CN108171200A - SAR image sorting technique based on SAR image statistical distribution and DBN - Google Patents
SAR image sorting technique based on SAR image statistical distribution and DBN Download PDFInfo
- Publication number
- CN108171200A CN108171200A CN201810031937.3A CN201810031937A CN108171200A CN 108171200 A CN108171200 A CN 108171200A CN 201810031937 A CN201810031937 A CN 201810031937A CN 108171200 A CN108171200 A CN 108171200A
- Authority
- CN
- China
- Prior art keywords
- boltzmann machine
- node
- limited boltzmann
- layer
- hidden
- 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.)
- Granted
Links
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 a kind of SAR image sorting techniques based on SAR image statistical distribution and DBN, mainly solve traditional depth confidence network DBN and classify for SAR image, are also easy to produce region consistency difference and the incomplete problem of marginal information.Its realization process is：SAR image to be sorted is pre-processed, obtains the input matrix of DBN；Design the DBN being made of 3 limited Boltzmann machines；The network designed with input matrix pre-training, obtains trained DBN；From the label of SAR image category label figure, set of pixels of the part with class label is randomly selected, and trained DBN is micro-adjusted using back-propagation algorithm；Using DBN after micro-adjustment, classification pixel-by-pixel is carried out to image to be classified, obtains classification results and output of painting.The present invention has the advantages that classification results are excellent, region consistency is good, marginal information is complete, can be applied to the terrain classification and target identification of SAR image.
Description
Technical field
The invention belongs to image processing field, specifically a kind of SAR image sorting technique can be applied to SAR images
Terrain classification and target identification.
Background technology
Synthetic aperture radar SAR is a kind of airborne radar or spaceborne radar that can generate high-definition picture, is had complete
It when round-the-clock feature, by extensively should be in the fields such as remote sensing and ground mapping.Diameter radar image, i.e. SAR image are
The radar signal that one width irradiates ground has the landforms image of light and dark color difference, extraordinary image black-and-white photograph through earth surface reflection.It is right
In the interpretation of SAR image, compared to the more of difficulty are wanted for optical imagery, the target to be realized of SAR image classification is to taking
SAR image, different ground species are labeled using different colors by designed sorting algorithm.
The statistical model of SAR image detection and the identification of ground object target, the elimination of coherent speckle noise, ground object target point
The application aspects such as class play very important effect.The statistical model of SAR image is roughly divided into according to the modeling process of SAR image
Two classes：Parametric model and nonparametric model.Parametric model refers to the parameter in unknown probability density function, by having
SAR image data, calculate the parameters of probability density function, the probability density function for then selecting fitting degree best
As the SAR image the statistical model of object area.For example, gamma distribution, Weibull distributions, Log-normal distributions etc..It is non-
Parametric model refers to directly select best probability density letter according to certain rule according to the data of the ground object area of SAR image
Number.
The feature extracting method of SAR image classification can be roughly divided into three types：Statistical method, transform domain method and
Method based on model.Wherein：
Statistical method, typical represent is the gray level co-occurrence matrixes feature proposed by U.Kandaswamy et al., referring to
U.Kandaswamy,D.A.Adjeroh,and M.C.Lee,Efficient texture analysis of SAR
imagery,IEEE Trans.Geosci.Remote Sens.,2005,43(9):2075-2083, this feature includes variance, right
Than degree, energy etc., since these information are only that SAR image is based on statistical feature, the spatial neighborhood for having ignored SAR image is special
Property, thus the classification results Space Consistency obtained with statistical method is poor.
Transform domain method including Fourier transformation, wavelet transformation and Garbor wavelet transformations, is mainly used for scheming SAR
As carrying out texture analysis, referring to G.Akbarizadeh, A new statistical-based kurtosis wavelet
energy feature for texture recognition of SAR Images,IEEE Trans.Geosci.Remote
Sens., 2002,50(11):4358–4368.These transform methods are due to that cannot obtain enough discriminant informations, dependent on big
When measuring the label information of high quality, thus being used for the classification of SAR image, it is impossible to obtain preferable classification results.
Method based on model, including the model based on Markov field and the model based on Bag-of-Words.Wherein
Based on the model of markov random file MRF, it is mainly used for the research to neighborhood territory pixel Space Consistency, in practical application
In, this method cannot effectively extract the texture information of image, and due to it is no fully excavate SAR image different resolution it
Between statistical relationship, only study the prior probability problem under single definition case, thus, be not suitable for multiresolution SAR figure
The classification of picture；And the Bag-of-Words models proposed based on Jie Feng et al., when for the classification of SAR image, then need
Low-level image feature is manually extracted, process is cumbersome.
Invention content
It is an object of the invention to be directed to the deficiency of above-mentioned prior art, propose it is a kind of based on SAR image statistical distribution and
The SAR image sorting technique of DBN, to improve the nicety of grading and region consistency of SAR image classification.
Realize the technical solution of the object of the invention including as follows：
1) SAR image to be sorted is read in, neighborhood is formed by a pixel in the image and surrounding pixel
Matrix takes a Neighborhood matrix to each pixel of image, with the input matrix of these Neighborhood matrixes composition neural network；
2) design forms depth confidence network DBN by 3 limited Boltzmann machines, wherein the 1st limited Boltzmann machine
It is that gamma is limited Boltzmann machine Gamma RBM, the 2nd and the 3rd limited Boltzmann machine are that Bernoulli Jacob-Bernoulli Jacob is limited glass
The graceful machine BBRBM of Wurz；Each limited Boltzmann machine is made of visual layers and hidden layer, and layer interior knot is connectionless, interlayer node
Full connection, the parameter set of each limited Boltzmann machine is { W, b, c }, and wherein W is the visible node layer of connection and hidden node
Weight matrix, b, c are respectively the biasing of visible layer and hidden layer；
3) network designed step 2 is trained, and obtains trained depth confidence network DBN：
3a) using input matrix as the input of depth confidence network DBN, the 1st limited Boltzmann machine is instructed in advance
Practice, obtain the output of the 1st limited Boltzmann machine, and preserve its weight matrix and biasing；
3b) using the output of the 1st limited Boltzmann machine as second limited Boltzmann of depth confidence network DBN
The input of machine, to the 2nd, limited Boltzmann machine carries out pre-training, obtains the output of the 2nd limited Boltzmann machine, and preserve
Its weight matrix and biasing；
3c) by the numerical value in the 3rd limited Boltzmann machine weight matrix be set as obey mean value be 0, variance 0.1
Normal distribution random number, hidden layer biasing c is set as random number in [0,1]；
4) according to the classification information marked in SAR image atural object classification reference chart, one is randomly selected to each classification
Divide the set of pixels with class label；
5) using back-propagation algorithm BP, using the set of pixels with class label to trained depth confidence network DBN
Parameter carry out the micro-adjustment for having supervision, the depth confidence network DBN after being adjusted；
6) using the depth confidence network DBN after adjustment, classified one by one to all pixels in image to be classified, obtained
Go out classification results；
7) in the SAR image for completing classification, according to red-green-blue, identical color is marked to identical classification,
Classification results figure after being painted simultaneously exports.
Compared with the prior art, the present invention has the following advantages：
1. the present invention is using upper improvement to deep learning method in SAR image, by the statistical distribution characteristic of SAR image
It is combined with classical deep learning model DBN, by the feature learning of multilayer, realizes automatically extracting for SAR image feature, gram
The difficulty of artificial extraction feature is taken.
2. the depth confidence network DBN that the present invention is made up of novel limited Boltzmann machine, by there is the fine tuning of supervision
Classify after whole, better classification performance can be reached.
Simulation result shows that classifying quality of the present invention is excellent, and region consistency is good, and marginal information is complete.
Description of the drawings
Fig. 1 is the realization flow chart of the present invention；
The SAR image for including three classes surface feature background to be sorted Fig. 2；
Fig. 3 is the truly substance markers figure of three classes surface feature background SAR image；
Fig. 4 is the classification simulation result figure to Fig. 2 with existing method；
Fig. 5 is classification simulation result figure of the present invention to Fig. 2.
Specific embodiment
Embodiment of the present invention and effect are described in further detail below in conjunction with attached drawing.
With reference to Fig. 1, implementation steps of the invention are as follows：
Step 1: being pre-processed to image to be classified, the input matrix of depth confidence network DBN is obtained.
1a) read in width SAR image to be sorted as shown in Figure 2；
E 1b) is set as a pixel in the image, is formed with pixel e and surrounding pixel a, b, c, d, f, g, h, i
The Neighborhood matrix Y of one 3*3：
Neighborhood matrix Y 1c) is converted into a row vector X：
X=[a d g b e h c f i],
1b and 1c 1d) are repeated to all pixels point in SAR image to be sorted, obtain representing the row of each pixel
Vector xi, i=1,2 ..., n, n represent SAR image number of pixels；
1e) with the input matrix D of these row vectors composition depth confidence network DBN：
D=[x1；x2...xi...xn]。
Step 2: the depth confidence network DBN that design is made of 3 limited Boltzmann machines.
1st limited Boltzmann machine is that gamma is limited Boltzmann machine Gamma RBM, the 2nd and the 3rd limited Bohr
Hereby graceful machine is that Bernoulli Jacob-Bernoulli Jacob is limited Boltzmann machine BBRBM；
Each limited Boltzmann machine is made of visual layers and hidden layer, and layer interior knot is connectionless, and interlayer node connects entirely,
The parameter set of each limited Boltzmann machine is { W, b, c }, and wherein W is the weights square of the visible node layer of connection and hidden node
Battle array, b, c are respectively the biasing of visible layer and hidden layer；
Three limited Boltzmann machines are stacked as follows to obtain depth confidence network DBN：
Visual layers the input phase with the 2nd limited Boltzmann machine is exported according to the hidden layer of the 1st limited Boltzmann machine
With feature, the visual layers of the hidden layer of the 1st limited Boltzmann machine and the 2nd limited Boltzmann machine are synthesized DBN's
The second layer；
And so on, the hidden layer of the 2nd limited Boltzmann machine and the visual layers of the 3rd limited Boltzmann machine are synthesized
For the third layer of DBN, using the hidden layer of the 3rd limited Boltzmann machine as the 4th layer of DBN, thus constitute layer and layer it
Between be connected with each other, the depth confidence network DBN with 1 visual layers and 3 hidden layers.
Step 3: the depth confidence network DBN designed step 2 is trained, trained depth confidence network is obtained
DBN。
3a) using input matrix D as the input of depth confidence network DBN, the 1st limited Boltzmann machine is instructed in advance
Practice, obtain the hidden layer output of the 1st limited Boltzmann machine, and preserve its weight matrix and biasing：
(3a1) is distributed according to the Gamma of SAR image for one group of specific data (v, h) and is limited Boltzmann machine
Energy theorem derives new energy function：
Wherein, v and h represents the visible layer and hidden layer of limited Boltzmann machine respectively, and L represents regarding for SAR image and counts, WijIt is
I-th row jth column element of weight matrix represents the weighted value of i-th of visible layer node of connection and j-th of hidden layer node；nvFor
The number of visual node layer, nhIt is the number of hidden node；viRepresent the state of i-th of visible layer node, hjRepresent j-th it is hidden
The state of layer node；biAnd cjRespectively visible node layer viWith hidden node hjBiasing；
The activation value formula that (3a2) is derived from j-th of hidden node by new energy function is：
Wherein sigmoid functions are the activation primitives in neural network, are defined as：
Sigmoid (x)=1/ (1+e-x)
The activation value formula of j-th of hidden layer node is represented in known visible layer state v(k)When, sampling obtain j-th it is hidden
Node layer value is 1 probability；
(3a3) calls CD-k algorithms to carry out k Gibbs sampling, that is, generates the random number R in [0,1]j, as P (hj=1 |
V) ＞ RjWhen, hj=1, otherwise hj=0, after being sampled to whole hidden nodes, obtain the limited Boltzmann of kth time iteration
The hidden layer state h of machine(k)；
(3a4) reconstructs the activation value of i-th of visible node layer after sampling obtains hidden layer state：
Wherein
Wherein Γ () represents Gamma functions, and the activation value formula of this i-th visible node layer is represented in known hidden layer state
h(k)When, sampling obtains the probability that i-th of visible layer nodal value is x, after to whole visible layer node samples, obtains visible
Layer state v(k+1)；
(3a5) set the training sample set of first limited Boltzmann machine as：
Wherein, nsFor the number of training sample, vtIt is t-th of training sample:T=1,2 ...,
ns；
(3a6) calculates the probability of t-th of training sample:
Wherein, θ is to be limited the parameter set that Boltzmann machine initializes, i.e. θ={ Wij,bi,cj, Z is limited Boltzmann
Machine matches subitem；
(3a7) is to P (vt| θ) it asks about Wij、biAnd cjPartial derivative, obtain the gradient of kth time iteration parameters：
Wherein Δ wij kRepresent the gradient of the weight matrix of first limited Boltzmann machine, Δ bi kRepresent that first is limited
The gradient of the visual layers biasing of Boltzmann machine, Δ cj kRepresent the gradient of the hidden layer biasing of first limited Boltzmann machine；
(3a8) obtains the weight after first limited Boltzmann machine training according to the gradient of kth time iteration parameters
Matrix Wij k, visual layers biasing bi k, hidden layer biasing cj k：
Wij k=Wij+ηΔwij k
bi k=bi+ηΔbi k
cj k=cj+ηΔcj k
Wherein η is the learning rate of parameters；
3b) using the hidden layer output of the 1st limited Boltzmann machine as second limited Bohr of depth confidence network DBN
The hereby visual layers input of graceful machine, to the 2nd, limited Boltzmann machine carries out pre-training, obtains the 2nd limited Boltzmann machine
Hidden layer exports, and preserves its weight matrix and biasing：
(3b1) is for one group of specific data (v(1),h(1)), second limited Boltzmann machine energy function in network
For：
Wherein v(1)And h(1)The visible layer and hidden layer of limited Boltzmann machine, W are represented respectively(1)To connect visible node layer
With the weight matrix of hidden node, Wij (1)It is the i-th row jth column element of weight matrix；nvFor the number of visual node layer, nhIt is
The number of hidden node；vi (1)Represent the state of i-th of visible layer node, hj (1)Represent the state of j-th of hidden layer node；bi (1)
And cj (1)Respectively visible node layer vi (1)With hidden node hj (1)Biasing；
The energy function of (3b2) based on (3b1), obtain second limited Boltzmann machine j-th of hidden node swash
Value formula living：
The activation value formula of the hidden layer node is represented in known visible layer state v{0}When, sampling obtains j-th of hidden node
It is worth the probability for 1；
(3b3) calls CD-k algorithms to carry out k Gibbs sampling, i.e., the random number in [0,1] is generated first, as P (hj (1)
=1 | v(0)) ＞ RjWhen, hj (1)=1, otherwise hj (1)=0；After being sampled to whole hidden nodes, second limited Bohr is obtained
The hereby hidden layer state h of graceful machine{0}, wherein k values are set as 1；
(3b4) reconstructs i-th of visible node layer of second limited Boltzmann machine after sampling obtains hidden layer state
Activation value formula is：
The activation value formula of the visible node layer is represented in known hidden layer state h{0}When, sampling obtains i-th of visible layer section
Point value is 1 probability, after to whole visible layer node samples, obtains visible layer state v{1}；
(3b5) set the training sample set of second limited Boltzmann machine as：
Wherein, nMFor the number of training sample, vmIt is m-th of training sample:M=1,2 ...,
nM；
(3b6) is in known second limited Boltzmann machine initiation parameter collection θ(1)Under conditions of, obtain m-th of training
The probability of sample:
Wherein, θ(1)={ Wij (0),bi (0),cj (0), Z(1)It itemizes for second matching for limited Boltzmann machine；
(3b7) is to P (vm| θ) it asks about Wij (1),bi (1),cj (1)Partial derivative, obtain the gradient of parameters：
Wherein Δ wij (0)Represent the gradient of the weight matrix of second limited Boltzmann machine, Δ bi (0)Represent second by
Limit the gradient of the visual layers biasing of Boltzmann machine, Δ cj (0)Represent the ladder of the hidden layer biasing of second limited Boltzmann machine
Degree；
The gradient that (3b8) is obtained according to (3a7) obtains the weight matrix W after second limited Boltzmann machine trainingij *,
Visual layers bias bi *C is biased with hidden layerj *：
bi *=bi (0)+λΔbi (0)
cj *=cj (0)+λΔcj (0)
Wherein λ represents the learning rate of parameters；
Initialize installation 3c) is carried out to the limited Boltzmann machine of third：By the 3rd limited Boltzmann machine weight matrix
In numerical value be set as obeying mean value be 0, the normal distribution random number that variance is 0.1, hidden layer biasing c is set as in [0,1]
Random number, obtained trained depth confidence network DBN since then.
Step 4: set of pixels of the extraction with class label.
The classification information marked in shown SAR image atural object category label figure according to fig. 3, from the pixel region of each classification
15% pixel with class label is randomly selected in domain, forms the set of pixels with category.
Step 5: using back-propagation algorithm BP, using the set of pixels with category to trained depth confidence network
DBN carries out the micro-adjustment for having supervision, the depth confidence network DBN after being adjusted.
5a) propagated forward process：Input matrix D travels to third from first limited Boltzmann machine, i.e., with previous
Input of the output of a limited Boltzmann machine hidden layer as next limited Boltzmann machine visual layers, it is limited to obtain third
The reality output of Boltzmann machine hidden layer；
5b) back-propagating process：By by 5a) reality output and anticipated output error successively back-propagation, i.e. handle
Error is successively limited Boltzmann machine to the first two and propagates, and the weights and biasing to trained depth confidence network DBN carry out
Micro-adjustment obtains the depth confidence network DBN after micro-adjustment；
It is described that the weights of trained depth confidence network DBN and biasing are micro-adjusted, it carries out as follows：
5b1) calculate the sensitivity of each limited Boltzmann machine j-th of node of hidden layer：
δj=οj(1-οj)(dj-οj), j=1,2 ..., nh,
Wherein, οjRepresent the reality output of j-th of node, djRepresent the anticipated output of j-th of node, nhRepresent hidden layer knot
The number of point；
5b2) according to the sensitivity of j-th of node of hidden layer, the sensitivity δ of l-th of limited Boltzmann machine hidden layer is calculatedi l：
Wherein yi lRepresent the reality output of l-th of limited Boltzmann machine hidden layer, wij lRepresent l-th of limited Boltzmann
The weight matrix of machine, δj l+1Represent the sensitivity of j-th of hidden layer node of l+1 limited Boltzmann machines；
5b3) the sensitivity of each limited Boltzmann machine hidden layer of basis, after obtaining each limited Boltzmann machine micro-adjustment
Weight matrix wij lWith biasing cj l：
wij l=wij l+ε×yi lδj l+1
cj l=cj l+ε×δj l+1
Wherein ε represents the learning rate of micro-adjustment.
Step 6: carrying out classification pixel-by-pixel to image to be classified, classification results are obtained.
Using the depth confidence network DBN having after supervising micro-adjustment, to the institute in SAR image to be sorted as shown in Figure 2
There is pixel to be classified one by one, export the class label of each pixel, obtain the classification results of whole SAR image.
Step 7: according to the classification results of step 6, the classification results figure after being painted.
It is identical to the pixel mark of the same category according to red-green-blue in whole SAR image for completing classification
Color, classification results figure after being painted simultaneously exports.
The effect of the present invention can be further illustrated by following emulation：
1st, simulated conditions
It is tradition DBN-SVM sorting techniques and the method for the present invention to emulate application method；
In emulation experiment, the method for the present invention and control methods are all to program to realize in MATLAB R2016b softwares.
2nd, emulation content and result
Emulation 1, classification experiments are carried out with traditional DBN-SVM sorting techniques to SAR image shown in Fig. 2, as a result such as Fig. 4
It is shown.
Emulation 2, carries out classification experiments, the results are shown in Figure 5 with the method for the present invention to SAR image shown in Fig. 2.
The truly substance markers figure of Fig. 4, classification results figure shown in fig. 5 and Fig. 3 are compared, it can from visual effect
To find out：
The classification results obtained using traditional DBN-SVM sorting techniques due to being influenced by coherent speckle noise, are accidentally divided
The phenomenon that than more serious, the pixel of classification error is more, and the marginal information of each region is imperfect, and region consistency is poor,
Classification accuracy is low；
In the classification results obtained using the method for the present invention, the region consistency of various regions species is improved, edge letter
Complete display is ceased, the pixel accidentally divided is less, and classification accuracy is high.
Claims (4)
1. the SAR image sorting technique based on SAR image statistical distribution characteristic and DBN, including：
1) SAR image to be sorted is read in, Neighborhood matrix is formed by a pixel in the image and surrounding pixel,
One Neighborhood matrix is taken to each pixel of image, with the input matrix of these Neighborhood matrixes composition neural network；
2) design forms depth confidence network DBN by 3 limited Boltzmann machines, wherein the 1st limited Boltzmann machine is gal
Horse is limited Boltzmann machine Gamma RBM, and the 2nd and the 3rd limited Boltzmann machine are that Bernoulli Jacob-Bernoulli Jacob is limited Bohr hereby
Graceful machine BBRBM；Each limited Boltzmann machine is made of visual layers and hidden layer, and layer interior knot is connectionless, and interlayer node connects entirely
It connects, the parameter set of each limited Boltzmann machine is { W, b, c }, and wherein W is the weights of the visible node layer of connection and hidden node
Matrix, b, c are respectively the biasing of visible layer and hidden layer；
3) network designed step 2 is trained, and obtains trained depth confidence network DBN：
3a) using input matrix as the input of depth confidence network DBN, to the 1st, limited Boltzmann machine carries out pre-training, obtains
To the output of the 1st limited Boltzmann machine, and preserve its weight matrix and biasing；
3b) using the output of the 1st limited Boltzmann machine as second limited Boltzmann machine of depth confidence network DBN
Input, to the 2nd, limited Boltzmann machine carries out pre-training, obtains the output of the 2nd limited Boltzmann machine, and preserve its power
Value matrix and biasing；
3c) by the numerical value in the 3rd limited Boltzmann machine weight matrix be set as obey mean value be 0, the normal state that variance is 0.1
Distribution random numbers, random number hidden layer biasing c being set as in [0,1]；
4) according to the classification information marked in SAR image atural object classification reference chart, a part of band is randomly selected to each classification
The set of pixels of class label；
5) using back-propagation algorithm BP, using the set of pixels with class label to the ginseng of trained depth confidence network DBN
Number carries out the micro-adjustment for having supervision, the depth confidence network DBN after being adjusted；
6) using the depth confidence network DBN after adjustment, classified one by one to all pixels in image to be classified, obtained point
Class result；
7) in the SAR image for completing classification, according to red-green-blue, identical color is marked to identical classification, is obtained
Classification results figure after colouring simultaneously exports.
2. according to the method described in claim 1, wherein step 3a) in using input matrix as the defeated of depth confidence network DBN
Enter, to the 1st, limited Boltzmann machine carries out pre-training, realizes as follows：
(3a1) according to the Gamma of SAR image is distributed and is limited the energy of Boltzmann machine for one group of specific data (v, h)
Formula derives new energy function：
Wherein v and h represents the visible layer and hidden layer of limited Boltzmann machine respectively, and L represents regarding for SAR image and counts, WijIt is weights
I-th row jth column element of matrix represents the weighted value of i-th of visible layer node of connection and j-th of hidden layer node；nvIt is visual
The number of node layer, nhIt is the number of hidden node；viRepresent the state of i-th of visible layer node, hjRepresent j-th of hidden layer knot
The state of point；biAnd cjRespectively visible node layer viWith hidden node hjBiasing；
The activation value that (3a2) is derived from j-th of hidden node by new energy function is：
Wherein sigmoid functions are the activation primitives in neural network, are defined as：
Sigmoid (x)=1/ (1+e-x)
The activation value formula is represented in known visible layer state v(k)When, sampling obtains the probability that j-th of hidden node value is 1；
(3a3) calls CD-k algorithms to carry out k Gibbs sampling, that is, the random number in [0,1] is generated, as P (hj=1 | v) ＞ Rj
When, hj=1, otherwise hj=0, after being sampled to whole hidden nodes, obtain kth time iteration limited Boltzmann machine it is hidden
Layer state h(k)；
(3a4) reconstructs the activation value of i-th of visible node layer after sampling obtains hidden layer state：
Wherein
Wherein Γ () represents Gamma functions, which represents in known hidden layer state h(k)When, sampling obtains i-th can
See the probability that node layer value is x, after to whole visible layer node samples, obtain visible layer state v(k+1)；
(3a5) set first limited Boltzmann machine training sample set as：
Wherein, nsFor the number of training sample, vtIt is t-th of training sample:
(3a6) calculates the probability of t-th of training sample:
Wherein θ is to be limited the parameter set that Boltzmann machine initializes, i.e. θ={ Wij,bi,cj, Z is matching for limited Boltzmann machine
Subitem；
(3a7) is to P (vt| θ) it asks about Wij、biAnd cjPartial derivative, obtain the gradient of kth time iteration parameters：
Wherein Δ wij kRepresent the gradient of the weight matrix of first limited Boltzmann machine, Δ bi kRepresent first limited Bohr
The hereby gradient of the visual layers biasing of graceful machine, Δ cj kRepresent the gradient of the hidden layer biasing of first limited Boltzmann machine；
(3a8) according to the gradient of kth time iteration parameters, the parameter W after being trainedij 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 η is the learning rate of parameters.
3. according to the method described in claim 1, wherein step 3b) using the hidden layer output of the 1st limited Boltzmann machine as
The input of second limited Boltzmann machine of depth confidence network DBN, to the 2nd, limited Boltzmann machine carries out pre-training,
Its realization is as follows：
(3b1) is for one group of specific data (v(1),h(1)), second limited Boltzmann machine energy function in network is：
Wherein v(1)And h(1)The visible layer and hidden layer of limited Boltzmann machine, W are represented respectively(1)To connect visible node layer and hidden
The weight matrix of node layer, Wij (1)It is the i-th row jth column element of weight matrix；nvFor the number of visual node layer, nhIt is hidden layer
The number of node；vi (1)Represent the state of i-th of visible layer node, hj (1)Represent the state of j-th of hidden layer node；bi (1)And cj (1)Respectively visible node layer vi (1)With hidden node hj (1)Biasing；
The energy function of (3b2) based on (3b1) obtains the activation value formula of j-th of hidden node：
The activation value formula is represented in known visible layer state v{0}When, sampling obtains the probability that j-th of hidden node value is 1；
(3b3) calls CD-k algorithms to carry out k Gibbs sampling, i.e., the random number in [0,1] is generated first, as P (hj (1)=1 | v(0)) ＞ RjWhen, hj (1)=1, otherwise hj (1)=0；After being sampled to whole hidden nodes, limited Bohr of kth time iteration is obtained
The hereby hidden layer state h of graceful machine{0}, wherein k values are set as 1；
(3b4) after sampling obtains hidden layer state, the activation value formula of i-th of visible node layer of reconstruct is：
The activation value formula is represented in known hidden layer state h{0}When, sampling obtains the probability that i-th of visible layer nodal value is 1, leads to
It crosses to after whole visible layer node samples, obtaining visible layer state v{1}；
(3b5) set the training sample set of second limited Boltzmann machine as：
Wherein, nMFor the number of training sample, vmIt is m-th of training sample:
(3b6) is in known second limited Boltzmann machine initiation parameter collection θ(1)Under conditions of, obtain m-th of training sample
Probability:
Wherein, θ(1)={ Wij (0),bi (0),cj (0), Z(1)It itemizes for second matching for limited Boltzmann machine；
(3b7) is to P (vm| θ) it asks about Wij (1),bi (1),cj (1)Partial derivative, obtain the gradient of parameters：
Wherein Δ wij (0)Represent the gradient of the weight matrix of second limited Boltzmann machine, Δ bi (0)Represent second limited glass
The gradient of the visual layers biasing of the graceful machine of Wurz, Δ cj (0)Represent the gradient of the hidden layer biasing of second limited Boltzmann machine；
The gradient that (3b8) is obtained according to (3a6) obtains the parameter W after second limited Boltzmann machine trainingij *, bi *And cj *：
Wij *=Wij (0)+λΔWij (0)
bi *=bi (0)+λΔbi (0)
cj *=cj (0)+λΔcj (0)
Wherein λ represents the learning rate of parameters.
4. according to the method described in claim 1, wherein step 5) uses the picture with class label using back-propagation algorithm BP
The parameter of the plain trained depth confidence network DBN of set pair carries out the micro-adjustment for having supervision, realizes as follows：
5a) propagated forward process：Input matrix travels to third from first limited Boltzmann machine, i.e., will be previous limited
Input of the output of Boltzmann machine hidden layer as next limited Boltzmann machine visual layers obtains the limited Bohr of third hereby
The reality output of graceful machine hidden layer；
5b) back-propagating process：By 5a) reality output and anticipated output error successively back-propagation, i.e., by error successively
Boltzmann machine is limited to the first two to propagate, and the weights and biasing of trained depth confidence network DBN are micro-adjusted, are obtained
Depth confidence network DBN after to micro-adjustment is realized as follows：
5b1) calculate the sensitivity of each limited Boltzmann machine j-th of node of hidden layer：
δj=οj(1-οj)(dj-οj),
Wherein, οjRepresent the reality output of j-th of node, djRepresent the anticipated output of j-th of node；
5b2) according to the sensitivity of j-th of node of hidden layer, the sensitivity of l-th of limited Boltzmann machine hidden layer is calculated：
Wherein δi lRepresent the sensitivity of l-th of limited Boltzmann machine hidden layer, yi lRepresent l-th limited Boltzmann machine hidden layer
Reality output, δj l+1Represent the sensitivity of l+1 limited Boltzmann machine hidden layers；
5b3) according to the sensitivity of each limited Boltzmann machine hidden layer, the power after each limited Boltzmann machine micro-adjustment is obtained
Value matrix wij lWith biasing cj l：
wij l=wij l+ε×yi lδj l+1
cj l=cj l+ε×δj l+1
Wherein ε represents the learning rate of micro-adjustment.
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 true CN108171200A (en) | 2018-06-15 |
| CN108171200B 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) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109507655A (en) * | 2018-12-11 | 2019-03-22 | 西北工业大学 | SAR Target Recognition Algorithms based on guiding reconstruct and norm constraint DBN |
| 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 |
| CN113344046A (en) * | 2021-05-24 | 2021-09-03 | 电子科技大学 | Method for improving SAR image ship classification precision |
Citations (7)
| 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 |
| US20150036922A1 (en) * | 2010-09-02 | 2015-02-05 | Edge 3 Technologies, Inc. | Method and Apparatus for Spawning Specialist Belief Propagation Networks For Adjusting Exposure Settings |
| US20160163035A1 (en) * | 2014-12-03 | 2016-06-09 | Kla-Tencor Corporation | Automatic Defect Classification Without Sampling and Feature Selection |
| 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 |
-
2018
- 2018-01-12 CN CN201810031937.3A patent/CN108171200B/en active Active
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20150036922A1 (en) * | 2010-09-02 | 2015-02-05 | Edge 3 Technologies, Inc. | Method and Apparatus for Spawning Specialist Belief Propagation Networks For Adjusting Exposure Settings |
| 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 |
| US20160163035A1 (en) * | 2014-12-03 | 2016-06-09 | Kla-Tencor Corporation | Automatic Defect Classification Without Sampling and Feature Selection |
| 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)
| Title |
|---|
| MINGYUAN ZHOU等: "The Poisson Gamma Belief Network", 《NIPS 2015》 * |
| YANHE GUO等: "Wishart RBM based DBN for polarimetric synthetic radar data classification", 《2015 IEEE INTERNATIONAL GEOSCIENCE AND REMOTE SENSING SYMPOSIUM (IGARSS)》 * |
| 柴瑞敏等: "一种改进的深度置信网络及其在自然图像分类中的应用", 《计算机应用与软件》 * |
| 梁鑫等: "基于深度学习神经网络的SAR图像目标识别算法", 《江汉大学学报（自然科学版）》 * |
| 高贵等: "高分辨SAR图像中杂波的统计特性分析", 《信号处理》 * |
Cited By (4)
| 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 |
| CN109507655A (en) * | 2018-12-11 | 2019-03-22 | 西北工业大学 | SAR Target Recognition Algorithms based on guiding reconstruct and norm constraint DBN |
| 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 |
Also Published As
| Publication number | Publication date |
|---|---|
| CN108171200B (en) | 2022-05-17 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108985238B (en) | Impervious surface extraction method and system combining deep learning and semantic probability | |
| CN107145830B (en) | Hyperspectral image classification method based on spatial information enhancing and deepness belief network | |
| Zhong et al. | An unsupervised artificial immune classifier for multi/hyperspectral remote sensing imagery | |
| Palmason et al. | Classification of hyperspectral data from urban areas using morphological preprocessing and independent component analysis | |
| CN109145992A (en) | Cooperation generates confrontation network and sky composes united hyperspectral image classification method | |
| CN107832797B (en) | Multispectral image classification method based on depth fusion residual error network | |
| CN108846426A (en) | Polarization SAR classification method based on the twin network of the two-way LSTM of depth | |
| CN107944370A (en) | Classification of Polarimetric SAR Image method based on DCCGAN models | |
| CN108830243A (en) | Hyperspectral image classification method based on capsule network | |
| CN107491793B (en) | Polarized SAR image classification method based on sparse scattering complete convolution | |
| CN105427309A (en) | Multiscale hierarchical processing method for extracting object-oriented high-spatial resolution remote sensing information | |
| CN112434745A (en) | Occlusion target detection and identification method based on multi-source cognitive fusion | |
| CN105894013B (en) | Classification of Polarimetric SAR Image method based on CNN and SMM | |
| CN105046268B (en) | Classification of Polarimetric SAR Image method based on Wishart depth networks | |
| CN104318245A (en) | Sparse depth network based polarization SAR (Synthetic Aperture Radar) image classification | |
| CN108171200A (en) | SAR image sorting technique based on SAR image statistical distribution and DBN | |
| CN105787517B (en) | Classification of Polarimetric SAR Image method based on the sparse self-encoding encoder of small echo | |
| CN107145860A (en) | Classification of Polarimetric SAR Image method based on spatial information and deep learning | |
| Plaza et al. | Nonlinear neural network mixture models for fractional abundance estimation in AVIRIS hyperspectral images | |
| CN108256557B (en) | Hyperspectral image classification method combining deep learning and neighborhood integration | |
| Paola et al. | Comparisons of neural networks to standard techniques for image classification and correlation | |
| CN107563420B (en) | The Classification of Polarimetric SAR Image method encoded certainly based on scattering energy and storehouse | |
| Babalola et al. | Soil surface texture classification using RGB images acquired under uncontrolled field conditions | |
| CN111460943A (en) | Remote sensing image ground object classification method and system | |
| Heidary et al. | Urban change detection by fully convolutional siamese concatenate network with attention |
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 |