CN108171200B - SAR image classification method based on SAR image statistical distribution and DBN - Google Patents

SAR image classification method based on SAR image statistical distribution and DBN Download PDF

Info

Publication number
CN108171200B
CN108171200B CN201810031937.3A CN201810031937A CN108171200B CN 108171200 B CN108171200 B CN 108171200B CN 201810031937 A CN201810031937 A CN 201810031937A CN 108171200 B CN108171200 B CN 108171200B
Authority
CN
China
Prior art keywords
boltzmann machine
layer
hidden layer
node
dbn
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
CN201810031937.3A
Other languages
Chinese (zh)
Other versions
CN108171200A (en
Inventor
侯彪
焦李成
梁亚敏
马晶晶
马文萍
王爽
白静
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 CN201810031937.3A priority Critical patent/CN108171200B/en
Publication of CN108171200A publication Critical patent/CN108171200A/en
Application granted granted Critical
Publication of CN108171200B publication Critical patent/CN108171200B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/10—Terrestrial scenes
    • G06V20/13—Satellite images
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/044—Recurrent networks, e.g. Hopfield networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • General Engineering & Computer Science (AREA)
  • Computing Systems (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computational Linguistics (AREA)
  • Biophysics (AREA)
  • Biomedical Technology (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Biology (AREA)
  • Astronomy & Astrophysics (AREA)
  • Remote Sensing (AREA)
  • Multimedia (AREA)
  • Image Analysis (AREA)

Abstract

The invention discloses an SAR image classification method based on SAR image statistical distribution and DBN, which mainly solves the problems that the traditional deep confidence network DBN is used for SAR image classification, and the regional consistency is poor and the edge information is incomplete. The realization process is as follows: preprocessing an SAR image to be classified to obtain an input matrix of a DBN; designing a DBN consisting of 3 restricted Boltzmann machines; pre-training the designed network by using an input matrix to obtain a trained DBN; randomly selecting part of pixel sets with category labels from labels of the SAR image category label graph, and performing micro-adjustment on the trained DBN by using a back propagation algorithm; and (4) carrying out pixel-by-pixel classification on the image to be classified by using the micro-adjusted DBN to obtain a classification result, and coloring and outputting the classification result. The method has the advantages of excellent classification result, good regional consistency and complete edge information, and can be applied to terrain classification and target identification of SAR images.

Description

SAR image classification method based on SAR image statistical distribution and DBN
Technical Field
The invention belongs to the field of image processing, and particularly relates to an SAR image classification method which can be applied to terrain classification and target identification of an SAR image.
Background
The synthetic aperture radar SAR is an airborne radar or a satellite-borne radar which can generate high-resolution images, has the characteristic of all weather all the day, and is widely applied to the fields of remote sensing, mapping and the like. Synthetic aperture radar images, SAR images, are topographical images of the ground surface that reflect radar signals that illuminate the ground surface and have differences in light and dark tones, much like black and white photographs. For interpretation of the SAR image, compared with an optical image, the SAR image classification is difficult, and the objective to be realized by the SAR image classification is to label different ground object types with different colors through a designed classification algorithm.
The statistical model of the SAR image plays a very important role in the aspects of detection and identification of the ground object target, elimination of speckle noise, classification of the ground object target and the like. Statistical models of SAR images are roughly classified into two categories according to the modeling process of the SAR images: parametric models and non-parametric models. The parametric model is to calculate each parameter of the probability density function through the existing SAR image data under the condition of unknown parameters of the probability density function, and then select the probability density function with the best fitting degree as a statistical model of the ground feature region of the SAR image. For example, gamma distribution, Weibull distribution, Log-normal distribution, and the like. The nonparametric model refers to that the best probability density function is directly selected according to a certain rule according to the data of the ground feature region of the SAR image.
The feature extraction method for the classification of the SAR image can be roughly classified into three types: statistical methods, transform domain methods, and model-based methods. Wherein:
a typical representative statistical method is a gray level co-occurrence matrix characteristic proposed by u.kandaswamy et al, see u.kandaswamy, d.a.adjeroh, and m.c.lee, effective texture analysis of SAR image, IEEE trans.geos.remote sens.,2005,43(9): 2075-2083, which includes variance, contrast, energy, etc., since these information are only features based on statistics of the SAR image, spatial neighborhood characteristics of the SAR image are ignored, and thus the spatial consistency of the classification result obtained by the statistical method is poor.
Transform domain methods, including fourier transforms, wavelet transforms and Garbor wavelet transforms, are mainly used for texture analysis of SAR Images, see e.g. akbaizadeh, a new static-based wavelet transform for texture recognition of SAR Images, IEEE trans. geosci. remote sens, 2002,50(11): 4358-one 4368. These transformation methods cannot obtain sufficient discrimination information and rely on a large amount of high-quality label information, and therefore, when used for classification of an SAR image, a good classification result cannot be obtained.
Model-based methods, including Markov field-based models and Bag-of-Words-based models. The method is mainly used for researching the spatial consistency of neighborhood pixels, in practical application, the method cannot effectively extract the texture information of the image, and is not suitable for classifying multi-resolution SAR images because the statistical relationship among different resolutions of the SAR images is not fully mined and the prior probability problem under the condition of only researching a single resolution is not researched; when the Bag-of-Words model proposed by Jie Feng et al is used for classification of SAR images, bottom layer features need to be extracted manually, and the process is complicated.
Disclosure of Invention
The invention aims to provide an SAR image classification method based on SAR image statistical distribution and DBN (direct binary decomposition) to overcome the defects of the prior art, so that the classification accuracy and the region consistency of SAR image classification are improved.
The technical scheme for realizing the aim of the invention comprises the following steps:
1) reading in an SAR image to be classified, forming a neighborhood matrix by one pixel in the image and pixels around the pixel, taking a neighborhood matrix for each pixel of the image, and forming an input matrix of a neural network by the neighborhood matrices;
2) designing a depth confidence network DBN consisting of 3 restricted Boltzmann machines, wherein the 1 st restricted Boltzmann machine is a Gamma restricted Boltzmann machine Gamma RBM, and the 2 nd and 3 rd restricted Boltzmann machines are Bernoulli-Bernoulli restricted Boltzmann machines BBRBM; each limited Boltzmann machine consists of a visible layer and a hidden layer, nodes in the layers are not connected, nodes between the layers are all connected, and a parameter set of each limited Boltzmann machine is { W, b, c }, wherein W is a weight matrix connecting nodes of the visible layer and nodes of the hidden layer, and b and c are offsets of the visible layer and the hidden layer respectively;
3) training the network designed in the step 2 to obtain a trained deep belief network DBN:
3a) the input matrix is used as the input of a deep confidence network DBN, the 1 st limited Boltzmann machine is pre-trained to obtain the output of the 1 st limited Boltzmann machine, and the weight matrix and the bias of the output are stored;
3b) the output of the 1 st limited Boltzmann machine is used as the input of the second limited Boltzmann machine of the deep confidence network DBN, the 2 nd limited Boltzmann machine is pre-trained to obtain the output of the 2 nd limited Boltzmann machine, and the weight matrix and the bias of the output are stored;
3c) setting the numerical value in the 3 rd limited Boltzmann machine weight matrix as a normally distributed random number with the mean value of 0 and the variance of 0.1, and setting the hidden layer bias c as a random number in [0,1 ];
4) according to the labeled category information in the ground object category reference map of the SAR image, randomly selecting a part of pixel sets with category labels for each category;
5) adopting a back propagation algorithm BP, and carrying out supervised micro-adjustment on the parameters of the trained deep belief network DBN by using a pixel set with a category label to obtain an adjusted deep belief network DBN;
6) classifying all pixels in the image to be classified one by using the adjusted depth confidence network DBN to obtain a classification result;
7) and marking the same color for the same category on the classified SAR image according to the three primary colors of red, green and blue to obtain a colored classification result image and outputting the colored classification result image.
Compared with the prior art, the invention has the following advantages:
1. the invention improves the application of the deep learning method to the SAR image, combines the statistical distribution characteristic of the SAR image with the classic deep learning model DBN, realizes the automatic extraction of the SAR image characteristic through multi-layer characteristic learning, and overcomes the difficulty of manually extracting the characteristic.
2. The invention can achieve better classification performance by classifying through a deep confidence network DBN formed by a novel limited Boltzmann machine after supervised fine adjustment.
Simulation results show that the method has the advantages of excellent classification effect, good regional consistency and complete edge information.
Drawings
FIG. 1 is a flow chart of an implementation of the present invention;
FIG. 2 is a SAR image to be classified containing three types of ground object backgrounds;
FIG. 3 is a real terrain labeling diagram of a three-class terrain background SAR image;
FIG. 4 is a graph of the results of a prior art simulation of the classification of FIG. 2;
FIG. 5 is a diagram of the classification simulation results of the present invention with respect to FIG. 2.
Detailed Description
Embodiments and effects of the present invention will be described in further detail below with reference to the accompanying drawings.
Referring to fig. 1, the implementation steps of the present invention are as follows:
firstly, preprocessing an image to be classified to obtain an input matrix of a Deep Belief Network (DBN).
1a) Reading in an SAR image to be classified as shown in fig. 2;
1b) let e be a pixel in the image, and form a3 × 3 neighborhood matrix Y with pixel e and its surrounding pixels a, b, c, d, f, g, h, i:
Figure GDA0001628675890000031
1c) converting the neighborhood matrix Y into a row vector X:
X＝[a d g b e h c f i]，
1d) repeating 1b and 1c on all pixel points in the SAR image to be classified to obtain a row vector x representing each pixel pointiN, n represents the number of pixels of the SAR image;
1e) the input matrix D of the deep belief network DBN is formed by the row vectors:
D＝[x1；x2...xi...xn]。
and step two, designing a deep confidence network DBN consisting of 3 limited Boltzmann machines.
The 1 st restricted Boltzmann machine is a Gamma restricted Boltzmann machine, a Gamma RBM, and the 2 nd and 3 rd restricted Boltzmann machines are Bernoulli-Bernoulli restricted Boltzmann machines, a BBRBM;
each limited Boltzmann machine consists of a visible layer and a hidden layer, nodes in the layers are not connected, nodes between the layers are all connected, and a parameter set of each limited Boltzmann machine is { W, b, c }, wherein W is a weight matrix connecting nodes of the visible layer and nodes of the hidden layer, and b and c are offsets of the visible layer and the hidden layer respectively;
stacking three restricted Boltzmann machines according to the following process to obtain a deep confidence network DBN:
according to the characteristics that the hidden layer output of the 1 st limited Boltzmann machine is the same as the visible layer input of the 2 nd limited Boltzmann machine, the hidden layer of the 1 st limited Boltzmann machine and the visible layer of the 2 nd limited Boltzmann machine are laminated into a second layer of the DBN;
by analogy, the hidden layer of the 2 nd limited boltzmann machine and the visible layer of the 3 rd limited boltzmann machine are combined into a third layer of the DBN, and the hidden layer of the 3 rd limited boltzmann machine is used as a fourth layer of the DBN, so that the depth confidence network DBN which is connected with each other layer and is provided with 1 visible layer and 3 hidden layers is formed.
And step three, training the deep confidence network DBN designed in the step 2 to obtain the trained deep confidence network DBN.
3a) Taking the input matrix D as the input of a deep confidence network DBN, pre-training a1 st limited Boltzmann machine to obtain hidden layer output of the 1 st limited Boltzmann machine, and storing a weight matrix and bias thereof:
(3a1) for a specific set of data (v, h), a new energy function is derived from the Gamma distribution of the SAR image and the energy formula of the restricted Boltzmann machine:
Figure GDA0001628675890000041
wherein v and h respectively represent a visible layer and a hidden layer of the restricted Boltzmann machine, L represents the view of the SAR image, and W represents the visual number of the SAR imageijThe ith row and the jth column elements of the weight matrix represent weight values connecting the ith visible layer node and the jth hidden layer node; n isvIs the number of nodes of the visual layer, nhIs the number of hidden layer nodes; v. ofiRepresents the state of the ith visible level node, hjRepresenting the state of the jth hidden layer node; biAnd cjRespectively visible layer node viAnd hidden layer node hjBias of (3);
(3a2) deriving an activation value formula of the jth hidden node from the new energy function as follows:
Figure GDA0001628675890000051
wherein the sigmoid function is an activation function in the neural network, which is defined as:
sigmoid(x)＝1/(1+e-x)
the activation value formula of the jth hidden layer node is expressed in the known visible layer state v(k)Then, sampling to obtain the probability that the j hidden layer node value is 1;
(3a3) calling CD-k algorithm to perform k times of Gibbs sampling, namely generating [0,1]Random number R of innerjWhen P (h)j＝1|v)＞RjWhen h is presentj1, otherwise hjObtaining the hidden layer state h of the restricted Boltzmann machine of the kth iteration after sampling all hidden layer nodes as 0(k)；
(3a4) After sampling to obtain a hidden layer state, reconstructing an activation value of an ith visible layer node:
Figure GDA0001628675890000052
wherein
Figure GDA0001628675890000053
Where Γ () represents a Gamma function, the activation value formula for the ith visible layer node is expressed in the known hidden state h(k)Sampling to obtain the probability that the ith visible layer node value is x, and sampling all visible layer nodes to obtain the visible layer state v(k+1)；
(3a5) Let the training sample set of the first constrained boltzmann machine be:
Figure GDA0001628675890000054
wherein n issTo train the number of samples, vtIs the t training sample:
Figure GDA0001628675890000055
(3a6) calculating the probability of the t training sample:
Figure GDA0001628675890000056
where θ is the parameter set initialized by the restricted boltzmann machine, i.e., θ ═ Wij,bi,cjZ is a distribution item of a restricted Boltzmann machine;
(3a7) for P (v)t| θ) to Wij、biAnd cjThe gradient of each parameter of the kth iteration is obtained:
Figure GDA0001628675890000061
Figure GDA0001628675890000062
Figure GDA0001628675890000063
wherein Δ wij kRepresenting the gradient of the weight matrix of the first restricted Boltzmann machine, Δ bi kGradient, Δ c, representing the visual layer bias of the first restricted Boltzmann machinej kA gradient representing a hidden layer bias of a first restricted boltzmann machine;
(3a8) obtaining a weight matrix W after the training of the first limited Boltzmann machine according to the gradient of each parameter of the kth iterationij kVisual layer bias bi kHidden layer bias cj k：
Wij k＝Wij+ηΔwij k
bi k＝bi+ηΔbi k
cj k＝cj+ηΔcj k
Wherein eta is the learning rate of each parameter;
3b) taking the hidden layer output of the 1 st limited Boltzmann machine as the visible layer input of the second limited Boltzmann machine of the deep confidence network DBN, pre-training the 2 nd limited Boltzmann machine to obtain the hidden layer output of the 2 nd limited Boltzmann machine, and storing the weight matrix and the bias thereof:
(3b1) for a specific set of data (v)(1),h(1)) The second limited boltzmann function in the network is:
Figure GDA0001628675890000064
wherein v is(1)And h(1)Respectively representing the visible and hidden layers of a confined Boltzmann machine, W(1)For a weight matrix connecting visible and hidden nodes, Wij (1)Is the ith row and the jth column element of the weight matrix; n isvIs the number of nodes of the visible layer, nhIs the number of hidden layer nodes; v. ofi (1)Represents the state of the ith visible level node, hj (1)Representing the state of the jth hidden layer node; bi (1)And cj (1)Respectively visible layer node vi (1)And hidden layer node hj (1)Bias of (c);
(3b2) based on the energy function of (3b1), obtaining an activation value formula of the j hidden layer node of the second limited Boltzmann machine:
Figure GDA0001628675890000071
the activation value formula of the hidden layer node is expressed in a known visible layer state v{0}Then, sampling to obtain the probability that the j hidden layer node value is 1;
(3b3) calling the CD-k algorithm to perform k times of Gibbs sampling, i.e. firstly generating [0,1]Random number of (b), when P (h)j (1)＝1|v(0))＞RjWhen h is presentj (1)1, otherwise hj (1)0; obtaining the hidden layer state h of a second limited Boltzmann machine after sampling all hidden layer nodes{0}Wherein the value of k is set to 1;
(3b4) after sampling to obtain a hidden layer state, reconstructing an activation value formula of an ith visible layer node of a second limited Boltzmann machine as follows:
Figure GDA0001628675890000072
the activation value formula of the visible layer node is expressed in a known hidden layer state h{0}Sampling to obtain the probability that the value of the ith visible layer node is 1, and carrying out the sampling on all visible layer nodesAfter point sampling, a visible layer state v is obtained{1}；
(3b5) Let the training sample set of the second constrained boltzmann machine be:
Figure GDA0001628675890000073
wherein n isMTo train the number of samples, vmIs the mth training sample:
Figure GDA0001628675890000074
Figure GDA0001628675890000075
(3b6) initializing parameter set theta in a second known limited Boltzmann machine(1)Under the condition (1), obtaining the probability of the mth training sample:
Figure GDA0001628675890000076
wherein, theta(1)＝{Wij (0),bi (0),cj (0)}，Z(1)A distribution term of a second limited Boltzmann machine;
(3b7) for P (v)m| θ) to Wij (1),bi (1),cj (1)The partial derivative of each parameter is obtained:
Figure GDA0001628675890000081
Figure GDA0001628675890000082
Figure GDA0001628675890000083
wherein Δ wij (0)Representing the gradient of the weight matrix of the second restricted Boltzmann machine, Δ bi (0)Gradient, Δ c, representing the visual layer bias of the second restricted Boltzmann machinej (0)A gradient representing a hidden layer bias of the second limited boltzmann machine;
(3b8) obtaining a weight matrix W after the training of a second limited Boltzmann machine according to the gradient obtained in the step (3a7)ij *Visual layer bias bi *And hidden layer bias cj *：
Wij *＝Wij (0)+λΔWij (0)
bi *＝bi (0)+λΔbi (0)
cj *＝cj (0)+λΔcj (0)
Wherein λ represents the learning rate of each parameter;
3c) initializing a third restricted boltzmann machine: and setting the numerical value in the 3 rd limited Boltzmann machine weight matrix as a normally distributed random number with the mean value of 0 and the variance of 0.1, and setting the hidden layer bias c as a random number in [0,1], thereby obtaining the trained deep confidence network DBN.
And step four, extracting the pixel set with the category label.
According to the labeled category information in the ground object category label map of the SAR image shown in the figure (3), 15% of pixel points with category labels are randomly selected from the pixel region of each category to form a pixel set with the category labels.
And step five, adopting a back propagation algorithm BP, and carrying out supervised micro-adjustment on the trained deep belief network DBN by using a pixel set with class labels to obtain an adjusted deep belief network DBN.
5a) And (3) forward propagation process: the input matrix D is propagated from the first limited Boltzmann machine to the third limited Boltzmann machine, namely the output of the hidden layer of the previous limited Boltzmann machine is used as the input of the visible layer of the next limited Boltzmann machine, and the actual output of the hidden layer of the third limited Boltzmann machine is obtained;
5b) and (3) backward propagation process: propagating errors of the actual output and the expected output of the step 5a) backwards layer by layer, namely propagating the errors forwards two limited Boltzmann machines layer by layer, and carrying out micro-adjustment on the weight and the bias of the trained deep belief network DBN to obtain a micro-adjusted deep belief network DBN;
the weight and the bias of the trained deep belief network DBN are subjected to micro adjustment according to the following steps:
5b1) calculating the sensitivity of the jth node of the hidden layer of each restricted Boltzmann machine:
δj＝οj(1-οj)(dj-οj),j＝1,2,...,nh，
wherein ojRepresenting the actual output of the j-th node, djRepresenting the expected output of the jth node, nhRepresenting the number of hidden layer nodes;
5b2) calculating the sensitivity delta of the hidden layer of the limited Boltzmann machine according to the sensitivity of the jth node of the hidden layeri l：
Figure GDA0001628675890000091
Wherein y isi lRepresenting the actual output of the implicit layer of the ith restricted boltzmann machine, wij lRepresenting the weight matrix, δ, of the ith restricted Boltzmann machinej l+1Representing the sensitivity of the jth hidden layer node of the (l + 1) th limited Boltzmann machine;
5b3) obtaining a weight matrix w of each restricted Boltzmann machine after micro-adjustment according to the sensitivity of the hidden layer of each restricted Boltzmann machineij lAnd bias cj l：
wij l＝wij l+ε×yi lδj l+1
cj l＝cj l+ε×δj l+1
Where epsilon represents the learning rate of the fine adjustment.
And step six, carrying out pixel-by-pixel classification on the image to be classified to obtain a classification result.
And (3) classifying all pixels in the SAR image to be classified one by using the depth confidence network DBN after supervision and fine adjustment, and outputting a class label of each pixel to obtain a classification result of the whole SAR image.
And step seven, obtaining a colored classification result graph according to the classification result of the step six.
And marking the same color for the pixel points of the same category on the whole classified SAR image according to the three primary colors of red, green and blue to obtain a colored classification result picture and outputting the colored classification result picture.
The effects of the present invention can be further illustrated by the following simulations:
1. simulation conditions
The simulation use method is a traditional DBN-SVM classification method and the method of the invention;
in simulation experiments, the method of the present invention and the comparative method were both implemented as programmed in MATLAB R2016b software.
2. Emulated content and results
Simulation 1, a conventional DBN-SVM classification method is used to perform a classification experiment on the SAR image shown in fig. 2, and the result is shown in fig. 4.
Simulation 2, the method of the invention is used to perform classification experiments on the SAR image shown in FIG. 2, and the result is shown in FIG. 5.
Comparing the classification result graphs shown in fig. 4 and 5 with the real ground object labeling graph shown in fig. 3, it can be seen from the visual effect that:
the classification result obtained by the traditional DBN-SVM classification method is influenced by coherent speckle noise, so that the phenomenon of wrong classification is serious, more pixels with wrong classification are obtained, the edge information of each region is incomplete, the region consistency is poor, and the classification accuracy is low;
in the classification result obtained by the method, the regional consistency of species in various regions is improved, the edge information is clear and complete, the number of wrongly classified pixels is small, and the classification accuracy is high.

Claims (3)

1. The SAR image classification method based on the SAR image statistical distribution characteristics and the DBN comprises the following steps:
1) reading in an SAR image to be classified, forming a neighborhood matrix by one pixel in the image and pixels around the pixel, taking a neighborhood matrix for each pixel of the image, and forming an input matrix of a neural network by the neighborhood matrices;
2) designing a depth confidence network DBN consisting of 3 restricted Boltzmann machines, wherein the 1 st restricted Boltzmann machine is a Gamma restricted Boltzmann machine Gamma RBM, and the 2 nd and 3 rd restricted Boltzmann machines are Bernoulli-Bernoulli restricted Boltzmann machines BBRBM; each limited Boltzmann machine consists of a visible layer and a hidden layer, nodes in the layers are not connected, nodes between the layers are all connected, and a parameter set of each limited Boltzmann machine is { W, b, c }, wherein W is a weight matrix connecting nodes of the visible layer and nodes of the hidden layer, and b and c are offsets of the visible layer and the hidden layer respectively;
3) training the network designed in the step 2 to obtain a trained deep belief network DBN:
3a) the input matrix is used as the input of a deep confidence network DBN, the 1 st limited Boltzmann machine is pre-trained to obtain the output of the 1 st limited Boltzmann machine, and the weight matrix and the bias of the output are stored; the implementation is as follows:
(3a1) for a specific set of data (v, h), a new energy function is derived from the Gamma distribution of the SAR image and the energy formula of the restricted Boltzmann machine:
Figure FDA0003513214460000011
wherein v and h represent the visible layer and the hidden layer of the restricted Boltzmann machine, respectively, L represents the view of the SAR image, and W represents the visual number of the SAR imageijIs the ith row and the jth column element of the weight matrix, and represents the weight connecting the ith visible layer node and the jth hidden layer nodeA value; n isvIs the number of nodes of the visual layer, nhIs the number of hidden layer nodes; v. ofiRepresents the state of the ith visible level node, hjRepresenting the state of the jth hidden layer node; biAnd cjRespectively visible layer node viAnd hidden layer node hjBias of (3);
(3a2) deriving the activation value of the jth hidden node from the new energy function as follows:
Figure FDA0003513214460000012
wherein the sigmoid function is an activation function in the neural network, which is defined as:
sigmoid(x)＝1/(1+e-x)
the activation value is formulated at a known visible layer state v(k)Then, sampling to obtain the probability that the j hidden layer node value is 1;
(3a3) calling CD-k algorithm to perform k times of Gibbs sampling, namely generating [0,1]Random number of (b), when P (h)j＝1|v)＞RjWhen h is presentj1, otherwise hjObtaining the hidden layer state h of the restricted Boltzmann machine of the kth iteration after sampling all hidden layer nodes as 0(k)；
(3a4) After sampling to obtain a hidden layer state, reconstructing an activation value of an ith visible layer node:
Figure FDA0003513214460000021
wherein
Figure FDA0003513214460000022
Where Γ () represents a Gamma function, the activation value formula represents at a known hidden state h(k)Sampling to obtain the probability that the ith visible layer node value is x, and sampling all visible layer nodes to obtain the visible layer state v(k+1)；
(3a5) Let the first restricted boltzmann machine training sample set as:
Figure FDA0003513214460000023
wherein n issTo train the number of samples, vtIs the t training sample:
Figure FDA0003513214460000024
(3a6) calculating the probability of the t training sample:
Figure FDA0003513214460000025
where θ is the parameter set initialized by the restricted boltzmann machine, i.e., θ ═ Wij,bi,cjZ is a distribution item of a restricted Boltzmann machine;
(3a7) for P (v)t| θ) to Wij、biAnd cjThe gradient of each parameter of the kth iteration is obtained:
Figure FDA0003513214460000026
Figure FDA0003513214460000027
Figure FDA0003513214460000028
wherein Δ wij kRepresenting the gradient of the weight matrix of the first restricted Boltzmann machine, Δ bi kGradient, Δ c, representing the visual layer bias of the first restricted Boltzmann machinej kHidden layer representing the first restricted boltzmann machineA gradient of bias;
(3a8) obtaining a trained parameter W according to the gradient of each parameter of the kth iterationij k+1，bi k+1，cj k+1：
Wij k+1＝Wij+ηΔwij k
bi k+1＝bi+ηΔbi k
cj k+1＝cj+ηΔcj k
Wherein eta is the learning rate of each parameter;
3b) the output of the 1 st limited Boltzmann machine is used as the input of the second limited Boltzmann machine of the deep confidence network DBN, the 2 nd limited Boltzmann machine is pre-trained to obtain the output of the 2 nd limited Boltzmann machine, and the weight matrix and the bias of the output are stored;
3c) setting the numerical value in the 3 rd limited Boltzmann machine weight matrix as a normally distributed random number with the mean value of 0 and the variance of 0.1, and setting the hidden layer bias c as a random number in [0,1 ];
4) according to the labeled category information in the ground object category reference map of the SAR image, randomly selecting a part of pixel sets with category labels for each category;
5) adopting a back propagation algorithm BP, and carrying out supervised micro-adjustment on the parameters of the trained deep belief network DBN by using a pixel set with a category label to obtain an adjusted deep belief network DBN;
6) classifying all pixels in the image to be classified one by using the adjusted depth confidence network DBN to obtain a classification result;
7) and marking the same color for the same category on the classified SAR image according to the three primary colors of red, green and blue to obtain a colored classification result image and outputting the colored classification result image.
2. The method of claim 1, wherein step 3b) pre-trains the 2 nd restricted boltzmann machine with the hidden layer output of the 1 st restricted boltzmann machine as an input to a second restricted boltzmann machine of the deep belief network DBN by:
(3b1) for a specific set of data (v)(1),h(1)) The second limited boltzmann function in the network is:
Figure FDA0003513214460000031
wherein v is(1)And h(1)Respectively representing the visible and hidden layers of a confined Boltzmann machine, W(1)For a weight matrix connecting visible and hidden nodes, Wij (1)Is the ith row and the jth column element of the weight matrix; n isvIs the number of nodes of the visual layer, nhIs the number of hidden layer nodes; v. ofi (1)Represents the state of the ith visible level node, hj (1)Representing the state of the jth hidden layer node; bi (1)And cj (1)Respectively visible layer node vi (1)And hidden layer node hj (1)Bias of (3);
(3b2) based on the energy function of (3b1), obtaining an activation value formula of the jth hidden node:
Figure FDA0003513214460000032
the activation value is formulated at a known visible layer state v{0}Then, sampling to obtain the probability that the j hidden layer node value is 1;
(3b3) calling the CD-k algorithm to perform k times of Gibbs sampling, i.e. firstly generating [0,1]Random number in P (h)j (1)＝1|v(0))＞RjWhen h is presentj (1)1, otherwise hj (1)0; obtaining the hidden layer state h of the restricted Boltzmann machine of the kth iteration after sampling all hidden layer nodes{0}Wherein the value of k is set to 1;
(3b4) after the hidden layer state is obtained through sampling, the activation value formula for reconstructing the ith visible layer node is as follows:
Figure FDA0003513214460000041
the activation value is formulated in a known hidden state h{0}Sampling to obtain the probability that the value of the ith visible layer node is 1, and sampling all visible layer nodes to obtain the visible layer state v{1}；
(3b5) Let the training sample set of the second constrained boltzmann machine be:
Figure FDA0003513214460000042
wherein n isMTo train the number of samples, vmIs the mth training sample:
Figure FDA0003513214460000043
(3b6) initializing parameter set theta in a second known limited Boltzmann machine(1)Under the condition (1), obtaining the probability of the mth training sample:
Figure FDA0003513214460000044
wherein, theta(1)＝{Wij (0),bi (0),cj (0)}，Z(1)A distribution term of a second limited Boltzmann machine;
(3b7) for P (v)m| θ) to Wij (1),bi (1),cj (1)The partial derivative of each parameter is obtained:
Figure FDA0003513214460000045
Figure FDA0003513214460000046
Figure FDA0003513214460000047
wherein Δ wij (0)Representing the gradient of the weight matrix of the second restricted Boltzmann machine, Δ bi (0)Gradient, Δ c, representing the visual layer bias of the second restricted Boltzmann machinej (0)A gradient representing a hidden layer bias of a second limited boltzmann machine;
(3b8) obtaining the parameter W after the training of the second limited Boltzmann machine according to the gradient obtained in (3a6)ij *，bi *And cj *：
Wij *＝Wij (0)+λΔWij (0)
bi *＝bi (0)+λΔbi (0)
cj *＝cj (0)+λΔcj (0)
Where λ represents the learning rate of each parameter.
3. The method of claim 1, wherein step 5) employs a back propagation algorithm BP, using a class-labeled set of pixels to perform supervised micro-tuning of parameters of a trained deep belief network DBN, which is implemented as follows:
5a) and (3) forward propagation process: the input matrix is transmitted from the first limited Boltzmann machine to the third limited Boltzmann machine, namely, the output of the hidden layer of the previous limited Boltzmann machine is used as the input of the visible layer of the next limited Boltzmann machine, and the actual output of the hidden layer of the third limited Boltzmann machine is obtained;
5b) and (3) backward propagation process: propagating errors of the actual output and the expected output of the step 5a) backwards layer by layer, namely propagating the errors forwards layer by two limited Boltzmann machines, and finely adjusting the weight and the bias of the trained deep belief network DBN to obtain the finely adjusted deep belief network DBN, wherein the implementation is as follows:
5b1) calculating the sensitivity of the jth node of the hidden layer of each restricted Boltzmann machine:
δj＝οj(1-οj)(dj-οj)，
wherein ojRepresenting the actual output of the j-th node, djRepresents the expected output of the jth node;
5b2) calculating the sensitivity of the hidden layer of the ith restricted Boltzmann machine according to the sensitivity of the jth node of the hidden layer:
Figure FDA0003513214460000051
wherein deltai lSensitivity, y, representing the hidden layer of the first restricted Boltzmann machinei lRepresenting the actual output of the hidden layer of the ith restricted Boltzmann machine, deltaj l+1The sensitivity of the (l + 1) limited Boltzmann machine hidden layer is represented;
5b3) obtaining a weight matrix w of each restricted Boltzmann machine after micro-adjustment according to the sensitivity of the hidden layer of each restricted Boltzmann machineij lAnd bias cj l：
wij l＝wij l+ε×yi lδj l+1
cj l＝cj l+ε×δj l+1
Where epsilon represents the learning rate of the fine adjustment.
CN201810031937.3A 2018-01-12 2018-01-12 SAR image classification method based on SAR image statistical distribution and DBN Active CN108171200B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201810031937.3A CN108171200B (en) 2018-01-12 2018-01-12 SAR image classification method based on SAR image statistical distribution and DBN

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201810031937.3A CN108171200B (en) 2018-01-12 2018-01-12 SAR image classification method based on SAR image statistical distribution and DBN

Publications (2)

Publication Number Publication Date
CN108171200A CN108171200A (en) 2018-06-15
CN108171200B true CN108171200B (en) 2022-05-17

Family

ID=62514696

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201810031937.3A Active CN108171200B (en) 2018-01-12 2018-01-12 SAR image classification method based on SAR image statistical distribution and DBN

Country Status (1)

Country Link
CN (1) CN108171200B (en)

Families Citing this family (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109596942A (en) * 2018-12-06 2019-04-09 华北电力大学 A kind of voltage sag reason recognition methods based on depth confidence network
CN109507655B (en) * 2018-12-11 2022-09-02 西北工业大学 SAR target recognition algorithm based on guiding reconstruction and norm constraint DBN
CN113344046A (en) * 2021-05-24 2021-09-03 电子科技大学 Method for improving SAR image ship classification precision

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104318245A (en) * 2014-10-20 2015-01-28 西安电子科技大学 Sparse depth network based polarization SAR (Synthetic Aperture Radar) image classification
CN105701503A (en) * 2016-01-07 2016-06-22 西安电子科技大学 Interferometric synthetic aperture radar (SAR) image classification method based on deep belief network (DBN) model
CN107132516A (en) * 2017-05-16 2017-09-05 电子科技大学 A kind of Radar range profile's target identification method based on depth confidence network
CN107407682A (en) * 2015-04-10 2017-11-28 社会福祉法人恩赐财团济生会 Differentiate the ill method of liver disease
CN107451565A (en) * 2017-08-01 2017-12-08 重庆大学 A kind of semi-supervised small sample deep learning image model classifying identification method

Family Cites Families (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2012030872A1 (en) * 2010-09-02 2012-03-08 Edge3 Technologies Inc. Method and apparatus for confusion learning
US10650508B2 (en) * 2014-12-03 2020-05-12 Kla-Tencor Corporation Automatic defect classification without sampling and feature selection

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104318245A (en) * 2014-10-20 2015-01-28 西安电子科技大学 Sparse depth network based polarization SAR (Synthetic Aperture Radar) image classification
CN107407682A (en) * 2015-04-10 2017-11-28 社会福祉法人恩赐财团济生会 Differentiate the ill method of liver disease
CN105701503A (en) * 2016-01-07 2016-06-22 西安电子科技大学 Interferometric synthetic aperture radar (SAR) image classification method based on deep belief network (DBN) model
CN107132516A (en) * 2017-05-16 2017-09-05 电子科技大学 A kind of Radar range profile's target identification method based on depth confidence network
CN107451565A (en) * 2017-08-01 2017-12-08 重庆大学 A kind of semi-supervised small sample deep learning image model classifying identification method

Non-Patent Citations (5)

* Cited by examiner, † Cited by third party
Title
The Poisson Gamma Belief Network;Mingyuan Zhou等;《NIPS 2015》;20151231;1-9 *
Wishart RBM based DBN for polarimetric synthetic radar data classification;Yanhe Guo等;《2015 IEEE International Geoscience and Remote Sensing Symposium (IGARSS)》;20151112;1841-1844 *
一种改进的深度置信网络及其在自然图像分类中的应用;柴瑞敏等;《计算机应用与软件》;20160915;第33卷(第09期);正文第1节、图2 *
基于深度学习神经网络的SAR图像目标识别算法;梁鑫等;《江汉大学学报（自然科学版）》;20160708;第44卷(第02期);第131-136页 *
高分辨SAR图像中杂波的统计特性分析;高贵等;《信号处理》;20080815(第04期);正文第2节 *

Also Published As

Publication number Publication date
CN108171200A (en) 2018-06-15

Similar Documents

Publication Publication Date Title
CN111368896B (en) Hyperspectral remote sensing image classification method based on dense residual three-dimensional convolutional neural network
Zhang et al. Hyperspectral unmixing via deep convolutional neural networks
CN110443143B (en) Multi-branch convolutional neural network fused remote sensing image scene classification method
Solberg Contextual data fusion applied to forest map revision
Palmason et al. Classification of hyperspectral data from urban areas using morphological preprocessing and independent component analysis
CN108171200B (en) SAR image classification method based on SAR image statistical distribution and DBN
CN107145830A (en) Hyperspectral image classification method with depth belief network is strengthened based on spatial information
Tian et al. Temporal updating scheme for probabilistic neural network with application to satellite cloud classification
CN105787510A (en) System and method for realizing subway scene classification based on deep learning
CN112347970B (en) Remote sensing image ground object identification method based on graph convolution neural network
CN110458192B (en) Hyperspectral remote sensing image classification method and system based on visual saliency
Lv et al. Deep learning model of image classification using machine learning
CN107491793B (en) Polarized SAR image classification method based on sparse scattering complete convolution
CN112434745A (en) Occlusion target detection and identification method based on multi-source cognitive fusion
Isobe et al. Deep convolutional encoder-decoder network with model uncertainty for semantic segmentation
CN108256557B (en) Hyperspectral image classification method combining deep learning and neighborhood integration
CN112749621A (en) Remote sensing image cloud layer detection method based on deep convolutional neural network
CN111639587A (en) Hyperspectral image classification method based on multi-scale spectrum space convolution neural network
CN114048810A (en) Hyperspectral image classification method based on multilevel feature extraction network
CN108460326B (en) Hyperspectral image semi-supervised classification method based on sparse expression graph
CN110807485A (en) Method for fusing two-classification semantic segmentation maps into multi-classification semantic map based on high-resolution remote sensing image
CN107464247B (en) Based on G0Distributed random gradient variational Bayesian SAR image segmentation method
Kouzani Road-sign identification using ensemble learning
CN103365985A (en) Class-adaptive polarimetric SAR (synthetic aperture radar) classification method
CN111914922B (en) Hyperspectral image classification method based on local convolution and cavity convolution

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