CN106067042B - Polarization SAR classification method based on semi-supervised depth sparseness filtering network - Google Patents
Polarization SAR classification method based on semi-supervised depth sparseness filtering network Download PDFInfo
- Publication number
- CN106067042B CN106067042B CN201610415914.3A CN201610415914A CN106067042B CN 106067042 B CN106067042 B CN 106067042B CN 201610415914 A CN201610415914 A CN 201610415914A CN 106067042 B CN106067042 B CN 106067042B
- Authority
- CN
- China
- Prior art keywords
- sample
- depth
- network
- sparseness filtering
- training
- 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
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
Abstract
The semi-supervised deep learning method based on semi-supervised sparseness filtering that the invention discloses a kind of.It is complicated to solve conventional depth learning method parameter regulation, the not high technical problem of nicety of grading when possessing lower label data, step includes: input polarimetric SAR image data to be sorted;Extract training sample and test sample;Seek Wishart neighbour's sample of training sample;Initialize the parameter of depth sparseness filtering network;To depth sparseness filtering network pre-training;Depth sparseness filtering network is finely tuned;Class prediction is carried out to test sample;Export the classification image and nicety of grading of polarimetric SAR image to be sorted.The present invention, which passes through, constructs novel depth sparseness filtering network model, and the method for semi-supervised regular terms is added during pre-training, reduces the complexity of deep learning network parameter adjusting, improves the precision of polarimetric SAR image terrain classification.It can be used for the technical fields such as environmental monitoring, earth resources survey and military system.
Description
Technical field
The invention belongs to technical field of image processing, in particular to a kind of polarimetric SAR image terrain classification method, specifically
A kind of polarization SAR classification method based on semi-supervised depth sparseness filtering network.Can be used for environmental monitoring, earth resources survey and
Military system etc..
Background technique
Machine learning (Machine Learning, ML) is a multi-field cross discipline, be related to probability theory, statistics,
The multiple subjects such as Approximation Theory, convextiry analysis, algorithm complexity theory.Specialize in the study that the mankind were simulated or realized to computer how
Behavior reorganizes the existing structure of knowledge and is allowed to constantly improve the performance of itself to obtain new knowledge or skills.It is polarizing
SAR image classification field, machine learning have had many breakthrough progress, such as Wishart maximum
Likelihood (WML), the methods of support vector machine (support vector machines, SVM).
Common machine learning method is time-consuming and laborious mostly with the artificial method for extracting feature, and can not necessarily take
Obtain satisfactory feature.Deep learning is a new field in machine learning research, it is a kind of simulation human brain progress
The neural network of analytic learning imitates the mechanism of human brain to explain data.For Classification of Polarimetric SAR Image, deep learning network
It can automatically learn from polarization SAR data to more abstract high-rise expression attribute or feature, the feature learnt can
More effectively to apply to the research such as terrain classification, environmental monitoring.
And existing deep learning model: stack self-encoding encoder (SAE) limits Boltzmann machine (RBM), depth confidence net
(DBN), convolutional neural networks (CNN) etc. require to adjust many parameters.Such as it is learning rate (learning rates), dynamic
(momentum), degree of rarefication penalty coefficient (sparsity penalties) etc. are measured, and final the determining of these parameters needs
It is obtained by cross validation, this just requires a great deal of time and energy.With the continuous development of remote sensing fields, environment prison
Demand of the application such as survey, earth resources survey, military system to polarimetric SAR image is handled increases, it is desirable to polarimetric SAR image
Although terrain classification obtains ideal as a result, deep learning has more apparent advantage in machine learning method, tradition
Deep learning network need a large amount of parameter regulation, this will consume a large amount of time, and improperly parameter selection will directly affect
The performance of deep learning network and final classification results, this just restricts deep learning method in the application of remote sensing fields.
Summary of the invention
It is an object of the invention to be directed to the deficiency of above-mentioned prior art, propose a kind of based on the sparse filter of semi-supervised depth
The polarization SAR classification method of wave network improves terrain classification accuracy.
The present invention is a kind of polarization SAR classification method based on semi-supervised depth sparseness filtering network, which is characterized in that packet
Include following steps:
(1) polarimetric SAR image data to be sorted, i.e. the coherence matrix T of polarimetric SAR image, according to polarization SAR figure are inputted
The atural object distributed intelligence of picture obtains label matrix Y, and the distribution of same atural object is indicated by same category label, not can determine that class
Other atural object is distributed in label matrix and is indicated with 0, generates sample matrix according to the coherence matrix T of polarimetric SAR imageN is the total number of sample, xiIndicate i-th of sample.
(2) training sample and test sample are extracted, according to the sample data X and label matrix Y of polarimetric SAR image, at random
Extract L training sample, M test sample, wherein L+M=N, according to the classification information of all samples, every class is randomly selected
For 1% sample as training sample, remaining is test sample.
(3) the Wishart neighbour's sample for seeking training sample seeks each training sample x in all sample datasi
(i=1,2...L) K Wishart neighbour's sample x corresponding toj(j=1,2...K), so far, basic data processing are completed.
(4) start to optimize depth network, initialize the basic parameter of depth network, random initializtion depth sparseness filtering net
The weight parameter W of network, the number of nodes that every layer of network of set depth sparseness filtering determine the whole knot of depth sparseness filtering network
Structure.
(5) to depth sparseness filtering network pre-training, training sample and its corresponding Wishart neighbour sample are sent to
Pre-training is carried out in depth sparseness filtering network, using the pre-training method of layer-by-layer greediness, the output of preceding layer is as later layer
Input, until the last one complete hidden layer of training, while semi-supervised neighbour is added in every layer of pre-training and keeps regular terms, with
Sparse filter optimizes the weight of network jointly, and the weight of depth sparseness filtering network is tentatively optimized.
(6) depth sparseness filtering network is finely tuned, using training sample and its label information, in conjunction with Softmax classifier
Depth sparseness filtering network is finely adjusted, the weight of network is advanced optimized, network is made to become more stable, so far, depth is dilute
Filter network optimization is dredged to complete.
(7) performance of test depth sparseness filtering network carries out class prediction to test sample, test sample is sent to
Depth sparseness filtering network is predicted using class label of the Softmax classifier to test sample, obtains each test specimens
This prediction classification.
(8) the classification results figure and nicety of grading for exporting polarimetric SAR image to be sorted, according to training sample and step
(7) it is predicted that the test sample of classification out, exports the final classification result of polarimetric SAR image to be sorted and calculate this point in
The precision of class.
Technical thought of the invention is: the sparse filter of single layer is expanded as a kind of novel depth sparseness filtering net
Network, in conjunction with semi-supervised thought, carries out feature learning to polarization SAR data, by dividing in the case where possessing a small amount of marker samples
Class device realizes terrain classification, improves classification accuracy.
The present invention has the advantage that
1, method of the present invention due to using deep learning, automatically learns to polarize using depth sparseness filtering network
The feature of SAR image, thus avoid the cumbersome of artificial learning characteristic in conventional method, and depth sparseness filtering network can be with
Learn the feature to polarimetric SAR image more abstract entities, these features are more conducive to the terrain classification of polarimetric SAR image.
2, the method for the present invention obtains depth sparseness filtering network, this depth due to expanding on the basis of sparse filter
Practising network has less parameter and stable performance, thus it is complicated effectively to compensate for conventional depth learning network parameter, it is difficult to
The defect of adjusting, suitably avoid because parameter regulation it is improper caused by the undesirable situation of classification results.
3, the method for the present invention due to use semi-supervised learning method, joined during pre-training it is semi-supervised just
Then item, thus improve that the classifier classification accuracy rate in the case where marker samples are less is lower and a large amount of unmarked samples
Caused by information waste the problem of.
Detailed description of the invention
Fig. 1 is implementation process schematic diagram of the invention;
Fig. 2 is the experimental result picture to polarization SAR geo-objects simulation image, and wherein Fig. 2 (a) is polarization SAR analogous diagram
Pauli exploded view, Fig. 2 (b) are the label figures of analogous diagram, and Fig. 2 (c) is the classification results figure using control methods DNN, Fig. 2 (d)
It is the classification results figure using control methods SAE, Fig. 2 (e) is the classification results figure using control methods DSF, and Fig. 2 (f) is to adopt
With the classification results figure of the method for the present invention.
Specific embodiment
With reference to the accompanying drawing to the detailed description of the invention
Embodiment 1
Because of the development of remote sensing technology, has in fields such as environmental monitoring, earth resources survey, military systems and widely answer
With the demand to polarimetric SAR image processing also continues to increase, and deep learning has more apparent excellent in machine learning method
Gesture, and traditional deep learning network needs a large amount of parameter regulation, can consume a large amount of time, thereby increases and it is possible to directly affect depth
The performance of learning network and final classification results, therefore, the present invention propose a kind of based on semi-supervised depth sparseness filtering net
The polarization SAR classification method of network includes the following steps: referring to Fig. 1
(1) polarimetric SAR image data to be sorted, i.e. the coherence matrix T of polarimetric SAR image, according to polarization SAR figure are inputted
The atural object distributed intelligence of picture obtains label matrix Y, and same atural object is regardless of being distributed, by same classification in label matrix
Label indicates that not can determine that the atural object of classification is distributed in label matrix is indicated with 0, raw by the coherence matrix T of polarimetric SAR image
At sample matrixN is the total number of sample, xiIndicate i-th of sample.
1a, take polarimetric SAR image polarization coherence matrix T upper angular position 6 elements modulus value as each picture
The primitive character of vegetarian refreshments;
1b, two-dimensional sample matrix is converted by T matrix using the reshape function in MATLAB softwareX
∈RP×N, P=6 is the dimension of sample matrix, and N is total sample number, and each column indicate a sample, this is to polarization SAR data
Basic handling.
(2) training sample and test sample are extracted, according to the sample data X and label matrix Y of polarimetric SAR image, at random
Extract L training sample, M test sample, wherein L+M=N, according to the classification information of all samples, every class is randomly selected
For 1% sample as training sample, remaining is test sample.
(3) the Wishart neighbour's sample for seeking training sample seeks each training sample x in all sample datasi
(i=1,2...L) K Wishart neighbour's sample x corresponding toj(j=1,2...K), since polarization SAR data itself are obeyed
Wishart distribution, Wishart more likely belong to identical classification between smaller two samples, so far, basic number
It is completed according to processing.
(4) start to optimize depth sparseness filtering network, initialize the basic parameter of depth sparseness filtering network, it is random initial
Change the weight parameter W of depth sparseness filtering network, the number of nodes that every layer of network of set depth sparseness filtering.
(5) to depth sparseness filtering network pre-training, training sample and its corresponding Wishart neighbour sample are sent to
Pre-training is carried out in depth sparseness filtering network, using the pre-training method of layer-by-layer greediness, the output of preceding layer is as later layer
Input semi-supervised neighbour is added in every layer of pre-training and keeps regular terms until the last one complete hidden layer of training, and it is sparse
The weight of the common optimization network of filtering.It is dilute that the canonical item constraint of training sample and its Wishart neighbour's sample can optimize depth
Filter network structure is dredged, makes it be more conducive to carrying out terrain classification to the polarization SAR data for obeying Wishart distribution, simultaneously
The use to unmarked sample is further enhanced in pre-training, can also be obtained very in the case where there are a small amount of marker samples
Good terrain classification precision.Sparseness filtering itself has less parameter to need to adjust, and being expanded using greedy algorithm is that depth is dilute
The new sample characteristics for having biggish advantage in terms of parameter regulation after filter network structure, and learning are dredged for atural object point
Class also has higher precision.
(6) depth sparseness filtering network is finely tuned, using training sample and its label information, in conjunction with Softmax classifier
Depth sparseness filtering network is finely adjusted, the weight of network is advanced optimized, keeps network more stable.
(7) class prediction is carried out to test sample, test sample is sent to depth sparseness filtering network, utilized
Softmax classifier predicts the class label of test sample, obtains the prediction classification of each test sample.
(8) the classification results figure and nicety of grading for exporting polarimetric SAR image to be sorted, according to training sample and step
(7) it is predicted that the test sample of classification out in, the final classification of polarimetric SAR image to be sorted is exported as a result, referring to fig. 2 (f),
And calculate the precision of this subseries.
The present invention expands on the basis of single layer sparse filter and obtains depth sparseness filtering network, depth of the invention
Practising network has less parameter regulation, it is complicated effectively to compensate for conventional depth learning network parameter, it is difficult to the defect of adjusting;This
Invention has stable performance, suitably avoid because parameter regulation it is improper caused by the undesirable situation of classification results.
Embodiment 2
Polarization SAR classification method based on semi-supervised depth sparseness filtering network is asked described in (3) with embodiment 1, step
Wishart neighbour's sample of each training sample, is to obtain as follows:
3a, training sample matrix areIt is sought between training sample and other samples using following formula
Wishart distance:
d(xi,xj)=ln ((xi)-1xj)+Tr((xj)-1xi)-q (x=1,2 ..., L, j=1,2 ..., N),
Wherein, the mark of Tr () representing matrix is integrated radar for sending and receiving, due to reciprocity, q=3;For
Send and receive is not integrated radar, q=4;
3b, using the sort function in MATLAB, the Wishart distance d (x that will be acquired in 3ai,xj) press absolute value ascending order
Arrangement finds corresponding preceding K sample, as training sample x K before takingiWishart neighbour's sample, be denoted as:
Embodiment 3
Polarization SAR classification method based on semi-supervised depth sparseness filtering network is with embodiment 1-2, described in step (4)
Initialize the parameter of depth network are as follows:
4a, depth sparseness filtering network hidden layer number be 3, every layer of number of nodes is respectively as follows: 25,100,50;
4b, the weight W for initializing depth sparseness filtering network1,W1∈RD×Q, D is the dimension of input signal, and Q is first
The number of nodes of hidden layer.
Depth sparseness filtering network of the invention has less parameter to need to adjust, and parameter initialization is convenient, network
In training process also only have simple regular terms parameter need to adjust compare with other deep learning methods time-consuming it is less.
Embodiment 4
Polarization SAR classification method based on semi-supervised depth sparseness filtering network is with embodiment 1-3, described in step (5)
To the pre-training process of depth sparseness filtering network are as follows:
5a, the pre-training sample for inputting depth sparseness filtering network, by training sample xi(i=1,2...L) and its it is corresponding
K Wishart neighbour's sampleIt is input in depth network as pre-training sample;
5b, learn new sample characteristics using first hidden layer, using traditional nonlinear s igmoid function as activation letter
Number: φ (z)=(1+exp (- z))-1, x ∈ RD×1It is input vector, then the output of first hidden layer are as follows:I-th
The output of sample is expressed as h1(xi), as new sample characteristics;
5c, optimization network weight W, each layer of sparseness filtering network solve following objective function using LBFGS algorithm, from
And optimize the weight W of network, objective function are as follows:
Wherein, in objective functionFor sparseness filtering objective function,It is half
It supervises neighbour and keeps regular terms, optimize the structure of network, λ using the neighbor relationships between sample referring to the method for manifold learning
For regular terms parameter,U(xi) indicate training sample xiCorresponding Wishart neighbour sample set;
5d, greedy algorithm train remaining hidden layer, the new sample characteristics h that will learn in 5b1It is sent to as new input
Next layer utilizes the objective function L in 5c by upper one layer of output as next layer of input using greedy learning algorithm1
(W) the weight W for successively optimizing each hidden layer, until the last one complete hidden layer of training, sample xiPass through the defeated of the last one hidden layer
It can be expressed as h out3(xi)。
The present invention combines to carry out depth network pre-training by sparseness filtering and semi-supervised regular terms, by the power of network
Weight parameter optimization has advanced optimized the structure of network to a suitable range, is more advantageous to obedience Wishart distribution
Polarization SAR data sample carries out terrain classification.
Embodiment 5
Polarization SAR classification method based on semi-supervised depth sparseness filtering network is with embodiment 1-4, described in step (6)
Fine tuning to depth sparseness filtering network has following objective function:
Wherein,It is mean square error item,It is weight attenuation term, the mesh of weight attenuation term
Be reduce weight amplitude, prevent over-fitting, yiIndicate training sample xiCorresponding class label, h (xi) it is that will train sample
This xiThe feature h learnt after entire depth network3(xi), it is sent to output obtained in Softmax classifier, β=3e-3
For weight attenuation parameter;
Above-mentioned objective function L is solved using gradient descent algorithm2(W), the weight W of sparseness filtering network is advanced optimized,
The fine tuning for realizing whole network, to construct depth sparseness filtering network, using the e-learning to feature would be even more beneficial to
Realize the terrain classification of polarimetric SAR image.
A more detailed example is given below, the present invention is further described:
Embodiment 6
Polarization SAR classification method based on semi-supervised depth sparseness filtering network is with embodiment 1-5, referring to Fig.1, the present invention
Specific implementation step it is as follows:
Step 1, input polarimetric SAR image data to be sorted input the relevant square of polarimetric SAR image referring to fig. 2 (a)
Battle array T, obtains label matrix Y according to the atural object distributed intelligence of polarimetric SAR image, referring to fig. 2 (b), Fig. 2 (b) is exactly by label square
The image that directly generates of battle array Y, different color lumps represents different atural object in image, and same atural object is distributed in label matrix
It is indicated by same category label, not can determine that the atural object of classification is distributed in label matrix is indicated with 0, according to polarization SAR figure
The coherence matrix T of picture generates sample matrixN is the total number of sample, xiIndicate i-th of sample.
This example uses polarization SAR geo-objects simulation image.Since the polarization coherence matrix T of polarization SAR data is Hamilton
Positive semidefinite matrix, it is possible to which the modulus value for extracting 6 elements of the upper angular position for the polarization coherence matrix T that dimension is 3 × 3 is made
For the primitive character of each pixel, two-dimensional sample moment is converted by T matrix using the reshape function in MATLAB software
Battle array, each column indicate a sample, and emulation data have 18000 samples, and the dimension of each sample is 6 dimensions.Each sample is corresponding
A pixel in polarimetric SAR image.
Step 2 extracts training sample and test sample, will according to the sample data X and label matrix Y of polarimetric SAR image
Sample data category extracts L training sample at random, M test sample, wherein L+M=N.Sample data is each according to it
Classification from place is extracted as training sample and test sample according to the ratio of 1:99 at random, and every class testing sample accounts for such
The 1% of sum.
Step 3 asks Wishart neighbour's sample of training sample to seek each training sample in all sample datas
xi(i=1,2...L) K Wishart neighbour's sample x corresponding toj(j=1,2...K).
3a, for training sample xi(i=1,2...L) asks it and other samples according to improved Wishart range formula
Wishart distance between this: d (xi,xj)=ln ((xi)-1xj)+Tr((xj)-1xi)-q (x=1,2 ..., L, j=1,
2 ..., N),
Wherein, the mark of Tr () representing matrix is integrated radar for sending and receiving, due to reciprocity, q=3, for
Sending and receiving is not integrated radar, q=4, and the polarization SAR geo-objects simulation image data that this example uses is by sending and connecing
The radar system for receiving one, which obtains, to be generated, and q takes 3.
3b, using the sort function in MATLAB, the Wishart distance d (x that will be acquired in 3ai,xj) press absolute value ascending order
Arrangement finds corresponding preceding K sample, as training sample x K before takingiWishart neighbour's sample, be denoted as:
Step 4, the basic parameter for initializing depth sparseness filtering network, the power of random initializtion depth sparseness filtering network
Weight parameter W, the number of nodes that every layer of network of set depth sparseness filtering.
Three the number of hidden nodes of depth sparseness filtering network are set, respectively 25,100,50;Initialize the weight of network
Matrix W1,W1∈RD×N, D is the dimension of input signal, and N is the number of nodes of first hidden layer.
Depth sparseness filtering network of the invention has less parameter and stable performance, effectively compensates for conventional depth
Learning network parameter is complicated, it is difficult to which the defect of adjusting is suitably avoided because classification results are undesirable caused by parameter regulation is improper
Situation.
Step 5, to depth sparseness filtering network pre-training.
5a, the pre-training sample for inputting depth sparseness filtering network, by training sample xiAnd its corresponding K Wishart
Neighbour's sampleIt is input in depth network as pre-training sample, i=1,2 ..., L;
5b, learn new sample characteristics using first hidden layer, using traditional nonlinear s igmoid function as activation letter
Number: φ (z)=(1+exp (- z))-1, x ∈ RD×1It is input vector, then the output of first hidden layer are as follows:I-th
The output of sample is expressed as h1(xi), as new sample characteristics;
5c, optimization network weight W, each layer of sparseness filtering network solve following objective function using LBFGS algorithm, from
And optimize the weight W of network, objective function are as follows:
Wherein, in objective functionFor sparseness filtering objective function,It is half
It supervises neighbour and keeps regular terms, optimize the structure of network, λ using the neighbor relationships between sample referring to the method for manifold learning
For regular terms parameter,U(xi) indicate training sample xiCorresponding Wishart neighbour sample set.
5d, greedy algorithm train remaining hidden layer, the new sample characteristics h that will learn in 5b1It is sent to as new input
Next layer utilizes the objective function L in 5c by upper one layer of output as next layer of input using greedy learning algorithm1
(W) the weight W for successively optimizing each hidden layer, until the last one complete hidden layer of training, sample xiPass through the defeated of the last one hidden layer
It can be expressed as h out3(xi)。
The present invention combines to carry out depth network pre-training by sparseness filtering and semi-supervised regular terms, by the power of network
Weight parameter optimization has advanced optimized the structure of network to a suitable range, is more advantageous to obedience Wishart distribution
Polarization SAR data sample carries out terrain classification.
Step 6 finely tunes depth sparseness filtering network.
After the completion of pre-training, the weight W of network has converged to reasonable range, further right with its using training sample
Class label information is answered, network is finely adjusted in conjunction with Softmax classifier:
Wherein,It is mean square error item,It is weight attenuation term, the mesh of weight attenuation term
Be reduce weight amplitude, prevent over-fitting, yiIndicate training sample xiCorresponding class label, h (xi) it is that will train sample
This xiThe feature h learnt after entire depth network3(xi), it is sent to output obtained in Softmax classifier, β=3e-3
For weight attenuation parameter;
Above-mentioned objective function L is solved using gradient descent algorithm2(W), the weight W of sparseness filtering network is advanced optimized,
Realize the fine tuning of whole network.
Step 7 carries out class prediction to test sample, and test sample is sent to depth sparseness filtering network, is utilized
Softmax classifier predicts the class label of test sample, obtains the prediction classification of each test sample.
7a, by test sample xjIt is input in the depth sparseness filtering network having had been built up, final feature is arrived in study
θj, j=1,2 ..., M;
7b, it the final feature learnt in 7a is sent in Softmax classifier carries out class prediction:
The output of Softmax classifier is y ∈ RP×1, P is expressed as classification number, test sample xjPrediction classification can be with table
It is shown as:
Wherein α is sample xjPrediction classification confidence level.
The classification synthetism picture mosaic and nicety of grading of step 8, output polarimetric SAR image to be sorted, according to training sample and
It is predicted that the test sample of classification out, exports the final classification result of polarimetric SAR image to be sorted and calculate this in step 7
The precision of classification.
(8a) utilizes the sample class of classifier prediction, corresponds to each pixel in polarimetric SAR image, by it is red,
Green, blue are used as three primary colours, are the colouring of each pixel according to color method in three primary colours, export result figure, referring to fig. 2 (f).
(8b) compares the category label that classifier is predicted with the true category label of test sample, obtains experiment
Classification accuracy rate.
Method of the present invention due to using deep learning, automatically learns to polarize using depth sparseness filtering network
The feature of SAR image, thus the cumbersome of artificial learning characteristic in conventional method is avoided, depth sparseness filtering network can learn
To the feature of polarimetric SAR image more abstract entities, these features are more conducive to the terrain classification of polarimetric SAR image.
Effect of the invention can be illustrated by emulation experiment:
Embodiment 7
Polarization SAR classification method based on semi-supervised depth sparseness filtering network with embodiment 1-6,
1. experiment condition
Hardware platform are as follows: Intel (R) Core (TM) i5-2410M CPU@2.30GHz, RAM 4.00GB;
Software platform are as follows: MATLAB R2011b;
Experiment selects 120 × 150 polarization SAR geo-objects simulation image to be tested, and has 18000 sample points, each sample
The dimension of this point is 6 dimensions, and classification number is 9, is labeled as Ci, i=1,2 ..., 9.In experiment, every class randomly selects 1% sample
As training sample, remaining is test sample.
2. experiment content and result
Present invention combination Softmax classifier classifies to polarization SAR geo-objects simulation figure, before same experimental setup
Put and be compared with other deep learning methods, wherein DNN be deep neural network, Fig. 2 (c) be by DNN to Fig. 2 (a) into
The result figure of row classification;SAE stack self-encoding encoder is additionally used in experiment, Fig. 2 (d) classifies to Fig. 2 (a) with SAE method
Result figure;Fig. 2 (e) is the result figure classified using DSF depth sparseness filtering network to Fig. 2 (a);SDSF is the present invention
Method.Table 1 is the terrain classification precision of the polarization SAR geo-objects simulation image respectively obtained using above-mentioned 4 kinds of methods and totally divided
Class precision.
The terrain classification precision (%) and overall classification accuracy (%) of table 1, various methods in analogous diagram
From table 1 it follows that in the case where training sample is 1%, the present invention and existing deep learning method
Compared to nicety of grading with higher.Referring to the simulation experiment result Fig. 2, the present invention has higher visuality.
In conclusion the polarization SAR classification method proposed by the present invention based on semi-supervised depth sparseness filtering network improves
Conventional depth learning method parameter regulation is excessive, requires training sample larger problem, can effectively improve polarization SAR figure
The nicety of grading of picture, and higher nicety of grading can be also obtained in the case where training sample is less.
The semi-supervised deep learning method based on semi-supervised sparseness filtering that the invention discloses a kind of.Solves conventional depth
Learning method parameter regulation is complicated, the not high technical problem of nicety of grading when possessing lower label data, step includes: 1,
Polarimetric SAR image data to be sorted, i.e. the coherence matrix T of polarimetric SAR image are inputted, according to the atural object of polarimetric SAR image point
Cloth information obtains label matrix Y, generates sample matrix X by the coherence matrix T of polarimetric SAR image;2, training sample and survey are extracted
Sample sheet extracts L training sample, M test specimens according to the sample data X and label matrix Y of polarimetric SAR image at random
This;3, the Wishart neighbour's sample for seeking training sample seeks each training sample x in all sample datasi(i=1,
2...L K Wishart neighbour's sample corresponding to);4, the parameter of depth sparseness filtering network is initialized, random initializtion is deep
Spend the weight parameter W of sparseness filtering network, the number of nodes that every layer of network of set depth sparseness filtering;5, depth sparseness filtering network
Training sample and its corresponding Wishart neighbour sample are sent in depth sparseness filtering network and carry out pre-training by pre-training,
Using the pre-training method of layer-by-layer greediness, input of the output of preceding layer as later layer, until the last one complete hidden layer of training,
Semi-supervised neighbour is added in every layer of pre-training simultaneously and keeps regular terms, optimizes the weight of network jointly with sparseness filtering;6,
Depth sparseness filtering network fine tuning, using training sample and its label information, in conjunction with Softmax classifier to depth sparseness filtering
Network is finely adjusted, and advanced optimizes the weight of network;7, class prediction is carried out to test sample, test sample is sent to depth
Sparseness filtering network is spent, is predicted using class label of the Softmax classifier to test sample, obtains each test sample
Prediction classification;8, the classification image and nicety of grading for exporting polarimetric SAR image to be sorted, according to training sample it is predicted that going out
The test sample of classification exports the final classification result of polarimetric SAR image to be sorted and calculates the precision of classification.
The present invention reduces the adjusting of deep learning network parameter by constructing novel depth sparseness filtering network model
Complexity, and the method that semi-supervised neighbour keeps regular terms is added during pre-training, reduce the number of training sample, leads to
Cross the accuracy that polarimetric SAR image terrain classification is improved by a small amount of training sample.It can be used for environmental monitoring, earth resource
The application such as exploration and military system.
Claims (5)
1. a kind of polarization SAR classification method based on semi-supervised depth sparseness filtering network, which is characterized in that including walking as follows
It is rapid:
(1) polarimetric SAR image data to be sorted, i.e. the coherence matrix T of polarimetric SAR image, according to polarimetric SAR image are inputted
Atural object distributed intelligence obtains label matrix Y, generates sample matrix by the coherence matrix T of polarimetric SAR imageN is sample
This total number, xiIndicate i-th of sample;
(2) training sample and test sample are extracted, it is random to extract according to the sample data X and label matrix Y of polarimetric SAR image
L training sample out, M test sample, wherein L+M=N, according to the classification information of all samples, every class randomly selects 1%
For sample as training sample, remaining is test sample;
(3) the Wishart neighbour's sample for seeking training sample seeks each training sample x in all sample datasi(i=1,
2...L K Wishart neighbour's sample x corresponding to)j(j=1,2...K);
(4) basic parameter of depth sparseness filtering network, the weight parameter of random initializtion depth sparseness filtering network are initialized
W, the number of nodes that every layer of network of set depth sparseness filtering;
(5) to depth sparseness filtering network pre-training, training sample and its corresponding Wishart neighbour sample are sent to depth
Pre-training is carried out in sparseness filtering network, using the pre-training method of layer-by-layer greediness, the output of preceding layer is as the defeated of later layer
Enter, until the last one complete hidden layer of training, while semi-supervised neighbour is added in every layer of pre-training and keeps regular terms, and it is sparse
Filter optimizes the weight of network jointly;
(6) depth sparseness filtering network is finely tuned, using training sample and its label information, in conjunction with Softmax classifier to depth
Degree sparseness filtering network is finely adjusted, and advanced optimizes the weight of network;
(7) class prediction is carried out to test sample, test sample is sent to depth sparseness filtering network, utilize Softmax points
Class device predicts the class label of test sample, obtains the prediction classification of each test sample;
(8) the classification results figure and nicety of grading for exporting polarimetric SAR image to be sorted, according to training sample and it is predicted that class out
Other test sample exports the final classification result of polarimetric SAR image to be sorted and calculates the precision of this subseries.
2. the polarization SAR classification method according to claim 1 based on semi-supervised depth sparseness filtering network, feature exist
In seeking Wishart neighbour's sample of each training sample described in step (3), be to obtain as follows:
3a, training sample matrix areIt is sought between training sample and other samples using following formula
Wishart distance:
d(xi,xj)=ln ((xi)-1xj)+Tr((xj)-1xi)-q (x=1,2 ..., L, j=1,2 ..., N),
Wherein, the mark of Tr () representing matrix is integrated radar for sending and receiving, due to reciprocity, q=3;For sending
It is not integrated radar, q=4 with receiving;
3b, using the sort function in MATLAB, the Wishart distance d (x that will be acquired in 3ai,xj) arranged by absolute value ascending order,
K before taking, corresponding preceding K sample is found, as training sample xiWishart neighbour's sample, be denoted as:
3. the polarization SAR classification method according to claim 1 based on semi-supervised depth sparseness filtering network, feature exist
In the parameter of initialization depth network described in step (4) are as follows:
4a, depth sparseness filtering network hidden layer number be 3, every layer of number of nodes is respectively as follows: 25,100,50;
4b, the weight W for initializing depth sparseness filtering network1,W1∈RD×Q, D is the dimension of input signal, and Q is first hidden layer
Number of nodes.
4. the polarization SAR classification method according to claim 1 based on semi-supervised depth sparseness filtering network, feature exist
In wherein to the pre-training process of depth sparseness filtering network described in step (5) are as follows:
5a, the pre-training sample for inputting depth sparseness filtering network, by training sample xiAnd its corresponding K Wishart neighbour's sample
ThisIt is input in depth network as pre-training sample, i=1,2 ..., L;
5b, learn new sample characteristics using first hidden layer, using traditional nonlinear s igmoid function as activation primitive:
φ (z)=(1+exp (- z))-1, x ∈ RD×1It is input vector, then the output of first hidden layer are as follows: h1=φ (W1 TX), i-th
The output of sample is expressed as h1(xi), as new sample characteristics;
5c, optimization network weight W, each layer of sparseness filtering network solves following objective function using LBFGS algorithm, thus excellent
Change the weight W of network, objective function are as follows:
Wherein, in objective functionFor sparseness filtering objective function,It is semi-supervised
Neighbour keeps regular terms, optimizes the structure of network using the neighbor relationships between sample referring to the method for manifold learning, λ is positive
Then item parameter,U(xi) indicate training sample xiCorresponding Wishart neighbour sample set;
5d, greedy algorithm train remaining hidden layer, the new sample characteristics h that will learn in 5b1It is sent to as new input next
Layer utilizes the objective function L in 5c by upper one layer of output as next layer of input using greedy learning algorithm1(W) according to
The weight W of each hidden layer of suboptimization, until the last one complete hidden layer of training, sample xiIt is indicated by the output of the last one hidden layer
For h3(xi)。
5. the polarization SAR classification method according to claim 1 based on semi-supervised depth sparseness filtering network, feature exist
In fine tuning described in step (6) to depth sparseness filtering network has following objective function:
Wherein,It is mean square error item,It is weight attenuation term, the purpose of weight attenuation term exists
In the amplitude for reducing weight, over-fitting, y are preventediIndicate training sample xiCorresponding class label, h (xi) it is by training sample xi
The feature h learnt after entire depth network3(xi), it is sent to output obtained in Softmax classifier, β=3e-3 is power
Weight attenuation parameter;
Above-mentioned objective function L is solved using gradient descent algorithm2(W), the weight W of sparseness filtering network is advanced optimized, is realized whole
The fine tuning of a network.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201610415914.3A CN106067042B (en) | 2016-06-13 | 2016-06-13 | Polarization SAR classification method based on semi-supervised depth sparseness filtering network |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201610415914.3A CN106067042B (en) | 2016-06-13 | 2016-06-13 | Polarization SAR classification method based on semi-supervised depth sparseness filtering network |
Publications (2)
Publication Number | Publication Date |
---|---|
CN106067042A CN106067042A (en) | 2016-11-02 |
CN106067042B true CN106067042B (en) | 2019-02-15 |
Family
ID=57421231
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201610415914.3A Active CN106067042B (en) | 2016-06-13 | 2016-06-13 | Polarization SAR classification method based on semi-supervised depth sparseness filtering network |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN106067042B (en) |
Families Citing this family (17)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106778564A (en) * | 2016-12-02 | 2017-05-31 | 中国船舶重工集团公司第七二四研究所 | Based on the naval vessels of one-dimensional picture Feature-level fusion under various visual angles and freighter sorting technique |
CN106886798A (en) * | 2017-03-10 | 2017-06-23 | 北京工业大学 | The image-recognizing method of the limited Boltzmann machine of the Gaussian Profile based on matrix variables |
CN108694408B (en) * | 2017-04-11 | 2023-05-09 | 西安邮电大学 | Driving behavior recognition method based on deep sparse filtering convolutional neural network |
CN107133653B (en) * | 2017-05-23 | 2020-11-03 | 西安电子科技大学 | High-resolution SAR image classification method based on deep ladder network |
CN107478418A (en) * | 2017-06-29 | 2017-12-15 | 南京航空航天大学 | A kind of rotating machinery fault characteristic automatic extraction method |
CN107358203B (en) * | 2017-07-13 | 2019-07-23 | 西安电子科技大学 | A kind of High Resolution SAR image classification method based on depth convolution ladder network |
CN107491734B (en) * | 2017-07-19 | 2021-05-07 | 苏州闻捷传感技术有限公司 | Semi-supervised polarimetric SAR image classification method based on multi-core fusion and space Wishart LapSVM |
CN108225750A (en) * | 2017-12-12 | 2018-06-29 | 南京航空航天大学 | A kind of rotary machinery fault diagnosis method based on the more correlations of fault signature |
CN108805167B (en) * | 2018-05-04 | 2022-05-13 | 江南大学 | Sparse depth confidence network image classification method based on Laplace function constraint |
CN108764489B (en) * | 2018-06-05 | 2021-02-02 | 北京百度网讯科技有限公司 | Model training method and device based on virtual sample |
CN111461285B (en) * | 2019-01-21 | 2024-03-05 | 京东科技控股股份有限公司 | Method and device for detecting electric equipment |
CN110109059B (en) * | 2019-03-27 | 2023-04-07 | 西安电子科技大学 | Radar radiation source signal identification method based on deep learning network |
CN110472483B (en) * | 2019-07-02 | 2022-11-15 | 五邑大学 | SAR image-oriented small sample semantic feature enhancement method and device |
CN111089720B (en) * | 2020-01-16 | 2021-10-08 | 山东科技大学 | Regularization sparse filtering method suitable for gear fault diagnosis under variable rotating speed |
CN112507479B (en) * | 2020-11-19 | 2022-03-15 | 电子科技大学 | Oil drilling machine health state assessment method based on manifold learning and softmax |
CN114724030B (en) * | 2022-04-06 | 2023-06-02 | 西安电子科技大学 | Polarization SAR ground object classification method based on contrast learning |
CN114444552B (en) * | 2022-04-08 | 2022-08-16 | 新黎明科技股份有限公司 | Weak fault diagnosis method, medium, electronic equipment and system for explosion-proof motor bearing |
Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103593676A (en) * | 2013-11-29 | 2014-02-19 | 重庆大学 | High-spectral remote-sensing image classification method based on semi-supervision sparse discriminant embedding |
CN104008394A (en) * | 2014-05-20 | 2014-08-27 | 西安电子科技大学 | Semi-supervision hyperspectral data dimension descending method based on largest neighbor boundary principle |
CN104408476A (en) * | 2014-12-08 | 2015-03-11 | 西安电子科技大学 | Deep sparse main component analysis-based polarimetric SAR image classification method |
CN104463193A (en) * | 2014-11-04 | 2015-03-25 | 西安电子科技大学 | Polarization SAR image classifying method based on depth sparsity ICA |
CN104504393A (en) * | 2014-12-04 | 2015-04-08 | 西安电子科技大学 | SAR (Synthetic Aperture Radar) image semi-supervised classification method based on integrated learning |
CN105139028A (en) * | 2015-08-13 | 2015-12-09 | 西安电子科技大学 | SAR image classification method based on hierarchical sparse filtering convolutional neural network |
CN105426923A (en) * | 2015-12-14 | 2016-03-23 | 北京科技大学 | Semi-supervised classification method and system |
-
2016
- 2016-06-13 CN CN201610415914.3A patent/CN106067042B/en active Active
Patent Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103593676A (en) * | 2013-11-29 | 2014-02-19 | 重庆大学 | High-spectral remote-sensing image classification method based on semi-supervision sparse discriminant embedding |
CN104008394A (en) * | 2014-05-20 | 2014-08-27 | 西安电子科技大学 | Semi-supervision hyperspectral data dimension descending method based on largest neighbor boundary principle |
CN104463193A (en) * | 2014-11-04 | 2015-03-25 | 西安电子科技大学 | Polarization SAR image classifying method based on depth sparsity ICA |
CN104504393A (en) * | 2014-12-04 | 2015-04-08 | 西安电子科技大学 | SAR (Synthetic Aperture Radar) image semi-supervised classification method based on integrated learning |
CN104408476A (en) * | 2014-12-08 | 2015-03-11 | 西安电子科技大学 | Deep sparse main component analysis-based polarimetric SAR image classification method |
CN105139028A (en) * | 2015-08-13 | 2015-12-09 | 西安电子科技大学 | SAR image classification method based on hierarchical sparse filtering convolutional neural network |
CN105426923A (en) * | 2015-12-14 | 2016-03-23 | 北京科技大学 | Semi-supervised classification method and system |
Non-Patent Citations (3)
Title |
---|
SEMI-SUPERVISED CLASSIFICATION BASED ON ANCHOR-SPATIAL GRAPH FOR LARGE POLARIMETRIC SAR DATA;Hongying Liu等;《IGARSS 2015》;20150731;1845-1848 |
Semisupervised Feature Extraction With Neighborhood Constraints for Polarimetric SAR Classification;Hongying Liu等;《IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING》;20160315;第9卷(第7期);3001-3015 |
基于半监督学习的SVM-Wishart 极化SAR 图像分类方法;滑文强等;《雷达学报》;20150228;第4卷(第1期);93-98 |
Also Published As
Publication number | Publication date |
---|---|
CN106067042A (en) | 2016-11-02 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN106067042B (en) | Polarization SAR classification method based on semi-supervised depth sparseness filtering network | |
CN107563422B (en) | A kind of polarization SAR classification method based on semi-supervised convolutional neural networks | |
CN107909101B (en) | Semi-supervised transfer learning character identifying method and system based on convolutional neural networks | |
CN109086700B (en) | Radar one-dimensional range profile target identification method based on deep convolutional neural network | |
CN107463966B (en) | Radar range profile's target identification method based on dual-depth neural network | |
CN111639679B (en) | Small sample learning method based on multi-scale metric learning | |
CN107194336A (en) | The Classification of Polarimetric SAR Image method of network is measured based on semi-supervised depth distance | |
CN108021947B (en) | A kind of layering extreme learning machine target identification method of view-based access control model | |
CN107358293A (en) | A kind of neural network training method and device | |
CN107132516A (en) | A kind of Radar range profile's target identification method based on depth confidence network | |
CN105975931A (en) | Convolutional neural network face recognition method based on multi-scale pooling | |
CN108846426A (en) | Polarization SAR classification method based on the twin network of the two-way LSTM of depth | |
CN103942749B (en) | A kind of based on revising cluster hypothesis and the EO-1 hyperion terrain classification method of semi-supervised very fast learning machine | |
CN105825511A (en) | Image background definition detection method based on deep learning | |
CN108681706A (en) | A kind of double source remotely-sensed data semantic segmentation method | |
CN108596203A (en) | Optimization method of the pond layer in parallel to pantograph carbon slide surface abrasion detection model | |
CN109376787B (en) | Manifold learning network and computer vision image set classification method based on manifold learning network | |
CN103177265B (en) | High-definition image classification method based on kernel function Yu sparse coding | |
CN106980831A (en) | Based on self-encoding encoder from affiliation recognition methods | |
CN111709477A (en) | Method and tool for garbage classification based on improved MobileNet network | |
CN106980830A (en) | One kind is based on depth convolutional network from affiliation recognition methods and device | |
CN109685071A (en) | Brain electricity classification method based on the study of common space pattern feature width | |
CN107862329A (en) | A kind of true and false target identification method of Radar range profile's based on depth confidence network | |
CN110287985A (en) | A kind of deep neural network image-recognizing method based on the primary topology with Mutation Particle Swarm Optimizer | |
CN109523514A (en) | To the batch imaging quality assessment method of Inverse Synthetic Aperture Radar ISAR |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
C06 | Publication | ||
PB01 | Publication | ||
C10 | Entry into substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |