CN107194336B - Polarized SAR image classification method based on semi-supervised depth distance measurement network - Google Patents
Polarized SAR image classification method based on semi-supervised depth distance measurement network Download PDFInfo
- Publication number
- CN107194336B CN107194336B CN201710336185.7A CN201710336185A CN107194336B CN 107194336 B CN107194336 B CN 107194336B CN 201710336185 A CN201710336185 A CN 201710336185A CN 107194336 B CN107194336 B CN 107194336B
- Authority
- CN
- China
- Prior art keywords
- samples
- sample
- network
- supervised
- distance measurement
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- 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
Abstract
The invention discloses a polarized SAR image classification method based on a semi-supervised depth distance measurement network. The technical problems that the traditional deep learning only considers the nonlinear relation of sample characteristics and the classification precision is not high when the number of marked samples is small are solved, and the method comprises the following steps: inputting polarized SAR image data to be classified; solving a neighbor sample with a marked sample; constructing a loss function of a semi-supervised large-boundary nearest neighbor algorithm; initializing parameters of a network; pre-training the network; fine-tuning the network; performing category prediction on the unmarked samples; and outputting a polarized SAR image classification result graph to be classified and classification precision. The invention overcomes the problems of classification precision and information waste of a large number of unmarked samples due to the insufficient marked samples by constructing a depth distance measurement network and adding a popular learning regular term in a large-boundary neighbor algorithm.
Description
Technical Field
The invention belongs to the technical field of image processing, and particularly relates to a polarized SAR image classification method, in particular to a polarized SAR image classification method based on a semi-supervised depth distance measurement network. The method can be used for environmental monitoring, earth resource surveying, military systems and the like.
Background
Machine Learning (ML), which is a sub-field of computer science, is an algorithm that is constructed from artificial intelligence, computer Learning theory and pattern recognition, and can learn knowledge from data and predict similar data. Machine learning can learn various attributes from raw data so as to have the capability of processing various similar problems, namely how to enable a computer to automatically acquire new knowledge in experience learning. In the field of classification and identification of polarized SAR images, machine learning has been a lot of breakthrough progress, and some effective classification methods are proposed successively. Such as wishartmaximulikeliod (wml), Support Vector Machines (SVM), and the like.
Most of common machine learning methods use a method for manually extracting features, are time-consuming and labor-consuming, and cannot necessarily obtain features beneficial to polarization SAR image classification. Deep learning is a new field rapidly developed in machine learning, and is a brand new feature extraction method developed on the basis of an artificial neural network. Deep learning has a unique feature extraction mechanism, and hierarchical features of data can be autonomously learned through a multi-layer network structure model. The complexity of manually extracting the features in the traditional machine learning method is avoided, and meanwhile, the learned hierarchical features can express the inherent attributes and characteristics of the data. These abstract features can improve the classification accuracy of various classification tasks. For the polarized SAR image, the deep learning can autonomously learn the abstract hierarchical features representing the intrinsic attributes of the polarized SAR image from the polarized SAR image data, and the extracted features can be very conveniently and effectively applied to ground feature classification, environment monitoring, target identification and the like.
For the classification task of the polarized SAR image, proper distance measurement selection is very important, and the discrimination performance of the classifier greatly depends on the selection of a distance measurement function. For example, the most common euclidean distance simply considers the true distance between data, does not consider the internal structure and specific attributes of the sample, and for some data, the reliability of the data cannot be guaranteed by a measurement mode similar to the euclidean distance, that is, the obtained data with a short distance are not very similar or the same data, and the distance measurement result may obtain a less-ideal result under the action of classifiers such as KNN, SVM and the like. The distance metric learning aims to obtain a proper distance metric mode by learning the internal structure and the attribute of the data, and in the distance metric mode, the sample interval with the same class label is correspondingly reduced, while the sample interval without the same label is correspondingly increased, so that a new feature space is obtained, and the data becomes more beneficial to classification.
The traditional machine learning method needs manual feature extraction and the obtained shallow features cannot fully reflect the inherent attributes of the data. The existing deep learning methods such as an auto-encoder, a deep belief network and the like adopt unsupervised pre-training methods, label sample guidance is not available, the pre-training effect is not ideal, and a large number of labeled samples are needed to perform back propagation fine adjustment on network parameters. For methods such as supervised convolutional neural networks, when the number of label samples is small, the network performance cannot be guaranteed, and the classification result is not ideal.
Disclosure of Invention
The invention aims to provide a polarimetric SAR image classification method based on a semi-supervised depth distance measurement network, which can obtain high classification precision even when the number of label samples is small, aiming at overcoming the defects of the prior art and being used for improving the image classification precision.
The invention relates to a polarized SAR image classification method based on a semi-supervised depth distance measurement network, which is characterized by comprising the following steps:
(1) inputting polarized SAR image data to be classified: namely a coherent matrix T of the polarized SAR image, a label matrix Y is obtained according to the ground feature distribution information of the polarized SAR image, the distribution of the same ground feature is represented by the same class label, the ground feature distribution which can not determine the class is represented by 0 in the label matrix Y, and a sample matrix is generated according to the coherent matrix T of the polarized SAR imageN is the total number of samples, xiRepresents the ith sample;
(2) selecting labeled and unlabeled samples: according to sample data X and a label matrix Y of the polarized SAR image, randomly selecting 1% of samples in each class as marked samples, and selecting the rest of samples as unmarked samples;
(3) finding neighbor samples with labeled samples: in all sample data, K of each marked sample is obtained1Marked Wishart homogeneous neighbor sample and K2A sample of unmarked Wishart neighbors;
(4) constructing a loss function of a semi-supervised large-boundary nearest neighbor algorithm: increasing a popular learning regular term on the basis of a large-boundary neighbor algorithm (LMNN), and improving a supervised large-boundary neighbor algorithm into a semi-supervised large-boundary neighbor algorithm to obtain a loss function of the semi-supervised large-boundary neighbor algorithm;
(5) basic parameters for initializing the deep network: randomly initializing a weight parameter W and a bias unit b of the depth distance measurement network, setting the number of nodes of each layer of the depth distance measurement network, and determining the overall structure of the depth distance measurement network;
(6) pre-training a depth distance measurement network: sending the marked sample and the Wishart neighbor sample corresponding to the marked sample into a depth distance measurement network for pre-training, using a semi-supervised large-boundary neighbor algorithm and adopting a layer-by-layer greedy pre-training method, taking the output of the previous layer as the input of the next layer until the last hidden layer is trained, optimizing the weight of the network, and primarily optimizing the weight of the depth distance measurement network;
(7) fine tuning the depth distance measurement network: utilizing the labeled samples and the label information thereof, and combining a Softmax classifier to finely adjust the depth distance measurement network, further optimizing the weight of the network, enabling the network to be more stable, and completing the weight optimization of the depth distance measurement network;
(8) and (3) performing class prediction on the unmarked sample: sending the unmarked samples into a depth distance measurement network, and predicting the class labels of the unmarked samples by using a Softmax classifier to obtain the prediction class of each unmarked sample;
(9) outputting a classification result graph and classification precision of the polarized SAR image to be classified: and outputting a final classification result of the polarized SAR image to be classified and calculating the precision of the classification according to the labeled sample and the non-labeled sample with the predicted category.
The invention combines a large-boundary nearest neighbor algorithm with a deep learning method and provides a polarized SAR image classification method based on a semi-supervised deep distance measurement network. The method of deep learning is utilized to avoid the complexity of manually extracting the features in the traditional machine learning method.
The invention has the following advantages:
1. the method adopts a deep learning method, and utilizes the deep distance measurement network to learn the characteristics of the polarized SAR image autonomously, so that the complexity of ergonomics learning characteristics in the traditional method is avoided, the deep distance measurement network is utilized to learn richer internal level characteristics of the polarized SAR image, the characteristics are more beneficial to the classification of the polarized SAR image, and the classification precision of the polarized SAR image is effectively improved.
2. According to the semi-supervised large-boundary nearest neighbor algorithm, the semi-supervised large-boundary nearest neighbor algorithm is obtained by adding the popular learning regular term on the basis of the large-boundary nearest neighbor algorithm, and the problems that the classification accuracy of a classifier is low under the condition of few labeled samples and information is wasted due to a large number of unlabeled samples are solved.
3. According to the invention, a semi-supervised distance measurement method and a deep learning method are combined to obtain a deep distance measurement network, the deep network can simultaneously learn linear and nonlinear characteristics of a sample, and the inherent attributes of the sample are fully described, so that the learned characteristics can obviously improve the classification accuracy.
Drawings
FIG. 1 is a schematic flow chart of an implementation of the present invention;
fig. 2 is a graph of experimental results of polarized SAR images for the flevioland region in the netherlands, in which fig. 2(a) is an exploded view of Pauli of the polarized SAR images, fig. 2(b) is a label graph, fig. 2(c) is a graph of classification results using the comparison method WDSN, fig. 2(d) is a graph of classification results using the comparison method WDBN, fig. 2(e) is a graph of classification results using the comparison method DSFN, and fig. 2(f) is a graph of classification results using the method of the present invention;
FIG. 3 is a schematic diagram of a semi-supervised large-boundary nearest neighbor algorithm.
Detailed Description
The invention is described in detail below with reference to the accompanying drawings:
example 1
Due to the development of remote sensing technology, the method is widely applied to the fields of environmental monitoring, earth resource surveying, military systems and the like, the requirement for polarized SAR image processing is continuously increased, and deep learning has obvious advantages in a machine learning method. Most of the traditional deep learning networks are unsupervised or supervised learning methods, when the number of marked samples is small, the unsupervised deep learning method can generate under-fitting conditions only depending on unsupervised pre-training, and the network performance cannot be effectively improved due to the fine adjustment of a small number of marked samples; for the supervised deep learning method, the conditions of insufficient network training and poor network performance occur. The traditional deep learning method does not consider the linear and nonlinear characteristics of the sample at the same time, and the characteristics obtained by learning cannot fully reflect the inherent attributes of the sample. The invention develops research and innovation aiming at the current situations, and provides a polarized SAR image classification method based on a semi-supervised depth distance measurement network, which is shown in figure 1 and comprises the following steps:
(1) inputting polarized SAR image data to be classified: namely a coherent matrix T of the polarized SAR image, a label matrix Y is obtained according to the ground feature distribution information of the polarized SAR image, the distribution of the same ground feature is represented by the same class label, the ground feature distribution which can not determine the class is represented by 0 in the label matrix Y, and a sample matrix is generated according to the coherent matrix T of the polarized SAR imageN is the total number of samples, xiThe ith sample is represented.
(2) Selecting labeled and unlabeled samples: according to sample data X and a label matrix Y of the polarized SAR image, 1% of samples are randomly selected from each type as marked samples, and the rest are unmarked samples.
(3) Finding neighbor samples with labeled samples: in all sample data, K of each marked sample is obtained1Marked Wishart homogeneous neighbor sample and K2Each unmarked Wishart neighbor sample has K1+K2And (4) each Wishart neighbor sample, and the basic data processing is completed.
(4) Constructing a loss function of a semi-supervised large-boundary nearest neighbor algorithm: the method is characterized in that a popular learning regular term is added on the basis of a large-boundary neighbor algorithm, the supervised large-boundary neighbor algorithm is improved into a semi-supervised large-boundary neighbor algorithm, a loss function of the semi-supervised large-boundary neighbor algorithm is obtained, part of unmarked samples are effectively utilized by the semi-supervised large-boundary neighbor algorithm, the demand on the marked samples is reduced, and the performance of the algorithm can be ensured under the condition that only a small number of marked samples exist.
(5) Basic parameters for initializing the deep network: randomly initializing a weight parameter W and a bias unit b of the depth distance measurement network, setting the number of nodes of each layer of the depth distance measurement network, and determining the overall structure of the depth distance measurement network.
(6) Pre-training a depth distance measurement network: and sending the marked sample and the Wishart neighbor sample corresponding to the marked sample into a depth distance measurement network for pre-training. And (3) utilizing a semi-supervised large-boundary nearest neighbor algorithm, adopting a layer-by-layer greedy pre-training method, taking the output of the previous layer as the input of the next layer until the last hidden layer is trained, optimizing the weight of the network, and primarily optimizing the weight of the depth distance measurement network.
(7) Fine tuning the depth distance measurement network: and (3) utilizing the labeled samples and the label information thereof and combining a Softmax classifier to finely adjust the depth distance measurement network, further optimizing the weight of the network, so that the network becomes more stable, and thus, completing the optimization of the depth distance measurement network.
(8) And (3) performing class prediction on the unmarked sample: and (3) sending the unmarked samples into a depth distance measurement network, and predicting the class labels of the unmarked samples by using a Softmax classifier to obtain the prediction class of each unmarked sample.
(9) Outputting a classification result graph and classification precision of the polarized SAR image to be classified: and (4) outputting the final classification result of the polarized SAR image to be classified and calculating the precision of the current classification according to the labeled sample and the unlabeled sample of which the classification is predicted in the step (8).
The technical idea of the invention is as follows: the semi-supervised large-boundary nearest neighbor algorithm is obtained by the large-boundary nearest neighbor algorithm under the action of a popular regular term, a deep distance measurement network is obtained by combining with a deep learning method, under the condition of having a small number of marked samples, feature learning is carried out on polarized SAR data, image classification is realized through a classifier, and the classification precision is improved.
Example 2
The method for classifying the polarized SAR image based on the semi-supervised depth distance measurement network is the same as that in the embodiment 1, and the Wishart neighbor sample of each marked sample is obtained in the step (3), and the method comprises the following steps:
3a, a matrix of marked samples of And (3) expressing the number of the marked samples, and solving the Wishart distance between each marked sample and the rest of samples by using the following formula:
d(xi,xj)＝ln((xi)-1xj)+Tr((xj)-1xi)-q，
wherein, Tr () represents the trace of the matrix, and for the radar whose transmission and reception are integrated, the constant q is 3 due to reciprocity; for a radar whose transmission and reception are not integral, the constant q is 4;
3b, utilizing the sort function in MATLAB to compare the Wishart distance d (x) obtained in the step 3ai,xj) Arranged in ascending order of absolute value, K is taken before1Marked similar neighbor sample xj(j＝1,2,···,K1)，K2An unlabeled neighbor sample xp(p＝1,2,···,K2) As marked sample xiThe Wishart neighbor samples.
Example 3
The method for classifying the polarized SAR image based on the semi-supervised depth distance measurement network is the same as the embodiment 1-2, and the process for constructing the loss function of the semi-supervised large boundary nearest neighbor algorithm in the step (4) comprises the following steps:
4a, solving a loss function of a large boundary nearest neighbor algorithm:
the distance square formula of the large-boundary nearest neighbor algorithm is as follows:
wherein L is a linear variation matrix, xjIs xiThe same class of (a) labeled sample. If there is a marked sample xiX of non-homogeneous samples oflThe following formula is satisfied:
||L(xi-xl)||2≤||L(xi-xj)||2+1，
then, xlReferred to as "imposters".
The large boundary nearest neighbor algorithm can be expressed in two parts: loss function epsilon between homogeneous samplespullLoss function epsilon between (L) and non-homogeneous samplespush(L)，εpull(L) is used to penalize large distance between the marked sample and its similar neighbor, i.e. reduce the distance between similar samples; epsilonpush(L) is the distance used to penalize small distances between marked samples and "imposters", i.e., increase the distance between non-homogeneous samples:
wherein the symbol yilRepresenting marked samples xiAnd marked sample xlClass relationship of (1), yil1 if and only if yi＝ylIs true, i.e. xiAnd xlSamples of the same type are obtained; otherwise, yil＝0。[z]+Max (z,0) is the standard hinge function.
Therefore, the loss function ε of the large-boundary nearest neighbor algorithm1(L) is:
ε1(L)＝(1-μ)εpull(L)+μεpush(L)，
wherein the loss parameter mu is [0,1 ]]Let the symbol xiijl＝[1+||L(xi-xj)||2-||L(xi-xl)||2]+And then:
4b, solving a loss function of the semi-supervised large-boundary nearest neighbor algorithm:
popular learning regular term is added in a loss function of a large-boundary neighbor algorithm, the use of the regular term increases the utilization of part of unmarked samples, the supervised large-boundary neighbor algorithm is improved into a semi-supervised learning method, and the popular learning regular term J is addedRIs aimed at penalizing the marked sample xiAnd a label-free sample xpWith large spacing therebetween, i.e. with a reduction in the marked sample xiAnd a label-free sample xpThe interval therebetween:
wherein, the symbolRepresenting marked samples xiSample x with its unmarked neighborspSimilarity between the two, the loss function epsilon of the semi-supervised large-boundary nearest neighbor algorithm2(L) is:
wherein | · | purple sweetFIs a Frobenius norm used for ensuring the maximum boundary, and is a Frobenius norm regular term coefficient,usually, λ ═ 1, and γ is a semi-supervised regularization term parameter.
On the basis of the large-boundary neighbor algorithm, the semi-supervised large-boundary neighbor algorithm is obtained by adding the popular learning regular term, so that the problems of low classification accuracy of a classifier under the condition of less labeled samples and information waste caused by a large number of unlabeled samples are solved, and the action principle of the semi-supervised large-boundary neighbor algorithm on the samples can be shown in FIG. 3.
Example 4
The method for classifying the polarized SAR image based on the semi-supervised depth distance measurement network is the same as the embodiment 1-3, and the parameters of the initialized depth network in the step (5) are as follows:
5a, the number of hidden layers of the depth distance measurement network is 3, and the number of nodes of each layer is respectively as follows: 150, 100, 50;
5b, randomly initializing a weight parameter W and a bias unit b of the deep sparse filter network, wherein the number of nodes of each hidden layer is Nk，t is the dimension of the input signal, N1Is the number of nodes in the first hidden layer of the network,representing a weight matrix between the kth hidden layer and the (k-1) th hidden layer of the network,the bias unit of the k hidden layer.
Example 5
The method for classifying the polarized SAR image based on the semi-supervised depth distance measurement network is the same as the embodiment 1-4, and the pre-training process of the depth distance measurement network in the step (6) is as follows:
6a, inputting a pre-training sample of the deep sparse filter network, and enabling a non-mark sample xiAnd its corresponding labeled and unlabeled wishirt neighbor samples are input into the deep network as pre-training samples,
6b, setting xi∈Rt×1Is the input vector, i ═ 1, 2. n represents the number of input samples, the output of the first hidden layer can be represented as:
wherein s (·) represents a nonlinear sigmoid function, and z is W1xi+b1Then s (z) ═ 1+ exp (-z)-1Will beThe output of the second hidden layer is obtained by inputting the data into the second hidden layer of the network:
and (3) sequentially training the greedy layer by layer, wherein the output of the kth hidden layer is as follows:
6c, optimizing the network weight W, and using a large-boundary nearest neighbor algorithm to enable the linear transformation matrix L to be equivalent to the weight W of the deep network, wherein the optimization objective function of the kth hidden layer is as follows:
wherein the content of the first and second substances,the whole optimization objective function can be solved by the traditional gradient descent algorithm, L-BFGS and other methods.
According to the method, the deep network pre-training is carried out through a semi-supervised large-boundary neighbor algorithm, the samples are subjected to characteristic transformation by using a distance measurement learning method, the distance between similar samples is reduced, the distance between non-similar samples is increased, the linear and nonlinear characteristics of the samples are simultaneously learned through distance measurement learning and deep learning by the deep distance measurement network, and finally the obtained characteristics are beneficial to improving the classification accuracy of the polarized SAR image.
A more detailed example is given below to further illustrate the invention:
example 6
The method for classifying the polarized SAR image based on the semi-supervised depth distance measurement network is the same as the embodiment 1-5, and referring to FIG. 1, the specific implementation steps of the invention are as follows:
step 1, inputting polarized SAR image data to be classified, referring to fig. 2(a), wherein fig. 2(a) is a Pauli exploded view of a polarized SAR image in a Fleviland area of the Netherlands, inputting a coherence matrix T of the polarized SAR image, obtaining a label matrix Y according to ground feature distribution information of the polarized SAR image, referring to fig. 2(b), wherein fig. 2(b) is an image directly generated by the label matrix Y, different color blocks in the image represent different ground features, the distribution of the same ground feature is represented by the same class label in the label matrix, the distribution of the ground feature which can not be determined is represented by 0 in the label matrix, and generating a sample matrix according to the coherence matrix T of the polarized SAR imageN is the total number of samples, xiThe ith sample is represented.
This example uses polarized SAR images in the Flevoland region of the netherlands. The image is obtained by the NASA/JPLAIRSAR system, the size of the image is 300 x 270, and the image comprises 6 different ground object categories. Because the polarized coherent matrix T of the polarized SAR data is a Hamilton semi-positive definite matrix, the module values of 6 elements at the upper triangular position of the polarized coherent matrix T with the dimension of 3 multiplied by 3 can be extracted as the original characteristics of each pixel point, the T matrix is converted into a two-dimensional sample matrix by utilizing a reshape function in MATLAB software, each column represents a sample, and the dimension of each sample is 6 dimensions. Each sample corresponds to a pixel point on the polarized SAR image.
Step 2, selecting marked samples and unmarked samples according to the polaritySampling data X and a label matrix Y of the SAR image are converted, and the sampling data are extracted randomly according to categoriesOne labeled sample and the rest unlabeled samples, and the labeled sample of each class accounts for 1 percent of the total number of the classes.
Step 3, solving the adjacent sample with the marked sample, and solving the K of each marked sample in all sample data1Marked Wishart homogeneous neighbor sample and K2A sample of unmarked wishirt neighbors.
3a, a matrix of marked samples ofAnd solving the Wishart distance between the Wishart distance and other samples according to an improved Wishart distance formula: d (x)i,xj)＝ln((xi)-1xj)+Tr((xj)-1xi)-q，
Wherein, Tr () represents the trace of the matrix, and for the radar whose transmission and reception are integrated, the constant q is 3 due to reciprocity, and for the radar whose transmission and reception are not integrated, the constant q is 4, and the polarized SAR image data in Flevoland area used in this example is obtained and generated by the radar system whose transmission and reception are integrated, and q is 3.
3b, utilizing the sort function in MATLAB to compare the Wishart distance d (x) obtained in the step 3ai,xj) Arranged in ascending order of absolute value, K is taken before1Marked similar neighbor sample xj(j＝1,2,···,K1)，K2An unlabeled neighbor sample xp(p＝1,2,···,K2) As marked sample xiThe Wishart neighbor samples.
Since the polarized SAR data obeys Wishart distribution, there is a labeled sample xiUnmarked Wishart neighbor sample xpPossibly, the samples are the same class samples, and x is used in the semi-supervised large-boundary nearest neighbor algorithm of the inventioniAnd xpPunishment is carried out on the distance between the two samples, so that part of unmarked samples can be effectively utilized, and the existence of the pair is reducedThe need to label the sample.
And 4, constructing a loss function of the semi-supervised large-boundary neighbor algorithm, adding a popular learning regular term on the basis of the large-boundary neighbor algorithm, and improving the supervised large-boundary neighbor algorithm into the semi-supervised large-boundary neighbor algorithm to obtain the loss function of the semi-supervised large-boundary neighbor algorithm.
4a, firstly obtaining a loss function epsilon of a large boundary nearest neighbor algorithm1(L)：
The distance square formula of the large-boundary nearest neighbor algorithm is as follows:
wherein L is a linear variation matrix, xjIs xiThe same class of (a) labeled sample. If there is a marked sample xiX of non-homogeneous samples oflThe following formula is satisfied:
||L(xi-xl)||2≤||L(xi-xj)||2+1，
then, xlReferred to as "imposters".
The large boundary nearest neighbor algorithm can be expressed in two parts: loss function epsilon between homogeneous samplespullLoss function epsilon between (L) and non-homogeneous samplespush(L)，εpull(L) is used to penalize large distance between the marked sample and its similar neighbor, i.e. reduce the distance between similar samples; epsilonpush(L) is the distance used to penalize small distances between marked samples and "imposters", i.e., increase the distance between non-homogeneous samples:
therefore, the loss function ε of the large-boundary nearest neighbor algorithm1(L) is:
ε1(L)＝(1-μ)εpull(L)+μεpush(L)，
wherein the loss parameter mu is [0,1 ]]Let the symbol xiijl＝[1+||L(xi-xj)||2-||L(xi-xl)||2]+Then the loss function epsilon of the large boundary nearest neighbor algorithm is obtained1(L)：
4b, solving the loss function epsilon of the semi-supervised large-boundary nearest neighbor algorithm2(L)：
Popular learning regular term is added in a loss function of a large-boundary neighbor algorithm, the use of the regular term increases the utilization of part of unmarked samples, the supervised large-boundary neighbor algorithm is improved into a semi-supervised learning method, and the popular learning regular term J is addedRIs aimed at penalizing the marked sample xiAnd a label-free sample xpWith large spacing therebetween, i.e. with a reduction in the marked sample xiAnd a label-free sample xpThe interval therebetween:
wherein, the symbolRepresenting marked samples xiSample x with its unmarked neighborspSimilarity between the two, the loss function epsilon of the semi-supervised large-boundary nearest neighbor algorithm2(L) is:
wherein | · | purple sweetFThe maximum boundary is guaranteed by the Frobenius norm, and λ is a Frobenius norm regularization term coefficient, where λ is 1 and γ is a semi-supervised regularization term parameter.
The semi-supervised large-boundary nearest neighbor algorithm combines the marked samples and part of unmarked samples, and overcomes the condition of low algorithm performance when the marked samples are insufficient.
Step 5, initializing basic parameters of the deep network: randomly initializing a weight parameter W and a bias unit b of the depth distance measurement network, setting the number of nodes of each layer of the depth distance measurement network, and determining the overall structure of the depth distance measurement network;
5a, the number of hidden layers of the depth distance measurement network is 3, and the number of nodes of each layer is respectively as follows: 150, 100, 50;
5b, randomly initializing a weight parameter W and a bias unit b of the deep sparse filter network, wherein the number of nodes of each hidden layer is Nk，t is the dimension of the input signal, N1Is the number of nodes in the first hidden layer of the network,representing a weight matrix between the kth hidden layer and the (k-1) th hidden layer of the network,the bias unit of the k hidden layer.
And 6, pre-training the depth distance measurement network.
6a, inputting a pre-training sample of the deep sparse filter network, and enabling a non-mark sample xiAnd its corresponding labeled and unlabeled wishirt neighbor samples are input into the deep network as pre-training samples,
6b, setting xi∈Rt×1Is the input vector, i ═ 1, 2. n represents the number of input samples, the output of the first hidden layer can be represented as:
wherein s (·) represents a nonlinear sigmoid function, and z is W1xi+b1Then s (z) ═ 1+ exp (-z)-1Will beThe output of the second hidden layer is obtained by inputting the data into the second hidden layer of the network:
and (3) sequentially training the greedy layer by layer, wherein the output of the kth hidden layer is as follows:
6c, optimizing the network weight W, and using a large-boundary nearest neighbor algorithm to enable the linear transformation matrix L to be equivalent to the weight W of the deep network, wherein the optimization objective function of the kth hidden layer is as follows:
wherein the content of the first and second substances,the entire optimization objective is solved by a conventional gradient descent algorithm.
The invention carries out deep network pre-training through a semi-supervised large-boundary neighbor algorithm, maps original data to a new feature space according to the thought of metric learning, and reduces the distance between similar samples and increases the distance between non-similar samples in the new feature space. The network can simultaneously learn linear and nonlinear characteristics of sample data, and the learned characteristics can be effectively used for classification of the polarized SAR images, so that the classification precision is improved.
And 7, fine adjustment of the depth distance measurement network.
After the pre-training is completed, the weight W of the network is already converged to a reasonable range, the network is further finely adjusted by using the labeled samples and the corresponding class label information thereof and combining a Softmax classifier, and the objective function phi (W) of the fine adjustment part can be expressed as:
where the former term is the mean square error term and the latter term is the weight attenuation term, which aims to reduce the magnitude of the weights and prevent over-fitting, yiRepresenting a training sample xiCorresponding class label, y (x)i) Will have a marked sample xiIn the prediction category obtained after the depth distance measurement network, β -3 e-3 is a weight attenuation parameter, and Φ (W) can be solved according to a gradient descent algorithm.
And 8, performing class prediction on the unmarked samples, sending the unmarked samples into a depth distance measurement network, and predicting class labels of the unmarked samples by using a Softmax classifier to obtain the prediction class of each unmarked sample.
7a, all the unmarked samples xuInputting the data into a constructed depth distance measurement network to obtain the final excitation output characteristicsNamely the output characteristics of the third hidden layer of the network;
7b, sending the features learned in the step 7a to a Softmax classifier for class prediction:
the output of the Softmax classifier is y ∈ RP×1P is expressed as the number of classes, unlabeled sample xuThe prediction categories of (a) may be expressed as:
where α is the sample xuThe prediction class confidence of.
And 9, outputting a classification result graph and classification precision of the polarized SAR image to be classified, outputting a final classification result of the polarized SAR image to be classified according to the labeled sample and the predicted type of unlabeled sample, and calculating the precision of the classification.
And 9a, coloring each pixel point by using red, green and blue as three primary colors according to the class label of the marked sample and the prediction class label of the unmarked sample according to a three-primary-color coloring method, and outputting a result graph, which is shown in a figure 2 (f).
And 9b, comparing the predicted class labels of the unmarked samples with the real class labels, and obtaining the classification precision by referring to the table 1.
The depth distance measurement network can simultaneously learn the linear and nonlinear characteristics of the sample, and improves the precision of the polarized SAR image classification.
The technical effects of the invention are explained in detail below with the combination of simulation experiments:
example 7
The classification method of the polarized SAR image based on the semi-supervised depth distance measurement network is the same as the embodiment 1-6,
the experimental conditions are as follows:
the hardware platform is as follows: intel (R) core (TM) i5-2410M CPU @2.30GHz and RAM 4.00 GB;
the software platform is as follows: MATLAB R2016 a;
the polarization SAR image in the Flevoland area of the Netherlands is selected for testing, the size of the image is 300 multiplied by 270, the dimension of each sample point is 6 dimensions, and the number of categories is 6. In the experiment, 1% of samples of each type were randomly selected as labeled samples, and the rest were unlabeled samples.
Experimental contents and results:
the invention combines a Softmax classifier to classify real polarized SAR images, and compares the classified real polarized SAR images with other deep learning methods on the premise of the same experimental setting, wherein WDSN is a deep network model with higher training speed, and FIG. 2(c) is a result graph of classifying FIG. 2(a) by WDSN; the second comparison method is WDBN, which is to apply the Wishart distance to the conventional deep belief network, and fig. 2(d) is a result diagram of classifying fig. 2(a) by the WDBN method; the DSFN is a deep sparse filtering network, which is a deep network model obtained by expanding the sparse filtering, and fig. 2(e) is a result graph of the DSFN classifying fig. 2 (a); FIG. 2(f) is a graph of the results of the SDMLN classification of FIG. 2(a) according to the method of the present invention. Table 1 shows the feature classification accuracy and the total classification accuracy of the polarized SAR images obtained by the above 4 methods, respectively.
TABLE 1 land feature Classification precision (%) and Total Classification precision (%) on Fleviland area charts for various methods
As can be seen from Table 1, the highest classification accuracy was found in the Bare soil, potato, while and barley classes of the present invention in the case of all the marked samples being 1%, and the total classification accuracy was 97.35% higher than that of the other comparative methods. The invention adopts the distance measurement method, has better classification precision for the block data with more perfect distribution, can learn the linear and nonlinear characteristics of the sample data, and keeps the neighbor relation of the neighbor samples, which promotes the invention to have higher classification accuracy.
Example 8
The classification method of the polarized SAR image based on the semi-supervised depth distance measurement network is the same as the embodiments 1-6, the simulation condition and the simulation content are the same as the embodiment 7,
referring to the simulation experiment results fig. 2, fig. 2(c) to fig. 2(f) are graphs of classification results obtained in the case of only 1% of labeled samples, in contrast to which the present invention has better visibility. Comparing the methods with the label graph, it is obvious that the final classification result of the present invention is better, fig. 2(f) has less noise points, each noise point represents a sample point with a classification error, which indicates that the method of the present invention has the higher classification accuracy, and it can also be seen from table 1. Fig. 2(c) is a diagram of the experimental results of WDSN, and the algorithm focuses on training speed, so that more noise appears with fewer labeled samples, and a poor classification result is presented. Fig. 2(d) and fig. 2(e) are the novel deep network models at present, and the visibility of the classification result is also significantly lower than that of the method of the present invention in the case of only 1% of labeled samples. The method combines the distance measurement and the deep learning method by using a semi-supervised method, and the learned characteristics are beneficial to classifying the polarized SAR images, so that the classification precision can be relatively improved under the condition of less labeled samples.
Referring to fig. 3, fig. 3 is a schematic diagram of the semi-supervised large-boundary nearest neighbor algorithm of the present invention. The triangles and the circles in the graph respectively represent two different classes of samples, wherein the samples in each class comprise marked samples and part of unmarked samples, the marked samples are represented by colors, the unmarked samples are represented by white, and the similar samples in the graph are in a close-proximity relationship with each other. The semi-supervised large-boundary nearest neighbor algorithm enables sample data to be mapped to a new feature space from one feature space under the combined action of distance measurement and a popular learning regular term. In the mapping process, the distance measurement method enables the distance between the similar marked samples to be reduced, and the distance between the non-similar marked samples to be increased. And the distance between the marked sample and part of unmarked neighbor samples is reduced by the popular learning regular term, so that part of unmarked information is fully utilized, the information waste is avoided, and the requirement of the algorithm on the marked sample is reduced. The semi-supervised large-boundary nearest neighbor algorithm finally properly separates the similar samples from the non-similar samples, the mapped samples are classified more easily under the action of the classifier, and the classification precision can be effectively improved.
The method for classifying the polarized SAR image based on the semi-supervised depth distance measurement network combines the distance measurement learning method and the deep learning method, can effectively extract the linear and nonlinear characteristics of sample data, solves the problem that the traditional deep learning method has higher demand on marked samples, and can also properly improve the classification precision of the polarized SAR image under the condition of only few marked samples.
In summary, the invention discloses a polarized SAR image classification method based on a semi-supervised depth distance measurement network. Solves the technical problems that the traditional deep learning only considers the nonlinear relation of the sample characteristics and the classification precision is not high when the marked samples are fewThe surgical problem comprises the following steps: 1. inputting polarized SAR image data to be classified, and generating a sample matrix X by a coherent matrix T of the polarized SAR image; 2. selecting marked samples and unmarked samples, and randomly selecting 1% of samples in each class as marked samples and the rest as unmarked samples according to sample data X and a label matrix Y of the polarized SAR image; 3. obtaining the adjacent samples with the marked samples, and obtaining the K of each marked sample in all the sample data1Marked Wishart homogeneous neighbor sample and K2A sample of unmarked Wishart neighbors; 4. constructing a loss function of a semi-supervised large-boundary neighbor algorithm, adding a popular learning regular term on the basis of the large-boundary neighbor algorithm, and improving the supervised large-boundary neighbor algorithm into the semi-supervised large-boundary neighbor algorithm to obtain the loss function of the semi-supervised large-boundary neighbor algorithm; 5. initializing parameters of a depth distance measurement network, randomly initializing a weight parameter W and a bias unit b of the depth distance measurement network, and setting the number of nodes of each layer of the depth distance measurement network; 6. pre-training a depth distance measurement network, sending a marked sample and a Wishart neighbor sample corresponding to the marked sample into the depth distance measurement network for pre-training, utilizing a semi-supervised large-boundary neighbor algorithm and adopting a layer-by-layer greedy pre-training method, taking the output of the previous layer as the input of the next layer until the last hidden layer is trained, and optimizing the weight of the network; 7. fine-tuning the depth distance measurement network, and using the labeled samples and the label information thereof in combination with a Softmax classifier to fine-tune the depth distance measurement network so as to further optimize the weight of the network; 8. performing category prediction on the unmarked samples; and outputting a classification result graph and classification precision of the polarized SAR image to be classified.
According to the invention, a semi-supervised large-boundary neighbor algorithm is obtained by constructing a deep distance measurement network model and adding a popular learning regular term method into the large-boundary neighbor algorithm, deep distance measurement learning is carried out by using a deep learning method, linear and nonlinear structures of samples can be effectively described, the deep network obtained by final training has more excellent feature extraction performance than the traditional deep network, and features extracted by the deep network are easy to meet the characteristic that the distance between similar samples is smaller than the distance between non-similar samples, so that the final classification precision can be effectively improved under the action of a classifier. The invention overcomes the problems that the classification precision is influenced due to the shortage of marked samples and the information waste of a large number of unmarked samples is caused, the learned characteristics fully describe the intrinsic attributes of the samples, and the method can be used in the technical fields of environmental monitoring, earth resource survey, military systems and the like.
Claims (4)
1. A polarized SAR image classification method based on a semi-supervised depth distance measurement network is characterized by comprising the following steps:
(1) inputting polarized SAR image data to be classified: namely a coherent matrix T of the polarized SAR image, a label matrix Y is obtained according to the ground feature distribution information of the polarized SAR image, the distribution of the same ground feature is represented by the same class label, the ground feature distribution which can not determine the class is represented by 0 in the label matrix Y, and a sample matrix is generated according to the coherent matrix T of the polarized SAR imageN is the total number of samples, xmRepresents the m-th sample;
(2) selecting labeled and unlabeled samples: according to sample data X and a label matrix Y of the polarized SAR image, randomly selecting 1% of samples in each class as marked samples, and selecting the rest of samples as unmarked samples;
(3) finding neighbor samples with labeled samples: in all sample data, K of each marked sample is obtained1Marked Wishart homogeneous neighbor sample and K2A sample of unmarked Wishart neighbors;
(4) constructing a loss function of a semi-supervised large-boundary nearest neighbor algorithm: increasing a popular learning regular term on the basis of a large-boundary neighbor algorithm, and improving the supervised large-boundary neighbor algorithm into a semi-supervised large-boundary neighbor algorithm to obtain a loss function of the semi-supervised large-boundary neighbor algorithm; the process of constructing the loss function of the semi-supervised large-boundary nearest neighbor algorithm comprises the following steps:
adding flows in a loss function of a large-boundary nearest neighbor algorithmThe method comprises the following steps of learning a regular term, wherein the use of the regular term increases the utilization of part of unmarked samples, and a supervised large-boundary nearest neighbor algorithm is improved into a semi-supervised learning method, and the popular learning regular term JRIs aimed at penalizing the marked sample xiAnd a label-free sample xpWith large spacing therebetween, i.e. with a reduction in the marked sample xiAnd a label-free sample xpThe interval therebetween:
wherein, the symbolRepresenting marked samples xiSample x with its unmarked neighborspSimilarity between the two, the loss function epsilon of the semi-supervised large-boundary nearest neighbor algorithm2(L) is:
wherein | · | purple sweetFIs a Frobenius norm used to ensure a maximum boundary, λ is a Frobenius norm regularized term coefficient, and is usually taken to be 1, γ is a semi-supervised regularized term parameter, yilRepresenting marked samples xiAnd marked sample xlClass relationship of (1), yil1 if and only if yi＝ylIs true, i.e. xiAnd xlIs a homogeneous sample, otherwise, yil＝0；[z]+Max (z,0) is the standard hinge function; loss parameter μ e [0,1 ∈ ]]The sign xiijl＝[1+||L(xi-xj)||2-||L(xi-xl)||2]+；
(5) Basic parameters for initializing the deep network: randomly initializing a weight parameter W and a bias unit b of the depth distance measurement network, setting the number of nodes of each layer of the depth distance measurement network, and determining the overall structure of the depth distance measurement network;
(6) pre-training a depth distance measurement network: sending the marked sample and the Wishart neighbor sample corresponding to the marked sample into a depth distance measurement network for pre-training, using a semi-supervised large-boundary neighbor algorithm and adopting a layer-by-layer greedy pre-training method, taking the output of the previous layer as the input of the next layer until the last hidden layer is trained, optimizing the weight of the network, and primarily optimizing the weight of the depth distance measurement network;
(7) fine tuning the depth distance measurement network: utilizing the labeled samples and the label information thereof, and combining a Softmax classifier to finely adjust the depth distance measurement network, further optimizing the weight of the network, enabling the network to be more stable, and completing the optimization of the depth distance measurement network;
(8) and (3) performing class prediction on the unmarked sample: sending the unmarked samples into a depth distance measurement network, and predicting the class labels of the unmarked samples by using a Softmax classifier to obtain the prediction class of each unmarked sample;
(9) outputting a classification result graph and classification precision of the polarized SAR image to be classified: and outputting a final classification result of the polarized SAR image to be classified and calculating the precision of the classification according to the labeled sample and the non-labeled sample with the predicted category.
2. The semi-supervised depth distance metric network-based polarimetric SAR image classification method as claimed in claim 1, wherein the step (3) of finding Wishart neighbor samples of each labeled sample comprises the following steps:
3a, a matrix of marked samples ofl represents the number of marked samples, and the Wishart distance between each marked sample and the rest of samples is obtained by the following formula:
d(xi,xg)＝ln((xi)-1xg)+Tr((xg)-1xi)-q，
where Tr () represents the trace of the matrix and the remaining samples are represented as xgI.e. byg is the serial number of the rest samples, and for the radar with integrated transmission and reception, the constant q is 3 due to reciprocity; for a radar whose transmission and reception are not integral, the constant q is 4;
3b, utilizing the sort function in MATLAB to calculate the Wishart distance d (x)i,xj) Arranged in ascending order of absolute value, K is taken before1Marked similar neighbor sample xj(j＝1,2,…,K1)，K2An unlabeled neighbor sample xp(p＝1,2,…,K2) As marked sample xiThe Wishart neighbor samples.
3. The semi-supervised depth distance metric network-based polarimetric SAR image classification method according to claim 1, wherein the parameters of the initialized depth network in the step (5) are as follows:
5a, the number of hidden layers of the depth distance measurement network is 3, and the number of nodes of each layer is respectively as follows: 150, 100, 50;
5b, randomly initializing a weight parameter W and a bias unit b of the deep sparse filter network, wherein the number of nodes of each hidden layer is Nk，t is the dimension of the input signal, N1Is the number of nodes in the first hidden layer of the network,representing a weight matrix between the kth hidden layer and the (k-1) th hidden layer of the network,the bias unit of the k hidden layer.
4. The semi-supervised depth distance measurement network-based polarimetric SAR image classification method according to claim 1, wherein the pre-training process for the depth distance measurement network in step (6) is as follows:
6a, input depth is thinPre-training samples of sparse filter network, and combining unmarked samples xgAnd corresponding marked and unmarked Wishart neighbor samples are used as pre-training samples to be input into the deep network, wherein g is 1, 2.
6b, setting xe∈Rt×1If the input vector is, e is 1,2,., n, n represents the number of pre-training input samples, and t is the dimension of the input signal, the output of the first hidden layer can be represented as:
wherein s (-) represents a non-linear sigmoid function, N1B is the node number of the first hidden layer of the network, and b is a bias unit; let z be W1xe+b1Then s (z) ═ 1+ exp (-z)-1Will beThe output of the second hidden layer is obtained by inputting the data into the second hidden layer of the network:
wherein N is2The number of nodes of the second hidden layer of the network; and (3) sequentially training the greedy layer by layer, wherein the output of the kth hidden layer is as follows:
wherein N iskThe number of nodes of the kth hidden layer of the network;
6c, optimizing the network weight W, and using a large-boundary nearest neighbor algorithm to enable the linear transformation matrix L to be equivalent to the weight W of the deep network, wherein the optimization objective function of the kth hidden layer is as follows:
wherein j is a labeled homogeneous neighbor sample,the entire optimization objective is solved by a conventional gradient descent algorithm.
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 CN107194336A (en) | 2017-09-22 |
| CN107194336B true 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) |
Families Citing this family (16)
| 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 |
| 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 |
| CN107944479B (en) * | 2017-11-16 | 2020-10-30 | 哈尔滨工业大学 | Disease prediction model establishing method 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 |
| 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 |
| 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 |
| CN109214386B (en) * | 2018-09-14 | 2020-11-24 | 京东数字科技控股有限公司 | 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 |
| US10980030B2 (en) * | 2019-03-29 | 2021-04-13 | Huawei Technologies Co., Ltd. | Method and apparatus for wireless communication using polarization-based signal space mapping |
| 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 |
| 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 |
| 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 |
| CN112257862B (en) * | 2020-09-30 | 2023-12-05 | 重庆大学 | Semi-supervised identification method based on relation network marking sample expansion |
| 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 |
Family Cites Families (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 |
| CN102208031B (en) * | 2011-06-17 | 2013-04-03 | 西安电子科技大学 | Freeman decomposition and homo-polarization rate-based polarized synthetic aperture radar (SAR) image classification method |
| CN103839073B (en) * | 2014-02-18 | 2017-02-08 | 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
Also Published As
| Publication number | Publication date |
|---|---|
| CN107194336A (en) | 2017-09-22 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107194336B (en) | Polarized SAR image classification method based on semi-supervised depth distance measurement network | |
| WO2021000903A1 (en) | End-to-end sar image recognition method and apparatus, and storage medium | |
| CN110135267B (en) | Large-scene SAR image fine target detection method | |
| CN106909924B (en) | Remote sensing image rapid retrieval method based on depth significance | |
| CN108038445B (en) | SAR automatic target identification method based on multi-view deep learning framework | |
| CN107145830B (en) | Hyperspectral image classification method based on spatial information enhancing and deepness belief network | |
| CN112150821B (en) | Lightweight vehicle detection model construction method, system and device | |
| CN113378632A (en) | Unsupervised domain pedestrian re-identification algorithm based on pseudo label optimization | |
| CN108052966A (en) | Remote sensing images scene based on convolutional neural networks automatically extracts and sorting technique | |
| CN112347970B (en) | Remote sensing image ground object identification method based on graph convolution neural network | |
| CN112990342B (en) | Semi-supervised SAR target recognition method | |
| CN110633708A (en) | Deep network significance detection method based on global model and local optimization | |
| CN106408030A (en) | SAR image classification method based on middle lamella semantic attribute and convolution neural network | |
| CN106023257A (en) | Target tracking method based on rotor UAV platform | |
| CN105069478A (en) | Hyperspectral remote sensing surface feature classification method based on superpixel-tensor sparse coding | |
| CN110414616B (en) | Remote sensing image dictionary learning and classifying method utilizing spatial relationship | |
| CN110287985B (en) | Depth neural network image identification method based on variable topology structure with variation particle swarm optimization | |
| CN114332578A (en) | Image anomaly detection model training method, image anomaly detection method and device | |
| CN114926693A (en) | SAR image small sample identification method and device based on weighted distance | |
| CN107423771B (en) | Two-time-phase remote sensing image change detection method | |
| CN116933141B (en) | Multispectral laser radar point cloud classification method based on multicore graph learning | |
| Wang et al. | Classification and extent determination of rock slope using deep learning | |
| CN115761240B (en) | Image semantic segmentation method and device for chaotic back propagation graph neural network | |
| CN113066537B (en) | Compound classification method based on graph neural network | |
| CN113139556B (en) | Manifold multi-view image clustering method and system based on self-adaptive composition |
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 |