阅读说明：本技术 一种基于卷积神经网络的散射场相位恢复方法 (Scattered field phase recovery method based on convolutional neural network ) 是由 吴亮 徐魁文 张璐 马振超 于 2020-05-28 设计创作，主要内容包括：本发明公开了一种基于卷积神经网络的散射场相位恢复方法。在电磁逆散射成像领域,全波数据反演算法需要用到全波数据,然而全波数据的实际测量相当困难；无相位反演算法仅需要使用无相位总场数据,无相位总场数据的实际测量要容易很多,但是无相位反演算法具有更高的非线性度,计算较为困难。本发明利用无相位总场恢复出散射场,对处理散射场获取困难的问题有一定的帮助。本发明所使用的CNN架构是U-net。U-net是一种U形结构的CNN,它通过先卷积然后反卷积的形式使得输入与输出的维度能保持一致。通过仿真数据和实测数据的测试,验证了该方法的可行性。(The invention discloses a scattered field phase recovery method based on a convolutional neural network. In the field of electromagnetic backscatter imaging, a full-wave data inversion algorithm needs to be used for full-wave data, but actual measurement of the full-wave data is quite difficult; the phase-free inversion algorithm only needs to use phase-free total field data, the actual measurement of the phase-free total field data is easy, but the phase-free inversion algorithm has higher nonlinearity and is difficult to calculate. The invention recovers the scattered field by using the phase-free total field, and has certain help for processing the problem of difficult acquisition of the scattered field. The CNN architecture used by the present invention is U-net. U-net is a CNN with a U-shaped structure, which enables the input and output dimensions to be consistent by means of convolution and then deconvolution. The feasibility of the method is verified through the test of simulation data and measured data.)

1. A scattered field phase recovery method based on a convolutional neural network is characterized by comprising the following steps: recovering the measured phase-free total field data into scattered field data through the trained CNN, generating a data set through theoretical calculation, establishing a relation between the phase-free total field data and the scattered field, and then establishing a U-net neural network; and after the U-net neural network training is finished, observing whether the loss function is converged.

2. The method of phase recovery of a fringe field based on a convolutional neural network as claimed in claim 1, wherein: the relationship between the phase-free total field data and the scattered field is established by:

assuming that an unknown object exists in the target region D in the free space background,in the method, the total number of the meshes of the target area after being divided is assumed to be N, and the position of each mesh is r_nN is 1,2,3, N; and outside this area, a transmitting antenna and a receiving antenna are installed, the position of the transmitting antenna being indicated by r_p,p＝1,2,…,N_iThe position of the receiving antenna is denoted as r_q,q＝1,2,…,N_rObtaining N_i×N_rScatter field data; the scattering body is composed of a non-magnetic and isotropic inhomogeneous medium, and a scattering field E is applied^sca(r_q) Solving for the dielectric constant distribution (r) in the target region D_n)：

Obtaining a total field integral equation by a Lippmann-Schwigger electric field integral equation:

wherein E^inc(r) represents the incident field at r located inside the region; chi (r) -₀)/₀As a function of contrast within the region; k is a radical of₀Represents the wave vector in free space; green function of green

integral equation of scattered field:

wherein E^sca(r_q) Is shown at r_qInformation of the scattered field received by the receiving antenna,the contrast source is the product of contrast and total field, defined as:

I(r)＝χ(r)E^tot(r) (3)

discretizing equations (1) - (3):

wherein ⊙ denotes the corresponding element multiplication, Green functionAs a discrete Green function G (r)_qR '), an integral operator of r'),as a discrete Green function G (r)_nR') integral operator; the induced current calculated by the formula (6) is substituted into the formula (5), and the required scattered field data is calculated. The formula is as follows:

whereinRepresenting an identity matrix.

The definition of the total field without phase is as follows:

wherein

3. The method of phase recovery of a fringe field based on a convolutional neural network as claimed in claim 2, wherein: the U-net neural network is built as follows:

the left side of the U-net neural network is a contraction path which consists of convolution and pooling operations, and each layer on the left side is subjected to convolution operation twice; each convolution is processed by Batch Normalization (BN) and an activation function (ReLU), then the maximum pooling merging operation enters the next layer, and in each down-sampling step, the picture size is reduced by half, and the number of characteristic channels is doubled; an expansion path is arranged on the right side of the U-net neural network and used for recovering a matrix, each step comprises up-sampling of a characteristic diagram, then deconvolution is carried out, the number of characteristic channels is halved, and meanwhile, the size of the matrix is doubled and used for recovering the matrix to the original size; the signature graph obtained by each convolutional layer of the U-net neural network is connected to the corresponding upsampling layer.

4. The method of phase recovery of a fringe field based on a convolutional neural network as claimed in claim 1, wherein: Mean-Square Error (MSE) is selected as the loss function of the U-net neural network.

Technical Field

The invention belongs to the technical field of phase recovery of electromagnetic field data, and particularly relates to a method for recovering a phase-free total field into a scattered field by utilizing a neural network.

Background

The research object of the electromagnetic field backscattering problem is an unknown object located in a certain area, field data outside the area are obtained through a certain method, and the information of the unknown object is inverted through an optimization algorithm.

In the past decades, there has been increasing interest in electromagnetic field backscattering techniques in order to obtain the geometric and physical properties of unknown objects in difficult to access areas. In practice, electromagnetic backscatter imaging techniques have wide application, such as non-destructive testing, petroleum surveying, geophysical, biomedical imaging, and subsurface imaging. Although the electromagnetic backscattering problem has a wide application prospect and research in the related field has been advanced, the application of the electromagnetic backscattering technology still has many problems. Pathobiology and nonlinearity have long been two major challenges for electromagnetic backscatter imaging. In order to reconstruct unknown scatterers within a region in the most efficient and reliable way, researchers have developed many backscatter imaging algorithms. Examples of such methods include Born Iterative Method (BIM), modified Born Iterative Method (DBIM), Contrast Source Inversion Method (CSI), Subspace Optimization algorithm (SOM), and Two-fold Subspace Method (TSOM).

A large number of experiments prove that the method is very effective and plays a great role in the field of electromagnetic field backscatter imaging. These methods all use fringe field data (full wave data) with amplitude and phase information to reconstruct the geometry and physics of the scatterers. Accurate measurement of the fringe field amplitude and phase has also been a difficulty. Phase measurement usually involves considerable difficulties and non-negligible hardware costs, and even when the frequency is high to some extent, phase information cannot be directly measured. Thus, the requirement to reconstruct scatterers from the scattered field data limits the application of this technique to many practical scenarios. To overcome this drawback, researchers have proposed methods for reconstructing scatterers using phase-free total field data. Such as Newton's Method, phase-free Subspace Optimization Method (PD-SOM), phase-free Data contrast source Inversion (PD-CSI), and phase-free Regularized contrast source Inversion (PD-MRCSI). But the phase-free algorithm has the disadvantages that the noise resistance is poorer and the nonlinearity is higher. Combining the advantages and disadvantages of the full-wave data inversion algorithm and the phase-free inversion algorithm, it is desirable to obtain the magnitude and phase information of the scattered field without direct measurement, which can be achieved by phase recovery techniques.

Convolutional Neural Networks (CNNs) are known to have strong nonlinear fitting capabilities and are increasingly used in various research fields. Similarly, CNNs can be used with phase recovery, where the input to the network is measured phase-free data and the output of the network is data with amplitude and phase information.

Disclosure of Invention

The invention aims to provide a scattering field phase recovery method based on CNN (continuous noise network) aiming at the advantages and disadvantages of a full-wave data inversion algorithm and a phase-free inversion algorithm. The CNN architecture used by the present invention is U-net. U-net is a CNN of U-shaped structure that enables input and output dimensions to be kept consistent by means of convolution followed by deconvolution, which is widely used in image processing. Firstly, CNN needs to be trained, then the phase-free total field data is used as the input of the network, and the required scattered field with amplitude and phase information is predicted and output through multilayer convolution and deconvolution.

The invention provides a method for recovering measured phase-free total field data into scattered field data through a trained CNN. Because the cost for acquiring real data is high, a data set is generated through theoretical calculation, the relation between the phase-free total field data and the scattered field is established, and then the U-net is established. The specific technical scheme is as follows:

a scattered field phase recovery method based on a convolutional neural network comprises the steps of recovering measured phase-free total field data into scattered field data through a trained CNN, generating a data set through theoretical calculation, establishing a relation between the phase-free total field data and a scattered field, and then establishing a U-net neural network; and after the U-net neural network training is finished, observing whether the loss function is converged.

Further, the relationship between the phase-free total field data and the scattered field is established by:

assuming that an unknown object exists in a target area under a free space backgroundIn the method, the total number of the meshes of the target area after being divided is assumed to be N, and the position of each mesh is r_nN is 1,2,3, N; and outside this area, a transmitting antenna and a receiving antenna are installed, the position of the transmitting antenna being indicated by r_p,p＝1,2,...,N_iThe position of the receiving antenna is denoted as r_q,q＝1,2,...,N_rObtaining N_i×N_rScatter field data; the scattering body is composed of a non-magnetic and isotropic inhomogeneous medium, and a scattering field E is applied^sca(r_q) Solving for the dielectric constant distribution (r) in the target region D_n)：

Obtaining a total field integral equation by a Lippmann-Schwigger electric field integral equation:

wherein E^inc(r) represents the incident field at r located inside the region; chi (r) -₀)/₀As a function of contrast within the region; k is a radical of₀Represents the wave vector in free space; green function of greenRepresenting the field generated by a point source located in space r' to a point r in the space surrounding it,representing a first class of zero-order hankel functions;

integral equation of scattered field:

wherein E^sca(r_q) Is shown at r_qInformation of the scattered field received by the receiving antenna,

the contrast source is the product of contrast and total field, defined as:

I(r)＝χ(r)E^tot(r) (3)

discretizing equations (1) - (3):

wherein ⊙ denotes the corresponding element multiplication, Green function

As a discrete Green function G (r)_qR '), an integral operator of r'),as a discrete Green function G (r)_nR') integral operator; the induced current calculated by the formula (6) is substituted into the formula (5), and the required scattered field data is calculated. The formula is as follows:

whereinRepresenting an identity matrix.

The definition of the total field without phase is as follows:

wherein

Representing the field information received by the receiving antenna in the absence of unknown scatterers.

Further, the building of the U-net neural network specifically includes:

the left side of the U-net neural network is a contraction path which consists of convolution and pooling operations, and each layer on the left side is subjected to convolution operation twice; each convolution is processed by Batch Normalization (BN) and an activation function (ReLU), then the maximum pooling merging operation enters the next layer, and in each down-sampling step, the picture size is reduced by half, and the number of characteristic channels is doubled; an expansion path is arranged on the right side of the U-net neural network and used for recovering a matrix, each step comprises up-sampling of a characteristic diagram, then deconvolution is carried out, the number of characteristic channels is halved, and meanwhile, the size of the matrix is doubled and used for recovering the matrix to the original size; the signature graph obtained by each convolutional layer of the U-net neural network is connected to the corresponding upsampling layer.

Further, Mean-Square Error (MSE) is selected as a loss function of the U-net neural network.

The invention has the beneficial effects that:

the invention mainly aims at the advantages and disadvantages of a full-wave data inversion algorithm and a phase-free inversion algorithm, and provides a scattering field phase recovery method based on CNN. In the field of electromagnetic backscatter imaging, a full-wave data inversion algorithm has more scatterer information (amplitude and phase information), and is lower in nonlinearity than a phase-free inversion algorithm, so that the calculation is easier. However, full-wave data inversion algorithms require the use of fringe field data with amplitude and phase information, and in practice accurate measurement of the fringe field is quite difficult. The invention can obtain the scattered field information by avoiding the method of directly measuring the scattered field, and the phase recovery result is more satisfactory, which has great significance for the subsequent research.

Drawings

FIG. 1 is a diagram of an experimental setup for measuring field information;

FIG. 2 is a diagram of a U-net structure;

FIG. 3 is a partial training sample presentation diagram for training a neural network;

FIG. 4 is a convergence diagram of a training neural network;

FIGS. 5a and 5b are the real scatterfield map and the resulting map of the phase-free total field restored to the scatterfield using U-net, respectively;

fig. 6a and 6b are graphs of experimental results of measured data of a real scattered field and a recovered scattered field, respectively.

Detailed Description

The invention will be further explained with reference to the drawings.

The invention provides a method for recovering measured phase-free total field data into scattered field data through a trained CNN. Because the cost for acquiring real data is high, a data set is generated through theoretical calculation, the relation between the phase-free total field data and the scattered field is established, and then the U-net is established. The main contents are as follows:

firstly, obtaining phase-free total field data and scattered field data through electromagnetic field correlation knowledge:

assuming that an unknown object exists in a target area under a free space background

And (4) the following steps. In general, the solution of the back scattering problem requires that the target region is subdivided according to a certain rule. Assuming that the total number of the meshes of the divided target area is N, and the position of each mesh is r_nN is 1,2,3. And outside this area, a transmitting antenna and a receiving antenna are installed (the transmitting antenna position is denoted by r)_p,p＝1,2,…,N_iReception dayLine position is denoted r_q,q＝1,2,...,N_r) Applying this arrangement N is obtained_i×N_rA scatter field data. Given that the scatterer is composed of a non-magnetic and isotropic inhomogeneous medium, the problem of backscatter imaging can be attributed to the application of a scattered field E^sca(r_q) Solving for the dielectric constant distribution (r) in the target region D_n)。

The total field integral equation can be obtained by the Lippmann-Schwigger electric field integral equation:

wherein E^inc(r) represents the incident field at r located inside the region; chi (r) -₀)/₀As a function of contrast within the region; k is a radical of₀Represents the wave vector in free space; green function of green

Representing the field generated by a point source located in space r' to a point r in the space surrounding it,representing a zero order first class hank function.

Integral equation of scattered field:

wherein E^sca(r_q) Is shown at r_qInformation of the scattered field received by the receiving antenna;

the contrast source is the product of contrast and total field, defined as:

I(r)＝χ(r)E^tot(r) (3)

discretizing equations (1) - (3):

wherein ⊙ denotes the corresponding element multiplication, Green functionAs a discrete Green function G (r)_qR '), an integral operator of r'),as a discrete Green function G (r)_nR') of the same. The induced current calculated by equation (6) is substituted into equation (5) to calculate the required scattered field data. The formula is as follows:

wherein

Representing an identity matrix.

The definition of the total field without phase is as follows:

wherein

Representing the field information received by the receiving antenna in the absence of unknown scatterers.

Secondly, building a U-net neural network:

the U-net is typically characterized by a "U" -shaped symmetrical network (as shown in FIG. 2). On the left side of the U-net is a shrink path, which consists mainly of convolution and pooling operations. Each layer on the left is subjected to two convolution operations. Each convolution is processed by Batch Normalization (BN) and activation function (ReLU), and the BN can effectively accelerate deep network training. Then the max-pooling merge operation proceeds to the next layer, and at each down-sampling step, the picture size is reduced by half while the number of feature channels is doubled. To the right of the U-net is an extended path. The extended path is mainly used to recover the matrix, each step of which consists of upsampling of the eigen-map, followed by deconvolution, halving the number of eigen-channels, while doubling the matrix size for recovering the matrix to the original size. The signature obtained from each convolutional layer of the U-net will be connected to the corresponding upsampling layer so that more information of the original matrix can be retained in subsequent calculations. The U-net architecture was originally used for medical segmentation, which can predict the value of each pixel well.

Derived from previous calculations:

scattered field calculation formula:

the phase-free total field calculation formula:

5000 real scatterers are randomly generated, and are randomly acquired from MNIST types of a common training set (as shown in FIG. 3), and the relative dielectric constant of the scatterers randomly changes from 1.1 to 1.5. Then, the scattered field and the total field without phase of all scatterers are calculated by formula (7) and formula (8), and the result is used as a training set, wherein 1000 are used as a test set. It should be noted that the scattered field is a complex matrix, and the total field without phase is a real matrix, so the real part and the imaginary part of the scattered field need to be separated, and a dimension is added to store the information. The phase-free total field is likewise increased by one dimension, the imaginary part of which is all 0. The dimensions of the processed scattered field and the phase-free total field are kept consistent, wherein the scattered field serves as a label.

The size formula of the matrix after convolution operation is as follows:

O＝(W-F+2P)/S+1 (9)

where O denotes the output matrix size, W denotes the input matrix size, F denotes the convolution kernel (Filter) size, P denotes the Padding value (Padding) size, and S denotes the step size. The matrix is reduced in size by a factor of two after the maximum pooling operation. When the convolution kernel size is 1 × 1, the matrix size is unchanged. Whereas deconvolution can restore the matrix to the size before convolution, upsampling can restore the matrix to the size before pooling. Therefore, the left side and the right side of the U-net are basically in a symmetrical structure, and the left side has the pooling times, and the right side has the corresponding up-sampling times.

Mean-Square Error (MSE) is selected as a loss function of the network, and part of parameters are set as follows:

the number of network layers: 3

Learning rate: 0.01

Batch size (batch size): 32

Training times are as follows: 200

After the neural network training is completed, it is necessary to observe whether the loss function converges. It is observed from fig. 4 that the convergence curve of the network tends to be flat after 200 times of training, and it is not significant to continue to increase the number of training times, so that 200 times of training are enough. After training is finished, a plurality of test samples are generated to test the network, the phase-free total field is used as the input of the network, and the output of the network is the recovered scattered field. Partial results are shown in fig. 5a and 5b, where the true scattered field has a high similarity to the scattered field recovered from the phase-free total field.