CN113095416A - Small sample SAR target classification method based on mixed loss and graph attention - Google Patents
Small sample SAR target classification method based on mixed loss and graph attention Download PDFInfo
- Publication number
- CN113095416A CN113095416A CN202110408623.2A CN202110408623A CN113095416A CN 113095416 A CN113095416 A CN 113095416A CN 202110408623 A CN202110408623 A CN 202110408623A CN 113095416 A CN113095416 A CN 113095416A
- Authority
- CN
- China
- Prior art keywords
- training
- test
- layer
- node
- vector
- 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
- 238000000034 method Methods 0.000 title claims abstract description 38
- 238000012549 training Methods 0.000 claims abstract description 308
- 238000012360 testing method Methods 0.000 claims abstract description 191
- 230000002708 enhancing effect Effects 0.000 claims abstract description 5
- 239000013598 vector Substances 0.000 claims description 168
- 230000009466 transformation Effects 0.000 claims description 22
- 238000013507 mapping Methods 0.000 claims description 8
- 210000002569 neuron Anatomy 0.000 claims description 8
- 238000003780 insertion Methods 0.000 claims description 7
- 230000037431 insertion Effects 0.000 claims description 7
- 238000011176 pooling Methods 0.000 claims description 7
- 101100149325 Escherichia coli (strain K12) setC gene Proteins 0.000 claims description 6
- 238000004364 calculation method Methods 0.000 claims description 6
- 238000010606 normalization Methods 0.000 claims description 6
- 230000007306 turnover Effects 0.000 claims description 6
- 238000004422 calculation algorithm Methods 0.000 claims description 5
- 238000005259 measurement Methods 0.000 claims description 5
- 238000012935 Averaging Methods 0.000 claims description 3
- 102100035115 Testin Human genes 0.000 claims description 3
- 101710070533 Testin Proteins 0.000 claims description 3
- 230000004913 activation Effects 0.000 claims description 3
- 238000010276 construction Methods 0.000 claims 1
- 230000008569 process Effects 0.000 abstract description 11
- 230000006870 function Effects 0.000 description 6
- 238000004088 simulation Methods 0.000 description 4
- 238000013135 deep learning Methods 0.000 description 3
- 230000000694 effects Effects 0.000 description 3
- 230000007547 defect Effects 0.000 description 2
- 239000000284 extract Substances 0.000 description 2
- 230000007246 mechanism Effects 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 230000010287 polarization Effects 0.000 description 2
- 238000012545 processing Methods 0.000 description 2
- 108010049931 Bone Morphogenetic Protein 2 Proteins 0.000 description 1
- 102100024506 Bone morphogenetic protein 2 Human genes 0.000 description 1
- 230000002776 aggregation Effects 0.000 description 1
- 238000004220 aggregation Methods 0.000 description 1
- 230000008901 benefit Effects 0.000 description 1
- 239000011159 matrix material Substances 0.000 description 1
- 230000009467 reduction Effects 0.000 description 1
Images
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
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Molecular Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Computational Biology (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- General Health & Medical Sciences (AREA)
- Evolutionary Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Probability & Statistics with Applications (AREA)
- Image Analysis (AREA)
- Radar Systems Or Details Thereof (AREA)
Abstract
The invention provides a small sample SAR target classification method based on mixed loss and graph attention, which comprises the following steps: acquiring a training sample set and a test sample set; constructing a network model H based on mixed loss and graph attention; performing iterative training on the H; and obtaining a target classification result of the small sample SAR image. The invention passes the classification loss value l of the training task setCAnd the embedding loss value l of the training task setEThe weighting and the mixed loss value l forming the training task set update the parameters of all the first convolution layers and all the second convolution layers in the embedded network module E and the parameters of all the first full-connection layers and all the second full-connection layers in the graph attention network module G, thereby enhancing the similarity between the characteristics of the same SAR target class and the difference between the characteristics of different SAR target classes, and the method can also improve the stability of the SAR target classesData enhancement effectively reduces the risk of overfitting in the model training process, and improves the classification precision of the small sample SAR target.
Description
Technical Field
The invention belongs to the technical field of radar image processing, relates to an SAR target classification method, and particularly relates to a small sample SAR target classification method based on mixed loss and drawing attention, which can be used for SAR target classification under the condition that the number of SAR images acquired by a target is small.
Background
The synthetic aperture radar SAR has the characteristics of all-time, all-weather, long range, high resolution and the like, and is widely used for target classification in military fields such as battlefield reconnaissance and the like because the two-dimensional high-resolution SAR image of the target contains rich information such as the shape, the size, the texture and the like of the target. The SAR target classification is an algorithm which is based on a computer system, extracts features after SAR image data of a target is obtained from a sensor, and gives target class attributes according to the extracted features. Although a large number of traditional SAR target classification methods based on manually selected features and designed classifiers have been generated, these traditional methods require setting a specific algorithm for a specific target according to a large amount of experience and strong professional knowledge, are time-consuming and difficult to popularize. In recent years, the SAR target classification method based on deep learning realizes data-driven SAR target classification, and the method can autonomously learn and extract the characteristics effective for classification from data and classify the target by using the characteristics, so that characteristics do not need to be manually selected, a classifier is not designed, professional knowledge is not needed to be stronger, and the method is easy to popularize into a new target class, so that excellent performance is obtained, and the method is widely researched and used by the industry.
However, some targets observed by the SAR are non-cooperative small-sample SAR targets, that is, the number of SAR images that can be acquired by these targets is small, and each target has only 1 image to ten or more images, whereas the SAR target classification method based on deep learning generally requires a large number of training samples to train a model to obtain a high classification accuracy on a test sample, and for small-sample SAR targets, the SAR target classification method based on deep learning has a problem of low classification accuracy due to the shortage of the training samples.
In order to solve the problem, in the prior art, a special model with low requirement on the number of samples is designed by improving a model structure to improve the classification accuracy of the small-sample SAR target. For example, a patent application with a publication number of CN111191718A entitled "small sample SAR target recognition method based on graph attention network" discloses a small sample SAR target recognition method based on graph attention network, which first obtains a small number of labeled SAR images and a large number of unlabeled SAR images of a target and performs noise reduction processing, then iteratively trains a self-encoder by using the denoised images to obtain feature vectors of all the SAR images, finally constructs an initial adjacency moment, and iteratively trains a graph attention network by using the initial adjacency moment and the feature vectors of all the SAR images, and the trained graph attention network can realize classification of the target in the unlabeled SAR images by using an attention mechanism. The graph attention network adopted by the method has less requirements on the SAR images with the labels in the class prediction process, and the application of the attention mechanism can improve the classification accuracy of the network, but the method has the defects that the obtained result is independently calculated and lost with the original SAR images after the self-encoder performs encoding and decoding operation on each SAR image, so that the extracted features of the same SAR target class have lower similarity, the difference between the features of different SAR target classes is weaker, and the classification accuracy of the model on the small-sample SAR target is still lower.
Disclosure of Invention
The invention aims to overcome the defects in the prior art, and provides a small sample SAR target classification method based on mixed loss and drawing attention, which is used for solving the technical problems of low classification accuracy caused by low similarity between the characteristics of the same SAR target category and weak difference between the characteristics of different SAR target categories in the prior art.
In order to achieve the purpose, the technical scheme adopted by the invention comprises the following steps:
(1) obtaining a training sample set
And test sample set
(1a) Acquiring a plurality of Synthetic Aperture Radar (SAR) images containing C different target categories, wherein each target category corresponds to M SAR images with the size of h multiplied by h, each SAR image contains 1 target, C is more than or equal to 10, M is more than or equal to 200, and h is 128;
(1b) marking the target category in each SAR image, and randomly selecting CtrainTotal of C for each object classtrainTaking x M SAR images and labels of each SAR image as training sample sets
Mixing the rest CtestTotal of C for each object classtestTaking x M SAR images and labels of each SAR image as test sample sets
wherein Ctrain+Ctest＝C，3≤Ctest≤5；
(2) Constructing a network model H based on mixed loss and graph attention:
constructing a mixed loss and graph attention-based network model H comprising a data enhancement module D, an embedded network module E, a node feature initialization module I and a graph attention network module G which are sequentially cascaded, wherein the embedded network module E comprises a plurality of first volume modules E which are sequentially cascadedCAnd a second convolution module ELEach first convolution module ECComprises a first convolution layer, a first batch of normalization layers, a Mish activation layer, a maximum pooling layer and a second convolution module E which are sequentially stackedLThe second convolution layer and the second batch of normalization layer are sequentially stacked; the graph attention network module G comprises a plurality of graph updating layers U which are sequentially cascaded, and each graph updating layer U comprises edge feature building modules U which are sequentially stackedEAttention weight calculation module UWAnd node feature update module UNEach edge feature building block UEComprises a plurality of first full-connection layers which are sequentially stacked, and each nodeFeature update module UNA second full-link layer is included;
(3) performing iterative training on a network model H based on mixed loss and graph attention:
(3a) initializing the iteration number to be N, wherein the maximum iteration number is N, N is more than or equal to 1000, and N is set to be 0;
(3b) from a training sample set
Randomly selecting a group containing CtestTotal of C for each object classtestMultiplying the SAR images by x M, and performing one-hot coding on the label of each SAR image to obtain C of each SAR imagetestDimension label vector, then from CtestRandomly selecting K SAR images contained in each target category and corresponding label vectors from the xM SAR images as training support sample sets
C is remainedtest(M-K) SAR images and corresponding label vectors serving as training query sample sets
After one-hot coding, the C-dimensional element in the tag vector of each SAR image indicates that the target in the SAR image belongs to CtestThe probability of the c-th object class of the object classes,
representing the a-th training support sample consisting of the SAR image and its corresponding label vector,
representing a b-th training query sample consisting of the SAR image and the corresponding label vector, wherein K is more than or equal to 1 and less than or equal to 10;
(3c) will train and support the sample set
With each training query sample
Combined into a training task
Obtaining a training task set
And will be
Forward propagation as input to the mixed-loss and graph attention based network model H:
(3c1) data enhancement module D pairs training task set
Performing data enhancement on each SAR image: performing power transformation on each SAR image, adding noise to the SAR image subjected to power transformation, performing turnover transformation on the SAR image subjected to noise addition, and performing rotation transformation on the SAR image subjected to turnover transformation to obtain an enhanced training task set
wherein ,
representing training tasks
A corresponding one of the enhanced training tasks is,
representing training tasks
Training support sample in (1)
The corresponding reinforced training support sample is used for training,
representing training query samples
Corresponding enhanced training query samples;
(3c2) embedded network module E pair enhanced training task set
Each of the enhanced training tasks in (1)
Mapping each SAR image to obtain a training embedded vector set
And using an embedded loss function LEBy training sets of embedded vectors
Computing training task set
Value of insertion loss lE：
wherein ,
representing an enhanced training task
Corresponding training embedded vector set, satisfies a ≠ CtestOf K +1
Presentation of enhanced training support samples
The corresponding training is embedded into the vector(s),
representing enhanced training query samples
Corresponding training embedding vectors, log (-) representing the logarithm based on the natural constant e, exp (-) representing the natural constanteA base exponent, Σ denotes a continuous sum,
representing a training task
Training support sample set in (1)
Each training embedded vector corresponding to each SAR image of the included c-th target class
The class center of the c-th object class obtained by averaging,
representing and training tasks
Training query sample in (1)
Targets in the contained SAR images belong to the class center of the same target class, d represents a measurement function, and d (p, q) | | p-q | | luminance2；
(3c3) Node characteristic initialization module I constructs a virtual label vector
And embedding a set of vectors for each training
In the condition that a is not equal to CtestEach training embedded vector of K +1
Splicing with the label vector of the corresponding SAR image, and simultaneously embedding each training vector group
Training embedded vector in
And virtual tag vector
Splicing to obtain a training node 1-layer feature group set
wherein ,
is represented by CtestA vector in which the element values of each dimension are all 1,
representing training embedded vector sets
The corresponding training node layer 1 feature set,
representing training embedded vectors
Corresponding training node 1 layer characteristics;
(3c4) graph attention network module G trains node layer 1 feature group set through input
To pair
Including each training node layer 1 feature set
Training node level 1 features in
Corresponding training query samples
Carrying out category prediction on targets in the SAR image to obtain a training prediction result vector set
wherein ,
to representTraining node layer 1 features
Corresponding dimension is CtestThe c-th element represents the 1-level feature of the training node
Corresponding training query samples
The prediction probability that the target in the included SAR image belongs to the c-th target class;
(3c5) using a classification loss function LCPredicting the result vector set by training
And training the query sample set
All the label vectors in (1) calculate the training task set
Is a classification loss value lC：
wherein ,
predicting outcome vectors for training
Value of the c-th element in (y)b,cPredicting outcome vectors for training
The value of the c-dimension element in the label vector of the corresponding SAR image;
(3d) for the training task set
Is a classification loss value lCAnd training task set
Value of insertion loss lEObtaining a training task set by calculating a weighted sum
Mixing loss value l, l ═ λ lC+(1-λ)lEThen, updating parameters of all first convolution layers and all second convolution layers embedded in the network module E and parameters of all first full-connection layers and all second full-connection layers in the attention network module G through a mixed loss value l by using a random gradient descent algorithm, wherein lambda is weight, and lambda is more than or equal to 0.7 and less than 1;
(3e) judging whether N is greater than or equal to N, if so, obtaining a trained network model H' based on the mixing loss and the graph attention, otherwise, enabling N to be N +1, and executing the step (3 b);
(4) obtaining a target classification result of the small sample SAR image:
(4a) for test sample set
Carrying out one-hot coding on the label of each SAR image to obtain C of each SAR imagetestDimension label vectors, then from the test sample set
C of (A)testRandomly selecting K SAR images contained in each target category and corresponding label vectors from the xM SAR images as a test support sample set
C is remainedtest(M-K) SAR images and corresponding label vectors serving as test query sample sets
wherein ,
representing the e-th test support sample consisting of the SAR image and its corresponding tag vector,
representing a g-th test query sample consisting of the SAR image and the corresponding label vector;
(4b) supporting the test with a sample set
With each test query sample
Combined into test tasks
Obtaining a set of test tasks
And will be
Forward propagation as input to the trained mixed-loss and attention-based network model H':
(4b1) trained embedded network module E' pair test task set
Each test task in (1)
Mapping each SAR image to obtain a test embedded vector group set
wherein ,
representing test tasks
Corresponding test embedded vector groups satisfy e ≠ CtestOf K +1
Presentation of test support samples
The corresponding test-embedded vector is inserted into the vector,
representing test query samples
A corresponding test embedding vector;
(4b2) node characteristic initialization module I constructs a virtual label vector
And embedding the vector set for each test
In the condition that e ≠ CtestEach test embedding vector of K +1
Splicing with the label vector of the corresponding SAR image, and simultaneously embedding each test into a vector group
Test embedded vector in
And virtual tag vector
Splicing to obtain a 1-layer feature group set of the test nodes
wherein ,
representing test embedding vector sets
The corresponding test point 1 layer feature set,
table test embedded vector
Corresponding test node 1 layer characteristics;
(4b3) the trained graph attention network module G' tests the node layer 1 feature group set through input
To pair
Including each test node a layer 1 feature set
Test node level 1 features in
Corresponding test query sample
Carrying out category prediction on targets in the SAR image to obtain a test prediction result vector set
Each test prediction vector
The dimension number corresponding to the medium maximum value is
Corresponding test query sample
Including a prediction category of the target in the SAR image, wherein,
representing test node level 1 features
Corresponding dimension is CtestThe element value of the c-th dimension represents the test node level 1 feature
Corresponding test query sample
The probability that the object in the included SAR image belongs to the c-th object class.
Compared with the prior art, the invention has the following advantages:
1. the invention passes the classification loss value l of the training task setCAnd the embedding loss value l of the training task setEThe mixed loss value l of the training task set formed by the weighted sum updates the parameters of all the first convolution layers and all the second convolution layers in the embedded network module E and the parameters of all the first full-connection layers and all the second full-connection layers in the graph attention network module G, thereby enhancing the similarity between the characteristics of the same SAR target class and the difference between the characteristics of different SAR target classes.
2. In the training process of the network model H based on the mixed loss and the attention of the graph, the data enhancement module D can effectively relieve the over-fitting risk of the model by enhancing the data of all SAR images; and the data enhancement module D, the embedded network module E and the node characteristic initialization module I can acquire the effective characteristics of each SAR image, the effective characteristics of each SAR image are combined with the corresponding label vector, data support can be provided for the class prediction of the attention network module G, and compared with the prior art, the classification precision of the small-sample SAR target is further improved.
Experimental results show that the method can obtain higher classification accuracy in small sample SAR target classification.
Drawings
FIG. 1 is a flow chart of an implementation of the present invention.
FIG. 2 is a flow chart of an implementation of the present invention for iterative training of a network model H based on mixed loss and graph attention.
Fig. 3 is a flow chart of an implementation of the present invention to obtain a target classification result of a small sample SAR image.
Detailed Description
The invention is described in further detail below with reference to the figures and the specific embodiments.
Referring to fig. 1, the present invention includes the steps of:
step 1) obtaining a training sample set
And test sample set
(1a) Acquiring a plurality of Synthetic Aperture Radar (SAR) images containing C different target categories, wherein each target category corresponds to M SAR images with the size of h multiplied by h, each SAR image contains 1 target, C is more than or equal to 10, M is more than or equal to 200, and h is 128;
(1b) marking the target category in each SAR image, and randomly selecting CtrainTotal of C for each object classtrainTaking x M SAR images and labels of each SAR image as training sample sets
Mixing the rest CtestTotal of C for each object classtestTaking x M SAR images and labels of each SAR image as test sample sets
wherein Ctrain+Ctest＝C，3≤Ctest≤5；
The invention trains the model by adopting the target class containing sufficient SAR images, and the trained model can have higher classification precision in the small sample SAR target classification of other classes, so the training sample set constructed in the step
And test sample set
The medium SAR targets are different in category;
step 2), constructing a network model H based on mixed loss and graph attention:
constructing a mixed loss and graph attention-based network model H comprising a data enhancement module D, an embedded network module E, a node feature initialization module I and a graph attention network module G which are sequentially cascaded, wherein the embedded network module E comprises five first convolution modules E which are sequentially cascadedCAnd a second convolution module ELEach first convolution module ECComprises a first convolution layer, a first batch of normalization layers, a Mish activation layer, a maximum pooling layer and a second convolution module E which are sequentially stackedLThe second convolution layer and the second batch of normalization layer are sequentially stacked; the graph attention network module G comprises three graph updating layers U which are sequentially cascaded, and each graph updating layer U comprises edge feature building modules U which are sequentially stackedEAttention weight calculation module UWAnd node feature update module UNEach edge feature building block UEComprises five first full-connection layers which are sequentially stacked, and each node feature updating module UNA second full-link layer is included;
the embedded network module E and the attention network module G have the following specific parameters:
first volume module E included in embedded network module ECThe sizes of the convolution kernels of the five first convolution layers in the three first convolution layers are all 3 multiplied by 3, the step lengths are all 1, the filling is all 1, the number of the convolution kernels of the first three first convolution layers is 64, the number of the convolution kernels of the fourth and fifth first convolution layers is 128, the sizes of the pooling kernels of the five maximum pooling layers are all 2 multiplied by 2, and the sliding step lengths are all 2; second convolution module ELThe size of the convolution kernel of the second convolution layer in (1) is 4 × 4, and the step size is 1;
the graph attention network module G comprises an edge feature building module U in each graph updating layer UEThe number of the neurons of the first four first full connection layers is 96, and the number of the neurons of the fifth first full connection layer is 1; node feature update module U included in the first two graph update layers UNThe number of the neurons of the second full-connection layer is 24, and the third graph updating layer U comprises a node feature updating module UNThe number of neurons in the second fully-connected layer in (1) is Ctest；
Step 3) iterative training is carried out on the network model H based on the mixed loss and the drawing attention, and the implementation steps are as shown in FIG. 2:
(3a) initializing the iteration number to be N, wherein the maximum iteration number is N, N is more than or equal to 1000, and N is set to be 0;
(3b) from a training sample set
Randomly selecting a group containing CtestTotal of C for each object classtestMultiplying the SAR images by x M, and performing one-hot coding on the label of each SAR image to obtain C of each SAR imagetestDimension label vector, then from CtestRandomly selecting K SAR images contained in each target category and corresponding label vectors from the xM SAR images as training support sample sets
C is remainedtest(M-K) SAR images and corresponding label vectors serving as training query sample sets
After one-hot coding, the C-dimensional element in the tag vector of each SAR image indicates that the target in the SAR image belongs to CtestThe probability of the c-th object class of the object classes,
representing the a-th training support sample consisting of the SAR image and its corresponding label vector,
representing a b-th training query sample consisting of the SAR image and the corresponding label vector, wherein K is more than or equal to 1 and less than or equal to 10;
in order to ensure that the trained model can have higher classification precision in the classification of small sample SAR targets of other classes, the invention simulates the test process in the training process and carries out the training on a sample set
Randomly selecting SAR images with the same SAR target category number in the testing process, and dividing all selected SAR images and corresponding label vectors into training support sample sets
And training the query sample set
Two subsets, a training support sample set
Has less SAR images and only CtestK images for simulating a small number of SAR images acquired by the small sample SAR target, training the query sample set
The SAR image in (1) is used for simulating the SAR image needing classification, and at the moment, a support sample set is trained
Training a query sample set for model prediction
The category of the medium SAR image provides data support;
(3c) will train and support the sample set
With each training query sample
Combined into a training task
At this time, each training task
Are all one smallSample SAR target problem, the model needs to support the sample set by training of known classes
To predict each training query sample
Class of medium SAR image, all training tasks
The combination can obtain a training task set
And will be
Forward propagation as input to the mixed-loss and graph attention based network model H:
(3c1) data enhancement module D pairs training task set
Performing data enhancement on each SAR image: performing power transformation on each SAR image, adding noise to the SAR image subjected to power transformation, performing turnover transformation on the SAR image subjected to noise addition, and performing rotation transformation on the SAR image subjected to turnover transformation to obtain an enhanced training task set
wherein ,
representing training tasks
A corresponding one of the enhanced training tasks is,
representing training tasks
Training support sample in (1)
The corresponding reinforced training support sample is used for training,
representing training query samples
Corresponding enhanced training query samples;
the training task set may be changed during each iteration using the data enhancement module D
Can increase the training task set
The risk of model overfitting in the training process is reduced. The data enhancement module D comprises the following concrete implementation steps: performing a power transformation on each SAR image B, i.e. B1＝(B/255)γX 255 to obtain SAR image B1For each SAR image B1Making an addition of noise, i.e. B2＝B1+ noise, obtaining SAR image B2For each SAR image B2Performing a flip-flop transformation, i.e. B3＝flip(B2) Obtaining an SAR image B3For each SAR image B3Performing a rotational transformation, i.e. B' ═ rot (B)3) Obtaining an enhanced SAR imageB', where γ represents a power, and γ is randomly taken to be [0.7,1.3 ]]The value within the range, noise, means compliance [ -alpha, alpha ]]Of uniformly distributed noise, alpha is randomly taken to be (0, 50)]The value in the range, film (DEG) indicates that the left and right or up and down overturn is carried out randomly, the probability of adopting each overturning mode is 1/2, rot (DEG) indicates that the clockwise rotation is carried out randomly by 90 degrees or 180 degrees or 270 degrees, and the probability of adopting each rotation angle is 1/3;
(3c2) embedded network module E pair enhanced training task set
Each of the enhanced training tasks in (1)
Mapping each SAR image to obtain a training embedded vector set
And using an embedded loss function LEBy training sets of embedded vectors
Computing training task set
Value of insertion loss lE：
wherein ,
representing an enhanced training task
Corresponding training embedded vector set, satisfies a ≠ CtestOf K +1
Presentation of enhanced training support samples
The corresponding training is embedded into the vector(s),
representing enhanced training query samples
The corresponding training embedding vector, log (-) represents the logarithm based on the natural constant e, exp (-) represents the exponent based on the natural constant e, Σ represents the running sum,
representing a training task
Training support sample set in (1)
Each training embedded vector corresponding to each SAR image of the included c-th target class
The class center of the c-th object class obtained by averaging,
representing and training tasks
Training query sample in (1)
Targets in the contained SAR images belong to the class center of the same target class, d represents a measurement function, and d (p, q) | | p-q | | luminance2；
In this step, the process of the embedded network module E mapping the SAR image into an embedded vector is equivalent to the process of the embedded network module E mapping the SAR image into an embedded vectorFeatures of each SAR image extracted, embedding loss by for each training task
Calculating the class center of each SAR target class in the SAR target class, and connecting each class center with a training query sample
Distance measurement is carried out on corresponding embedded vectors, and the query sample is trained under the constraint of an embedded loss function
The corresponding embedded vector is closer to the class center belonging to the same class and is further away from the class center belonging to the different classes, and after multiple iterations, the embedded network module E can achieve the effects of higher similarity between the extracted features of the same SAR target class and stronger difference between the extracted features of the different SAR target classes, so that the attention network module G of the subsequent graph is improved for each training query sample
The category prediction precision of the medium SAR image;
(3c3) node characteristic initialization module I constructs a virtual label vector
And embedding a set of vectors for each training
In the condition that a is not equal to CtestEach training embedded vector of K +1
Splicing with the label vector of the corresponding SAR image, and simultaneously embedding each training vector group
Training embedded vector in
And virtual tag vector
Splicing to obtain a training node 1-layer feature group set
wherein ,
is represented by CtestA vector in which the element values of each dimension are all 1,
representing training embedded vector sets
The corresponding training node layer 1 feature set,
representing training embedded vectors
Corresponding training node 1 layer characteristics;
in this step, the sample is supported by each enhanced training
Corresponding training embedded vector
And
the label vectors in (1) are spliced, and a vector can be embedded for each training
Adding class information to obtain 1-layer characteristics of all training nodes
Each training query sample may be paired for the model
The class prediction of the medium SAR image provides data support; for each training query sample for which a prediction class is required
The corresponding training embedded vector of the SAR image is
And virtual tag vector
Stitching to ensure 1-level features per training node
Are equal in dimension;
(3c4) graph attention network module G trains node layer 1 feature group set through input
To pair
Including each training node layer 1 feature set
Training node level 1 features in
Corresponding training query samples
Carrying out category prediction on targets in the SAR image to obtain a training prediction result vector set
wherein ,
representing training node layer 1 features
Corresponding dimension is CtestThe c-th element represents the 1-level feature of the training node
Corresponding training query samples
The prediction probability that the target in the included SAR image belongs to the c-th target class;
in this step, the graph attention network module G may be assembled according to the training node layer 1 feature group
1-layer features for each training node
Corresponding training query samples
Carrying out category prediction on the SAR image; the graph attention network module G firstly trains the task
Each SAR image in the SAR image is regarded as a node, two opposite directional edges are connected between each two nodes, namely a fully-connected directed graph is formed, each node and each edge have the characteristics, and the initial characteristics of each node/the input characteristics of the 1 st graph updating layer U are the 1-layer characteristics of the training nodes
The feature is updated for multiple times through each graph updating layer U, and the ith graph updating layer U is used for inputting training nodes and i-layer features
Updated to obtain the i +1 layer characteristics of the training nodes
Features of each edge in the ith graph update layer U
The attention weight is obtained by calculating the characteristics of the corresponding nodes in the ith graph updating layer U, the characteristics of each edge correspond to one attention weight, the attention weight can be used for guiding the corresponding nodes to aggregate the characteristics from other nodes, the node characteristics of the corresponding nodes are updated accordingly, and after the characteristics of the nodes are updated for three times, the attention network module G can be used for searching the training query sample
Training node 4-layer characteristics corresponding to SAR image in (synthetic aperture radar)
Converting to its predictive label; the specific implementation steps of the whole process are as follows:
(3c41) initializing the iteration number to be i, and noting that the graph force network module G includes three graph updating layers U, so that the maximum iteration number is 3, and making i equal to 1;
(3c42) the ith graph updates the edge of layer UFeature building module UEUsing training node i-layer feature set sets
Each training node in (1) i-layer feature set
All included training node i-layer features
Constructing edge features to obtain a training edge i-layer feature group set
wherein ,
representing training node i-layer feature set
The corresponding training edge i-layer feature set,
representing training node i-layer feature set
The i-layer characteristics of the h-th training node,
representing training node i-layer feature set
The i-layer feature of the j-th training node,
representing i-layer features by training nodes
And training node i-layer features
The obtained training edge i-layer characteristic, abs (-) represents taking the absolute value of each element,
representing the edge feature building Module U included in the ith graph update layer UEA plurality of first full connection layers sequentially laminated in the above order;
(3c43) attention weight calculation module U of ith map update layer UWFor training edge i-layer feature group set
Each training edge i-layer feature set in (1)
All training edge i-layer features included
Calculating attention weight to obtain training i-layer attention weight group set
wherein ,
representing training edge i-layer feature set
Correspondingly training the i-layer attention weight group,
representing training edge i-layer features
The corresponding training i-layer attention weight,
from 1 to C representing removal of htestA set of positive integers of (M-K);
in this step, the attention weight calculation module UWCalculating attention weights among the nodes according to the characteristics of each edge, wherein the weights are normalized, namely the sum of the attention weights of one node to all other nodes is 1;
(3c44) node characteristic updating module U of ith graph updating layer UNUsing training node i-layer feature set sets
And training the i-layer attention weight set
Updating the node characteristics to obtain a training node i +1 layer characteristic group set
wherein ,
representing training node i-layer feature set
The corresponding training node i +1 level feature set,
representing training node i-layer features
Corresponding training nodes are i +1 layer characteristics, a | b represents that after the vector b is spliced in the vector a,
indicating the node characteristic update module U included by inputting n into the ith graph update layer UNThe second fully connected layer of (a);
in this step, the node characteristic update module UNBy pouringIn the process, each node can aggregate the characteristics from all other nodes, the aggregated characteristics are spliced with the characteristics of the node, and the spliced result is input into a second full-connection layer to obtain the updated characteristics, namely the characteristics of the i layer of the training node
Updating to training node i +1 layer characteristics
Node feature aggregation guided by attention weights, per training query sample
The node corresponding to the SAR image in the system can acquire the category information from the nodes of the same category, and is converted into the prediction category after being updated for multiple times;
(3c45) judging whether i is true or not, if so, collecting the obtained training node 4-layer feature group set
Each training node in (4) layer feature set
Included training node 4-level features
Performing softmax transformation to obtain a vector set of training prediction results
Otherwise, let i be i +1, and perform step (3c42), wherein,
for training node 4-level features
A corresponding training predictor vector;
(3c5) using a classification loss function LCPredicting the result vector set by training
And training the query sample set
All the label vectors in (1) calculate the training task set
Is a classification loss value lC：
wherein ,
predicting outcome vectors for training
Value of the c-th element in (y)bC is a vector of training predictors
The value of the c-dimension element in the label vector of the corresponding SAR image;
(3d) for the training task set
Is a classification loss value lCAnd training task set
Value of insertion loss lEObtaining a training task set by calculating a weighted sum
Mixing loss value l, l of＝λlC+(1-λ)lEThen, updating parameters of all first convolution layers and all second convolution layers embedded in the network module E and parameters of all first full-connection layers and all second full-connection layers in the attention network module G through a mixed loss value l by using a random gradient descent algorithm, wherein lambda is weight, and lambda is more than or equal to 0.7 and less than 1;
in this step, the classification loss value l is setCAnd an insertion loss value lEThe weighted and obtained mixed loss value l has the effects of enhancing the similarity between the characteristics of the same SAR target class and the difference between the characteristics of different SAR target classes while updating the parameters of the whole model, so that the classification precision is improved;
(3e) judging whether N is greater than or equal to N, if so, obtaining a trained network model H' based on the mixing loss and the graph attention, otherwise, enabling N to be N +1, and executing the step (3 b);
step 4) obtaining a target classification result of the small sample SAR image, wherein the implementation steps are as shown in FIG. 2:
(4a) for test sample set
Carrying out one-hot coding on the label of each SAR image to obtain C of each SAR imagetestDimension label vectors, then from the test sample set
C of (A)testRandomly selecting K SAR images contained in each target category and corresponding label vectors from the xM SAR images as a test support sample set
C is remainedtest(M-K) SAR images and corresponding label vectors serving as test query sample sets
wherein ,
representing the e-th test support sample consisting of the SAR image and its corresponding tag vector,
representing a g-th test query sample consisting of the SAR image and the corresponding label vector;
(4b) supporting the test with a sample set
With each test query sample
Combined into test tasks
Obtaining a set of test tasks
And will be
Forward propagation as input to the trained mixed-loss and attention-based network model H':
(4b1) trained embedded network module E' pair test task set
Each test task in (1)
Mapping each SAR image to obtain a test embedded vector group set
wherein ,
representing test tasks
Corresponding test embedded vector groups satisfy e ≠ CtestOf K +1
Presentation of test support samples
The corresponding test-embedded vector is inserted into the vector,
representing test query samples
A corresponding test embedding vector;
(4b2) node characteristic initialization module I constructs a virtual label vector
And embedding the vector set for each test
In the condition that e ≠ CtestEach test embedding vector of K +1
Splicing with the label vector of the corresponding SAR image, and simultaneously embedding each test into a vector group
Test embedded vector in
And virtual tag vector
Splicing to obtain a 1-layer feature group set of the test nodes
wherein ,
representing test embedding vector sets
The corresponding test point 1 layer feature set,
table test embedded vector
Corresponding test node 1 layer characteristics;
(4b3) the trained graph attention network module G' tests the node layer 1 feature group set through input
To pair
Including 1-level characteristics per test nodeGroup of
Test node level 1 features in
Corresponding test query sample
Carrying out category prediction on targets in the SAR image to obtain a test prediction result vector set
Each test prediction vector
The dimension number corresponding to the medium maximum value is
Corresponding test query sample
Including a prediction category of the target in the SAR image, wherein,
representing test node level 1 features
Corresponding dimension is CtestThe element value of the c-th dimension represents the test node level 1 feature
Corresponding test query sample
The probability that the target included in the SAR image belongs to the c-th target class;
in this step, test data is obtainedSet of measurement vectors
Similar to the step (3c4), only the input data of the trained graph attention network module G' is adjusted.
The technical effects of the present invention are further explained by combining simulation experiments as follows:
1. simulation experiment conditions and contents:
the hardware platform of the simulation experiment is as follows: the GPU is NVIDIA GeForce RTX 3090, and the software platform is as follows: the operating system is windows 10. The data set of the simulation experiment is a public MSTAR data set, wherein the SAR sensor is in a high-resolution beam-gathering type, the resolution is 0.3m multiplied by 0.3m, the SAR sensor works in an X wave band, the polarization mode is HH polarization, the pitch angles are respectively 17 degrees, the azimuth angles are continuously changed from 0 degree to 360 degrees, the interval is about 5 degrees, and the size after image cutting is 128 multiplied by 128 degrees. The consolidated MSTAR data set contains 10 classes of ground military vehicle targets, i.e., C-10, model numbers 2S1, BMP-2, BRDM-2, BTR-60, BTR-70, D-7, T-62, T-72, ZIL-131, ZSU-234, and each class of target has 210 SAR images, i.e., M-210.
In order to compare the small sample SAR target classification accuracy with the existing small sample SAR target identification method based on the graph attention network, 1050 total SAR images of 5 target categories and the label of each SAR image are selected from the MSTAR data set as a training sample set, namely CtrainSelecting 1050 total SAR images of the remaining 5 target categories and labels of each SAR image as a test sample set, namely Ctest5. Meanwhile, the number K of training/testing support samples sampled by each target class in each training/testing task is 10, and the number M-K of training/testing query samples is 200. The classification of the target classes in the training sample set and the testing sample set and the number of the SAR images of each class of target are shown in Table 1:
TABLE 1
The classification confusion matrix and the classification average accuracy of the small sample SAR target recognition method based on the graph attention network are compared and simulated, and the result is shown in table 2:
TABLE 2
As can be seen from Table 2, the average accuracy of the small sample SAR target classification of the invention is improved by 5.1% compared with the prior art.
The foregoing description is only exemplary of the invention and is not intended to limit the invention, and it will be apparent to those skilled in the art that various changes and modifications in form and detail may be made without departing from the principles and arrangements of the invention, but these changes and modifications are within the scope of the invention as defined in the appended claims.
Claims (4)
1. A small sample SAR target classification method based on mixed loss and graph attention is characterized by comprising the following steps:
(1) obtaining a training sample set
And test sample set
(1a) Acquiring a plurality of Synthetic Aperture Radar (SAR) images containing C different target categories, wherein each target category corresponds to M SAR images with the size of h multiplied by h, each SAR image contains 1 target, C is more than or equal to 10, M is more than or equal to 200, and h is 128;
(1b) for the purpose in each SAR imageMarking the mark class and randomly selecting the mark class containing CtrainTotal of C for each object classtrainTaking x M SAR images and labels of each SAR image as training sample sets
Mixing the rest CtestTotal of C for each object classtestTaking x M SAR images and labels of each SAR image as test sample sets
wherein Ctrain+Ctest＝C，3≤Ctest≤5；
(2) Constructing a network model H based on mixed loss and graph attention:
constructing a mixed loss and graph attention-based network model H comprising a data enhancement module D, an embedded network module E, a node feature initialization module I and a graph attention network module G which are sequentially cascaded, wherein the embedded network module E comprises a plurality of first volume modules E which are sequentially cascadedCAnd a second convolution module ELEach first convolution module ECComprises a first convolution layer, a first batch of normalization layers, a Mish activation layer, a maximum pooling layer and a second convolution module E which are sequentially stackedLThe second convolution layer and the second batch of normalization layer are sequentially stacked; the graph attention network module G comprises a plurality of graph updating layers U which are sequentially cascaded, and each graph updating layer U comprises edge feature building modules U which are sequentially stackedEAttention weight calculation module UWAnd node feature update module UNEach edge feature building block UEComprises a plurality of first full-connection layers which are stacked in sequence, and each node feature updating module UNA second full-link layer is included;
(3) performing iterative training on a network model H based on mixed loss and graph attention:
(3a) initializing the iteration number to be N, wherein the maximum iteration number is N, N is more than or equal to 1000, and N is set to be 0;
(3b) from a training sample set
Randomly selecting a group containing CtestTotal of C for each object classtestMultiplying the SAR images by x M, and performing one-hot coding on the label of each SAR image to obtain C of each SAR imagetestDimension label vector, then from CtestRandomly selecting K SAR images contained in each target category and corresponding label vectors from the xM SAR images as training support sample sets
C is remainedtest(M-K) SAR images and corresponding label vectors serving as training query sample sets
After one-hot coding, the C-dimensional element in the tag vector of each SAR image indicates that the target in the SAR image belongs to CtestThe probability of the c-th object class of the object classes,
representing the a-th training support sample consisting of the SAR image and its corresponding label vector,
representing a b-th training query sample consisting of the SAR image and the corresponding label vector, wherein K is more than or equal to 1 and less than or equal to 10;
(3c) will train and support the sample set
With each training query sample
Combined into a training task
Obtaining a training task set
And will be
Forward propagation as input to the mixed-loss and graph attention based network model H:
(3c1) data enhancement module D pairs training task set
Performing data enhancement on each SAR image: performing power transformation on each SAR image, adding noise to the SAR image subjected to power transformation, performing turnover transformation on the SAR image subjected to noise addition, and performing rotation transformation on the SAR image subjected to turnover transformation to obtain an enhanced training task set
wherein ,
representing training tasks
A corresponding one of the enhanced training tasks is,
representing training tasks
Training support sample in (1)
The corresponding reinforced training support sample is used for training,
representing training query samples
Corresponding enhanced training query samples;
(3c2) embedded network module E pair enhanced training task set
Each of the enhanced training tasks in (1)
Mapping each SAR image to obtain a training embedded vector set
And using an embedded loss function LEBy training sets of embedded vectors
Computing training task set
Value of insertion loss lE：
wherein ,
representing an enhanced training task
Corresponding training embedded vector set, satisfies a ≠ CtestOf K +1
Presentation of enhanced training support samples
The corresponding training is embedded into the vector(s),
representing enhanced training query samples
The corresponding training embedding vector, log (-) represents the logarithm based on the natural constant e, exp (-) represents the exponent based on the natural constant e, Σ represents the running sum,
representing a training task
Training support sample set in (1)
Each training embedded vector corresponding to each SAR image of the included c-th target class
The class center of the c-th object class obtained by averaging,
representing and training tasks
Training query sample in (1)
Targets in the contained SAR images belong to the class center of the same target class, d represents a measurement function, and d (p, q) | | p-q | | luminance2；
(3c3) Node characteristic initialization module I constructs a virtual label vector
And embedding a set of vectors for each training
In the condition that a is not equal to CtestEach training embedded vector of K +1
Splicing with the label vector of the corresponding SAR image, and simultaneously embedding each training vector group
Training embedded vector in
And virtual tag vector
Splicing to obtain a training node 1-layer feature group set
wherein ,
is represented by CtestA vector in which the element values of each dimension are all 1,
representing training embedded vector sets
The corresponding training node layer 1 feature set,
representing training embedded vectors
Corresponding training node 1 layer characteristics;
(3c4) graph attention network module G trains node layer 1 feature group set through input
To pair
Including each training node layer 1 feature set
Training node level 1 features in
Corresponding training query samples
Carrying out category prediction on targets in the SAR image to obtain a training prediction result vector set
wherein ,
representing training node layer 1 features
Corresponding dimension is CtestThe c-th element represents the 1-level feature of the training node
Corresponding training query samples
The prediction probability that the target in the included SAR image belongs to the c-th target class;
(3c5) using a classification loss function LCPredicting the result vector set by training
And training the query sample set
All the label vectors in (1) calculate the training task set
Is a classification loss value lC：
wherein ,
predicting outcome vectors for training
Value of the c-th element in (y)b,cPredicting outcome vectors for training
The value of the c-dimension element in the label vector of the corresponding SAR image;
(3d) for the training task set
Is a classification loss value lCAnd training task set
Value of insertion loss lEObtaining a training task set by calculating a weighted sum
Mixing loss value l, l ═ λ lC+(1-λ)lEThen, updating parameters of all first convolution layers and all second convolution layers embedded in the network module E and parameters of all first full-connection layers and all second full-connection layers in the attention network module G through a mixed loss value l by using a random gradient descent algorithm, wherein lambda is weight, and lambda is more than or equal to 0.7 and less than 1;
(3e) judging whether N is greater than or equal to N, if so, obtaining a trained network model H' based on the mixing loss and the graph attention, otherwise, enabling N to be N +1, and executing the step (3 b);
(4) obtaining a target classification result of the small sample SAR image:
(4a) For test sample set
Carrying out one-hot coding on the label of each SAR image to obtain C of each SAR imagetestDimension label vectors, then from the test sample set
C of (A)testRandomly selecting K SAR images contained in each target category and corresponding label vectors from the xM SAR images as a test support sample set
C is remainedtest(M-K) SAR images and corresponding label vectors serving as test query sample sets
wherein ,
representing the e-th test support sample consisting of the SAR image and its corresponding tag vector,
representing a g-th test query sample consisting of the SAR image and the corresponding label vector;
(4b) supporting the test with a sample set
With each test query sample
Combined into test tasks
Obtaining a set of test tasks
And will be
Forward propagation as input to the trained mixed-loss and attention-based network model H':
(4b1) trained embedded network module E' pair test task set
Each test task in (1)
Mapping each SAR image to obtain a test embedded vector group set
wherein ,
representing test tasks
Corresponding test embedded vector groups satisfy e ≠ CtestOf K +1
Presentation of test support samples
The corresponding test-embedded vector is inserted into the vector,
representing test query samples
A corresponding test embedding vector;
(4b2) node characteristic initialization module I constructs a virtual label vector
And embedding the vector set for each test
In the condition that e ≠ CtestEach test embedding vector of K +1
Splicing with the label vector of the corresponding SAR image, and simultaneously embedding each test into a vector group
Test embedded vector in
And virtual tag vector
Splicing to obtain a 1-layer feature group set of the test nodes
wherein ,
representing test embedding vector sets
The corresponding test point 1 layer feature set,
table test embedded vector
Corresponding test node 1 layer characteristics;
(4b3) the trained graph attention network module G' tests the node layer 1 feature group set through input
To pair
Including each test node a layer 1 feature set
Test node level 1 features in
Corresponding test query sample
IncludedCarrying out category prediction on the target in the SAR image to obtain a test prediction result vector set
Each test prediction vector
The dimension number corresponding to the medium maximum value is
Corresponding test query sample
Including a prediction category of the target in the SAR image, wherein,
representing test node level 1 features
Corresponding dimension is CtestThe element value of the c-th dimension represents the test node level 1 feature
Corresponding test query sample
The probability that the object in the included SAR image belongs to the c-th object class.
2. The mixed-loss and attention-based small-sample SAR target classification method according to claim 1, characterized in that the mixed-loss and attention-based network model H in the step (2) is embedded in a first convolution module E included in a network module ECIs five, the figure notes the number of figure update layers U comprised by the force network module GFor three, each graph updating layer U comprises an edge feature building module UEThe number of the first full connection layers in (1) is five, and the specific parameters of the embedded network module E and the graph attention network module G are as follows:
first volume module E included in embedded network module ECThe sizes of the convolution kernels of the five first convolution layers in the three first convolution layers are all 3 multiplied by 3, the step lengths are all 1, the filling is all 1, the number of the convolution kernels of the first three first convolution layers is 64, the number of the convolution kernels of the fourth and fifth first convolution layers is 128, the sizes of the pooling kernels of the five maximum pooling layers are all 2 multiplied by 2, and the sliding step lengths are all 2; second convolution module ELThe size of the convolution kernel of the second convolution layer in (1) is 4 × 4, and the step size is 1;
the graph attention network module G comprises an edge feature building module U in each graph updating layer UEThe number of the neurons of the first four first full connection layers is 96, and the number of the neurons of the fifth first full connection layer is 1; node feature update module U included in the first two graph update layers UNThe number of the neurons of the second full-connection layer is 24, and the third graph updating layer U comprises a node feature updating module UNThe number of neurons in the second fully-connected layer in (1) is Ctest。
3. The method for classifying small-sample SAR target based on mixed loss and attention of claim 1, wherein the data enhancement module D in step (3c1) is used for training task set
The method for enhancing the data of each SAR image comprises the following specific steps: performing a power transformation on each SAR image B, i.e. B1＝(B/255)γX 255 to obtain SAR image B1For each SAR image B1Making an addition of noise, i.e. B2＝B1+ noise, obtaining SAR image B2For each SAR image B2Performing a flip-flop transformation, i.e. B3＝flip(B2) Obtaining an SAR image B3For each SAR image B3To make a rotary changeAlternatively, i.e. B' ═ rot (B)3) Obtaining an enhanced SAR image B', wherein gamma represents power and is randomly selected to be [0.7,1.3 ]]The value within the range, noise, means compliance [ -alpha, alpha ]]Of uniformly distributed noise, alpha is randomly taken to be (0, 50)]The value within the range, film (DEG) indicates that the left and right or up and down are randomly turned, the probability of each turning mode is 1/2, rot (DEG) indicates that the clockwise rotation is randomly carried out by 90 DEG or 180 DEG or 270 DEG, and the probability of each rotation angle is 1/3.
4. The method for classifying SAR targets based on mixed loss and graph attention of claim 1, wherein the graph attention network module G in step (3c4) is used for training node layer 1 feature set sets
Including each training node layer 1 feature set
Training node level 1 features in
Performing category prediction, comprising the following steps:
(3c41) initializing the iteration times to be i, setting the maximum iteration time to be 3, and setting i to be 1;
(3c42) edge feature construction module U of ith graph updating layer UEUsing training node i-layer feature set sets
Each training node in (1) i-layer feature set
All included training node i-layer features
Constructing edge characteristics to obtain training edge i-layer characteristicsGroup collection
wherein ,
representing training node i-layer feature set
The corresponding training edge i-layer feature set,
representing training node i-layer feature set
The i-layer characteristics of the h-th training node,
representing training node i-layer feature set
The i-layer feature of the j-th training node,
show the result of trainingTraining node i-layer features
And training node i-layer features
The obtained training edge i-layer characteristic, abs (-) represents taking the absolute value of each element,
representing the edge feature building Module U included in the ith graph update layer UEA plurality of first full connection layers sequentially laminated in the above order;
(3c43) attention weight calculation module U of ith map update layer UWFor training edge i-layer feature group set
Each training edge i-layer feature set in (1)
All training edge i-layer features included
Calculating attention weight to obtain training i-layer attention weight group set
wherein ,
representing training edge i-layer feature set
Correspondingly training the i-layer attention weight group,
representing training edge i-layer features
The corresponding training i-layer attention weight,
from 1 to C representing removal of htestA set of positive integers of (M-K);
(3c44) node characteristic updating module U of ith graph updating layer UNUsing training node i-layer feature set sets
And training the i-layer attention weight set
Updating the node characteristics to obtain a training node i +1 layer characteristic group set
wherein ,
representing training node i-layer feature set
The corresponding training node i +1 level feature set,
representing training node i-layer features
Corresponding training nodes are i +1 layer characteristics, a | b represents that after the vector b is spliced in the vector a,
indicating the node characteristic update module U included by inputting n into the ith graph update layer UNThe second fully connected layer of (a);
(3c45) judging whether i is true or not, if so, collecting the obtained training node 4-layer feature group set
Each training node in (4) layer feature set
Included training node 4-level features
Performing softmax transformation to obtain a vector set of training prediction results
Otherwise, let i be i +1, and perform step (3c42), wherein,
for training node 4-level features
A corresponding training predictor vector.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110408623.2A CN113095416B (en) | 2021-04-16 | 2021-04-16 | Small sample SAR target classification method based on mixing loss and graph meaning force |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110408623.2A CN113095416B (en) | 2021-04-16 | 2021-04-16 | Small sample SAR target classification method based on mixing loss and graph meaning force |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN113095416A true CN113095416A (en) | 2021-07-09 |
| CN113095416B CN113095416B (en) | 2023-08-18 |
Family
ID=76677939
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110408623.2A Active CN113095416B (en) | 2021-04-16 | 2021-04-16 | Small sample SAR target classification method based on mixing loss and graph meaning force |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113095416B (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113592008A (en) * | 2021-08-05 | 2021-11-02 | 哈尔滨理工大学 | System, method, equipment and storage medium for solving small sample image classification based on graph neural network mechanism of self-encoder |
| CN113655479A (en) * | 2021-08-16 | 2021-11-16 | 西安电子科技大学 | Small sample SAR target classification method based on deformable convolution and double attention |
| CN115131580A (en) * | 2022-08-31 | 2022-09-30 | 中国科学院空天信息创新研究院 | Space target small sample identification method based on attention mechanism |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110516561A (en) * | 2019-08-05 | 2019-11-29 | 西安电子科技大学 | SAR image target recognition method based on DCGAN and CNN |
| CN111191718A (en) * | 2019-12-30 | 2020-05-22 | 西安电子科技大学 | Small sample SAR target identification method based on graph attention network |
| WO2021000906A1 (en) * | 2019-07-02 | 2021-01-07 | 五邑大学 | Sar image-oriented small-sample semantic feature enhancement method and apparatus |
-
2021
- 2021-04-16 CN CN202110408623.2A patent/CN113095416B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2021000906A1 (en) * | 2019-07-02 | 2021-01-07 | 五邑大学 | Sar image-oriented small-sample semantic feature enhancement method and apparatus |
| CN110516561A (en) * | 2019-08-05 | 2019-11-29 | 西安电子科技大学 | SAR image target recognition method based on DCGAN and CNN |
| CN111191718A (en) * | 2019-12-30 | 2020-05-22 | 西安电子科技大学 | Small sample SAR target identification method based on graph attention network |
Non-Patent Citations (2)
| Title |
|---|
| 何凯;冯旭;高圣楠;马希涛;: "基于多尺度特征融合与反复注意力机制的细粒度图像分类算法", 天津大学学报(自然科学与工程技术版), no. 10 * |
| 刘晨;曲长文;周强;李智;李健伟;: "基于卷积神经网络迁移学习的SAR图像目标分类", 现代雷达, no. 03 * |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113592008A (en) * | 2021-08-05 | 2021-11-02 | 哈尔滨理工大学 | System, method, equipment and storage medium for solving small sample image classification based on graph neural network mechanism of self-encoder |
| CN113655479A (en) * | 2021-08-16 | 2021-11-16 | 西安电子科技大学 | Small sample SAR target classification method based on deformable convolution and double attention |
| CN113655479B (en) * | 2021-08-16 | 2023-07-07 | 西安电子科技大学 | Small sample SAR target classification method based on deformable convolution and double attentions |
| CN115131580A (en) * | 2022-08-31 | 2022-09-30 | 中国科学院空天信息创新研究院 | Space target small sample identification method based on attention mechanism |
Also Published As
| Publication number | Publication date |
|---|---|
| CN113095416B (en) | 2023-08-18 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN113095416A (en) | Small sample SAR target classification method based on mixed loss and graph attention | |
| CN111199214B (en) | Residual network multispectral image ground object classification method | |
| CN103955702B (en) | SAR image terrain classification method based on depth RBF network | |
| CN105931253B (en) | A kind of image partition method being combined based on semi-supervised learning | |
| CN110309868A (en) | In conjunction with the hyperspectral image classification method of unsupervised learning | |
| CN110633708A (en) | Deep network significance detection method based on global model and local optimization | |
| CN113486981A (en) | RGB image classification method based on multi-scale feature attention fusion network | |
| CN112085059B (en) | Breast cancer image feature selection method based on improved sine and cosine optimization algorithm | |
| CN107992891A (en) | Based on spectrum vector analysis multi-spectral remote sensing image change detecting method | |
| CN107292341A (en) | Adaptive multi views clustering method based on paired collaboration regularization and NMF | |
| CN109446894A (en) | The multispectral image change detecting method clustered based on probabilistic segmentation and Gaussian Mixture | |
| CN113095409A (en) | Hyperspectral image classification method based on attention mechanism and weight sharing | |
| CN113655479B (en) | Small sample SAR target classification method based on deformable convolution and double attentions | |
| CN106446806B (en) | Semi-supervised face identification method based on the sparse reconstruct of fuzzy membership and system | |
| CN114169442A (en) | Remote sensing image small sample scene classification method based on double prototype network | |
| CN115311478A (en) | Federal image classification method based on image depth clustering and storage medium | |
| CN115631396A (en) | YOLOv5 target detection method based on knowledge distillation | |
| CN111179272B (en) | Rapid semantic segmentation method for road scene | |
| CN113420593B (en) | Small sample SAR automatic target recognition method based on hybrid inference network | |
| CN114998688A (en) | Large-view-field target detection method based on YOLOv4 improved algorithm | |
| WO2022100607A1 (en) | Method for determining neural network structure and apparatus thereof | |
| CN115116539A (en) | Object determination method and device, computer equipment and storage medium | |
| CN109063750B (en) | SAR target classification method based on CNN and SVM decision fusion | |
| CN111222534A (en) | Single-shot multi-frame detector optimization method based on bidirectional feature fusion and more balanced L1 loss | |
| CN116824485A (en) | Deep learning-based small target detection method for camouflage personnel in open scene |
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 |