CN104463207A - Knowledge self-encoding network and polarization SAR image terrain classification method thereof - Google Patents
Knowledge self-encoding network and polarization SAR image terrain classification method thereof Download PDFInfo
- Publication number
- CN104463207A CN104463207A CN201410741792.8A CN201410741792A CN104463207A CN 104463207 A CN104463207 A CN 104463207A CN 201410741792 A CN201410741792 A CN 201410741792A CN 104463207 A CN104463207 A CN 104463207A
- Authority
- CN
- China
- Prior art keywords
- network
- knowledge
- data
- particle
- classification
- 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
- 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
- G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Abstract
The invention discloses a knowledge self-encoding network and a polarization SAR image terrain classification method of the knowledge self-encoding network. The problems that an existing polarization SAR image terrain classification method has too many manual markers and is low in classification correct rate are mainly solved. The implementation process mainly includes the steps that the knowledge self-encoding network is constructed, features, at different depth levels, of input data are extracted layer by layer, and the polarization SAR image data Wish art distance is used as priori knowledge for guiding terrain classification; during network learning, an orthogonal PSO algorithm is used for optimizing network parameters to obtain a classification network; the data to be classified are input into the network to obtain classification results. As the knowledge self-encoding network is constructed, the features of the data are automatically extracted, manual participation is avoided, and therefore uncertain factors are eliminated; the priori knowledge is used for guiding the classification results, so that the classification correct rate is increased; the orthogonal PSO algorithm is used for network optimization, and the training speed is increased. The polarization SAR image terrain classification method can be used for data classification, image classification, scene classification, target recognition and data predictive analysis.
Description
Technical field
The invention belongs to technical field of image processing, in particular to the method for polarization SAR image terrain classification, specifically knowledge autoencoder network and a Classification of Polarimetric SAR Image method thereof, can be used for Data classification, Images Classification, scene classification, target identification and data prediction analysis.
Background technology
Polarization SAR image terrain classification technology is mainly for polarimetric synthetic aperture radar (PolarimetrieSynthetie Aperture Radar, write a Chinese character in simplified form POLSAR) become image data to carry out terrain classification, be the basis that polarization SAR image is understood further and applied.Polarimetric synthetic aperture radar is a kind of multiparameter, multichannel imaging radar system, it is by measuring the Complete polarimetry echo in each resolution element of ground, and then obtain the polarization information of target, there are scattering matrix, polarization coherence matrix or Kelmaugh matrix etc. at present for information representation.Compared with conventional radar image, polarimetric synthetic aperture radar can provide more terrestrial object information and characteristic of division.Classification of Polarimetric SAR Image is basis in polarization SAR market demand and important step, also be technological difficulties, mostly the sorting technique of Polarimetric SAR Image conventional is at present the method for decomposing based on polarization characteristic, in this method, the artificial composition participated in is more, can bring the many uncertain factors such as not comprehensive described data.
A kind of polarization SAR data classification method based on hybrid classifer is disclosed in the patent " polarization SAR data classification method and system based on hybrid classifer " (number of patent application: 201310310179, publication number: CN103366184A) of Wuhan University's application.The method is first by carrying out polarization decomposing to polarization scattering matrix, obtain initial polarization feature, then employing decision tree classifier selects the polarization characteristic for classifying from initial polarization characteristic, the polarization characteristic finally will selected, adopts support vector machine classifier to polarization SAR Data classification.The method combines the advantage of decision tree classifier and support vector machine classifier, but, the method is Shortcomings also, nicety of grading do not have too large raising compared to support vector machine classifier, complicated operation, and only make use of scattering signatures, be not enough to represent actual atural object, therefore, many to the point divided wrong on polarization SAR terrain classification.
Patent " a kind of detection method based on marking area in the natural image of degree of depth study " (patent No.: 201310739026 of Northwestern Polytechnical University's application, publication number: CN103810503A) in disclose a kind of degree of depth that utilizes and learn to carry out the detection method of marking area in natural image, in the training stage, first the picture choosing some on natural image database extracts essential characteristic, composing training sample, then degree of depth learning model is utilized to learn extracting feature again, thus obtain the more abstract enhanced advanced feature more having separating capacity, finally with the features training sorter learning to arrive.At test phase, for any width test pattern, first extract essential characteristic, then the depth model trained is utilized, extract enhanced advanced feature, finally utilize sorter to carry out conspicuousness whether prediction, and using the saliency value of the predicted value of each pixel as this point.The method obtains the conspicuousness knowledge in natural image by the method that the degree of depth learns, but, the deficiency that the method exists is, the specific aim described for the knowledge feature of different pieces of information is not strong, the priori not making full use of data instructs classification results, thus causes accuracy in classification results not high.
Summary of the invention
The object of the invention is to the deficiency for prior art, overcoming in above-mentioned prior art needs artificial participation easily to cause probabilistic problem for during polarization SAR image terrain classification to polarization SAR image data feature extraction aspect, propose one and automatically extract feature efficiently, and instruct classification direction by the Wishart distance of polarization SAR image data as priori, enrich the feature interpretation information of data, accuracy is high, portable strong knowledge autoencoder network and polarization SAR image terrain classification method thereof.
Object of the present invention and thinking are described below: by building knowledge autoencoder network, automatically carrying out feature extraction to data, avoiding artificial participation, eliminate uncertain factor; Instruct classification results by priori, improve classification accuracy rate; With orthogonal PSO algorithm to network optimizing, accelerate training speed.The present invention is compared with other data classification methods in prior art, and can overcome for the uncertainty caused owing to manually participating in too much in characteristic extraction procedure in prior art, automaticity is high, and classification accuracy rate is high, portable strong.
For achieving the above object, technical scheme of the present invention comprises as lower network and method:
First the present invention provides a kind of knowledge autoencoder network, and network is a kind of polynary neural network structure, it is characterized in that: described network is made up of 1 autoencoder network Net1 and 1 knowledge network Net2 cascade; Net1 specifically has 1 input layer, a n own coding feature abstraction layer and 1 output layer cascade to form, and the output layer of note Net1 is output layer 1; Net2 has 1 stratum of intellectual and 1 output layer cascade is formed, and the output layer of note Net2 is output layer 2; Output layer 1 input that form Net2 in parallel with the stratum of intellectual of Net2 of Net1, and by direct for this input and output layer 2 cascade.
Realization of the present invention is also: the 1st level that is direct and own coding feature abstraction layer of the input layer in Net1 joins, by the successively effect of own coding feature abstraction layer 1 ~ n layer, obtain the advanced features f1 of input data under different levels, f2, fn, wherein fn is designated as the advanced features for classifying in autoencoder network Net1; The n-th layer of own coding feature abstraction layer directly with output layer 1 cascade, fn is input to the sorter of output layer 1, obtains the preliminary classification result of input data; Data in stratum of intellectual are designated as the knowledge feature fk of knowledge network Net2, the preliminary classification result that the output layer 1 of Net1 obtains is in parallel with fk forms the whole level feature of Net2 for classifying, this ultimate feature is input to output layer 2, the classification results be input data obtained.
A kind of polarization SAR image terrain classification method of the present invention or knowledge based autoencoder network, is characterized in that, is run under above-mentioned network, includes following steps:
Step 1 inputs polarization SAR image data, sets up polarization SAR image terrain classification training set U and test set V according to this input image data;
Step 2 builds the knowledge autoencoder network described in claim 1-2, the knowledge autoencoder network of not yet being trained;
Step 3 divides the level of this network, carries out layering Active Learning to own coding feature abstraction layer, and training data used is the sample in training set U; In the process of successively Active Learning, with orthogonal PSO algorithm, optimizing is carried out to the parameter of network:
First 3a sets Population Size is N, and iteration stopping condition is that loss function value L reaches minimum value Tmin;
3b generation quantity is that the population of N is settled in space at random, and find the individual extreme value under current state and population extreme value, setting network loss function is as follows, as the mark whether decision network is stable:
Wherein, y is the class label having label data, h
w,bx () is the output of input x after network, what this formula represented is the difference predicting label and class label;
3c judges whether the loss function L of current network is less than minimum value Tmin, if be less than minimum value Tmin, then exits circulation, performs step 4; If be not less than minimum value Tmin, perform step 3d, start renewal and calculating that iteration carries out this learning algorithm;
3d, for each particle in population, carries out renewal to their speed and individuality according to following formula and calculates:
Wherein,
for the speed after particle renewal,
for particle present speed, ω is the weight that particle keeps existing speed, c
1for particle is to the cognition of displacement state,
for the optimal location that particle is current,
C
2for particle is to the cognition of population group movement state,
for the optimal location that population colony is current;
3e, according to upgrading the particle obtained, the loss function L that computational grid is current, according to the size of loss function L, upgrades individual optimum and population optimum;
Between the individuality optimum of 3f for each particle, the thought according to orthogonal test is carried out orthogonal to it, form orthogonal after new particle populations;
They, for each particle in new particle populations, are acted on current autoencoder network by 3g, calculate the loss function L of current network;
3h judges whether loss function L is less than minimum value Tmin, if be not less than minimum value Tmin, then turns to step 3d to proceed; If be less than minimum value Tmin, then exit circulation, perform step 4;
The current optimized parameter of the own coding feature abstraction layer that layering Active Learning obtains by step 4 is assigned to the respective layer at different levels of autoencoder network Net1 respectively, obtains initial autoencoder network Net1; And the sample in training set U is input to the initial autoencoder network Net1 obtained, learn by the parameter of orthogonal PSO algorithm to initial autoencoder network Net1 and adjust, obtaining an autoencoder network Net1 for preliminary classification;
Step 5 trains the parameter of stratum of intellectual in Net2, the preliminary classification result that sample in training set U is obtained by Net1 and knowledge feature fk Input knowledge network N et2 in parallel, being learnt by the parameter of orthogonal PSO algorithm to knowledge network Net2, obtaining a knowledge network Net2 for finally classifying;
Step 6 is configured for the knowledge autoencoder network of classifying by Net1 and the Net2 cascade trained, and all for polarization SAR image data to be sorted are inputted this knowledge autoencoder network, obtains polarization SAR image terrain classification result.
Realization of the present invention is also: the data set wherein described in step 1, form with the coherence matrix coefficient in the T matrix of the input vector expression polarization SAR of each scatterer, this coherence matrix T is for the complex matrix being described as a 3*3 of each pixel, in this complex matrix, the real part of each element and imaginary part are extracted respectively and form the input vector of M*1 dimensional vector as this pixel, and calculate the multiple Wishart distance at this pixel and all kinds of center as knowledge feature fk, output vector is the class label belonging to these data; All have label data to form test set V, and random selecting has 3% of label data as training set U.
Realization of the present invention is also: the structure knowledge autoencoder network wherein described in step 2, builds knowledge autoencoder network according to the input layer number set, the degree of depth number of plies, own coding feature abstraction node layer number, stratum of intellectual's nodes and output layer nodes.
Realization of the present invention is also: between the individuality optimum for each particle wherein described in step 3f, the thought according to orthogonal test is carried out orthogonal to it, form orthogonal after new particle populations:
3f1 is the population of N for population quantity, calculates the value of the loss function L of each particle effect lower network, by L according to sorting successively from small to large;
3f2 chooses the less N/2 of loss function in a sequence particle, retains its data respectively tieed up and forms;
Their N/2 less with loss function respectively particles, for larger N/2 the particle of loss function in sequence, carry out orthogonal by 3f3, close to global optimum while of making each dimension all, obtain a new N/2 particle;
All new particles that 3f4 is obtained by 3f2 and 3f3 form orthogonal after new particle populations, its population quantity is N.
The present invention provide not only a kind of knowledge autoencoder network, and provides for a kind of outstanding orthogonal PSO algorithm during this e-learning.Described network is used for the application of polarization SAR image terrain classification, not only increases the diversity of polarization SAR sorting technique method choice, and improve the accuracy of polarization SAR image terrain classification result.
The present invention has the following advantages compared with prior art:
1, the present invention builds the advanced features that knowledge autoencoder network successively extracts input data, is completed by Web-based Self-regulated Learning, and overcome in classic method and need artificial setting to extract the complicacy of feature, automaticity is higher;
2, the priori of polarization SAR data is incorporated in network and instructs the categorised decision of network by the present invention, overcomes the understanding deviation that network oneself study produces, by the guidance of priori, the classify accuracy of network is improved greatly;
3, the present invention adopts a kind of orthogonal PSO algorithm of novel close globally optimal solution to carry out optimizing to the parameter learning of knowledge autoencoder network, trains, greatly reduce the time of network training in the mode of global search to network.
Accompanying drawing explanation
Fig. 1 is process flow diagram of the present invention;
Fig. 2 is knowledge autoencoder network design of graphics of the present invention;
Fig. 3 is that the present invention emulates the polarization SAR image terrain classification result obtained to polarization SAR data;
Fig. 4 is that the present invention and other method emulate the polarization SAR image terrain classification accuracy column comparison diagram obtained to polarization SAR data.
Embodiment
Below in conjunction with accompanying drawing, the present invention is described in further detail.
Embodiment 1:
The present invention proposes a kind of knowledge autoencoder network and polarization SAR image terrain classification method thereof, in the following description, suppose that pending data are expressed as double precision datum through process, all data input vectors have obtained the statistical distribution characteristic of sample by normalized; Suppose that the label of pending data represents to be represented according to the mode of 0/1, the node that can realize network output layer is expressed; Suppose to set the degree of depth number of plies of knowledge autoencoder network and the number of each layer depth interior joint according to the characteristic sum scale of input data; Suppose that network struction and study run on the computer being configured to Intel (R) Core (TM) 2Duo CPU T5550@1.83GHz, 3.00GB RAM, its spatial cache enough and not occupied; Suppose that software platform used is MatlabR2012a, the scale of mass data can be run.
During embodiment below describes, emulate the L-band full polarimetric SAR data that the data used are the regional true region, farmland of NASA/JPL lab A IRSAR system Flevoland in the middle part of the Holland of acquisition in 1989, its input is data from the coherence matrix T of this polarization SAR image, this coherence matrix T is for the complex matrix being described as a 3*3 of each pixel, the real part of each element in this complex matrix and imaginary part are extracted the formation as this pixel input vector respectively, and extract the input of multiple Wishart distance as stratum of intellectual, totally 16 dimensions, simultaneously owing to including 6 kinds of different atural object classifications in analogous diagram, it is then the terrain classification result of 6 classes to the classification output of this polarization SAR image data.
First the present invention is a kind of knowledge autoencoder network, and network is a kind of polynary neural network structure, and see Fig. 2, network is made up of 1 autoencoder network Net1 and 1 knowledge network Net2 cascade; Net1 specifically has 1 input layer, a n own coding feature abstraction layer and 1 output layer cascade to form, and the output layer of note Net1 is output layer 1; Net2 has 1 stratum of intellectual and 1 output layer cascade is formed, and the output layer of note Net2 is output layer 2; Output layer 1 input that form Net2 in parallel with the stratum of intellectual of Net2 of Net1, and by direct for this input and output layer 2 cascade.Own coding feature abstraction layer of the present invention is n layer, and choosing of n is relevant to actual experiment or actual conditions: when n is larger, easily causes the e-learning time longer, crosses conference and causes the unnecessary of time to expend; When n is less, the time is shorter, if n is too small, then can cause the reduction of Classification of Polarimetric SAR Image accuracy.According to actual conditions in this example, determine that the number of plies n of own coding feature abstraction layer is defined as 3.
Embodiment 2:
The general structure of knowledge autoencoder network and formation are with embodiment 1, input layer in Net1 the 1st level that is direct and own coding feature abstraction layer joins, by the successively effect of own coding feature abstraction layer 1 ~ n layer, obtain the advanced features f1 of input data under different levels, f2,, fn, wherein fn is designated as the advanced features for classifying in autoencoder network Net1; The n-th layer of own coding feature abstraction layer directly with output layer 1 cascade, fn is input to the sorter of output layer 1, obtains the preliminary classification result of input data; Data in stratum of intellectual are designated as the knowledge feature fk of knowledge network Net2, the preliminary classification result that the output layer 1 of Net1 obtains is in parallel with fk forms the whole level feature of Net2 for classifying, this ultimate feature is input to output layer 2, the classification results be input data obtained.
In this example, input layer is set to 15, the number of plies of own coding feature abstraction layer is set to 3 layers, successively nodes is set to 200,150,80, the result representative that each layer of own coding feature abstraction layer obtains inputs the advanced features f1 of data under different levels, f2, f3, wherein f3 is designated as the advanced features for classifying in autoencoder network Net1.Stratum of intellectual's nodes is set to 6, and its input data are the knowledge feature fk of knowledge network Net2, and the preliminary classification result obtained by fk and Net1 output layer forms the whole level feature of output layer for classifying of Net2 jointly.
Embodiment 3:
A kind of polarization SAR image terrain classification method of the present invention or knowledge based autoencoder network, runs under above-mentioned knowledge autoencoder network, and its general structure and formation are with embodiment 1-2, and polarization SAR image terrain classification process includes following steps:
Step 1 inputs polarization SAR image data, polarization SAR image terrain classification training set U and test set V is set up according to this input image data, by training set U, network parameter is trained, by test set V, network performance is tested, contribute to evaluation and the performance of network.
Step 2 builds above-mentioned knowledge autoencoder network, the knowledge autoencoder network of not yet being trained.The structure parameter of this network determines according to the feature of pending data, pointed, can realize the terrain classification of polarization SAR image better.
Step 3 divides the level of this network, carries out layering Active Learning to own coding feature abstraction layer, and training data used is the sample in training set U; In the process of successively Active Learning, with orthogonal PSO algorithm, optimizing is carried out to the parameter of network:
First 3a sets Population Size is N, and iteration stopping condition is that loss function value L reaches minimum value Tmin;
3b generation quantity is that the population of N is settled in space at random, and find the individual extreme value under current state and population extreme value, setting network loss function is as follows, as the mark whether decision network is stable:
Wherein, y is the class label having label data, h
w,bx () is the output of input x after network, what this formula represented is the difference predicting label and class label;
3c judges whether the loss function L of current network is less than minimum value Tmin, if be less than minimum value Tmin, then exits circulation, performs step 4; If be not less than minimum value Tmin, perform step 3d, start renewal and calculating that iteration carries out this learning algorithm;
3d, for each particle in population, carries out renewal to their speed and individuality according to following formula and calculates:
Wherein,
for the speed after particle renewal,
for particle present speed, ω is the weight that particle keeps existing speed, is set to 0.8, c in this example
1for particle is to the cognition of displacement state, in this example, be set to 2,
for the optimal location that particle is current, c
2for particle is to the cognition of population group movement state, in this example, be set to 2,
for the optimal location that population colony is current;
3e, according to upgrading the particle obtained, the loss function L that computational grid is current, according to the size of loss function L, upgrades individual optimum and population optimum;
Between the individuality optimum of 3f for each particle, the thought according to orthogonal test is carried out orthogonal to it, form orthogonal after new particle populations;
They, for each particle in new particle populations, are acted on current autoencoder network by 3g, calculate the loss function L of current network;
3h judges whether loss function L is less than minimum value Tmin, if be not less than minimum value Tmin, then turns to step 3d to proceed; If be less than minimum value Tmin, then exit circulation, perform step 4.
The current optimized parameter of the own coding feature abstraction layer that layering Active Learning obtains by step 4 is assigned to the respective layer at different levels of autoencoder network Net1 respectively, obtains initial autoencoder network Net1; And the sample in training set U is input to the initial autoencoder network Net1 obtained, learn by the parameter of orthogonal PSO algorithm to initial autoencoder network Net1 and adjust, obtaining an autoencoder network Net1 for preliminary classification.
Step 5 trains the parameter of stratum of intellectual in Net2, the preliminary classification result that sample in training set U is obtained by Net1 and knowledge feature fk Input knowledge network N et2 in parallel, being learnt by the parameter of orthogonal PSO algorithm to knowledge network Net2, obtaining a knowledge network Net2 for finally classifying.
Step 6 is configured for the knowledge autoencoder network of classifying by Net1 and the Net2 cascade trained, and all for polarization SAR image data to be sorted are inputted this knowledge autoencoder network, obtains polarization SAR image terrain classification result, see Fig. 3.
Embodiment 4:
The polarization SAR image terrain classification method of knowledge based autoencoder network is with embodiment 3, the data set mentioned in step 1, form with the coherence matrix coefficient in the T matrix of the input vector expression polarization SAR of each scatterer, this coherence matrix T is for the complex matrix being described as a 3*3 of each pixel, in this complex matrix, the real part of each element and imaginary part are extracted respectively and form the input vector of M*1 dimensional vector as this pixel, and calculate the multiple Wishart distance at this pixel and all kinds of center as knowledge feature fk, output vector is the class label belonging to these data, all have label data to form test set V, and random selecting has 3% of label data as training set U.In this example, data to be sorted have 81000, wherein have label data totally 41659 as Data classification test set, random selecting 1200 has exemplar as Data classification training set, the benefit of such choosing is with less sample training network parameter to realize predicting the classification compared with multisample, both the time of network training had been reduced, can obtain higher classification performance again, be a kind of selection mode preferably.
Embodiment 5:
The polarization SAR image terrain classification method of knowledge based autoencoder network is with embodiment 3-4, the structure knowledge autoencoder network mentioned in step 2, builds knowledge autoencoder network according to the input layer number set, the degree of depth number of plies, own coding feature abstraction node layer number, stratum of intellectual's nodes and output layer nodes.In this example, input node is set to 15, the each node layer number of own coding feature abstraction layer of autoencoder network Net1 is set to 200,150,80 successively, owing to including 6 kinds of different atural object classifications in analogous diagram, therefore output layer 1 nodes is set to 6, in knowledge network Net2, stratum of intellectual's nodes is set to 6, the nodes of output layer 2 is set to 6, the result that 6 corresponding nodes obtain is the classification results to input data.The structure parameter of this network determines according to the feature of pending data, pointed, can realize the terrain classification of polarization SAR image better.Embodiment 6:
The polarization SAR image terrain classification method of knowledge based autoencoder network is with embodiment 3-5, and between the individuality optimum for each particle described in step 3f, the thought according to orthogonal test is carried out orthogonal to it, form orthogonal after new particle populations:
3f1 is the population of N for population quantity, calculates the value of the loss function L of each particle effect lower network, by L according to sorting successively from small to large;
3f2 chooses the less N/2 of loss function in a sequence particle, retains its data respectively tieed up and forms;
Their N/2 less with loss function respectively particles, for larger N/2 the particle of loss function in sequence, carry out orthogonal by 3f3, close to global optimum while of making each dimension all, obtain a new N/2 particle;
All new particles that 3f4 is obtained by 3f2 and 3f3 form orthogonal after new particle populations, its population quantity is N.
The population quantity N set in this example is 40, is therefore 20 particles that 20 particles that loss function is less and loss function are larger respectively for the number of particles chosen in sequence in 3f2 and 3f3.
Embodiment 7:
The polarization SAR image terrain classification method of knowledge based autoencoder network, with embodiment 3-6, arranges three kinds of contrast algorithms, proves the advantages such as the classification accuracy rate of this method is high, learning time is fast by the present invention to polarization SAR image terrain classification result in this example.
The present invention can be verified by following emulation experiment:
1. simulated conditions:
Emulation experiment adopts data: the L-band full polarimetric SAR data in the true region, farmland, Flevoland area, Holland middle part that NASA/JPL lab A IRSAR system obtained in 1989, size is 270 × 300, mainly comprises 6 kinds of atural object classifications.
Hardware platform: Intel (R) Core (TM) 2Duo CPU T5550@1.83GHz, 3.00GB RAM;
Software platform: Matlab R2012a;
2. emulate content and analysis:
The method such as the present invention and H/a-wishart method, Freeman method and own coding device is used to contrast, all polarization SAR image data is classified, record training time, test duration, all kinds of accuracy and total accuracy, and analysis is compared to properties.
Emulation 1,10 subseries experiments are carried out by the L-band full polarimetric SAR data of the inventive method to the true region, farmland, Flevoland area, Holland middle part that NASA/JPL lab A IRSAR system obtained in 1989, calculate the average correct classification rate of 10 experiments, as the final accuracy of Classification of Polarimetric SAR Image, result is as Fig. 3.
Fig. 3 illustrates the classification results of the present invention for all data to be sorted of this width polarization SAR image data, wherein Fig. 3 (a) represents for polarization SAR image raw data atural object, Fig. 3 (b) for raw data label figure, Fig. 3 (c) for the present invention represents for the pcolor of the classification results of all data to be sorted of this width polarization SAR image data.As seen from Figure 3, the present invention more than 90%, has good terrain classification effect for Polarimetric SAR Image average correct classification rate of all categories.
Emulation 2, contrast experiment is carried out by the L-band full polarimetric SAR data of the methods such as the inventive method and existing H/a-wishart method, Freeman method and own coding device to the true region, farmland, Flevoland area, Holland middle part that NASA/JPL lab A IRSAR system obtained in 1989, calculate the average correct classification rate of 10 experiments, as the final accuracy of Classification of Polarimetric SAR Image, its quantitative analysis results contrast table is as shown in table 1, and column comparison diagram is as Fig. 4.
Table 1 Classification of Polarimetric SAR Image Comparative result
The classification results row that the final evaluating data of the present invention obtained in this example and other three kinds of methods obtain in Table 1, control methods is respectively two kinds of classical way H/a-wishart methods in polarization SAR image terrain classification and Freeman method, and the classical way own coding device in degree of depth study.The contrast of the training time of each method, test duration, all kinds of terrain classification accuracy and overall accuracy of classifying is followed successively by from top to bottom in table 1.Because H/a-wishart method and Freeman method belong to the implication that unsupervised approaches so there is no training time and test duration, therefore can to score method and own coding device method in the time, the training time of this method is only 2861.10s, compared to own coding device method 3185.10s be advantageous time, as seen by the importance of orthogonal PSO Algorithm for Training network.And in the test duration, both time phase differences are very few.Contrast the classification accuracy rate of all kinds of atural object, from the nicety of grading result of each class, for the classification that discrimination is high, the nicety of grading of four kinds of methods is equally matched.Distinguish for H/a-wishart method the 3rd class potato class that effective and Freeman distinguishes weak effect, and distinguish for Freeman the 6th class barley class that effective and H/a distinguishes weak effect, own coding device and this method all can reach good effect.Especially, in the first kind exposed soil class that conventional sorting methods cannot be distinguished, own coding device classifying quality only has 15.99%, and the accuracy of orthogonal PSO knowledge autoencoder network reaches 91.28%, has had significant lifting to the effect of such terrain classification.From the total accuracy of classification, the more traditional two kinds of methods of this method improve nearly 30 percentage points, and comparatively own coding device method improves nearly 10 percentage point, and the introducing of visible stratum of intellectual has a kind of very large help for degree of accuracy.And the classification accuracy rate of classification results this method and other three kinds of methods obtained is drawn as histogram shows in the diagram, the wherein classification accuracy rate that obtained by the left-to-right H/a-wishart of being followed successively by method, Freeman method, own coding device method and the present invention of histogram, Fig. 4 can find out that the classification accuracy rate of the present invention compared to other three kinds of methods in classification accuracy rate is high intuitively.
From Fig. 3, Fig. 4 and table 1, the present invention in polarization SAR image terrain classification accuracy far away higher than other method.
In sum, the present invention is by carrying out guidance to different types of areas to the introducing of priori to degree of depth network, and reach and obtain better classification results to polarization SAR image data, accuracy improves a lot.
To sum up, knowledge autoencoder network of the present invention and polarization SAR image terrain classification method thereof, mainly solve existing for the problem that polarization SAR image terrain classification method handmarking workload is comparatively large and classification accuracy rate is lower.It realizes mainly: build knowledge autoencoder network, successively extracts input data in the feature of different depth level, instructs terrain classification by polarization SAR image data Wishart distance as priori; With orthogonal PSO algorithm, sorter network is obtained to network parameter optimizing when e-learning; Data input network to be sorted is obtained classification results.The present invention, by building knowledge autoencoder network, automatically carries out feature extraction to data, avoids artificial participation, eliminates uncertain factor; Instruct classification results by priori, improve classification accuracy rate; With orthogonal PSO algorithm to network optimizing, accelerate training speed.The present invention can be used for Data classification, Images Classification, scene classification, target identification and data prediction analysis.
Claims (6)
1. a knowledge autoencoder network, is a kind of polynary neural network structure, it is characterized in that: described network is made up of 1 autoencoder network Net1 and 1 knowledge network Net2 cascade; Net1 specifically has 1 input layer, a n own coding feature abstraction layer and 1 output layer cascade to form, and the output layer of note Net1 is output layer 1; Net2 has 1 stratum of intellectual and 1 output layer cascade is formed, and the output layer of note Net2 is output layer 2; Output layer 1 input that form Net2 in parallel with the stratum of intellectual of Net2 of Net1, and by direct for this input and output layer 2 cascade.
2. a kind of knowledge autoencoder network according to claim 1, it is characterized in that, input layer in Net1 the 1st level that is direct and own coding feature abstraction layer joins, by the successively effect of own coding feature abstraction layer 1 ~ n layer, obtain the advanced features f1 of input data under different levels, f2 ... fn, wherein fn is designated as the advanced features for classifying in autoencoder network Net1; The n-th layer of own coding feature abstraction layer directly with output layer 1 cascade, fn is input to the sorter of output layer 1, obtains the preliminary classification result of input data; Data in stratum of intellectual are designated as the knowledge feature fk of knowledge network Net2, the preliminary classification result that the output layer 1 of Net1 obtains is in parallel with fk forms the whole level feature of Net2 for classifying, this ultimate feature is input to output layer 2, the classification results be input data obtained.
3. a polarization SAR image terrain classification method for knowledge based autoencoder network, is characterized in that, is run under the network described in claim 1-2, includes following steps:
Step 1 inputs polarization SAR image data, sets up polarization SAR image terrain classification training set U and test set V according to this input image data;
Step 2 builds the knowledge autoencoder network described in claim 1-2, the knowledge autoencoder network of not yet being trained;
Step 3 divides the level of this network, carries out layering Active Learning to own coding feature abstraction layer, and training data used is the sample in training set U; In the process of successively Active Learning, with orthogonal PSO algorithm, optimizing is carried out to the parameter of network:
First 3a sets Population Size is N, and iteration stopping condition is that loss function value L reaches minimum value Tmin;
3b generation quantity is that the population of N is settled in space at random, and find the individual extreme value under current state and population extreme value, setting network loss function is as follows, as the mark whether decision network is stable:
Wherein, y is the class label having label data, h
w,bx () is the output of input x after network, what this formula represented is the difference predicting label and class label;
3c judges whether the loss function L of current network is less than minimum value Tmin, if be less than minimum value Tmin, then exits circulation, performs step 4; If be not less than minimum value Tmin, perform step 3d, start renewal and calculating that iteration carries out this learning algorithm;
3d, for each particle in population, carries out renewal to their speed and individuality according to following formula and calculates:
Wherein,
for the speed after particle renewal,
for particle present speed, ω is the weight that particle keeps existing speed, c
1for particle is to the cognition of displacement state,
for the optimal location that particle is current, c
2for particle is to the cognition of population group movement state,
for the optimal location that population colony is current;
3e, according to upgrading the particle obtained, the loss function L that computational grid is current, according to the size of loss function L, upgrades individual optimum and population optimum;
Between the individuality optimum of 3f for each particle, the thought according to orthogonal test is carried out orthogonal to it, form orthogonal after new particle populations;
They, for each particle in new particle populations, are acted on current autoencoder network by 3g, calculate the loss function L of current network;
3h judges whether loss function L is less than minimum value Tmin, if be not less than minimum value Tmin, then turns to step 3d to proceed; If be less than minimum value Tmin, then exit circulation, perform step 4;
The current optimized parameter of the own coding feature abstraction layer that layering Active Learning obtains by step 4 is assigned to the respective layer at different levels of autoencoder network Net1 respectively, obtains initial autoencoder network Net1; And the sample in training set U being input to the initial autoencoder network Net1 obtained, the parameter starting from coding network Net1 with orthogonal PSO algorithm carries out learning and adjusting, and obtains an autoencoder network Net1 for preliminary classification;
Step 5 trains the parameter of stratum of intellectual in Net2, the preliminary classification result that sample in training set U is obtained by Net1 and knowledge feature fk Input knowledge network N et2 in parallel, being learnt by the parameter of orthogonal PSO algorithm to knowledge network Net2, obtaining a knowledge network Net2 for finally classifying;
Step 6 is configured for the knowledge autoencoder network of classifying by Net1 and the Net2 cascade trained, and all for polarization SAR image data to be sorted are inputted this knowledge autoencoder network, obtains polarization SAR image terrain classification result.
4. the polarization SAR image terrain classification method of knowledge based autoencoder network according to claim 3, it is characterized in that, data set wherein described in step 1, form with the coherence matrix coefficient in the T matrix of the input vector expression polarization SAR of each scatterer, this coherence matrix T is for the complex matrix being described as a 3*3 of each pixel, in this complex matrix, the real part of each element and imaginary part are extracted respectively and form the input vector of M*1 dimensional vector as this pixel, and calculate the multiple Wishart distance at this pixel and all kinds of center as knowledge feature fk, output vector is the class label belonging to these data, all have label data to form test set V, and random selecting has 3% of label data as training set U.
5. the polarization SAR image terrain classification method of knowledge based autoencoder network according to claim 3, it is characterized in that, structure knowledge autoencoder network wherein described in step 2, builds knowledge autoencoder network according to the input layer number set, the degree of depth number of plies, own coding feature abstraction node layer number, stratum of intellectual's nodes and output layer nodes.
6. the polarization SAR image terrain classification method of knowledge based autoencoder network according to claim 3, it is characterized in that, between the individuality optimum for each particle wherein described in step 3f, the thought according to orthogonal test is carried out orthogonal to it, form orthogonal after new particle populations:
3f1 is the population of N for population quantity, calculates the value of the loss function L of each particle effect lower network, by L according to sorting successively from small to large;
3f2 chooses the less N/2 of loss function in a sequence particle, retains its data respectively tieed up and forms;
Their N/2 less with loss function respectively particles, for larger N/2 the particle of loss function in sequence, carry out orthogonal by 3f3, close to global optimum while of making each dimension all, obtain a new N/2 particle;
All new particles that 3f4 is obtained by 3f2 and 3f3 form orthogonal after new particle populations, its population quantity is N.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201410741792.8A CN104463207B (en) | 2014-12-05 | 2014-12-05 | Knowledge autoencoder network and its polarization SAR image terrain classification method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201410741792.8A CN104463207B (en) | 2014-12-05 | 2014-12-05 | Knowledge autoencoder network and its polarization SAR image terrain classification method |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN104463207A true CN104463207A (en) | 2015-03-25 |
| CN104463207B CN104463207B (en) | 2017-08-25 |
Family
ID=52909221
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201410741792.8A Active CN104463207B (en) | 2014-12-05 | 2014-12-05 | Knowledge autoencoder network and its polarization SAR image terrain classification method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN104463207B (en) |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106228130A (en) * | 2016-07-19 | 2016-12-14 | 武汉大学 | Remote sensing image cloud detection method of optic based on fuzzy autoencoder network |
| CN106772371A (en) * | 2016-11-21 | 2017-05-31 | 上海卫星工程研究所 | Polarimetric calibration parameter requirements analysis method based on polarimetric SAR interferometry classification application |
| CN107229944A (en) * | 2017-05-04 | 2017-10-03 | 青岛科技大学 | Semi-supervised active identification method based on cognitive information particle |
| CN107392122A (en) * | 2017-07-07 | 2017-11-24 | 西安电子科技大学 | Polarization SAR silhouette target detection method based on multipolarization feature and FCN CRF UNEs |
| CN108319667A (en) * | 2018-01-22 | 2018-07-24 | 上海星合网络科技有限公司 | The knowledge hierarchy methods of exhibiting and device of multidimensional |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US8209080B2 (en) * | 2009-04-27 | 2012-06-26 | Toyota Motor Engineering & Manufacturing North America, Inc. | System for determining most probable cause of a problem in a plant |
| CN103955702A (en) * | 2014-04-18 | 2014-07-30 | 西安电子科技大学 | SAR image terrain classification method based on depth RBF network |
| CN104156736A (en) * | 2014-09-05 | 2014-11-19 | 西安电子科技大学 | Polarized SAR image classification method on basis of SAE and IDL |
-
2014
- 2014-12-05 CN CN201410741792.8A patent/CN104463207B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US8209080B2 (en) * | 2009-04-27 | 2012-06-26 | Toyota Motor Engineering & Manufacturing North America, Inc. | System for determining most probable cause of a problem in a plant |
| CN103955702A (en) * | 2014-04-18 | 2014-07-30 | 西安电子科技大学 | SAR image terrain classification method based on depth RBF network |
| CN104156736A (en) * | 2014-09-05 | 2014-11-19 | 西安电子科技大学 | Polarized SAR image classification method on basis of SAE and IDL |
Non-Patent Citations (1)
| Title |
|---|
| 王寿彪等: "基于知识的合成孔径雷达图像目标识别研究", 《图书情报工作 2012年增刊（1）》 * |
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106228130A (en) * | 2016-07-19 | 2016-12-14 | 武汉大学 | Remote sensing image cloud detection method of optic based on fuzzy autoencoder network |
| CN106228130B (en) * | 2016-07-19 | 2019-09-10 | 武汉大学 | Remote sensing image cloud detection method of optic based on fuzzy autoencoder network |
| CN106772371A (en) * | 2016-11-21 | 2017-05-31 | 上海卫星工程研究所 | Polarimetric calibration parameter requirements analysis method based on polarimetric SAR interferometry classification application |
| CN107229944A (en) * | 2017-05-04 | 2017-10-03 | 青岛科技大学 | Semi-supervised active identification method based on cognitive information particle |
| CN107229944B (en) * | 2017-05-04 | 2021-05-07 | 青岛科技大学 | Semi-supervised active identification method based on cognitive information particles |
| CN107392122A (en) * | 2017-07-07 | 2017-11-24 | 西安电子科技大学 | Polarization SAR silhouette target detection method based on multipolarization feature and FCN CRF UNEs |
| CN108319667A (en) * | 2018-01-22 | 2018-07-24 | 上海星合网络科技有限公司 | The knowledge hierarchy methods of exhibiting and device of multidimensional |
| CN108319667B (en) * | 2018-01-22 | 2021-03-05 | 上海星合网络科技有限公司 | Multidimensional knowledge system display method and device |
Also Published As
| Publication number | Publication date |
|---|---|
| CN104463207B (en) | 2017-08-25 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111259930B (en) | General target detection method of self-adaptive attention guidance mechanism | |
| CN105184309B (en) | Classification of Polarimetric SAR Image based on CNN and SVM | |
| CN104992184B (en) | A kind of multiclass image classification method based on semi-supervised extreme learning machine | |
| CN108009509A (en) | Vehicle target detection method | |
| CN104156734B (en) | A kind of complete autonomous on-line study method based on random fern grader | |
| CN110472817A (en) | A kind of XGBoost of combination deep neural network integrates credit evaluation system and its method | |
| CN107633226B (en) | Human body motion tracking feature processing method | |
| CN105005789B (en) | A kind of remote sensing images terrain classification method of view-based access control model vocabulary | |
| CN101847263B (en) | Unsupervised image division method based on multi-target immune cluster integration | |
| CN107862275A (en) | Human bodys' response model and its construction method and Human bodys' response method | |
| CN104463207A (en) | Knowledge self-encoding network and polarization SAR image terrain classification method thereof | |
| CN111833322B (en) | Garbage multi-target detection method based on improved YOLOv3 | |
| CN104331716A (en) | SVM active learning classification algorithm for large-scale training data | |
| CN108229550A (en) | A kind of cloud atlas sorting technique that network of forests network is cascaded based on more granularities | |
| CN104484681A (en) | Hyperspectral remote sensing image classification method based on space information and ensemble learning | |
| CN104680193B (en) | Online objective classification method and system based on quick similitude network integration algorithm | |
| CN108447057A (en) | SAR image change detection based on conspicuousness and depth convolutional network | |
| CN110287985B (en) | Depth neural network image identification method based on variable topology structure with variation particle swarm optimization | |
| CN102542293A (en) | Class-I extraction and classification method aiming at high-resolution SAR (Synthetic Aperture Radar) image scene interpretation | |
| CN114998220B (en) | Tongue image detection and positioning method based on improved Tiny-YOLO v4 natural environment | |
| CN111368660A (en) | Single-stage semi-supervised image human body target detection method | |
| CN105740914A (en) | Vehicle license plate identification method and system based on neighboring multi-classifier combination | |
| CN108446616A (en) | Method for extracting roads based on full convolutional neural networks integrated study | |
| Chen et al. | Agricultural remote sensing image cultivated land extraction technology based on deep learning | |
| CN112633257A (en) | Potato disease identification method based on improved convolutional neural network |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |