CN107464247B

CN107464247B - Based on G0Distributed random gradient variational Bayesian SAR image segmentation method

Info

Publication number: CN107464247B
Application number: CN201710702367.1A
Authority: CN
Inventors: 刘芳; 孙宗豪; 焦李成; 李婷婷; 郝红侠; 古晶; 马文萍; 陈璞花
Original assignee: Xidian University
Current assignee: Xidian University
Priority date: 2017-08-16
Filing date: 2017-08-16
Publication date: 2021-09-21
Anticipated expiration: 2037-08-16
Also published as: CN107464247A

Abstract

The invention discloses a method based on G⁰Distribution ofThe random gradient variational Bayes SAR image segmentation method comprises the steps of extracting a sketch of an SAR image according to an initial sketch model; dividing the SAR image into a mixed pixel subspace, a homogeneous pixel subspace and a structural pixel subspace according to the regional map; estimating G for each region of extreme inhomogeneity in the mixed pixel subspace⁰Distribution parameter, using G-based⁰The distributed random gradient variational Bayesian model learns the structural characteristics of the mixed pixel, thereby realizing unsupervised segmentation of the mixed pixel subspace; and correspondingly segmenting the homogeneous pixel subspace and the structural pixel subspace, fusing segmentation results of the three subspaces, and finally obtaining an SAR image segmentation result. Because the invention assumes the prior distribution and the approximate posterior distribution of hidden variables in the model as G meeting the extremely inhomogeneous region⁰And the corresponding analytic expressions are deduced for learning, so that the clustering accuracy of the extremely inhomogeneous regions in the mixed pixel subspace is improved.

Description

Based on G0Distributed random gradient variational Bayesian SAR image segmentation method

Technical Field

The invention belongs to the technical field of image processing, and particularly relates to a G-based image processing method⁰The distributed random gradient variational Bayes SAR image segmentation method can be applied to accurately segmenting different areas of the synthetic aperture radar SAR and detecting and identifying targets in the SAR image.

Background

Synthetic Aperture Radar (SAR) is an important development in the field of remote sensing technology, and is used for acquiring high-resolution images of the earth surface. Compared with other types of imaging technologies, the SAR has very important advantages, is not influenced by atmospheric conditions such as cloud layers, rainfall or heavy fog and the like and illumination intensity, and can acquire high-resolution remote sensing data all day long and all weather. The SAR technology has important guiding significance for many fields such as military, agriculture, geography and the like.

The image segmentation refers to a process of dividing an image into a plurality of mutually disjoint areas according to characteristics such as color, gray scale and texture. The current common methods for image segmentation include: edge detection based methods, threshold based methods, region growing and watershed based methods, clustering based methods, and the like. Due to the unique imaging mechanism of the SAR, the SAR image contains a lot of speckle noise, so that the traditional methods of many optical images cannot be directly used for the segmentation of the SAR image. Traditional segmentation methods for SAR images include some methods based on clustering such as K-means, FCM, etc., and other supervised and semi-supervised methods. The features are often extracted by manual experience, but the quality of the extracted features has an important influence on the segmentation result of the SAR image. For the supervised and semi-supervised methods, label data are needed, the label data of the SAR image are few, and the cost for acquiring the label data is high. The Bayesian network has unique advantages in the aspects of expression and inference of uncertain knowledge, and the variational Bayesian inference network can not only perform unsupervised training without label data, but also effectively learn the structural characteristics implied by each pixel space, and has great significance in effective segmentation of SAR images.

An effective MSTAR SAR image segmentation method (Wuhan university school report: page 1377-page 1380 of No. 10 of 2015 of information science edition) is provided in a paper published by Wuhan university. The method comprises the steps of firstly carrying out over-segmentation operation on an image to be processed to obtain an over-segmented image area. Secondly, extracting the characteristics of the image region level and the pixel level of the image after the segmentation to obtain a characteristic vector for representing the image, and establishing the model proposed by the method by using a space hidden Dirichlet allocation model (sLDA) and a Markov Random Field (MRF) for the MSTARSAR image to obtain an energy functional. And finally, optimizing the energy functional by using a Graph-Cut algorithm and a Branch-and-Bound algorithm to obtain a final segmentation result. The method has the disadvantages that when the characteristic vector of the SAR image is obtained, the pixel-level characteristics of the SAR image are used, and the specific structural characteristics of the SAR image due to the correlation among pixels are not automatically learned, so that the structural characteristics really representing the ground feature characteristics of the SAR image are not fully utilized, and the segmentation result is not accurate enough.

Disclosure of Invention

The technical problem to be solved by the present invention is to provide a G-based solution to the above-mentioned deficiencies in the prior art⁰The distributed random gradient variational Bayes SAR image segmentation method is used for improving the accuracy of SAR image segmentation.

The invention adopts the following technical scheme:

based on G⁰The distributed random gradient variational Bayes SAR image segmentation method comprises the steps of extracting a sketch of an SAR image according to an initial sketch model; then dividing the SAR image into a mixed pixel subspace, a homogeneous pixel subspace and a structural pixel subspace according to the regional image; estimating G for each region of extreme inhomogeneity in the mixed pixel subspace⁰Distribution parameter, using G-based⁰The distributed random gradient variational Bayesian model learns the structural characteristics of the mixed pixel, and unsupervised segmentation of the mixed pixel subspace is realized; and correspondingly segmenting the homogeneous pixel subspace and the structural pixel subspace, fusing segmentation results of the three subspaces, and finally obtaining an SAR image segmentation result.

Preferably, the specific steps are as follows:

s1, inputting a synthetic aperture radar SAR image, and establishing a sketch model of the synthetic aperture radar SAR image;

s2, performing regional processing on the sketch of the synthetic aperture radar SAR image by adopting a sketch line regional method to obtain a regional diagram of the synthetic aperture radar SAR image comprising an aggregation region, a sketch-free region and a structural region;

s3, mapping a region map comprising a gathering region, a no-sketch region and a structural region into the SAR image to obtain a mixed pixel subspace, a homogeneous pixel subspace and a structural pixel subspace of the SAR image;

s4, obtaining the condition that each region meets G by using the intensity values of all pixel points in each region and adopting a Mellin transform parameter estimation method for extremely uneven regions in the mixed pixel subspace⁰Estimated values of three parameters α, γ, n required for the distribution;

s5, constructing a random gradient variational Bayesian network model for the mixed pixel subspace;

s6, performing feature learning on the mixed pixel subspace;

s7, segmenting SAR image mixed pixel subspace;

s8, extracting line targets by adopting a visual semantic rule, and then segmenting a structural pixel subspace by using a structural region segmentation method of a polynomial hidden model based on a geometric structure window to obtain a segmentation result of the structural pixel subspace;

s9, dividing the homogeneous pixel subspace by adopting a homogeneous region division method of a polynomial hidden model based on self-adaptive window selection to obtain a division result of the homogeneous pixel subspace;

and S10, combining the segmentation results of the mixed pixel subspace, the homogeneous pixel subspace and the structural pixel subspace to obtain the final segmentation result of the SAR image.

Preferably, step S1 specifically includes the following steps:

s101, randomly selecting one number within the range of [100,150] as the total number of the templates;

s102, constructing a template with edges and lines formed by pixel points and having different directions and scales, constructing an anisotropic Gaussian function by using the direction and scale information of the template, calculating a weighting coefficient of each pixel point in the template through the Gaussian function, and counting the weighting coefficients of all the pixel points in the template, wherein the number of the scales is 3-5, and the number of the directions is 18;

s103, calculating the average value of pixel points in the synthetic aperture radar SAR image corresponding to the template area coordinates:

wherein mu represents the mean value of all pixel points in the synthetic aperture radar SAR image corresponding to the coordinates of the template region, sigma represents the summation operation, g represents the coordinates corresponding to any pixel point in the omega region of the template, epsilon represents the symbol, w_gRepresenting the weight coefficient, w, of the pixel point in the omega-th region of the template at the coordinate g_gHas a value range of w_g∈[0,1]，A_gRepresenting the value of a pixel point in the synthetic aperture radar SAR image corresponding to the pixel point in the omega-th area of the template at the coordinate g;

s104, calculating a variance value of pixel points in the synthetic aperture radar SAR image corresponding to the template area coordinates:

the method comprises the steps that v represents the variance value of all pixel points in a synthetic aperture radar SAR image corresponding to the coordinates of a template area;

s105, calculating a response value of each pixel point in the synthetic aperture radar SAR image to a ratio operator:

wherein R represents the response value of each pixel point in the synthetic aperture radar SAR image to a ratio operator, min {. cndot } represents minimum operation, a and b respectively represent two different areas in a template, and mu_aRepresents the mean value, mu, of all the pixels in the template region a_bRepresenting the mean value of all pixel points in the template region b;

s106, calculating a response value of each pixel in the synthetic aperture radar SAR image to the correlation operator:

wherein C represents each in the SAR image of the synthetic aperture radarThe response value of each pixel to the correlation operator,

representing square root operations, a and b representing two different regions in the template, v_aExpress the variance value, v, of all the pixels in the template region a_bRepresents the variance value, mu, of all the pixels in the template region b_aRepresents the mean value, mu, of all the pixels in the template region a_bRepresenting the mean value of all pixel points in the template region b;

s107, calculating the response value of each pixel point in the synthetic aperture radar SAR image to each template:

wherein F represents the response value of each pixel point in the synthetic aperture radar SAR image to each template,

representing square root operation, wherein R and C respectively represent response values of pixel points in the synthetic aperture radar SAR image to a ratio operator and pixel points in the synthetic aperture radar SAR image to a correlation operator;

s108, judging whether the constructed template is equal to the total number of the selected templates, if so, executing the step S102, otherwise, executing the step S109;

s109, selecting a template with the maximum response value from the templates as a template of the SAR image, taking the maximum response value of the template as the intensity of a pixel point in the SAR image, taking the direction of the template as the direction of the pixel point in the SAR image, and obtaining a sideline response graph and a gradient graph of the SAR image;

s110, calculating the intensity value of the synthetic aperture radar SAR image intensity map to obtain an intensity map:

wherein I represents the intensity value of the synthetic aperture radar SAR image intensity map, r represents the value in the synthetic aperture radar SAR image edge response map, and t represents the value in the synthetic aperture radar SAR image gradient map;

s111, detecting the intensity map by adopting a non-maximum value inhibition method to obtain a suggested sketch;

s112, selecting the pixel point with the maximum strength in the suggested sketch, and connecting the pixel points which are communicated with the pixel point with the maximum strength in the suggested sketch to form a suggested line segment to obtain a suggested sketch;

s113, calculating the coding length gain of the sketch line in the suggested sketch:

wherein CLG represents coding length gain of sketch line in the suggested sketch, Sigma represents summation operation, J represents number of pixel points in current sketch line neighborhood, A_jExpressing the observed value of the jth pixel point in the current sketch line neighborhood, A_j,0Indicating the estimation value of the jth pixel point in the sketch line neighborhood under the condition that the current sketch line can not represent structural information, ln (·) represents the logarithm operation with e as the base, A_j,1The estimation value of the jth pixel point in the adjacent area of the sketch line is expressed under the condition that the current sketch line can express the structural information;

s114, randomly selecting one number within the range of [5,50] as a threshold value T;

and S115, selecting the recommended sketch lines with CLG > T in all the recommended sketch lines, combining the recommended sketch lines into a sketch of the SAR image, and extracting the sketch of the SAR image from the sketch model.

Preferably, step S2 specifically includes:

s201, dividing sketch lines into gathering sketch lines representing gathering ground objects and boundary sketch lines, line target sketch lines and isolated target sketch lines representing boundaries, line targets and isolated targets according to the concentration degree of sketch line segments in a sketch map of the synthetic aperture radar SAR image;

s202, according to the histogram statistics of the concentration degree of the sketch line segments, selecting the sketch line segments with the concentration degree equal to the optimal concentration degree as a seed line segment set { E_kK is 1,2,.. m }, wherein E_kRepresenting any sketch line segment in the seed line segment set, k representing the label of any sketch line segment in the seed line segment set, m representing the total number of the seed line segments, and {. represents the set operation;

s203, using the unselected line segment added into the seed line segment set as a base point, and recursively solving the line segment set by using the base point;

s204, constructing a circular element with the radius as the upper bound of the optimal concentration degree interval, expanding the line segments in the line segment set by using the circular element, corroding the expanded line segment set from outside to inside, and obtaining a concentration area with a sketch point as a unit on a sketch map;

s205, constructing a geometric structure window with the size of 5 multiplied by 5 by taking each pixel point of each sketch line as the center to obtain a structural area for the sketch lines representing the boundary, the line target and the isolated target;

s206, taking the part of the sketch except the aggregation region and the structural region as a non-sketch region;

and S207, combining the gathering area, the non-sketch area and the structural area in the sketch to obtain an area map of the synthetic aperture radar SAR image comprising the gathering area, the non-sketch area and the structural area.

Preferably, step S5 is specifically as follows:

intermediate variables from the input layer to the hidden layer of the stochastic gradient variation Bayesian network model:

wherein h is_φIntermediate variables representing the input layer to the hidden layer of the stochastic gradient variational Bayesian network model,

representing the input layer to intermediate variable h of a stochastic gradient variational Bayesian network model_φM represents the number of neurons in the hidden layer, m is 50, n represents the number of neurons in the input layer, n is 441,

to represent

A corresponding offset vector;

approximate posterior probability of stochastic gradient variational bayesian network model:

q_φ(z|x)～G⁰(α_φ,γ_φ)

wherein q is_φ(z | x) represents the approximate posterior probability, G, of a stochastic gradient variational Bayesian network model⁰(α_φ,γ_φ) Denotes the uniformity as alpha_φScale of gamma_φG of (A)⁰Distribution, G⁰Probability density formula of distribution

-α,γ,n,I(x,y)＞0

Wherein, I (x, y) is the image pixel intensity value, and n is the equivalent view; gamma is a scale parameter; alpha is the uniformity; Γ (x) is a Gamma function, defined in the real number domain as:

when the equivalent visual number n is 1, the distribution becomes Beta-Prime distribution, and the expression is:

-α,γ,I(x,y)＞0

wherein,

intermediate variable h from input layer to hidden layer of representing stochastic gradient variation Bayesian network model_φAnd-alpha_φThe connection weight of (a) is set,

to represent

The corresponding offset vector is set to the offset vector,

intermediate variable h representing input layer to hidden layer_φAnd gamma_φThe connection weight of (a) is set,

to represent

A corresponding offset vector;

intermediate variables from hidden layer to reconstructed layer of the stochastic gradient variation Bayesian network model:

wherein h is_θIntermediate variables representing hidden layer to reconstruction layer of the stochastic gradient variation Bayes network model,

hidden layer to intermediate variable h_θThe connection weight of (a) is set,

to represent

A corresponding offset vector;

conditional probability of stochastic gradient variant bayesian network model:

wherein,

representing conditional probability of a stochastic gradient variational Bayesian network model, G⁰(α_θ,γ_θ) Denotes the uniformity as alpha_θScale of gamma_θThe normal distribution of (c),

intermediate variable h representing hidden to reconstructed layers_θAnd-alpha_θThe connection weight of (a) is set,

to represent

The corresponding offset vector is set to the offset vector,

intermediate variable h representing hidden to reconstructed layers_θAnd gamma_θThe connection weight of (a) is set,

to represent

A corresponding offset vector;

lower bound of variation of stochastic gradient variation bayesian network model:

wherein L (theta, phi) represents a variation lower bound of the stochastic gradient variation Bayes network model, phi represents a variation parameter of the stochastic gradient variation Bayes network model,

theta represents a generation parameter of the stochastic gradient variation bayesian network model,

D_KL(q_φ(z|x)||p_θ(z)) represents q_φ(z | x) and p_θ(z) relative entropy between z representing hidden layer variables of the stochastic gradient variational Bayesian network model, p_θ(z) represents the prior probability of the hidden layer variable z, Σ (-) represents the summation operation, L represents the number of times the hidden layer variable z was gaussian sampled, log (-) represents the logarithm operation, z (-) represents the sum of the hidden layer variable z^lExpressing the result of the first Gaussian sampling of z, and the value of the result is expressed by a formula

Get, wherein, <' > indicates a dot product operation, ∈^lRepresenting auxiliary variables, epsilon, of Gaussian samples^lN (0, I) indicates that the Gaussian sampled auxiliary variable satisfies the standard normal distribution.

Preferably, two cases of equivalent view n ≠ 1 and n ≠ 1 are considered:

when the equivalent view n is 1, the following is calculated:

-D_KL(q_φ(z|x)||p_θ(z))＝∫P^β'(z；-α,γ)logP^β'(z；c,γ₁)-P^β'(z；-α,γ)logP^β'(z；-α,γ)dz

wherein p is a priori_θ(z) satisfies- α ═ c, γ ═ γ₁Beta-Primer distribution, c, gamma of₁Are all positive numbers, are known quantities derived from the image, and approximate a posteriori q_φ(z | x) satisfiesBeta-Primer distribution, where z ∈ [ a, b ]]A is more than 0 and less than b and less than or equal to 1, and the normalized image pixel intensity value is represented;

when the equivalent view n ≠ 1, it is calculated as follows:

wherein p is a priori_θ(z) satisfies- α ═ c, γ ═ γ₁G of (A)⁰Distribution, c, gamma₁Are all positive numbers, which are known quantities derived from the image; approximate posterior q_φ(z | x) satisfies G⁰Distribution, wherein z ∈ [ a, b ]]And a is more than 0 and less than b and less than or equal to 1, which represents the normalized image pixel intensity value.

Preferably, step S6 includes the steps of:

s601, performing area division on a mixed pixel subspace of the synthetic aperture radar SAR image according to spatial connectivity, and if only one non-connected area exists, executing the step S7;

s602, sampling each unconnected area at intervals according to a 21 x 21 window to obtain a plurality of image block samples corresponding to each area;

s603, for each non-communicated area, generating a group of non-uniform ground object distribution G corresponding to each area⁰A distributed random number;

s604, initializing a connection weight of the stochastic gradient variant Bayesian network by using a group of random numbers corresponding to each region for each region, and obtaining the initialized stochastic gradient variant Bayesian network;

s605, initializing each unconnected area, then using the image block sample as an input layer of the random gradient variation Bayesian network, and training the initialized random gradient variation Bayesian network by using a random gradient variation Bayesian inference method according to the following steps to obtain the trained random gradient variation Bayesian network;

s606, for each region which is not communicated with each other, the weight of the trained random gradient variational Bayesian network is taken

As a feature set for that region.

Preferably, step S603 specifically includes:

the first step is as follows: calculating uneven ground object distribution G of synthetic aperture radar SAR image⁰Probability density of distribution:

wherein, P (I (x, y)) represents the probability density of the uneven ground feature distribution of the synthetic aperture radar SAR image, I (x, y) represents the intensity value of the pixel point with coordinates (x, y), n represents the equivalent view of the synthetic aperture radar SAR image, α represents the shape parameter of the synthetic aperture radar SAR image, γ represents the scale parameter of the synthetic aperture radar SAR image, and Γ (·) represents a gamma function, the value of which is obtained by the following formula:

wherein u represents an independent variable, ^ represents an integral operation, t represents an integral variable, and a mixed pixel subspace region R is randomly selected_iThe 50 image block samples form a matrix A of 441 × 50;

the second step is that: non-uniform terrain distribution G through SAR images using matrix A⁰The probability density function of the distribution generates a matrix B of 441 multiplied by 50, the data in the matrix B satisfies the uneven ground object distribution G of the SAR image of the synthetic aperture radar⁰And (4) distribution.

Preferably, step S604 specifically includes:

the first step is as follows: taking the matrix B as an input layer x to an intermediate variable h of the stochastic gradient variation Bayesian network model_φConnection weight of

Second step of: randomly selecting 50 rows from the matrix B to form a matrix C of 50 multiplied by 50, and taking the matrix C as an intermediate variable h of the stochastic gradient variational Bayesian network model_φTo-alpha_φConnection weight of

Taking the matrix C as an intermediate variable h of the stochastic gradient variation Bayesian network model_φTo gamma_φConnection weight

Taking the matrix C as an implicit layer z to an intermediate variable h of the stochastic gradient variation Bayesian network model_θConnection weight of

The third step: taking the transposition of the matrix B as an intermediate variable h of the stochastic gradient variational Bayesian network model_θTo-alpha_θConnection weight of

Taking the transposition of the matrix B as an intermediate variable h of the stochastic gradient variational Bayesian network model_θTo gamma_θConnection weight of

Preferably, step S605 specifically includes:

firstly, initializing the prior probability of a hidden layer of a stochastic gradient variational Bayesian network model to G⁰Distribution probability, the approximate posterior probability of the stochastic gradient variation Bayesian network model is initialized to G⁰And distributing the probability to obtain an analytic expression of a variation lower bound of the random gradient variation Bayesian network model as follows:

(a) when the equivalent view n is 1,

(b) when the equivalent view number n ≠ 1,

wherein F'_m＝(1/(α-(m-1))((nb)^m-1log(nb+γ₁)(nb+γ₁)^m-1-α

-(na)^m-1log(na+γ₁)(na+γ₁)^m-1-α-(m-1)F′_m-1-G'_m-1))

G'_m＝1/(α-m)·((n^b)^m(n^b+γ)^α-m-(na)^m(na+γ)^α-m-mG'_m-1)；

And secondly, updating generation parameters of the stochastic gradient variation Bayesian network model:

wherein, theta^t+1Represents the generation parameter theta of the stochastic gradient variation Bayesian network model after the t +1 iteration^tRepresenting the generation parameters of the stochastic gradient variation Bayesian network model after the t-th iteration,

an operation of obtaining a partial derivative of a parameter theta of L (theta, phi);

and thirdly, updating variation parameters of the stochastic gradient variation Bayesian network model:

wherein phi is^t+1Represents the variation parameter phi of the stochastic gradient variation Bayes network model after the t +1 iteration^tRepresenting the variation parameters of the stochastic gradient variation Bayesian network model after the t-th iteration,

an operation of partial derivation of a parameter phi of L (theta, phi) is expressed;

step four, judging whether the times of the variation lower bound invariance reaches a threshold value of 100, if so, executing the step five; otherwise, executing the second step;

and fifthly, finishing the training of the stochastic gradient variational Bayesian network.

Compared with the prior art, the invention has at least the following beneficial effects:

the invention is based on G⁰The distributed random gradient variational Bayes SAR image segmentation method comprises the steps of extracting a sketch of an SAR image according to an initial sketch model; then dividing the SAR image into a mixed pixel subspace, a homogeneous pixel subspace and a structural pixel subspace according to the regional image; estimating G for each region of extreme inhomogeneity in the mixed pixel subspace⁰Distribution parameter, using G-based⁰The distributed random gradient variational Bayesian model learns the structural characteristics of the mixed pixel, and unsupervised segmentation of the mixed pixel subspace is realized; the invention sets a random gradient variational Bayesian network, carries out unsupervised training through each non-communicated region of the mixed pixel subspace, takes the weight of the trained network as the structural characteristic of each non-communicated region, overcomes the defect that the prior art uses the pixel level characteristic of the SAR image to obtain the characteristic vector of the SAR image, but does not learn the special structural characteristic of the SAR image due to the correlation between the pixels, and can automatically extract the structural characteristic of the SAR image.

Further, in the SAR image sketch, a black line is a sketch line. It can be seen that the clustered regions have many clustered sketch lines, and the sketch lines representing edges and line objects are not clustered. This means that the sketch lines of different areas have different properties, on the basis of which the degree of concentration is used to divide the sketch lines into aggregated sketch lines and non-aggregated sketch lines.

Further, an aggregation area is extracted on the aggregated sketch lines, a structural area is extracted on the non-aggregated sketch lines, and the rest area is the area without the sketch lines. Therefore, the SAR image can be divided into an aggregation region, a non-sketch line region and a structural region based on the sketch map, the obtained region map is a sparse representation of the SAR image, and the SAR image is reduced to a space with a relatively single structure and used for guiding SAR image segmentation.

Furthermore, when the invention is used for carrying out feature learning on each region which is not communicated with each other in the mixed pixel subspace, the pixel intensity value of each region is respectively used for estimating G⁰Parameters of distribution, then by G⁰The probability density function of the distribution generates random data to initialize the weight of the network, and overcomes the defect that the network is initialized by random distribution in the depth self-coding network for automatically extracting the image characteristics in the prior art without catching the intrinsic characteristics of the SAR image, so that the intrinsic characteristics of the ground features of the SAR image can be effectively learned by adopting the method and the device, and the accuracy of SAR image segmentation is improved.

Further, since the present invention uses G⁰Distribution as a distribution of hidden layer variable priors and conditional probabilities, G⁰The distribution can better depict the statistical characteristics of the SAR image, and compared with the Gaussian hypothesis, the distribution can better depict the characteristics of the SAR image, so that the method can learn the characteristics of the ground features of the SAR image, and further improve the accuracy of SAR image segmentation.

In summary, the present invention assumes both the hidden variable prior distribution and the approximate posterior distribution in the model as G satisfying the extremely heterogeneous region⁰And the corresponding analytic expressions are deduced for learning, so that the clustering accuracy of extremely heterogeneous regions in the mixed pixel subspace is improved.

The technical solution of the present invention is further described in detail by the accompanying drawings and embodiments.

Drawings

FIG. 1 is a flow chart of the present invention;

FIG. 2 is a Bayesian inference model diagram of the present invention;

FIG. 3 is a simulation of the present invention;

FIG. 4 is a diagram illustrating simulation results of the present invention.

Detailed Description

The invention provides a method based on G⁰The distributed random gradient variational Bayes SAR image segmentation method comprises the steps of extracting a sketch of an SAR image according to an initial sketch model; dividing the SAR image into a mixed pixel subspace, a homogeneous pixel subspace and a structural pixel subspace according to the regional map; estimating G for each region of extreme inhomogeneity in the mixed pixel subspace⁰Distribution parameter, using G-based⁰The distributed random gradient variational Bayesian model learns the structural characteristics of the mixed pixel, thereby realizing unsupervised segmentation of the mixed pixel subspace; and correspondingly segmenting the homogeneous pixel subspace and the structural pixel subspace, fusing segmentation results of the three subspaces, and finally obtaining an SAR image segmentation result.

Referring to FIG. 1, the present invention is based on G⁰The distributed random gradient variational Bayes SAR image segmentation method specifically comprises the following steps:

and S1, sketching SAR images.

And inputting a synthetic aperture radar SAR image.

Establishing a sketch model of the SAR image according to the following steps:

s103, calculating the average value of pixel points in the synthetic aperture radar SAR image corresponding to the template area coordinates according to the following formula:

wherein, mu represents and modeThe method comprises the steps that the mean value of all pixel points in a synthetic aperture radar SAR image corresponding to the coordinates of a plate area is sigma, summation operation is represented, g represents the corresponding coordinates of any pixel point in the omega-th area of a template, epsilon represents the symbol belonging to, and w represents the symbol belonging to_gRepresenting the weight coefficient, w, of the pixel point in the omega-th region of the template at the coordinate g_gHas a value range of w_g∈[0,1]，A_gRepresenting the value of a pixel point in the synthetic aperture radar SAR image corresponding to the pixel point in the omega-th area of the template at the coordinate g;

s104, calculating a variance value of pixel points in the synthetic aperture radar SAR image corresponding to the template area coordinates according to the following formula:

s105, calculating the response value of each pixel point in the synthetic aperture radar SAR image to the ratio operator according to the following formula:

s106, calculating a response value of each pixel in the synthetic aperture radar SAR image to the correlation operator according to the following formula:

where C denotes per pixel in the synthetic aperture radar SAR image for the correlation operatorIn response to the value of the response(s),

s107, calculating the response value of each pixel point in the SAR image to each template according to the following formula:

s108, judging whether the constructed template is equal to the total number of the selected templates, if so, executing S102, otherwise, executing S109;

s110, calculating the intensity value of the synthetic aperture radar SAR image intensity map according to the following formula to obtain the intensity map:

s113, calculating the coding length gain of the sketch line in the suggested sketch according to the following formula:

and S115, selecting the recommended sketch lines with CLG > T in all the recommended sketch lines, and combining the recommended sketch lines into a sketch of the SAR image.

And extracting a sketch map of the SAR image from the sketch model.

The synthetic aperture radar SAR image sketch model used by the invention is a model proposed by Jie-Wu et al in 2014 in an article Local maximum geographic region search for SAR specific reduction with sketch-based geographic region search function in IEEE Transactions on Geoscience and Remote Sensing journal.

And S2, obtaining an area map by sketch.

And carrying out regional processing on the sketch map of the synthetic aperture radar SAR image by adopting a sketch line regional method to obtain a regional map of the synthetic aperture radar SAR image comprising an aggregation region, a sketch-free region and a structural region.

S201, dividing sketch lines into gathering sketch lines representing gathering ground objects and boundary sketch lines, line target sketch lines and isolated target sketch lines representing boundaries, line targets and isolated targets according to the concentration degree of sketch line segments in a sketch map of the synthetic aperture radar SAR image.

S202, according to the histogram statistics of the concentration degree of the sketch line segments, selecting the sketch line segments with the concentration degree equal to the optimal concentration degree as a seed line segment set { E_kK is 1,2,.. m }, wherein E_kRepresenting any sketch line segment in the seed line segment set, k representing the label of any sketch line segment in the seed line segment set, m representing the total number of the seed line segments, and {. represents the set operation.

S203, using the unselected line segment added into the seed line segment set as the base point, and recursively solving the line segment set by the base point.

S204, constructing a circular element with the radius as the upper bound of the optimal concentration degree interval, expanding the line segments in the line segment set by using the circular element, corroding the expanded line segment set from outside to inside, and obtaining a concentration area with a sketch point as a unit on a sketch map.

S205, a geometric structure window of 5 × 5 in size is constructed with each pixel point of each sketch line as the center, for the sketch lines representing the boundary, line object, and isolated object, to obtain a structure region.

S206, the parts except the aggregation areas and the structural areas in the sketch are taken as non-sketch areas.

S3, dividing pixel subspace

And mapping a region map comprising the aggregation region, the no-sketch region and the structural region into the synthetic aperture radar SAR image to obtain a mixed pixel subspace, a homogeneous pixel subspace and a structural pixel subspace of the synthetic aperture radar SAR image.

S4, obtaining the condition that each region meets G by using the intensity values of all pixel points in each region and adopting a Mellin transform parameter estimation method for extremely uneven regions in the mixed pixel subspace⁰The estimated values of the three parameters α, γ, n required for the distribution.

And S5, constructing a stochastic gradient variational Bayesian network model.

Calculating intermediate variables from the input layer to the hidden layer of the stochastic gradient variation Bayesian network model according to the following formula:

to represent

The corresponding offset vector.

Calculating the approximate posterior probability of the stochastic gradient variational Bayesian network model according to the following formula:

q_φ(z|x)～G⁰(α_φ,γ_φ)

-α,γ,n,I(x,y)＞0

-α,γ,I(x,y)＞0

wherein,

to represent

The corresponding offset vector is set to the offset vector,

to represent

The corresponding offset vector.

Calculating intermediate variables from the hidden layer to the reconstruction layer of the stochastic gradient variation Bayesian network model according to the following formula:

hidden layer to intermediate variable h_θThe connection weight of (a) is set,

to represent

The corresponding offset vector.

The conditional probability of the stochastic gradient variational bayesian network model is calculated according to the following formula:

wherein,

to represent

The corresponding offset vector is set to the offset vector,

to represent

The corresponding offset vector.

Calculating the variation lower bound of the stochastic gradient variation Bayesian network model according to the following formula:

Get, wherein, <' > indicates a dot product operation, ∈^lRepresenting auxiliary variables, epsilon, of Gaussian samples^lN (0, I), indicating that the Gaussian sampled auxiliary variable satisfies the standard normal distribution.

D in the lower bound (1) of the variation_KL(q_φ(z|x)||p_θ(z)) the present invention is based on q_φ(z | x) and p_θ(z) are all G⁰The distribution assumption pushes down to derive an analytical expression rather than using a gaussian distribution because the extremely inhomogeneous region of the SAR image satisfies G⁰And (4) distribution. In addition G⁰The distribution differs in the form in the case where the equivalent view n is equal to 1 and not equal to 1, so the invention considers the case where the equivalent view n is equal to 1 and n is equal to 1, respectively:

when the equivalent view n is 1, the calculation formula is as follows:

-D_KL(q_φ(z|x)||p_θ(z))＝∫P^β'(z；-α,γ)logP^β'(z；c,γ₁)-P^β'(z；-α,γ)logP^β'(z；-α,γ)dz (2)

wherein p is a priori_θ(z) satisfies- α ═ c, γ ═ γ₁Beta-Primer distribution, c, gamma of₁Are all positive numbers and are known quantities derived from the image. Approximate posterior q_φ(z | x) satisfies a Beta-Primer distribution where z ∈ [ a, b ]]And a is more than 0 and less than b and less than or equal to 1, and the normalized image pixel intensity value is also a known quantity. The derivation process is omitted here, and the final form is directly given:

when the equivalent view number n ≠ 1, the calculation formula is as follows:

wherein p is a priori_θ(z) satisfies- α ═ c, γ ═ γ₁G of (A)⁰Distribution, c, gamma₁Are all positive numbers and are known quantities derived from the image. Approximate posterior q_φ(z | x) satisfies G⁰Distribution, where z ∈ [ a, b ]]And a is more than 0 and less than b and less than or equal to 1, and the normalized image pixel intensity value is also a known quantity. The derivation process is omitted here, and the final form is directly given:

-D_KL(q_φ(z|x)||p_θ(z))＝logn+clogγ₁-logΓ(n-α)+logΓ(n)+logΓ(-α)-αlogγ+(n-1)nⁿΓ(n-α)/(γ^αΓ(n)Γ(-α))·(1/(n(α-(n-1)))·(lnb·b^n-1/(γ₁+nb)^n-1-α

-lna·a^n-1/(γ₁+na)^n-1-α-lnb·b^n-1/(γ+nb)^n-1-α+lna·a^n-1/(γ+na)^n-1-α))

-(n+c)Γ(n-α)/(γ^αΓ(n)Γ(-α))·(1/(α-(n-2))·((nb)^n-2log(nb+γ₁)(nb+γ₁)^n-2-α-(na)^n-2log(na+γ₁)(na+γ₁)^n-2-α-(n-1)F′_n-2-G'_n-2))

+(n-α)Γ(n-α)/(γ^αΓ(n)Γ(-α))(1/(α-(n-2))·((nb)^n-2log(nb+γ)(nb+γ)^n-2-α

-(na)^n-2log(na+γ)(na+γ)^n-2-α-(n-1)F′_n-2-G'_n-2))

wherein F'_m＝(1/(α-(m-1))((nb)^m-1log(nb+γ₁)(nb+γ₁)^m-1-α

-(na)^m-1log(na+γ₁)(na+γ₁)^m-1-α-(m-1)F′_m-1-G'_m-1))，

G'_m＝1/(α-m)·((nb)^m(nb+γ)^α-m-(na)^m(na+γ)^α-m-mG'_m-1),m≥1。

And S6, feature learning of the mixed pixel subspace.

S601, carrying out region division on the mixed pixel subspace of the synthetic aperture radar SAR image according to spatial connectivity, and if only one region is not connected with each other, executing the step S7.

And S602, sampling each unconnected area at intervals according to a 21 x 21 window to obtain a plurality of image block samples corresponding to each area.

S603, for each non-communicated area, generating a group of non-uniform ground object distribution G corresponding to each area⁰Random number of distribution:

the first step is as follows: according to the following formula, calculating the uneven ground object distribution G of the SAR image⁰Probability density of distribution:

wherein u represents an independent variable, ^ represents an integral operation, and t represents an integral variable.

Randomly selecting a blend pixel subspace region R_iOf the image block samples, constitute a matrix a of 441 × 50.

The second step is that: using the matrix A, byUneven ground object distribution G of aperture-forming radar SAR image⁰The probability density function of the distribution generates a matrix B of 441 multiplied by 50, the data in the matrix B satisfies the uneven ground object distribution G of the SAR image of the synthetic aperture radar⁰And (4) distribution.

S604, initializing the connection weight of the stochastic gradient variant Bayes network by using a group of random numbers corresponding to each region for each region, so as to obtain the initialized stochastic gradient variant Bayes network:

The second step is that: randomly selecting 50 rows from the matrix B to form a matrix C of 50 multiplied by 50, and taking the matrix C as an intermediate variable h of the stochastic gradient variational Bayesian network model_φTo-alpha_φConnection weight of

S605, initializing each unconnected area, then using the image block sample as an input layer of the stochastic gradient variation Bayesian network, and training the initialized stochastic gradient variation Bayesian network by using a stochastic gradient Bayesian inference method according to the following steps to obtain the trained stochastic gradient variation Bayesian network:

step 1, initializing prior probability of hidden layer of stochastic gradient variation Bayesian network model to G⁰Distribution probability, the approximate posterior probability of the stochastic gradient variation Bayesian network model is initialized to G⁰And distributing the probability to obtain an analytic expression of a variation lower bound of the random gradient variation Bayesian network model as follows:

(a) when the equivalent view n is 1,

(b) when the equivalent view number n ≠ 1,

wherein F'_m＝(1/(α-(m-1))((nb)^m-1log(nb+γ₁)(nb+γ₁)^m-1-α

-(na)^m-1log(na+γ₁)(na+γ₁)^m-1-α-(m-1)F′_m-₁-G'_m-1))

G'_m＝1/(α-m)·((nb)^m(nb+γ)^α-m-(na)^m(na+γ)^α-m-mG'_m-1)

And 2, updating generation parameters of the stochastic gradient variation Bayesian network model according to the following formula:

and 3, updating variation parameters of the stochastic gradient variation Bayesian network model according to the following formula:

step 4, judging whether the times of the variation lower bound invariance reaches a threshold value 100, if so, executing step 5; otherwise, executing the step 2;

and 5, finishing the training of the stochastic gradient variational Bayesian network.

As a feature set for that region.

And S7, segmenting the SAR image mixed pixel subspace.

And splicing the feature sets of all the unconnected areas, and taking the spliced feature sets as a codebook.

And respectively calculating the inner product of all the characteristics of each unconnected area and each characteristic in the codebook to obtain the projection vector of all the characteristics of each area on the codebook.

And carrying out maximum value convergence on all projection vectors of each unconnected area to obtain a structural feature vector corresponding to each area.

And clustering the structural feature vectors of all the unconnected areas by using a hierarchical clustering algorithm to obtain a segmentation result of the mixed pixel subspace.

And S8, dividing the structural pixel subspace.

And extracting line targets by adopting a visual semantic rule, and then segmenting the structural pixel subspace by using a structural region segmentation method of a polynomial hidden model based on a geometric structure window to obtain a segmentation result of the structural pixel subspace.

The method for segmenting the structural region used in the present invention is a Model proposed by Fan-Liu et al in 2016 in the article SAR Image Segmentation base on Hierarchical Visual search and Adaptive neighbor patent Model, published in the journal of IEEE transactions on geographic and Remote Sensing.

And S9, dividing the homogeneous pixel subspace.

And (3) dividing the homogeneous pixel subspace by adopting a homogeneous region division method of a polynomial hidden model based on self-adaptive window selection to obtain a division result of the homogeneous pixel subspace.

The homogeneous region Segmentation method Based on Adaptive window selection polynomial hidden Model is a Model proposed by Fan-Liu et al in 2016 in the article SAR Image Segmentation Based on high performance Visual search and Adaptive neighbor multinominal Model published in the journal of IEEE transactions on science and Remote Sensing.

The effect of the present invention will be further described with reference to the simulation diagram.

1. Simulation conditions are as follows:

the hardware conditions of the simulation of the invention are as follows: a smart sensing and image understanding laboratory graphics workstation; the synthetic aperture radar SAR image used by the simulation of the invention is as follows: and the resolution of the X wave band is 1 meter in a Pyramid graph.

2. Simulation content:

the simulation experiment of the present invention is to segment the Pyramid map in the SAR image, as shown in fig. 3 (a). The figure is derived from a synthetic aperture radar SAR image with a resolution of 1 meter in the X-band.

The Pyramid image shown in fig. 3(a) is rasterized by the SAR image rasterization step of the present invention, and a sketch image shown in fig. 3(b) is obtained.

The pixel rendering shown in fig. 3(b) is regionalized by the pixel subspace division step of the present invention, resulting in a region map as shown in fig. 3 (c). The white space in fig. 3(c) indicates the aggregation region, and the rest are the non-sketch region and the texture region. From the region map shown in fig. 3(c), a Pyramid image mixed pixel subspace map as shown in fig. 3(d) can be obtained.

Based on G by adopting the invention⁰The distributed stochastic gradient variant bayesian network segments the Pyramid mixed pixel subspace map shown in fig. 3(d) to obtain the cluster structure of the mixed pixel subspace shown in fig. 4(a), wherein the gray areas represent unprocessed terrain spaces, the remaining areas of the same color represent the same terrain space, and the areas of different colors represent different terrain spaces.

By using the merge segmentation result step of the present invention, the mixed pixel subspace segmentation result, the homogeneous pixel subspace segmentation result and the structural pixel subspace segmentation result shown in fig. 4(a) are merged to obtain fig. 4(c), and fig. 4(c) is a final segmentation result diagram of the Pyramid image shown in fig. 3 (a).

3. Simulation effect analysis:

fig. 4(a) is a mixed pixel subspace segmentation result graph of the Pyramid image by the method of the present invention, and fig. 4(b) is a mixed pixel subspace segmentation result graph obtained by a gaussian-distributed random gradient variant bayesian network. FIG. 4(c) is a final segmentation result of the Pyramid image according to the method of the present invention, and FIG. 4(d) is a final segmentation result based on the assumption of Gaussian distribution. The comparison can be used for drawing a conclusion that the method is more reasonable for the segmentation result of the mixed pixel subspace. The method provided by the invention is used for segmenting the SAR image, so that the accuracy of SAR image segmentation is improved.

The above-mentioned contents are only for illustrating the technical idea of the present invention, and the protection scope of the present invention is not limited thereby, and any modification made on the basis of the technical idea of the present invention falls within the protection scope of the claims of the present invention.

Claims

1. Based on G⁰The distributed random gradient variational Bayes SAR image segmentation method is characterized in that a sketch of an SAR image is extracted according to an initial sketch model; then dividing the SAR image into a mixed pixel subspace, a homogeneous pixel subspace and a structural pixel subspace according to the regional image; estimating G for each region of extreme inhomogeneity in the mixed pixel subspace⁰Distribution parameter, using G-based⁰The distributed random gradient variational Bayesian model learns the structural characteristics of the mixed pixel, and unsupervised segmentation of the mixed pixel subspace is realized; correspondingly segmenting the homogeneous pixel subspace and the structural pixel subspace, fusing segmentation results of the three subspaces, and finally obtaining an SAR image segmentation result, wherein the method comprises the following specific steps:

s1, inputting a synthetic aperture radar SAR image, and establishing a sketch model of the synthetic aperture radar SAR image, which specifically comprises the following steps:

wherein R represents the response value of each pixel point in the synthetic aperture radar SAR image to a ratio operator, min {. cndot } represents minimum operation, a and b respectively represent two different areas in a template, and mu_aRepresenting a template regionMean value of all pixels in a, mu_bRepresenting the mean value of all pixel points in the template region b;

wherein C represents the response value of each pixel in the synthetic aperture radar SAR image to the correlation operator,

s115, selecting the recommended sketch lines with CLG > T in all the recommended sketch lines, combining the recommended sketch lines into a sketch of the SAR image, and extracting the sketch of the SAR image from the sketch model;

s2, performing regional processing on the sketch map of the synthetic aperture radar SAR image by adopting a sketch line regional method to obtain a regional map of the synthetic aperture radar SAR image comprising an aggregation region, a sketch-free region and a structural region, wherein the method specifically comprises the following steps:

s207, combining the gathering area, the non-sketch area and the structural area in the sketch to obtain an area map of the synthetic aperture radar SAR image comprising the gathering area, the non-sketch area and the structural area;

s5, constructing a stochastic gradient variational Bayesian network model for the mixed pixel subspace, which comprises the following steps:

to represent

A corresponding offset vector;

wherein,

to represent

The corresponding offset vector is set to the offset vector,

to represent

A corresponding offset vector;

hidden layer to intermediate variable h_θThe connection weight of (a) is set,

to represent

A corresponding offset vector;

conditional probability of stochastic gradient variant bayesian network model:

wherein,

bayes expressing stochastic gradient variationalConditional probability of network model, G⁰(α_θ,γ_θ) Denotes the uniformity as alpha_θScale of gamma_θThe normal distribution of (c),

to represent

The corresponding offset vector is set to the offset vector,

to represent

A corresponding offset vector;

D_KL(q_φ(z|x)||p_θ(z)) represents q_φ(z | x) and p_θ(z) relative entropy between z representing hidden layer variables of the stochastic gradient variational Bayesian network model, p_θ(z) represents the prior probability of the hidden layer variable z,

indicating a summation operation, L indicates the number of gaussian samples taken by the hidden layer variable z,

denotes a logarithmic operation, z^lExpressing the result of the first Gaussian sampling of z, and the value of the result is expressed by a formula

To obtain a mixture of, among others,

represents a dot product operation, ε^lRepresenting a gaussian sampled auxiliary variable,

representing that the auxiliary variable of Gaussian sampling meets the standard normal distribution;

two cases of equivalent view n ≠ 1 and n ≠ 1 are considered:

when the equivalent view n is 1, the following is calculated:

-D_KL(q_φ(z|x)||p_θ(z))＝∫P^β'(z；-α,γ)logP^β'(z；c,γ₁)-P^β'(z；-α,γ)^logP^β'(z；-α,γ)dz

wherein p is a priori_θ(z) satisfies- α ═ c, γ ═ γ₁Beta-Primer distribution, c, gamma of₁Are all positive numbers, are known quantities derived from the image, and approximate a posteriori q_φ(z | x) satisfies a Beta-Primer distribution where z ∈ [ a, b ]]And a is more than 0 and less than b and less than or equal to 1 to represent the normalized image pixel intensity value；

When the equivalent view n ≠ 1, it is calculated as follows:

wherein p is a priori_θ(z) satisfies- α ═ c, γ ═ γ₁G of (A)⁰Distribution, c, gamma₁Are all positive numbers, which are known quantities derived from the image; approximate posterior q_φ(z | x) satisfies G⁰Distribution, wherein z ∈ [ a, b ]]A is more than 0 and less than b and less than or equal to 1, and the normalized image pixel intensity value is represented;

s6, learning the characteristics of the mixed pixel subspace, comprising the following steps:

s603, for each non-communicated area, generating a group of non-uniform ground object distribution G corresponding to each area⁰The distributed random numbers specifically include:

wherein P (I (x, y)) represents the probability density of uneven ground feature distribution of the SAR image, I (x, y) represents the intensity value of a pixel point with coordinates of (x, y), n represents the equivalent view of the SAR image, alpha represents the shape parameter of the SAR image, gamma represents the scale parameter of the SAR image,

represents a gamma function whose value is given by:

the second step is that: non-uniform terrain distribution G through SAR images using matrix A⁰The probability density function of the distribution generates a matrix B of 441 multiplied by 50, the data in the matrix B satisfies the uneven ground object distribution G of the SAR image of the synthetic aperture radar⁰Distributing;

s604, initializing a connection weight of the stochastic gradient variant bayesian network with a group of random numbers corresponding to each region for each region, to obtain an initialized stochastic gradient variant bayesian network, which specifically includes:

Taking the matrix C as the hidden layer z to middle variation of the stochastic gradient variation Bayesian network modelAmount h_θConnection weight of

S605, initializing each unconnected region, then using the image block sample as an input layer of the stochastic gradient variation Bayesian network, and training the initialized stochastic gradient variation Bayesian network by using a stochastic gradient Bayesian inference method according to the following steps to obtain the trained stochastic gradient variation Bayesian network, wherein the step S605 specifically comprises:

(a) when the equivalent view n is 1,

(b) when the equivalent view number n ≠ 1,

wherein F'_m＝(1/(α-(m-1))((nb)^m-1log(nb+γ₁)(nb+γ₁)^m-1-α-(na)^m-1log(na+γ₁)(na+γ₁)^m-1-α-(m-1)F'_m-1-G'_m-1))

fifthly, finishing the training of the stochastic gradient variational Bayesian network;

As a feature set for the region;

s7, segmenting SAR image mixed pixel subspace;