CN114219960A - Space target ISAR image classification method under small sample condition of XGboost based on multi-learner optimization - Google Patents
Space target ISAR image classification method under small sample condition of XGboost based on multi-learner optimization Download PDFInfo
- Publication number
- CN114219960A CN114219960A CN202111544817.1A CN202111544817A CN114219960A CN 114219960 A CN114219960 A CN 114219960A CN 202111544817 A CN202111544817 A CN 202111544817A CN 114219960 A CN114219960 A CN 114219960A
- Authority
- CN
- China
- Prior art keywords
- network
- convolution
- training
- xgboost
- isar
- 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.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 68
- 238000005457 optimization Methods 0.000 title claims description 19
- 238000012545 processing Methods 0.000 claims abstract description 41
- 238000000605 extraction Methods 0.000 claims abstract description 39
- 239000013598 vector Substances 0.000 claims abstract description 32
- 230000007246 mechanism Effects 0.000 claims abstract description 28
- 239000002775 capsule Substances 0.000 claims abstract description 26
- 238000011176 pooling Methods 0.000 claims abstract description 18
- 238000004422 calculation algorithm Methods 0.000 claims abstract description 11
- 238000009987 spinning Methods 0.000 claims abstract description 10
- 238000012549 training Methods 0.000 claims description 98
- 238000012795 verification Methods 0.000 claims description 42
- 239000011159 matrix material Substances 0.000 claims description 38
- 230000008569 process Effects 0.000 claims description 38
- 238000010586 diagram Methods 0.000 claims description 37
- 238000010606 normalization Methods 0.000 claims description 36
- 230000006870 function Effects 0.000 claims description 29
- 238000004364 calculation method Methods 0.000 claims description 23
- 230000004913 activation Effects 0.000 claims description 20
- 238000013528 artificial neural network Methods 0.000 claims description 13
- 239000013604 expression vector Substances 0.000 claims description 9
- 102100040412 Amyloid beta A4 precursor protein-binding family B member 1-interacting protein Human genes 0.000 claims description 7
- 101000964223 Homo sapiens Amyloid beta A4 precursor protein-binding family B member 1-interacting protein Proteins 0.000 claims description 7
- 238000001914 filtration Methods 0.000 claims description 5
- 238000004088 simulation Methods 0.000 claims description 5
- 230000008859 change Effects 0.000 claims description 4
- 238000003066 decision tree Methods 0.000 claims description 4
- 230000010354 integration Effects 0.000 claims description 4
- 238000007781 pre-processing Methods 0.000 claims description 4
- 238000012935 Averaging Methods 0.000 claims description 3
- 230000017105 transposition Effects 0.000 claims description 3
- 238000012360 testing method Methods 0.000 description 21
- 238000003384 imaging method Methods 0.000 description 6
- 210000002569 neuron Anatomy 0.000 description 4
- 238000013136 deep learning model Methods 0.000 description 3
- 230000009466 transformation Effects 0.000 description 3
- 230000003213 activating effect Effects 0.000 description 2
- 238000013135 deep learning Methods 0.000 description 2
- 238000013461 design Methods 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- PXFBZOLANLWPMH-UHFFFAOYSA-N 16-Epiaffinine Natural products C1C(C2=CC=CC=C2N2)=C2C(=O)CC2C(=CC)CN(C)C1C2CO PXFBZOLANLWPMH-UHFFFAOYSA-N 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000000295 complement effect Effects 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 238000013527 convolutional neural network Methods 0.000 description 1
- 238000012217 deletion Methods 0.000 description 1
- 230000037430 deletion Effects 0.000 description 1
- 230000001419 dependent effect Effects 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 238000002372 labelling Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000009467 reduction Effects 0.000 description 1
- 238000010200 validation analysis Methods 0.000 description 1
- 230000000007 visual effect 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
- 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)
- Computing Systems (AREA)
- Software Systems (AREA)
- Molecular Biology (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- Probability & Statistics with Applications (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Image Analysis (AREA)
Abstract
A spatial target ISAR image classification method based on XGboost optimized by multiple learners under a small sample condition belongs to the technical field of ISAR image processing. The method aims to solve the problem that the space target ISAR image recognition rate is low under the existing small sample condition. Aiming at ISAR spinning space target images to be classified, the method classifies the ISAR spinning space target images by utilizing an XGboost model optimized based on multiple learners; the XGboost model is an XGboost network model formed by integrating a stack network by utilizing an XGboost algorithm, and parallel basic learners in the stack network comprise a graph convolution network, a capsule network and an Alexenet network added with an attention mechanism module based on rotation invariance; the capsule network comprises a basic feature extraction module, a vector feature extraction module and a dynamic routing layer, and does not contain a pooling layer. The method is mainly used for classifying ISAR images.
Description
Technical Field
The invention relates to a spatial target ISAR image classification method under a small sample condition of XGboost based on multi-learner optimization, and belongs to the technical field of ISAR image processing.
Background
Recently, deep learning models have enjoyed great success in Inverse Synthetic Aperture Radar (ISAR) image recognition. However, most deep learning models fail to achieve satisfactory results under small sample conditions due to the contradiction between the parameter space of the deep learning model and the insufficiency of target image labeling samples in the ISAR space.
For the ISAR image of the space target, due to the problems of target inversion, shielding and the like, the image may have the problems of deletion or different expression forms and the like, and is influenced by the angular flicker phenomenon, speckle noise and the like, the reduction of the image quality exists, so that the identification of the ISAR image is influenced; for space targets, the same type of targets are completely different in image representation due to structural differences of the targets, imaging plane changes are caused due to rotation changes of the targets relative to the radar, the radar can obtain different sides of the targets, and small sample problems can occur due to the reasons;
the existing deep learning network generally needs a large amount of labeled sample data, and even the network in the aspect of small sample learning has the problem of low recognition rate, so that a network model special for ISAR image recognition is lacked.
Disclosure of Invention
Aiming at the problem that the existing identification rate of space target ISAR images under the condition of small samples is low, the invention provides an XGboost space target ISAR image classification method under the condition of small samples based on multi-learner optimization.
The method for classifying the ISAR images of the space targets under the condition of small samples of the XGboost based on multi-learner optimization aims at ISAR spinning space target images to be classified and utilizes an XGboost model based on multi-learner optimization to classify the ISAR images of the space targets;
the XGboost model based on multi-learner optimization is an XGboost network model formed by integrating stacked networks by utilizing an XGboost algorithm;
the stack network is based on meta-learner integration, and parallel basic learners in the stack network comprise a graph convolution network, a capsule network and an Alexnet network added with an attention mechanism module based on rotation invariance; the Alexnet network added with the attention mechanism module based on the rotation invariance is a network model formed by adding the attention mechanism module based on the rotation invariance to the back of a second convolution layer of the Alexnet network; the capsule network comprises a basic feature extraction module, a vector feature extraction module and a dynamic routing layer, and does not contain a pooling layer.
Further, the method is characterized in that the ISAR spinning space target image to be classified needs to be preprocessed before the space target ISAR image classification is carried out by utilizing the XGboost model based on multi-learner optimization, and the preprocessing process comprises the following steps:
firstly, carrying out enhanced Lee filtering on the space target ISAR image;
and then carrying out contrast enhancement and energy normalization processing.
Further, the specific structure of the capsule network is as follows:
(A) basic feature extraction module:
the basic feature extraction module comprises three convolution units, wherein each convolution unit comprises a convolution layer, a batch normalization layer and a ReLU activation layer;
(B) the vector feature extraction module:
the vector feature extraction module comprises 8 convolutional layers, wherein the 8 convolutional layers are in parallel relation, and the 8 convolutional layers do not use an activation function; the sizes of the convolution layers are the same, and the input feature graphs are respectively convolved;
then, composing the obtained result into a tensor form;
(C) dynamic routing layer
The dynamic routing layer is realized by adopting a dynamic routing method.
Further, the Alexnet network incorporating the attention mechanism module based on rotation invariance is as follows:
the attention mechanism module based on the rotation invariance comprises a rotation processing module and a space attention mechanism module;
a rotation processing module: based on an Alexnet network, performing rotation operation on the output characteristic diagram of the ReLU activation function of the second convolution layer to obtain a rotated characteristic diagram;
the spatial attention mechanism module comprises three convolution units, namely Conv1, Conv2 and Conv 3;
conv 1: the rotation characteristic diagram is firstly activated through the convolution layer and then through the ReLU;
conv2 comprises Conv2-1 and Conv2-2 subunits, the characteristic diagrams processed by Conv1 are respectively sent into Conv2-1 and Conv2-2 subunits, and in the Conv2-1 and Conv2-2 subunits, the convolution layers are firstly passed through and then deactivated by Softmax, and an output characteristic diagram F is obtained2-1And F2-2;
The Conv3 unit comprises Conv3-1 and Conv3-2 subunits, a rotation characteristic diagram obtained by the rotation processing module is directly and respectively sent into the Conv3-1 and Conv3-2 subunits, the convolution layers in the Conv3-1 and Conv3-2 subunits are firstly processed, then batch normalization processing is carried out, and G is obtained3-j,j=1,2;
For feature map F2-jWas subjected to transposition to obtain F'2-jIs prepared from'2-jAnd a characteristic diagram G3-jAnd random initialization coefficient alphajMultiplying and adding to obtain an output characteristic diagram
Finally, averaging the output characteristic diagrams obtained by rotating all the angles to obtain the output characteristic diagram of the RIAM moduleK represents the number of rotation operations performed by the rotation processing module.
Further, the graph convolution network is as follows:
the graph convolution network comprises a feature extraction module and a graph neural network processing module;
(a) a feature extraction module:
the feature extraction module comprises three convolution units, each convolution unit comprising: the system comprises a convolution layer, a batch normalization layer, a maximum pooling layer and a ReLU activation function layer;
the last convolution unit outputs the feature expression of the image in a vector form, and combines the feature expression with the one-hot coding formed by the image category as the input of the graph neural network module;
(b) the figure neural network module:
the graph neural network module comprises two units for calculating the adjacent matrix and two graph convolution units; the first calculation adjacent matrix unit is connected with the first graph convolution unit, and then the second calculation adjacent matrix unit and the second graph convolution unit are connected;
the input of the first calculation adjacent matrix unit is the subtraction of node expression vectors of the feature extraction module to obtain a two-dimensional feature map; the input of the second calculation adjacent matrix unit is a characteristic graph obtained by subtracting the node expression vectors output by the first graph convolution unit; the two calculation adjacent matrix units respectively comprise three convolution modules, the first two convolution modules respectively comprise a convolution layer, a batch normalization layer and a ReLU activation layer, and the last convolution module only comprises one convolution layer;
the first graph convolution unit comprises a full connection layer and a batch normalization layer; the second graph convolution unit contains two fully connected layers and one batch normalization layer, and the output of the second graph convolution unit is equal to the classification probability of five types of targets.
Further, the process of integrating the XGBoost network model formed by the stacked network by using the XGBoost algorithm includes the steps of:
2.4.1 computing weights of the underlying learner
XGboost networkIs D { xi,α,li,ei,αWhere i 1,2, n denotes the number of meta-features, xi,αIs a meta feature, liIs a label, ei,αIs the weight of the element feature, alpha is 1,2,3 is the serial number of the basic learner, and introduces the weight q corresponding to the learnerαThe weight expression of the meta-feature is as follows:
wherein x isi,α,1、……,xi,α,NRespectively being meta-features xi,αThe elements of (1);is the mean of the meta-features; n is the total number of meta-features, N is the number of elements in a meta-feature;
2.4.2 calculation of weight to loss function partial derivatives
Defining the loss function L, including the predicted valuesAnd a label liCross entropy loss of l, and tree structure fkCorresponding regularization loss Ω ();
the loss function L is defined as:
l is defined as follows:
wherein γ is the regularization parameter for the number of leaves, T is the number of leaf nodes, λ is the regularization parameter for the leaf weights, and w represents the weights of the leaves;
l relates to qαPartial derivatives of (a):
2.4.3 adaptively updating weights
Updating the learner's weights with the partial derivatives and update rate η calculated above:
updating the weight corresponding to the meta-feature by the change of the weight of the learner, and realizing the self-adaptive updating of the weight;
2.4.4 calculation of network gain
The gain of the network is calculated from the first and second derivatives of the predicted values, gi,αAnd hi,αAre the first and second derivatives of the meta-feature, and L and R are the left and right nodes of the decision tree. The gain is defined as:
and (5) splitting the tree by the maximum gain to obtain a final classification result.
Further, the training process of the XGBoost network model includes the following steps:
meta-feature output of graph convolution network: taking the classification probability of the data in the verification set as training meta-characteristics;
adding a rotation invariance-based attention mechanism module to the meta-feature output of the Alexnet network: combining the classification confidence degrees of the verification sample sets to form training meta-features;
meta-feature output of capsule network: taking the L2 norm of the output vector to obtain the classification probability, and taking the classification probability of the data in the verification sample set as the training element characteristic;
and training the XGboost network optimized by the multiple learners by using the training element characteristics.
Further, the training process of the graph volume network, the capsule network and the Alexnet network is as follows:
the Alexnet network model training process comprises the following steps:
in the training process, each class of the training sample library is equally divided into K parts, wherein K-1 part is a training sample set, the rest part is a verification sample set, and K-fold cross verification is carried out to ensure that K parts of data are verified;
training an Alexnet network model by utilizing a training sample set and a verification sample set, and combining the classification confidence degrees of the verification sample set to form training meta-features;
the graph convolution network model training process comprises the following steps:
dividing a training sample library into a training set and a verification set, wherein each class of the training set has qshot images, and each class of the verification set has K-qshot images; in the training process, 5 xqshot +1 images of the network are input in each iteration, wherein 1 means that one image is taken from a verification set in sequence to be used as verification, and the label of the image is known and is used for supervised learning;
training the graph convolution network model by using a training set and a verification set, and taking the classification probability of data in the verification set as training meta-features;
the capsule network model training process comprises the following steps:
dividing a training sample library into a training set and a verification set, training a capsule network model by using the training set and the verification set, obtaining an L2 norm of an output vector to obtain a classification probability, and taking the classification probability of data in the verification sample set as a training meta-feature;
the ISAR spin space target sample is stored in the training sample library.
Further, the ISAR spin space target sample is obtained based on a three-dimensional space target model and a range-Doppler algorithm simulation.
Further, the ISAR spin space target samples stored in the training sample library are preprocessed samples, and the preprocessing process includes the following steps:
firstly, carrying out enhanced Lee filtering on the space target ISAR image;
and then carrying out contrast enhancement, energy normalization and data expansion, wherein the data expansion comprises left-right turning and up-down turning.
The invention has the beneficial effects that:
the method of the invention uses the thought of an integration network as a reference, uses three complementary small sample learning networks, namely a network of Alexenet + RIAM module which focuses on the ISAR image target position and semantic information, a graph convolution network which focuses on the deep characteristic relation of the image and a capsule network which focuses on the relative position relation of key parts in a space target, and finally uses an innovative XGboost decision tree optimized by a multi-learner to integrate the network.
The invention comprises three parts, namely an ISAR spinning space target sample, a stacked network architecture and target class estimation; the ISAR spinning space target sample part is responsible for simulating and generating a large number of ISAR images under the space target motion state to form a training sample library and a testing sample library, the stacking network architecture provides a design structure of the network, and the target type estimation part is based on the stacking network and conducts training test on the ISAR spinning space target sample library to achieve a target type estimation function.
By adopting the stack network model trained by the training sample library, the target class estimation can be carried out on the ISAR space target images under the condition of a small amount of marked samples in the ISAR observation scene, the correct classification probability of basic learners such as a graph convolution network and the like on the targets is improved, the actual requirements are met, and the implementation is convenient.
According to the invention, scattering point information required by ISAR imaging is obtained through the three-dimensional model of the space target and electromagnetic parameter simulation, the ISAR imaging result is approximately consistent with the actually measured ISAR image, a small number of ISAR simulation images of the five types of targets are obtained, the requirement of the text stacking network method on samples is ensured, and the smooth application of the algorithm is ensured.
The invention innovatively provides a stacked network for ISAR spin space target type estimation, the input of the stacked network is an ISAR space target image with a small number of marked samples, the output is the category to which the target belongs, wherein the introduction of an attention mechanism module based on rotation invariance and the proposal of an XGboost method optimized by multiple learners provide guarantee for the improvement of the recognition rate, and certain method superiority is provided.
The invention utilizes ISAR imaging technology and deep learning technology to better meet the requirements of complex ISAR target class estimation in different imaging scenes, and the stacked network keeps certain stability on the visual angle change, affine transformation and noise of the input image.
Drawings
FIG. 1 is an overall flow chart of a spatial target ISAR image classification method under a small sample condition of XGboost based on multi-learner optimization according to the invention;
FIG. 2 is a schematic diagram of a RIAM module design;
FIG. 3 is an architecture diagram of a graph convolution network in a base learner;
FIG. 4 is an architecture diagram of the capsule network in the base learner;
FIG. 5 is a flow chart of obtaining a training sample library and a testing sample library.
Detailed Description
The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
The invention is further described with reference to the following drawings and specific examples, which are not intended to be limiting.
First embodiment this embodiment will be described with reference to fig. 1 to 5,
the method for classifying the ISAR images of the space targets under the condition of the small samples of the XGboost based on the optimization of the multiple learners in the embodiment comprises the following steps:
s1, obtaining an ISAR spin space target sample sequence based on the three-dimensional space target model and the range-Doppler algorithm simulation, and grouping the ISAR spin space target samples into a training sample library and a test sample library;
after obtaining the ISAR spin space target sample sequence, image data needs to be preprocessed: firstly, carrying out enhanced Lee filtering on the space target ISAR image to remove speckle noise; then contrast enhancement, energy normalization and data expansion are carried out, wherein the data expansion comprises left-right overturning and up-down overturning;
in the process of grouping ISAR spinning space target samples into a training sample library and a testing sample library, a K/N value is required to be set, the value represents the proportion of the number of images in the training sample library to the number of images in the testing sample library, and when the K/N is equal to 1, the number of images in the training sample library is equal to the number of images in the testing sample library;
s2, constructing a stack network based on meta-learner integration, wherein a basic learner of the stack network comprises an Alexnet network, a graph convolution network and a capsule network, wherein the Alexnet network is added with an attention mechanism module based on rotation invariance; the basic learner of the stacked network is a learner working in parallel;
the last meta-learner integrates the three basic learners by adopting an XGboost algorithm optimized by multiple learners;
the Alexnet network added with the attention mechanism module based on the rotation invariance comprises the following components: an attention mechanism module based on rotation invariance is added behind a second convolution layer of the Alexnet network, and due to the fact that ISAR space target image imaging mechanism causes different display angles of the same target in an image, the rotation invariance of feature extraction is enhanced for the network, and the identification effect is enhanced; the construction process of the Alexnet network added with the attention mechanism module based on the rotation invariance comprises the following steps:
2.1 constructing Alexnet network model with attention mechanism module based on rotation invariance:
2.1.1 equally dividing each class of the training sample library into K parts in the training process, wherein K-1 part is a training sample set, the remaining part is a verification sample set, and performing K-fold cross verification to ensure that K parts of data are verified;
2.1.2 attention mechanism module RIAM based on rotation invariance comprises a rotation processing module and a space attention mechanism module;
a rotation processing module: based on Alexnet network, carrying out a series of rotation operations on the output characteristic diagram of the ReLU activation function of the second convolution layer to become a rotation characteristic diagram;is a series of rotary operations, usingThe function makes an angle of lambda on the input characteristic diagramiE, K, obtaining a rotated characteristic diagram x, x e R{C×W×H}Wherein, R represents a real number set, W multiplied by H is the size of the characteristic diagram, C is the number of characteristic channels and is consistent with the number of input characteristic diagram channels;
the spatial attention mechanism module comprises three convolution units, namely Conv1, Conv2 and Conv 3;
firstly, setting the convolution kernel size and the channel number of Conv1, and deactivating the rotation characteristic diagram through the convolution layer and then through the ReLU, as shown in the following formula:
F(x)=ReLU(Conv1(Rλ(x))) (1)
conv2 includes Conv2-1 and Conv2-2 subunits, and the Conv1 processed feature maps are respectively sent to Conv2-1 and Conv2-2 subunits, and for the Conv2-1 and Conv2-2 subunits, the input feature map is F (x), the size of a convolution kernel is set, and the number of channels is consistent with the number of input feature channels. In Conv2-1 and Conv2-2 subunits, the output characteristic diagram F is obtained by firstly passing through the convolution layer and then deactivating by Softmax2-1And F2-2It is expressed as follows:
F2-j=Softmax(Conv2-j(F(x))),j=1,2 (2)
the Conv3 unit comprises Conv3-1 and Conv3-2 subunits, a rotation characteristic diagram obtained by the rotation processing module is directly and respectively sent into the Conv3-1 and Conv3-2 subunits, a convolution layer in the Conv3-1 and Conv3-2 subunits is firstly passed, hole convolution is used, the size of a convolution kernel is 3 x 3, the expansion step is set to be 2, filling is set to keep the size of an output characteristic diagram to be W x H, (so that a RIAM module can be added on the premise of not changing an Alexnet structure); then, the batch normalization process is performed as follows:
G3-j=BN(Conv3-j(Rλ(x))),j=1,2 (3)
for feature map F2-jWas subjected to transposition to obtain F'2-jIs prepared from'2-jAnd a characteristic diagram G3-jAnd random initialization coefficient alphajThe multiplication and addition result in an output profile, expressed as:
and finally, averaging the output characteristic diagrams obtained by rotating all the angles to obtain the output characteristic diagram of the RIAM module, wherein the output characteristic diagram is expressed as follows:
after the constructed attention mechanism module based on the rotation invariance is placed in a second convolution layer of Alexnet, because the requirement on the rotation invariance of the characteristics in too deep layers is not high, the characteristics extracted by the first convolution layer are too sensitive to the rotation, and the final classification result is poor after the first convolution layer;
2.1.3, combining the classification confidence degrees of the verification sample set to form training meta-features; and inputting the test sample library serving as a test set into the constructed Alexnet network for training to obtain the test meta-characteristics.
2.2 construct graph convolution network:
2.2.1 data processing:
the training sample library is divided into a training set and a verification set, wherein each class of the training set has qshot images, and each class of the verification set has K-qshot images; in the training process, 5 xqshot +1 images of the network are input in each iteration, wherein 1 means that one image is taken from a verification set in sequence to serve as verification, and the label of the image is known and is used for supervised learning.
2.2.2 the graph convolution network comprises a feature extraction module and a graph neural network processing module;
on the whole, the graph convolution network has 9 convolution layers, wherein the number of the feature extraction modules is 3, and the number of two calculation adjacent matrix units of the graph neural network module is 3 respectively; the convolution kernel size in the first convolution layer is 9 x 9 pixels, the step length is 2 pixels, and 32 convolution kernels are in total; the convolution kernel size in the second convolution layer is 5 x 5 pixels, the step size is 1 pixel, and 64 convolution kernels are calculated; the convolution kernel size in the third convolution layer is 3 x 3 pixels, the step size is 1 pixel, and 128 convolution kernels are totally arranged; the convolution kernel size in the fourth convolution layer is 3 x 3 pixels, the step size is 1 pixel, and 64 convolution kernels are totally arranged; the convolution kernel size in the fifth convolution layer is 3 x 3 pixels, the step length is 1 pixel, and 32 convolution kernels are in total; the convolution kernel size in the sixth convolution layer is 3 x 3 pixels, the step length is 1 pixel, and 1 convolution kernel is total; the convolution kernel size in the seventh convolution layer is 3 x 3 pixels, the step size is 1 pixel, and 64 convolution kernels are totally arranged; the convolution kernel size in the eighth convolution layer is 3 x 3 pixels, the step length is 1 pixel, and 32 convolution kernels are in total; the size of the convolution kernel in the ninth convolution layer is 3 multiplied by 3 pixels, the step length is 1 pixel, and 1 convolution kernel is total; only after the last convolution layer of the adjacent matrix unit is calculated, the activation function is not existed, and the activation function of each of the rest convolution layers is the ReLU activation function.
The total number of 3 pooling layers in the graph convolution network is maximum pooling, the size is 2 x 2 pixels, and the step length is 2 pixels.
(a) A feature extraction module:
the feature extraction module comprises three convolution units, each convolution unit comprising: a convolution layer, a batch normalization layer, a maximum pooling layer and a ReLU activation function layer;
the convolution layer input in the first convolution unit is a training set image and a verification set image, and the input of the next two convolution units is a feature map output by the previous convolution unit;
processing the weight coefficient by batch normalization;
performing pooling operation by using the maximum pooling layer;
activating the weight coefficient by utilizing a ReLU function;
and the last convolution unit outputs the feature expression of the image, the feature expression exists in a vector form, and the one-hot codes formed by the feature expression unit and the class of the image are combined and are recorded as node representation vectors of the feature extraction module to be used as the input of the graph neural network module.
(b) The figure neural network module:
the graph neural network module comprises two units for calculating the adjacent matrix and two graph convolution units; the first calculation adjacent matrix unit is connected with the first graph convolution unit, and then the second calculation adjacent matrix unit and the second graph convolution unit are connected;
the first input of the calculation adjacent matrix unit is a two-dimensional characteristic graph obtained by subtracting the node expression vectors of the characteristic extraction module (the initial input is a series of images, the node expression vectors are obtained after the images are processed in pairs, and the node expression vectors are subtracted from each other); the input of the second calculation adjacent matrix unit is a feature graph obtained by subtracting the node expression vectors output by the first graph convolution unit (similar to the processing of subtracting the node expression vectors of the first calculation adjacent matrix unit); the two calculation adjacent matrix units respectively comprise three convolution modules, the first two convolution modules respectively comprise a convolution layer, a batch normalization layer and a ReLU activation layer, and the last convolution module only comprises one convolution layer;
the first graph convolution unit comprises a full connection layer and a batch normalization layer; the second graph convolution unit contains two fully connected layers and one batch normalization layer, and the output of the second graph convolution unit is equal to the classification probability of five types of targets.
2.2.4 output of Meta features:
and taking the classification probability of the data in the verification set as a training meta-feature, and taking the classification probability of the data in the test sample library as a test meta-feature.
2.3 constructing a capsule network:
2.3.1 data processing:
the training sample library is subdivided into a training set and a validation set.
2.3.2 the capsule network comprises a basic feature extraction module, a vector feature extraction module and a dynamic routing layer;
on the whole, the capsule network has 11 convolution layers, wherein 3 basic feature extraction modules and 8 vector feature extraction modules are provided; the convolution kernel size in the first convolution layer is 9 x 9 pixels, the step size is 2 pixels, and 64 convolution kernels are obtained in total; the convolution kernel size in the second convolution layer is 9 x 9 pixels, the step size is 2 pixels, and 128 convolution kernels are totally arranged; the convolution kernel size in the third convolution layer is 7 x 7 pixels, the step size is 1 pixel, and 128 convolution kernels are totally arranged; the sizes of convolution kernels from the fourth convolution layer to the eleventh convolution layer are all 9 x 9 pixels, the step length is 2 pixels, the convolution kernels are all 32 layers, the 8 convolution layers are in parallel relation, and vector features are formed in new dimensions; the activation functions of the first three convolutional layers are all ReLU activation functions, and the rest convolutional layers do not use activation functions. The capsule network does not contain a pooling layer.
(A) Basic feature extraction module:
the basic feature extraction module aims to convert the image into a feature map form through a convolution layer; the basic feature extraction module comprises three convolution units, wherein each convolution unit comprises a convolution layer, a batch normalization layer and a ReLU activation layer;
the input of the convolutional layer is a training set image and a test set image;
processing the weight coefficient by batch normalization;
activating the weight coefficient by utilizing a ReLU function;
does not contain a pooling layer;
(B) the vector feature extraction module:
the vector feature extraction module comprises 8 convolution layers, the size of each convolution layer is the same, the input feature graph is convolved respectively, and the obtained results form a tensor form;
the input of the convolutional layer is a training set image and a test set image;
setting the size of convolution kernel, the number of channels and the step length of the convolution layer respectively;
combining the output results in a new dimension, and outputting a feature vector;
the length of the characteristic vector changes after the characteristic vector passes through the attitude transformation matrix;
(C) dynamic routing layer
Similar to the full connection layer in the convolutional neural network (the dynamic routing and the full connection layer work on different principles and have the same function), the dynamic routing method is used for realizing the dynamic routing in the capsule network;
the number of output vectors is consistent with the type number of the space target;
output of 2.3.3-dimensional features
And taking the L2 norm of the output vector to obtain the classification probability, taking the classification probability of the data in the verification sample set as the training meta-feature, and taking the classification probability of the data in the test sample set as the test meta-feature.
2.4 building a multi-learner optimized XGboost network:
the XGboost network optimized by the multiple learners comprises the steps of calculating the weight of a basic learner, calculating the partial derivative of a loss function by the weight, and calculating the self-adaptive updating weight and the network gain;
2.4.1 computing weights of the underlying learner
The input to the network is D { xi,α,li,ei,α} of whichWhere i 1,2, n denotes the number of training meta-features, xi,αIs a training meta-feature, liIs a label, ei,αIs the weight of the element feature, alpha is 1,2,3 is the serial number of the basic learner, and introduces the weight q corresponding to the learnerαThe weight expression of the meta-feature is as follows:
wherein x isi,α,1、……,xi,α,NRespectively being meta-features xi,αThe elements of (1);is the mean of the meta-features; n is the total number of meta-features, N is the number of elements in a meta-feature;
2.4.2 calculation of weight to loss function partial derivatives
Defining the loss function L, including the predicted valuesAnd a label liCross entropy loss of l, and tree structure fkCorresponding regularization loss Ω ();
the loss function L is defined as:
l is defined as follows:
wherein γ is the regularization parameter for the number of leaves, T is the number of leaf nodes, λ is the regularization parameter for the leaf weights, and w represents the weights of the leaves;
l relates to qαPartial derivatives of (a):
2.4.3 adaptively updating weights
Updating the learner's weights with the partial derivatives and update rate η calculated above:
updating the weight corresponding to the meta-feature by the change of the weight of the learner, and realizing the self-adaptive updating of the weight;
2.4.4 calculation of network gain
The gain of the network is calculated from the first and second derivatives of the predicted values, gi,αAnd hi,αAre the first and second derivatives of the meta-feature, and L and R are the left and right nodes of the decision tree. The gain is defined as:
and (5) splitting the tree by the maximum gain to obtain a final classification result.
S3, training the basic learner by adopting a training sample library and a testing sample library respectively to obtain training element characteristics and testing element characteristics;
training the XGboost network optimized by the multiple learners by using the training element characteristics to obtain an XGboost network model, and testing the testing element characteristics to obtain a result with higher recognition rate;
the training sample and the test sample are preprocessed ISAR images;
in the training process, the pixel size of the Alexene network input ISAR sample of the stack network is 224 multiplied by 224, the pixel size of the graph convolution network input ISAR sample is 32 multiplied by 32, the pixel size of the capsule network input ISAR sample is 128 multiplied by 128, and the class 5 space targets are included.
For the graph convolution network, the input image size is 32 × 32 pixels;
the ISAR image is processed by the characteristic extraction layer and then transmitted to the first convolution layer;
the first convolution layer processes the input image and outputs 32 feature maps with the size of 32 multiplied by 32 pixels;
after batch normalization and ReLU processing, the first pooling layer processes the feature map output by the first convolution layer and outputs a three-dimensional feature matrix of 16 × 16 × 32 pixels;
the second convolution layer processes the three-dimensional feature matrix output by the first pooling layer and outputs 64 feature maps with the size of 16 x 16 pixels;
after batch normalization and ReLU processing, the second pooling layer processes the feature map output by the second convolution layer and outputs a three-dimensional feature matrix of 8 multiplied by 64 pixels;
the third convolution layer processes the three-dimensional feature matrix output by the second pooling layer and outputs 128 feature maps with the size of 8 x 8 pixels;
the third pooling layer processes the feature map output by the third convolution layer after batch normalization and ReLU processing, and outputs a three-dimensional feature matrix of 4 x 128 pixels;
converting the feature map into vectors, adding one-hot codes as vertex feature representation, wherein the size of the vertex feature representation is 21 multiplied by 1 multiplied by 128, and subtracting the vertex representation to obtain an adjacent feature matrix, and the size of the adjacent feature matrix is 21 multiplied by 128 pixels;
the fourth convolution layer processes the adjacent feature matrix and outputs 64 feature maps with the size of 21 multiplied by 21 pixels;
processing the feature map output by the fourth convolutional layer by the fifth convolutional layer after batch normalization and ReLU processing, and outputting 32 feature maps with the size of 21 x 21 pixels;
processing the feature map output by the fifth convolutional layer by the sixth convolutional layer after batch normalization and ReLU processing, and outputting 1 feature map with the size of 21 multiplied by 21 pixels;
the first full-connection layer comprises 16 neurons and is fully connected with one-dimensional vectors mapped by the three-dimensional characteristic matrix output by the sixth convolutional layer;
outputting vertex representation vectors with the size of 16 multiplied by 1 after the processing of a full connection layer and a batch normalization layer; the vertex represents the adjacent characteristic matrix obtained by subtraction, and the size of the adjacent characteristic matrix is 16 multiplied by 1 pixels;
the seventh convolution layer processes the adjacent feature matrix and outputs 64 feature maps with the size of 16 multiplied by 16 pixels;
processing the feature map output by the seventh convolutional layer by the eighth convolutional layer after batch normalization and ReLU processing, and outputting 32 feature maps with the size of 16 multiplied by 16 pixels;
the ninth convolutional layer after batch normalization and ReLU processing processes the feature map output by the eighth convolutional layer and outputs 1 feature map with the size of 16 multiplied by 16 pixels;
the second full-connection layer comprises 16 neurons and is fully connected with the one-dimensional vectors mapped by the three-dimensional characteristic matrix output by the fourth pooling layer;
the second full-junction layer comprises 5 neurons and is fully connected with the 16 neurons of the first full-junction layer;
for the capsule network, the input image size is 128 × 128 pixels;
the first convolution layer processes the input image and outputs 64 feature maps with the size of 60 multiplied by 60 pixels;
batch normalization and ReLU processing;
the second convolution layer processes the three-dimensional feature matrix and outputs 128 feature maps with the size of 26 multiplied by 26 pixels;
batch normalization and ReLU processing;
the third convolution layer processes the three-dimensional characteristic matrix and outputs 128 characteristic graphs with the size of 20 multiplied by 20 pixels;
batch normalization and ReLU processing;
processing the three-dimensional feature matrix by the fourth to eleventh convolutional layers to obtain a feature vector with the size of 6 × 6 × 8 × 32 formed by the feature map on a new dimension;
1152 eigenvectors with the size of 1 × 16 are obtained after processing by 6 × 6 × 32 transformation matrixes with the size of 8 × 16;
5 characteristic vectors of 1 multiplied by 16 are obtained by utilizing a dynamic routing algorithm;
the L2 norm of the feature vector is computed to obtain the classification probability.
During training, the number of training samples in each batch in the Alexnet network is 32, 2800 batches are trained, the learning rate before the full link layer is 0.0001, the full link layer is 0.001, and the optimization method is SGD; the number of training samples in each batch in the GCN is 16, 6000 batches of training are performed in total, and the learning rate is 0.01; the number of training samples in each batch in the CapsNet is 64, 100 batches are trained together, and the learning rate is 0.001; the depth of the XGboost optimized by the multiple learners is 20, the number of the trees is set to 2000, and the classification number is 5.
Although the invention herein has been described with reference to particular embodiments, it is to be understood that these embodiments are merely illustrative of the principles and applications of the present invention. It is therefore to be understood that numerous modifications may be made to the illustrative embodiments and that other arrangements may be devised without departing from the spirit and scope of the present invention as defined by the appended claims. It should be understood that features described in different dependent claims and herein may be combined in ways different from those described in the original claims. It is also to be understood that features described in connection with individual embodiments may be used in other described embodiments.
Claims (10)
1. The method for classifying the ISAR images of the space targets under the condition of small samples of the XGboost based on multi-learner optimization is characterized in that the ISAR images to be classified are classified by utilizing an XGboost model based on multi-learner optimization aiming at ISAR spinning space target images to be classified;
the XGboost model based on multi-learner optimization is an XGboost network model formed by integrating stacked networks by utilizing an XGboost algorithm;
the stack network is based on meta-learner integration, and parallel basic learners in the stack network comprise a graph convolution network, a capsule network and an Alexnet network added with an attention mechanism module based on rotation invariance; the Alexnet network added with the attention mechanism module based on the rotation invariance is a network model formed by adding the attention mechanism module based on the rotation invariance to the back of a second convolution layer of the Alexnet network; the capsule network comprises a basic feature extraction module, a vector feature extraction module and a dynamic routing layer, and does not contain a pooling layer.
2. The method for classifying ISAR (inverse synthetic aperture radar) images of space targets under the condition of small samples based on XGboost optimized by multiple learners as claimed in claim 1, wherein the ISAR spinning space target images to be classified need to be preprocessed before the ISAR images of space targets are classified by the XGboost model optimized by the multiple learners, and the preprocessing process comprises the following steps:
firstly, carrying out enhanced Lee filtering on the space target ISAR image;
and then carrying out contrast enhancement and energy normalization processing.
3. The spatial target ISAR image classification method under the condition of small samples based on XGboost optimized by multiple learners according to claim 1 or 2, wherein the specific structure of the capsule network is as follows:
(A) basic feature extraction module:
the basic feature extraction module comprises three convolution units, wherein each convolution unit comprises a convolution layer, a batch normalization layer and a ReLU activation layer;
(B) the vector feature extraction module:
the vector feature extraction module comprises 8 convolutional layers, wherein the 8 convolutional layers are in parallel relation, and the 8 convolutional layers do not use an activation function; the sizes of the convolution layers are the same, and the input feature graphs are respectively convolved;
then, composing the obtained result into a tensor form;
(C) dynamic routing layer
The dynamic routing layer is realized by adopting a dynamic routing method.
4. The method for classifying ISAR (inverse synthetic aperture radar) images of space targets under the condition of small samples based on XGboost optimized by multiple learners as claimed in claim 3, wherein the Alexenet network added with the attention mechanism module based on rotation invariance is as follows:
the attention mechanism module based on the rotation invariance comprises a rotation processing module and a space attention mechanism module;
a rotation processing module: based on an Alexnet network, performing rotation operation on the output characteristic diagram of the ReLU activation function of the second convolution layer to obtain a rotated characteristic diagram;
the spatial attention mechanism module comprises three convolution units, namely Conv1, Conv2 and Conv 3;
conv 1: the rotation characteristic diagram is firstly activated through the convolution layer and then through the ReLU;
conv2 comprises Conv2-1 and Conv2-2 subunits, the characteristic diagrams processed by Conv1 are respectively sent into Conv2-1 and Conv2-2 subunits, and in the Conv2-1 and Conv2-2 subunits, the convolution layers are firstly passed through and then deactivated by Softmax, and an output characteristic diagram F is obtained2-1And F2-2;
The Conv3 unit comprises Conv3-1 and Conv3-2 subunits, a rotation characteristic diagram obtained by the rotation processing module is directly and respectively sent into the Conv3-1 and Conv3-2 subunits, the convolution layers in the Conv3-1 and Conv3-2 subunits are firstly processed, then batch normalization processing is carried out, and G is obtained3-j,j=1,2;
For feature map F2-jWas subjected to transposition to obtain F'2-jIs prepared from'2-jAnd a characteristic diagram G3-jAnd random initialization coefficient alphajMultiplying and adding to obtain an output characteristic diagram
5. The method for spatial target ISAR image classification under small sample condition based on XGboost optimized by multiple learners according to claim 4, wherein the graph convolution network is as follows:
the graph convolution network comprises a feature extraction module and a graph neural network processing module;
(a) a feature extraction module:
the feature extraction module comprises three convolution units, each convolution unit comprising: the system comprises a convolution layer, a batch normalization layer, a maximum pooling layer and a ReLU activation function layer;
the last convolution unit outputs the feature expression of the image in a vector form, and combines the feature expression with the one-hot coding formed by the image category as the input of the graph neural network module;
(b) the figure neural network module:
the graph neural network module comprises two units for calculating the adjacent matrix and two graph convolution units; the first calculation adjacent matrix unit is connected with the first graph convolution unit, and then the second calculation adjacent matrix unit and the second graph convolution unit are connected;
the input of the first calculation adjacent matrix unit is the subtraction of node expression vectors of the feature extraction module to obtain a two-dimensional feature map; the input of the second calculation adjacent matrix unit is a characteristic graph obtained by subtracting the node expression vectors output by the first graph convolution unit; the two calculation adjacent matrix units respectively comprise three convolution modules, the first two convolution modules respectively comprise a convolution layer, a batch normalization layer and a ReLU activation layer, and the last convolution module only comprises one convolution layer;
the first graph convolution unit comprises a full connection layer and a batch normalization layer; the second graph convolution unit contains two fully connected layers and one batch normalization layer, and the output of the second graph convolution unit is equal to the classification probability of five types of targets.
6. The ISAR image classification method for the space targets under the condition of the small samples of the XGboost based on the multi-learner optimization according to claim 5, wherein the process of integrating the XGboost network model formed by the stacked networks by utilizing the XGboost algorithm comprises the following steps:
2.4.1 computing weights of the underlying learner
The input to the XGboost network is D { xi,α,li,ei,αWhere i 1,2, n denotes the number of meta-features, xi,αIs a meta feature, liIs a label, ei,αIs the weight of the element feature, alpha is 1,2,3 is the serial number of the basic learner, and introduces the weight q corresponding to the learnerαThe weight expression of the meta-feature is as follows:
wherein x isi,α,1、……,xi,α,NRespectively being meta-features xi,αThe elements of (1);is the mean of the meta-features; n is the total number of meta-features, N is the number of elements in a meta-feature;
2.4.2 calculation of weight to loss function partial derivatives
Defining the loss function L, including the predicted valuesAnd a label liCross entropy loss of l, and tree structure fkCorresponding regularization loss Ω ();
the loss function L is defined as:
l is defined as follows:
wherein γ is the regularization parameter for the number of leaves, T is the number of leaf nodes, λ is the regularization parameter for the leaf weights, and w represents the weights of the leaves;
l relates to qαPartial derivatives of (a):
2.4.3 adaptively updating weights
Updating the learner's weights with the partial derivatives and update rate η calculated above:
updating the weight corresponding to the meta-feature by the change of the weight of the learner, and realizing the self-adaptive updating of the weight;
2.4.4 calculation of network gain
The gain of the network is calculated from the first and second derivatives of the predicted values, gi,αAnd hi,αAre the first and second derivatives of the meta-feature, and L and R are the left and right nodes of the decision tree. The gain is defined as:
and (5) splitting the tree by the maximum gain to obtain a final classification result.
7. The method for classifying ISAR (inverse synthetic aperture radar) images of space targets under the condition of small samples of XGboost based on multi-learner optimization according to claim 6, wherein the training process of the XGboost network model comprises the following steps:
meta-feature output of graph convolution network: taking the classification probability of the data in the verification set as training meta-characteristics;
adding a rotation invariance-based attention mechanism module to the meta-feature output of the Alexnet network: combining the classification confidence degrees of the verification sample sets to form training meta-features;
meta-feature output of capsule network: taking the L2 norm of the output vector to obtain the classification probability, and taking the classification probability of the data in the verification sample set as the training element characteristic;
and training the XGboost network optimized by the multiple learners by using the training element characteristics.
8. The ISAR image classification method for the space target under the condition of the small sample of the XGboost based on the multi-learner optimization according to the claim 7, wherein the training processes of the graph convolution network, the capsule network and the Alexnet network are as follows:
the Alexnet network model training process comprises the following steps:
in the training process, each class of the training sample library is equally divided into K parts, wherein K-1 part is a training sample set, the rest part is a verification sample set, and K-fold cross verification is carried out to ensure that K parts of data are verified;
training an Alexnet network model by utilizing a training sample set and a verification sample set, and combining the classification confidence degrees of the verification sample set to form training meta-features;
the graph convolution network model training process comprises the following steps:
dividing a training sample library into a training set and a verification set, wherein each class of the training set has qshot images, and each class of the verification set has K-qshot images; in the training process, 5 xqshot +1 images of the network are input in each iteration, wherein 1 means that one image is taken from a verification set in sequence to be used as verification, and the label of the image is known and is used for supervised learning;
training the graph convolution network model by using a training set and a verification set, and taking the classification probability of data in the verification set as training meta-features;
the capsule network model training process comprises the following steps:
dividing a training sample library into a training set and a verification set, training a capsule network model by using the training set and the verification set, obtaining an L2 norm of an output vector to obtain a classification probability, and taking the classification probability of data in the verification sample set as a training meta-feature;
the ISAR spin space target sample is stored in the training sample library.
9. The ISAR image classification method for the space target under the condition of the small sample of the XGboost based on the multi-learner optimization according to claim 8, wherein the ISAR spin space target sample is obtained based on a three-dimensional space target model and a range-Doppler algorithm simulation.
10. The ISAR image classification method for the spatial target under the condition of the small samples of the XGboost based on the multi-learner optimization according to claim 9, wherein ISAR spin spatial target samples stored in the training sample library are preprocessed samples, and the preprocessing process comprises the following steps:
firstly, carrying out enhanced Lee filtering on the space target ISAR image;
and then carrying out contrast enhancement, energy normalization and data expansion, wherein the data expansion comprises left-right turning and up-down turning.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202111544817.1A CN114219960A (en) | 2021-12-16 | 2021-12-16 | Space target ISAR image classification method under small sample condition of XGboost based on multi-learner optimization |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202111544817.1A CN114219960A (en) | 2021-12-16 | 2021-12-16 | Space target ISAR image classification method under small sample condition of XGboost based on multi-learner optimization |
Publications (1)
Publication Number | Publication Date |
---|---|
CN114219960A true CN114219960A (en) | 2022-03-22 |
Family
ID=80703091
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202111544817.1A Pending CN114219960A (en) | 2021-12-16 | 2021-12-16 | Space target ISAR image classification method under small sample condition of XGboost based on multi-learner optimization |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN114219960A (en) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN114897835A (en) * | 2022-05-16 | 2022-08-12 | 中国矿业大学 | Coal product ash content real-time detection system based on image |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108717568A (en) * | 2018-05-16 | 2018-10-30 | 陕西师范大学 | A kind of image characteristics extraction and training method based on Three dimensional convolution neural network |
CN110020682A (en) * | 2019-03-29 | 2019-07-16 | 北京工商大学 | A kind of attention mechanism relationship comparison net model methodology based on small-sample learning |
AU2020103901A4 (en) * | 2020-12-04 | 2021-02-11 | Chongqing Normal University | Image Semantic Segmentation Method Based on Deep Full Convolutional Network and Conditional Random Field |
CN112766340A (en) * | 2021-01-11 | 2021-05-07 | 中山大学 | Depth capsule network image classification method and system based on adaptive spatial mode |
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 |
-
2021
- 2021-12-16 CN CN202111544817.1A patent/CN114219960A/en active Pending
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108717568A (en) * | 2018-05-16 | 2018-10-30 | 陕西师范大学 | A kind of image characteristics extraction and training method based on Three dimensional convolution neural network |
CN110020682A (en) * | 2019-03-29 | 2019-07-16 | 北京工商大学 | A kind of attention mechanism relationship comparison net model methodology based on small-sample learning |
AU2020103901A4 (en) * | 2020-12-04 | 2021-02-11 | Chongqing Normal University | Image Semantic Segmentation Method Based on Deep Full Convolutional Network and Conditional Random Field |
CN112766340A (en) * | 2021-01-11 | 2021-05-07 | 中山大学 | Depth capsule network image classification method and system based on adaptive spatial mode |
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 |
Non-Patent Citations (1)
Title |
---|
陈英义;龚川洋;刘烨琦;方晓敏;: "基于FTVGG16卷积神经网络的鱼类识别方法", 农业机械学报, no. 05, 28 February 2019 (2019-02-28) * |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN114897835A (en) * | 2022-05-16 | 2022-08-12 | 中国矿业大学 | Coal product ash content real-time detection system based on image |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109584248B (en) | Infrared target instance segmentation method based on feature fusion and dense connection network | |
CN112446398B (en) | Image classification method and device | |
Zhang et al. | Scene classification via a gradient boosting random convolutional network framework | |
Kanthi et al. | A 3D-deep CNN based feature extraction and hyperspectral image classification | |
CN110210515B (en) | Image data multi-label classification method | |
CN112639828A (en) | Data processing method, method and equipment for training neural network model | |
CN110097178A (en) | It is a kind of paid attention to based on entropy neural network model compression and accelerated method | |
CN112561027A (en) | Neural network architecture searching method, image processing method, device and storage medium | |
CN116052016A (en) | Fine segmentation detection method for remote sensing image cloud and cloud shadow based on deep learning | |
CN110222718B (en) | Image processing method and device | |
CN107316004A (en) | Space Target Recognition based on deep learning | |
Napoli et al. | Simplified firefly algorithm for 2d image key-points search | |
Singh et al. | Steganalysis of digital images using deep fractal network | |
CN112837215B (en) | Image shape transformation method based on generation countermeasure network | |
CN113627472A (en) | Intelligent garden defoliating pest identification method based on layered deep learning model | |
CN110660020A (en) | Image super-resolution method of countermeasure generation network based on fusion mutual information | |
CN112699838B (en) | Hyperspectral mixed pixel nonlinear blind decomposition method based on spectral diagnosis characteristic weighting | |
CN114913379B (en) | Remote sensing image small sample scene classification method based on multitasking dynamic contrast learning | |
CN114077891B (en) | Training method of style conversion model and training method of virtual building detection model | |
Han et al. | Multi-scale feature network for few-shot learning | |
CN116097307A (en) | Image processing method and related equipment | |
CN114170659A (en) | Facial emotion recognition method based on attention mechanism | |
Li et al. | A method to establish a synthetic image dataset of stored-product insects for insect detection | |
CN117237547B (en) | Image reconstruction method, reconstruction model processing method and device | |
CN111191704A (en) | Foundation cloud classification method based on task graph convolutional network |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination |