure BDA0001769551240000051
wherein, mumRepresents the mean value of all pixel values in the mth sliding window, sigma represents the summation operation, i represents the serial number of the pixel point row in the mth sliding window, j represents the serial number of the pixel point column in the mth sliding window, yijThe pixel value, n, of the ith row and the jth column of pixel points in the mth sliding windowmRepresenting the total number of all pixel points in the mth sliding window.
And step 2, calculating the variance of all pixel values in each sliding window according to the following formula:
Figure BDA0001769551240000052
wherein the content of the first and second substances,
Figure BDA0001769551240000053
representing the variance of all pixel values within the mth sliding window.
And 3, calculating skewness of all pixel values in each sliding window according to the following formula:
Figure BDA0001769551240000054
wherein phimRepresenting the skewness of all pixel values within the mth sliding window.
And 4, calculating the peak states of all pixel values in each sliding window according to the following formula:
Figure BDA0001769551240000055
wherein, KmRepresenting the peak state of all pixel values in the mth sliding window.
And 5, calculating the entropy of all pixel values in each sliding window according to the following formula:
Figure BDA0001769551240000056
wherein, TmRepresenting the entropy of all pixel values within the mth sliding window.
And 6, forming the histogram characteristics of the sliding window by the mean value, the variance, the skewness, the kurtosis and the entropy of all pixel values in each sliding window.
And sliding the wavelet-transformed image by taking one pixel point as an interval in the wavelet-transformed image by using a window with the radius of 7 pixel points, and calculating all gray level co-occurrence matrix texture characteristics in each sliding window.
The specific steps for calculating all gray level co-occurrence matrix texture features in each sliding window are as follows:
step 1, forming a pixel pair by one pixel and adjacent pixels in each sliding window, dividing the occurrence frequency of the same pixel pair in each sliding window by the occurrence frequency of other pixel pairs in the sliding window, and forming a gray level co-occurrence matrix of the sliding window by all quotient values.
And 2, calculating the contrast of all elements in each gray level co-occurrence matrix according to the following formula:
Figure BDA0001769551240000061
wherein the content of the first and second substances,
Figure BDA0001769551240000062
is shown as
Figure BDA0001769551240000063
The contrast of all elements in the gray level co-occurrence matrix, u represents the serial number of the row in the gray level co-occurrence matrix, v represents the serial number of the column in the gray level co-occurrence matrix,
Figure BDA0001769551240000064
is shown as
Figure BDA0001769551240000065
And the element values of the ith row and the vth column in the gray level co-occurrence matrix.
And 3, calculating the correlation of all elements in each gray level co-occurrence matrix according to the following formula:
Figure BDA0001769551240000066
wherein the content of the first and second substances,
Figure BDA0001769551240000067
is shown as
Figure BDA0001769551240000068
The correlation of all elements in the gray level co-occurrence matrix.
And 4, calculating the homogeneity of all elements in each gray level co-occurrence matrix according to the following formula:
Figure BDA0001769551240000069
wherein the content of the first and second substances,
Figure BDA00017695512400000610
is shown as
Figure BDA00017695512400000611
Homogeneity of all elements within the gray level co-occurrence matrix.
And 5, forming the characteristic texture characteristics of the gray level co-occurrence matrix by the mean value, the contrast, the correlation and the homogeneity of elements in each gray level co-occurrence matrix.
And forming local feature vectors by using all histogram features and gray level co-occurrence matrix texture features of the wavelet-transformed image.
And forming the feature vector of the high-order multi-scale conditional random field CRF by using all local feature vectors and edge intensities of the wavelet-transformed image.
And 4, obtaining the local class conditional probability of the high-order multi-scale conditional random field CRF.
And inputting the feature vector of the high-order multi-scale conditional random field CRF into a multi-class Support Vector Machine (SVM) classifier, and outputting the classifier as the local class conditional probability of the high-order multi-scale conditional random field CRF.
And 5, preliminarily segmenting the SAR image.
And initially segmenting the SAR image by utilizing a method for maximizing the local class conditional probability.
And 6, calculating the edge probability of each pixel point.
And calculating the edge probability of each pixel point high-order multi-scale conditional random field CRF in the current segmented synthetic aperture radar SAR image by using a mean field estimation method.
The mean field estimation method comprises the following specific steps:
step 1, taking the local class conditional probability as an initial value of local mean value field energy of the segmented multi-scale synthetic aperture radar SAR image.
Step 2, calculating local radical energy between each pixel point in the segmented multi-scale synthetic aperture radar SAR image and each pixel point in the neighborhood system according to the following formula:
Figure BDA0001769551240000071
wherein Q is{s,t}Expressing the local radical energy alpha between the s-th pixel point in the segmented multi-scale synthetic aperture radar SAR image and the t-th pixel point in the neighborhood systemHExpressing the correlation parameter between the s-th pixel point of the segmented multi-scale synthetic aperture radar SAR image and the t-th pixel point in the horizontal neighborhood system, the initial value of the parameter is 1, the parameter value is updated by using an iterative condition estimation ICE method in step (9), the horizontal neighborhood system expresses the set of two pixel points which are adjacent to the s-th pixel point left and right,
Figure BDA0001769551240000072
expressing a constant, and when the segmented multi-scale synthetic aperture radar SAR image is in the category of the s pixel point in the n scale
Figure BDA0001769551240000073
And the category of the t-th pixel point in the neighborhood system
Figure BDA0001769551240000074
When the phase of the mixture is the same as the phase of the mixture,
Figure BDA0001769551240000075
when the label of the s pixel point of the segmented multi-scale synthetic aperture radar SAR image in the n scale
Figure BDA0001769551240000076
Label of t-th pixel point in its neighborhood system
Figure BDA0001769551240000077
At a different time,
Figure BDA0001769551240000078
Figure BDA0001769551240000079
representing an arbitrary symbol, e representing belonging to a symbol, NHA horizontal neighborhood system, alpha, representing the s-th pixel in the segmented SAR imageVExpressing the correlation parameter between the s-th pixel point in the segmented multi-scale synthetic aperture radar SAR image and the t-th pixel point in the vertical neighborhood system, wherein the initial value of the parameter is 1, updating the parameter value by using an iterative condition estimation ICE method in step (9), and NVAnd the vertical neighborhood system represents the vertical neighborhood system of the s-th pixel point in the segmented multi-scale synthetic aperture radar SAR image, and the vertical neighborhood system represents the set of two pixel points which are vertically adjacent to the s-th pixel point.
And 3, calculating the local mean field energy of each pixel point of the segmented multi-scale synthetic aperture radar SAR image in each scale according to the following formula:
Figure BDA0001769551240000081
wherein the content of the first and second substances,
Figure BDA0001769551240000086
representing the local mean value field energy, N, of the s-th pixel point of the segmented multi-scale synthetic aperture radar SAR image in the N-th scalesThe method comprises the steps of representing a neighborhood system of an s-th pixel point in a segmented multi-scale synthetic aperture radar SAR image, representing a set of four pixel points which are adjacent to the s-th pixel point in the vertical and horizontal directions, k representing the number of classes in the neighborhood system, L representing the total number of the segmented multi-scale synthetic aperture radar SAR image classes, L representing the number of the segmented multi-scale synthetic aperture radar SAR image classes, and elRepresenting the class of the s-th pixel point in the segmented multi-scale synthetic aperture radar SAR image, ekRepresents the category of the t-th pixel point in the neighborhood system,
Figure BDA0001769551240000087
the local mean field energy of the t-th pixel point in the nth scale of the neighborhood system is represented,
Figure BDA0001769551240000082
and values are taken in a synthetic aperture radar SAR image category set according to category numbers.
And 4, executing the step 3 twice to obtain the local mean field energy of each pixel point of the segmented multi-scale synthetic aperture radar SAR image in each scale, and taking the local mean field energy as the estimated value of the edge probability of each pixel point of the segmented multi-scale synthetic aperture radar SAR image in each scale.
And 5, calculating the inter-scale potential energy of each pixel point of the segmented multi-scale synthetic aperture radar SAR image in each scale according to the following formula:
Figure BDA0001769551240000083
wherein the content of the first and second substances,
Figure BDA0001769551240000084
representing the s pixel point of the segmented multi-scale synthetic aperture radar SAR image in the n scale and the n +1 scale
Figure BDA0001769551240000085
The inter-scale potential energy of each pixel point, eta represents the inter-scale model parameter of the segmented multi-scale synthetic aperture radar SAR image, the initial value of the parameter is 1, the parameter value is updated by using an iterative condition estimation ICE method in the step (9),
Figure BDA0001769551240000091
representing the segmented multi-scale synthetic aperture radar SAR image in the (n + 1) th scale
Figure BDA0001769551240000092
The category of each pixel.
And 6, calculating the conditional prior marginal probability of each pixel point of the segmented multi-scale synthetic aperture radar SAR image in each scale according to the following formula:
Figure BDA0001769551240000093
wherein the content of the first and second substances,
Figure BDA0001769551240000094
and expressing the conditional prior marginal probability of the s-th pixel point of the segmented multi-scale synthetic aperture radar SAR image in the n-th scale.
And 7, calculating unitary potential energy of the conditional random field CRF model of each pixel point of the segmented multi-scale synthetic aperture radar SAR image in each scale according to the following formula:
Figure BDA0001769551240000095
wherein the content of the first and second substances,
Figure BDA0001769551240000099
expressing the unary potential energy of the conditional random field CRF model of the s-th pixel point of the segmented multi-scale synthetic aperture radar SAR image in the n-th scale, and log expressing the logarithm operation with 10 as the base,
Figure BDA00017695512400000910
and expressing the local class conditional probability of the s-th pixel point of the segmented multi-scale synthetic aperture radar SAR image in the n-th scale.
And 8, calculating the scattering statistics of each pixel point of the segmented multi-scale synthetic aperture radar SAR image in each scale according to the following formula:
Figure BDA0001769551240000096
wherein the content of the first and second substances,
Figure BDA0001769551240000097
representing scattering statistics of s pixel point of segmented multi-scale synthetic aperture radar SAR image in n scale, betalThe shape parameter of the generalized Gamma Gamma distribution model is represented, the parameter value is calculated by a logarithmic cumulant MoLC method, and lambda islIndex shape parameters representing Gamma Gamma distribution model, the parameter values are calculated by a logarithmic cumulant MoLC method,
Figure BDA0001769551240000098
representing the strength alpha of the s-th pixel point of the segmented multi-scale synthetic aperture radar SAR image in the n-th scalelAnd the scale parameter represents a scale parameter of the generalized Gamma distribution model, the parameter value is calculated by a logarithm cumulant MoLC method, and Gamma (·) represents a Gamma function.
And 9, calculating unitary potential energy of the generalized conditional random field GCRF model of each pixel point of the segmented multi-scale synthetic aperture radar SAR image in each scale according to the following formula:
Figure BDA0001769551240000101
wherein the content of the first and second substances,
Figure BDA0001769551240000102
and representing unary potential energy of the generalized conditional random field GCRF model of the s-th pixel point of the segmented multi-scale synthetic aperture radar SAR image in the n-th scale.
And step 10, sliding a window with the radius of 7 pixel points at intervals in the segmented synthetic aperture radar SAR image, and performing one-time window sliding operation on the segmented multi-scale synthetic aperture radar SAR image.
And 11, calculating the probability of each category in each sliding window according to the following formula:
Figure BDA0001769551240000103
wherein the content of the first and second substances,
Figure BDA0001769551240000104
is shown as
Figure BDA00017695512400001012
The probability of the ith class within the sliding window,
Figure BDA0001769551240000105
is shown as
Figure BDA00017695512400001013
The number of the ith class in a sliding window,
Figure BDA0001769551240000106
is shown as
Figure BDA00017695512400001014
Single radical C in a sliding window1The total number of (c).
And 12, calculating the probability of each category adjacent to other categories in each sliding window according to the following formula:
Figure BDA0001769551240000107
wherein the content of the first and second substances,
Figure BDA0001769551240000108
is shown as
Figure BDA00017695512400001015
Probability that the ith category is adjacent to the epsilon category in each sliding window, pi represents the product operation, tau represents the type of a binary group, the binary group represents a pair of pixel points which are vertically, horizontally and adjacently connected with one pixel point,
Figure BDA0001769551240000109
is shown as
Figure BDA00017695512400001016
A left binary group C consisting of the first class and the epsilon class in the sliding window2(1) The left binary group represents a pair of pixel points left adjacent to the first category,
Figure BDA00017695512400001010
is shown as
Figure BDA00017695512400001017
Right binary group C formed by the first category and the epsilon category in the sliding window2(2) Right binary radical represents a pair of pixel points right adjacent to the l-th class,
Figure BDA00017695512400001011
is shown as
Figure BDA00017695512400001018
A lower binary group C consisting of the first category and the epsilon second category in the sliding window2(3) The lower binary group represents a pair of pixel points adjacent to the first category,
Figure BDA0001769551240000111
is shown as
Figure BDA00017695512400001112
The upper binary group C formed by the first category and the epsilon category in the sliding window2(4) The upper binary group represents a pair of pixel points adjacent to the first class,
Figure BDA0001769551240000112
is shown as
Figure BDA00017695512400001113
The total number of left binary clusters of the ith class within the individual sliding window,
Figure BDA0001769551240000113
is shown as
Figure BDA00017695512400001114
The total number of the first category of right binary radicals in each sliding window,
Figure BDA0001769551240000114
is shown as
Figure BDA00017695512400001115
The total number of binary groups under the ith class in each sliding window,
Figure BDA0001769551240000115
is shown as
Figure BDA00017695512400001116
Total number of binary groups on the l-th class within each sliding window.
And step 13, calculating the high-order potential energy of each pixel point in each sliding window according to the following formula:
Figure BDA0001769551240000116
wherein the content of the first and second substances,
Figure BDA0001769551240000117
is shown as
Figure BDA00017695512400001117
The high-order potential energy of the s-th pixel point in each sliding window.
And 14, calculating the mean field energy of each pixel point of the segmented multi-scale synthetic aperture radar SAR image in each scale according to the following formula:
Figure BDA0001769551240000118
wherein the content of the first and second substances,
Figure BDA0001769551240000119
and representing the mean field energy of the s-th pixel point of the segmented multi-scale synthetic aperture radar SAR image in the nth scale.
And step 15, calculating the edge probability of each pixel point of the segmented multi-scale synthetic aperture radar SAR image in each scale according to the following formula:
Figure BDA00017695512400001110
wherein the content of the first and second substances,
Figure BDA00017695512400001111
and representing the marginal probability of the s-th pixel point of the segmented multi-scale synthetic aperture radar SAR image in the n-th scale.
And 7, calculating the joint posterior marginal probability of each pixel point.
And calculating the joint posterior marginal probability of the high-order multi-scale conditional random field CRF of each pixel point in the current segmented synthetic aperture radar SAR image by adopting a bottom-to-top recursion method and utilizing the marginal probability of each pixel point.
The bottom-up recursion method comprises the following specific steps:
step 1, taking the edge probability of each pixel point of the segmented multi-scale synthetic aperture radar SAR image in the 0 th scale as the local posterior edge probability of each pixel point of the segmented multi-scale synthetic aperture radar SAR image in the 0 th scale.
Step 2, calculating the local posterior marginal probability of each pixel point of the segmented multi-scale synthetic aperture radar SAR image in each scale according to the following formula:
Figure BDA0001769551240000121
wherein the content of the first and second substances,
Figure BDA0001769551240000122
the local posterior marginal probability of the s-th pixel point of the segmented multi-scale synthetic aperture radar SAR image in the nth scale is represented,srepresenting the descendant of the s-th pixel in the multi-scale synthetic aperture radar SAR image, representing the s-th pixel point of the multi-scale synthetic aperture radar SAR image in the scale lower than the n-th scale,
Figure BDA0001769551240000123
representing the class of the q generation in the filial generation of the s pixel point of the segmented multi-scale synthetic aperture radar SAR image,
Figure BDA0001769551240000124
representing the posterior marginal probability of the q generation in the s pixel point descendant of the segmented multi-scale synthetic aperture radar SAR image,
Figure BDA0001769551240000125
representing the q-th generation conditional prior marginal probability in the s-th pixel point descendant of the segmented multi-scale synthetic aperture radar SAR image,
Figure BDA0001769551240000126
and expressing the prior marginal probability of the q generation in the s pixel point descendant of the segmented multi-scale synthetic aperture radar SAR image.
And 3, executing the step 2 twice to obtain the local posterior marginal probability of each pixel point of the segmented multi-scale synthetic aperture radar SAR image in each scale.
And 4, calculating the joint posterior marginal probability of each pixel point and the parent pixel point of the segmented multi-scale synthetic aperture radar SAR image in each scale according to the following formula:
Figure BDA0001769551240000127
wherein the content of the first and second substances,
Figure BDA0001769551240000128
expressing the combined posterior marginal probability of the s-th pixel point of the segmented multi-scale synthetic aperture radar SAR image in the nth scale and the parent pixel point thereof, wherein the parent represents the s-th pixel point of the multi-scale synthetic aperture radar image SAR in the nth scale,
Figure BDA0001769551240000129
represents a parent of the s-th pixel in the multi-scale synthetic aperture radar SAR image,
Figure BDA00017695512400001210
representing the nth of the segmented multi-scale synthetic aperture radar SAR image in the (n + 1) th scale
Figure BDA00017695512400001211
The prior edge probability of an individual pixel point,
Figure BDA00017695512400001212
representing the nth of the segmented multi-scale synthetic aperture radar SAR image in the (n + 1) th scale
Figure BDA0001769551240000131
A category of individual pixels.
And 8, calculating the posterior marginal probability of each pixel point.
And calculating the posterior marginal probability of the high-order multi-scale conditional random field CRF of each pixel point of the currently segmented synthetic aperture radar SAR image by adopting a top-down recursion method and utilizing the combined posterior marginal probability of each pixel point.
The top-down recursion method comprises the following specific steps:
step 1, setting an initial value of posterior marginal probability of each pixel point of the segmented multi-scale synthetic aperture radar SAR image in the 2 nd scale as the local posterior marginal probability of each pixel point of the segmented multi-scale synthetic aperture radar SAR image in the 2 nd scale.
Step 2, calculating the posterior marginal probability of each pixel point of the segmented multi-scale synthetic aperture radar SAR image in each scale according to the following formula:
Figure BDA0001769551240000132
wherein the content of the first and second substances,
Figure BDA0001769551240000135
representing the posterior marginal probability of the s-th pixel point of the segmented multi-scale synthetic aperture radar SAR image in the n-th scale,
Figure BDA0001769551240000133
representing the segmented multi-scale synthetic aperture radar SAR image in the (n + 1) th scale
Figure BDA0001769551240000134
Posterior marginal probability of each pixel point.
And 3, executing the step 2 twice to obtain the posterior marginal probability of each pixel point of the segmented multi-scale synthetic aperture radar SAR image in all scales.
And 9, segmenting the SAR image.
And performing parameter estimation by using an Iterative Condition Estimation (ICE) method, and segmenting the Synthetic Aperture Radar (SAR) image.
The specific steps of the iterative condition estimation ICE method are as follows:
step 1, forming a model parameter of a synthetic aperture radar SAR image segmentation method based on high-order multi-scale conditional random field CRF semi-supervision by using horizontal related parameters, vertical related parameters, inter-scale related parameters of a multi-scale synthetic aperture radar SAR image, scale parameters of generalized Gamma distribution, shape parameters and index shape parameters.
And 2, according to model parameters of the current synthetic aperture radar SAR image segmentation method based on the high-order multi-scale conditional random field CRF semi-supervision, taking the category corresponding to the maximum value of the posterior marginal probability of each pixel point of the multi-scale synthetic aperture radar SAR image in each scale as the category of the pixel point, and obtaining the current segmentation result of the synthetic aperture radar SAR image.
And 3, calculating the frequency of each second-order neighborhood system in the multi-scale synthetic aperture radar SAR image in the whole multi-scale synthetic aperture radar SAR image by using the currently segmented image according to the following formula:
Figure BDA0001769551240000141
wherein the content of the first and second substances,
Figure BDA0001769551240000142
representing the second-order neighborhood system s' centered on the s-th pixel in the multi-scale synthetic aperture radar SAR image, the frequency appearing in the whole synthetic aperture radar SAR image,
Figure BDA0001769551240000143
and H represents the total number of all second-order neighborhood systems in the whole multi-scale synthetic aperture radar SAR image.
And 4, calculating the association vector of each pixel point in the multi-scale synthetic aperture radar SAR image according to the following formula:
Figure BDA0001769551240000144
wherein R issRepresenting the relevance vector of the s-th pixel point in the multi-scale synthetic aperture radar SAR image, wherein I (·,) represents a constant, when two elements in the parenthesis take equal values, I (·) 1, otherwise, I (·) 1, t1Representing the left pixel point, t, in the neighborhood system with the s-th pixel as the center in the multi-scale synthetic aperture radar SAR image2Representing the upper pixel point, t, in the neighborhood system with the s-th pixel as the center in the multi-scale synthetic aperture radar SAR image3Representing the right pixel point, t, in the neighborhood system with the s-th pixel as the center in the multi-scale synthetic aperture radar SAR image4And representing a lower pixel point in the neighborhood system with the s-th pixel as the center in the multi-scale synthetic aperture radar SAR image.
And 5, calculating horizontal related parameters, vertical related parameters and inter-scale related parameters in the multi-scale synthetic aperture radar SAR image:
Figure BDA0001769551240000145
wherein R isbExpressing the relevance vector of the b-th pixel point in the multi-scale synthetic aperture radar SAR image, T expressing a transposed symbol, theta1Representing a vector consisting of horizontal related parameters, vertical related parameters, and inter-scale related parameters in a multi-scale synthetic aperture radar SAR image,
Figure BDA0001769551240000151
and the frequency of the second-order neighborhood system b 'taking the b-th pixel in the image as the center in the x-th' type of the multi-scale synthetic aperture radar SAR image in the whole synthetic aperture radar SAR image is represented.
And 6, performing Mellin transform on the probability density function of the generalized Gamma distribution to obtain a first second characteristic function of the generalized Gamma distribution.
And 7, calculating a second class characteristic function of the generalized Gamma distribution according to the following formula:
ξ(ω)＝lnγ(ω)
where ξ (ω) is the second class of characteristic function of the generalized Gamma distribution, ω is the argument of the probability density function of the generalized Gamma distribution, and γ (ω) is the first second class of characteristic function of the generalized Gamma distribution.
And 8, calculating first-order logarithmic cumulant, second-order logarithmic cumulant and third-order logarithmic cumulant of a second class characteristic function of generalized Gamma distribution according to the following formula:
Figure BDA0001769551240000152
step 9, establishing a first moment model of parameter estimation according to the following formula:
Figure BDA0001769551240000153
wherein ψ (0, λ)l) The parameter k 'representing the multi-gamma poly gamma function is 0, and x' is lambdalCase of (a), nsRepresenting the number of pixel points of the segmented multi-scale synthetic aperture radar SAR,
Figure BDA0001769551240000154
and representing the strength of the segmented multi-scale synthetic aperture radar SAR at the s-th pixel point.
Step 10, establishing a second moment model of parameter estimation according to the following formula:
Figure BDA0001769551240000155
wherein ψ (1, λ)l) The parameter k 'representing the multi-gamma poly gamma function is 1, and x' is lambdalThe situation of (c)1First-order logarithm of second class characteristic function representing generalized Gamma Gamma distributionThe cumulative amount.
And 11, establishing a third moment model for parameter estimation according to the following formula:
Figure BDA0001769551240000161
wherein ψ (2, λ)l) The parameter k 'representing the multi-gamma poly gamma function is 2, and the parameter x' is lambdalThe situation of time;
and 12, solving scale parameters, shape parameters and index shape parameters of generalized Gamma distribution according to a first moment, a second moment and a third moment model formula of parameter estimation.
Step 13, judging whether the maximum iteration number is 6, if so, executing step 14; otherwise, the step 2 is executed after adding 1 to the current iteration number.
And step 14, respectively averaging the parameters obtained by 6 iterations to obtain model parameters of the synthetic aperture radar SAR image segmentation method based on the semi-supervised high-order multi-scale conditional random field CRF.
Step 10, judging whether the current iteration number is 10, if so, executing step 11; otherwise, step 6 is executed after adding 1 to the current iteration number.
And step 11, ending the division.
And finishing the segmentation of the SAR image to obtain a segmentation result of the SAR image semi-supervised by the high-order multi-scale conditional random field CRF.
The effect of the present invention will be further described with reference to simulation experiments.
1. Simulation experiment conditions are as follows:
the hardware test platform of the simulation experiment of the invention is as follows: the processor is a CPU intel Core i7-7700, the main frequency is 4.2GHz, and the internal memory is 16 GB; the software platform is as follows: windows 10 family version, 64-bit operating system, MATLAB R2016 a.
2. Simulation content and simulation result analysis:
the simulation experiment 1 is to simulate a synthetic aperture radar SAR image combined by three homogeneous areas by adopting the method disclosed by the invention, as shown in figure 2, the simulation experiment 2 is to respectively simulate a synthetic aperture radar SAR image of a river channel by adopting the method disclosed by the invention and a generalized conditional random field CRF method, as shown in figure 3, and the simulation experiment 3 is to simulate an ESAR image of a Parvolun area by adopting the method disclosed by the invention and the generalized conditional random field CRF method, as shown in figure 4.
Simulation experiment 1: fig. 2(a) is an original drawing of a simulation experiment 1 of the present invention, which is a simulated image generated from a pavloney area near munich, germany using ESAR L band HV data of the aerospace center, germany. And manually selecting three types of uniform areas from the simulated image, and then putting the areas into a preset ground truth to obtain the synthetic aperture radar SAR image combined by the three types of uniform areas. Fig. 2(b) is a reference diagram of the original image after being divided, and fig. 2(c) is a simulation result diagram of the division of fig. 2(a) by using the method of the present invention.
Simulation experiment 2: fig. 3(a) is an original drawing of a simulation experiment 2 according to the present invention, in which a synthetic aperture radar SAR image of a river channel is generated from L-band data of a river channel acquired by a jet propulsion laboratory in the united states, fig. 3(b) is a simulation result drawing obtained by dividing fig. 3(a) by using a generalized conditional random field CRF method according to the related art, and fig. 3(c) is a simulation result drawing obtained by dividing fig. 3(a) by using a method according to the present invention.
Simulation experiment 3: fig. 4(a) is an original drawing of a simulation experiment 3 of the present invention, in which a synthetic aperture radar SAR image is generated from L-band data of the pavlonen region acquired by an ESAR radar system of the german space center, fig. 4(b) is a simulation result drawing obtained by dividing fig. 4(a) by using a conventional generalized conditional random field CRF method, and fig. 4(c) is a simulation result drawing obtained by dividing fig. 4(a) by using the method of the present invention.
As can be seen from fig. 2(a), the original image has strong noise, and comparing the simulation result of the method of the present invention, fig. 2(c), with the reference image of the original image after segmentation, it can be seen that the noise interference has less influence on the method of the present invention, and comparing the three homogeneous regions segmented in the simulation result of the method of the present invention, fig. 2(c), with the reference image 2(b), it can be seen that the simulation result of the method of the present invention has strong region consistency and less erroneous segmentation. As can be seen from the original image 3(a), the image mainly comprises three areas, namely farmland, land and river channel, and the simulation result of the method of the invention, namely a graph 3(c), is compared with the simulation result of the generalized conditional random field CRF method in the prior art, namely a graph 3(b), so that the simulation result of the method of the invention is obviously superior to the boundary positioning of the river channel and the farmland to the conventional generalized conditional random field CRF method. Comparing the simulation result of the method of the present invention, namely the graph 4(c), with the simulation result of the generalized conditional random field CRF method of the prior art, namely the graph 4(b), the simulation result of the method of the present invention, namely the graph 4(a), which is a complex regional graph containing urban areas and mountain areas, can be seen to be smoother, the structural information of the image is maintained more accurately, and more detailed information is contained.
In conclusion, the simulation of the invention verifies the correctness, validity and reliability of the invention.

Claims (8)

1. A synthetic aperture radar SAR image segmentation method based on high-order multi-scale conditional random field CRF semi-supervision is characterized in that high-order potential energy is introduced to calculate edge probability; calculating the joint posterior marginal probability by using a bottom-up recursion method; calculating posterior marginal probability by using a top-down recursion method; the method comprises the following steps:
(1) inputting a Synthetic Aperture Radar (SAR) image;
(2) performing wavelet transformation on an input Synthetic Aperture Radar (SAR) image;
(3) obtaining a feature vector of a high-order multi-scale conditional random field CRF:
(3a) calculating the edge strength between each pixel point in each scale and each pixel point in the adjacent domain system of the image after wavelet transformation by using an exponential weighted average ratio operator;
(3b) sliding a window with the radius of 5 pixel points at intervals in the wavelet-transformed image, performing sliding window operation on the wavelet-transformed image, and calculating histogram features of all pixel values in each sliding window;
(3c) sliding a window with the radius of 7 pixel points at intervals in the wavelet-transformed image, performing sliding window operation on the wavelet-transformed image, and calculating all gray level co-occurrence matrix texture features in each sliding window;
(3d) forming local feature vectors by using all histogram features and gray level co-occurrence matrix texture features of the wavelet-transformed image;
(3e) forming a feature vector of a high-order multi-scale conditional random field CRF by using all local feature vectors and edge intensities of the wavelet-transformed image;
(4) obtaining the local class conditional probability of the high-order multi-scale conditional random field CRF:
inputting the feature vector of the high-order multi-scale conditional random field CRF into a multi-class Support Vector Machine (SVM) classifier, and outputting the classifier as the local class conditional probability of the high-order multi-scale conditional random field CRF;
(5) initially segmenting a Synthetic Aperture Radar (SAR) image:
initially segmenting the SAR image by utilizing a method of maximizing the local class conditional probability;
(6) calculating the edge probability of each pixel:
calculating the edge probability of each pixel point high-order multi-scale conditional random field CRF in the current segmented synthetic aperture radar SAR image by using a mean field estimation method;
(7) calculating the combined posterior marginal probability of each pixel point:
calculating the joint posterior marginal probability of each pixel point high-order multi-scale conditional random field CRF in the current segmented synthetic aperture radar SAR image by using the marginal probability of each pixel point by adopting a bottom-up recursion method;
(8) calculating the posterior marginal probability of each pixel point:
calculating the posterior marginal probability of each pixel point of the currently segmented synthetic aperture radar SAR image in a high-order multi-scale conditional random field CRF by using a top-down recursion method and the combined posterior marginal probability of each pixel point;
(9) segmenting the SAR image:
performing parameter estimation by using an Iterative Condition Estimation (ICE) method, and segmenting a Synthetic Aperture Radar (SAR) image;
(10) judging whether the current iteration number is 10, if so, executing the step (11); otherwise, adding 1 to the current iteration number and then executing the step (6);
(11) and (5) finishing the segmentation:
and finishing the segmentation of the SAR image to obtain a segmentation result of the SAR image semi-supervised by the high-order multi-scale conditional random field CRF.
2. The SAR image segmentation method based on the CRF semi-supervised of the higher-order multi-scale conditional random field (3a) as claimed in claim 1, wherein the formula of the exponentially weighted average ratio operator in the step (3a) is as follows:
Figure FDA0002646660880000021
wherein the content of the first and second substances,
Figure FDA0002646660880000022
the method is characterized in that the method represents the edge strength between the s-th pixel point of the multi-scale synthetic aperture radar SAR image in the nth scale and the t-th pixel point of the multi-scale synthetic aperture radar SAR image in a neighborhood system, the value range of n is {0,1,2}, exp (·) represents exponential operation with a natural constant as a base,
Figure FDA0002646660880000023
representing the pixel value of the s-th pixel point of the multi-scale synthetic aperture radar SAR image in the nth scale,
Figure FDA0002646660880000024
and expressing the pixel value of the t-th pixel point of the multi-scale synthetic aperture radar SAR image in the nth scale.
3. The SAR image segmentation method based on the CRF semi-supervised of the higher-order multi-scale Conditional Random Field (CRF) as claimed in claim 1, wherein the specific steps of calculating the histogram features of all the pixel values in each sliding window in the step (3b) are as follows:
first, the mean value of all pixel values in each sliding window is calculated according to the following formula:
Figure FDA0002646660880000031
wherein, mumRepresents the mean value of all pixel values in the mth sliding window, sigma represents the summation operation, i represents the serial number of the pixel point row in the mth sliding window, j represents the serial number of the pixel point column in the mth sliding window, yijThe pixel value, n, of the ith row and the jth column of pixel points in the mth sliding windowmRepresenting the total number of all pixel points in the mth sliding window;
secondly, calculating the variance of all pixel values in each sliding window according to the following formula:
Figure FDA0002646660880000032
wherein the content of the first and second substances,
Figure FDA0002646660880000033
representing the variance of all pixel values within the mth sliding window;
thirdly, calculating the skewness of all pixel values in each sliding window according to the following formula:
Figure FDA0002646660880000034
wherein phimRepresenting the skewness of all pixel values in the mth sliding window;
fourthly, calculating the peak state of all pixel values in each sliding window according to the following formula:
Figure FDA0002646660880000035
wherein, KmRepresenting the peak state of all pixel values in the mth sliding window;
fifthly, calculating the entropy of all pixel values in each sliding window according to the following formula:
Figure FDA0002646660880000036
wherein, TmRepresenting the entropy of all pixel values within the mth sliding window;
and sixthly, forming the histogram characteristics of the sliding window by the mean value, the variance, the skewness, the kurtosis and the entropy of all the pixel values in each sliding window.
4. The SAR image segmentation method based on the CRF semi-supervised of the higher-order multi-scale conditional random field is characterized in that the specific steps of calculating all the texture features of the gray level co-occurrence matrix in each sliding window in the step (3c) are as follows:
step one, forming a pixel pair by one pixel and adjacent pixels in each sliding window, dividing the occurrence frequency of the same pixel pair in each sliding window by the occurrence frequency of other pixel pairs in the sliding window, and forming a gray level co-occurrence matrix of the sliding window by all quotient values;
secondly, calculating the contrast of all elements in each gray level co-occurrence matrix according to the following formula:
Figure FDA0002646660880000041
wherein the content of the first and second substances,
Figure FDA0002646660880000046
is shown as
Figure FDA0002646660880000047
The contrast of all elements in the gray level co-occurrence matrix, u represents the serial number of the row in the gray level co-occurrence matrix, v represents the serial number of the column in the gray level co-occurrence matrix,
Figure FDA0002646660880000042
is shown as
Figure FDA0002646660880000048
Element values of the ith row and the vth column in the gray level co-occurrence matrix;
thirdly, calculating the correlation degree of all elements in each gray level co-occurrence matrix according to the following formula:
Figure FDA0002646660880000043
wherein the content of the first and second substances,
Figure FDA0002646660880000049
is shown as
Figure FDA00026466608800000410
The correlation degree of all elements in the gray level co-occurrence matrix;
fourthly, calculating the homogeneity of all elements in each gray level co-occurrence matrix according to the following formula:
Figure FDA0002646660880000044
wherein the content of the first and second substances,
Figure FDA00026466608800000411
is shown as
Figure FDA00026466608800000412
Homogeneity of all elements in the gray level co-occurrence matrix;
and fifthly, forming the characteristic texture characteristics of the gray level co-occurrence matrix by the mean value, the contrast, the correlation and the homogeneity of elements in each gray level co-occurrence matrix.
5. The synthetic aperture radar SAR image segmentation method based on the high-order multi-scale conditional random field CRF semi-supervised according to the claim 2, wherein the mean field estimation method in the step (6) comprises the following specific steps:
step one, taking the local class conditional probability as an initial value of local mean field energy of a segmented synthetic aperture radar SAR image;
secondly, calculating local radical energy between each pixel point in the segmented multi-scale synthetic aperture radar SAR image and each pixel point in the neighborhood system according to the following formula:
Figure FDA0002646660880000045
wherein Q is{s,t}Expressing the local radical energy alpha between the s-th pixel point in the segmented multi-scale synthetic aperture radar SAR image and the t-th pixel point in the neighborhood systemHExpressing the correlation parameter between the s-th pixel point of the segmented multi-scale synthetic aperture radar SAR image and the t-th pixel point in the horizontal neighborhood system, the initial value of the parameter is 1, updating the parameter value by using an iterative condition estimation ICE method in step (9),
Figure FDA0002646660880000051
expressing a constant, and when the segmented multi-scale synthetic aperture radar SAR image is in the category of the s pixel point in the n scale
Figure FDA0002646660880000052
And the category of the t-th pixel point in the neighborhood system
Figure FDA0002646660880000053
When the phase of the mixture is the same as the phase of the mixture,
Figure FDA0002646660880000054
when the label of the s pixel point of the segmented multi-scale synthetic aperture radar SAR image in the n scale
Figure FDA0002646660880000055
Label of t-th pixel point in its neighborhood system
Figure FDA0002646660880000056
At a different time,
Figure FDA0002646660880000057
Figure FDA0002646660880000058
representing an arbitrary symbol, e representing belonging to a symbol, NHA horizontal neighborhood system, alpha, representing the s-th pixel in the segmented SAR imageVExpressing the correlation parameter between the s-th pixel point in the segmented multi-scale synthetic aperture radar SAR image and the t-th pixel point in the vertical neighborhood system, wherein the initial value of the parameter is 1, updating the parameter value by using an iterative condition estimation ICE method in step (9), and NVA vertical neighborhood system for expressing the s-th pixel point in the segmented multi-scale synthetic aperture radar SAR image;
thirdly, calculating the local mean field energy of each pixel point of the segmented multi-scale synthetic aperture radar SAR image in each scale according to the following formula:
Figure FDA0002646660880000059
wherein the content of the first and second substances,
Figure FDA00026466608800000510
representing the local mean value field energy, N, of the s-th pixel point of the segmented multi-scale synthetic aperture radar SAR image in the N-th scalesRepresenting segmented multiscale composite poresK represents the number of the categories in the neighborhood system, L represents the total number of the categories of the segmented multi-scale synthetic aperture radar SAR image, L represents the number of the categories of the segmented multi-scale synthetic aperture radar SAR image, and elRepresenting the ith category of the s pixel point in the segmented multi-scale synthetic aperture radar SAR image, ekRepresents the category of the t-th pixel point in the neighborhood system,
Figure FDA00026466608800000511
the local mean field energy of the t-th pixel point in the nth scale of the neighborhood system is represented,
Figure FDA0002646660880000061
values are taken in a synthetic aperture radar SAR image category set according to category numbers;
fourthly, executing the third step twice to obtain local mean value field energy of each pixel point of the segmented multi-scale synthetic aperture radar SAR image in each scale, and taking the local mean value field energy as an estimated value of prior marginal probability of each pixel point of the segmented multi-scale synthetic aperture radar SAR image in each scale;
fifthly, calculating the inter-scale potential energy of each pixel point of the segmented multi-scale synthetic aperture radar SAR image in each scale according to the following formula:
Figure FDA0002646660880000062
wherein the content of the first and second substances,
Figure FDA0002646660880000063
representing the s pixel point of the segmented multi-scale synthetic aperture radar SAR image in the n scale and the n +1 scale
Figure FDA0002646660880000064
The inter-scale potential energy of each pixel point, eta represents the multi-scale synthetic hole after segmentationThe inter-scale related parameters of the SAR image of the radar have the initial value of 1, the parameter values are updated by using an iterative condition estimation ICE method in step (9),
Figure FDA0002646660880000065
representing the segmented multi-scale synthetic aperture radar SAR image in the (n + 1) th scale
Figure FDA0002646660880000066
The category of each pixel point;
sixthly, calculating the conditional prior marginal probability of each pixel point of the segmented multi-scale synthetic aperture radar SAR image in each scale according to the following formula:
Figure FDA0002646660880000067
wherein the content of the first and second substances,
Figure FDA0002646660880000068
expressing the conditional prior marginal probability of an s-th pixel point of the segmented multi-scale synthetic aperture radar SAR image in an n-th scale;
and seventhly, calculating unitary potential energy of the conditional random field CRF model of each pixel point of the segmented multi-scale synthetic aperture radar SAR image in each scale according to the following formula:
Figure FDA0002646660880000069
wherein f iss nExpressing the unary potential energy of the conditional random field CRF model of the s pixel point of the segmented multi-scale synthetic aperture radar SAR image in the n scale, wherein log expresses the logarithm operation with 10 as the base, Ps nExpressing the local class conditional probability of an s-th pixel point of the segmented multi-scale synthetic aperture radar SAR image in an n-th scale;
and eighthly, calculating the scattering statistics of each pixel point of the segmented multi-scale synthetic aperture radar SAR image in each scale according to the following formula:
Figure FDA0002646660880000071
wherein the content of the first and second substances,
Figure FDA0002646660880000072
representing scattering statistics of s pixel point of segmented multi-scale synthetic aperture radar SAR image in n scale, betalThe shape parameter of the generalized Gamma Gamma distribution model is represented, and the parameter value is calculated by the logarithm cumulant MoLC method in the step (9), lambdalIndex shape parameter representing Gamma Gamma distribution model, the parameter value is calculated by the logarithm cumulant MoLC method in step (9),
Figure FDA0002646660880000073
representing the strength alpha of the s-th pixel point of the segmented multi-scale synthetic aperture radar SAR image in the n-th scalelA scale parameter representing a generalized Gamma distribution model, wherein the parameter value is calculated by a logarithm cumulant MoLC method in the step (9), and Gamma (·) represents a Gamma function;
and ninthly, calculating unitary potential energy of the generalized conditional random field GCRF model of each pixel point of the segmented multi-scale synthetic aperture radar SAR image in each scale according to the following formula:
Figure FDA0002646660880000074
wherein the content of the first and second substances,
Figure FDA0002646660880000075
representing unary potential energy of a generalized conditional random field GCRF model of an s-th pixel point of the segmented multi-scale synthetic aperture radar SAR image in the nth scale;
tenthly, sliding a window with the radius of 7 pixel points at intervals in the segmented synthetic aperture radar SAR image, and performing one-time window sliding operation on the segmented multi-scale synthetic aperture radar SAR image;
step ten, calculating the probability of each category in each sliding window according to the following formula;
Figure FDA0002646660880000076
wherein the content of the first and second substances,
Figure FDA0002646660880000077
is shown as
Figure FDA0002646660880000078
The probability of the ith class within the sliding window,
Figure FDA0002646660880000079
is shown as
Figure FDA00026466608800000710
The number of the ith class in a sliding window,
Figure FDA00026466608800000711
is shown as
Figure FDA00026466608800000712
Single radical C in a sliding window1The total number of (c);
the twelfth step, according to the following formula, calculates the probability of each category adjacent to other categories in each sliding window:
Figure FDA0002646660880000081
wherein the content of the first and second substances,
Figure FDA0002646660880000082
is shown as
Figure FDA0002646660880000083
Probability that the ith class is adjacent to the epsilon-th class in the sliding window, pi represents the product operation, tau represents the type of the binary group,
Figure FDA0002646660880000084
is shown as
Figure FDA0002646660880000085
A left binary group C consisting of the first class and the epsilon class in the sliding window2(1) The total number of (a) and (b),
Figure FDA0002646660880000086
is shown as
Figure FDA0002646660880000087
Right binary group C formed by the first category and the epsilon category in the sliding window2(2) The total number of (a) and (b),
Figure FDA0002646660880000088
is shown as
Figure FDA0002646660880000089
A lower binary group C consisting of the first category and the epsilon second category in the sliding window2(3) The total number of (a) and (b),
Figure FDA00026466608800000810
is shown as
Figure FDA00026466608800000811
The upper binary group C formed by the first category and the epsilon category in the sliding window2(4) The total number of (a) and (b),
Figure FDA00026466608800000812
is shown as
Figure FDA00026466608800000813
The total number of left binary clusters of the ith class within the individual sliding window,
Figure FDA00026466608800000814
is shown as
Figure FDA00026466608800000815
The total number of the first category of right binary radicals in each sliding window,
Figure FDA00026466608800000816
is shown as
Figure FDA00026466608800000817
The total number of binary groups under the ith class in each sliding window,
Figure FDA00026466608800000818
is shown as
Figure FDA00026466608800000819
The total number of binary groups in the first category within each sliding window;
step thirteen, calculating the high-order potential energy of each pixel point in each sliding window according to the following formula:
Figure FDA00026466608800000820
wherein the content of the first and second substances,
Figure FDA00026466608800000821
is shown as
Figure FDA00026466608800000822
The high-order potential energy of the s-th pixel point in each sliding window;
fourthly, calculating the mean field energy of each pixel point of the segmented multi-scale synthetic aperture radar SAR image in each scale according to the following formula:
Figure FDA00026466608800000823
wherein the content of the first and second substances,
Figure FDA00026466608800000824
representing the mean field energy of the s-th pixel point of the segmented multi-scale synthetic aperture radar SAR image in the n-th scale;
fifthly, calculating the edge probability of each pixel point of the segmented multi-scale synthetic aperture radar SAR image in each scale according to the following formula:
Figure FDA00026466608800000825
wherein the content of the first and second substances,
Figure FDA00026466608800000826
and representing the marginal probability of the s-th pixel point of the segmented multi-scale synthetic aperture radar SAR image in the n-th scale.
6. The synthetic aperture radar SAR image segmentation method based on the high-order multi-scale conditional random field CRF semi-supervised according to the claim 5, wherein the bottom-up recursion method in the step (7) comprises the following steps:
step one, taking the edge probability of each pixel point of the segmented multi-scale synthetic aperture radar SAR image in the 0 th scale as the local posterior edge probability of each pixel point of the segmented multi-scale synthetic aperture radar SAR image in the 0 th scale;
secondly, calculating the local posterior marginal probability of each pixel point of the segmented multi-scale synthetic aperture radar SAR image in each scale according to the following formula:
Figure FDA0002646660880000091
wherein the content of the first and second substances,
Figure FDA0002646660880000092
the local posterior marginal probability of the s-th pixel point of the segmented multi-scale synthetic aperture radar SAR image in the nth scale is represented,srepresenting the descendant of the s-th pixel in the multi-scale synthetic aperture radar SAR image, representing the s-th pixel point of the multi-scale synthetic aperture radar SAR image in the scale lower than the n-th scale,
Figure FDA0002646660880000093
representing the class of the q generation in the filial generation of the s pixel point of the segmented multi-scale synthetic aperture radar SAR image,
Figure FDA0002646660880000094
representing the posterior marginal probability of the q generation in the s pixel point descendant of the segmented multi-scale synthetic aperture radar SAR image,
Figure FDA0002646660880000095
representing the q-th generation conditional prior marginal probability in the s-th pixel point descendant of the segmented multi-scale synthetic aperture radar SAR image,
Figure FDA0002646660880000096
expressing the prior marginal probability of the q generation in the s pixel point descendant of the segmented multi-scale synthetic aperture radar SAR image;
thirdly, the second step is executed twice to obtain the local posterior marginal probability of each pixel point of the segmented multi-scale synthetic aperture radar SAR image in each scale;
fourthly, calculating the combined posterior marginal probability of each pixel point and the parent pixel point of the segmented multi-scale synthetic aperture radar SAR image in each scale according to the following formula:
Figure FDA0002646660880000101
wherein the content of the first and second substances,
Figure FDA0002646660880000102
expressing the combined posterior marginal probability of the s-th pixel point of the segmented multi-scale synthetic aperture radar SAR image in the nth scale and the parent pixel point thereof, wherein the parent represents the s-th pixel point of the multi-scale synthetic aperture radar image SAR in the nth scale,
Figure FDA0002646660880000109
represents a parent of the s-th pixel in the multi-scale synthetic aperture radar SAR image,
Figure FDA0002646660880000103
representing the nth of the segmented multi-scale synthetic aperture radar SAR image in the (n + 1) th scale
Figure FDA0002646660880000104
The prior edge probability of an individual pixel point,
Figure FDA0002646660880000105
representing the nth of the segmented multi-scale synthetic aperture radar SAR image in the (n + 1) th scale
Figure FDA00026466608800001010
A category of individual pixels.
7. The SAR image segmentation method based on the CRF semi-supervised of the higher-order multi-scale Conditional Random Field (CRF) according to claim 6, wherein the top-down recursive method in the step (8) comprises the following specific steps:
firstly, setting an initial value of posterior marginal probability of each pixel point of a segmented multi-scale synthetic aperture radar SAR image in a 2 nd scale as a local posterior marginal probability of each pixel point of the segmented multi-scale synthetic aperture radar SAR image in the 2 nd scale;
secondly, calculating the posterior marginal probability of each pixel point of the segmented multi-scale synthetic aperture radar SAR image in each scale according to the following formula:
Figure FDA0002646660880000106
wherein the content of the first and second substances,
Figure FDA0002646660880000107
representing the posterior marginal probability of the s-th pixel point of the segmented multi-scale synthetic aperture radar SAR image in the n-th scale,
Figure FDA0002646660880000108
representing the segmented multi-scale synthetic aperture radar SAR image in the (n + 1) th scale
Figure FDA00026466608800001011
Posterior marginal probability of each pixel point;
and thirdly, executing the second step twice to obtain the posterior marginal probability of each pixel point of the segmented multi-scale synthetic aperture radar SAR image in all scales.
8. The synthetic aperture radar SAR image segmentation method based on the high-order multi-scale conditional random field CRF semi-supervised according to the claim 1, wherein the iterative condition estimation ICE method in the step (9) comprises the following specific steps:
firstly, horizontal related parameters, vertical related parameters, inter-scale related parameters and scale parameters alpha of generalized Gamma Gamma distribution of the multi-scale synthetic aperture radar SAR image are obtainedlShape parameter λlIndexing the shape parameter betalComposition based on higher order polyModel parameters of a synthetic aperture radar SAR image segmentation method of scale conditional random field CRF semi-supervision;
secondly, according to model parameters of a current synthetic aperture radar SAR image segmentation method based on high-order multi-scale conditional random field CRF semi-supervision, taking the category corresponding to the maximum value of the posterior marginal probability of each pixel point of the multi-scale synthetic aperture radar SAR image in each scale as the category of the pixel point, and obtaining the current segmentation result of the synthetic aperture radar SAR image;
thirdly, according to the following formula, calculating the frequency of each second-order neighborhood system in the multi-scale synthetic aperture radar SAR image in the whole multi-scale synthetic aperture radar SAR image by using the current segmented image:
Figure FDA0002646660880000111
wherein the content of the first and second substances,
Figure FDA0002646660880000112
representing the second-order neighborhood system s' centered on the s-th pixel in the multi-scale synthetic aperture radar SAR image, the frequency appearing in the whole synthetic aperture radar SAR image,
Figure FDA0002646660880000113
expressing the times of occurrence of the chi-type second-order neighborhood system s' taking the s-th pixel in the multi-scale synthetic aperture radar SAR image as the center in the whole synthetic aperture radar SAR image, and expressing the total number of all the second-order neighborhood systems in the whole multi-scale synthetic aperture radar SAR image by H;
fourthly, calculating the association vector of each pixel point in the multi-scale synthetic aperture radar SAR image according to the following formula:
Figure FDA0002646660880000114
wherein R issRepresenting the relevance vector of the s-th pixel point in the multi-scale synthetic aperture radar SAR image, wherein I (·,) represents a constant, when two elements in the parenthesis take equal values, I (·) 1, otherwise, I (·) 1, t1Representing the left pixel point, t, in the neighborhood system with the s-th pixel as the center in the multi-scale synthetic aperture radar SAR image2Representing the upper pixel point, t, in the neighborhood system with the s-th pixel as the center in the multi-scale synthetic aperture radar SAR image3Representing the right pixel point, t, in the neighborhood system with the s-th pixel as the center in the multi-scale synthetic aperture radar SAR image4Representing a lower pixel point in a neighborhood system with an s-th pixel as a center in a multi-scale synthetic aperture radar SAR image;
fifthly, calculating horizontal related parameters, vertical related parameters and inter-scale related parameters in the multi-scale synthetic aperture radar SAR image according to the following formula:
Figure FDA0002646660880000121
wherein R isbExpressing the relevance vector of the b-th pixel point in the multi-scale synthetic aperture radar SAR image, T expressing a transposed symbol, theta1Representing a vector consisting of horizontal correlation parameters, vertical correlation parameters, inter-scale correlation parameters in a multi-scale synthetic aperture radar SAR image,
Figure FDA0002646660880000122
representing the frequency of the second-order neighborhood system b ' in the x ' th synthetic aperture radar SAR image by taking the b ' th pixel in the image as the center in the whole synthetic aperture radar SAR image;
sixthly, performing Mellin transform on the probability density function of the generalized Gamma distribution to obtain a first second characteristic function of the generalized Gamma distribution;
and seventhly, calculating a second class characteristic function of the generalized Gamma distribution according to the following formula:
ξ(ω)＝lnγ(ω)
where ξ (ω) is the second class of characteristic function of the generalized Gamma distribution, ω is the argument of the probability density function of the generalized Gamma distribution, and γ (ω) is the first second class of characteristic function of the generalized Gamma distribution;
and eighthly, calculating first-order logarithmic cumulant, second-order logarithmic cumulant and third-order logarithmic cumulant of a second class characteristic function of generalized Gamma distribution according to the following formula:
Figure FDA0002646660880000123
and step nine, establishing a first moment model of parameter estimation according to the following formula:
Figure FDA0002646660880000124
wherein ψ (0, λ)l) The parameter k 'representing the multi-gamma poly gamma function is 0, and x' is lambdalCase of (a), nsRepresenting the number of pixel points of the segmented multi-scale synthetic aperture radar SAR,
Figure FDA0002646660880000125
representing the strength of the segmented multi-scale synthetic aperture radar SAR at the s-th pixel point;
step ten, establishing a second moment model of parameter estimation according to the following formula:
Figure FDA0002646660880000131
wherein ψ (1, λ)l) The parameter k 'representing the multi-gamma poly gamma function is 1, and x' is lambdalThe situation of (c)1A first order logarithmic cumulant of a second class of characteristic functions representing a generalized Gamma distribution;
step ten, establishing a third moment model of parameter estimation according to the following formula:
Figure FDA0002646660880000132
wherein ψ (2, λ)l) The parameter k 'representing the multi-gamma poly gamma function is 2, and the parameter x' is lambdalThe situation of time;
the twelfth step, solving scale parameters, shape parameters and index shape parameters of generalized Gamma distribution according to a first moment, a second moment and a third moment model formula of parameter estimation;
step thirteen, judging whether the maximum iteration times is 6, if so, executing the step fourteen; otherwise, the second step is executed after the current iteration times are added by 1;
and fourteenth, respectively averaging the parameters obtained by 6 iterations to obtain model parameters of the synthetic aperture radar SAR image segmentation method based on the semi-supervised high-order multi-scale conditional random field CRF.
CN201810943115.2A 2018-08-17 2018-08-17 SAR image segmentation method based on high-order multi-scale CRF semi-supervision Active CN109345537B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201810943115.2A CN109345537B (en) 2018-08-17 2018-08-17 SAR image segmentation method based on high-order multi-scale CRF semi-supervision

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201810943115.2A CN109345537B (en) 2018-08-17 2018-08-17 SAR image segmentation method based on high-order multi-scale CRF semi-supervision

Publications (2)

Publication Number Publication Date
CN109345537A CN109345537A (en) 2019-02-15
CN109345537B true CN109345537B (en) 2021-02-12

Family

ID=65291709

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201810943115.2A Active CN109345537B (en) 2018-08-17 2018-08-17 SAR image segmentation method based on high-order multi-scale CRF semi-supervision

Country Status (1)

Country Link
CN (1) CN109345537B (en)

Families Citing this family (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110363201B (en) * 2019-07-10 2021-06-01 上海交通大学 Weak supervision semantic segmentation method and system based on collaborative learning
CN110517277B (en) * 2019-08-05 2022-12-06 西安电子科技大学 SAR image segmentation method based on PCANet and high-order CRF

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101950363A (en) * 2010-08-19 2011-01-19 武汉大学 SAR image monitoring and classifying method based on conditional random field model
CN103606164A (en) * 2013-12-03 2014-02-26 西安电子科技大学 SAR image segmentation method based on high-dimensional triple Markov field
CN104732552A (en) * 2015-04-09 2015-06-24 西安电子科技大学 SAR image segmentation method based on nonstationary condition field
CN106683109A (en) * 2016-12-28 2017-05-17 西安电子科技大学 SAR image segmentation method based on semantic condition random field model
US9679226B1 (en) * 2012-03-29 2017-06-13 Google Inc. Hierarchical conditional random field model for labeling and segmenting images
CN107680113A (en) * 2017-10-27 2018-02-09 武汉大学 The image partition method of multi-layer segmentation network based on Bayesian frame edge prior

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101950363A (en) * 2010-08-19 2011-01-19 武汉大学 SAR image monitoring and classifying method based on conditional random field model
US9679226B1 (en) * 2012-03-29 2017-06-13 Google Inc. Hierarchical conditional random field model for labeling and segmenting images
CN103606164A (en) * 2013-12-03 2014-02-26 西安电子科技大学 SAR image segmentation method based on high-dimensional triple Markov field
CN104732552A (en) * 2015-04-09 2015-06-24 西安电子科技大学 SAR image segmentation method based on nonstationary condition field
CN106683109A (en) * 2016-12-28 2017-05-17 西安电子科技大学 SAR image segmentation method based on semantic condition random field model
CN107680113A (en) * 2017-10-27 2018-02-09 武汉大学 The image partition method of multi-layer segmentation network based on Bayesian frame edge prior

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
Adaptive Hybrid Conditional Random Field Model for SAR Image Segmentation;Fan Wang等;《IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING》;20170131;第55卷(第1期);第537-550页 *

Also Published As

Publication number Publication date
CN109345537A (en) 2019-02-15

Similar Documents

Publication Publication Date Title
CN108388927B (en) Small sample polarization SAR terrain classification method based on deep convolution twin network
CN109086700B (en) Radar one-dimensional range profile target identification method based on deep convolutional neural network
CN110210486B (en) Sketch annotation information-based generation countermeasure transfer learning method
CN104915676B (en) SAR image sorting technique based on further feature study and watershed
CN103413146B (en) Polarimetric SAR Image sophisticated category method based on Freeman entropys and self study
CN108447057B (en) SAR image change detection method based on significance and depth convolution network
CN107491734B (en) Semi-supervised polarimetric SAR image classification method based on multi-core fusion and space Wishart LapSVM
CN105069796B (en) SAR image segmentation method based on small echo both scatternets
CN103914705B (en) Hyperspectral image classification and wave band selection method based on multi-target immune cloning
CN108764310B (en) SAR target recognition method based on multi-scale multi-feature depth forest
CN104794730B (en) SAR image segmentation method based on super-pixel
CN105117736B (en) Classification of Polarimetric SAR Image method based on sparse depth heap stack network
CN104166859A (en) Polarization SAR image classification based on SSAE and FSALS-SVM
CN102999762B (en) Decompose and the Classification of Polarimetric SAR Image method of spectral clustering based on Freeman
CN103839073A (en) Polarization SAR image classification method based on polarization features and affinity propagation clustering
CN114595732B (en) Radar radiation source sorting method based on depth clustering
CN112560966B (en) Polarized SAR image classification method, medium and equipment based on scattering map convolution network
CN109345537B (en) SAR image segmentation method based on high-order multi-scale CRF semi-supervision
CN109446894A (en) The multispectral image change detecting method clustered based on probabilistic segmentation and Gaussian Mixture
CN105160353A (en) Polarimetric SAR data ground object classification method based on multiple feature sets
CN110490894A (en) Background separating method before the video decomposed based on improved low-rank sparse
CN107392863A (en) SAR image change detection based on affine matrix fusion Spectral Clustering
Fang et al. Polsar image classification based on complex-valued convolutional long short-term memory network
CN113392871B (en) Polarized SAR (synthetic aperture radar) ground object classification method based on scattering mechanism multichannel expansion convolutional neural network
CN109272515B (en) Unsupervised SAR image segmentation method based on high-order multi-scale CRF

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