CN114626987A - Electromagnetic backscattering imaging method of deep expansion network based on physics - Google Patents
Electromagnetic backscattering imaging method of deep expansion network based on physics Download PDFInfo
- Publication number
- CN114626987A CN114626987A CN202210307192.5A CN202210307192A CN114626987A CN 114626987 A CN114626987 A CN 114626987A CN 202210307192 A CN202210307192 A CN 202210307192A CN 114626987 A CN114626987 A CN 114626987A
- Authority
- CN
- China
- Prior art keywords
- network
- deep
- image
- induced current
- dimension
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 238000003384 imaging method Methods 0.000 title claims abstract description 15
- 238000012549 training Methods 0.000 claims abstract description 13
- 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 38
- 239000011159 matrix material Substances 0.000 claims description 34
- 230000008602 contraction Effects 0.000 claims description 12
- 238000012545 processing Methods 0.000 claims description 12
- 230000006698 induction Effects 0.000 claims description 7
- 239000011541 reaction mixture Substances 0.000 claims description 6
- 230000000694 effects 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
- 238000005516 engineering process Methods 0.000 abstract description 5
- 230000006870 function Effects 0.000 description 16
- 238000012360 testing method Methods 0.000 description 14
- 238000010586 diagram Methods 0.000 description 7
- 238000013528 artificial neural network Methods 0.000 description 4
- 238000013135 deep learning Methods 0.000 description 4
- 230000009471 action Effects 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
- 230000007310 pathophysiology Effects 0.000 description 2
- 238000005070 sampling Methods 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
- 238000013527 convolutional neural network Methods 0.000 description 1
- 238000000354 decomposition reaction Methods 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 238000000790 scattering method Methods 0.000 description 1
- 230000007480 spreading Effects 0.000 description 1
- 238000013519 translation Methods 0.000 description 1
Images
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 backscattering imaging method based on a physical deep-developed network, which comprises the following steps: 1, constructing mixed input data, including data acquisition and preprocessing, and enriching network input information; 2, in a network structure building stage, designing a network structure of a deep expansion technology by using the deep expansion technology in combination with a traditional subspace optimization iterative algorithm; designing a loss function, and jointly optimizing a network by using the image structure similarity loss and the pixel-by-pixel loss in the objective function, particularly using the induced current loss and the scattered field loss; 4, by training the deep expansion network, the induced current, the scattering field and the contrast of the target scatterer can be quickly reconstructed with high quality. The deep-expansion network based on physics provided by the invention can effectively replace the traditional SOM iterative algorithm, enrich the physical knowledge of the network and realize rapid and high-precision electromagnetic backscatter imaging.
Description
Technical Field
The invention belongs to the technical field of electromagnetic backscatter imaging, and particularly relates to electromagnetic backscatter imaging by effectively replacing a traditional SOM iteration method with a deep learning method.
Background
Electromagnetic backscattering is the determination of properties, such as position, shape and physical parameters, of unknown scatterers within a spatial region from the distribution of the scattered field within that spatial region. However, the electromagnetic backscattering problem (ISP) presents a high degree of nonlinearity and pathophysiology, the nonlinearity is due to multiple scattering effects between the measured scattering field and the scatterer, and the pathophysiology is mainly due to the fact that small perturbations of the observed data will cause large errors in the solution. The ISP can be solved by conventional objective function methods, which mainly include back propagation method (BP) of linear approximation, modified born iteration method (DBIM) of nonlinear iteration, Contrast Source Inversion (CSI), Subspace Optimization (SOM), and the like. In the traditional method, fluctuation physics is integrated into a model, but the linear method is rough in imaging, and the iterative method is high in calculation cost and long in time consumption.
In recent years, due to the strong learning mapping capability and the fast solving speed of the deep learning network, researchers successfully apply the deep learning network to solving the electromagnetic backscattering problem. For example, the Direct Inversion Scheme (DIS) proposed by Wei et al is a typical algorithm for mapping scatterers from a scattering field to a target scatterer using a neural network, but it can only reconstruct some simple scatterers within a training set. Li et al propose a 'DeepNIS' algorithm based on a rewinding product neural network by analogy of the relation between the traditional nonlinear iterative method and 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 cascaded network exclusively for learning the fuzzy part of the induced current. Huang et al simplify the backscattering problem to an image translation problem, first use the back-propagation method to obtain a coarse image, and then use the neural network to achieve high resolution reconstruction of the image. The test results in the above paper show that the current depth inverse scattering method is superior to the traditional nonlinear optimization method in both imaging quality and speed.
The above approach is limited by the quality of the input and the prior of the scatterer boundary type, and the generalization capability is limited especially when the network lacks physical knowledge guidance. The deep backscattering method needs to take into account physical model consistency and data consistency. The method makes up the gap between the traditional objective function method and the deep learning method driven by data, and effectively embeds physical knowledge into the deep neural network to realize high-quality imaging, which is a key technical problem.
Disclosure of Invention
The invention aims to solve the key technical problems and provides an electromagnetic backscattering imaging method based on a physical deep unfolding network, so that the deep unfolding technology, an SOM iteration framework and the existing physical knowledge can be effectively combined, the network learns the physical knowledge and enhances the generalization capability of a model, the rapid and high-precision electromagnetic backscattering imaging is realized, and the high-quality reconstruction of the induced current, the scattering field and the contrast of a scatterer is further realized.
The invention adopts the following technical scheme for solving the technical problems:
the invention discloses an electromagnetic backscattering imaging method based on a physical deep unfolding network, which is characterized by comprising the following steps of:
step one, constructing mixed input data, including data acquisition and preprocessing;
step 1.1, the 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 the R receiving antennas simultaneously measure a scattered field;
step 1.2, during forward performance, calculating the target scatterer by adopting a moment methodAnalog induced current ofAnd simulating the scattered field
Step 1.3, during inversion, dispersing the region of interest D into a grid with the size of M multiplied by M, and forming M by the same2A sub-grid;
step 1.4, the simulated scattered field is decomposed by using singular valuesProcessing to obtain dimension [ M2,T]Deterministic partial currentWherein, T represents T transmitting antenna channels;
step 1.5, changing deterministic partial CurrentAnd obtaining the dimension of [ T, M]Three-dimensional current image matrix of
Step 1.6, three-dimensional current image matrixAdding a dimension for storing the three-dimensional current image matrixReal and imaginary parts of (a), thereby obtaining a dimension of [ T, N ]1,M,M]Deterministic current matrixWherein, N1Representing said deterministic currentThe number of real and imaginary channels of (1);
step 1.7, utilizing a back propagation method to simulate the scattered fieldProcessing to generate dimension [ M, M]Low resolution scatterer image χBP;
Step 1.8, correcting the low-resolution scatterer image chiBPAdding a dimension for storing the low resolution scatterer image χBPTo obtain an imaginary part of dimension [ N ]2,M,M]Three-dimensional image matrix of
Step 1.9, the three-dimensional image matrix is processedAdding one dimension for storing T three-dimensional image matrixesTo obtain a dimension of [ T, N2,M,M]Low resolution contrast image ofWherein N is2Representing the low resolution contrast imageThe number of imaginary channels of (a);
step 1.10, the deterministic current is appliedAnd the low resolution contrastDegree imageSplicing in the second dimension to obtain the dimension [ T, N, M]Mixed input data x of (2); wherein N is N1+N2Representing mixed input data x1The number of real and imaginary channels of (a);
step two, building a deep expansion network PθAnd mixing the input data x1As a deep developed network PθBy said deep-developed network PθOutputting the target scattererApproximately true full induced current of
Step 2.1, cascading K subnetworks { P }θ,k|k∈[1,K]Form a deep-developed network Pθ(ii) a Wherein, Pθ,kRepresents the kth cascaded subnetwork; and the kth cascaded subnetwork Pθ,kAdopting a U-net structure comprising a contraction path and an expansion path;
the contraction path is formed by adding a maximum pooling layer after two convolution blocks in sequence, wherein each convolution block is formed by a convolution layer with convolution kernel size of a multiplied by a, a BN layer and a ReLU activation function;
the expansion path is formed by adding two convolution blocks after one deconvolution operation in sequence, the deconvolution operation is formed by a deconvolution layer with a convolution kernel of b multiplied by b, and the structure of the convolution blocks is the same as that of the contraction path;
step 2.2, when k is 1, the mixed input data x1Inputting the deep developed network PθAnd passes through the kth sub-network Pθ,kThe feature map f with the dimension of c x c is obtained by the contraction path processing of (1)kThen the k-th sub-network P is output after the processing of the expanded pathθ,kPredicted induced current
According to the induced currentObtaining the target scatterer by using a state equation, a data equation and a contrast updating formula of the SOMK predicted total field ofPredicted scatter field kAnd the k-th predicted contrast image
When K is 2,3, K, the K-1 st sub-network Pθ,k-1Output induced current matrixAnd the k-1 st scattered field image matrixSplicing in the second dimension to obtain the kth mixed input data xkAnd passes through the kth sub-network Pθ,kOutput the k-th predicted induced current matrixThereby consisting of the Kth sub-network Pθ,KOutputting the Kth predicted induced current matrixAnd as a deep deployment network PθApproximate real complete induction current of output
Then according to the complete induction current which is approximate to the realFurther obtaining the target scatterer by using a state equation, a data equation and a contrast updating formula of the SOMIncluding predicting the total fieldPredicting the scattered fieldAnd predicting the contrast image
Step three, designing a loss function and establishing a deep expansion network PθThe optimization objective of (2);
step 3.1, constructing a deep expansion network P by using the formula (1)θTarget loss function L ofP:
LP=LJ+LE+λ1LSSIM+λ2LMSE (1)
In the formula (1), LJRepresents an induced current loss and is obtained by the formula (2); l isERepresents the scattered field loss and is obtained by the formula (3); l isSSIMRepresents a loss of contrast image quality and is obtained by equation (3); l isMSERepresents the pixel-by-pixel loss and is obtained by the formula (4); lambda1,λ2Is a hyper-parameter used to balance the effects of image quality loss and pixel-by-pixel loss;
in the formula (2), the reaction mixture is,target scatterer corresponding to j-th transmitting antennaApproximately real induced current matrix of Target scatterer corresponding to j-th transmitting antennaThe simulated induced current of (a);
in the formula (3), the reaction mixture is,representing the scattered field predicted by the depth expanded network corresponding to the ith receiving antenna,target scatterer corresponding to the first receiving antennaThe true fringe field of (a);
in the formula (4), SSIM represents loss of image structural similarity;
in the formula (5), N represents the number of pixels of one contrast image;representing a predicted contrast imageThe corresponding contrast value of the qth pixel point of (1);representing a target scattererThe corresponding contrast value of the qth pixel point of (1);
fourthly, reconstructing induced current, scattering field and contrast of the scatterer by training a deep unfolding network;
based on the mixed input data x1For the deep developed network PθLearning is performed and the loss function L is calculatedPIn the process, the network parameter theta is continuously optimized, so that the induced current, the scattering field and the contrast ratio image output by network reconstruction are gradually fitted to the physical quantity corresponding to the real scatterer, and an optimal network model is obtained and is used for realizing the reconstruction of the induced current, the scattering field and the contrast ratio image with high quality.
Compared with the prior art, the invention has the beneficial effects that:
1. the invention provides an electromagnetic backscattering imaging method of a deep expansion network based on physics, which uses two characteristic information of induction current and contrast to carry out mixed input 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 deep expansion technology to combine the traditional SOM iteration method and the existing physical knowledge to establish physical network mapping, uses the characteristic information such as induced current, scattered field and contrast to carry out comprehensive constraint, effectively replaces the traditional SOM iteration method, contains rich physical knowledge, and improves the network prediction accuracy and generalization capability.
3. In addition to a contrast loss function, a current loss function and a scattered field loss function are particularly introduced into the target 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 expanded subnetwork structure of the present invention;
FIG. 3 is a diagram showing a result of reconstructing a MNIST handwritten digital data set according to the present invention;
FIG. 4 is a graph showing the reconstruction results of Test # 1 induced current and scattered field with 10% noise added according to the present invention;
FIG. 5 is a diagram showing the reconstruction result of "Austria" data set with different proportions of noise added according to the present invention;
FIG. 6 is a graph showing the reconstruction results of induced current and scattered field of "Austria" with 25% noise added in accordance with the present invention;
FIG. 7 is a diagram showing the result of experimental data reconstruction with a frequency of 3GHz according to the present invention;
FIG. 8 is a graph showing the reconstruction results of induced current and scattered field of the experimental data with the 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 x1As a deep developed network PθOf the approximate real induced current matrix of the scattererAs a deep developed network PθAn output of (d); the traditional SOM method is effectively replaced by using a deep expansion technology in combination with the traditional SOM iterative process, and one-time iterative process of the SOM is mapped into a deep expansion network PθEach module updates the induced current and the contrast simultaneously, and the deep spreading network PθRealize complete inductionBesides the high-quality reconstruction of the current, the scattering field and the contrast can be further obtained through calculation of a state equation, a data equation and a contrast updating formula of the 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 16 transmitting antennas T and 32 receiving antennas R, and places target scatterers in a square region of interest DThe transmitting antenna sequentially transmits plane wave signals to the interested region D, and the scattering fields are measured simultaneously by 32 receiving antennas; assuming an operating frequency of 400MHZ, the region of interest D is 2.0 meters by 2.0 meters.
Step 1.2, placing a target scatterer in the region of interest D during forward operationTo avoid the occurrence of the "inverse crime" phenomenon in the inverse problem, for each incidence, the target scatterers were calculated on a 100 × 100 grid using the moment methodAnalog induced current ofAnd simulating the scattered field
Step 1.3, in inversion, dispersing the region of interest D into a grid with a size of M × M-64 × 64, and forming M together24096 sub-grids;
step 1.4, simulating a scattering field by using singular value decompositionProcessing to obtain dimension [ M2,T]=[4096,16]IndeedQualitative partial current
Step 1.5, changing deterministic partial CurrentAnd obtaining the dimension of [ T, M]=[16,64,64]Three-dimensional current image matrix of
Step 1.6, three-dimensional current image matrixAdding a dimension for storing a three-dimensional current image matrixTo obtain the real and imaginary components of dimension [16,2, 64%]Deterministic current matrix
Step 1.7, simulating the scattered field by using a back propagation methodProcessing to generate dimension [ M, M]Low resolution scatterer image χBP;
Step 1.8, aiming at low-resolution scatterer image chiBPAdding one dimension for storing low-resolution scatterer images xBPTo obtain an imaginary part of dimension N2,M,M]=[1,64,64]Three-dimensional image matrix of
Step 1.9, three-dimensional image matrixAdding one dimension for storing T-16 three-dimensional image matrixesTo obtain a dimension of [ T, N2,M,M]=[16,1,64,64]Low resolution contrast image of
Step 1.10, deterministic CurrentAnd low resolution contrast imagesSplicing in the second dimension to obtain the dimension [16,3,64]Mixed input data x of1;
The invention adopts MNIST handwritten figures added with random circles as a training set, each scatterer is assumed to be uniform and has no loss, the relative dielectric constant is randomly distributed between 1.5 and 2.5, and the background is a free space; wherein the training set scattered field does not contain noise, and the test set scattered field is added with 10% of Gaussian noise. 5000 pieces of MNIST training sets added with a random circle are randomly selected as training sets, 2500 pieces of MNIST training sets are selected as verification sets, and 1500 pieces of MNIST testing sets added with a random circle are randomly selected as tests. 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, dielectric constant of 2 and configuration frequency of 3GHz are generated respectively;
step two, building a deep expansion network PθAnd mixing the input data x1As a deep developed network PθBy a deep-developed network PθOutput target scattererApproximately true full induced current of
Step 2.1, cascading K sub-networks { P }θ,k|k∈[1,K]Form a deep-expanded network Pθ(ii) a Wherein, Pθ,kRepresents the kth cascaded subnetwork; in the present embodiment, K ═ 4; and the kth cascaded subnetwork Pθ,kAdopting a U-net structure comprising a contraction path and an expansion 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-scaling network are 3 and 2, respectively.
The contraction path is formed by adding a maximum pooling layer after two convolution blocks in sequence, wherein each convolution block is formed by a convolution layer with convolution kernel size of 3 multiplied by 3, a BN layer and a ReLU activation function; the maximum pooling layer performs down-sampling on the output feature map of the convolution block, the size of the feature map is reduced by half and the number of channels is multiplied by 2 after each down-sampling;
the expansion path is formed by adding two convolution blocks after one deconvolution operation in sequence, the deconvolution operation is formed by a deconvolution layer with a convolution kernel size of 2 multiplied by 2, and the convolution blocks have the same structure as the contraction path; after each deconvolution, the size of the feature graph is doubled and the number of channels is halved, and meanwhile, after each deconvolution operation, feature information of different depths is fused by using jump connection and feature graphs with the same size from corresponding contraction paths. At the last layer of the network, a 1 × 1 convolutional layer is provided, and this operation can convert the 64-channel features into the required number of feature channels;
step 2.2, when k is 1, mixing the input data x1Input deep unfolding network PθAnd passes through the kth sub-network Pθ,kThe feature map f having a dimension of c × c 16 × 16 is obtained by the narrowing-down path processing of (1)kThen the k-th sub-network P is output after the processing of the expanded pathθ,kPredicted induced current
According to the induced currentObtaining the target scatterer by using a state equation, a data equation and a contrast updating formula of the SOMK predicted total field ofK predicted scatter fieldAnd the k-th predicted contrast image
When K is 2,3,.., K, the K-1 st sub-network Pθ,k-1Output induced current matrixAnd a k-1 st fringe field image matrixSplicing in the second dimension to obtain the kth mixed input data xkAnd passes through the kth sub-network Pθ,kOutput the k-th predicted induced current matrixThereby consisting of the Kth sub-network Pθ,KOutputting the Kth predicted induced current matrixAnd as a deep deployment network PθApproximate real complete induction current of output
Then according to the complete induction current which is approximate to the realFurther obtaining the target scatterer by using a state equation, a data equation and a contrast updating formula of the SOMIncluding predicting the total fieldPredicting the scattered fieldAnd predicting the contrast image
Step three, designing a loss function and establishing a deep expansion network PθThe optimization objective of (2);
deep-developed network PθAnd continuously obtaining an optimized physical parameter theta in the back propagation process of the loss function, and guiding the network to better learn induced current, scattering field and contrast ratio reconstruction of the target scatterer.
Step 3.1, constructing a deep expansion network P by using the formula (1)θTarget loss function L ofP:
LP=LJ+LE+λ1LSSIM+λ2LMSE (1)
In the formula (1), LJRepresents an induced current loss and is obtained by the formula (2); l isERepresents the scattered field loss and is obtained by the formula (3); l is a radical of an alcoholSSIMRepresents a loss of contrast image quality and is obtained by equation (3); l isMSERepresents the pixel-by-pixel loss and is obtained by equation (4); lambda [ alpha ]1,λ2Is a hyper-parameter to balance the effects of image quality loss and pixel-by-pixel loss;
in the formula (2), the reaction mixture is,target scatterer corresponding to j-th transmitting antennaApproximately real induced current matrix of Target scatterer corresponding to j-th transmitting antennaThe simulated induced current of (a);
in the formula (3), the reaction mixture is,representing the scattered field predicted by the depth expanded network corresponding to the ith receiving antenna,target scatterer corresponding to the first receiving antennaThe true fringe field of (a);
in the formula (4), SSIM represents loss of image structural similarity;
in the formula (5), in this embodiment, the number N of pixels in one contrast image is 4096;representing a predicted contrast imageThe corresponding contrast value of the qth pixel point of (1);representing a target scattererThe corresponding contrast value of the qth pixel point of (1);
fourthly, reconstructing induced current, scattering field and contrast of the scatterer by training a deep unfolding network;
based on mixed input data x1For deep developed network PθLearning is carried out, the nonlinear difficulty of network learning is reduced, the predicted induced current, scattering field and contrast output by a deep expansion network and the physical quantity of a real scatterer are used as constraints, and a loss function L is calculatedPAnd optimizing a network parameter theta, and selecting an optimal network model for realizing high-quality induced current, scattered field and contrast image reconstruction, thereby ensuring the consistency of the physical model and the data consistency.
The deep-unfolding network uses an Adam optimizer, setting beta1=0.9,β2At 0.999, the batch size is set to 1, the training epoch is set to 40, the first 20 epochs maintain an initial learning rate of 0.0002, the learning rate decays linearly from the 21 st epoch until the last epoch learning rate drops to 0. Continuously adjusting network parameters according to the error of the verification set after the mth training, and simultaneously adjusting the hyper-parameter lambda in the experimental process1And λ2Balancing the weights between losses until the losses of the network converge, and selecting an optimal model to achieve high-quality reconstruction of unknown scatterers。
The present invention uses Structural Similarity (SSIM) and Root Mean Square Error (RMSE) as evaluation indicators for both intra-and cross-dataset testing. The method provided by the invention is compared with a method for reconstructing the relative dielectric constant by using a U-net network, wherein the target loss function used by the U-net network is LU=LSSIMFor the sake of comparison, the method of reconstructing an image using a U-net network is referred to as U-net in the description of the result.
The method of the present invention performs training and testing on handwritten digital data sets and directly performs testing on "Austria" and experimental data. In the following figures, GT represents a contrast image of a real target scatterer; SOM denotes a reconstructed image using a conventional SOM iterative algorithm; u-net represents the reconstruction of images using a U-net network; the reconstruction result in the MNIST Test set is shown in fig. 3, the first graph in fig. 4 is a current distribution diagram of Test # 1 under the action of the first transmitting antenna, from left to right, sequentially including input induced current, predicted complete induced current and simulated induced current, the first row is real part current, and the second row is imaginary part current; the second graph shows the fitting results of the scattered field from each transmit antenna received by all receive antennas, the first line being the real field and the second line being the imaginary field. Meanwhile, the invention also carries out cross-dataset test, and the reconstruction results on the 'Austria' dataset and the experimental data are respectively shown in FIG. 5, FIG. 6, FIG. 7 and FIG. 8. As shown in FIG. 5, the reconstructed results of "Austria" data set with dielectric constant of 2 under different scales of noise, wherein Test #5-Test # 8 are sequentially added with 10%, 20%, 25% and 30% of white Gaussian noise; as shown in fig. 6, the current distribution diagram of Test # 7 under the action of the first transmitting antenna is sequentially an input induced current, a predicted complete induced current and a simulated induced current from left to right, wherein the first row is a real part current, and the second row is an imaginary part current; the second graph shows the fitting results of the scattered field from each transmit antenna received by all receive antennas, the first line being the real field and the second line being the imaginary field. The result of the experimental data reconstruction with a frequency of 3GHz is shown in fig. 7, in which the dotted line represents the position of the real image; fig. 8 shows a current distribution diagram of experimental data under the action of a first transmitting antenna, which sequentially includes, from left to right, an input induced current, a predicted complete induced current, and a simulated induced current, where a first row is a real part current and a second row is an imaginary part current; the second graph shows the fitting results of the scattered field from each transmit antenna received by all receive antennas, the first line being the real field and the second line being the imaginary field.
According to the reconstruction result, the method can rapidly and accurately reconstruct physical quantities such as scatterer induced current, scattering field, contrast ratio and the like, the deep expansion network trained by the method effectively replaces the traditional SOM iterative method, the consistency and the data consistency of the physical model are ensured, the obvious advantages are shown on 'Austria' data sets and experimental data of different noises, and the fact that the trained physical model learns physical knowledge and has better generalization capability is also proved.
Claims (1)
1. An electromagnetic backscattering imaging method based on a physical depth expansion network is characterized by comprising the following steps:
step one, constructing mixed input data, including data acquisition and preprocessing;
step 1.1, the 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 the R receiving antennas simultaneously measure a scattered field;
step 1.2, during forward performance, calculating the target scatterer by adopting a moment methodAnalog induced current ofAnd simulating the scattered field
Step 1.3, during inversion, dispersing the region of interest D into a grid with the size of M multiplied by M, and forming M by the same2A sub-grid;
step 1.4, the simulated scattered field is decomposed by using singular valuesProcessing to obtain dimension [ M2,T]Deterministic partial currentWherein, T represents T transmitting antenna channels;
step 1.5, changing deterministic partial CurrentAnd obtaining the dimension of [ T, M]Three-dimensional current image matrix of
Step 1.6, three-dimensional current image matrixAdding a dimension for storing the three-dimensional current image matrixReal and imaginary parts of (a), thereby obtaining a dimension of [ T, N ]1,M,M]Deterministic current matrixWherein N is1Representing said deterministic currentThe number of real and imaginary channels of (1);
step 1.7, utilizing a back propagation method to simulate the scattered fieldProcessing to generate dimension [ M, M]Low resolution scatterer image χBP;
Step 1.8, aiming at the low-resolution scatterer image xBPAdding a dimension for storing the low resolution scatterer image χBPTo obtain an imaginary part of dimension [ N ]2,M,M]Of the three-dimensional image matrix
Step 1.9, the three-dimensional image matrix is processedAdding one dimension for storing T three-dimensional image matrixesTo obtain a dimension of [ T, N2,M,M]Low resolution contrast image ofWherein N is2Representing the low resolution contrast imageThe number of imaginary channels of (a);
step 1.10, the deterministic current is appliedAnd the low resolution contrast imageSplicing in the second dimension to obtain the dimension of [ T, N,M,M]Mixed input data x of (2); wherein N is N1+N2Representing mixed input data x1The number of real and imaginary channels of (a);
step two, building a deep expansion network PθAnd mixing the input data x1As a deep developed network PθBy said deep-developed network PθOutputting the target scattererApproximately true full induced current of
Step 2.1, cascading K sub-networks { P }θ,k|k∈[1,K]Form a deep-expanded network Pθ(ii) a Wherein, Pθ,kRepresents the kth cascaded subnetwork; and the kth cascaded subnetwork Pθ,kAdopting a U-net structure comprising a contraction path and an expansion path;
the contraction path is formed by adding a maximum pooling layer after two convolution blocks in sequence, wherein each convolution block is formed by a convolution layer with convolution kernel size of a multiplied by a, a BN layer and a ReLU activation function;
the expansion path is formed by adding two convolution blocks after one deconvolution operation in sequence, the deconvolution operation is formed by a deconvolution layer with a convolution kernel size of b x b, and the convolution blocks have the same structure as the contraction path;
step 2.2, when k is 1, the mixed input data x1Inputting the deep developed network PθAnd passes through the kth sub-network Pθ,kThe feature map f with the dimension of c x c is obtained by the contraction path processing of (1)kThen the k-th sub-network P is output after the processing of the expanded pathθ,kPredicted induced current
According to the induced currentObtaining the target scatterer by using a state equation, a data equation and a contrast updating formula of the SOMK predicted total field ofPredicted scatter field kAnd the k-th predicted contrast image
When K is 2,3,.., K, the K-1 st sub-network Pθ,k-1Output induced current matrixAnd the k-1 st scattered field image matrixSplicing in the second dimension to obtain the kth mixed input data xkAnd passes through the kth sub-network Pθ,kOutput the k-th predicted induced current matrixThereby consisting of the Kth sub-network Pθ,KOutputting the Kth predicted induced current matrixAnd as a deep deployment network PθApproximate real complete induction current of output
Then according to the complete induction current which is approximate to the realFurther obtaining the target scatterer by using a state equation, a data equation and a contrast updating formula of the SOMIncluding predicting the total fieldPredicting the scattered fieldAnd predicting the contrast image
Step three, designing a loss function and establishing a deep expansion network PθThe optimization objective of (2);
step 3.1, constructing a deep expansion network P by using the formula (1)θTarget loss function L ofP:
LP=LJ+LE+λ1LSSIM+λ2LMSE (1)
In the formula (1), LJRepresents an induced current loss and is obtained by the formula (2); l isERepresents the scattered field loss and is obtained by the formula (3); l isSSIMRepresents a loss of contrast image quality and is obtained by equation (3); l isMSERepresents the pixel-by-pixel loss and is obtained by equation (4); lambda [ alpha ]1,λ2Is a hyper-parameter to balance the effects of image quality loss and pixel-by-pixel loss;
in the formula (2), the reaction mixture is,target scatterer corresponding to j-th transmitting antennaApproximately real induced current matrix of Target scatterer corresponding to j-th transmitting antennaThe simulated induced current of (a);
in the formula (3), the reaction mixture is,representing the scattered field predicted by the depth expanded network corresponding to the ith receiving antenna,target scatterer corresponding to the first receiving antennaThe true fringe field of (a);
in the formula (4), SSIM represents loss of image structural similarity;
in the formula (5), N represents the number of pixels of one contrast image;representing a predicted contrast imageThe corresponding contrast value of the qth pixel point of (1);representing a target scattererThe corresponding contrast value of the qth pixel point of (1);
fourthly, reconstructing induced current, scattering field and contrast of the scatterer by training a deep unfolding network;
based on the mixed input data x1For the deep developed network PθLearning is performed and the loss function L is calculatedPIn the process, the network parameter theta is continuously optimized, so that the induced current, the scattering field and the contrast ratio image output by network reconstruction are gradually fitted to the physical quantity corresponding to the real scatterer, and an optimal network model is obtained and is used for realizing the reconstruction of the induced current, the scattering field and the contrast ratio image with high quality.
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 true CN114626987A (en) | 2022-06-14 |
CN114626987B 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 |
---|
王龙刚;钟威;阮恒心;贺凯;李廉林;: "大尺度电磁散射与逆散射问题的深度学习方法", 电波科学学报, no. 05 * |
Also Published As
Publication number | Publication date |
---|---|
CN114626987B (en) | 2024-02-20 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110074813B (en) | Ultrasonic image reconstruction method and system | |
CN111814875B (en) | Ship sample expansion method in infrared image based on pattern generation countermeasure network | |
Yonel et al. | Deep learning for passive synthetic aperture radar | |
CN109683161B (en) | Inverse synthetic aperture radar imaging method based on depth ADMM network | |
CN103713288B (en) | Sparse Bayesian reconstruct linear array SAR formation method is minimized based on iteration | |
CN107193003B (en) | Sparse singular value decomposition scanning radar foresight imaging method | |
Wang et al. | TPSSI-Net: Fast and enhanced two-path iterative network for 3D SAR sparse imaging | |
CN110726992B (en) | SA-ISAR self-focusing method based on structure sparsity and entropy joint constraint | |
CN112906300B (en) | Polarization SAR soil humidity inversion method based on double-channel convolutional neural network | |
CN109343060B (en) | ISAR imaging method and system based on deep learning time-frequency analysis | |
CN107015225A (en) | A kind of SAR platform elemental height error estimation based on self-focusing | |
CN111948652B (en) | SAR intelligent parameterized super-resolution imaging method based on deep learning | |
CN114117886A (en) | Water depth inversion method for multispectral remote sensing | |
CN109557540A (en) | Total variation regularization relevance imaging method based on target scattering coefficient nonnegativity restrictions | |
CN109991602A (en) | ISAR image resolution enhancement method based on depth residual error network | |
CN114442092B (en) | SAR deep learning three-dimensional imaging method for distributed unmanned aerial vehicle | |
CN113378472B (en) | Mixed boundary electromagnetic backscattering imaging method based on generation countermeasure network | |
CN115453528A (en) | Method and device for realizing segmented observation ISAR high-resolution imaging based on rapid SBL algorithm | |
Hu et al. | FCNN-based ISAR sparse imaging exploiting gate units and transfer learning | |
CN115018943B (en) | Electromagnetic backscatter imaging method based on training-free depth cascade network | |
CN114626987A (en) | Electromagnetic backscattering imaging method of deep expansion network based on physics | |
CN114994674B (en) | Intelligent microwave staring correlated imaging method and equipment and storage medium | |
Lorintiu et al. | Compressed sensing reconstruction of 3D ultrasound data using dictionary learning | |
CN113311429A (en) | 1-bit radar imaging method based on countermeasure sample | |
CN115100538A (en) | Three-dimensional water quality spatial interpolation method based on remote sensing image |
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 |