CN113486917B - Radar HRRP small sample target recognition method based on metric learning - Google Patents
Radar HRRP small sample target recognition method based on metric learning Download PDFInfo
- Publication number
- CN113486917B CN113486917B CN202110535025.1A CN202110535025A CN113486917B CN 113486917 B CN113486917 B CN 113486917B CN 202110535025 A CN202110535025 A CN 202110535025A CN 113486917 B CN113486917 B CN 113486917B
- Authority
- CN
- China
- Prior art keywords
- hrrp
- layer
- category
- neural network
- convolutional neural
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Classifications
-
- 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
- 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
-
- 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
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
- Y02A90/00—Technologies having an indirect contribution to adaptation to climate change
- Y02A90/10—Information and communication technologies [ICT] supporting adaptation to climate change, e.g. for weather forecasting or climate simulation
Abstract
The invention discloses a radar HRRP small sample target recognition method based on metric learning, which comprises the steps of firstly constructing a multi-target HRRP sample set, training out a feature extraction model through a convolutional neural network based on a feature adaptation layer and a conversion operation layer improvement, calculating a center point of each category by using the extracted features, constructing a loss function by using a metric function, and realizing non-cooperative small sample target recognition by using a full-connection layer training classifier based on gradient optimization by using HRRP feature data of non-cooperative small sample targets after feature extraction. The convolutional neural network is improved by constructing the feature adaptation layer and the conversion operation layer, so that the generalization capability of the feature extractor is further improved.
Description
Technical Field
The invention belongs to the technical field of radar target recognition, and particularly relates to a radar HRRP small sample target recognition method based on measurement learning.
Background
In many fields such as military exploration, due to many factors such as low observation rate, the data sample size of non-cooperative targets is often extremely small, which is an important cause of inaccurate recognition accuracy of the radar to the non-cooperative (out-of-library) targets.
At present, a great amount of data with labels are often required for training by a fire and heat deep learning method, and the problem of non-cooperative target identification cannot be directly solved by a general deep learning method due to the serious shortage of sample size of the non-cooperative target with labels, and the problem of the non-cooperative target identification is also caused to be typical small sample identification. Compared with the existing category, the number of the samples of the newly added category is far smaller than that of the existing category, so that the model learning problem of unbalanced samples in general deep learning is caused, and the recognition accuracy of non-cooperative targets is directly low.
The traditional target recognition method mostly builds a shallow probability model through known target samples in a library based on a statistical method, realizes target recognition through template matching, and the traditional target recognition method based on a deep learning technology carries out supervised training through a large amount of tag data so as to acquire a target recognition model, the mechanism often needs to mark a large amount of tag data, and for new types, the model often needs to be retrained, the calculation cost is extremely high, and the recognition accuracy of small sample non-cooperative targets is seriously insufficient, so that the method has extremely limited, and the prior art proposes a method for building a dynamic adjustment layer through a preprocessed HRRP (High Resolution Range Profile, high-resolution range profile) sample; then, dividing HRRP by utilizing sliding window size, modeling time sequence correlation of samples by bidirectional stacking RNNs, adjusting importance degree of hidden layer state by adopting a multi-level attention mechanism, and finally, classifying targets by Softmax to realize target recognition, for example, the application of patent application with publication number of CN 111736125A, named as a radar target recognition method based on the attention mechanism and the bidirectional stacking cyclic neural network, discloses a target recognition method based on deep learning. The method comprises the steps of preprocessing to reduce sensitivity in an HRRP sample and establishing a dynamic adjustment layer; selecting a sliding window size to split the HRRP; adjusting the importance degree of each segmentation sequence through an importance network; modeling the time sequence correlation of the samples through the two-way stacking RNNs, and extracting high-level features of the samples; the importance degree of the hidden layer state is adjusted by adopting a multi-level attention mechanism; target classification was performed by Softmax. The method successfully solves the problems that the radar HRRP feature extraction process is unsupervised, the information is damaged, and the feature extraction method is selected highly depending on the cognition and experience accumulation of researchers on the HRRP data, and plays a certain role in improving the accuracy of radar HRRP target identification. However, this method has the following disadvantages: 1. the method is effective only for the problem of low accuracy of identifying small sample targets in the library, and does not consider the problem of identifying small sample targets in a non-cooperative way (a radar target identifying method based on an attention mechanism and a bidirectional stacking cyclic neural network); 2. the method still adopts the deep learning existing model training mechanism based on a large amount of label data, and for the problems of low observation rate of non-cooperative (out-of-library) targets and extremely small sample size with labels, the direct use of the model training mechanism can cause unbalance of model learning, so that the recognition rate of the non-cooperative targets is not guaranteed.
Therefore, providing a method capable of improving accuracy of identifying non-cooperative targets is a need to be solved.
Disclosure of Invention
In order to solve the problems in the prior art, the invention provides a radar HRRP small sample target identification method based on measurement learning. The technical problems to be solved by the invention are realized by the following technical scheme:
a radar HRRP small sample target identification method based on metric learning comprises the following steps:
step 4, inputting the intra-library cooperative target HRRP training sample set into a convolutional neural network to obtain a center of each category, wherein the convolutional neural network comprises three cascaded convolutional neural network modules, each convolutional neural network module comprises a cascaded convolutional layer, a Relu layer and a pooling layer, the convolutional neural network further comprises a characteristic adaptation layer and three conversion operation layers, and the characteristic adaptation layer is respectively connected with the three convolutional neural network modules through the three conversion operation layers;
step 5, obtaining a loss function according to the center of each category and the output result of the convolutional neural network;
step 6, back propagation is carried out by utilizing the loss function, so that the loss function is converged to obtain a feature extractor;
step 7, extracting features of the multi-category non-cooperative target small sample training set through the feature extractor to obtain feature data of the non-cooperative targets of the categories;
and 8, training the feature data of the non-cooperative targets of the multiple categories by adopting a gradient optimized full-connection layer network to obtain a classifier, so as to perform target identification by using the classifier.
In one embodiment of the present invention, the step 1 includes:
step 1.1, carrying out average division on azimuth angles of 0-90 degrees under the same pitch angle to obtain n angular domains;
step 1.2, continuously acquiring n radar echo signals of multiple categories in the angular domain, and dividing the acquired radar echo signals into multiple sections of sub echo signals on average;
and 1.3, carrying out FFT processing on the sub-echo signals of each category to obtain the one-dimensional range profile signals, wherein all the one-dimensional range profile signals of different categories form the multi-category HRRP sample set.
In one embodiment of the present invention, the step 2 includes:
step 2.1, processing the HRRP sample set of each category by using an energy normalization method to obtain an energy normalized HRRP sample set;
and 2.2, performing alignment treatment on the HRRP sample set subjected to energy normalization by using a gravity center alignment method to obtain an effective HRRP sample set of each category.
In one embodiment of the present invention, the step 3 includes:
and randomly extracting the effective HRRP sample sets of the K categories to obtain a cooperative target HRRP training sample set in the library.
In one embodiment of the present invention, the feature adaptation layer includes five cascaded convolutional neural network modules, an average pooling layer is connected after the five cascaded convolutional neural network modules, and then an average value is obtained from the output of the average pooling layer, so as to construct a coding layer, and the coding layer is then cascaded with a first linear layer, a first relu layer, a second linear layer, a second relu layer, a third linear layer and a third relu layer in sequence.
In one embodiment of the present invention, the transformation operation layer performs linear change on the feature adaptation layer, and the result of the linear change is input to the corresponding convolutional neural network module.
In one embodiment of the present invention, the step 5 includes:
step 5.1, obtaining the distance between the center of each category and the convolutional neural network based on a distance measurement function constructed based on the Euclidean distance of the center of each category and the output result of the convolutional neural network;
step 5.2, obtaining probability distribution of the intra-library cooperative target HRRP training sample set according to the distances between the center of each category and the convolutional neural network based on a softmax function;
and 5.3, obtaining the loss function according to the probability distribution.
In one embodiment of the present invention, the step 7 includes:
step 7.1, pair K c Randomly extracting the effective HRRP sample sets of the individual categories to obtain a multi-category non-cooperative target small sample test set and a multi-category non-cooperative target small sample training set, wherein K and K are c Non-intersecting;
and 7.2, extracting the characteristics of the multi-category non-cooperative target small sample training set through the characteristic extractor to obtain characteristic data of the non-cooperative targets of a plurality of categories.
In one embodiment of the present invention, after the step 8, the method further includes:
and based on the recognition accuracy evaluation model, evaluating the recognition accuracy of the classifier by using the multi-category non-cooperative target small sample test set to obtain an evaluation result.
The invention has the beneficial effects that:
the convolutional neural network is improved by constructing the feature adaptation layer and the conversion operation layer, so that the generalization capability of the feature extractor is further improved.
According to the method, a metric function is constructed by the Euclidean distance calculation method, a loss function is constructed by the metric function, and the distance between the prediction center of the training sample and the label is measured, so that the feature extraction model is continuously optimized, and compared with the prior art, the accuracy of the target recognition model is effectively improved;
according to the invention, a plurality of small sample recognition tasks are constructed to perform model fine adjustment in batches, so that a feature extraction model is trained, and the problem of unbalance of model learning under the condition that the new class of target sample size is limited by the existing feature extraction method based on deep learning is solved.
The present invention will be described in further detail with reference to the accompanying drawings and examples.
Drawings
Fig. 1 is a schematic flow chart of a radar HRRP small sample target recognition method based on metric learning according to an embodiment of the present invention;
FIG. 2 is a flow chart of another method for identifying radar HRRP small sample targets based on metric learning according to an embodiment of the invention;
FIG. 3 is a schematic diagram of a convolutional neural network provided by an embodiment of the present invention;
fig. 4 is a schematic diagram of a part of an aircraft target three-dimensional model and a simulated one-dimensional range profile according to an embodiment of the present invention.
Detailed Description
The present invention will be described in further detail with reference to specific examples, but embodiments of the present invention are not limited thereto.
Example 1
Referring to fig. 1 and fig. 2, fig. 1 is a schematic flow chart of a method for identifying a radar HRRP small sample target based on metric learning according to an embodiment of the present invention, and fig. 2 is a schematic flow chart of another method for identifying a radar HRRP small sample target based on metric learning according to an embodiment of the present invention. The embodiment of the invention provides a radar HRRP small sample target recognition method based on metric learning, which comprises the following steps of 1 to 8, wherein:
In this embodiment, the HRRP sample set of multiple classes includes multiple HRRP sample sets of different classes, and multiple classes refer to different classes of aircraft, such as jac 38, gust F3.
In a specific embodiment, step 1 specifically includes steps 1.1 to 1.3, wherein:
and 1.1, carrying out average division on azimuth angles of 0-90 degrees under the same pitch angle to obtain n angular domains.
Preferably, n has a value in the range of 5 to 20, for example, n has a value of 10.
Step 1.2, continuously acquiring multiple types of radar echo signals in n angular domains, and dividing the acquired radar echo signals into multiple sections of sub-echo signals on average.
In this embodiment, the number of sub-echo signals in each angular region is L/n, where L is the total sample size of the sub-echo signals, for example, the number of sub-echo signals in each angular region is 320 segments, where L is equal to or greater than 2000.
And 1.3, performing FFT (Fast Fourier Transform ) processing on the sub-echo signals of each category to obtain one-dimensional range profile signals, wherein all the one-dimensional range profile signals of different categories form a multi-category HRRP sample set.
In this embodiment, each sub-echo signal obtained in step 1.2 is subjected to FFT processing, and then a one-dimensional range profile signal is obtained, so that each class includes a plurality of one-dimensional range profile signals, and all the one-dimensional range profile signals form a multi-class HRRP sample set.
And 2, processing the HRRP sample set of each category to obtain an effective HRRP sample set of each category.
In a specific embodiment, step 2 specifically includes steps 2.1-2.2, wherein:
and 2.1, processing the HRRP sample set of each category by using an energy normalization method to obtain an energy normalized HRRP sample set.
In this embodiment, due to the influence of factors such as the transmitter power, the gain of the transmitting and receiving antennas, the target distance, and the antennas, the signals of the obtained high-resolution range profile are different, so that the energy normalization processing method is used to normalize each one-dimensional range profile signal in the effective HRRP sample set obtained in step 1, so as to obtain an HRRP sample set after energy normalization.
And 2.2, performing alignment treatment on the HRRP sample set subjected to energy normalization by using a gravity center alignment method to obtain an effective HRRP sample set of each category.
In this embodiment, aiming at the problem of translational sensitivity, a gravity center alignment method in an absolute alignment method is adopted to perform alignment processing on the one-dimensional range profile signals in the HRRP sample set after normalization in step 2.1, so that an effective HRRP sample set after preprocessing can be obtained.
And 3, constructing an intra-library cooperative target HRRP training sample set by using the effective HRRP sample sets of all the classes, wherein K classes are shared by the classes of the intra-library cooperative target HRRP training sample set.
Specifically, the valid HRRP sample sets of the K categories are randomly extracted to obtain a cooperative target HRRP training sample set in the library.
In this embodiment, K classes are selected from the valid HRRP sample sets of all classes, and then a number of samples are randomly extracted from the valid HRRP sample sets of each class, so that the samples extracted from all classes form a intra-library collaborative target HRRP training sample set H, h= { (x) 1 ，y 1 ),...,(x N ,y N ) -wherein there are K samples, N samples, x i ∈R D ，R D As a D-dimensional real number set, y i E { 1..k } is a label, which is used to label the aircraft class to which the sample corresponds, for example, target samples of 45 classes of different classes of aircrafts are constructed, 3200 samples are taken in each class, and the number of step frequency points of each sample is 256.
Step 4, please refer to fig. 3, the intra-library cooperative target HRRP training sample set is input to a convolutional neural network to obtain the center of each class, the convolutional neural network comprises three cascaded convolutional neural network modules, each convolutional neural network module comprises a cascaded convolutional layer, a Relu layer and a pooling layer, the convolutional neural network further comprises a characteristic adaptation layer and three conversion operation Layers, the characteristic adaptation layer is respectively connected with the three convolutional neural network modules through the three conversion operation Layers, wherein the three cascaded convolutional neural network modules are respectively marked as Block1, block2 and Block3 in fig. 3, and the three conversion operation Layers are respectively marked as Film Layers1, film Layers2 and Film Layers3 in fig. 3.
In a specific embodiment, step 4 specifically includes steps 4.1-4.2, wherein:
and 4.1, inputting the intra-library cooperative target HRRP training sample set into a convolutional neural network to obtain an output result.
In this embodiment, three cascaded convolutional neural network modules are built, each convolutional neural network module includes a cascaded convolutional layer, a Relu layer and a pooling layer, and the calculation process of each convolutional neural network module is shown in the following formula:
wherein f j Representing the output of each convolutional neural network module, b j Representing the corresponding rule of nonlinear function of each convolutional neural network module, j= {1,2,3}, S K The sample size of the targets representing the kth category.
In order to enable three convolutional neural network modules of the convolutional neural network to better learn multi-modal information of different HRRP recognition task data, the embodiment constructs a feature extractor added with a feature adaptation layer and a conversion operation layer, wherein two vectors gamma are output to each convolutional neural network module through the feature adaptation layer operation j And beta j The specific formula is as follows:
wherein f adapation Parameters representing the feature adaptation layer. Then use vector gamma j And beta j And connecting a conversion operation layer to calculate after each convolutional neural network module, wherein the calculation is specifically shown as the following formula:
D j ＝F(f j ；γ j ,β j )＝γ j f j +β j
wherein D is j Representing the output value after conversion of each layer of convolutional neural network module, f j Representing results of the intra-library collaborative target HRRP training sample set after passing through the convolutional neural network module.
Further, the feature adaptation layer comprises five cascaded convolutional neural network modules, an average pooling layer is connected after the five cascaded convolutional neural network modules, an average value is obtained from the output of the average pooling layer, and then a coding layer is constructed, and the coding layer is sequentially cascaded with a first linear layer, a first relu layer, a second linear layer, a second relu layer, a third linear layer and a third relu layer.
The feature adaptation layer of this embodiment constructs five cascaded convolutional neural network modules, the intra-library cooperative target HRRP training sample set sequentially passes through the five cascaded convolutional neural network modules, then the output of the last convolutional neural network module enters an average pooling layer, and then the average value of the output of the average pooling layer is calculated, so as to construct a coding layer, and the output of the coding layer is defined as z j ＝g f (S), j= {1,2,3}. Output z of the coding layer j First enter the first linear layer and output as z j1 Then (z) j +z j1 ) Enter the first relu layer output z j2 ，z j2 Into a second linear layer, output is z j3 Then (z) j2 +z j3 ) Into the second relu layer output z j4 ，z j4 Into the third linear layer, output is z j5 Then (z) j4 +z j5 ) Into the third relu layer output z j6 After z j6 And entering a jth convolutional neural network module of the convolutional neural network through a conversion operation layer.
Further, concatenating the first linear layer, the first relu layer, the second linear layer, the second relu layer, the third linear layer, and the third relu layer performs a linear transformation on a final output of the encoded layer, as shown in the following formula:
wherein, the liquid crystal display device comprises a liquid crystal display device,
representing the output of the third relu layer, i.e. z of the above j6 。
Thereafter, enter
Conversion operation layer calculates vector gamma j And beta j Vector gamma j And beta j The calculation method of (2) is shown as follows:
wherein r represents a set of vectors subject to a normal distribution with a mean of 0 and a variance of 0.001, the dimensions of which are equal to z j Same, h represents a unit vector, dimension and
the same applies.
Therefore, the conversion operation layer inputs the output of the conversion operation layer to the corresponding convolutional neural network module of the convolutional neural network, and the final output value of the convolutional neural network module which passes through the three cascade connection of the characteristic adaptation layer and the conversion operation layer is marked as D 3 At f φ (-) represents the algorithm of the feature extractor constructed by three cascaded convolutional neural network modules, a feature adaptation layer and a conversion operation layer, and the calculation process can be expressed as follows:
and 4.2, obtaining the center of each category according to the output result of the convolutional neural network based on the center calculation model of each category.
Specifically, the output result D of the convolutional neural network is utilized 3 Acquiring center c of each category k ，x i ∈R D ，D 3 ∈R M The per class center calculation model is shown as follows:
wherein c k ∈R M ，R M Representing an M-dimensional real set.
And step 5, obtaining a loss function according to the output result of the center and the convolutional neural network of each category.
In a specific embodiment, step 5 specifically includes steps 5.1 to 5.3, wherein:
and 5.1, obtaining the distance between the center of each category and the convolutional neural network by a distance measurement function constructed based on the Euclidean distance of the output result of the center of each category and the convolutional neural network.
Specifically, a distance metric function is constructed using Euclidean distance, as shown in the following equation,
wherein f φ Representing features extracted through a convolutional neural network, M representing the dimensions of the features, and calculating the pass f using the distance metric function f (.) distance between the sample point at which feature extraction is performed and the center of each category.
And 5.2, obtaining probability distribution of the intra-library cooperative target HRRP training sample set according to the distances between the center of each category and the convolutional neural network based on the softmax function.
Specifically, the probability distribution of the test sample is obtained by a modified softmax function, which is specifically shown in the following formula:
wherein p is f (y=k|x) is the probability distribution of the intra-library collaborative target HRRP training sample set, c k′ Is the center of the other class samples than the currently calculated class k sample.
And 5.3, obtaining the loss function according to probability distribution.
Define the loss function as J (f) = -logp f (y=k|x), this formula can be further modified into the following formula:
J(f)＝d(f f (x)，c k )+log∑ k′ exp(-d(f f (x)，c k ))。
and 6, back propagation is carried out by using the loss function, so that the loss function is converged, and the feature extractor is obtained.
Specifically, the loss function is used for calculating the loss value, the loss value trained by the model is back-propagated until the loss value converges to the minimum, and therefore the convolutional neural network model with stronger generalization performance can be obtained, and then the final convolutional neural network is used as a final feature extractor.
And 7, extracting features of the multi-category non-cooperative target small sample training set through a feature extractor to obtain feature data of the non-cooperative targets of the categories.
In a specific embodiment, step 7 specifically includes steps 7.1-7.2, wherein:
step 7.1, pair K c Randomly extracting each class of effective HRRP sample set to obtain a multi-class non-cooperative target small sample test set and a multi-class non-cooperative target small sample training set, wherein K and K are c Does not intersect.
In this embodiment, K is first selected from the valid HRRP sample set of all the classes obtained in step 2 c Class, then from K c Several samples are randomly extracted in the valid HRRP sample set for each of the classes such that the samples extracted for all classes constitute a multi-class non-cooperative target small sample training set, e.g., N c Class 5, N and N c Disjoint, i.e. the 5 classes of aircraft targets are quite different from the 35 classes of targets used for previous training, are new target classes that have not been seen, e.g. 1, 5, 10 samples are randomly extracted from the valid HRRP sample set of each class to form a multi-class non-cooperative target small sample training set, and then N is taken from c Randomly extracting a plurality of samples from the valid HRRP sample set of each class in the class, so that the samples extracted from all classes form a multi-class non-cooperative target small sample test set, for example, randomly extracting 15 samples from the valid HRRP sample set of each class form a multi-class non-cooperative target small sample training set, forming a multi-class non-cooperative target small sample task set by the multi-class non-cooperative target small sample training set and the multi-class non-cooperative target small sample test set, defining the small sample task set as
In which there is N c Non-collaborative off-library sample-like +.>
Referred to as meta-training set (i.e. multi-class non-cooperative target small sample training set), +.>
Sample size S1, 5, 10 were tested separately for each category,/o>
Referred to as meta-test set (i.e. multi-class non-cooperative target small sample test set), +.>
Sample size Q15,/for each class>
A training set and test set of one test task is described, together with J training tasks, e.g., a total of 600 small sample recognition tasks.
And 7.2, extracting features of the multi-category non-cooperative target small sample training set through a feature extractor to obtain feature data of the non-cooperative targets of the categories.
Specifically, the feature extractor trained in the steps performs feature extraction on the multi-category non-cooperative target small sample training set so as to obtain feature data of the non-cooperative targets of the categories.
And 8, training the feature data of the non-cooperative targets of the multiple categories by adopting a gradient optimized full-connection layer network to obtain a classifier, so as to perform target identification by using the classifier.
Specifically, the classifier training is performed by adopting the gradient-optimized full-connection layer network through the feature data of the non-cooperative targets of the multiple categories, which is acquired in the step 7, and the process is as follows:
wherein is L ce Represents a cross entropy loss function, phi c Representing parameters of the full connection layer classifier.
In a specific embodiment, after step 8, the method may further include:
based on the recognition accuracy evaluation model, the recognition accuracy of the classifier is evaluated by utilizing a multi-class non-cooperative target small sample test set to obtain an evaluation result, wherein the recognition accuracy evaluation model is as follows:
wherein L is meta Representing the loss function, m representing the average class accuracy (Mean Class Accuracy, MCA), the specific calculation method is:
wherein M represents average class accuracy, N represents class number, and M i Representing the sample size of class i, T i The number of correct classifications in class i is indicated.
The technical effects of the invention are further described by the following actual detection method:
1) In the embodiment, the modeling platform CPU is i7-4770, the main frequency of the CPU is 3.40GHz, the memory is 32GB, the video card is NVIDA GeForce GTX1070, the video memory is 6G, and the system is Windows 7 (64 bit).
2) The compiling environment is Pycharm Community 2020.1 and Matlab R2015a, python is 3.7.1. The frames used were: pyTorch 1.1.0, cuda 9.0, the main libraries used were: numpy 1.19.1, scikit-learn 0.23.2,Scipy 1.5.2,Matplotlib 3.3.2.
3) And (3) data simulation: in order to verify the effectiveness of the HRRP-based non-cooperative target small sample identification method, firstly, a 50-class aircraft 3D model is constructed through three-dimensional drawing software Solidworks2018 (for convenience, a required aircraft 3D model can also be directly downloaded from a fly away simulation functional network), then electromagnetic simulation is carried out on the 50-class aircraft model through high-frequency electromagnetic computing software CST study SUITE to obtain broadband electromagnetic scattering data of an aircraft target, fourier transformation (FFT) is carried out on the data, and further a one-dimensional range image (HRRP) of the aircraft model is obtained, and simulation parameters are shown in table 1. Electromagnetic calculation is carried out on 50 types of airplanes under 84-degree pitch angles, 3200 HRRP samples are arranged on each pitch angle of each type of target, the azimuth angle of each type of target covers 10.05-90 degrees, and fig. 4 shows a three-dimensional model and a corresponding one-dimensional range profile of part of airplanes in the 50 types of airplanes.
Table 1 electromagnetic scattering calculation parameters
In order to overcome the problem of HRRP sensitivity, the method carries out energy normalization and center of gravity alignment treatment on the obtained one-dimensional range profile data to obtain a preprocessed one-dimensional range profile;
4) From the class 50 targets, 35 classes are selected as training samples (all samples in each class), 5 classes are selected as test samples (all samples in each class), and the 5 classes are not intersected with the 35 classes of training samples.
The training set sample size S is 1, 5 and 10 respectively, that is, sample data with dimensions of 10000×35×1×256, 10000×35×5×256, 10000×35×10×256 are respectively constructed, that is, 10000 small sample training tasks, and 35 types of samples in each small sample training task randomly extract 1, 5 and 10 samples from 3200 samples in each type respectively.
The verification set is constructed as 10000×35×15×256 sample data, namely 10000 small sample verification tasks, 35 classes in each small sample training task randomly extract 15 samples from 3200 samples in each class, and the samples are not intersected with the training set samples.
Sample data of the meta-training set sample size S of the test set, namely 600×5×1×256, 600×5×5×256 and 600×5×10×256, are respectively taken as 1, 5 and 10, i.e. 600 small sample training tasks, and 5 types (non-cooperative targets) in each small sample training task randomly extract 1, 5 and 10 samples from 3200 samples in each type.
The meta-test set in the test set is constructed as 600×5×15×256 sample data, i.e., 600 small sample test tasks, and 5 classes (non-cooperative targets) in each small sample training task randomly extract 15 samples (and) from 3200 samples in each class, respectively, for evaluating the classifier performance of the new class trained on the support set. There is no overlap between samples of the meta training set and the meta test set.
5) And (3) setting a model structure: a training phase and a testing phase, each small sample training task trains a feature extractor for the task. The feature extractor adopts a three-layer convolution neural network: the number of input channels of the first layer is 1, the number of output channels is 32, the step size is 9, and the padding mode is zero padding; the output channel of the second layer is 64, the step length is 9, and the padding mode is zero padding; the third layer output channel is 128, the step size is 9, and the padding mode is zero padding. Finally, the convolution output characteristics are pulled Cheng Xiangliang through the full connection layer of 300 by the flat operation. The classifier in the training stage is a full connection layer from 300 dimension to 35 dimension; finally, euclidean distance is adopted to measure the distance between the sample point and each class representation, and the probability distribution is obtained by utilizing softmax;
6) Comparison analysis the recognition results of non-cooperative (ex-library) targets under the conditions of K equal to 1, 5 and 10 samples, and compared with the conventional SVM, logistic, AGC and CNN-FC methods, the experimental results were as follows:
table 2 comparison of accuracy results of small sample identification methods
In one recognition task, the model recognizes 5 types of non-cooperative aircraft targets, and the recognition accuracy is 20% under the condition that the model is not trained to make random guesses. As shown in Table 2, SVM and Logistic have the capability of identifying small samples of non-cooperative targets, and the identification capability is obviously improved as the number of samples increases. Taking SVM as an example, the performance of each class of 10 samples is improved by more than 19% compared with the performance of each class of 1 sample; the AGC has a poor covariance matrix estimation due to a much higher sample dimension than the number of samples, and its performance is much lower than the baseline model of the other two shallow layers. The recognition accuracy of the CNN-FC method based on the non-cooperative target fine-tuning classifier is about 20% under the condition that the number of samples in each class is not more than 10, so that the CNN-FC method does not have recognition capability. On the other hand, compared with the shallow model SVM and Logistic, the CNN-Metric-FC method provided by the invention has the advantages that the accuracy rate is improved by more than 14%, 10% and 3% in experiments of 1 sample, 5 samples and 10 samples, so that the CNN-Metric-FC method provided by the invention has higher identification accuracy rate and higher effectiveness and robustness compared with methods such as SVM and Logistic under the condition of small samples with non-cooperative targets.
The technical idea of the invention is as follows: firstly, a multi-target HRRP sample set is constructed, a feature extraction model is trained through a convolutional neural network based on the improvement of a feature adaptation layer and a conversion operation layer, the center point of each category is calculated by using the extracted features, then a loss function is constructed by using a metric function, and a full-connection layer training classifier based on gradient optimization is utilized by using HRRP feature data of a non-cooperative small sample target after feature extraction, so that non-cooperative small sample target identification is realized.
The convolutional neural network is improved by constructing the feature adaptation layer and the conversion operation layer, so that the generalization capability of the feature extractor is further improved.
According to the method, a metric function is constructed by the Euclidean distance calculation method, a loss function is constructed by the metric function, and the distance between the prediction center of the training sample and the label is measured, so that the feature extraction model is continuously optimized, and compared with the prior art, the accuracy of the target recognition model is effectively improved;
according to the invention, a plurality of small sample recognition tasks are constructed to perform model fine adjustment in batches, so that a feature extraction model is trained, and the problem of unbalance of model learning under the condition that the new class of target sample size is limited by the existing feature extraction method based on deep learning is solved.
In the description of the present specification, a description referring to terms "one embodiment," "some embodiments," "examples," "specific examples," or "some examples," etc., means that a particular feature, structure, material, or characteristic described in connection with the embodiment or example is included in at least one embodiment or example of the present invention. In this specification, schematic representations of the above terms are not necessarily directed to the same embodiment or example. Furthermore, the particular features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples. Further, one skilled in the art can engage and combine the different embodiments or examples described in this specification.
Although the present application has been described herein in connection with various embodiments, other variations to the disclosed embodiments can be understood and effected by those skilled in the art in practicing the claimed application, from a review of the figures, the disclosure, and the appended claims. In the claims, the word "comprising" does not exclude other elements or steps, and the "a" or "an" does not exclude a plurality. A single processor or other unit may fulfill the functions of several items recited in the claims. The mere fact that certain measures are recited in mutually different dependent claims does not indicate that a combination of these measures cannot be used to advantage.
The foregoing is a further detailed description of the invention in connection with the preferred embodiments, and it is not intended that the invention be limited to the specific embodiments described. It will be apparent to those skilled in the art that several simple deductions or substitutions may be made without departing from the spirit of the invention, and these should be considered to be within the scope of the invention.
Claims (6)
1. The radar HRRP small sample target identification method based on metric learning is characterized by comprising the following steps of:
step 1, constructing a multi-category HRRP sample set, wherein each category of the HRRP sample set in the multi-category HRRP sample set comprises a plurality of one-dimensional range profile signals; carrying out average division on azimuth angles of 0-90 degrees under the same pitch angle to obtain n angular domains; continuously acquiring n radar echo signals of multiple categories in the angular domain, and dividing the acquired radar echo signals into multiple sections of sub echo signals on average; carrying out FFT processing on the sub echo signals of each category to obtain the one-dimensional range profile signals, wherein all the one-dimensional range profile signals of different categories form the multi-category HRRP sample set;
step 2, processing each class of HRRP sample set to obtain an effective HRRP sample set of each class;
step 3, constructing an intra-library cooperative target HRRP training sample set by utilizing the effective HRRP sample sets of all the classes, wherein K classes are shared by the classes of the intra-library cooperative target HRRP training sample set;
step 4, inputting the intra-library cooperative target HRRP training sample set into a convolutional neural network to obtain a center of each category, wherein the convolutional neural network comprises three cascaded convolutional neural network modules, each convolutional neural network module comprises a cascaded convolutional layer, a Relu layer and a pooling layer, the convolutional neural network further comprises a characteristic adaptation layer and three conversion operation layers, and the characteristic adaptation layer is respectively connected with the three convolutional neural network modules through the three conversion operation layers; the characteristic adaptation layer comprises five cascaded convolutional neural network modules, an average pooling layer is connected behind the five cascaded convolutional neural network modules, the average value of the output of the average pooling layer is obtained, and then a coding layer is constructed, and the coding layer is sequentially cascaded with a first linear layer, a first relu layer, a second linear layer, a second relu layer, a third linear layer and a third relu layer; the conversion operation layer carries out linear change on the characteristic adaptation layer, and a result of the linear change is input to the corresponding convolutional neural network module;
step 5, obtaining the distance between the center of each category and the convolutional neural network based on a distance measurement function constructed based on the Euclidean distance of the center of each category and the output result of the convolutional neural network; based on a softmax function, obtaining probability distribution of the intra-library cooperative target HRRP training sample set according to the distance between the center of each category and the convolutional neural network; obtaining a loss function according to the probability distribution;
step 6, back propagation is carried out by utilizing the loss function, so that the loss function is converged to obtain a feature extractor;
step 7, extracting features of the multi-category non-cooperative target small sample training set through the feature extractor to obtain feature data of the non-cooperative targets of the categories;
and 8, training the feature data of the non-cooperative targets of the multiple categories by adopting a gradient optimized full-connection layer network to obtain a classifier, so as to perform target identification by using the classifier.
2. The method for identifying a radar HRRP small sample target according to claim 1 wherein said step 2 comprises:
step 2.1, processing the HRRP sample set of each category by using an energy normalization method to obtain an energy normalized HRRP sample set;
and 2.2, performing alignment treatment on the HRRP sample set subjected to energy normalization by using a gravity center alignment method to obtain an effective HRRP sample set of each category.
3. The method for identifying a radar HRRP small sample target according to claim 1 wherein said step 3 comprises:
and randomly extracting the effective HRRP sample sets of the K categories to obtain a cooperative target HRRP training sample set in the library.
4. The method for identifying a radar HRRP small sample target according to claim 1 wherein said step 4 comprises:
step 4.1, inputting the intra-library cooperative target HRRP training sample set into a convolutional neural network to obtain an output result;
and 4.2, obtaining the center of each category according to the output result of the convolutional neural network based on the center calculation model of each category.
5. The method for identifying a radar HRRP small sample target according to claim 1 wherein said step 7 comprises:
step 7.1, pair K c Randomly extracting the effective HRRP sample sets of the individual categories to obtain a multi-category non-cooperative target small sample test set and a multi-category non-cooperative target small sample training set, wherein K and K are c Non-intersecting;
and 7.2, extracting the characteristics of the multi-category non-cooperative target small sample training set through the characteristic extractor to obtain characteristic data of the non-cooperative targets of a plurality of categories.
6. The method for radar HRRP small sample target identification of claim 5 wherein after step 8, further comprising:
and based on the recognition accuracy evaluation model, evaluating the recognition accuracy of the classifier by using the multi-category non-cooperative target small sample test set to obtain an evaluation result.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110535025.1A CN113486917B (en) | 2021-05-17 | 2021-05-17 | Radar HRRP small sample target recognition method based on metric learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110535025.1A CN113486917B (en) | 2021-05-17 | 2021-05-17 | Radar HRRP small sample target recognition method based on metric learning |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN113486917A CN113486917A (en) | 2021-10-08 |
| CN113486917B true CN113486917B (en) | 2023-06-02 |
Family
ID=77933374
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110535025.1A Active CN113486917B (en) | 2021-05-17 | 2021-05-17 | Radar HRRP small sample target recognition method based on metric learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113486917B (en) |
Families Citing this family (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114488140B (en) * | 2022-01-24 | 2023-04-25 | 电子科技大学 | Small sample radar one-dimensional image target recognition method based on deep migration learning |
| CN114444608B (en) * | 2022-02-08 | 2024-03-29 | 中国电信股份有限公司 | Data set quality evaluation method and device, electronic equipment and storage medium |
Citations (15)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109375186A (en) * | 2018-11-22 | 2019-02-22 | 中国人民解放军海军航空大学 | Radar target identification method based on the multiple dimensioned one-dimensional convolutional neural networks of depth residual error |
| CN109977871A (en) * | 2019-03-27 | 2019-07-05 | 中国人民解放军战略支援部队航天工程大学 | A kind of Satellite Targets recognition methods based on wideband radar data and GRU neural network |
| CN110321999A (en) * | 2018-03-30 | 2019-10-11 | 北京深鉴智能科技有限公司 | Neural computing figure optimization method |
| CN110391022A (en) * | 2019-07-25 | 2019-10-29 | 东北大学 | A kind of deep learning breast cancer pathological image subdivision diagnostic method based on multistage migration |
| CN111461063A (en) * | 2020-04-24 | 2020-07-28 | 武汉大学 | Behavior identification method based on graph convolution and capsule neural network |
| CN111596292A (en) * | 2020-04-02 | 2020-08-28 | 杭州电子科技大学 | Radar target identification method based on importance network and bidirectional stacking recurrent neural network |
| CN111736125A (en) * | 2020-04-02 | 2020-10-02 | 杭州电子科技大学 | Radar target identification method based on attention mechanism and bidirectional stacked cyclic neural network |
| CN111798980A (en) * | 2020-07-10 | 2020-10-20 | 哈尔滨工业大学（深圳） | Complex medical biological signal processing method and device based on deep learning network |
| CN111986112A (en) * | 2020-08-19 | 2020-11-24 | 北京航空航天大学 | Deep full-convolution neural network image denoising method of soft attention mechanism |
| CN112307996A (en) * | 2020-11-05 | 2021-02-02 | 杭州电子科技大学 | Fingertip electrocardiogram identity recognition device and method |
| CN112417760A (en) * | 2020-11-20 | 2021-02-26 | 哈尔滨工程大学 | Warship control method based on competitive hybrid network |
| CN112541355A (en) * | 2020-12-11 | 2021-03-23 | 华南理工大学 | Few-sample named entity identification method and system with entity boundary class decoupling |
| CN112764024A (en) * | 2020-12-29 | 2021-05-07 | 杭州电子科技大学 | Radar target identification method based on convolutional neural network and Bert |
| CN112784930A (en) * | 2021-03-17 | 2021-05-11 | 西安电子科技大学 | CACGAN-based HRRP identification database sample expansion method |
| EP3819659A1 (en) * | 2019-11-08 | 2021-05-12 | MBDA Deutschland GmbH | Communication module for components of tactical air conditioning systems |
-
2021
- 2021-05-17 CN CN202110535025.1A patent/CN113486917B/en active Active
Patent Citations (15)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110321999A (en) * | 2018-03-30 | 2019-10-11 | 北京深鉴智能科技有限公司 | Neural computing figure optimization method |
| CN109375186A (en) * | 2018-11-22 | 2019-02-22 | 中国人民解放军海军航空大学 | Radar target identification method based on the multiple dimensioned one-dimensional convolutional neural networks of depth residual error |
| CN109977871A (en) * | 2019-03-27 | 2019-07-05 | 中国人民解放军战略支援部队航天工程大学 | A kind of Satellite Targets recognition methods based on wideband radar data and GRU neural network |
| CN110391022A (en) * | 2019-07-25 | 2019-10-29 | 东北大学 | A kind of deep learning breast cancer pathological image subdivision diagnostic method based on multistage migration |
| EP3819659A1 (en) * | 2019-11-08 | 2021-05-12 | MBDA Deutschland GmbH | Communication module for components of tactical air conditioning systems |
| CN111736125A (en) * | 2020-04-02 | 2020-10-02 | 杭州电子科技大学 | Radar target identification method based on attention mechanism and bidirectional stacked cyclic neural network |
| CN111596292A (en) * | 2020-04-02 | 2020-08-28 | 杭州电子科技大学 | Radar target identification method based on importance network and bidirectional stacking recurrent neural network |
| CN111461063A (en) * | 2020-04-24 | 2020-07-28 | 武汉大学 | Behavior identification method based on graph convolution and capsule neural network |
| CN111798980A (en) * | 2020-07-10 | 2020-10-20 | 哈尔滨工业大学（深圳） | Complex medical biological signal processing method and device based on deep learning network |
| CN111986112A (en) * | 2020-08-19 | 2020-11-24 | 北京航空航天大学 | Deep full-convolution neural network image denoising method of soft attention mechanism |
| CN112307996A (en) * | 2020-11-05 | 2021-02-02 | 杭州电子科技大学 | Fingertip electrocardiogram identity recognition device and method |
| CN112417760A (en) * | 2020-11-20 | 2021-02-26 | 哈尔滨工程大学 | Warship control method based on competitive hybrid network |
| CN112541355A (en) * | 2020-12-11 | 2021-03-23 | 华南理工大学 | Few-sample named entity identification method and system with entity boundary class decoupling |
| CN112764024A (en) * | 2020-12-29 | 2021-05-07 | 杭州电子科技大学 | Radar target identification method based on convolutional neural network and Bert |
| CN112784930A (en) * | 2021-03-17 | 2021-05-11 | 西安电子科技大学 | CACGAN-based HRRP identification database sample expansion method |
Non-Patent Citations (4)
| Title |
|---|
| Fast and Flexible Multi-Task Classification Using Conditional Neural Adaptive Processes;James Requeima et al;《arXiv:1906.07697v2》;20200108;第1-24页 * |
| Radar HRRP Target Recognition Based on Concatenated Deep Neural Networks;K. Liao et al;《IEEE Access》;20180531;第29211-29218页 * |
| The performance of multi-layer neural network on face recognition system;M. J. Yashaswini et al;《2016 2nd International Conference on Contemporary Computing and Informatics (IC3I)》;20161231;第414-420页 * |
| 基于中心排序损失及弱监督定位的细粒度目标检索;郑侠武;《中国优秀硕士学位论文全文数据库信息科技辑》;20190715;第2019年卷(第7期);I138-993 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN113486917A (en) | 2021-10-08 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108133232B (en) | Radar high-resolution range profile target identification method based on statistical dictionary learning | |
| CN107728142B (en) | Radar high-resolution range profile target identification method based on two-dimensional convolutional network | |
| CN108008385B (en) | Interference environment ISAR high-resolution imaging method based on management loading | |
| CN113486917B (en) | Radar HRRP small sample target recognition method based on metric learning | |
| CN112764024B (en) | Radar target identification method based on convolutional neural network and Bert | |
| CN106443632B (en) | The radar target identification method of multitask Factor Analysis Model is kept based on label | |
| CN107766893B (en) | Target identification method based on label multilevel coding neural network | |
| CN112990334A (en) | Small sample SAR image target identification method based on improved prototype network | |
| CN114564982B (en) | Automatic identification method for radar signal modulation type | |
| CN110619352A (en) | Typical infrared target classification method based on deep convolutional neural network | |
| Wan et al. | Recognizing the HRRP by combining CNN and BiRNN with attention mechanism | |
| CN113625227B (en) | Attention transformation network-based radar high-resolution range profile target recognition method | |
| CN109932717B (en) | ISAR high-resolution imaging method based on environmental statistics modeling | |
| CN110988804A (en) | Radar radiation source individual identification system based on radar pulse sequence | |
| CN107133648B (en) | One-dimensional range profile identification method based on adaptive multi-scale fusion sparse preserving projection | |
| CN110161480A (en) | Radar target identification method based on semi-supervised depth probabilistic model | |
| CN106951822B (en) | One-dimensional range profile fusion identification method based on multi-scale sparse preserving projection | |
| CN113109780B (en) | High-resolution range profile target identification method based on complex number dense connection neural network | |
| CN111401168A (en) | Multi-layer radar feature extraction and selection method for unmanned aerial vehicle | |
| Li et al. | Supervised domain adaptation for few-shot radar-based human activity recognition | |
| CN112835008B (en) | High-resolution range profile target identification method based on attitude self-adaptive convolutional network | |
| CN111458688B (en) | Three-dimensional convolution network-based radar high-resolution range profile target recognition method | |
| CN113239959A (en) | Radar HRRP target identification method based on decoupling representation variational self-coding machine | |
| Chang et al. | Wireless physical-layer identification assisted 5G network security | |
| CN109766899B (en) | Physical feature extraction and SVM SAR image vehicle target recognition method |
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 |