CN109993050B

CN109993050B - Synthetic aperture radar image identification method

Info

Publication number: CN109993050B
Application number: CN201811430191.XA
Authority: CN
Inventors: 占荣辉; 田壮壮; 张军; 欧建平; 陈诗琪
Original assignee: National University of Defense Technology
Current assignee: National University of Defense Technology
Priority date: 2018-11-28
Filing date: 2018-11-28
Publication date: 2019-12-27
Anticipated expiration: 2038-11-28
Also published as: CN109993050A

Abstract

The invention discloses a synthetic aperture radar image identification method, and aims to solve the problem that the existing SAR image identification method is inaccurate in identification. The technical scheme is based on a convolutional neural network, and the weight of a convolutional kernel is introduced into a convolutional layer. Firstly, a neural network model is constructed according to an SAR image, and the SAR image is transmitted in a network forward direction to obtain a probability predicted value; matching and comparing the probability predicted value with the class label of the image to obtain a target loss function; parameters in the neural network model are then adjusted using a back propagation algorithm. And finally, identifying the SAR image to be identified through the trained network to obtain an identification result. By adopting the method, the dependence of the identification process on the expert experience can be avoided, and the subjectivity and the breaking property of the identification are avoided; in addition, the invention introduces the weight of the convolution kernel in the convolution layer of the convolution neural network, and can carry out self-adaptive adjustment on the weight of the convolution kernel, thereby further improving the accuracy of target classification and identification.

Description

Synthetic aperture radar image identification method

Technical Field

The invention belongs to the field of image processing, and relates to a Synthetic Aperture Radar (SAR) image recognition method, in particular to an SAR image recognition method based on a deep learning framework.

Background

The Synthetic Aperture Radar (SAR) has the characteristics of high resolution, long distance, all-time, all-weather and the like, compared with the traditional radar, the new system expands the dimension of radar imaging and can provide two-dimensional scattering information of a target, so that the SAR has wide application in a plurality of fields such as battlefield monitoring, weapon guidance, topographic mapping and the like. The conventional SAR image recognition is mainly implemented by a feature extraction and pattern classification method based on artificial experience, for example, document 1: qun Zhao, Jose c. principal, "Support vector machines for SAR automatic target registration", ieee transactions on Aerospace and Electronic Systems, 2001, 37 (2): 643-655(Qun Zhao et al, "Support Vector Machine for SAR automatic target recognition" published in 2001 of the institute of electrical and electronics engineers, aerospace and electronic systems, proceedings of 37 th), it is proposed to use Support Vector Machine (SVM) to perform recognition and classification of targets.

Document 2: jayaraman J.Thiagarajan, Karthikeyan N.Ramamurthy, et al "spark representation for automatic target classification in SAR images", 4th International Symposium on Communications, Control and Signal Processing (ISCSP), 2010: 1-4 ("sparse representation for SAR image automatic target recognition" published by Jayaraman J. Thiagarajan et al on the 4th Commission on International discussions of communication, control and Signal processing 2010) propose to use a normalized set of training vectors to construct a sparse representation dictionary and use the dictionary to find a local linear approximation of the underlying class manifold to compute a sparse representation of the test set.

Document 3: ganggang Dong, Wang Na, et al "spark representation of monogenic signal: with application to target recognition in SAR images ", IEEE Signalprocessing Letters, 2014, 21 (8): 952-956 (the sparse representation of the monogenic signal for target recognition of SAR image published by gangging Dong et al in 2014 at 21 st phase of signal processing promulgation of the institute of electrical and electronics engineers) proposes to extract the monogenic signal of the SAR image, generate an enhanced monogenic feature vector by uniformly down-sampling, normalizing and cascading the monogenic signal, and finally input the monogenic feature vector to a Sparse Representation Classifier (SRC) for training and testing.

Document 4: ganggang Dong, GangyaoKuang "Target recognition in SAR image vision classification on Riemannian semiconductors", IEEE Geoscience and Remote sensing letters, 2015, 12 (1): 199-. The method comprises the steps of firstly representing a target image through a monogenic signal, and constructing a covariance matrix by calculating the correlation among monogenic components. Due to the symmetry and the positive nature of the covariance matrix, a connected Riemannian manifold can be constructed and can be converted into a tangent vector space. Then, the obtained tangent vector is used as a feature vector, and classification is performed by using a Sparse Representation Classifier (SRC).

The above methods usually need to manually extract target features and select a classifier, the recognition effect of the methods depends greatly on manual experience, and feature extraction and classifier design are two relatively independent links. As a data-driven method, the convolutional neural network can automatically learn effective characteristic information from data through machine training and simultaneously complete the classification and identification of targets.

Document 5: the use of a full convolutional neural network to reduce training parameters to avoid overfitting in the case of undersized training samples was proposed in Sizhe Chen, Haipeng Wang, et al, "Target classification using the deep dependent networks for SAR images," IEEE Transactions on geoscience and remove Sensing,2016,54(8):4806- "deep convolutional network for SAR image object classification," published by Sizhe Chen et al in the 2016 institute of Electrical and electronics Engineers, Earth science and telemetry, 54 th.

Document 6: in Jun Ding, Bo Chen et al, "connected neural network with data amplification for SAR target recognition", IEEE Geoscience and Remote sensing letters,2016,13(3), 364-.

Although the convolutional neural network is applied to the SAR image recognition field, the influence of the relationship between different convolutional kernels on the generation of the characteristic diagram is not considered, so that the further improvement of the recognition effect is influenced. The synthetic aperture radar image identification method provided by the invention is based on the convolutional neural network, and the relation between convolutional kernels is constructed by introducing the convolutional kernel weight into the convolutional layer, so that the effective characteristics are enhanced, the non-effective characteristics are inhibited, and the classification performance of the convolutional neural network is improved. No published document relates to the introduction of convolution kernel weights in convolution layers of a convolutional neural network and the application thereof in SAR image recognition.

Disclosure of Invention

The invention aims to provide a synthetic aperture radar image identification method, and solves the problem of inaccurate identification caused by the fact that the relation between different convolution kernels is not considered in the existing SAR image identification method based on a convolution neural network.

The invention introduces the weight of convolution kernel in convolution layer based on convolution neural network. Firstly, a neural network model is constructed according to an SAR image, and the SAR image is transmitted in a network forward direction to obtain a probability predicted value; matching and comparing the probability predicted value with the class label of the image to obtain a target loss function; parameters in the neural network model are then adjusted using a back propagation algorithm. And finally, identifying the SAR image to be identified through the trained network to obtain an identification result.

The invention mainly comprises the following steps:

first, build for trainingThe SAR image database consists of N SAR images and category labels of the N SAR images; n is the number of SAR images in an SAR image database, and each SAR image only comprises one target; SAR image is denoted as G₁,…G_n,…G_N，G_nIs of size W_n×H_n，W_nIs G_nWidth of (H)_nIs G_nHigh of (d); class label denoted L₁,…L_n,…L_N，L_nIs a one-dimensional matrix containing C elements corresponding to C object classes, L_nThe middle element representing the real category is assigned with 1, and the rest elements are assigned with 0; c is the number of the image categories, C is a positive integer, and C is less than or equal to N; .

Second step, for G₁,…G_n,…G_NThe pretreatment is carried out, and the method comprises the following steps:

2.1 initializing variable n ═ 1;

2.2 if G_nAll the pixels in the pixel array are complex data, and 2.3 is converted; if G is_nAll the pixels in the image are real number data, and 2.4 turns;

2.3 pairs of G_nW of (2)_n×H_nEach pixel is subjected to modulus extraction and converted into a real number, and the method comprises the following steps:

2.3.1 let the row variable p be 1;

2.3.2 let column variable q be 1;

2.3.3 orderG_n(p, q) is G_nThe value of the upper (p, q) point, a is G_n(p, q) the real part of the complex data, b being G_n(p, q) the imaginary part of the complex data.

2.3.4q＝q+1；

2.3.5 determining whether q is equal to or less than W_nIf yes, turning to 2.3.3; if not, go to 2.3.6.

2.3.6p＝p+1；

2.3.7 determining whether p is equal to or less than H_nIf yes, turning to 2.3.2; if not, G will be described_nThe conversion is to a real image, 2.4.

2.4 Manual determination G₁,…,G_n,…,G_NPosition of the target and size of the target(Is G_nThe width of the medium target is greater than the target width,is G_nHigh for medium targets).

2.5 pairs of G₁,…G_n,…G_NCutting to make all SAR images have uniform size, and making uniform size be W^G×H^G。W^GAnd H^GAre all positive integers, and the method comprises the following steps: get1.1 to 2 times the maximum value of (A) as W^GGet it1.1 to 2 times the maximum value of (A) as H^GWith the object as the center, press W on the image^G×H^GAnd (5) cutting.

2.6 Generation of random sequences rn from 1 to N Using random sequence Generation function (e.g., randderm function in MATLAB)₁…,rn_n,…,rn_NUsing random sequence as index, for G₁,…G_n,…G_N，L₁,…L_n,…L_NReading is performed to make the random image readRandom class label for readingObtaining a random image sequence G'₁,…G′_n,…,G′_NAnd a random class tag L'₁,…,L′_n,…,L′_N。

2.7 to G'₁,…G′_n,…,G′_N，L′₁,…,L′_n,…,L′_NGrouping is carried out, the number of the images and the number of the labels in each group are in, in is more than or equal to 1 and less than or equal to N, and in belongs to [1,64 ]]，G′₁,…G′_n,…,G′_NIs totally BN group, L'₁,…,L′_n,…,L′_NAnd also into BN groups.WhereinIndicating rounding up.

Third step, according to W^G×H^GConstructing a neural network model, wherein the neural network model at least comprises a convolution layer, and a pooling layer can be constructed in order to reduce the dimension of the characteristic diagram in propagation; in order to improve the robustness of the network model, a discarding layer can be constructed; to get a better mapping of the final features, a fully connected layer can be constructed.

3.1 initialize the layer number variable, make the layer number variable cn of the convolutional layer equal to 1, the layer number variable pn of the pooling layer equal to 1, the layer number variable dn of the discard layer equal to 1, and the layer number variable fn of the all-connected layers equal to 1.

3.2 calculating the size of the convolution layer output characteristic diagram, wherein the method comprises the following steps:

3.2.1 if cn is 1, let CW^cn＝W^G,CH^cn＝H^G,CD^cn1, otherwise, 3.2.2, where CW^cnAn input feature map CG for the cn-th convolution layer^cnWidth of (C, CH)^cnIs CG^cnHigh, CD^cnIs CG^cnDepth of (CW)^cn×CH^cn×CD^cnRepresentation CG^cnIs measured in the dimension (d).

3.2.2 pairs of CG^cnBuilding convolutional layers with the convolutional kernel of the cn convolutional layer of sizeThe number of convolution kernels of the cn convolution layer is KN^cnStep size of convolution kernel at sliding isThe zero element filling size is x^cn. Order toRepresenting the kn of the cn-th convolutional layer^cn(1≤kn^cn≤KN^cn) A number of convolution kernels, each of which is a convolution kernel,representing the kn of the cn-th convolutional layer^cnThe bias voltage is set to be equal to the bias voltage,representing the kn of the cn-th convolutional layer^cnThe weights of the convolution kernels. Order toOutputting a feature map for a convolutional layerThe size of (d) is as follows:

in convolutional layers, the size of the convolutional kernelIs usually in the range ofAnd isAnd isKN when cn is 1^cnIs gotThe value range is usually KN^cn∈[10,20]When cn ≠ 1, KN^cnIs usually KN^cn∈[KN^cn-1,2×KN^cn-1]. Step size of convolution kernel during slidingUsually set to 1 or 2, zero element fill size

3.3 construct pooling layer, calculate the size of the pooling layer output characteristic map, the method is:

3.3.1 input profile of pn-th pooling layerOrder toWherein PW^cnIs PG^pnWidth of (D), PH^cnIs PG^pnHigh, PD of^cnIs PG^pnDepth of (1), then PW^pn×PH^pn×PD^pnRepresents PG^pnIs measured in the dimension (d).

3.3.2 for PG^pnConstructing the pooling layer such that the size of the sliding window in the pn-th pooling layer isThe step size of the sliding window during sliding isOrder toOutputting a feature map for the pooling layerThe size of (d) is as follows:

wherein,usually set to 2 or 3, step size

3.4, constructing a discarding layer, and calculating the size of a feature map of the discarding layer, wherein the method comprises the following steps:

3.4.1 input feature map for the dn-th discard layerOrder toIn which DW^dnRepresents DG^dnWidth of (D), DH^dnRepresents DG^dnHigh, DD^dnRepresents DG^dnDepth of (2), then DW^dn×DH^dn×DD^dnRepresents DG^dnIs measured in the dimension (d).

3.4.2 pairs DG^dnBuild a discard layer, orderIndicating the drop probability of the dn-th drop layer, typically orderOrder toOutputting a feature map for a discard layerThe size of (d) is as follows:

3.5 determinationthf is the discard level output profileIs a positive integer, is usually set to 3000, and if satisfied, let cn be cn +1, pn be pn +1, and dn be dn +1, and then letAnd orderTurning to step 3.2; if not, it indicates that the dimension of the output feature map is small, so step 3.6 is performed.

3.6 calculate the size of the fully-connected layer output feature vector.

3.6.1 let fn be 1, let the input feature vector FG for the fn-th fully-connected layer^fnDimension of (2)

3.6.2 FW for dimension size^fnFG of (1)^fnConstructing a full connection layer, and making a weight matrix A of the full connection layer^fnHas a dimension ofOffset fb^fnHas a dimension ofWhereinOutputting feature vectors for full connection layersIs usually set

3.7 calculate the feature vector size of the discarded layer.

3.7.1 let dn be dn + 1.

3.7.2 order DW^dn＝1,DH^dn＝1,In which DW^dn×DH^dn×DD^dn＝DD^dnInput feature vector DG representing the dn-th discarded layer^dnOf size of (1), wherein

3.7.3 for dimension size DW^dn×DH^dn×DD^dnDG of (1)^dnBuild a discard layer, orderIndicating the drop probability of the dn-th drop layer, typically orderOrder toOutputting a feature map for a discard layerThe size of (d) is as follows:

3.7.4 determinationthd is the discard layer output feature mapThe second threshold of (2) is a positive integer, usually set to 1000, and if satisfied, let fn be fn +1, dn be dn +1,and orderTurning to step 3.6.1; if not, go to step 3.8.

3.8 for dimension size ofIs/are as followsConstructing a full connection layer, and making a weight matrix A of the full connection layer^fnHas a dimension ofOffset fb^fnIs C.

3.9 let CN ═ PN, DN ═ DN, FN ═ FN, i.e. the neural network model has CN convolution layers, PN pooling layers, DN dropping layers and FN +1 full connection layers.

And fourthly, initializing model parameters of the neural network by adopting an Xavier method which is proposed in page 2.3 of page 2 and is published by Gloot Xavier et al in 13 th artificial intelligence and statistics international conference in 2010 and used for understanding difficulty in training the deep forward neural network. The specific operation mode is as follows:

4.1 convolution kernels in pairs of convolution layersInitialization is performed.

4.1.1 initializing the number of layers variable cn of the convolutional layer to 1.

4.1.2 initializing the number variable kn of the convolution kernel of the cn-th convolution layer^cn＝1。

4.1.3 Generation of random matrices using random functions (e.g., rand functions in MATLAB)Wherein the value range of each element is within (0,1),has a dimension of

4.1.4 pairsInitializing to enable:

wherein, the formula (5) representsIs initialized toThe element at the corresponding position is subtracted by 0.5 and multiplied byThe operation of the matrix is in this sense hereinafter.

4.1.5 order kn^cn＝kn^cn+1, decision kn^cn≤KN^cnIf yes, turn to 4.1.3, and if not, turn to 4.1.6.

4.1.6 making CN ═ CN +1, judging CN ≤ CN, if yes, turning to 4.1.2; if not, indicating the convolution kernelInitialization is complete, 4.2.

4.2 weights for convolution kernels in convolutional layerInitialization is performed.

4.2.1 initialize the number of layers of convolutional layers variable cn 1.

4.2.2 initializing kn^cn＝1。

4.2.3 Using randomizersNumbers (e.g., rand functions in MATLAB) generate random number matrices The value range of each element in the (1) is within (0).

4.2.4 pairsInitializing to enable:

4.2.5 order kn^cn＝kn^cn+1, decision kn^cn≤KN^cnIf yes, turning to 4.2.3; if not, 4.2.6 is performed.

4.2.6 making CN equal to CN +1, judging CN less than or equal to CN, if yes, turning to 4.2.2; if not, the initialization of the weight of the convolution kernel is completed, and 4.3 is executed.

4.3 weight matrix A in full connection layer¹,…,A^fn,…,A^FN+1Initialization is performed.

4.3.1 initialize fn to 1.

4.3.2 Generation of random number matrix RA Using random function (e.g., rand function in MATLAB)^fn，RA^fnThe value range of each element in the (1) is within (0).

4.3.3 pairs of A^fnInitializing to enable:

4.3.4, making FN equal to FN +1, judging FN is less than or equal to FN +1, if yes, turning to 4.3.2, if not, turning to 4.4.

4.4 bias in the pairs of convolution kernelsAnd offset fb in the full connection layer¹,…,fb^fn,…,fb^FN+1Initialization is performed to assign all biases to 0.

And fifthly, in the training stage, parameters in the neural network model need to be updated through continuous iteration of forward propagation and backward propagation. The number of initialization iterations en is 1.

And sixthly, initializing the group number bn to be 1.

In the seventh step, the initialization variable n ═ bn-1 × in + 1.

Eighth step, to G'_nAnd (4) carrying out forward propagation by adopting a forward propagation method, namely carrying out forward propagation on the nth SAR image in the constructed neural network model to obtain the probability prediction value of the SAR image category. And in the transmission process, the output of each layer is the characteristic extracted by the neural network.

8.1 initialize the layer number variable, make the layer number variable cn of the convolutional layer equal to 1, the layer number variable pn of the pooling layer equal to 1, the layer number variable dn of the discard layer equal to 1, and the layer number variable fn of the all-connected layers equal to 1.

8.2 calculating the output characteristic diagram of the convolution layer, the method is as follows:

8.2.1 if cn is 1, let us sayWhereinAnd inputting the feature map of the nth input image at the cn convolutional layer, and otherwise, executing 8.2.2.

8.2.2 pairsAnd (3) carrying out zero element filling, and specifically operating as follows:

8.2.2.1 initialize a size of (CW)^cn+2×χ^cn)×(CH^cn+2×χ^cn)×CD^cnIs/are as followsThe matrix is all zeros.

8.2.2.2 initializing cd^cn＝1，cd^cnIs the coordinate of the third dimension on the feature map.

8.2.2.3 initialize ch^cn＝1，ch^cnIs the second dimension coordinate on the feature map.

8.2.2.4 initializing cw^cn＝1，cw^cnIs the coordinate of the first dimension on the feature map.

8.2.2.5 orderWhereinTo representIn the coordinate (cw)^cn+χ^cn,ch^cn+χ^cn,cd^cn) The values above are the same as below.

8.2.2.6 order cw^cn＝cw^cn+1, decision cw^cn≤CW^cnIf yes, go to 8.2.2.5, if not, go to 8.2.2.7.

8.2.2.7 order ch^cn＝ch^cn+1, determining ch^cn≤CH^cnIf yes, go to 8.2.2.4, if not, go to 8.2.2.8.

8.2.2.8 order cd^cn＝cd^cn+1, decision cd^cn≤CD^cnIf yes, go to 8.2.2.3, if not, indicating that the assignment is complete, execute 8.2.3.

8.2.3 initializing kn^cn＝1。

8.2.4 calculating the kth^cnA convolution kernel andresult of convolution ofThe calculation method is as follows:

wherein,to representIn the coordinate (cw)^cn,ch^cn,kn^cn) The value of (a); (kw^cn,kh^cn) Is the position coordinate of the convolution kernel, (cw)^cn,ch^cn,cd^cn) For inputting feature mapsThe position coordinates of (a).

8.2.5 activation function σ (-) pairsThe value in (3) is subjected to nonlinear processing to obtain a convolution result after the nonlinear processingCommon activation functions include sigmoid function (sigmoid), hyperbolic tangent function (tanh), and modified Linear element (Rectified Linear Unit, ReLU) proposed in "ImageNet Classification with deep neural network" published by Alex Krizhevsky et al at the Ness International conference in 2012.

8.2.6 will utilize an S-shaped functionMapping to 0-1 to obtain the weight of the mapped convolution kernel

8.2.7 utilizeTo pairWeighting to obtain the output of the convolution layerThe specific operation is as follows:

8.2.7.1 initializing kn^cn＝1。

8.2.7.2 initialize ch^cn＝1。

8.2.7.3 initializing cw^cn＝1。

8.2.7.4 order

8.2.7.5 order cw^cn＝cw^cn+1, decision cw^cn≤CW^cn+2×χ^cnIf yes, go to 8.2.7.4, if not, go to 8.2.7.6.

8.2.7.6 order ch^cn＝ch^cn+1, determining ch^cn≤CH^cn+2×χ^cnIf yes, go to 8.2.7.3, if not, go to 8.2.7.7.

8.2.7.7 order kn^cn＝kn^cn+1, decision kn^cn≤KN^cnIf yes, go to 8.2.7.2, if not, indicating that the assignment is complete, execute 8.3.

8.3 calculating the output characteristic diagram of the pooling layer, wherein the method comprises the following steps:

8.3.1 input profile of the pn-th pooling layer

8.3.2 according to the size of the sliding WindowAnd step size of sliding window during slidingFor PG^pnPerforming a sliding window operation to obtainAnd (4) a region.

8.3.3 pairsIn equal areas respectively pooled, whereinThe coordinates of the position of the output feature map of the pooling layer are represented, and the position of the output feature map corresponding to the region of the input feature map is represented by the coordinates. The specific operation is as follows:

wherein,as a coordinate relative to the sliding window area, usually a function of taking the maximum valueOr taking the mean functionIf the maximum function is taken, the coordinates of the sliding window area at the position of the maximum are recordedRotating 8.4; if the average function is taken, the operation is directly changed to 8.4.

8.4 calculating the output characteristic graph of the discarding layer, wherein the method comprises the following steps:

8.4.1 input feature map for the dn-th discard layer

8.4.2 Generation of 1 through DD using random sequence Generation function (e.g., randderm function in MATLAB)^dnRandom sequence of (2)Random sequence is used as index, and dimension is DD^dnIs assigned to the one-dimensional matrix phi. Order toIs 0, orderIs 1. WhereinIndicating a rounding down.

8.4.3 coupling phi with DG^dnMultiplication. The specific operation is as follows:

8.4.3.1 initialize dd^dn＝1。

8.4.3.2 initialize dh^dn＝1。

8.4.3.3 initialize dw^dn＝1。

8.4.3.4 order

8.4.3.5 order dw^dn＝dw^dn+1, decision dw^dn≤DW^dnIf yes, go to 8.4.3.4, if not, go to 8.4.3.6.

8.4.3.6 order dh^dn＝dh^dn+1, decision dh^dn≤DH^dnIf yes, go to 8.4.3.3, if not, go to 8.4.3.7.

8.4.3.7 order dd^dn＝dd^dn+1, decision dd^dn≤DD^dnIf yes, go to 8.4.3.2, if not, indicating that the assignment is complete, execute 8.5.

8.5 let CN be CN +1, pn be pn +1, dn be dn +1, and dn be dn +1, determine CN ≦ CN, if satisfied, go to 8.2, if not, execute 8.6.

8.6 calculating the output characteristic vector of the full connection layer, the method is as follows:

8.6.1 feature map output by the dn-1 st layer drop layer if fn is 1Converting into one-dimensional feature vector as input feature vector FG of full connection layer^fnThe method comprises the following specific operation steps:

8.6.1.1 initialization

8.6.1.2 initialization

8.6.1.3 initialization

8.6.1.4 order:

8.6.1.5 orderDeterminationIf so, go to 8.6.1.4, if not, go to 8.6.1.6.

8.6.1.6 orderDeterminationIf so, go to 8.6.1.3, if not, go to 8.6.1.7.

8.6.1.7 orderDeterminationIf so, go to 8.6.1.2, if not, indicating that the assignment is complete, execute 8.6.2.

8.6.2 calculating the weight matrix A^fnAnd FG^fnAnd with an offset fb^fnAdding to obtain a feature vector FG between convolution layers^fn', the calculation is as follows:

FG^fn′＝A^fn×FG^fn+fb^fn (11)

8.6.3 Using the activation function σ (-) on FG^fnThe values in' are subjected to nonlinear processing to obtain nonlinear processed characteristic vectors

8.7 calculating the output feature vector of the discarding layer, the method is:

8.7.1 input feature map for the dn-th discard level

8.7.2 use random sequence generation function (such as randderm function in MATLAB) to generate 1 to DD^dnRandom sequence of (2)Random sequence is used as index, and dimension is DD^dnIs assigned to the one-dimensional matrix phi. Order toIs 0, orderIs 1. WhereinIndicating a rounding down.

8.7.3 will phi and DG^dnMultiplication. The specific operation is as follows:

8.7.3.1 initialize dd^dn＝1。

8.7.3.2 order

8.7.3.3 order dd^dn＝dd^dn+1, decision dd^dn≤DD^dnIf yes, go to 8.7.3.2, if not, indicating that the assignment is complete, execute 8.8.

8.8 let FN be FN +1 and dn be dn +1, determine FN is not more than FN, if satisfied, go to 8.6, if not, execute 8.9.

8.9 calculating the output characteristic vector of the full connection layer, the method is as follows:

8.9.1 calculating the weight matrix A^fnAnd FG^fnAnd with an offset fb^fnAdding to obtain feature vector (FG) in the middle of the full connection layer^fn) ', the calculation is as follows:

(FG^fn)′＝A^fn×FG^fn+fb^fn (12)

8.9.2 utilize a softmax function pair (FG)^fn) ' the values in (1) are non-linearized to obtain a fully connected layer resultTo be provided withFor example, the calculation is as follows:

the probability prediction value of the class c is the image.

The ninth step is based onAnd L'_nCalculating a loss function J of the nth SAR image_n：

L′_n(cc) represents L'_nThe cc element of (1); 9.1, let n be n +1, judge BN be BN, if satisfied, execute 9.2; if not, 9.3 is executed.

9.2 judging that N is less than or equal to N, if so, turning to the eighth step; if not, executing the tenth step.

9.3 judging that n is less than or equal to bn x in, if yes, turning to the eighth step, if not, executing the tenth step.

The tenth step is that the loss functions obtained by the SAR images in the bn group are averaged, and the loss function JJ of the bn group in the en iteration is calculated_{(en-1)×in+bn}(i.e., the average of the loss functions of the intra-group SAR images), BN is determined, and if satisfied, 10.1 is passed, and if not, 10.2 is passed.

10.1

10.2

The tenth step, determining the loss function JJ of the bn th group in the en-th iteration_{(en-1)×in+bn}JJM where JJM is a set loss function threshold (typically less than 0.01), if met, indicating that forward propagation is complete, go to the seventeenth step, if not, go to the twelfth step.

A twelfth step of mixing J_nAnd carrying out backward propagation in the network so as to adjust the neural network model parameters of the convolution kernel and the weight and the offset of the convolution kernel in the convolution layer and the weight matrix and the offset in the full-connection layer. The method comprises the following specific steps:

12.1 initialize the layer number variable, make the layer number variable CN of the convolution layer equal to CN, the layer number variable PN of the pooling layer equal to PN, the layer number variable DN of the discarding layer equal to DN, and the layer number variable FN of the all-connected layer equal to FN + 1.

12.2 calculation of J_nFor weight matrix A in the full connection layer¹,…,A^fn,…,A^FN+1And an offset fb¹,…,fb^fn,…,fb^FN+1Partial derivatives of (a).

12.2.1 if FN ═ FN +1, thenIf not, then,wherein softmax 'and S' represent derivatives of the softmax function and sigmoid function.

12.2.2 calculating J_nFor weight matrix A^fnAnd J_nTo offset fb^fnPartial derivatives of (a).

12.2.3, making fn equal to fn-1, judging that fn is equal to or more than 1, if so, converting to 12.2; if not, go to 12.3.

12.3 if PN ═ PN, converting J into_nFor FG^fn+1Partial derivatives ofIs converted intoFor three-dimensional matrix, the specific operation steps are as follows, otherwise 12.4 is turned.

12.3.1 initialization

12.3.2 initialization

12.3.3 initialization

12.3.4 order:

12.3.5determinationIf so, go to 12.3.4; if not, go to 12.3.6.

12.3.6DeterminationIf so, go to 12.3.3; if not, go to 12.3.7.

12.3.7DeterminationIf so, go to 12.3.2; if not, the assignment is complete, and 12.4 is turned.

12.4 use of J_nTo pairPartial derivatives ofCalculating a loss function J_nFor PG^pnInPartial derivatives of elements within the equal regions.

12.4.1 if in 8.3.2As a function of the maximum, go to 12.4.2, and if it is a function of the average, go to 12.4.3.

12.4.2 order J_nFor PG^pnThe middle position isThe partial derivatives of the elements of (a) are:

J_nfor PG^pnThe partial derivative of the remaining elements in (a) is 0.

12.4.3 order J_nFor PG^pnPartial derivatives of all elements in

12.5 according to 8.3.1,so can make

12.6 calculation of J_nFor convolution kernel in convolution layerWeights of convolution kernelsAnd bias in convolution kernelThe method comprises the following steps:

12.6.1 calculating J_nFor convolution kernelPartial derivatives of (a).

12.6.2 calculating J_nWeight to convolution kernelPartial derivatives of (a).

12.6.3 calculating J_nFor bias in convolution kernelPartial derivatives of (a).

12.6.4, making cn ═ cn-1, pn ═ pn-1, dn ═ dn-1, judging cn ≥ 1, if not, it is said that propagation is completed, and going to the thirteenth step; if so, go to 12.6.5.

12.6.5 calculating J_nTo pairPartial derivatives ofThe calculation method is as follows:

12.6.5.1 pairsAnd (3) carrying out zero element filling, and specifically operating as follows:

12.6.5.1.1 initialize a size ofAll-zero matrix JC of^cn+1。

12.6.5.1.2 initializationIs composed ofOf the third dimension.

12.6.5.1.3 initializationIs composed ofOf the second dimension.

12.6.5.1.4 initializationIs composed ofOf the first dimension of (a).

12.6.5.1.5 order

12.6.5.1.6DeterminationIf so, go to 12.6.5.1.5, if not, go to 12.6.5.1.7.

12.6.5.1.7DeterminationIf so, go to 12.6.5.1.4, if not, go to 12.6.5.1.8.

12.6.5.1.8DeterminationIf so, go to 12.6.5.1.3, if not, indicating that the assignment is complete, go to 12.6.5.2.

12.6.5.2 will beRotate clockwise by 180 degrees to obtainRotated matrix

12.6.5.3 calculates:

12.6.5.4 to 12.4.

Thirteenth, let n be n +1, judge BN be BN, if satisfy, go to 13.1; if not, go to 13.2.

13.1 judging that N is less than or equal to N, and if the N is less than or equal to N, turning to 12.1.2; if not, go to 13.3.

13.2 judging that n is less than or equal to bn x in, if so, turning to 12.1.2; if not, go to 13.3.

13.3 obtaining loss function for SAR images in the bn group vs. convolution kernel in convolution layerWeights of convolution kernels in convolutional layersBias in convolution kernelsWeight matrix A in full connection layer¹,…,A^fn,…,A^FN+1Bias in the full connection layerFb setting¹,…,fb^fn,…,fb^FN+1Respectively averaging partial derivatives of (A) by the following steps:

13.3.1, judging BN is BN, if so, turning to 13.3.2; if not, go to 13.3.3.

13.3.2

Turning to the fourteenth step;

13.3.3

a fourteenth step, using the output of the thirteenth step, using "Adam: adam algorithm updating convolution kernel provided in random optimization methodWeights of convolution kernelsBias in convolution kernelsWeight matrix A in full connection layer¹,…,A^fn,…,A^FN+1Bias fb in the full connection layer¹,…,fb^fn,…,fb^FN+1The method specifically comprises the following steps:

14.1 order

Wherein beta is₁Is a given first hyperparameter, usually set to β₁＝0.9。

14.2 order

Wherein beta is₂Is a given second hyperparameter, usually set to β₂＝0.999。

14.3 order

14.4 updateA^fnAnd fb^fnOrder:

where η is the initial learning rate and is typically set to η ≦ 0.01.

The fifteenth step, making BN be BN +1, judging that BN is less than or equal to BN, and if so, turning to the seventh step; if not, go to the sixteenth step.

Sixthly, making EN equal to EN +1, judging that EN is less than or equal to EN, and if the EN is less than or equal to EN, turning to the sixth step; if not, the training is finished, and the seventeenth step is executed.

Seventeenth, using the trained network model to identify the SAR image TG to be identified.

17.1 forward propagating the image in the network.

17.1.1 initializing the layer number variable dn of the discarded layer to 1

17.1.2 orderAll of which are 0,.

17.2 adopting the forward propagation method of the eighth step to forward propagate TG in the network to obtain

17.3 searchThe position cm of the medium maximum, i.e.:

and 17.4, outputting the cm-th category in the C categories in the label as the identification result.

Description of the drawings: in actual use, in addition to the convolution layer, other layers may be constructed as desired. For pooling layers, if the feature map dimension during propagation is too large, pooling layers can be used to reduce the feature map dimension. For the discard layer, if the training sample is too small or an overfitting phenomenon occurs during the training process, the discard layer may be used. For a fully connected layer, in order to get a better mapping of features, a fully connected layer may be constructed. The invention provides the construction and use methods of the convolution layer, the pooling layer, the discarding layer and the full-connection layer, but a user can construct the corresponding layer according to the actual requirement, and only the convolution layer needs to be constructed.

The invention has the beneficial effects that:

(1) the training of the convolutional neural network model (namely the eighth step to the sixteenth step) is completely and automatically completed through machine learning, so that the dependence of the recognition process on expert experience is avoided, the subjectivity and the force breaking property of recognition are also avoided, and the accuracy of image recognition is improved.

(2) The method for constructing the neural network model for the SAR image provided in the third step can construct a more effective network model for SAR images with different sizes. Compared with the traditional method for subjectively constructing the neural network model, the method provides a guide for constructing the neural network model.

(3) And in the eighth step, a weight value of a convolution kernel is introduced into a convolution layer of the convolution neural network, and the aim of weighting the features extracted by the convolution kernel is fulfilled by weighting the convolution kernel. The weight of the features extracted by the convolution kernel in the output feature graph can be adjusted through the weight, the representation capability of the features is enhanced, the recognition result is influenced finally, and the accuracy of target classification recognition is further improved.

(4) In the twelfth step, a method for performing backward propagation on the convolution layer containing the weight of the convolution kernel is provided, and the weight of the convolution kernel can be adaptively adjusted. The self-adaptive adjustment avoids the subjectivity of manual setting, and a more appropriate weight can be obtained according to training.

(5) In the fourteenth step, by introducing the Adam algorithm, the network parameters can be adaptively adjusted according to the gradient obtained by training, the defect that the learning rate in the original algorithm is fixed is avoided, and the network model can complete training more quickly and stably.

Drawings

FIG. 1 is a general flow diagram of the present invention;

FIG. 2 is a flow chart of a third step of constructing a neural network model according to the present invention;

FIG. 3 is a flow chart of an eighth forward propagation method of the present invention;

fig. 4 is a flowchart of the twelfth step back propagation method of the present invention.

Detailed Description

FIG. 1 is a general flow diagram of the present invention. The invention is further described in connection with experiments:

in the first step, a SAR image database for training is constructed. In the experiment, SAR image data for training in the MSTAR dataset, N is 2747, W_n×H_n＝128×128。

Secondly, preprocessing SAR images to obtain 2747W images^G×H^G96 × 96 SAR images. Here, in is set to 30, BN is set to 92, that is, the number of images and tags in each group is 30, and the images are grouped into 92 groups in total.

And thirdly, constructing a neural network model according to the preprocessed SAR image. The construction flow is shown in FIG. 3.

3.2 calculate the size of the convolutional layer output feature map.

3.3, building a pooling layer, and calculating the size of an output characteristic diagram of the pooling layer.

3.4 constructing a discarding layer, and calculating the size of the feature map of the discarding layer.

3.5 determinationthf is the discard level output profileIf the first threshold value of (1) is satisfied, let cn be cn +1, pn be pn +1, dn be dn +1, and then let And orderTurning to step 3.2; if not, go to step 3.6.

3.6 calculate the size of the fully-connected layer output feature vector.

3.7 calculate the feature vector size of the discarded layer.

3.7.1 let dn be dn + 1.

3.7.4 determinationthd is the discard layer output feature mapIf the second threshold value of (2) is satisfied, let fn be fn +1, dn be dn +1,and orderTurning to step 3.6; if not, go to step 3.8.

3.8 for dimension size ofIs/are as followsConstructing a full connection layer, and making a weight matrix A of the full connection layer^fnHas a dimension ofOffset fb^fnDimension of (2)Is C.

In the experiment, thf is first set to 3000, and then according to 3.2, the parameters of the first convolutional layer are set to:KN¹＝16，χ¹the output feature map size is 96 × 96 × 16, which is 3; according to 3.3, the parameters of the first pooling layer are set as follows:the size of the output characteristic graph is 48 multiplied by 16; according to 3.4, the parameters of the first discarded layer are set as:the size of the output characteristic graph is 48 multiplied by 16; a decision of 48 × 48 × 16 > thf is made from 3.5, so return to 3.2 and continue building convolutional layers, pooling layers, and discard layers according to the above steps as follows:

the parameters of the second convolutional layer are set as follows:KN²＝32，χ²the output feature map size is 48 × 48 × 32, which is 3; the parameters of the second pooling layer were:the output feature map size is 24 × 24 × 32; the parameters of the second discarded layer are:the output feature map has a size of 24 × 24 × 32, because24X 32 > thf, so 3.2 was returned.

The parameters for the third convolutional layer were set as:KN³＝64，χ³2, the output feature map size is 24 × 24 × 64; the parameters of the third pooling layer are:the output feature map size is 12 × 12 × 64; the parameters of the third discard layer are:the output feature map size is 12 × 12 × 64, since 12 × 12 × 64 > thf, 3.2 is returned.

The parameters for the fourth convolutional layer are set as follows:KN⁴＝128，χ⁴2, the output feature map size is 12 × 12 × 128; the parameters of the fourth pooling layer are:the output feature map size is 6 × 6 × 128; the parameters of the fourth discard layer are:the output feature map size is 6 × 6 × 128, since 6 × 6 × 128 > thf, 3.2 is returned.

The parameters for the fifth convolutional layer were set as:KN⁵＝128，χ⁵1, the output feature map size is 6 × 6 × 128; the parameters of the fifth pooling layer are:the size of the output characteristic map is 3 multiplied by 128; the parameters of the fifth discard layer are:the output feature map size is 3 × 3 × 128, since 3 × 3 × 128 < thf, proceed to the next step.

Set thd to 1000, then according to 3.6, set the parameters of the first fully-connected layer as:the output feature vector size is 512; according to 3.7, the parameters of the sixth discard layer are set as:the size of the output feature vector is 512, according to 3.7.4, the judgment result shows that 512 is less than thd, a full connection layer is constructed according to 3.8, image data in an experiment has 10 types of targets, and the setting parameters are as follows: c ═ 10. In the finally constructed network, CN ═ 5, PN ═ 5, DN ═ 6, and FN ═ 1.

And fourthly, initializing model parameters of the neural network by adopting an Xavier method. .

Fifthly, setting EN to 300, and initializing the iteration number EN to 1;

sixthly, initializing a group number bn to be 1;

seventhly, initializing a variable n which is (bn-1) x in + 1;

eighth step, as shown in FIG. 3, for G'_nAnd carrying out forward propagation by adopting a forward propagation method to obtain a probability prediction value of the SAR image category.

8.2 calculate the output characteristic map of the convolutional layer.

8.3 calculate the output profile of the pooling layer.

8.4 calculate the output profile of the discarded layer.

8.6 calculate the output feature vector of the fully-connected layer.

8.7 calculate the output feature vector of the drop layer.

8.9 calculate the output feature vector of the fully-connected layer.

In the experiment, the output characteristic graphs of each convolution layer, each pooling layer and each discarded layer are calculated according to the calculation methods of 8.2, 8.3 and 8.4 and the judgment method of 8.5; the output feature vectors of the first fully-connected layer and the sixth pooled layer are calculated according to the calculation methods of 8.6, 8.7 and the decision method of 8.8. Then according to the calculation method of 8.9, obtaining

The ninth step is based onAnd L'_nCalculating a loss function J of the nth SAR image according to the formula (14)_n：

9.1, let n be n +1, judge BN be BN, if satisfied, execute 9.2; if not, 9.3 is executed.

The tenth step is that the loss functions obtained by the SAR images in the bn group are averaged, and the loss function JJ of the bn group in the en iteration is calculated_{(en-1)×in+bn}；

A tenth step of determining the bn-th group in the en-th iterationLoss function JJ_{(en-1)×in+bn}JJM where JJM is the set loss function threshold, here set to 0.01, if satisfied, indicating that forward propagation is complete, go to the seventeenth step, if not, go to the twelfth step;

the twelfth step, as shown in FIG. 4, is to mix J_nAnd carrying out backward propagation in the network, and adjusting the neural network model parameters of the convolution kernel and the weight and the bias of the convolution kernel in the convolution layer, and the weight matrix and the bias in the full-connection layer.

12.2.2 calculation of J according to equation (17)_nFor weight matrix A^fnAnd J_nTo offset fb^fnPartial derivatives of (a).

12.3 if PN ═ PN, converting J into_nFor FG^fn+1Partial derivatives ofIs converted intoAnd (4) turning the three-dimensional matrix back to 12.4, otherwise, directly turning to 12.4.

12.5 order

12.6.1 calculating J according to equation (20)_nFor convolution kernelPartial derivatives of (a).

12.6.2 calculating J according to equation (21)_nWeight to convolution kernelPartial derivatives of (a).

12.6.3 calculating J according to equation (22)_nFor bias in convolution kernelPartial derivatives of (a).

12.6.5 calculating J_nTo pairPartial derivatives ofAnd 12.4. turning.

In the experiment, calculating partial derivatives of the loss function to the weight matrix and the bias in the full-connection layer according to 12.2; according to 12.2.3, fn is judged to be more than or equal to 1, if the condition is satisfied, the rotation is 12.2, and if the condition is not satisfied, the rotation is 12.3. According to 12.3, judging PN as PN, if satisfying, converting the partial derivative of the loss function to the characteristic diagram into three-dimension, if not, directly converting into 12.4. The partial derivatives of the loss function for the elements in each pooled region in the feature map are calculated according to 12.4. The partial derivatives of the loss function to the convolution kernel in the convolutional layer, the weights of the convolution kernel and the bias in the convolution kernel are calculated according to 12.6.

Thirteenth, let n be n +1, judge BN be BN, if satisfy, go to 13.1; if not, 13.2 is switched;

13.3 obtaining loss function for SAR images in the bn group vs. convolution kernel in convolution layerWeights of convolution kernels in convolutional layersBias in convolution kernelsWeight matrix A in full connection layer¹,…,A^fn,…,A^FN+1Bias fb in the full connection layer¹,…,fb^fn,…,fb^FN+1The partial derivatives of (a) are respectively averaged.

Fourteenth, updating the convolution kernel using Adam algorithmWeights of convolution kernelsBias in convolution kernelsWeight matrix A in full connection layer¹,…,A^fn,…,A^FN+1Bias fb in the full connection layer¹,…,fb^fn,…,fb^FN+1。

Fifteenth, making BN equal to BN +1, judging that BN is less than or equal to BN, and if so, turning to the seventh step; if not, turning to the sixteenth step;

sixthly, making EN equal to EN +1, judging that EN is less than or equal to EN, and if the EN is less than or equal to EN, turning to the sixth step; if not, indicating that the training is finished, and executing a seventeenth step;

the eighth step to the tenth step are to carry out forward propagation and backward propagation on the image in the model; the thirteenth step and the fourteenth step are to update the parameters of the model. And the eighth step to the fourteenth step are iterative processes, and whether the iteration is ended is judged according to the fifteenth step and the sixteenth step. And finally, when en is 301, the requirement that en is less than or equal to 300 is not met, so that the training is completed, and the model finally used for recognition is obtained.

Seventeenth, using the trained neural network model to identify the SAR image to be identified. In the experiment, 3203 SAR images used for testing in the MSTAR data set are adopted to test the neural network model. If the output category is the same as the actual category, the recognition is determined to be correct. The final correct recognition rate was 98.39%, demonstrating the effectiveness of the present invention.

Claims

1. A synthetic aperture radar image identification method is characterized by comprising the following steps:

first, build for trainingThe SAR image database consists of N SAR images and category labels of the N SAR images; n is the number of SAR images in an SAR image database, and each SAR image only comprises one target; SAR image is denoted as G₁,…G_n,…G_N，G_nIs of size W_n×H_n，W_nIs G_nWidth of (H)_nIs G_nHigh of (d); class label denoted L₁,…L_n,…L_N，L_nIs a one-dimensional matrix containing C elements corresponding to C object classes, L_nThe middle element representing the real category is assigned with 1, and the rest elements are assigned with 0; c is the number of the image categories, C is a positive integer, and C is less than or equal to N;

2.1 initializing variable n ═ 1;

2.3 mixing G_nW of (2)_n×H_nEach pixel is converted into a real number, and the real number is converted into 2.4;

2.4 determination of G₁,…,G_n,…,G_NPosition of the target and size of the target Is G_nThe width of the medium target is greater than the target width,is G_nHigh for medium targets;

2.5 pairs of G₁,…G_n,…G_NCutting to make all SAR images have uniform size, and making uniform size be W^G×H^G，W^GAnd H^GAre all positive integers;

2.6 Generation of random sequences rn from 1 to N Using random sequence Generation function₁…,rn_n,…,rn_NUsing random sequence as index, for G₁,…G_n,…G_N，L₁,…L_n,…L_NReading is performed to make the random image readRandom class label for readingObtaining a random image sequence G'₁,…G′_n,…,G′_NAnd a random class tag L'₁,…,L′_n,…,L′_N；

2.7 to G'₁,…G′_n,…,G′_N，L′₁,…,L′_n,…,L′_NGrouping is carried out, the number of images and the number of labels in each group are in, and in is more than or equal to 1 and less than or equal to N and G'₁,…G′_n,…,G′_NIs totally BN group, L'₁,…,L′_n,…,L′_NAnd is also divided into a group BN, wherein,whereinRepresents rounding up;

third step, according to W^G×H^GConstructing a neural network model, wherein the method comprises the following steps:

3.1 initializing the layer number variable cn, setting the layer number variable cn of the convolutional layer to 1, setting the layer number variable pn of the pooling layer to 1, setting the layer number variable dn of the discarded layer to 1, and setting the layer number variable fn of the all-connected layer to 1;

3.2.1 if cn is 1, let CW^cn＝W^G,CH^cn＝H^G,CD^cn1, otherwise, 3.2.2, where CW^cnAn input feature map CG for the cn-th convolution layer^cnWidth of (C, CH)^cnIs CG^cnHigh, CD^cnIs CG^cnDepth of (CW)^cn×CH^cn×CD^cnRepresentation CG^cnThe dimension size of (d);

3.2.2 pairs of CG^cnBuilding convolutional layers with the convolutional kernel of the cn convolutional layer of sizeThe number of convolution kernels of the cn convolution layer is KN^cnStep size of convolution kernel at sliding isThe zero element filling size is x^cn(ii) a Order toRepresenting the kn of the cn-th convolutional layer^cnA convolution kernel of 1 or more kn^cn≤KN^cn，Representing the kn of the cn-th convolutional layer^cnThe bias voltage is set to be equal to the bias voltage,representing the kn of the cn-th convolutional layer^cnThe weight of each convolution kernel; order toOutputting a feature map for a convolutional layerThe size of (d) is as follows:

3.3.1 input profile of pn-th pooling layerOrder toWherein PW^cnIs PG^pnWidth of (D), PH^cnIs PG^pnHigh, PD of^cnIs PG^pnDepth of (1), then PW^pn×PH^pn×PD^pnRepresents PG^pnThe dimension size of (d);

3.4.1 input feature map for the dn-th discard layerOrder toIn which DW^dnRepresents DG^dnWidth of (D), DH^dnRepresents DG^dnHigh, DD^dnRepresents DG^dnDepth of (2), then DW^dn×DH^dn×DD^dnRepresents DG^dnThe dimension size of (d);

3.4.2 pairs DG^dnBuild a discard layer, orderRepresents the drop probability of the dn-th drop layer, orderOrder toOutputting a feature map for a discard layerThe size of (d) is as follows:

3.5 determinationthf is the discard level output profileIf the first threshold value of (a) is a positive integer, let cn be cn +1, pn be pn +1, and dn be dn +1, and then letAnd orderTurning to step 3.2; if not, executing step 3.6;

3.6 calculate the size of the full-connection layer output feature vector:

3.6.2 FW for dimension size^fnFG of (1)^fnConstructing a full connection layer, and making a weight matrix A of the full connection layer^fnHas a dimension ofOffset fb^fnHas a dimension ofWhereinOutputting feature vectors for full connection layersSize, setting of

3.7 calculate the eigenvector size of the discarded layer:

3.7.1 let dn be dn + 1;

3.7.2 order DW^dn＝1,DH^dn＝1,Then DW^dn×DH^dn×DD^dnInput feature vector DG for the dnth discarded layer^dnOf size of (1), wherein

3.7.3 for dimension size DW^dn×DH^dn×DD^dnDG of (1)^dnConstruction ofDiscard the layer, orderRepresents the drop probability of the dn-th drop layer, orderOrder toOutputting a feature map for a discard layerThe size of (d) is as follows:

3.7.4 determinationthd is the discard layer output feature mapIf the second threshold value of (2) is a positive integer, let fn be fn +1, dn be dn +1,and orderTurning to step 3.6.1; if not, executing step 3.8;

3.8 for dimension size ofIs/are as followsConstructing a full connection layer, and making the weight moment of the full connection layerArray A^fnHas a dimension ofOffset fb^fnThe dimension of (a) is C;

3.9 make CN ═ PN, DN ═ DN, FN ═ FN, that is, the neural network model has CN convolution layers, PN pooling layers, DN discarding layers and FN +1 full connection layers;

fourthly, initializing model parameters of the neural network to obtain convolution kernels in the initialized convolution layerWeights of convolution kernels in initialized convolution layerWeight matrix A in initialized full connection layer¹,…,A^fn,…,A^FN+1Bias in initialized convolution kernelAnd offset fb in the full connection layer¹,…,fb^fn,…,fb^FN+1；

Fifthly, initializing the iteration number en as 1;

sixthly, initializing a group number bn to be 1;

seventhly, initializing a variable n which is (bn-1) x in + 1;

eighth step, to G'_nAdopting a forward propagation method to carry out forward propagation, namely carrying out forward propagation on the nth SAR image in the constructed neural network model to obtain a probability prediction value of the SAR image category, wherein the method comprises the following steps:

8.1 initializing the layer number variable, setting the layer number variable cn of the convolutional layer to 1, the layer number variable pn of the pooling layer to 1, the layer number variable dn of the discard layer to 1, and the layer number variable fn of the all-connected layer to 1;

8.2.1 if cn is 1, let us sayWhereinInputting the feature map of the nth input image at the cn convolution layer, otherwise, executing 8.2.2;

8.2.2 pairsCarrying out zero element filling to obtain a filled characteristic diagram

8.2.3 initializing kn^cn＝1；

wherein,to representIn the coordinate (cw)^cn,ch^cn,kn^cn) The value of (a); (kw^cn,kh^cn) Is the position coordinate of the convolution kernel, (cw)^cn,ch^cn,cd^cn) For inputting feature mapsThe position coordinates of (a);

8.2.5 activation function σ (-) pairsThe value in (3) is subjected to nonlinear processing to obtain a convolution result after the nonlinear processing

8.2.7 utilizeTo pairWeighting to obtain the output of the convolution layer

8.3.1 input profile of the pn-th pooling layer

8.3.2 according to the size of the sliding WindowAnd step size of sliding window during slidingFor PG^pnPerforming a sliding window operation to obtainAn area;

8.3.3 pairsAre respectively pooled in whichThe position coordinates of the output characteristic diagram of the pooling layer are represented, and meanwhile, the coordinates are used for representing that the position of the output characteristic diagram corresponds to the area of the input characteristic diagram, and the specific operation is as follows:

wherein,as a coordinate relative to the sliding window area, as a function of taking the maximumOr taking the mean functionIf the maximum function is taken, the coordinates of the sliding window area at the position of the maximum are recordedRotating 8.4; if get the flatThe mean function is directly converted to 8.4;

8.4.1 input feature map for the dn-th discard layer

8.4.2 Generation of 1 to DD Using random sequence Generation function^dnRandom sequence of (2)Random sequence is used as index, and dimension is DD^dnAssigning the one-dimensional matrix phi; order toIs 0, orderIs 1; whereinRepresents rounding down;

8.4.3 coupling phi with DG^dnMultiplying;

8.5 making CN ═ CN +1, pn ═ pn +1, dn ═ dn +1, judging CN ≦ CN, if yes, turning to 8.2, if no, executing 8.6;

8.6.1 feature map output by the dn-1 st layer drop layer if fn is 1Converting into one-dimensional feature vector as input feature vector FG of full connection layer^fn；

8.6.2 calculating the weight matrix A^fnAnd FG^fnAnd with an offset fb^fnAdding to obtain a feature vector FG between convolution layers^fn′The calculation method is as follows:

FG^fn′＝A^fn×FG^fn+fb^fn (11)；

8.6.3 Using the activation function σ (-) on FG^fn′The value in (3) is subjected to nonlinear processing to obtain a characteristic vector after the nonlinear processing

8.7.1 input feature map for the dn-th discard level

8.7.2 use random sequence generation function to generate 1 to DD^dnRandom sequence of (2)Random sequence is used as index, and dimension is DD^dnAssigning the one-dimensional matrix phi; order toIs 0, orderIs 1;

8.7.3 will phi and DG^dnMultiplication, specifically:

8.7.3.1 initialize dd^dn＝1；

8.7.3.2 order

8.7.3.3 order dd^dn＝dd^dn+1, decision dd^dn≤DD^dnIf yes, 8.7.3.2 is turned to, if no, the assignment is finished, and 8.8 is executed;

8.8 making FN ═ FN +1 and dn ═ dn +1, judging FN is less than or equal to FN, if it is satisfied, turning to 8.6, if it is not satisfied, executing 8.9;

(FG^fn)′＝A^fn×FG^fn+fb^fn (12)

8.9.2 utilize a softmax function pair (FG)^fn) ' the values in (1) are non-linearized to obtain a fully connected layer result The calculation method of (c) is as follows:

the probability prediction value of the image as the class c is obtained;

L′_n(cc) represents L'_nThe cc element of (1);

9.1, let n be n +1, judge BN be BN, if satisfied, execute 9.2; if not, executing 9.3;

9.2 judging that N is less than or equal to N, if so, turning to the eighth step; if not, executing the tenth step;

9.3 judging that n is less than or equal to bn x in, if so, turning to the eighth step, and if not, executing the tenth step;

the tenth step is that the loss functions obtained by the SAR images in the bn group are averaged, and the loss function JJ of the bn group in the en iteration is calculated_{(en-1)×in+bn}If so, turning to 10.1, and if not, turning to 10.2;

10.1

10.2

the tenth step, determining the loss function JJ of the bn th group in the en-th iteration_{(en-1)×in+bn}JJM, wherein JJM is a loss function threshold, if yes, the forward propagation is completed, the seventeenth step is carried out, and if not, the twelfth step is carried out;

a twelfth step of mixing J_nBackward propagation is carried out in the network, neural network model parameters of convolution kernels and weights and offsets of the convolution kernels in the convolution layers and weight matrixes and offsets in the full-connection layers are adjusted, and the method specifically comprises the following steps:

12.1 initializing a layer number variable, wherein the layer number variable CN of the convolutional layer is CN, the layer number variable PN of the pooling layer is PN, the layer number variable DN of the discarded layer is DN, and the layer number variable FN of the fully-connected layer is FN + 1;

12.2 calculation of J_nFor weight matrix A in the full connection layer¹,…,A^fn,…,A^FN+1And an offset fb¹,…,fb^fn,…,fb^FN+1Partial derivatives of (d);

12.2.1 if FN ═ FN +1, thenIf not, then,wherein T represents a matrix A^(fn+1)Transpose of (1), softmax 'and S' represent derivatives of the softmax function and sigmoid function;

12.2.2 calculating J_nFor weight matrix A^fnAnd J_nTo offset fb^fnPartial derivatives of (a):

12.2.3, making fn equal to fn-1, judging that fn is equal to or more than 1, if so, converting to 12.2; if not, turning to 12.3;

12.3 if PN ═ PN, converting J into_nFor FG^fn+1Partial derivatives ofConversion into three dimensions12.4, turning; otherwise, directly rotating to 12.4;

12.4 use of J_nTo pairPartial derivatives ofCalculating a loss function J_nFor PG^pnInPartial derivatives of the elements;

12.4.1 if in 8.3.2The maximum function is taken, and the operation is switched to 12.4.2, and if the maximum function is taken as the average function, the operation is switched to 12.4.3;

J_nfor PG^pnThe partial derivatives of the other elements are 0;

12.4.3 order J_nFor PG^pnPartial derivatives of all elements in

12.5 order

12.6.1 calculating J_nFor convolution kernelWhere σ' (. cndot.) is the derivative of the activation function σ ():

12.6.2 calculating J_nWeight to convolution kernelPartial derivatives of (a):

12.6.3 calculating J_nFor bias in convolution kernelPartial derivatives of (a):

12.6.4, making cn ═ cn-1, pn ═ pn-1, dn ═ dn-1, judging cn ≥ 1, if not, it is said that propagation is completed, and going to the thirteenth step; if so, go to 12.6.5;

12.6.5.1 pairsCarrying out zero element filling;

12.6.5.2 will beRotate clockwise by 180 degrees to obtainRotated matrix

12.6.5.3 calculates:

12.6.5.4 to 12.4;

13.1 judging that N is less than or equal to N, and if the N is less than or equal to N, turning to 12.1.2; if not, turning to 13.3;

13.2 judging that n is less than or equal to bn x in, if so, turning to 12.1.2; if not, turning to 13.3;

13.3 obtaining loss function for SAR images in the bn group vs. convolution kernel in convolution layerWeights of convolution kernels in convolutional layersBias in convolution kernelsWeight matrix A in full connection layer¹,…,A^fn,…,A^FN+1Bias fb in the full connection layer¹,…,fb^fn,…,fb^FN+1Respectively averaging partial derivatives of (A) by the following steps:

13.3.1, judging BN is BN, if so, turning to 13.3.2; if not, go to 13.3.3;

13.3.2

turning to the fourteenth step;

13.3.3

fourteenth, updating the convolution kernel using the output of the thirteenth stepWeights of convolution kernelsBias in convolution kernelsWeight matrix A in full connection layer¹,…,A^fn,…,A^FN+1Bias fb in the full connection layer¹,…,fb^fn,…,fb^FN+1The method specifically comprises the following steps:

14.1 order

Wherein beta is₁Is a given first hyper-parameter;

14.2 order

Wherein beta is₂Is a given second hyperparameter;

14.3 order

14.4 updateA^fnAnd fb^fnOrder:

wherein η is the initial learning rate;

the fifteenth step, making BN be BN +1, judging that BN is less than or equal to BN, and if so, turning to the seventh step; if not, turning to the sixteenth step;

seventeenth step, using the trained neural network model to identify the SAR image TG to be identified, wherein the method comprises the following steps:

17.1 forward-propagating the image in the network:

17.1.1 initializing the layer number variable dn of the discarded layer to 1;

17.1.2 orderAre all 0;

17.3 searchThe position cm of the medium maximum, i.e.:

2. The method of claim 1, wherein step 2.3 is performed by using G_nW of (2)_n×H_nThe method for respectively converting each pixel into a real number comprises the following steps:

2.3.1 let the row variable p be 1;

2.3.2 let column variable q be 1;

2.3.3 orderG_n(p, q) is G_nThe value of the upper (p, q) point, a is G_n(p, q) the real part of the complex data, b being G_n(p, q) an imaginary part of the complex data;

2.3.4q＝q+1；

2.3.5 determining whether q is equal to or less than W_nIf yes, turning to 2.3.3; if not, go to 2.3.6;

2.3.6p＝p+1；

2.3.7 determining whether p is equal to or less than H_nIf yes, turning to 2.3.2; if not, G will be described_nThe conversion is finished for a real image.

3. The method of claim 1, wherein 2.5 of the pairs G₁,…G_n,…G_NThe method for cutting comprises the following steps: get1.1 to 2 times the maximum value of (A) as W^GGet it1.1 to 2 times the maximum value of (A) as H^GWith the object as the center, press W on the image^G×H^GAnd (5) cutting.

4. The method according to claim 1, wherein the random sequence generating function refers to randderm function in MATLAB; random functions refer to rand functions in MATLAB.

5. The synthetic aperture radar image recognition method of claim 1, wherein in the convolutional layer,has a value range ofAnd isAnd isKN when cn is 1^cnHas a value range of KN^cn∈[10,20]When cn ≠ 1, KN^cnHas a value range of KN^cn∈[KN^cn-1,2×KN^cn-1](ii) a Step size of convolution kernel during slidingSet to 1 or 2, zero element fill sizeIn the above-mentioned pooling layer the water is added,set to 2 or 3, step size

6. The synthetic aperture radar image recognition method of claim 1, wherein in e [1,64 ∈ is said](ii) a The discard layer outputs a feature mapIs set to 3000, the layer output profile is discardedSet to 1000; the loss function threshold JJM is less than 0.01; beta is the same as₁0.9; beta is the same as₂0.999; eta is less than or equal to 0.01.

7. The synthetic aperture radar image recognition method of claim 1, wherein the fourth step initializes model parameters of the neural network using an Xavier method by:

4.1 convolution kernels in pairs of convolution layersAnd (3) initializing:

4.1.1 initializing the number of layers of the convolutional layer variable cn to 1;

4.1.2 initializing the number variable kn of the convolution kernel of the cn-th convolution layer^cn＝1；

4.1.3 Generation of random matrices Using random functionsWherein the value range of each element is within (0,1),has a dimension of

4.1.4 pairsInitializing to enable:

wherein, the formula (5) representsIs initialized toThe element at the corresponding position is subtracted by 0.5 and multiplied byAll matrix operations are in this sense;

4.1.5 order kn^cn＝kn^cn+1, decision kn^cn≤KN^cnIf yes, turning to 4.1.3, and if not, turning to 4.1.6;

4.1.6 making CN ═ CN +1, judging CN ≤ CN, if yes, turning to 4.1.2; if not, indicating the convolution kernelAfter the initialization is finished, 4.2 is switched;

4.2 weights for convolution kernels in convolutional layerAnd (3) initializing:

4.2.1 initializing the number of layers of the convolutional layer variable cn to 1;

4.2.2 initializing kn^cn＝1；

4.2.3 Generation of random number matrix Using random functionThe value range of each element in the (1) is within (0);

4.2.4 pairsInitializing to enable:

4.2.5 order kn^cn＝kn^cn+1, decision kn^cn≤KN^cnIf yes, turning to 4.2.3; if not, 4.2.6 is executed;

4.2.6 making CN equal to CN +1, judging CN less than or equal to CN, if yes, turning to 4.2.2; if not, the initialization of the weight value of the convolution kernel is completed, and 4.3 is executed;

4.3 weight matrix A in full connection layer¹,…,A^fn,…,A^FN+1And (3) initializing:

4.3.1 initialize fn to 1;

4.3.2 Generation of random number matrix RA Using random function^fn，RA^fnThe value range of each element in the (1) is within (0);

4.3.3 pairs of A^fnInitializing to enable:

4.3.4, making FN be FN +1, judging FN to be not more than FN +1, if yes, turning to 4.3.2, if not, turning to 4.4;

4.4 bias in the pairs of convolution kernelsAnd offset fb in the full connection layer¹,…,fb^fn,…,fb^FN ⁺¹Initialization is performed to assign all biases to 0.

8. A synthetic aperture radar image recognition method according to claim 1, wherein the pair of steps 8.2.2The method for carrying out zero element filling comprises the following steps:

8.2.2.1 initialize a size of (CW)^cn+2×χ^cn)×(CH^cn+2×χ^cn)×CD^cnIs/are as followsThe matrix is all zero;

8.2.2.2 initializing cd^cn＝1，cd^cnCoordinates of a third dimension on the feature map;

8.2.2.3 initialize ch^cn＝1，ch^cnCoordinates for a second dimension on the feature map;

8.2.2.4 initializing cw^cn＝1，cw^cnCoordinates that are a first dimension on the feature map;

8.2.2.5 orderWhereinTo representIn the coordinate (cw)^cn+χ^cn,ch^cn+χ^cn,cd^cn) The value of (a);

8.2.2.6 order cw^cn＝cw^cn+1, decision cw^cn≤CW^cnIf yes, go to 8.2.2.5, if not, go to 8.2.2.7;

8.2.2.7 order ch^cn＝ch^cn+1, determining ch^cn≤CH^cnIf yes, go to 8.2.2.4, if not, go to 8.2.2.8;

8.2.2.8 order cd^cn＝cd^cn+1, decision cd^cn≤CD^cnIf yes, go to 8.2.2.3, otherwise, end.

9. The method of claim 1, wherein 8.2.7 said pairs are used for image recognitionWeighting to obtain the output of the convolution layerThe method comprises the following steps:

8.2.7.1 initializing kn^cn＝1；

8.2.7.2 initialize ch^cn＝1；

8.2.7.3Initializing cw^cn＝1；

8.2.7.4 order

8.2.7.5 order cw^cn＝cw^cn+1, decision cw^cn≤CW^cn+2×χ^cnIf yes, go to 8.2.7.4, if not, go to 8.2.7.6;

8.2.7.6 order ch^cn＝ch^cn+1, determining ch^cn≤CH^cn+2×χ^cnIf yes, go to 8.2.7.3, if not, go to 8.2.7.7;

8.2.7.7 order kn^cn＝kn^cn+1, decision kn^cn≤KN^cnIf yes, go to 8.2.7.2, otherwise, end.

10. The method of claim 1, wherein the step of 8.4.3 is performed by combining Φ and DG^dnThe multiplication method comprises the following steps:

8.4.3.1 initialize dd^dn＝1；

8.4.3.2 initialize dh^dn＝1；

8.4.3.3 initialize dw^dn＝1；

8.4.3.4 order

8.4.3.5 order dw^dn＝dw^dn+1, decision dw^dn≤DW^dnIf yes, go to 8.4.3.4, if not, go to 8.4.3.6;

8.4.3.6 order dh^dn＝dh^dn+1, decision dh^dn≤DH^dnIf yes, go to 8.4.3.3, if not, go to 8.4.3.7;

8.4.3.7 order dd^dn＝dd^dn+1, decision dd^dn≤DD^dnIf yes, go to 8.4.3.2, otherwise, end.

11.The method of claim 1, wherein the step 8.6.1 is implemented by outputting the feature map of the layer dn-1 discarding layerThe method of converting into a one-dimensional feature vector is:

8.6.1.1 initialization

8.6.1.2 initialization

8.6.1.3 initialization

8.6.1.4 order:

8.6.1.5 orderDeterminationIf yes, go to 8.6.1.4, if not, execute 8.6.1.6;

8.6.1.6 orderDeterminationIf yes, go to 8.6.1.3, if not, execute 8.6.1.7;

8.6.1.7 orderDeterminationIf so, the routine proceeds to 8.6.1.2, and if not, the routine ends.

12. The method of claim 1, wherein step 12.3 is performed by using J_nFor FG^fn+1Partial derivatives ofConversion into three dimensionsThe method comprises the following steps:

12.3.1 initialization

12.3.2 initialization

12.3.3 initialization

12.3.4 order:

12.3.5determinationIf so, go to 12.3.4; if not, go to 12.3.6;

12.3.6determinationIf so, go to 12.3.3; if not, go to 12.3.7;

12.3.7determinationIf so, go to 12.3.2; if not, ending.

13. The method of claim 1, wherein 12.6.5.1 said pairs are used for image recognitionThe method for carrying out zero element filling comprises the following steps:

12.6.5.1.1 initialize a size ofAll-zero matrix JC of^cn+1；

12.6.5.1.2 initialization Is composed ofCoordinates of a third dimension above;

12.6.5.1.3 initialization Is composed ofA second dimension of coordinates of;

12.6.5.1.4 initialization Is composed ofA coordinate of a first dimension above;

12.6.5.1.5 order

12.6.5.1.6DeterminationIf yes, go to 12.6.5.1.5, if not, go to 12.6.5.1.7;

12.6.5.1.7determinationIf yes, go to 12.6.5.1.4, if not, go to 12.6.5.1.8;

12.6.5.1.8determinationIf so, the routine proceeds to 12.6.5.1.3, and if not, the routine ends.