CN107194336A - The Classification of Polarimetric SAR Image method of network is measured based on semi-supervised depth distance - Google Patents
The Classification of Polarimetric SAR Image method of network is measured based on semi-supervised depth distance Download PDFInfo
- Publication number
- CN107194336A CN107194336A CN201710336185.7A CN201710336185A CN107194336A CN 107194336 A CN107194336 A CN 107194336A CN 201710336185 A CN201710336185 A CN 201710336185A CN 107194336 A CN107194336 A CN 107194336A
- Authority
- CN
- China
- Prior art keywords
- mrow
- sample
- network
- msub
- msup
- 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
- G06F18/22—Matching criteria, e.g. proximity measures
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Engineering & Computer Science (AREA)
- Software Systems (AREA)
- Artificial Intelligence (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Medical Informatics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Astronomy & Astrophysics (AREA)
- Remote Sensing (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of Classification of Polarimetric SAR Image method that network is measured based on semi-supervised depth distance.Non-linear relation and the not high technical problem of nicety of grading when marked sample is less that conventional depth study only considers sample characteristics are solved, its step includes：Input polarimetric SAR image data to be sorted；Seek neighbour's sample of marked sample；Build the loss function of semi-supervised big Boundary Nearest algorithm；Initialize the parameter of network；To network pre-training；Network is finely tuned；Class prediction is carried out to unmarked sample；Output Classification of Polarimetric SAR Image result figure and nicety of grading to be sorted.The present invention measures network by building depth distance, popular study regular terms is added in big Boundary Nearest algorithm, overcome the information waste problem of marked sample deficiency influence nicety of grading and a large amount of unmarked samples, the feature that the present invention learns fully portrays sample inherent attribute, available for technical fields such as earth resources survey and military systems.
Description
Technical field
The invention belongs to technical field of image processing, more particularly to a kind of Classification of Polarimetric SAR Image method is specifically one kind
The Classification of Polarimetric SAR Image method of network is measured based on semi-supervised depth distance.Available for environmental monitoring, earth resources survey and
Military system etc..
Background technology
Machine learning (Machine Learning, ML) is from artificial intelligence, calculated as the subdomains of computer science
The research of the machine theories of learning and pattern-recognition is set out, and constructing one kind can be carried out from data learning knowledge and to set of metadata of similar data
The algorithm of prediction.Machine learning can learn its each attribute to possess the energy for handling various Similar Problems from initial data
Power, is how to make computer is automatic in empirical learning to obtain the new ability of new knowledge.Led in the Classification and Identification of Polarimetric SAR Image
Domain, machine learning has had many breakthrough progress, some effectively sorting technique be suggested in succession.For example
Wishartmaximumlikelihood (WML), support vector machine (support vector machines, SVM) etc..
Common machine learning method mostly with the artificial method for extracting feature, is wasted time and energy, and can not necessarily take
The feature of Classification of Polarimetric SAR Image must be conducive to.Deep learning is the frontier developed rapidly in machine learning, is in people
A kind of brand-new feature extracting method developed on the basis of artificial neural networks.Deep learning has exclusive feature extraction
Mechanism, can be by the level characteristics of the network structure model of multilayer independently learning data.Avoid conventional machines study side
Legal person's work extracts the cumbersome of feature, while the level characteristics learnt can more express the inherent attribute and characteristic of data.These
Abstract feature can improve the nicety of grading of various classification tasks.For Polarimetric SAR Image, deep learning can be autonomous
Ground is from polarimetric SAR image data learning to the abstraction hierarchy feature for characterizing its inherent attribute, and these features extracted can
It is effectively applied to terrain classification, environmental monitoring, target identification etc. with very convenient.
For Classification of Polarimetric SAR Image task, suitable distance metric selection is highly important, the identification of grader
The selection of distance metric function can be largely dependent upon.Such as most common Euclidean distance, just simply simply considers number
Actual distance between, does not account for the internal structure and specific object of sample, for some data, similar Euclidean distance
Such metric form does not ensure that its reliability, that is, the data closer to the distance tried to achieve be not it is much like be same in other words
Class data, such distance metric result may obtain not ideal result in the presence of the graders such as KNN, SVM.Away from
Purpose from metric learning is that and learnt by immanent structure and attribute to data, obtains a kind of suitable distance degree
Amount mode, under this distance metric mode, the sample spacing with identical category label can be reduced accordingly, without identical
The sample spacing of label can accordingly increase, and obtain a new feature space, so that data become advantageously in classification.
Traditional machine learning method needs the artificial shallow-layer feature extracted feature and obtained to fully demonstrate data
Inherent attribute.And existing deep learning method such as self-encoding encoder, depth confidence net etc. use unsupervised pre-training method, do not have
The sample for having label is instructed, and pre-training effect is not ideal, it is desired nonetheless to a large amount of to have exemplar to carry out reversely network parameter
Propagate fine setting.For there is the methods such as the convolutional neural networks of supervision, when exemplar is less, network performance cannot be ensured,
Classification results are often not ideal.
The content of the invention
It is an object of the invention to the deficiency for above-mentioned prior art, it is proposed that one kind is when exemplar is less
Also the higher Classification of Polarimetric SAR Image method that network is measured based on semi-supervised depth distance of nicety of grading can be obtained, for carrying
Hi-vision nicety of grading.
The present invention is a kind of Classification of Polarimetric SAR Image method that network is measured based on semi-supervised depth distance, and its feature exists
In, including have the following steps：
(1) polarimetric SAR image data to be sorted is inputted：That is the coherence matrix T of Polarimetric SAR Image, according to polarization SAR figure
The atural object distributed intelligence of picture obtains label matrix Y, and the distribution of same atural object is represented by same category label, it is impossible to determine class
Other atural object is distributed in label matrix Y to be represented with 0, and sample matrix is generated according to the coherence matrix T of Polarimetric SAR ImageN is the total number of sample, xiRepresent i-th of sample；
(2) selection marked sample and unmarked sample：According to the sample data X and label matrix Y of Polarimetric SAR Image,
1% sample is randomly selected per class as marked sample, remaining is unmarked sample；
(3) neighbour's sample of marked sample is sought：In all sample datas, the K of each marked sample is asked for1It is individual
There are the mark similar neighbour's samples of Wishart and K2Individual unmarked Wishart neighbours sample；
(4) loss function of semi-supervised big Boundary Nearest algorithm is built：On the basis of big Boundary Nearest algorithm (LMNN)
The popular study regular terms of increase, is semi-supervised big Boundary Nearest algorithm by the big Boundary Nearest algorithm improvement for having supervision, obtains
The loss function of semi-supervised big Boundary Nearest algorithm；
(5) basic parameter of depth network is initialized：Random initializtion depth distance measurement network weight parameter W and partially
Unit b is put, the nodes of every layer of set depth distance metric network determine that depth distance measures the overall structure of network；
(6) network pre-training is measured to depth distance：Marked sample and its corresponding Wishart neighbours sample are sent into
Pre-training is carried out into depth distance measurement network, using semi-supervised big Boundary Nearest algorithm, using successively greedy pre-training
Method, the output of preceding layer, until having trained last hidden layer, optimizes the weight of network, depth as the input of later layer
The weight of distance metric network is tentatively optimized；
(7) to depth distance measurement network fine setting：Using having exemplar and its label information, with reference to Softmax classification
Device to depth distance measurement network be finely adjusted, further optimization network weight, network is become more stable, complete depth away from
From the power optimization of measurement network；
(8) class prediction is carried out to unmarked sample：Unmarked sample is sent to depth distance measurement network, utilized
Softmax graders are predicted to the class label of unmarked sample, obtain the prediction classification of each unmarked sample；
(9) the classification results figure and nicety of grading of Polarimetric SAR Image to be sorted are exported：According to marked sample and prediction
Go out the unmarked sample of classification, export the final classification result of Polarimetric SAR Image to be sorted and calculate the precision of this subseries.
Big Boundary Nearest algorithm is combined by the present invention with deep learning method, it is proposed that it is a kind of be based on semi-supervised depth away from
From the Classification of Polarimetric SAR Image method of measurement network.Avoid conventional machines learning method artificial using the method for deep learning
Extract the cumbersome of feature.
The invention has the advantages that：
1st, the present invention measures network independently to learn polarization as a result of the method for deep learning using depth distance
The feature of SAR image, thus the cumbersome of artificial learning characteristic in conventional method is avoided, utilize depth distance measurement e-learning
To the inherent level characteristics of Polarimetric SAR Image more horn of plenty, these features are advantageously in the classification of Polarimetric SAR Image, effectively
Improve the nicety of grading of Polarimetric SAR Image.
2nd, the present invention has obtained semi-supervised big on the basis of big Boundary Nearest algorithm by adding popular study regular terms
Boundary Nearest algorithm, improves the grader classification accuracy rate in the case where marker samples are less relatively low and a large amount of unmarked
The problem of information waste that sample is caused.
3rd, semi-supervised distance metric method is combined by the present invention with deep learning method, has obtained depth distance measurement net
Network, this depth network can simultaneously learning sample linear and nonlinear characteristic, fully feature the inherent attribute of sample, thus
The feature learnt can significantly improve nicety of grading.
Brief description of the drawings
Fig. 1 is the implementation process schematic diagram of the present invention；
Fig. 2 is the experimental result picture of the Polarimetric SAR Image to Dutch Flevoland areas, and wherein Fig. 2 (a) is polarization SAR
The Pauli exploded views of image, Fig. 2 (b) is label figure, and Fig. 2 (c) is the classification results figure using control methods WDSN, Fig. 2 (d)
It is the classification results figure using control methods WDBN, Fig. 2 (e) is the classification results figure using control methods DSFN, and Fig. 2 (f) is
Using the classification results figure of the inventive method；
Fig. 3 is the schematic diagram of semi-supervised big Boundary Nearest algorithm.
Embodiment
Below in conjunction with the accompanying drawings to the detailed description of the invention：
Embodiment 1
Because of the development of remote sensing technology, widely should have in fields such as environmental monitoring, earth resources survey, military systems
With the demand to Polarimetric SAR Image processing is also continued to increase, and deep learning has more obvious excellent in machine learning method
Gesture.It is unsupervised when marker samples are less and traditional deep learning network is generally learning method that is unsupervised or having supervision
If deep learning method rely only on unsupervised pre-training and be likely to occur poor fitting situation, and a small amount of marker samples fine setting
Network performance can not effectively be lifted；It is even more to occur that network training is insufficient for the deep learning method for having supervision, net
The situation of network poor-performing.Traditional deep learning method is without linear processes feature simultaneously in view of sample, study
Obtained feature can not fully demonstrate the inherent attribute of sample.Research and innovation are expanded for these present situations present invention, is proposed
A kind of Classification of Polarimetric SAR Image method that network is measured based on semi-supervised depth distance, referring to Fig. 1, including is had the following steps：
(1) polarimetric SAR image data to be sorted is inputted：That is the coherence matrix T of Polarimetric SAR Image, according to polarization SAR figure
The atural object distributed intelligence of picture obtains label matrix Y, and the distribution of same atural object is represented by same category label, it is impossible to determine class
Other atural object is distributed in label matrix Y to be represented with 0, and sample matrix is generated according to the coherence matrix T of Polarimetric SAR ImageN is the total number of sample, xiRepresent i-th of sample.
(2) selection marked sample and unmarked sample：According to the sample data X and label matrix Y of Polarimetric SAR Image,
1% sample is randomly selected per class as marked sample, remaining is unmarked sample.
(3) neighbour's sample of marked sample is sought：In all sample datas, the K of each marked sample is asked for1It is individual
There are the mark similar neighbour's samples of Wishart and K2Individual unmarked Wishart neighbours sample, each marked sample has K1+K2It is individual
Wishart neighbour's samples, so far, basic data processing are completed.
(4) loss function of semi-supervised big Boundary Nearest algorithm is built：Increase stream on the basis of big Boundary Nearest algorithm
Row study regular terms, is semi-supervised big Boundary Nearest algorithm by the big Boundary Nearest algorithm improvement for having supervision, obtains semi-supervised
The loss function of big Boundary Nearest algorithm, semi-supervised big Boundary Nearest algorithm is effectively utilized part unmarked sample, reduces
To the demand of marked sample so that algorithm also ensure that performance in the case of only a small amount of marked sample.
(5) basic parameter of depth network is initialized：Random initializtion depth distance measurement network weight parameter W and partially
Unit b is put, the nodes of every layer of set depth distance metric network determine that depth distance measures the overall structure of network.
(6) network pre-training is measured to depth distance：Marked sample and its corresponding Wishart neighbours sample are sent into
Pre-training is carried out into depth distance measurement network.Using semi-supervised big Boundary Nearest algorithm, using successively greedy pre-training
Method, the output of preceding layer, until having trained last hidden layer, optimizes the weight of network, depth as the input of later layer
The weight of distance metric network is tentatively optimized.
(7) to depth distance measurement network fine setting：Using having exemplar and its label information, with reference to Softmax classification
Device is finely adjusted to depth distance measurement network, and further the weight of optimization network, makes network become more stable, so far, completes
Depth distance measures the network optimization.
(8) class prediction is carried out to unmarked sample：Unmarked sample is sent to depth distance measurement network, utilized
Softmax graders are predicted to the class label of unmarked sample, obtain the prediction classification of each unmarked sample.
(9) the classification results figure and nicety of grading of Polarimetric SAR Image to be sorted are exported：According to marked sample and step
(8) it is predicted that going out the unmarked sample of classification in, export the final classification result of Polarimetric SAR Image to be sorted and calculate this
The precision of classification.
The present invention technical thought be：Big Boundary Nearest algorithm is obtained into semi-supervised big side in the presence of popular regular terms
Boundary's nearest neighbor algorithm, is combined with deep learning method and obtains depth distance measurement network, possessing the situation of a small amount of marker samples
Under, feature learning is carried out to polarization SAR data, image classification is realized by grader, nicety of grading is improved.
Embodiment 2
Measured based on semi-supervised depth distance in the Classification of Polarimetric SAR Image method be the same as Example 1 of network, step (3)
Wishart neighbour's samples of each marked sample are sought, including are had the following steps：
3a, marked sample matrix are Marked sample number is represented, is asked for often using following formula
Wishart distances between individual marked sample and remaining sample：
d(xi,xj)=ln ((xi)-1xj)+Tr((xj)-1xi)-q,
Wherein, the mark of Tr () representing matrix, is integral radar for sending and receiving, due to reciprocity, constant q=3；
It is not integral radar, constant q=4 for sending and receiving；
3b, using the sort functions in MATLAB, by the Wishart tried to achieve in 3a apart from d (xi,xj) press absolute value ascending order
Arrangement, takes preceding K1Individual markd similar neighbour's sample xj(j=1,2, K1), K2Individual unmarked neighbour's sample xp(p
=1,2, K2), it is used as marked sample xiWishart neighbour's samples.
Embodiment 3
Measured based on semi-supervised depth distance in the Classification of Polarimetric SAR Image method be the same as Example 1-2 of network, step (4)
The process of the loss function of the described semi-supervised big Boundary Nearest algorithm of structure includes：
4a, the loss function for asking for big Boundary Nearest algorithm：
The square distance formula of big Boundary Nearest algorithm is：
Wherein, L is linear change matrix, xjIt is xiSimilar marked sample.If there is marked sample xiIt is non-same
The x of class samplel, meet equation below：
||L(xi-xl)||2≤||L(xi-xj)||2+ 1,
Then, xlIt is referred to as " jactitator ".
Big Boundary Nearest algorithm can be expressed as two parts：Loss function ε between similar samplepull(L) it is and non-similar
Loss function ε between samplepush(L), εpull(L) it is big between marked sample and its generic neighbour for punishing
Spacing, that is, reduce the distance between similar sample；εpush(L) it is for punishing between small between marked sample and " jactitator "
Away from increasing the distance between non-similar sample：
Wherein, symbol yilRepresent marked sample xiWith marked sample xlClass relations, yil=1 and if only if yi=
ylShi Chengli, i.e. xiAnd xlFor similar sample；Otherwise, yil=0.[z]+=max (z, 0) is the hinge function of standard.
So, the loss function ε of big Boundary Nearest algorithm1(L) it is：
ε1(L)=(1- μ) εpull(L)+μεpush(L),
Wherein, loss parameter μ ∈ [0,1], makes symbol ξijl=[1+ | | L (xi-xj)||2-||L(xi-xl)||2]+, then：
4b, the loss function for asking for semi-supervised big Boundary Nearest algorithm：
Popular study regular terms is added in the loss function of big Boundary Nearest algorithm, the use of regular terms is added to portion
Divide the utilization of unmarked sample, be semi-supervised learning method by the big Boundary Nearest algorithm improvement for having supervision, add the epidemiology
Practise regular terms JRPurpose be to punish marked sample xiWith unmarked sample xpBetween large-spacing, that is, reduce have mark
Remember sample xiWith unmarked sample xpBetween interval：
Wherein, symbolRepresent marked sample xiWith its unmarked neighbour's sample xpBetween
The loss function ε of similarity, then semi-supervised big Boundary Nearest algorithm2(L) it is：
Wherein, | | | |FIt is Frobenius norms, for ensureing maximum boundary, λ is Frobenius norm canonical term systems
Number, it is semi-supervised regular terms parameter generally to take λ=1, γ.
The present invention has obtained semi-supervised big side on the basis of big Boundary Nearest algorithm by adding popular study regular terms
Boundary's nearest neighbor algorithm, it is relatively low to improve the grader classification accuracy rate in the case where marker samples are less, and a large amount of unmarked samples
Originally the problem of information waste caused, semi-supervised big Boundary Nearest algorithm can be as shown in Figure 3 to the action principle of sample.
Embodiment 4
Measured based on semi-supervised depth distance in the Classification of Polarimetric SAR Image method be the same as Example 1-3 of network, step (5)
It is described initialization depth network parameter be：
5a, the hidden layer number of depth distance measurement network are 3, and every layer of nodes are respectively：150,100,50；
5b, random initializtion depth sparseness filtering network weight parameter W and bias unit b, the nodes of each hidden layer
It is divided into Nk,T is the dimension of input signal, N1It is the nodes of first hidden layer of network,Represent
Weight matrix between -1 hidden layer of k-th of hidden layer and kth of network,For the bias unit of k-th of hidden layer.
Embodiment 5
Measured based on semi-supervised depth distance in the Classification of Polarimetric SAR Image method be the same as Example 1-4 of network, step (6)
To depth distance measure network pre-training process be：
6a, the pre-training sample for inputting depth sparseness filtering network, by unmarked sample xiAnd its it is corresponding have mark and
Unmarked Wishart neighbours sample is input in depth network as pre-training sample,
6b, set xi∈Rt×1It is input vector, i=1,2 ..., n.N represents input sample number, then first hidden layer
Output can be expressed as：
Wherein, s () represents nonlinear sigmoid functions, makes z=W1xi+b1, then s (z)=(1+exp (- z))-1, willSecond hidden layer of network is input to, second hidden layer is obtained and is output as：
Successively greedy training is gone down successively, then k-th of hidden layer is output as：
6c, optimization network weight W, using big Boundary Nearest algorithm, linear transformation matrix L are equivalent to the power of depth network
W is weighed, the optimization object function of k-th of hidden layer is：
Wherein,Whole optimization object function can pass through
The methods such as traditional gradient descent algorithm, L-BFGS are solved.
The present invention carries out depth network pre-training by semi-supervised big Boundary Nearest algorithm, utilizes learning distance metric
Method cause sample to carry out eigentransformation, the distance between similar sample is reduced, the increase of the distance between non-similar sample,
Depth distance measure network by learning distance metric and deep learning come and meanwhile learning sample linear processes feature, most
The feature obtained eventually is conducive to improving the nicety of grading of Polarimetric SAR Image.
A more detailed example is given below, the present invention is further described:
Embodiment 6
The Classification of Polarimetric SAR Image method be the same as Example 1-5 of network, reference picture 1, sheet are measured based on semi-supervised depth distance
The specific implementation step of invention is as follows：
Step 1, input polarimetric SAR image data to be sorted, referring to Fig. 2 (a), Fig. 2 (a) is Dutch Flevoland
The Pauli exploded views of the Polarimetric SAR Image in area, input the coherence matrix T of the Polarimetric SAR Image, according to the ground of Polarimetric SAR Image
Thing distributed intelligence obtains label matrix Y, referring to Fig. 2 (b), the image that Fig. 2 (b) is exactly directly generated by label matrix Y, in image
Different color lumps represents different atural objects, and being distributed in label matrix for same atural object is represented by same category label, no
It can determine that the atural object of classification is distributed in label matrix to be represented with 0, sample moment generated according to the coherence matrix T of Polarimetric SAR Image
Battle arrayN is the total number of sample, xiRepresent i-th of sample.
This example uses Polarimetric SAR Image regional Holland Flevoland.The image is by NASA/JPLAIRSAR systems
Obtain, image size is 300 × 270, contain 6 kinds of different atural object classifications in image.Due to the polarization phase of polarization SAR data
Dry matrix T is Hamilton positive semidefinite matrix, it is possible to which it is 3 × 3 polarization coherence matrix T upper angular position to extract dimension
The modulus value of 6 elements is turned T matrixes using the reshape functions in MATLAB softwares as the primitive character of each pixel
The sample matrix of two dimension is turned to, each row represent a sample, and the dimension of each sample is 6 dimensions.Each sample correspondence polarization SAR
A pixel on image.
Step 2, selection marked sample and unmarked sample, according to the sample data X and label matrix of Polarimetric SAR Image
Y, sample data category is extracted at randomIndividual marked sample, remaining is unmarked sample, the marked sample per class
Account for the 1% of such sum.
Step 3, the neighbour's sample for seeking marked sample, in all sample datas, ask for each marked sample
K1It is individual to have the mark similar neighbour's samples of Wishart and K2Individual unmarked Wishart neighbours sample.
3a, marked sample matrix areAccording to improved Wishart range formulas, it and other samples are asked
Between Wishart distances：d(xi,xj)=ln ((xi)-1xj)+Tr((xj)-1xi)-q,
Wherein, the mark of Tr () representing matrix, is integral radar for sending and receiving, due to reciprocity, constant q=3,
It is not integral radar, constant q=4, the regional polarization SAR figures of Dutch Flevoland that this example is used for sending and receiving
As data are that, by transmission and the radar system acquisition generation for receiving one, q takes 3.
3b, using the sort functions in MATLAB, by the Wishart tried to achieve in 3a apart from d (xi,xj) press absolute value ascending order
Arrangement, takes preceding K1Individual markd similar neighbour's sample xj(j=1,2, K1), K2Individual unmarked neighbour's sample xp(p
=1,2, K2), it is used as marked sample xiWishart neighbour's samples.
Because polarization SAR data obey Wishart distributions, so marked sample xiUnmarked Wishart neighbours sample
This xpIt is similar sample to be likely to, by x in the semi-supervised big Boundary Nearest algorithm of the present inventioniAnd xpThe distance between carry out
Punishment, can effectively utilize part unmarked sample, reduce the demand to marked sample.
Step 4, the loss function for building semi-supervised big Boundary Nearest algorithm, increase on the basis of big Boundary Nearest algorithm
Prevalence study regular terms, is semi-supervised big Boundary Nearest algorithm by the big Boundary Nearest algorithm improvement for having supervision, obtains half prison
Superintend and direct the loss function of big Boundary Nearest algorithm.
4a, the loss function ε for asking for big Boundary Nearest algorithm first1(L)：
The square distance formula of big Boundary Nearest algorithm is：
Wherein, L is linear change matrix, xjIt is xiSimilar marked sample.If there is marked sample xiIt is non-same
The x of class samplel, meet equation below：
||L(xi-xl)||2≤||L(xi-xj)||2+ 1,
Then, xlIt is referred to as " jactitator ".
Big Boundary Nearest algorithm can be expressed as two parts：Loss function ε between similar samplepull(L) it is and non-similar
Loss function ε between samplepush(L), εpull(L) it is big between marked sample and its generic neighbour for punishing
Spacing, that is, reduce the distance between similar sample；εpush(L) it is for punishing between small between marked sample and " jactitator "
Away from increasing the distance between non-similar sample：
So, the loss function ε of big Boundary Nearest algorithm1(L) it is：
ε1(L)=(1- μ) εpull(L)+μεpush(L),
Wherein, loss parameter μ ∈ [0,1], makes symbol ξijl=[1+ | | L (xi-xj)||2-||L(xi-xl)||2]+, then
To the loss function ε of big Boundary Nearest algorithm1(L)：
4b, the loss function ε for asking for semi-supervised big Boundary Nearest algorithm2(L)：
Popular study regular terms is added in the loss function of big Boundary Nearest algorithm, the use of regular terms is added to portion
Divide the utilization of unmarked sample, be semi-supervised learning method by the big Boundary Nearest algorithm improvement for having supervision, add the epidemiology
Practise regular terms JRPurpose be to punish marked sample xiWith unmarked sample xpBetween large-spacing, that is, reduce have mark
Remember sample xiWith unmarked sample xpBetween interval：
Wherein, symbolRepresent marked sample xiWith its unmarked neighbour's sample xpBetween
The loss function ε of similarity, then semi-supervised big Boundary Nearest algorithm2(L) it is：
Wherein, | | | |FIt is Frobenius norms, for ensureing maximum boundary, λ is Frobenius norm canonical term systems
Number, it is semi-supervised regular terms parameter generally to take λ=1, γ.
Marked sample and part unmarked sample are combined by semi-supervised big Boundary Nearest algorithm, have overcome mark sample
Algorithm performance relatively low situation when this is not enough.
Step 5, the basic parameter for initializing depth network：Random initializtion depth distance measures the weight parameter W of network
With bias unit b, the nodes of every layer of set depth distance metric network determine that depth distance measures the overall structure of network；
5a, the hidden layer number of depth distance measurement network are 3, and every layer of nodes are respectively：150,100,50；
5b, random initializtion depth sparseness filtering network weight parameter W and bias unit b, the nodes of each hidden layer
It is divided into Nk,T is the dimension of input signal, N1It is the nodes of first hidden layer of network,Represent
Weight matrix between -1 hidden layer of k-th of hidden layer and kth of network,For the bias unit of k-th of hidden layer.
Step 6, to depth distance measure network pre-training.
6a, the pre-training sample for inputting depth sparseness filtering network, by unmarked sample xiAnd its it is corresponding have mark and
Unmarked Wishart neighbours sample is input in depth network as pre-training sample,
6b, set xi∈Rt×1It is input vector, i=1,2 ..., n.N represents input sample number, then first hidden layer
Output can be expressed as：
Wherein, s () represents nonlinear sigmoid functions, makes z=W1xi+b1, then s (z)=(1+exp (- z))-1, willSecond hidden layer of network is input to, second hidden layer is obtained and is output as：
Successively greedy training is gone down successively, then k-th of hidden layer is output as：
6c, optimization network weight W, using big Boundary Nearest algorithm, linear transformation matrix L are equivalent to the power of depth network
W is weighed, the optimization object function of k-th of hidden layer is：
Wherein,Whole optimization aim passes through traditional ladder
Descent algorithm is spent to solve.
The present invention carries out depth network pre-training by semi-supervised big Boundary Nearest algorithm, according to the thought of metric learning
Initial data is mapped to new feature space, in new feature space, the distance between similar sample is reduced, non-similar sample
The distance between increase.Network can be while the linear and nonlinear characteristic of learning sample data, the feature learnt can
The classification of Polarimetric SAR Image is efficiently used for, nicety of grading is improved.
Step 7, to depth distance measurement network fine setting.
After the completion of pre-training, the weight W of network has converged to rational scope, further utilizes marked sample and its
Correspondence class label information, is finely adjusted with reference to Softmax graders to network, and the object function Φ (W) for finely tuning part can be with
It is expressed as：
Wherein, previous item is mean square error, and latter is weight attenuation term, and the purpose of weight attenuation term is to reduce power
The amplitude of weight, prevents over-fitting, yiRepresent training sample xiCorresponding class label, y (xi) it is by marked sample xiDepth away from
From obtained prediction classification after measurement network, β=3e-3 is weight attenuation parameter, Φ (W) can according to gradient descent algorithm come
Solve.
Step 8, to unmarked sample carry out class prediction, by unmarked sample be sent to depth distance measurement network, profit
The class label of unmarked sample is predicted with Softmax graders, the prediction classification of each unmarked sample is obtained.
7a, by all unmarked sample xuIt is input in the depth distance measurement network having had been built up, obtains final
Excitation output characteristicThe namely output characteristic of the 3rd hidden layer of network；
7b, by 7a learnings to feature be sent in Softmax graders and carry out class prediction：
The output of Softmax graders is y ∈ RP×1, P is expressed as classification number, unmarked sample xuPrediction classification can be with
It is expressed as：
Wherein α is sample xuPrediction classification confidence level.
Step 9, the classification results figure of output Polarimetric SAR Image to be sorted and nicety of grading, according to marked sample and
It is predicted that going out the unmarked sample of classification, export the final classification result of Polarimetric SAR Image to be sorted and calculate this subseries
Precision.
9a, the class label according to marked sample and unmarked sample prediction class label, by red, green, indigo plant
Color is painted as three primary colours according to color method in three primary colours for each pixel, output result figure, referring to Fig. 2 (f).
9b, the prediction category label of unmarked sample and its real category label contrasted, referring to table 1, drawn
Nicety of grading.
Depth distance measurement network of the present invention can be while the linear and nonlinear characteristic of learning sample, improves polarization
The precision of SAR image classification.
The technique effect of the present invention is described in detail again with reference to emulation experiment：
Embodiment 7
The Classification of Polarimetric SAR Image method be the same as Example 1-6 of network is measured based on semi-supervised depth distance,
Experiment condition：
Hardware platform is：Intel(R)Core(TM)i5-2410M [email protected]、RAM 4.00GB；
Software platform is：MATLAB R2016a；
Experiment is tested from the Polarimetric SAR Image in Holland Flevoland areas, and image size is 300 × 270, often
The dimension of individual sample point is 6 dimensions, and classification number is 6.In experiment, 1% sample is randomly selected per class as marked sample, remaining
For unmarked sample.
Experiment content and result：
The present invention classifies with reference to Softmax graders to real Polarimetric SAR Image, before same Setup Experiments
Put and be compared with other deep learning methods, wherein WDSN is a kind of faster depth network model of training speed, Fig. 2
(c) it is the result figure classified by WDSN to Fig. 2 (a)；Second control methods is WDBN, is by Wishart distance applications
In traditional depth confidence net, the result figure that Fig. 2 (d) is classified with WDBN methods to Fig. 2 (a)；DSFN is that depth is sparse
Filter network, is that sparseness filtering expand to a kind of obtained depth network model, Fig. 2 (e) is that DSFN is carried out to Fig. 2 (a)
The result figure of classification；Fig. 2 (f) is the result figure that the inventive method SDMLN is classified to Fig. 2 (a).Table 1 is to use above-mentioned 4 kinds
The terrain classification precision and overall classification accuracy for the Polarimetric SAR Image that method is respectively obtained.
The terrain classification precision (%) and overall classification accuracy (%) of table 1, various methods on Flevoland plats
From table 1 it follows that in the case where marked sample is 1%, the present invention Bare soil, potato,
Wheat and barley classes suffer from highest nicety of grading, and total nicety of grading also will than other control methods for 97.35%
It is high.Present invention employs the method for distance metric, for being distributed more carefully and neatly done block data, there is preferably classification essence
Degree, this method can be with the linear and nonlinear characteristic of learning sample data, and keeps the neighbor relationships of neighbour's sample, and these are all
The present invention has been promoted to have higher classification accuracy rate.
Embodiment 8
Based on semi-supervised depth distance measure network Classification of Polarimetric SAR Image method be the same as Example 1-6, simulated conditions and
Emulation content be the same as Example 7,
Referring to the simulation experiment result Fig. 2, Fig. 2 (c) to Fig. 2 (f) is obtained in the case of only 1% marked sample
The classification results figure obtained, by contrast, the present invention have preferably visual.Each method and label figure are contrasted, can be obvious
Find out, preferably, Fig. 2 (f) possesses less noise to the final classification results of the present invention, each noise represents classification error
Sample point, illustrates that the inventive method has this higher nicety of grading, from table 1 it is also seen that.Fig. 2 (c) is WDSN experiment knot
Fruit is schemed, and the algorithm focuses on training speed, thus occurs in that more noise in the case of less marked sample, is presented
Go out poor classification results.Fig. 2 (d) and Fig. 2 (e) is the depth network model of current more novelty, has mark sample only 1%
In the case of this, the visuality of classification results is also intended to be significantly lower than the inventive method.Illustrate that the inventive method is utilized semi-supervised
Method, distance metric is combined with deep learning method, and the feature learnt is conducive to classifying to Polarimetric SAR Image,
In the case where marked sample is less, raising nicety of grading that also can be relative.
Referring to Fig. 3, Fig. 3 is the schematic diagram of the semi-supervised big Boundary Nearest algorithm of the present invention.Triangle and circle point in figure
Two kinds of different classes of samples are not represented, wherein marked sample and part unmarked sample are included in the sample of each classification,
Marked sample represents that unmarked sample is represented with white with colour, neighbor relationships each other between similar sample in figure.The present invention
Semi-supervised big Boundary Nearest algorithm causes sample data from one under the collective effect of distance metric and popular study regular terms
Feature space is mapped to new feature space.In mapping process, distance metric method causes markd similar neighbour's sample
The distance between reduce, the increase of the distance between non-similar marked sample.And popular study regular terms causes marked sample
Reduce with the distance between the unmarked neighbour's sample in its part, take full advantage of the unmarked information in part, it is to avoid information
Waste and reduce demand of the algorithm to marked sample.Semi-supervised big Boundary Nearest algorithm most at last similar sample with it is non-same
Class sample is suitably separated, and the sample after mapping is easier to make for classification in the presence of grader, can effectively improve point
Class precision.
It is proposed by the present invention that the Classification of Polarimetric SAR Image method of network is measured by distance metric based on semi-supervised depth distance
Study and deep learning method are combined, and can efficiently extract the linear and nonlinear characteristic of sample data, and the present invention improves
The problem of conventional deep learning method is larger to marked sample demand, in the case of the only less marked sample,
The nicety of grading of Polarimetric SAR Image can be properly increased.
In summary, a kind of Classification of Polarimetric SAR Image that network is measured based on semi-supervised depth distance disclosed by the invention
Method.Solve non-linear relation and the nicety of grading when marked sample is less that conventional depth study only considers sample characteristics
Not high technical problem, its step includes：1st, polarimetric SAR image data to be sorted, the coherence matrix of Polarimetric SAR Image are inputted
T generation sample matrix X；2nd, selection marked sample and unmarked sample, according to the sample data X and label of Polarimetric SAR Image
Matrix Y, 1% sample is randomly selected per class as marked sample, remaining is unmarked sample；3rd, marked sample is sought
Neighbour's sample, in all sample datas, asks for the K of each marked sample1It is individual to have the mark similar neighbour's samples of Wishart
And K2Individual unmarked Wishart neighbours sample；4th, the loss function of semi-supervised big Boundary Nearest algorithm is built, in big Boundary Nearest
Increase popular study regular terms on the basis of algorithm, the big Boundary Nearest algorithm improvement for having supervision is near for semi-supervised big border
Adjacent algorithm, obtains the loss function of semi-supervised big Boundary Nearest algorithm；5th, initialization depth distance measures the parameter of network, at random
Initialize weight parameter W and bias unit b that depth distance measures network, the node of every layer of set depth distance metric network
Number；6th, network pre-training is measured to depth distance, marked sample and its corresponding Wishart neighbours sample is sent to depth
Pre-training is carried out in distance metric network, it is preceding using successively greedy pre-training method using semi-supervised big Boundary Nearest algorithm
One layer of output, until having trained last hidden layer, optimizes the weight of network as the input of later layer；7th, to depth distance
Network fine setting is measured, with having exemplar and its label information, depth distance measurement network is carried out with reference to Softmax graders
The weight of fine setting, further optimization network；8th, class prediction is carried out to unmarked sample；Output Polarimetric SAR Image to be sorted
Classification results figure and nicety of grading.
The present invention measures network model by building depth distance, and popular study is being added in big Boundary Nearest algorithm just
Then the method for item, has obtained semi-supervised big Boundary Nearest algorithm, and the distance degree of profound level is carried out using deep learning method
Amount study, can effectively describe the linear and nonlinear organization of sample, can make finally to train obtained depth network to have
Than conventional depth network more outstanding feature extraction performance, the depth network extraction to feature be content with very little between similar sample
Property away from less than non-similar sample spacing increase, thus final classification essence can be effectively improved in the presence of grader
Degree.The information waste problem of nicety of grading and a large amount of unmarked samples is influenceed instant invention overcomes marked sample deficiency, is learned
The feature practised fully features sample inherent attribute, available for technologies such as environmental monitoring, earth resources survey and military systems
Field.
Claims (5)
1. it is a kind of based on semi-supervised depth distance measure network Classification of Polarimetric SAR Image method, it is characterised in that including just like
Lower step：
(1) polarimetric SAR image data to be sorted is inputted：That is the coherence matrix T of Polarimetric SAR Image, according to Polarimetric SAR Image
Atural object distributed intelligence obtains label matrix Y, and the distribution of same atural object is represented by same category label, it is impossible to determine classification
Atural object is distributed in label matrix Y to be represented with 0, and sample matrix is generated according to the coherence matrix T of Polarimetric SAR ImageN
It is the total number of sample, xiRepresent i-th of sample；
(2) selection marked sample and unmarked sample：According to the sample data X and label matrix Y of Polarimetric SAR Image, per class
1% sample is randomly selected as marked sample, remaining is unmarked sample；
(3) neighbour's sample of marked sample is sought：In all sample datas, the K of each marked sample is asked for1It is individual to have mark
Remember the similar neighbour's samples of Wishart and K2Individual unmarked Wishart neighbours sample；
(4) loss function of semi-supervised big Boundary Nearest algorithm is built：Increase epidemiology on the basis of big Boundary Nearest algorithm
Regular terms is practised, is semi-supervised big Boundary Nearest algorithm by the big Boundary Nearest algorithm improvement for having supervision, obtains semi-supervised big side
The loss function of boundary's nearest neighbor algorithm；
(5) basic parameter of depth network is initialized：The weight parameter W of random initializtion depth distance measurement network and biasing are single
First b, the nodes of every layer of set depth distance metric network determine that depth distance measures the overall structure of network；
(6) network pre-training is measured to depth distance：Marked sample and its corresponding Wishart neighbours sample are sent to depth
Pre-training is carried out in degree distance metric network, using semi-supervised big Boundary Nearest algorithm, using successively greedy pre-training method,
The output of preceding layer, until having trained last hidden layer, optimizes the weight of network, depth distance degree as the input of later layer
The weight of amount network is tentatively optimized；
(7) to depth distance measurement network fine setting：Using having exemplar and its label information, with reference to Softmax graders pair
Depth distance measurement network is finely adjusted, and further the weight of optimization network, makes network become more stable, completes depth distance degree
Measure the network optimization；
(8) class prediction is carried out to unmarked sample：Unmarked sample is sent to depth distance measurement network, utilized
Softmax graders are predicted to the class label of unmarked sample, obtain the prediction classification of each unmarked sample；
(9) the classification results figure and nicety of grading of Polarimetric SAR Image to be sorted are exported：According to marked sample and predicting class
Other unmarked sample, exports the final classification result of Polarimetric SAR Image to be sorted and calculates the precision of this subseries.
2. the Classification of Polarimetric SAR Image method according to claim 1 that network is measured based on semi-supervised depth distance, it is special
Levy and be, seek Wishart neighbour's samples of each marked sample described in step (3), including have the following steps：
3a, marked sample matrix are Marked sample number is represented, asks for each having using following formula
Wishart distances between marker samples and remaining sample：
d(xi,xj)=ln ((xi)-1xj)+Tr((xj)-1xi)-q,
Wherein, the mark of Tr () representing matrix, is integral radar for sending and receiving, due to reciprocity, constant q=3；For
It is not integral radar, constant q=4 to send and receive；
3b, using the sort functions in MATLAB, by the Wishart tried to achieve apart from d (xi,xj) by the arrangement of absolute value ascending order, before taking
K1Individual markd similar neighbour's sample xj(j=1,2 ..., K1), K2Individual unmarked neighbour's sample xp(p=1,2 ..., K2),
It is used as marked sample xiWishart neighbour's samples.
3. the Classification of Polarimetric SAR Image method according to claim 1 that network is measured based on semi-supervised depth distance, it is special
Levy and be, the process of the loss function of the semi-supervised big Boundary Nearest algorithm of structure described in step (4) includes：
Popular study regular terms is added in the loss function of big Boundary Nearest algorithm, the use of regular terms is added to part nothing
The utilization of marker samples, is semi-supervised learning method, prevalence study regular terms by the big Boundary Nearest algorithm improvement for having supervision
JRPurpose be to punish marked sample xiWith unmarked sample xpBetween large-spacing, that is, reduce marked sample xi
With unmarked sample xpBetween interval：
<mrow>
<msub>
<mi>J</mi>
<mi>R</mi>
</msub>
<mo>=</mo>
<msub>
<mi>&theta;</mi>
<mrow>
<mi>i</mi>
<mi>p</mi>
</mrow>
</msub>
<munder>
<mo>&Sigma;</mo>
<mrow>
<mi>i</mi>
<mo>,</mo>
<mi>p</mi>
</mrow>
</munder>
<mo>|</mo>
<mo>|</mo>
<mi>f</mi>
<mrow>
<mo>(</mo>
<msub>
<mi>x</mi>
<mi>i</mi>
</msub>
<mo>)</mo>
</mrow>
<mo>-</mo>
<mi>f</mi>
<mrow>
<mo>(</mo>
<msub>
<mi>x</mi>
<mi>p</mi>
</msub>
<mo>)</mo>
</mrow>
<mo>|</mo>
<msup>
<mo>|</mo>
<mn>2</mn>
</msup>
<mo>=</mo>
<msub>
<mi>&theta;</mi>
<mrow>
<mi>i</mi>
<mi>p</mi>
</mrow>
</msub>
<munder>
<mo>&Sigma;</mo>
<mrow>
<mi>i</mi>
<mo>,</mo>
<mi>p</mi>
</mrow>
</munder>
<mo>|</mo>
<mo>|</mo>
<mi>L</mi>
<mrow>
<mo>(</mo>
<msub>
<mi>x</mi>
<mi>i</mi>
</msub>
<mo>-</mo>
<msub>
<mi>x</mi>
<mi>p</mi>
</msub>
<mo>)</mo>
</mrow>
<mo>|</mo>
<msup>
<mo>|</mo>
<mn>2</mn>
</msup>
<mo>,</mo>
</mrow>
Wherein, symbolRepresent marked sample xiWith its unmarked neighbour's sample xpBetween it is similar
Spend, then the loss function ε of semi-supervised big Boundary Nearest algorithm2(L) it is：
<mrow>
<msub>
<mi>&epsiv;</mi>
<mn>2</mn>
</msub>
<mrow>
<mo>(</mo>
<mi>L</mi>
<mo>)</mo>
</mrow>
<mo>=</mo>
<mrow>
<mo>(</mo>
<mn>1</mn>
<mo>-</mo>
<mi>&mu;</mi>
<mo>)</mo>
</mrow>
<munder>
<mo>&Sigma;</mo>
<mrow>
<mi>i</mi>
<mo>,</mo>
<mi>j</mi>
<mo>&RightArrow;</mo>
<mi>i</mi>
</mrow>
</munder>
<mo>|</mo>
<mo>|</mo>
<mi>L</mi>
<mrow>
<mo>(</mo>
<msub>
<mi>x</mi>
<mi>i</mi>
</msub>
<mo>-</mo>
<msub>
<mi>x</mi>
<mi>j</mi>
</msub>
<mo>)</mo>
</mrow>
<mo>|</mo>
<msup>
<mo>|</mo>
<mn>2</mn>
</msup>
<mo>+</mo>
<mi>&mu;</mi>
<munder>
<mo>&Sigma;</mo>
<mrow>
<mi>i</mi>
<mo>,</mo>
<mi>j</mi>
<mo>&RightArrow;</mo>
<mi>i</mi>
</mrow>
</munder>
<munder>
<mo>&Sigma;</mo>
<mi>l</mi>
</munder>
<mrow>
<mo>(</mo>
<mn>1</mn>
<mo>-</mo>
<msub>
<mi>y</mi>
<mrow>
<mi>i</mi>
<mi>l</mi>
</mrow>
</msub>
<mo>)</mo>
</mrow>
<msub>
<mi>&xi;</mi>
<mrow>
<mi>i</mi>
<mi>j</mi>
<mi>l</mi>
</mrow>
</msub>
<mo>+</mo>
<msub>
<mi>&gamma;&theta;</mi>
<mrow>
<mi>i</mi>
<mi>p</mi>
</mrow>
</msub>
<munder>
<mo>&Sigma;</mo>
<mrow>
<mi>i</mi>
<mo>,</mo>
<mi>p</mi>
</mrow>
</munder>
<mo>|</mo>
<mo>|</mo>
<mi>L</mi>
<mrow>
<mo>(</mo>
<msub>
<mi>x</mi>
<mi>i</mi>
</msub>
<mo>-</mo>
<msub>
<mi>x</mi>
<mi>p</mi>
</msub>
<mo>)</mo>
</mrow>
<mo>|</mo>
<msup>
<mo>|</mo>
<mn>2</mn>
</msup>
<mo>+</mo>
<mfrac>
<mi>&lambda;</mi>
<mn>2</mn>
</mfrac>
<mo>|</mo>
<mo>|</mo>
<mi>L</mi>
<mo>|</mo>
<msubsup>
<mo>|</mo>
<mi>F</mi>
<mn>2</mn>
</msubsup>
<mo>,</mo>
</mrow>
Wherein, | | | |FIt is Frobenius norms, for ensureing maximum boundary, λ is Frobenius norm regularization coefficients, is led to
It is semi-supervised regular terms parameter often to take λ=1, γ.
4. the Classification of Polarimetric SAR Image method according to claim 1 that network is measured based on semi-supervised depth distance, it is special
Levy and be, the parameter that depth network is initialized described in step (5) is：
5a, the hidden layer number of depth distance measurement network are 3, and every layer of nodes are respectively：150,100,50；
5b, random initializtion depth sparseness filtering network weight parameter W and bias unit b, the nodes of each hidden layer are divided into
Nk,T is the dimension of input signal, N1It is the nodes of first hidden layer of network,Represent network
- 1 hidden layer of k-th of hidden layer and kth between weight matrix,For the bias unit of k-th of hidden layer.
5. the Classification of Polarimetric SAR Image method according to claim 1 that network is measured based on semi-supervised depth distance, it is special
Levy and be, wherein described in step (6) is to the pre-training process that depth distance measures network：
6a, the pre-training sample for inputting depth sparseness filtering network, by unmarked sample xiAnd its it is corresponding have mark and it is unmarked
Wishart neighbours sample is input in depth network as pre-training sample,
6b, set xi∈Rt×1It is input vector, i=1,2 ..., n.N represents input sample number, then the output of first hidden layer
It can be expressed as：
<mrow>
<msubsup>
<mi>h</mi>
<mi>i</mi>
<mn>1</mn>
</msubsup>
<mo>=</mo>
<mi>s</mi>
<mrow>
<mo>(</mo>
<msup>
<mi>W</mi>
<mn>1</mn>
</msup>
<msub>
<mi>x</mi>
<mi>i</mi>
</msub>
<mo>+</mo>
<msup>
<mi>b</mi>
<mn>1</mn>
</msup>
<mo>)</mo>
</mrow>
<mo>&Element;</mo>
<msup>
<mi>R</mi>
<mrow>
<msub>
<mi>N</mi>
<mn>1</mn>
</msub>
<mo>&times;</mo>
<mn>1</mn>
</mrow>
</msup>
<mo>,</mo>
</mrow>
Wherein, s () represents nonlinear sigmoid functions, makes z=W1xi+b1, then s (z)=(1+exp (- z))-1, willSecond hidden layer of network is input to, second hidden layer is obtained and is output as：
<mrow>
<msubsup>
<mi>h</mi>
<mi>i</mi>
<mn>2</mn>
</msubsup>
<mo>=</mo>
<mi>s</mi>
<mrow>
<mo>(</mo>
<msup>
<mi>W</mi>
<mn>2</mn>
</msup>
<msup>
<msub>
<mi>h</mi>
<mi>i</mi>
</msub>
<mn>1</mn>
</msup>
<mo>+</mo>
<msup>
<mi>b</mi>
<mn>2</mn>
</msup>
<mo>)</mo>
</mrow>
<mo>&Element;</mo>
<msup>
<mi>R</mi>
<mrow>
<msub>
<mi>N</mi>
<mn>2</mn>
</msub>
<mo>&times;</mo>
<mn>1</mn>
</mrow>
</msup>
<mo>,</mo>
</mrow>
Successively greedy training is gone down successively, then k-th of hidden layer is output as：
<mrow>
<msubsup>
<mi>h</mi>
<mi>i</mi>
<mi>k</mi>
</msubsup>
<mo>=</mo>
<mi>s</mi>
<mrow>
<mo>(</mo>
<msup>
<mi>W</mi>
<mi>k</mi>
</msup>
<msubsup>
<mi>h</mi>
<mi>i</mi>
<mrow>
<mi>k</mi>
<mo>-</mo>
<mn>1</mn>
</mrow>
</msubsup>
<mo>+</mo>
<msup>
<mi>b</mi>
<mi>k</mi>
</msup>
<mo>)</mo>
</mrow>
<mo>&Element;</mo>
<msup>
<mi>R</mi>
<mrow>
<msub>
<mi>N</mi>
<mi>k</mi>
</msub>
<mo>&times;</mo>
<mn>1</mn>
</mrow>
</msup>
<mo>;</mo>
</mrow>
6c, optimization network weight W, using big Boundary Nearest algorithm, linear transformation matrix L are equivalent to the weight W of depth network,
The optimization object function of k-th of hidden layer is：
<mrow>
<munder>
<mi>min</mi>
<mi>W</mi>
</munder>
<mrow>
<mo>(</mo>
<mn>1</mn>
<mo>-</mo>
<mi>&mu;</mi>
<mo>)</mo>
</mrow>
<munder>
<mo>&Sigma;</mo>
<mrow>
<mi>i</mi>
<mo>,</mo>
<mi>j</mi>
<mo>&RightArrow;</mo>
<mi>i</mi>
</mrow>
</munder>
<mo>|</mo>
<mo>|</mo>
<msup>
<mi>W</mi>
<mi>k</mi>
</msup>
<mrow>
<mo>(</mo>
<msubsup>
<mi>h</mi>
<mi>i</mi>
<mrow>
<mi>k</mi>
<mo>-</mo>
<mn>1</mn>
</mrow>
</msubsup>
<mo>-</mo>
<msubsup>
<mi>h</mi>
<mi>j</mi>
<mrow>
<mi>k</mi>
<mo>-</mo>
<mn>1</mn>
</mrow>
</msubsup>
<mo>)</mo>
</mrow>
<mo>|</mo>
<msup>
<mo>|</mo>
<mn>2</mn>
</msup>
<mo>+</mo>
<mi>&mu;</mi>
<munder>
<mo>&Sigma;</mo>
<mrow>
<mi>i</mi>
<mo>,</mo>
<mi>j</mi>
<mo>&RightArrow;</mo>
<mi>i</mi>
<mo>,</mo>
<mi>l</mi>
</mrow>
</munder>
<mrow>
<mo>(</mo>
<mn>1</mn>
<mo>-</mo>
<msub>
<mi>y</mi>
<mrow>
<mi>i</mi>
<mi>l</mi>
</mrow>
</msub>
<mo>)</mo>
</mrow>
<msubsup>
<mi>&xi;</mi>
<mrow>
<mi>i</mi>
<mi>j</mi>
<mi>l</mi>
</mrow>
<mi>k</mi>
</msubsup>
<mo>+</mo>
<msub>
<mi>&gamma;&theta;</mi>
<mrow>
<mi>i</mi>
<mi>p</mi>
</mrow>
</msub>
<munder>
<mo>&Sigma;</mo>
<mrow>
<mi>i</mi>
<mo>,</mo>
<mi>p</mi>
</mrow>
</munder>
<mo>|</mo>
<mo>|</mo>
<msup>
<mi>W</mi>
<mi>k</mi>
</msup>
<mrow>
<mo>(</mo>
<msubsup>
<mi>h</mi>
<mi>i</mi>
<mrow>
<mi>k</mi>
<mo>-</mo>
<mn>1</mn>
</mrow>
</msubsup>
<mo>-</mo>
<msubsup>
<mi>h</mi>
<mi>p</mi>
<mrow>
<mi>k</mi>
<mo>-</mo>
<mn>1</mn>
</mrow>
</msubsup>
<mo>)</mo>
</mrow>
<mo>|</mo>
<msup>
<mo>|</mo>
<mn>2</mn>
</msup>
<mo>+</mo>
<mfrac>
<mi>&lambda;</mi>
<mn>2</mn>
</mfrac>
<mo>|</mo>
<mo>|</mo>
<msup>
<mi>W</mi>
<mi>k</mi>
</msup>
<mo>|</mo>
<msubsup>
<mo>|</mo>
<mi>F</mi>
<mn>2</mn>
</msubsup>
<mo>,</mo>
</mrow>
Wherein,Whole optimization aim passes through under traditional gradient
Algorithm is dropped to solve.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710336185.7A CN107194336B (en) | 2017-05-11 | 2017-05-11 | Polarized SAR image classification method based on semi-supervised depth distance measurement network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710336185.7A CN107194336B (en) | 2017-05-11 | 2017-05-11 | Polarized SAR image classification method based on semi-supervised depth distance measurement network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN107194336A true CN107194336A (en) | 2017-09-22 |
| CN107194336B CN107194336B (en) | 2019-12-24 |
Family
ID=59873232
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710336185.7A Active CN107194336B (en) | 2017-05-11 | 2017-05-11 | Polarized SAR image classification method based on semi-supervised depth distance measurement network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107194336B (en) |
Cited By (16)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107633236A (en) * | 2017-09-28 | 2018-01-26 | 北京达佳互联信息技术有限公司 | Picture material understanding method, device and server |
| CN107870321A (en) * | 2017-11-03 | 2018-04-03 | 电子科技大学 | Radar range profile's target identification method based on pseudo label study |
| CN107944479A (en) * | 2017-11-16 | 2018-04-20 | 哈尔滨工业大学 | Disease forecasting method for establishing model and device based on semi-supervised learning |
| CN108280485A (en) * | 2018-02-01 | 2018-07-13 | 北京科技大学 | A kind of non-rigid method for searching three-dimension model based on spectrogram Wavelet Descriptor |
| CN108280460A (en) * | 2017-12-04 | 2018-07-13 | 西安电子科技大学 | Based on the SAR vehicle target recognition methods for improving convolutional neural networks |
| CN108846381A (en) * | 2018-07-06 | 2018-11-20 | 西安电子科技大学 | SAR image change detection based on maximal margin metric learning |
| CN109214386A (en) * | 2018-09-14 | 2019-01-15 | 北京京东金融科技控股有限公司 | Method and apparatus for generating image recognition model |
| CN110161480A (en) * | 2019-06-18 | 2019-08-23 | 西安电子科技大学 | Radar target identification method based on semi-supervised depth probabilistic model |
| CN111583321A (en) * | 2019-02-19 | 2020-08-25 | 富士通株式会社 | Image processing apparatus, method and medium |
| CN111639714A (en) * | 2020-06-01 | 2020-09-08 | 贝壳技术有限公司 | Method, device and equipment for determining attributes of users |
| CN111695468A (en) * | 2020-06-01 | 2020-09-22 | 西安电子科技大学 | Polarization SAR terrain classification method based on K-shot learning |
| CN111695466A (en) * | 2020-06-01 | 2020-09-22 | 西安电子科技大学 | Semi-supervised polarization SAR terrain classification method based on feature mixup |
| CN112257862A (en) * | 2020-09-30 | 2021-01-22 | 重庆大学 | Semi-supervised identification method based on relational network marker sample expansion |
| CN112364920A (en) * | 2020-11-12 | 2021-02-12 | 西安电子科技大学 | Thyroid cancer pathological image classification method based on deep learning |
| CN113128620A (en) * | 2021-05-11 | 2021-07-16 | 北京理工大学 | Semi-supervised domain self-adaptive picture classification method based on hierarchical relationship |
| CN113557704A (en) * | 2019-03-29 | 2021-10-26 | 华为技术有限公司 | Method and apparatus for wireless communication using polarization-based signal space mapping |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102208031A (en) * | 2011-06-17 | 2011-10-05 | 西安电子科技大学 | Freeman decomposition and homo-polarization rate-based polarized synthetic aperture radar (SAR) image classification method |
| US8125370B1 (en) * | 2007-04-16 | 2012-02-28 | The United States Of America As Represented By The Secretary Of The Navy | Polarimetric synthetic aperture radar signature detector |
| CN103839073A (en) * | 2014-02-18 | 2014-06-04 | 西安电子科技大学 | Polarization SAR image classification method based on polarization features and affinity propagation clustering |
-
2017
- 2017-05-11 CN CN201710336185.7A patent/CN107194336B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US8125370B1 (en) * | 2007-04-16 | 2012-02-28 | The United States Of America As Represented By The Secretary Of The Navy | Polarimetric synthetic aperture radar signature detector |
| CN102208031A (en) * | 2011-06-17 | 2011-10-05 | 西安电子科技大学 | Freeman decomposition and homo-polarization rate-based polarized synthetic aperture radar (SAR) image classification method |
| CN103839073A (en) * | 2014-02-18 | 2014-06-04 | 西安电子科技大学 | Polarization SAR image classification method based on polarization features and affinity propagation clustering |
Cited By (28)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107633236B (en) * | 2017-09-28 | 2019-01-22 | 北京达佳互联信息技术有限公司 | Picture material understanding method, device and server |
| CN107633236A (en) * | 2017-09-28 | 2018-01-26 | 北京达佳互联信息技术有限公司 | Picture material understanding method, device and server |
| CN107870321B (en) * | 2017-11-03 | 2020-12-29 | 电子科技大学 | Radar one-dimensional range profile target identification method based on pseudo-label learning |
| CN107870321A (en) * | 2017-11-03 | 2018-04-03 | 电子科技大学 | Radar range profile's target identification method based on pseudo label study |
| CN107944479A (en) * | 2017-11-16 | 2018-04-20 | 哈尔滨工业大学 | Disease forecasting method for establishing model and device based on semi-supervised learning |
| CN108280460B (en) * | 2017-12-04 | 2021-07-27 | 西安电子科技大学 | SAR vehicle target identification method based on improved convolutional neural network |
| CN108280460A (en) * | 2017-12-04 | 2018-07-13 | 西安电子科技大学 | Based on the SAR vehicle target recognition methods for improving convolutional neural networks |
| CN108280485B (en) * | 2018-02-01 | 2020-12-01 | 北京科技大学 | Non-rigid three-dimensional model retrieval method based on spectrogram wavelet descriptor |
| CN108280485A (en) * | 2018-02-01 | 2018-07-13 | 北京科技大学 | A kind of non-rigid method for searching three-dimension model based on spectrogram Wavelet Descriptor |
| CN108846381B (en) * | 2018-07-06 | 2022-03-18 | 西安电子科技大学 | SAR image change detection method based on maximum edge metric learning |
| CN108846381A (en) * | 2018-07-06 | 2018-11-20 | 西安电子科技大学 | SAR image change detection based on maximal margin metric learning |
| CN109214386A (en) * | 2018-09-14 | 2019-01-15 | 北京京东金融科技控股有限公司 | Method and apparatus for generating image recognition model |
| CN111583321A (en) * | 2019-02-19 | 2020-08-25 | 富士通株式会社 | Image processing apparatus, method and medium |
| CN113557704A (en) * | 2019-03-29 | 2021-10-26 | 华为技术有限公司 | Method and apparatus for wireless communication using polarization-based signal space mapping |
| CN110161480A (en) * | 2019-06-18 | 2019-08-23 | 西安电子科技大学 | Radar target identification method based on semi-supervised depth probabilistic model |
| CN110161480B (en) * | 2019-06-18 | 2023-04-21 | 西安电子科技大学 | Radar target identification method based on semi-supervised depth probability model |
| CN111695466B (en) * | 2020-06-01 | 2023-03-24 | 西安电子科技大学 | Semi-supervised polarization SAR terrain classification method based on feature mixup |
| CN111695468A (en) * | 2020-06-01 | 2020-09-22 | 西安电子科技大学 | Polarization SAR terrain classification method based on K-shot learning |
| CN111695466A (en) * | 2020-06-01 | 2020-09-22 | 西安电子科技大学 | Semi-supervised polarization SAR terrain classification method based on feature mixup |
| CN111639714A (en) * | 2020-06-01 | 2020-09-08 | 贝壳技术有限公司 | Method, device and equipment for determining attributes of users |
| CN111639714B (en) * | 2020-06-01 | 2021-07-23 | 贝壳找房(北京)科技有限公司 | Method, device and equipment for determining attributes of users |
| CN111695468B (en) * | 2020-06-01 | 2023-04-18 | 西安电子科技大学 | Polarization SAR terrain classification method based on K-shot learning |
| CN112257862B (en) * | 2020-09-30 | 2023-12-05 | 重庆大学 | Semi-supervised identification method based on relation network marking sample expansion |
| CN112257862A (en) * | 2020-09-30 | 2021-01-22 | 重庆大学 | Semi-supervised identification method based on relational network marker sample expansion |
| CN112364920A (en) * | 2020-11-12 | 2021-02-12 | 西安电子科技大学 | Thyroid cancer pathological image classification method based on deep learning |
| CN112364920B (en) * | 2020-11-12 | 2023-05-23 | 西安电子科技大学 | Thyroid cancer pathological image classification method based on deep learning |
| CN113128620B (en) * | 2021-05-11 | 2022-10-21 | 北京理工大学 | Semi-supervised domain self-adaptive picture classification method based on hierarchical relationship |
| CN113128620A (en) * | 2021-05-11 | 2021-07-16 | 北京理工大学 | Semi-supervised domain self-adaptive picture classification method based on hierarchical relationship |
Also Published As
| Publication number | Publication date |
|---|---|
| CN107194336B (en) | 2019-12-24 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107194336A (en) | The Classification of Polarimetric SAR Image method of network is measured based on semi-supervised depth distance | |
| CN106067042B (en) | Polarization SAR classification method based on semi-supervised depth sparseness filtering network | |
| CN103955702B (en) | SAR image terrain classification method based on depth RBF network | |
| Mitra et al. | Self-organizing neural network as a fuzzy classifier | |
| CN105184312B (en) | A kind of character detecting method and device based on deep learning | |
| CN108021947B (en) | A kind of layering extreme learning machine target identification method of view-based access control model | |
| CN105975931A (en) | Convolutional neural network face recognition method based on multi-scale pooling | |
| CN108984745A (en) | A kind of neural network file classification method merging more knowledge mappings | |
| CN107563422A (en) | A kind of polarization SAR sorting technique based on semi-supervised convolutional neural networks | |
| CN108182450A (en) | A kind of airborne Ground Penetrating Radar target identification method based on depth convolutional network | |
| CN107239802A (en) | A kind of image classification method and device | |
| CN107871136A (en) | The image-recognizing method of convolutional neural networks based on openness random pool | |
| CN106682569A (en) | Fast traffic signboard recognition method based on convolution neural network | |
| CN106991666B (en) | A kind of disease geo-radar image recognition methods suitable for more size pictorial informations | |
| CN107203810A (en) | A kind of precipitation Forecasting Methodology based on depth network | |
| CN107657243A (en) | Neutral net Radar range profile's target identification method based on genetic algorithm optimization | |
| CN109359693A (en) | A kind of Power Quality Disturbance Classification Method | |
| CN108921285A (en) | Single-element classification method in sequence based on bidirectional valve controlled Recognition with Recurrent Neural Network | |
| CN108446766A (en) | A kind of method of quick trained storehouse own coding deep neural network | |
| CN104408481A (en) | Deep wavelet neural network-based polarimetric SAR (synthetic aperture radar) image classification method | |
| CN110009030A (en) | Sewage treatment method for diagnosing faults based on stacking meta learning strategy | |
| US20240190442A1 (en) | Complex network-based complex environment model, cognition system, and cognition method of autonomous vehicle | |
| CN106529570A (en) | Image classification method based on deep ridgelet neural network | |
| CN105184370B (en) | A kind of river mouth river basin water quality evaluation method based on SOM sorting techniques | |
| CN106960185A (en) | The Pose-varied face recognition method of linear discriminant depth belief network |
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 |