CN114626987B - Electromagnetic backscatter imaging method based on physical depth expansion network - Google Patents
Electromagnetic backscatter imaging method based on physical depth expansion network Download PDFInfo
- Publication number
- CN114626987B CN114626987B CN202210307192.5A CN202210307192A CN114626987B CN 114626987 B CN114626987 B CN 114626987B CN 202210307192 A CN202210307192 A CN 202210307192A CN 114626987 B CN114626987 B CN 114626987B
- Authority
- CN
- China
- Prior art keywords
- network
- image
- induced current
- dimension
- loss
- 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
- 238000003384 imaging method Methods 0.000 title claims abstract description 15
- 238000012549 training Methods 0.000 claims abstract description 12
- 238000005457 optimization Methods 0.000 claims abstract description 6
- 238000007781 pre-processing Methods 0.000 claims abstract description 4
- 238000000034 method Methods 0.000 claims description 42
- 239000011159 matrix material Substances 0.000 claims description 28
- 230000006870 function Effects 0.000 claims description 20
- 238000012545 processing Methods 0.000 claims description 8
- 125000003275 alpha amino acid group Chemical group 0.000 claims description 6
- 150000001875 compounds Chemical class 0.000 claims description 6
- 230000008602 contraction Effects 0.000 claims description 6
- 230000000694 effects Effects 0.000 claims description 4
- 230000006698 induction Effects 0.000 claims description 4
- 238000011176 pooling Methods 0.000 claims description 4
- 230000004913 activation Effects 0.000 claims description 3
- 230000008569 process Effects 0.000 claims description 3
- 230000001939 inductive effect Effects 0.000 claims 1
- 238000004422 calculation algorithm Methods 0.000 abstract description 6
- 238000005516 engineering process Methods 0.000 abstract description 5
- 238000010276 construction Methods 0.000 abstract 1
- 238000012360 testing method Methods 0.000 description 14
- 238000010586 diagram Methods 0.000 description 10
- 238000013528 artificial neural network Methods 0.000 description 4
- 230000009471 action Effects 0.000 description 3
- 238000013135 deep learning Methods 0.000 description 3
- 238000013507 mapping Methods 0.000 description 3
- 238000004364 calculation method Methods 0.000 description 2
- 238000012804 iterative process Methods 0.000 description 2
- 239000000203 mixture Substances 0.000 description 2
- 238000012795 verification Methods 0.000 description 2
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- FFBHFFJDDLITSX-UHFFFAOYSA-N benzyl N-[2-hydroxy-4-(3-oxomorpholin-4-yl)phenyl]carbamate Chemical compound OC1=C(NC(=O)OCC2=CC=CC=C2)C=CC(=C1)N1CCOCC1=O FFBHFFJDDLITSX-UHFFFAOYSA-N 0.000 description 1
- 238000013527 convolutional neural network Methods 0.000 description 1
- 238000000354 decomposition reaction Methods 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 238000005259 measurement Methods 0.000 description 1
- 230000000452 restraining effect Effects 0.000 description 1
- 238000004088 simulation Methods 0.000 description 1
- 238000013519 translation Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T3/00—Geometric image transformations in the plane of the image
- G06T3/40—Scaling of whole images or parts thereof, e.g. expanding or contracting
- G06T3/4038—Image mosaicing, e.g. composing plane images from plane sub-images
-
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T3/00—Geometric image transformations in the plane of the image
- G06T3/40—Scaling of whole images or parts thereof, e.g. expanding or contracting
- G06T3/4046—Scaling of whole images or parts thereof, e.g. expanding or contracting using neural networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T5/00—Image enhancement or restoration
- G06T5/90—Dynamic range modification of images or parts thereof
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20084—Artificial neural networks [ANN]
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Computation (AREA)
- Computational Linguistics (AREA)
- Molecular Biology (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Health & Medical Sciences (AREA)
- Data Mining & Analysis (AREA)
- General Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Image Processing (AREA)
- Monitoring And Testing Of Transmission In General (AREA)
Abstract
The invention discloses an electromagnetic backscatter imaging method based on a physical deep expansion network, which comprises the following steps: 1, constructing mixed input data, including data acquisition and preprocessing, and enriching network input information; 2, in the network structure construction stage, a depth expansion technology is combined with a traditional subspace optimization iterative algorithm to design a network structure of the depth expansion technology; 3, designing a loss function, and optimizing a network together by using the image structure similarity loss and pixel-by-pixel loss in the objective function, particularly using the induced current loss and the scattered field loss; 4, through training the depth expansion network, high-quality reconstruction of induced current, scattered field and contrast of the target scatterer can be rapidly performed. The physical-based deep expansion network provided by the invention can effectively replace the traditional SOM iterative algorithm, enrich the physical knowledge of the network and realize quick and high-precision electromagnetic backscatter imaging.
Description
Technical Field
The invention belongs to the technical field of electromagnetic backscatter imaging, and particularly relates to an electromagnetic backscatter imaging method which uses a deep learning method to effectively replace a traditional SOM iteration method.
Background
Electromagnetic backscatter is the determination of properties of an unknown scatterer within a spatial region, such as information on position, shape, and physical parameters, from the scattered field distribution of that spatial region. However, electromagnetic backscattering problems (ISPs) exhibit a high degree of nonlinearity, which arises from the multiple scattering effect between the measurement scattering field and the scatterer, and morbidity, which arises mainly from the large errors in solution caused by small perturbations in the observed data. ISP can be solved by traditional objective function methods, mainly comprising a back propagation method (BP) of linear approximation, a deformation-born iteration method (DBIM) of nonlinear iteration, contrast Source Inversion (CSI), subspace Optimization (SOM) and the like. The traditional method integrates fluctuation physics into a model, but the linear method has rough imaging, and the iterative method has large calculation cost and long time consumption.
In recent years, due to the strong learning mapping capability and high solving speed, the deep learning network is successfully applied to solving the electromagnetic backscatter problem by researchers. For example, the Direct Inversion Scheme (DIS) proposed by Wei et al is a typical algorithm that uses neural networks to achieve a mapping from a scattered field to a target scatterer, but it can reconstruct some simple scatterers only within the training set. Li et al propose a 'deep' algorithm based on a deconvoluted neural network by analogy to the linkage of the traditional nonlinear iterative method with CNN. Some studies convert the target domain from the contrast domain to the current domain. The 'ICLM' algorithm proposed by Wei et al uses a cascade network dedicated to learning the ambiguous portion of the induced current. Huang et al reduce the back-scattering problem to an image translation problem, first use a back-propagation method to obtain a coarse image, and then use a neural network to realize high-resolution reconstruction of the image. The test results in the above papers show that the current depth backscattering method is superior to the traditional nonlinear optimization method in both imaging quality and speed.
The above-described method is limited by the quality of the input and a priori of the type of scatterer boundaries, and in particular the generalization ability is limited when the network lacks physical knowledge guidance. The depth backscattering approach requires consideration of physical model consistency and data consistency. The gap between the traditional objective function method and the data-driven deep learning method is made up, and the realization of high-quality imaging by effectively embedding physical knowledge into a deep neural network is a key technical problem.
Disclosure of Invention
The invention provides an electromagnetic backscatter imaging method based on a physical depth expansion network to more effectively combine the depth expansion technology, an SOM iteration frame and the prior physical knowledge, so that the network learns the physical knowledge and enhances the model generalization capability, thereby realizing quick and high-precision electromagnetic backscatter imaging and further realizing high-quality reconstruction of induced current, scattered field and contrast of a scatterer.
The invention adopts the following technical scheme for solving the technical problems:
the invention discloses an electromagnetic backscatter imaging method based on a physical depth expansion network, which is characterized by comprising the following steps:
step one, constructing mixed input data, including data acquisition and preprocessing;
step 1.1, an electromagnetic scattering system adopts T transmitting antennas and R receiving antennas, and a target scatterer is placed in a square region of interest DThe transmitting antennas sequentially transmit plane wave signals to the region of interest D, and R receiving antennas simultaneously measure scattered fields;
step 1.2, calculating the target scatterer by adopting a moment method during forward modelingIs +.>And simulated fringe field->
Step 1.3, inverting the region of interestDomain D is discretized into a grid of size M x M, conformal M 2 A sub-grid;
step 1.4, decomposing the simulated scattered field by singular valuesProcessing to obtain a dimension [ M ] 2 ,T]Deterministic partial current +.>Wherein T represents T transmitting antenna channels;
step 1.5, varying deterministic portion currentAnd obtain the dimension of [ T, M ]]Three-dimensional current image matrix->
Step 1.6, matrix three-dimensional current imageAdding a dimension for storing the three-dimensional current image matrixTo obtain the real part and the imaginary part of the dimension [ T, N ] 1 ,M,M]Deterministic current matrix->Wherein N is 1 Representing the deterministic current +.>The real part and the imaginary part channel number;
step 1.7, using a back propagation method on the simulated fringe fieldProcessing to generate dimension as[M,M]Is a low resolution scatterer image χ BP ;
Step 1.8, χ on the low-resolution scatterer image BP Adding one dimension for storing the low resolution scatterer image χ BP To obtain the imaginary part of the dimension N 2 ,M,M]Three-dimensional image matrix of (a)
Step 1.9, matrix the three-dimensional imageAdding one dimension for storing T three-dimensional image matrixes +.>Obtaining the dimension of [ T, N ] 2 ,M,M]Is +.>Wherein N is 2 Representing the low resolution contrast imageIs the imaginary channel number of (a);
step 1.10, the deterministic current is processedAnd said low resolution contrast image +.>Splicing in the second dimension to obtain a piece of dimension [ T, N, M ]]Is a mixed input data x; wherein n=n 1 +N 2 Representing mixed input data x 1 The number of real and imaginary channels;
step two, building a deep expansion network P θ And mixes the input data x 1 As a deep-spread network P θ Is extended by the depth by the input of the network P θ Output stationThe target scattererIs approximately true of the complete induced current +.>
Step 2.1, consisting of K cascaded subnetworks { P ] θ,k |k∈[1,K]Form deep expansion network P θ The method comprises the steps of carrying out a first treatment on the surface of the Wherein P is θ,k Representing a kth cascaded subnetwork; and the kth cascaded subnetwork P θ,k Adopting a U-net structure, comprising a contracted path and an expanded path;
the contraction path is formed by sequentially adding a maximum pooling layer after two convolution blocks, wherein the convolution blocks consist of a convolution layer with a convolution kernel size of a multiplied by a, a BN layer and a ReLU activation function;
the expansion path is formed by sequentially adding two convolution blocks after deconvolution operation, the deconvolution operation is formed by deconvolution layers with the convolution kernel size of b multiplied by b, and the convolution blocks have the same structure as the contraction path;
step 2.2, when k=1, the mixed input data x 1 Inputting the depth expansion network P θ And pass through the kth sub-network P θ,k Is processed to obtain a feature map f with dimension c×c k Then the kth sub-network P is output after the processing of the expansion path θ,k Predicted induced current
According to the induced currentObtaining the target scatterer by using a contrast updating formula of a state equation, a data equation and SOM>The kth predictive total field of +.>Kth predicted Scattering field->And kth prediction contrast image +.>
When k=2, 3..k, K-1 th subnetwork P θ,k-1 Output induced current matrixAnd k-1 th fringe field image matrix +.>Splicing in the second dimension to obtain kth mixed input data x k And pass through the kth sub-network P θ,k Outputs a kth predicted induced current matrix +.>Thereby by the Kth sub-network P θ,K Outputting a K-th predicted induced current matrix +.>And as a deep-spread network P θ The output approximately true complete induction current +.>
Then according to the approximate real complete induced currentThe target scatterer is further obtained by using a state equation, a data equation and a contrast updating formula of SOM>Including predictive total field +.>Predicted fringe field->Prediction contrast image +.>
Step three, designing a loss function, and establishing a deep expansion network P θ Is an optimization objective of (1);
step 3.1, constructing a deep expansion network P by utilizing the method (1) θ Target loss function L of (2) P :
L P =L J +L E +λ 1 L SSIM +λ 2 L MSE (1)
In the formula (1), L J Represents an induced current loss and is obtained by the formula (2); l (L) E Represents the loss of the scattered field and is obtained by the formula (3); l (L) SSIM Representing a loss of contrast image quality and obtained by formula (3); l (L) MSE Representing pixel-by-pixel loss and obtained by equation (4); lambda (lambda) 1 ,λ 2 Is a super parameter to balance the effects of image quality loss and pixel-by-pixel loss;
in the formula (2), the amino acid sequence of the compound,indicating the target scatterer corresponding to the jth transmitting antenna>Is a near real induced current matrix of (a) Indicating the target scatterer corresponding to the jth transmitting antenna>Is a simulated induced current of (a);
in the formula (3), the amino acid sequence of the compound,representing the predicted fringe field of the depth-spread network corresponding to the first receive antenna, +.>Indicating the target scatterer corresponding to the first receiving antenna->Is a true scatter field of (2);
in the formula (4), SSIM represents an image structural similarity loss;
in the formula (5), N represents the number of pixels of one contrast image;representing predictive contrast image +.>A contrast value corresponding to the q-th pixel point; />Representing object scatterer->A contrast value corresponding to the q-th pixel point;
fourthly, performing diffuser induced current, a diffuser field and contrast reconstruction through training a depth unfolding network;
based on the mixed input data x 1 For the deep expansion network P θ Learning and calculating the loss function L P In the process of (2), the network parameter theta is continuously optimized, so that the induced current, the scattered field and the contrast image output by the network reconstruction are gradually fitted to the physical quantity corresponding to the real scattering body, thereby obtaining an optimal network model for realizing high-quality induced current, scattered field and contrast image reconstruction.
Compared with the prior art, the invention has the beneficial effects that:
1. the invention provides an electromagnetic backscatter imaging method based on a physical deep expansion network, which uses two characteristic information of induced current and contrast to carry out mixed input into the network, enriches the input characteristic information, reduces the nonlinear difficulty of network learning, and improves the network reconstruction quality and the network reconstruction efficiency.
2. The method utilizes the depth expansion technology to combine the traditional SOM iteration method and the prior physical knowledge to establish the physical network mapping, and uses the characteristic information such as induced current, scattered field, contrast and the like to carry out comprehensive constraint.
3. In the invention, besides the contrast loss function, the current loss function and the scattered field loss function are particularly introduced into the objective function, so that the consistency of a physical model and the consistency of data are ensured, and the generalization capability of a physical network is better enhanced.
Drawings
FIG. 1 is a flow chart of the method of the present invention;
FIG. 2 is a diagram of a deep spread sub-network structure according to the present invention;
FIG. 3 is a diagram showing the reconstruction result of MNIST handwriting digital data set according to the present invention;
FIG. 4 is a diagram showing the results of Test #1 induced current and fringe field reconstruction with 10% noise added in accordance with the present invention;
fig. 5 is a diagram showing the reconstruction result of an "Austria" dataset with different proportions of noise added according to the present invention;
fig. 6 is a diagram showing the induced current and fringe field reconstruction results of the "Austria" invention with 25% noise added;
FIG. 7 is a diagram showing the reconstruction result of experimental data with the frequency of 3 GHz;
fig. 8 is a diagram showing the results of induced current and scattered field reconstruction of experimental data with a frequency of 3GHz according to the present invention.
Detailed Description
In this embodiment, an electromagnetic backscatter imaging method based on a physical deep expansion network is to mix input data x 1 As a deep-spread network P θ To approximate the actual induced current matrix of the scattererAs a deep-spread network P θ An output of (2); the traditional SOM method is effectively replaced by combining the depth expansion technology with the traditional SOM iterative process, and one iterative process of the SOM is mapped into the depth expansion network P θ Each module updates the induced current and contrast simultaneously, and the network P is deeply deployed θ Besides realizing high-quality reconstruction of complete induced current, the scattered field and the contrast can be further obtained through calculation of a state equation, a data equation and a contrast updating formula of SOM. Specifically, as shown in fig. 1, the method comprises the following steps:
step one, constructing mixed input data, including data acquisition and preprocessing;
step 1.1, the electromagnetic scattering system adopts t=16 transmitting antennas, r=32 receiving antennas, and the electromagnetic scattering system is placed in a square region of interest DTarget-placed scattererThe transmitting antennas sequentially transmit plane wave signals to the region of interest D, and R=32 receiving antennas measure scattered fields simultaneously; assuming an operating frequency of 400MHZ, the region of interest D is 2.0 meters x 2.0 meters.
Step 1.2, during forward modeling, placing a target scatterer in the region of interest DTo avoid the occurrence of the "reverse CRIME" phenomenon in the inverse problem, for each incidence, the target scatterer is calculated by means of a moment method on a 100X 100 grid>Is +.>And simulated fringe field->
In the step 1.3, during inversion, the region of interest D is discretized into a grid with a size of mxm=64×64, and is formed into M in a conformal manner 2 =4096 sub-grids;
step 1.4, simulation of the scattered field Using singular value decompositionProcessing to obtain a dimension [ M ] 2 ,T]=[4096,16]Deterministic partial current +.>
Step 1.5, varying deterministic portion currentAnd obtain the dimension of [ T, M ]]=[16,64,64]Three-dimensional current image matrix->
Step 1.6, matrix three-dimensional current imageAdding one dimension for storing three-dimensional current image matrix +.>To the real and imaginary parts of [16,2,64,64 ] of the dimension]Deterministic current matrix->
Step 1.7, using a back propagation method to simulate a scattered fieldProcessing to generate a dimension of [ M, M ]]Is a low resolution scatterer image χ BP ;
Step 1.8, χ on low resolution scatterer image BP Adding one dimension for storing low resolution scatterer image χ BP To obtain the imaginary part of the dimension N 2 ,M,M]=[1,64,64]Three-dimensional image matrix of (a)
Step 1.9, matrix the three-dimensional imageAdding one dimension for storing t=16 three-dimensional image matrices +.>Obtaining the dimension of [ T, N ] 2 ,M,M]=[16,1,64,64]Is +.>
Step 1.10, deterministic CurrentAnd low resolution contrast image->Splicing in the second dimension to obtain a dimension of [16,3,64,64 ]]Is input data x of a mixture of (a) 1 ;
According to the invention, the MNIST handwriting numbers added with random circles are used as training sets, each scatterer is assumed to be uniform and lossless, the relative dielectric constants are randomly distributed between 1.5 and 2.5, and the background is free space; wherein the training set field is noise free and the test set field is 10% gaussian noise added. 5000 pieces were randomly selected as training sets from the MNIST training set with one added random circle, 2500 pieces were selected as verification sets, and 1500 pieces were randomly selected as tests from the MNIST test set with one added random circle. In order to verify the effectiveness of the model under different shapes and different noise levels and the generalization capability of the model to real data, four pieces of experimental data with different shapes, different proportions of noise added and dielectric constants of 2 and 3GHz configuration frequency are respectively generated;
step two, building a deep expansion network P θ And mix the input data x 1 As a deep-spread network P θ Is input by the deep expansion network P θ Output target scattererIs approximately true of the complete induced current +.>
Step 2.1, consisting of K cascaded subnetworks { P ] θ,k |k∈[1,K]Form deep expansion network P θ The method comprises the steps of carrying out a first treatment on the surface of the Wherein P is θ,k Representing a kth cascaded subnetwork; in this embodiment, k=4; and the kth cascaded subnetwork P θ,k By usingA U-net structure including a contracted path and an expanded path; the network structure of the sub-network is shown in fig. 2; in this embodiment, the number of input channels and the number of output channels of the deep-spread network are 3 and 2, respectively.
The contraction path is formed by sequentially adding a maximum pooling layer after two convolution blocks, wherein the convolution blocks are formed by a convolution layer with the convolution kernel size of 3 multiplied by 3, a BN layer and a ReLU activation function; the maximum pooling layer downsamples the output feature map of the convolution block, the feature map size is reduced by half after each downsampling, and the channel number is multiplied by 2;
the expansion path is formed by sequentially adding two convolution blocks after deconvolution operation, wherein the deconvolution operation is formed by deconvolution layers with the convolution kernel size of 2 multiplied by 2, and the convolution blocks have the same structure as the contraction path; the feature map size is doubled and the channel number is halved after each deconvolution, and the feature maps with consistent sizes from the corresponding shrink paths are added by using jump connection after each deconvolution operation, so that feature information with different depths is fused. At the last layer of the network is a 1 x 1 convolutional layer, which converts the 64-channel features into the number of feature channels needed;
step 2.2, mixing the input data x when k=1 1 Input depth expansion network P θ And pass through the kth sub-network P θ,k Is processed to obtain a feature map f with dimension c×c=16×16 k Then the kth sub-network P is output after the processing of the expansion path θ,k Predicted induced current
According to the induced currentObtaining a target scatterer by using a contrast updating formula of a state equation, a data equation and SOM>The kth predictive total field of +.>Kth predicted Scattering field->And kth prediction contrast image +.>
When k=2, 3..k, K-1 th subnetwork P θ,k-1 Output induced current matrixAnd k-1 th fringe field image matrix +.>Splicing in the second dimension to obtain kth mixed input data x k And pass through the kth sub-network P θ,k Outputs a kth predicted induced current matrix +.>Thereby by the Kth sub-network P θ,K Outputting a K-th predicted induced current matrix +.>And as a deep-spread network P θ The output approximately true complete induction current +.>
Then according to the approximate real complete induced currentThe target scatterer +.>Other physical information of (a) including predictionGeneral field->Predicted fringe field->Prediction contrast image +.>
Step three, designing a loss function, and establishing a deep expansion network P θ Is an optimization objective of (1);
deep expansion network P θ And the optimized physical parameter theta is continuously obtained in the back propagation process of the loss function, so that the network is guided to learn the induction current, the scattering field and the contrast reconstruction of the target scatterer better.
Step 3.1, constructing a deep expansion network P by utilizing the method (1) θ Target loss function L of (2) P :
L P =L J +L E +λ 1 L SSIM +λ 2 L MSE (1)
In the formula (1), L J Represents an induced current loss and is obtained by the formula (2); l (L) E Represents the loss of the scattered field and is obtained by the formula (3); l (L) SSIM Representing a loss of contrast image quality and obtained by formula (3); l (L) MSE Representing pixel-by-pixel loss and obtained by equation (4); lambda (lambda) 1 ,λ 2 Is a super parameter to balance the effects of image quality loss and pixel-by-pixel loss;
in the formula (2), the amino acid sequence of the compound,indicating the target scatterer corresponding to the jth transmitting antenna>Is a near real induced current matrix of (a) Indicating the target scatterer corresponding to the jth transmitting antenna>Is a simulated induced current of (a);
in the formula (3), the amino acid sequence of the compound,representing the predicted fringe field of the depth-spread network corresponding to the first receive antenna, +.>Indicating the target scatterer corresponding to the first receiving antenna->Is a true scatter field of (2);
in the formula (4), SSIM represents an image structural similarity loss;
in equation (5), in the present embodiment, the number of pixels in one contrast image n=4096;representation predictionContrast image->A contrast value corresponding to the q-th pixel point; />Representing object scatterer->A contrast value corresponding to the q-th pixel point;
fourthly, performing diffuser induced current, a diffuser field and contrast reconstruction through training a depth unfolding network;
based on mixed input data x 1 For deep expansion network P θ Learning is carried out, nonlinear difficulty of network learning is reduced, predicted induced current, scattered field and contrast outputted by a depth expansion network are used for restraining physical quantities of a real scatterer, and a loss function L is calculated P And optimizing the network parameter theta, selecting an optimal network model for realizing high-quality induced current, scattered field and contrast image reconstruction, and ensuring the consistency of a physical model and data.
Depth expansion network uses Adam optimizer to set beta 1 =0.9,β 2 =0.999, batch size is set to 1, training epochs is set to 40, the first 20 epochs keep the initial learning rate at 0.0002, learning rate decays linearly from 21 st epochs until the last epoch learning rate drops to 0. Continuously adjusting network parameters according to the verification set error after the mth training, and adjusting the super parameter lambda in the experimental process 1 And lambda (lambda) 2 And balancing weights among the losses until each loss of the network converges, and selecting an optimal model to realize high-quality reconstruction of the unknown scatterer.
The present invention uses Structural Similarity (SSIM) and Root Mean Square Error (RMSE) as evaluation indicators for intra-dataset and cross-dataset testing. The method provided by the invention is compared with a method for reconstructing relative dielectric constant by using a U-net network, wherein the target loss function used by the U-net network is as followsL U =L SSIM For ease of comparison, the method of reconstructing an image using a U-net network is denoted as U-net in the results presentation.
The method of the invention performs training and testing on the handwriting digital data set and directly performs testing on 'Austria' and experimental data. In the following figures, GT represents a contrast image of a real target scatterer; SOM represents a reconstructed image using a conventional SOM iterative algorithm; u-net means reconstructing an image using a U-net network; the reconstruction result of the MNIST Test set is shown in fig. 3, the first graph in fig. 4 shows a current distribution diagram of test#1 under the action of the first transmitting antenna, which sequentially includes, from left to right, an input induced current, a predicted complete induced current and an analog induced current, where the first row is a real part current, and the second row is an imaginary part current; the second plot shows the fitting of the scattered field from each transmit antenna received by all receive antennas, the first row being the real field and the second row being the imaginary field. Meanwhile, the invention also carries out cross-data set testing, and the reconstruction results on an Austria data set and experimental data are respectively shown in fig. 5, fig. 6, fig. 7 and fig. 8. As shown in fig. 5, the reconstruction results of the "Austria" dataset with dielectric constant of 2 at different ratio noise, with Test #5-Test #8 added with 10%, 20%, 25% and 30% gaussian white noise in order; as shown in fig. 6, the current distribution diagram of test#7 under the action of the first transmitting antenna is that, from left to right, the input induced current, the predicted complete induced current and the analog induced current are sequentially shown, the first row is the real part current, and the second row is the imaginary part current; the second plot shows the fitting of the scattered field from each transmit antenna received by all receive antennas, the first row being the real field and the second row being the imaginary field. Experimental data reconstruction results at a frequency of 3GHz as shown in fig. 7, wherein the dashed line represents the position of the real image; as shown in fig. 8, the current distribution diagram of experimental data under the action of the first transmitting antenna is shown, from left to right, sequentially including an input induced current, a predicted complete induced current and an analog induced current, where the first row is a real current and the second row is an imaginary current; the second plot shows the fitting of the scattered field from each transmit antenna received by all receive antennas, the first row being the real field and the second row being the imaginary field.
According to the reconstruction result, the method provided by the invention can rapidly reconstruct physical quantities such as induced current, scattered field, contrast and the like of a scatterer with high precision, a depth expansion network trained by the method effectively replaces a traditional SOM iteration method, physical model consistency and data consistency are ensured, obvious advantages are shown on Austra data sets and experimental data of different noises, and meanwhile, the trained physical model is proved to learn physical knowledge and have better generalization capability.
Claims (1)
1. An electromagnetic backscatter imaging method based on a physical depth expansion network, comprising the steps of:
step one, constructing mixed input data, including data acquisition and preprocessing;
step 1.1, an electromagnetic scattering system adopts T transmitting antennas and R receiving antennas, and a target scatterer is placed in a square region of interest DThe transmitting antennas sequentially transmit plane wave signals to the region of interest D, and R receiving antennas simultaneously measure scattered fields;
step 1.2, calculating the target scatterer by adopting a moment method during forward modelingIs +.>And simulated fringe field->
Step 1.3, during inversion, discretizing the region of interest D into a grid with a size of M×M, and forming M in a conformal manner 2 A sub-grid;
step 1.4, decomposing the simulated scattered field by singular valuesProcessing to obtain a dimension [ M ] 2 ,T]Deterministic partial current +.>Wherein T represents T transmitting antenna channels;
step 1.5, varying deterministic portion currentAnd obtain the dimension of [ T, M ]]Three-dimensional current image matrix of (2)
Step 1.6, matrix three-dimensional current imageAdding one dimension for storing the three-dimensional current image matrix +.>To obtain the real part and the imaginary part of the dimension [ T, N ] 1 ,M,M]Deterministic current matrix->Wherein N is 1 Representing the deterministic current +.>The real part and the imaginary part channel number;
step 1.7, using a back propagation method on the simulated fringe fieldThe treatment is carried out in such a way that,generating dimension [ M, M ]]Is a low resolution scatterer image χ BP ;
Step 1.8, χ on the low-resolution scatterer image BP Adding one dimension for storing the low resolution scatterer image χ BP To obtain the imaginary part of the dimension N 2 ,M,M]Three-dimensional image matrix of (a)
Step 1.9, matrix the three-dimensional imageAdding one dimension for storing T three-dimensional image matrixes +.>Obtaining the dimension of [ T, N ] 2 ,M,M]Is +.>Wherein N is 2 Representing said low resolution contrast image +.>Is the imaginary channel number of (a);
step 1.10, the deterministic current is processedAnd said low resolution contrast image +.>Splicing in the second dimension to obtain a piece of dimension [ T, N, M ]]Is a mixed input data x; wherein n=n 1 +N 2 Representing mixed input data x 1 The number of real and imaginary channels;
step two, building a deep expansion network P θ And mixes the input data x 1 As a deep-spread network P θ Is extended by the depth by the input of the network P θ Outputting the target scattererIs approximately true of the complete induced current +.>
Step 2.1, consisting of K cascaded subnetworks { P ] θ,k |k∈[1,K]Form deep expansion network P θ The method comprises the steps of carrying out a first treatment on the surface of the Wherein P is θ,k Representing a kth cascaded subnetwork; and the kth cascaded subnetwork P θ,k Adopting a U-net structure, comprising a contracted path and an expanded path;
the contraction path is formed by sequentially adding a maximum pooling layer after two convolution blocks, wherein the convolution blocks consist of a convolution layer with a convolution kernel size of a multiplied by a, a BN layer and a ReLU activation function;
the expansion path is formed by sequentially adding two convolution blocks after deconvolution operation, the deconvolution operation is formed by deconvolution layers with the convolution kernel size of b multiplied by b, and the convolution blocks have the same structure as the contraction path;
step 2.2, when k=1, the mixed input data x 1 Inputting the depth expansion network P θ And pass through the kth sub-network P θ,k Is processed to obtain a feature map f with dimension c×c k Then the kth sub-network P is output after the processing of the expansion path θ,k Predicted induced current
According to the induced currentObtaining the target scatterer by using a contrast updating formula of a state equation, a data equation and SOM>The kth predictive total field of +.>Kth predicted Scattering field->And kth prediction contrast image +.>
When k=2, 3..k, K-1 th subnetwork P θ,k-1 Output induced current matrixAnd k-1 th fringe field image matrix +.>Splicing in the second dimension to obtain kth mixed input data x k And pass through the kth sub-network P θ,k Outputs a kth predicted induced current matrix +.>Thereby by the Kth sub-network P θ,K Outputting a K-th predicted induced current matrix +.>And as a deep-spread network P θ The output approximately true complete induction current +.>
Then according to the approximate real complete induced currentThe target scatterer is further obtained by using a state equation, a data equation and a contrast updating formula of SOM>Including predictive total field +.>Predicted fringe field->Prediction contrast image +.>
Step three, designing a loss function, and establishing a deep expansion network P θ Is an optimization objective of (1);
step 3.1, constructing a deep expansion network P by utilizing the method (1) θ Target loss function L of (2) P :
L P =L J +L E +λ 1 L SSIM +λ 2 L MSE (1)
In the formula (1), L J Represents an induced current loss and is obtained by the formula (2); l (L) E Represents the loss of the scattered field and is obtained by the formula (3); l (L) SSIM Representing a loss of contrast image quality and obtained by formula (3); l (L) MSE Representing pixel-by-pixel loss and obtained by equation (4); lambda (lambda) 1 ,λ 2 Is a super parameter to balance the effects of image quality loss and pixel-by-pixel loss;
in the formula (2), the amino acid sequence of the compound,indicating the target scatterer corresponding to the jth transmitting antenna>Is approximately true of the inductive current matrix-> Indicating the target scatterer corresponding to the jth transmitting antenna>Is a simulated induced current of (a);
in the formula (3), the amino acid sequence of the compound,representing the predicted fringe field of the depth-spread network corresponding to the first receive antenna, +.>Indicating the target scatterer corresponding to the first receiving antenna->Is a true scatter field of (2);
in the formula (4), SSIM represents an image structural similarity loss;
in the formula (5), N represents the number of pixels of one contrast image;representing predictive contrast image +.>A contrast value corresponding to the q-th pixel point; />Representing object scatterer->A contrast value corresponding to the q-th pixel point;
fourthly, performing diffuser induced current, a diffuser field and contrast reconstruction through training a depth unfolding network;
based on the mixed input data x 1 For the deep expansion network P θ Learning and calculating the loss function L P In the process of (2), the network parameter theta is continuously optimized, so that the induced current, the scattered field and the contrast image output by the network reconstruction are gradually fitted to the physical quantity corresponding to the real scattering body, thereby obtaining an optimal network model for realizing high-quality induced current, scattered field and contrast image reconstruction.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202210307192.5A CN114626987B (en) | 2022-03-25 | 2022-03-25 | Electromagnetic backscatter imaging method based on physical depth expansion network |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202210307192.5A CN114626987B (en) | 2022-03-25 | 2022-03-25 | Electromagnetic backscatter imaging method based on physical depth expansion network |
Publications (2)
Publication Number | Publication Date |
---|---|
CN114626987A CN114626987A (en) | 2022-06-14 |
CN114626987B true CN114626987B (en) | 2024-02-20 |
Family
ID=81904718
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202210307192.5A Active CN114626987B (en) | 2022-03-25 | 2022-03-25 | Electromagnetic backscatter imaging method based on physical depth expansion network |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN114626987B (en) |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111428407A (en) * | 2020-03-23 | 2020-07-17 | 杭州电子科技大学 | Electromagnetic scattering calculation method based on deep learning |
CN111610374A (en) * | 2020-05-28 | 2020-09-01 | 杭州电子科技大学 | Scattered field phase recovery method based on convolutional neural network |
CN111999731A (en) * | 2020-08-26 | 2020-11-27 | 合肥工业大学 | Electromagnetic backscattering imaging method based on perception generation countermeasure network |
CN113378472A (en) * | 2021-06-23 | 2021-09-10 | 合肥工业大学 | Mixed boundary electromagnetic backscattering imaging method based on generation countermeasure network |
-
2022
- 2022-03-25 CN CN202210307192.5A patent/CN114626987B/en active Active
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111428407A (en) * | 2020-03-23 | 2020-07-17 | 杭州电子科技大学 | Electromagnetic scattering calculation method based on deep learning |
CN111610374A (en) * | 2020-05-28 | 2020-09-01 | 杭州电子科技大学 | Scattered field phase recovery method based on convolutional neural network |
CN111999731A (en) * | 2020-08-26 | 2020-11-27 | 合肥工业大学 | Electromagnetic backscattering imaging method based on perception generation countermeasure network |
CN113378472A (en) * | 2021-06-23 | 2021-09-10 | 合肥工业大学 | Mixed boundary electromagnetic backscattering imaging method based on generation countermeasure network |
Non-Patent Citations (1)
Title |
---|
大尺度电磁散射与逆散射问题的深度学习方法;王龙刚;钟威;阮恒心;贺凯;李廉林;;电波科学学报(第05期);全文 * |
Also Published As
Publication number | Publication date |
---|---|
CN114626987A (en) | 2022-06-14 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109085589B (en) | Sparse aperture ISAR imaging phase self-focusing method based on image quality guidance | |
CN107193003B (en) | Sparse singular value decomposition scanning radar foresight imaging method | |
CN103713288B (en) | Sparse Bayesian reconstruct linear array SAR formation method is minimized based on iteration | |
Zhang et al. | Resolution enhancement for large-scale real beam mapping based on adaptive low-rank approximation | |
CN108008385B (en) | Interference environment ISAR high-resolution imaging method based on management loading | |
Wang et al. | TPSSI-Net: Fast and enhanced two-path iterative network for 3D SAR sparse imaging | |
Pu | SAE-Net: A deep neural network for SAR autofocus | |
CN110726992B (en) | SA-ISAR self-focusing method based on structure sparsity and entropy joint constraint | |
CN111948652B (en) | SAR intelligent parameterized super-resolution imaging method based on deep learning | |
Dai et al. | SRCNN-based enhanced imaging for low frequency radar | |
CN109932717A (en) | ISAR high-resolution imaging method based on environmental statistics modeling | |
CN114442092B (en) | SAR deep learning three-dimensional imaging method for distributed unmanned aerial vehicle | |
CN115453528A (en) | Method and device for realizing segmented observation ISAR high-resolution imaging based on rapid SBL algorithm | |
CN117192548A (en) | Sparse ISAR high-resolution imaging method based on depth expansion | |
CN113378472B (en) | Mixed boundary electromagnetic backscattering imaging method based on generation countermeasure network | |
Gao et al. | Fast super-resolution 3D SAR imaging using an unfolded deep network | |
Wei et al. | Learning-based split unfolding framework for 3-D mmW radar sparse imaging | |
CN114626987B (en) | Electromagnetic backscatter imaging method based on physical depth expansion network | |
Hu et al. | FCNN-based ISAR sparse imaging exploiting gate units and transfer learning | |
CN117671384A (en) | Hyperspectral image classification method | |
CN115018943B (en) | Electromagnetic backscatter imaging method based on training-free depth cascade network | |
CN114994674B (en) | Intelligent microwave staring correlated imaging method and equipment and storage medium | |
CN116243313A (en) | SAR rapid intelligent sparse self-focusing technology based on distance partition | |
Xu et al. | Backward projection imaging of through-wall radar based on airspace nonuniform sampling | |
Oral et al. | Plug-and-play reconstruction with 3D deep prior for complex-valued near-field MIMO imaging |
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 |