CN117807876A - Multi-mapping electromagnetic inversion method and system based on CycleGAN - Google Patents
Multi-mapping electromagnetic inversion method and system based on CycleGAN Download PDFInfo
- Publication number
- CN117807876A CN117807876A CN202311841903.8A CN202311841903A CN117807876A CN 117807876 A CN117807876 A CN 117807876A CN 202311841903 A CN202311841903 A CN 202311841903A CN 117807876 A CN117807876 A CN 117807876A
- Authority
- CN
- China
- Prior art keywords
- data
- cyclegan
- domain
- electromagnetic
- representing
- 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
Classifications
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
- G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01V—GEOPHYSICS; GRAVITATIONAL MEASUREMENTS; DETECTING MASSES OR OBJECTS; TAGS
- G01V3/00—Electric or magnetic prospecting or detecting; Measuring magnetic field characteristics of the earth, e.g. declination, deviation
- G01V3/08—Electric or magnetic prospecting or detecting; Measuring magnetic field characteristics of the earth, e.g. declination, deviation operating with magnetic or electric fields produced or modified by objects or geological structures or by detecting devices
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01V—GEOPHYSICS; GRAVITATIONAL MEASUREMENTS; DETECTING MASSES OR OBJECTS; TAGS
- G01V3/00—Electric or magnetic prospecting or detecting; Measuring magnetic field characteristics of the earth, e.g. declination, deviation
- G01V3/12—Electric or magnetic prospecting or detecting; Measuring magnetic field characteristics of the earth, e.g. declination, deviation operating with electromagnetic waves
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/0475—Generative networks
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; 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
- G06N3/094—Adversarial learning
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- General Physics & Mathematics (AREA)
- Remote Sensing (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Software Systems (AREA)
- General Engineering & Computer Science (AREA)
- Computational Linguistics (AREA)
- Computing Systems (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Geophysics (AREA)
- Data Mining & Analysis (AREA)
- General Life Sciences & Earth Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Health & Medical Sciences (AREA)
- Geology (AREA)
- Mathematical Physics (AREA)
- Environmental & Geological Engineering (AREA)
- Electromagnetism (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Medical Informatics (AREA)
- Computer Hardware Design (AREA)
- Geometry (AREA)
- Radar Systems Or Details Thereof (AREA)
Abstract
The invention discloses a CycleGAN-based multi-mapping electromagnetic inversion method and a CycleGAN-based multi-mapping electromagnetic inversion system, wherein the method comprises the following steps: constructing a two-dimensional electromagnetic backscatter model; collecting scattering data of different angles of a target area and preprocessing the data; constructing a CycleGAN frame and a corresponding loss function, wherein the CycleGAN frame receives scattering data of different angles of a target area; training a CycleGAN framework; inputting scattering data of different angles of a target area acquired in real time into a trained CycleGAN frame to generate a reconstructed image of the target area; the invention has the advantages that: the method has the advantages of simple process, no need of spending a large amount of labor cost, consideration of the non-uniform distribution condition of the scatterers, more comprehensive extraction of scattering data aiming at the physical characteristics of the electromagnetic model, and improvement of the reconstruction effect.
Description
Technical Field
The invention relates to the field of electromagnetic backscattering, in particular to a multi-mapping electromagnetic inversion method and system based on CycleGAN.
Background
The electromagnetic backscattering problem, which is a core topic in the electromagnetic detection field, relates to reconstructing electrical property parameters (such as conductivity, permeability and permittivity) of a medium target in a region of interest from a forward model of the measurement data of a receiving antenna and the propagation of electromagnetic waves. The core of the process is that the geometry and the spatial position of a medium scatterer in a target area are revealed by analyzing received scattered field data, and a corresponding medium model is constructed. As a non-contact and nondestructive detection method, the method is widely applied to the fields of military detection, medical imaging, geographical exploration and the like.
Aiming at the nonlinear and pathological problems of electromagnetic backscattering, the general traditional methods comprise a linear approximation optimization algorithm and an iterative inversion algorithm. The linear approximation optimization algorithm simply and approximately equivalent the total field as the incident field ignores the multiple scattering effect, and the algorithm is applicable to the limitation because the linear approximation optimization algorithm can not accurately and quantitatively reconstruct the strong scatterer. Common iterative algorithms include extended Born algorithms, contrast source inversion methods, subspace-based optimization algorithms, and the like. The iterative algorithm relies on Yu Gelin functions, and the green function must be constructed case by case in the face of complex situations, so that the algorithm lacks versatility and robustness, and the calculation cost and the time cost are high.
With the deep learning method introduced into electromagnetic inversion, the traditional formula derivation process is replaced by a black box form, and the multi-layer network approximation derivation solves the problem of nonlinearity, and simultaneously solves the limitation of the electromagnetic inversion algorithm facing various situations. However, the common deep learning algorithm mostly adopts a direct inversion mode, for example, a hybrid input method for solving the electromagnetic backscattering problem based on deep learning disclosed in Chinese patent publication No. CN111488549A lacks physical characteristics aiming at an electromagnetic model, and only general information about geometry and electromagnetic characteristics of a scatterer can be extracted; in the initial parameter setting, the spatial and geometric characteristics of the scatterers and the dielectric constants are the global smooth average value, so that the non-uniform distribution condition of the scatterers is ignored; the pathogenicity problem of electromagnetic inversion is not well addressed either, there may be no exact solution or the solution is not unique for a given scatter data, meaning that even minor variations or errors may lead to completely different reconstruction results.
In summary, the conventional algorithm formula reasoning process is too complicated, a great deal of labor cost is required, the function needs to be reconstructed in the face of complex conditions, and the conventional algorithm has limitations, such as the conventional iterative algorithm cannot flexibly select optimal parameters in each iteration; the existing direct inversion algorithm based on deep learning can only extract general information features of scatterers, and the reasoning process is not specific to physical characteristics of an electromagnetic model, so that certain resources are wasted, and the reconstruction effect is poor.
Disclosure of Invention
The technical problem to be solved by the invention is that the electromagnetic inversion method in the prior art is too complicated, the non-uniform distribution condition of the scatterers is ignored without aiming at the physical characteristics of an electromagnetic model, and the reconstruction effect is poor.
The invention solves the technical problems by the following technical means: the multi-mapping electromagnetic inversion method based on the CycleGAN comprises the following steps:
step one, constructing a two-dimensional electromagnetic inverse scattering model;
step two, based on the constructed two-dimensional electromagnetic backscattering model, collecting scattering data of different angles of a target area and carrying out data preprocessing;
thirdly, constructing a CycleGAN frame and constructing a corresponding loss function, wherein the CycleGAN frame receives scattering data of different angles of a target area;
training the CycleGAN frame, and obtaining a trained CycleGAN frame when the loss function converges and reaches a set number of rounds;
and fifthly, inputting scattering data of different angles of the target area acquired in real time into a trained CycleGAN frame to generate reconstruction contrast data of the target area.
Further, the first step includes:
the two-dimensional electromagnetic backscatter model comprises a scatterer, a plurality of emission sources and a plurality of receivers, wherein the scatterer is surrounded by the emission sources, the receivers are arranged at intervals of 360 degrees around the scatterer, and when an incident electromagnetic wave emitted by the emission sources encounters the scatterer to be scattered to form a scattered field, the receivers arranged at intervals of 360 degrees around the scatterer collect scattered data.
Still further, the first step further includes:
the whole forward process of electromagnetic backscatter is represented by two equations—a fringe field calculation and a total electric field calculation, the first equation, denoted as the data equation, describing the fringe field as the re-radiated field of the scatterer within the target domain, the first equation is as follows
Wherein E is s (r) is the fringe field at r, r' is the domain point and the source point, representing the position vector of the receiver and the source point of the target domain, k, respectively 0 Representing wavenumbers in free space, D obj The target domain of the scatterer is represented, G (r, r ') represents the green function, describing the value of the electromagnetic field generated by the source point at the receiving point, χ (r ') is a contrast function defined as χ (r ') =ε r (r') -1, wherein ε r (r') is the dielectric constant, the contrast function reflects the difference in electromagnetic properties between the object and the surrounding environment, E t (r ') represents the total electric field at r';
the second equation is a state equation, which describes the target domain D obj The field interaction between the medium scatterer blocks, the second equation is as follows
E t (r ') represents the total electric field at r', E in (r) represents incident electric field data.
Further, the second step includes:
m receivers are arranged around the scattering body at intervals of 90 degrees in 360 degrees, 4*M receivers are used for carrying out standardized processing on data received by the receivers, and magnitude differences among different data are eliminated; and performing singular value decomposition on the scattered data, selecting a eigenvector corresponding to the maximum singular value, and combining the low-frequency component with the eigenvector corresponding to the extracted maximum singular value to obtain the processed scattered data.
Further, the CycleGAN framework in the third step includes a forward derivation part and a reverse derivation part, wherein the forward derivation part includes 4 parallel generators G 1 ～G 4 And a discriminator D, generator G 1 ～G 4 Respectively receiving scattering data acquired by receivers corresponding to 4 angle intervals, fusing the scattering features through a coding structure, mapping the synthesized features to a potential space, generating reconstruction data in the potential space by a decoder, and judging the reconstruction data by a discriminator D; the reverse deriving part comprises a generator G and 4 discriminators D 1 ～D 4 The generator G receives the results of the forward deriving part of the decoder, is mapped back to the same potential space, reconstructs the scatter input by the 4 decoders of the reverse deriving part, and is then passed through the 4 discriminators D 1 ～D 4 And judging and reconstructing input.
Further, the loss function in the third step includes a countermeasure loss, and the countermeasure loss of the forward derivation section is
Where t represents the target domain data (contrast function), s represents the source domain data (scattering data),representing mathematical expectations following a target domain data distribution, D T (t) represents the discrimination result of the target domain data t,/->Representing mathematical expectations obeying the source domain data distribution, f S→T (s) represents a mapping of source domain data s converted from a source domain to a target domain;
the countermeasures of the reverse deriving part are that
Wherein i represents the i-th reconstructed input data sequence number, s i Representing the i-th reconstructed input source data, n=4,representing mathematical expectations of obeying the 4-angle source domain data distribution, +.>Discrimination values representing 4 reconstructed input data, +.>Mathematical expectations indicative of compliance with the target domain data distribution, +.>Representing a mapping of the target domain data t converted from the target domain to the source domain.
Still further, the loss function in step three includes a potential consistency loss, the potential consistency loss of the forward derived portion being
Wherein,representing mathematical expectations obeying the source domain data distribution, h S→L (s) represents the mapping of source data s from source domain to potential domain, h T→L () Representing the mapping from the target domain to the potential domain 1 Representation calculation l 1 A norm;
the potential consistency loss of the reverse derivation part is that
Wherein f T→S (t) represents the mapping of the reconstructed data t from the target domain to the source domain.
Still further, the loss function in step three includes a cyclic consistency loss, the cyclic consistency loss of the forward derived portion being
Wherein f T→S () Representing the mapping of the target domain to the source domain, f S→T () Representing a mapping of a source domain to a target domain, s representing source domain data;
the cycle consistency loss of the reverse derivation part is that
Where t represents target domain data.
The invention also provides a CycleGAN-based multi-mapping electromagnetic inversion system, comprising:
the first model building module is used for building a two-dimensional electromagnetic backscatter model;
the data processing module is used for acquiring scattering data of different angles of the target area based on the constructed two-dimensional electromagnetic backscattering model and carrying out data preprocessing;
the second model building module is used for building a CycleGAN frame and building a corresponding loss function, and the CycleGAN frame receives scattering data of different angles of a target area;
the training module is used for training the CycleGAN frame, and when the loss function converges and reaches the set number of rounds, the trained CycleGAN frame is obtained;
and the image reconstruction module is used for inputting scattering data of different angles of the target area acquired in real time into the trained CycleGAN frame to generate contrast data of the target area.
Further, the first model building module is further configured to:
the two-dimensional electromagnetic backscatter model comprises a scatterer, a plurality of emission sources and a plurality of receivers, wherein the scatterer is surrounded by the emission sources, the receivers are arranged at intervals of 360 degrees around the scatterer, and when an incident electromagnetic wave emitted by the emission sources encounters the scatterer to be scattered to form a scattered field, the receivers arranged at intervals of 360 degrees around the scatterer collect scattered data.
Still further, the first model building module is further configured to:
the whole forward process of electromagnetic backscatter is represented by two equations—a fringe field calculation and a total electric field calculation, the first equation, denoted as the data equation, describing the fringe field as the re-radiated field of the scatterer within the target domain, the first equation is as follows
Wherein E is s (r) is the fringe field at r, r' is the domain point and the source point, representing the position vector of the receiver and the source point of the target domain, k, respectively 0 Representing wavenumbers in free space, D obj The target domain of the scatterer is represented, G (r, r ') represents the green function, describing the value of the electromagnetic field generated by the source point at the receiving point, χ (r ') is a contrast function defined as χ (r ') =ε r (r') -1, wherein ε r (r') is the dielectric constant, the contrast function reflects the difference in electromagnetic properties between the object and the surrounding environment, E t (r ') represents the total electric field at r';
the second equation is a state equation, which describes the target domain D obj The field interaction between the medium scatterer blocks, the second equation is as follows
E t (r ') represents the total electric field at r', E in (r) represents incident electric field data.
Still further, the data processing module is further configured to:
m receivers are arranged around the scattering body at intervals of 90 degrees in 360 degrees, 4*M receivers are used for carrying out standardized processing on data received by the receivers, and magnitude differences among different data are eliminated; and performing singular value decomposition on the scattered data, selecting a eigenvector corresponding to the maximum singular value, and combining the low-frequency component with the eigenvector corresponding to the extracted maximum singular value to obtain the processed scattered data.
Further, the CycleGAN framework in the second model building module includes a forward derivation part and a reverse derivation part, the forward derivation part includes 4 parallel generators G 1 ～G 4 And a discriminator D, generator G 1 ～G 4 Respectively receiving scattering data acquired by receivers corresponding to 4 angle intervals, fusing the scattering features through a coding structure, mapping the synthesized features to a potential space, generating reconstruction data in the potential space by a decoder, and judging the reconstruction data by a discriminator D; the reverse deriving part comprises a generator G and 4 discriminators D 1 ～D 4 The result of the decoder of the forward derivation of the generator G is mapped back to the same potential space, the scatter input is reconstructed by the 4 decoders of the reverse derivation, and then by the 4 discriminators D 1 ～D 4 And judging and reconstructing input.
Further, the loss function in the second model building block includes a counterloss, and the counterloss of the forward derivation section is
Where t represents target domain data, s represents source domain data,representing mathematical expectations following a target domain data distribution, D T (t) represents the discrimination result of the target domain data t,/->Representing mathematical expectations obeying the source domain data distribution, f S→T (s) represents a mapping of source domain data s converted from a source domain to a target domain;
the countermeasures of the reverse deriving part are that
Wherein i represents the i-th reconstructed input data sequence number, s i Representing the i-th reconstructed input source data, n=4,representing mathematical expectations of obeying the 4-angle source domain data distribution, +.>Discrimination values representing 4 reconstructed input data, +.>Mathematical expectations indicative of compliance with the target domain data distribution, +.>Representing a mapping of the target domain data t converted from the target domain to the source domain.
Further, the loss function in the second model building block includes a potential consistency loss, and the potential consistency loss of the forward derivation part is that
Wherein,representing mathematical expectations obeying the source domain data distribution, h S→L (s) represents the mapping of source data s from source domain to potential domain, h T→L () Representing the mapping from the target domain to the potential domain 1 Representation calculation l 1 A norm;
the potential consistency loss of the reverse derivation part is that
Wherein f T→S (t) represents the mapping of the reconstructed data t from the target domain to the source domain.
Still further, the loss function in the second model building block includes a cyclic consistency loss, and the cyclic consistency loss of the forward derivation section is
Wherein f T→S () Representing the mapping of the target domain to the source domain, f S→T () Representing a mapping of a source domain to a target domain, s representing source domain data;
the cycle consistency loss of the reverse derivation part is that
The invention has the advantages that:
(1) According to the invention, the CycleGAN frame is constructed and trained, so that the image reconstruction of the target area is carried out, the process is simple, a large amount of labor cost is not required, the CycleGAN frame receives scattering data of the target area at different angles, the scattering data are not globally smooth average values, the non-uniform distribution condition of the scattering bodies is considered, the scattering data are extracted more comprehensively aiming at the physical characteristics of an electromagnetic model, and the reconstruction effect is improved.
(2) According to the invention, the situation of uneven distribution of an actual scatterer is considered through multi-angle input, the reconstruction pathogenicity is solved by combining a characteristic vector corresponding to a low-frequency component and an extracted maximum singular value as processed scattering data through a dominant current method, the mapping relation between a scattered field and a dielectric constant is ensured by integrally circularly generating an antagonism network (namely a cyclegaN framework), the introduced potential domain concept solves the problems of the generality and the roughness of direct inversion, the unique electromagnetic characteristic of the scatterer in deduction is maintained, and the resource waste of irrelevant reasoning is avoided.
Drawings
FIG. 1 is a schematic diagram of a two-dimensional electromagnetic backscatter model in a CycleGAN-based multi-mapping electromagnetic inversion method according to an embodiment of the present invention;
fig. 2 is a schematic diagram of a CycleGAN frame in a CycleGAN-based multi-mapping electromagnetic inversion method according to an embodiment of the present invention.
Detailed Description
For the purpose of making the objects, technical solutions and advantages of the embodiments of the present invention more apparent, the technical solutions in the embodiments of the present invention will be clearly and completely described in the following in conjunction with the embodiments of the present invention, and it is apparent that the described embodiments are some embodiments of the present invention, but not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.
Example 1
The multi-mapping electromagnetic inversion method based on the CycleGAN comprises the following steps:
s1, constructing a two-dimensional electromagnetic inverse scattering model; the specific process is as follows:
as shown in fig. 1, the two-dimensional electromagnetic backscatter model includes a scatterer 1, a plurality of emission sources 2 and a plurality of receivers 3, the scatterer 1 is surrounded by the emission sources 2, the receivers 3 are arranged at 360-degree intervals around the scatterer 1, when an incident electromagnetic wave emitted by the emission sources 2 encounters the scatterer 1 and is scattered to form a scattered field, and the receivers 3 arranged at 360-degree intervals around the scatterer 1 collect scattered data. The dielectric constant can be deduced from the scattering data, and the dielectric constant effectively distinguishes the electromagnetic properties of different materials. The model mathematical expression is as follows:
the above equation describes the scattering effect of an object on an incident electromagnetic wave. The object is referred to as a diffuser 1, and each point in the object can be regarded as a secondary radiation source, the strength of the diffuse field of which depends on the electromagnetic properties of the point (indicated by the contrast function) and the total electric field at the point, the diffuse field being the additive effect of the fields generated by these secondary radiation sources at the location of the receiver 3. The total electric field at each point in the object domain is composed of two parts: the sum of the direct incident electric field and the contribution of all scattering points to this point is formulated as
Where r, r' denote the position vector of the receiver 3 and the source point of the target domain, E, respectively in (r) represents incident electric field data, k 0 Representing wavenumbers in free space, D obj The target domain representing the scatterer 1, G (r, r ') represents the gray function describing the value of the electromagnetic field generated by the source point at the receiving point, χ (r ') is a contrast function defined as χ (r ') =ε r (r') -1, wherein ε r (r') is the dielectric constant, the contrast function reflects the difference in electromagnetic properties between the object and the surrounding environment, E t And (r ') represents the total electric field at r'.
The final implementation target domain takes the form of discrete samples, and the total electric field and contrast function are also known as piecewise constants.
S2, based on the constructed two-dimensional electromagnetic backscatter model, acquiring the scatter data of different angles of the target area and carrying out data preprocessing; the specific process is as follows:
the scattering body 1 is characterized in that each receiver 3 is arranged at every 90 degrees around 360 degrees, 4 receivers 3 are uniformly divided into 4 parts to serve as multi-branch input, and the difference of the spatial distribution of the network learning is facilitated. Extracting effective features with dominant unknown quantity from 4-branch test data by using a dominant current algorithm, collecting conditions with unknown and non-unique solutions in the dominant current, specifically, carrying out standardized processing on data received by a receiver 3, and eliminating magnitude differences among different data; and performing singular value decomposition on the scattered data, selecting a eigenvector corresponding to the maximum singular value, and combining the low-frequency component with the eigenvector corresponding to the extracted maximum singular value to obtain the processed scattered data. The low frequency component typically contains global information about the overall structure and electromagnetic properties of the scatterer 1.
S3, constructing a cycleGAN frame and constructing a corresponding loss function, wherein the cycleGAN frame receives scattering data of different angles of a target area; the specific process is as follows:
as shown in fig. 2, the CycleGAN framework includes a forward derivation section including 4 parallel generators G, and a reverse derivation section 1 ～G 4 And a discriminator D, generator G 1 ～G 4 Receiving the scattering data collected by the receivers 3 corresponding to the 4 angle intervals respectively, fusing the scattering features through the coding structure, mapping the synthesized features to a potential space, generating reconstruction data in the potential space by a decoder, and judging the reconstruction data by a discriminator D; the reverse deriving part comprises a generator G and 4 discriminators D 1 ～D 4 The result of the decoder of the forward derivation of the generator G is mapped back to the same potential space, the scatter input is reconstructed by the 4 decoders of the reverse derivation, and then by the 4 discriminators D 1 ～D 4 And judging and reconstructing input. The asymmetric forward and reverse training mode ensures the cyclic consistency of network derivation, while the potential domain regards different inputs as different representations of the same concept, relieving the coarseness and generality of direct inversion. Positive directionThe codec network architecture is adopted for reverse derivation. S and T represent the source domain and the target domain, respectively, L represents the potential domain located in transit between the source domain and the target domain, and h is used for converting the source domain into the potential domain S→L Representing the transformation h of the potential domain into the target domain L→T The representation is f for the conversion of the source domain to the target domain as a whole S→T And (3) representing.
To constrain the fitting direction of the network, constructing an fight loss function, a cyclic consistency function, a potential consistency loss function, the fight loss of the forward derivation part being
Where t represents target domain data, s represents source domain data,representing mathematical expectations following a target domain data distribution, D T (t) represents the discrimination result of the target domain data t,/->Representing mathematical expectations obeying the source domain data distribution, f S→T (s) represents a mapping of source domain data s converted from a source domain to a target domain; the target discriminator cannot distinguish between the true reference contrast and the generated contrast by calculating the mathematical expectation E of the discrimination value.
The countermeasures of the reverse deriving part are that
Wherein i represents the i-th reconstructed input data sequence number, s i Representing the i-th reconstructed input source data, n=4,representing mathematical expectations of obeying the 4-angle source domain data distribution, +.>Discrimination values representing 4 reconstructed input data, +.>Mathematical expectations indicative of compliance with the target domain data distribution, +.>Representing a mapping of the target domain data t converted from the target domain to the source domain. The inverse derivation results in a plurality of mapping source data, so that the penalty is accumulated for a plurality of discrimination values.
Wherein,representing mathematical expectations obeying the source domain data distribution, h S→L (s) represents the mapping of source data s from source domain to potential domain, h T→L () Representing the mapping from the target domain to the potential domain 1 Representation calculation l 1 A norm;
the potential consistency loss of the reverse derivation part is that
Wherein f T→S (t) represents the mapping of the reconstructed data t from the target domain to the source domain.
The potential consistency loss of forward and reverse deductions ensures that the source domain and the target domain have the same potential space, namely the same electromagnetic property.
The cycle consistency loss of the forward derivation part is that
Wherein f T→S () Representing the mapping of the target domain to the source domain, f S→T () Representing a mapping of a source domain to a target domain, s representing source domain data;
the cycle consistency loss of the reverse derivation part is that
The forward and reverse derived cyclic consistency loss ensures that the input data can still maintain the scatterer 1 characteristics after cyclic mapping. Wherein forward deriving the loop consistency loss calculates a multi-source fusion feature.
S4, training the CycleGAN frame, and obtaining the trained CycleGAN frame when the loss function converges and reaches the set number of rounds; the specific process is as follows:
first in the 4-branch generator G 1 ～G 4 Respectively inputting the scatter receipts acquired by the 4-angle interval receiver 3, carrying out feature extraction and fusion, conveying the scatter receipts into a coding and decoding network to realize reconstruction, and performing supervision training by using the corresponding real contrast of the scatterer 1 at an output part. The iterative training process repeats the scatter domain to contrast domain and contrast domain to scatter domain conversion. And when the loss function converges and reaches the set number of rounds, obtaining a trained electromagnetic backscatter model. And importing trained network parameters, inputting processed multi-angle scattering data, and generating a clear high-contrast reconstructed image.
S5, inputting scattering data of different angles of the target area acquired in real time into a trained CycleGAN frame, and generating a reconstructed image of the target area.
Through the technical scheme, the multiple-input part adopts the data of the receivers 3 with different angles, the spatial difference is considered to extract the characteristics, and the problems of non-unique quantity of the scatterers 1, uneven pixel distribution and the like are solved. The data received at different angles adopts a dominant current scheme to extract effective information, singular value decomposition is carried out on a green function of the scattered data, and the most main scattering characteristics, namely dominant current components (the components reflect the main information in the scattered data and ignore secondary characteristics which lead to unstable solution) are identified and extracted. The dominant current and the low-frequency component are combined to supplement global information, so that the problem of electromagnetic inversion morbidity is solved. The network architecture adopts a CycleGAN mode, the input and output repeated mapping can effectively learn the reasoning route closest to the physical model, and the setting of the potential domain is helpful for the network to pay attention to the unique electromagnetic characteristics of the same scatterer 1, so that the waste of network parameter resources to reasoning of irrelevant information is avoided.
Example 2
Based on embodiment 1, embodiment 2 of the present invention further provides a CycleGAN-based multi-mapping electromagnetic inversion system, comprising:
the first model building module is used for building a two-dimensional electromagnetic backscatter model;
the data processing module is used for acquiring scattering data of different angles of the target area based on the constructed two-dimensional electromagnetic backscattering model and carrying out data preprocessing;
the second model building module is used for building a CycleGAN frame and building a corresponding loss function, and the CycleGAN frame receives scattering data of different angles of a target area;
the training module is used for training the CycleGAN frame, and when the loss function converges and reaches the set number of rounds, the trained CycleGAN frame is obtained;
and the image reconstruction module is used for inputting scattering data of different angles of the target area acquired in real time into the trained CycleGAN frame to generate contrast data of the target area.
Specifically, the first model building module is further configured to:
the two-dimensional electromagnetic backscatter model comprises a scatterer 1, a plurality of emission sources 2 and a plurality of receivers 3, wherein the scatterer 1 is surrounded by the emission sources 2, the receivers 3 are arranged at 360-degree intervals around the scatterer 1, when an incident electromagnetic wave emitted by the emission sources 2 encounters the scatterer 1 to be scattered to form a scattered field, and the receivers 3 arranged at 360-degree intervals around the scatterer 1 collect scattered data.
More specifically, the first model building module is further configured to:
the whole forward process of electromagnetic backscatter is represented by two equations—a fringe field calculation and a total electric field calculation, the first equation, denoted as the data equation, describing the fringe field as the re-radiated field of the scatterer within the target domain, the first equation is as follows
Wherein E is s (r) is the fringe field at r, r' is the domain point and the source point, representing the position vector of the receiver and the source point of the target domain, k, respectively 0 Representing wavenumbers in free space, D obj The target domain of the scatterer is represented, G (r, r ') represents the green function, describing the value of the electromagnetic field generated by the source point at the receiving point, χ (r ') is a contrast function defined as χ (r ') =ε r (r') -1, wherein ε r (r') is the dielectric constant, the contrast function reflects the difference in electromagnetic properties between the object and the surrounding environment, E t (r ') represents the total electric field at r';
the second equation is a state equation, which describes the target domain D obj The field interaction between the medium scatterer blocks, the second equation is as follows
E t (r ') represents the total electric field at r', E in (r) represents incident electric field data.
More specifically, the data processing module is further configured to:
m receivers 3 are arranged around the scattering body 1 at intervals of 90 degrees in 360 degrees, 4*M receivers 3 are arranged in total, data received by the receivers 3 are subjected to standardized processing, and magnitude differences among different data are eliminated; and performing singular value decomposition on the scattered data, selecting a eigenvector corresponding to the maximum singular value, and combining the low-frequency component with the eigenvector corresponding to the extracted maximum singular value to obtain the processed scattered data.
More specifically, the CycleGAN framework in the second model building block includes a forward derivation sectionAnd a reverse deriving section including 4 parallel generators G 1 ～G 4 And a discriminator D, generator G 1 ～G 4 Receiving the scattering data collected by the receivers 3 corresponding to the 4 angle intervals respectively, fusing the scattering features through the coding structure, mapping the synthesized features to a potential space, generating reconstruction data in the potential space by a decoder, and judging the reconstruction data by a discriminator D; the reverse deriving part comprises a generator G and 4 discriminators D 1 ～D 4 The result of the decoder of the forward derivation of the generator G is mapped back to the same potential space, the scatter input is reconstructed by the 4 decoders of the reverse derivation, and then by the 4 discriminators D 1 ～D 4 And judging and reconstructing input.
More specifically, the loss function in the second model building block includes a counterloss, and the counterloss of the forward derivation section is
Where t represents target domain data, s represents source domain data,representing mathematical expectations following a target domain data distribution, D T (t) represents the discrimination result of the target domain data t,/->Representing mathematical expectations obeying the source domain data distribution, f S→T (s) represents a mapping of source domain data s converted from a source domain to a target domain;
the countermeasures of the reverse deriving part are that
Wherein i represents the i-th reconstructed input data sequence number, s i Representing the i-th reconstructed input source data, n=4,representing mathematical expectations of obeying the 4-angle source domain data distribution, +.>Discrimination values representing 4 reconstructed input data, +.>Mathematical expectations indicative of compliance with the target domain data distribution, +.>Representing a mapping of the target domain data t converted from the target domain to the source domain.
More specifically, the loss function in the second model building block includes a potential consistency loss, and the potential consistency loss of the forward derivation section is that
Wherein,representing mathematical expectations obeying the source domain data distribution, h S→L (s) represents the mapping of source data s from source domain to potential domain, h T→L () Representing the mapping from the target domain to the potential domain 1 Representation calculation l 1 A norm;
the potential consistency loss of the reverse derivation part is that
Wherein f T→S (t) represents the mapping of the reconstructed data t from the target domain to the source domain.
More specifically, the loss function in the second model building module includes a cyclic consistency loss, and the cyclic consistency loss of the forward derivation part is that
Wherein f T→S () Representing the mapping of the target domain to the source domain, f S→T () Representing a mapping of a source domain to a target domain, s representing source domain data;
the cycle consistency loss of the reverse derivation part is that
The above embodiments are only for illustrating the technical solution of the present invention, and are not limiting; although the invention has been described in detail with reference to the foregoing embodiments, it will be understood by those of ordinary skill in the art that: the technical scheme described in the foregoing embodiments can be modified or some technical features thereof can be replaced by equivalents; such modifications and substitutions do not depart from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims (10)
1. The multi-mapping electromagnetic inversion method based on the CycleGAN is characterized by comprising the following steps of:
step one, constructing a two-dimensional electromagnetic inverse scattering model;
step two, based on the constructed two-dimensional electromagnetic backscattering model, collecting scattering data of different angles of a target area and carrying out data preprocessing;
thirdly, constructing a CycleGAN frame and constructing a corresponding loss function, wherein the CycleGAN frame receives scattering data of different angles of a target area;
training the CycleGAN frame, and obtaining a trained CycleGAN frame when the loss function converges and reaches a set number of rounds;
and fifthly, inputting scattering data of different angles of the target area acquired in real time into a trained CycleGAN frame to generate comparison data of the target area.
2. The CycleGAN-based multi-map electromagnetic inversion method of claim 1, wherein step one comprises:
the two-dimensional electromagnetic backscatter model comprises a scatterer, a plurality of emission sources and a plurality of receivers, wherein the scatterer is surrounded by the emission sources, the receivers are arranged at intervals of 360 degrees around the scatterer, and when an incident electromagnetic wave emitted by the emission sources encounters the scatterer to be scattered to form a scattered field, the receivers arranged at intervals of 360 degrees around the scatterer collect scattered data.
3. The CycleGAN-based multi-map electromagnetic inversion method of claim 2 wherein step one further comprises:
the whole forward process of electromagnetic backscatter is represented by two equations—a fringe field calculation and a total electric field calculation, the first equation, denoted as the data equation, describing the fringe field as the re-radiated field of the scatterer within the target domain, the first equation is as follows
Wherein E is s (r) is the fringe field at r, r' is the domain point and the source point, representing the position vector of the receiver and the source point of the target domain, k, respectively 0 Representing wavenumbers in free space, D obj The target domain of the scatterer is represented, G (r, r ') represents the green function, describing the value of the electromagnetic field generated by the source point at the receiving point, χ (r ') is a contrast function defined as χ (r ') =ε r (r') -1, wherein ε r (r') is the dielectric constant, the contrast function reflects the difference in electromagnetic properties between the object and the surrounding environment, E t (r ') represents the total electric field at r';
the second equation is a state equation, which describes the target domain D obj Field interaction between the diffuser blocks, the secondThe process is as follows
E t (r ') represents the total electric field at r', E in (r) represents incident electric field data.
4. The CycleGAN-based multi-map electromagnetic inversion method of claim 2, wherein the second step comprises:
m receivers are arranged around the scattering body at intervals of 90 degrees in 360 degrees, 4*M receivers are used for carrying out standardized processing on data received by the receivers, and magnitude differences among different data are eliminated; and performing singular value decomposition on the scattered data, selecting a eigenvector corresponding to the maximum singular value, and combining the low-frequency component with the eigenvector corresponding to the extracted maximum singular value to obtain the processed scattered data.
5. The CycleGAN-based multi-mapping electromagnetic inversion method according to claim 4, wherein the CycleGAN framework in step three comprises a forward derivation section and a reverse derivation section, the forward derivation section comprising 4 parallel generators G 1 ～G 4 And a discriminator D, generator G 1 ～G 4 Respectively receiving scattering data acquired by receivers corresponding to 4 angle intervals, fusing the scattering features through a coding structure, mapping the synthesized features to a potential space, generating reconstruction data in the potential space by a decoder, and judging the reconstruction data by a discriminator D; the reverse deriving part comprises a generator G and 4 discriminators D 1 ～D 4 The result of the decoder of the forward derivation of the generator G is mapped back to the same potential space, the scatter input is reconstructed by the 4 decoders of the reverse derivation, and then by the 4 discriminators D 1 ～D 4 And judging and reconstructing input.
6. The CycleGAN-based multi-map electromagnetic inversion method of claim 5 wherein the loss function in step three comprises a fight loss, the fight loss of the forward derived portion being
Where t represents target domain data, s represents source domain data,representing mathematical expectations following a target domain data distribution, D T (t) represents the discrimination result of the target domain data t,/->Representing mathematical expectations obeying the source domain data distribution, f S→T (s) represents a mapping of source domain data s converted from a source domain to a target domain;
the countermeasures of the reverse deriving part are that
Wherein i represents the i-th reconstructed input data sequence number, s i Representing the i-th reconstructed input source data, n=4,representing mathematical expectations of obeying the 4-angle source domain data distribution, +.>Discrimination values representing 4 reconstructed input data, +.>Mathematical expectations indicative of compliance with the target domain data distribution, +.>Representing a mapping of the target domain data t converted from the target domain to the source domain.
7. The CycleGAN-based multi-map electromagnetic inversion method of claim 6 wherein the loss function in step three comprises a potential consistency loss, the potential consistency loss of the forward derived portion being
Wherein,representing mathematical expectations obeying the source domain data distribution, h S→L (s) represents the mapping of source data s from source domain to potential domain, h T→L () Representing the mapping from the target domain to the potential domain 1 Representation calculation l 1 A norm;
the potential consistency loss of the reverse derivation part is that
Wherein f T→S (t) represents the mapping of the reconstructed data t from the target domain to the source domain.
8. The CycleGAN-based multi-map electromagnetic inversion method of claim 7 wherein the step three loss function comprises a cyclic uniformity loss, the cyclic uniformity loss of the forward derived portion being
Wherein f T→S () Representing objectsMapping of domains to source domains, f S→T () Representing a mapping of a source domain to a target domain, s representing source domain data;
the cycle consistency loss of the reverse derivation part is that
9. The multi-mapping electromagnetic inversion system based on the CycleGAN is characterized by comprising:
the first model building module is used for building a two-dimensional electromagnetic backscatter model;
the data processing module is used for acquiring scattering data of different angles of the target area based on the constructed two-dimensional electromagnetic backscattering model and carrying out data preprocessing;
the second model building module is used for building a CycleGAN frame and building a corresponding loss function, and the CycleGAN frame receives scattering data of different angles of a target area;
the training module is used for training the CycleGAN frame, and when the loss function converges and reaches the set number of rounds, the trained CycleGAN frame is obtained;
and the image reconstruction module is used for inputting scattering data of different angles of the target area acquired in real time into the trained CycleGAN frame to generate contrast data of the target area.
10. The CycleGAN-based multi-mapping electromagnetic inversion system of claim 9 wherein the first model building module is further configured to:
the two-dimensional electromagnetic backscatter model comprises a scatterer, a plurality of emission sources and a plurality of receivers, wherein the scatterer is surrounded by the emission sources, the receivers are arranged at intervals of 360 degrees around the scatterer, and when an incident electromagnetic wave emitted by the emission sources encounters the scatterer to be scattered to form a scattered field, the receivers arranged at intervals of 360 degrees around the scatterer collect scattered data.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311841903.8A CN117807876B (en) | 2023-12-27 | 2023-12-27 | CycleGAN-based multi-mapping electromagnetic inversion method and system |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311841903.8A CN117807876B (en) | 2023-12-27 | 2023-12-27 | CycleGAN-based multi-mapping electromagnetic inversion method and system |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN117807876A true CN117807876A (en) | 2024-04-02 |
| CN117807876B CN117807876B (en) | 2025-06-03 |
Family
ID=90429279
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202311841903.8A Active CN117807876B (en) | 2023-12-27 | 2023-12-27 | CycleGAN-based multi-mapping electromagnetic inversion method and system |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN117807876B (en) |
Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20100250141A1 (en) * | 2009-03-13 | 2010-09-30 | Michael Tompkins | Subsurface electrical conductivity contrast reconstruction using narrow-band electromagnetic data and fullwave imaging |
| KR20130010156A (en) * | 2011-07-18 | 2013-01-28 | (주)뮤트로닉스 | Analysis method of microwave inverse scattering using ultrasonic wave |
| WO2015022475A1 (en) * | 2013-08-14 | 2015-02-19 | Toshiba Research Europe Limited | A system and method for analysing data from a microwave inverse scattering apparatus |
| 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 |
| US20220028038A1 (en) * | 2020-07-22 | 2022-01-27 | Zhejiang University | Method and system for directed transfer of cross-domain data based on high-resolution remote sensing images |
| CN114255293A (en) * | 2021-12-22 | 2022-03-29 | 杭州电子科技大学 | Rapid imaging method for solving highly nonlinear inverse scattering problem based on deep learning |
| CN114298905A (en) * | 2021-12-22 | 2022-04-08 | 杭州电子科技大学 | Real-time imaging method for solving highly nonlinear inverse scattering problem based on deep learning |
| WO2022200931A1 (en) * | 2021-03-22 | 2022-09-29 | Agco Corporation | Deep learning for electromagnetic imaging of stored commodities |
-
2023
- 2023-12-27 CN CN202311841903.8A patent/CN117807876B/en active Active
Patent Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20100250141A1 (en) * | 2009-03-13 | 2010-09-30 | Michael Tompkins | Subsurface electrical conductivity contrast reconstruction using narrow-band electromagnetic data and fullwave imaging |
| KR20130010156A (en) * | 2011-07-18 | 2013-01-28 | (주)뮤트로닉스 | Analysis method of microwave inverse scattering using ultrasonic wave |
| WO2015022475A1 (en) * | 2013-08-14 | 2015-02-19 | Toshiba Research Europe Limited | A system and method for analysing data from a microwave inverse scattering apparatus |
| US20220028038A1 (en) * | 2020-07-22 | 2022-01-27 | Zhejiang University | Method and system for directed transfer of cross-domain data based on high-resolution remote sensing images |
| CN111999731A (en) * | 2020-08-26 | 2020-11-27 | 合肥工业大学 | Electromagnetic backscattering imaging method based on perception generation countermeasure network |
| WO2022200931A1 (en) * | 2021-03-22 | 2022-09-29 | Agco Corporation | Deep learning for electromagnetic imaging of stored commodities |
| 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 |
| CN114255293A (en) * | 2021-12-22 | 2022-03-29 | 杭州电子科技大学 | Rapid imaging method for solving highly nonlinear inverse scattering problem based on deep learning |
| CN114298905A (en) * | 2021-12-22 | 2022-04-08 | 杭州电子科技大学 | Real-time imaging method for solving highly nonlinear inverse scattering problem based on deep learning |
Non-Patent Citations (2)
| Title |
|---|
| CHE LIU ET AL.: "Towards Intelligent Electromagnetic Inverse Scattering Using Deep Learning Techniques and Information Metasurfaces", IEEE JOURNAL OF MICROWAVES, vol. 3, no. 1, 31 January 2023 (2023-01-31), pages 509 - 522 * |
| 甘泳机;尹成友;范启蒙;李安琪;: "电大目标的对比度源逆散射成像方法研究", 微波学报, no. 04, 8 July 2020 (2020-07-08) * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN117807876B (en) | 2025-06-03 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Wei et al. | Uncertainty quantification in inverse scattering problems with Bayesian convolutional neural networks | |
| CN109063741B (en) | Energy spectrum analysis method based on Hilbert curve transformation and deep learning | |
| CN115661069B (en) | A hyperspectral anomaly detection method and computer device | |
| Zhang et al. | Revisiting rotation averaging: Uncertainties and robust losses | |
| CN109215025B (en) | Infrared weak and small target detection method based on non-convex rank approach minimization | |
| Nakashima et al. | Learning to drop points for lidar scan synthesis | |
| Song et al. | Inversion of rough surface parameters from SAR images using simulation-trained convolutional neural networks | |
| Liu et al. | Group sparse representation based dictionary learning for SAR image despeckling | |
| Zhao et al. | Neutron image denoising and deblurring based on generative adversarial networks | |
| Song et al. | GM-APD lidar single-source data self-guided: Obtaining high-resolution depth map | |
| Xing et al. | VBIM-Net: Variational Born iterative network for inverse scattering problems | |
| CN117807876B (en) | CycleGAN-based multi-mapping electromagnetic inversion method and system | |
| CN112507826A (en) | End-to-end ecological change monitoring method, terminal, computer equipment and medium | |
| Tahraoui et al. | Land change detection in Sentinel-2 images using IR-MAD and deep neural network | |
| CN120314941A (en) | Video SAR depth imaging method and system based on interpretable sparse tensor recovery | |
| CN117710268B (en) | EMIS reconstruction method and system based on composite deep learning | |
| Xu et al. | Towards Understanding the Benefits of Neural Network Parameterizations in Geophysical Inversions: A Study With Neural Fields | |
| CN118070648A (en) | An electromagnetic inverse scattering method, device and storage medium combining SOM and complex phase | |
| Guo et al. | Physics embedded machine learning for electromagnetic data imaging | |
| Guo et al. | Lidar reflective tomography of the target under incomplete view state | |
| Wu et al. | A 3D reconstruction of terahertz images based on the FCTMVSNet algorithm | |
| Li et al. | GEO spatial correlation SAR imaging framework using MBO2D-SC algorithm for sea surface wave imaging under prolonged synthetic aperture time | |
| Carpio | Seeing the invisible: Digital holography | |
| Cao et al. | Infrared small target detection based on the tensor model | |
| Tang et al. | Shape reconstruction of acoustic obstacle with linear sampling method and neural network |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |