Expanded random irradiation phase-free microwave imaging method for generative adversarial network augmentation

Abstract

The invention discloses an expanded random irradiation phase-free microwave imaging method for generative adversarial network augmentation. An echo measurement value is obtained according to scattered echo processing, a real backscattering coefficient is obtained according to target scatterer processing, and a microwave imaging data set is constructed by matching with a fixed observation matrix; inputting the microwave imaging data set into a random irradiation microwave imaging network to obtain a trained random irradiation microwave imaging network; and inputting an echo measurement value of a to-be-measured target and the observation matrix into the trained network to obtain a backscattering coefficient of the target scatterer, and constructing a target scatterer image. According to the invention, the initial value generated by the generative adversarial network with physical characteristics is obviously superior to the initial value generated by the traditional spectrum method in the aspect of iteration stability, and benefits from the high-dimensional input after feature enhancement and the screening capability of the gating module in the iteration process; and the gated recurrent neural network can be quickly and accurately solved by virtue of a small number of iterations, so that high-quality image reconstruction is realized.

Description

Technical Field

[0001] The present invention belongs to the field of electromagnetic inverse scattering, and in particular relates to an expanded random irradiation phaseless microwave imaging method augmented by a generative adversarial network. Background Art

[0002] Random irradiation microwave imaging is a key research area in the field of inverse scattering. Its core mechanism is to invert the characteristic information of target scatterers by combining random fields generated by spatially diverse excitations with scattered field information. This technology has broad application prospects in various fields, particularly in industrial nondestructive testing, such as internal defect identification in composite materials. Compared with traditional radar imaging and electromagnetic tomography, random irradiation microwave imaging offers advantages such as tunable imaging resolution, reduced system architecture complexity, and strong environmental robustness. Because electromagnetic waves have a certain degree of penetration and are non-destructive to the objects being detected, random irradiation microwave imaging can achieve nondestructive detection of targets, thereby ensuring their integrity and safety. Random irradiation microwave imaging can be applied to a variety of complex media and environments, such as metals, dielectrics, biological tissues, random media, and through-wall radar, demonstrating strong adaptability and flexibility. In recent years, artificial intelligence methods, represented by deep learning, have achieved breakthroughs in computational paradigms. Leveraging data processing capabilities, distributed computing architectures, and adaptive optimization theory, modern machine learning models have demonstrated superior nonlinear feature extraction capabilities and unprecedented computational efficiency. This breakthrough in intelligent technology not only reshapes the research paradigm in the fields of computer vision and speech recognition, but also spawns innovative solutions in the field of random irradiation electromagnetic imaging.

[0003] Given the difficulty of measuring phase information, utilizing phase-free signals offers greater practical value in engineering applications. When solving the inverse scattering problem, target reconstruction based on phase-free data can be transformed into a quadratic non-convex optimization problem. Classical solutions include the Riemann optimization method and the Gauss-Newton method. However, due to the reliance on manual feature extraction and low computational efficiency of traditional algorithms, actual inversion results are less than ideal. Summary of the Invention

[0004] To address the aforementioned issues in the prior art, the present invention provides an expanded random irradiation phaseless microwave imaging method augmented by a generative adversarial network. This method combines the advantages of traditional iterative algorithms for solving quadratic non-convex optimization problems. Using a generative adversarial network, the method performs phase recovery on echo measurements to obtain reliable initial values ​​for iteration. This initial value is then iterated using a gated recurrent neural network. This method can improve imaging efficiency and quality.

[0005] The technical solutions adopted in the present invention are as follows:

[0006] The present invention is used to obtain an image of a target scatterer, comprising the following steps:

[0007] S1. The random irradiation microwave imaging system collects scattered echoes from the target scatterer, processes the scattered echoes to obtain echo measurements and constructs an observation matrix. Furthermore, the true backscatter coefficients are obtained based on the target scatterer processing, and a microwave imaging dataset is constructed based on the echo measurements, the true backscatter coefficients, and the observation matrix.

[0008] S2. Construct a random irradiation microwave imaging network including a generative adversarial network and a gated recurrent neural network, input the microwave imaging dataset into the random irradiation microwave imaging network for training, and obtain a trained random irradiation microwave imaging network;

[0009] S3. Collect the scattered echoes of the target scatterer to be measured and process them to obtain the measured echo measurement value, input the measured echo measurement value and the observation matrix into the trained random irradiation microwave imaging network to obtain the backscattering coefficient of the target scatterer to be measured, and then construct the image of the target scatterer to be measured based on the backscattering coefficient of the target scatterer to be measured.

[0010] The echo measurement value is the amplitude value of the scattered echo.

[0011] The method adopts a random irradiation microwave imaging system, which includes a transmitting antenna array and a receiving antenna. The transmitting antenna array transmits electromagnetic waves with a preset frequency and random phase modulation to irradiate the target scatterer. The electromagnetic waves interact with the target scatterer and generate scattered echoes, which are received by the receiving antenna.

[0012] The S1 is specifically:

[0013] 1) Divide the preset imaging area into several regions, place the same scatterer unit in each region, and transmit electromagnetic waves with a preset frequency f1 and different random phase modulation to the scatterer unit M times in sequence. The receiving antenna receives M groups of scattered echoes, and constructs the echo measurement value of the scatterer unit in the current region based on the M groups of scattered echoes. Then, the echo measurement value corresponding to each region is obtained to construct the observation matrix;

[0014] 2) Using a transmitting antenna array, an electromagnetic wave with a preset frequency f1 and a different random phase modulation is transmitted M times to target scatterers of different shapes. The receiving antenna receives M groups of scattered echoes, and the echo measurement value of the target scatterer is constructed based on the M groups of scattered echoes. The target scatterer is processed to obtain the true backscattering coefficient of the target scatterer;

[0015] 3) The observation matrix is ​​combined with the echo measurement values ​​and true backscatter coefficients obtained from target scatterers of different shapes to form several samples and then form a microwave imaging data set.

[0016] The random irradiation microwave imaging network includes a generative adversarial network and a gated recurrent neural network. The microwave imaging data set is input into the generative adversarial network for phase retrieval processing to obtain an initial value of the backscattering coefficient. The initial value of the backscattering coefficient is then input into the gated recurrent neural network for processing to obtain an estimated value of the backscattering coefficient of the target scatterer, which serves as the output of the random irradiation microwave imaging network.

[0017] The generative adversarial network includes a generator model G and a discriminator model D, wherein the generator model G receives the echo measurement value as input and outputs an initial value of the backscatter coefficient, and the discriminator model D receives the true backscatter coefficient and the initial value of the backscatter coefficient output by the generator model G, and outputs a true or false label probability value of the initial value of the backscatter coefficient;

[0018] The loss function of the generative adversarial network is set according to the following formula:

[0019]

[0020]

[0021]

[0022] Among them, E represents the expected value of the variable in the sample set, D represents the discriminator model, G represents the generator model, and P data Represents the sample distribution of the training data set, f, g int and H represent the true backscatter coefficient, echo measurement value and observation matrix respectively, λ GP Represents the gradient penalty strength, GP represents the gradient penalty term, are samples interpolated between f and the output of the generator, is the discriminator in the generator pair Output The gradient at , ‖‖2 is the Euclidean norm.

[0023] The generator model G is mainly composed of several generating layers and an output layer connected in sequence, each generating layer is mainly composed of a transposed convolution layer, a normalization layer and a Relu activation function connected in sequence, and the output layer is mainly composed of a transposed convolution layer and a Tanh activation function connected in sequence;

[0024] The input of the transposed convolution layer in the first generation layer is used as the input of the generator model G. The transposed convolution layers in the second generation layer to the last generation layer all take the output of the Relu activation function of the previous generation layer as input. The output of the Relu activation function of the last generation layer is input to the transposed convolution layer of the output layer, and the output of the Tanh activation function of the output layer is used as the output of the generator model G.

[0025] The discriminator model D is mainly composed of an import layer, several discrimination layers and an output layer connected in sequence. The import layer is mainly composed of a convolutional layer and a LeakyRelu activation function connected in sequence. The discrimination layer is mainly composed of a convolutional layer, a normalization layer and a LeakyRelu activation function connected in sequence. The output layer is mainly composed of a convolutional layer and a Tanh activation function connected in sequence.

[0026] The input of the convolutional layer in the import layer is used as the input of the discriminator model D. The output of the LeakyRelu activation function in the import layer is input to the convolutional layer in the first discrimination layer. The convolutional layers in the second discrimination layer to the last discrimination layer all use the output of the LeakyRelu activation function of the previous discrimination layer as input. The output of the LeakyRelu activation function in the last discrimination layer is input to the convolutional layer of the output layer, and the output of the Tanh activation function of the output layer is used as the output of the discriminator model D.

[0027] The gated recurrent neural network is mainly composed of L iterative layers connected in sequence, and each iterative layer is composed of an encoder, a gated recurrent unit and a decoder connected in sequence;

[0028] The input of the first iterative layer is the initial value of the backscattering coefficient, the echo measurement value and the observation matrix. The input of the second to the Lth iterative layer are all echo measurement values, the observation matrix and the estimated backscattering coefficient output by the previous iterative layer. The encoder of each iterative layer concatenates its own input and the gradient calculated based on its own input to obtain a high-dimensional tensor with feature enhancement, which is input as the output of the encoder of the current iterative layer to the gated recurrent unit of the current iterative layer. The output of the encoder of the current iterative layer and the hidden state of the gated recurrent unit in the previous iterative layer are used as the input of the gated recurrent unit in the current iterative layer. The output of the gated recurrent unit in the current iterative layer is input to the decoder of the current iterative layer. The output of the decoder of the current iterative layer is used as part of the input of the next iterative layer or the output of the gated recurrent neural network.

[0029] The loss function L3 of the gated recurrent neural network is set according to the following formula:

[0030]

[0031] Where, f represents the true backscattering coefficient, f L represents the estimated backscatter coefficient output by the gated recurrent neural network, Represents the square of the Euclidean norm.

[0032] The beneficial effects of the present invention are:

[0033] This paper uses a generative adversarial network and unsupervised learning to enable the network to learn the physical properties of a random irradiation microwave imaging model, thereby improving the reliability of the initial value of the backscatter coefficient. This paper also uses a gated recurrent neural network and supervised learning to improve the accuracy of the network's iterative inversion, achieving high-quality image reconstruction with only a small number of iterations.

[0034] This paper proposes a novel algorithmic architecture: It leverages a generative adversarial mechanism to improve the reliability of initial values ​​for iterations, while also simulating traditional algorithms to construct a mapping network for efficient iterative solutions to electromagnetic inverse problems. This provides a novel methodology for target reconstruction in complex media environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 It is a flow chart of the overall implementation of the present invention.

[0036] Figure 2 Schematic diagram of the transmitting antenna array in the random illumination microwave imaging system of the present invention.

[0037] Figure 3 This is a training flow chart of the neural network described in the present invention.

[0038] Figure 4 This is a model structure diagram of the generative adversarial network described in the present invention.

[0039] Figure 5 This is a model structure diagram of the gated recurrent neural network described in the present invention.

[0040] Figure 6 This is a visualization diagram of the inversion results of the present invention for a target scatterer with a resolution of 16×16.

[0041] Figure 7 This is a visualization diagram of the inversion results of the present invention for a target scatterer with a resolution of 32×32. DETAILED DESCRIPTION

[0042] The following is a general description of the proposed method of random irradiation phaseless microwave imaging with generative adversarial network augmentation in conjunction with the accompanying drawings of the present invention. Figure 1 shown.

[0043] The method of the present invention adopts a random irradiation microwave imaging system with a preset operating frequency of 10 GHz.

[0044] Microwave imaging methods such as Figure 1 The following steps are shown:

[0045] S1. The random irradiation microwave imaging system collects scattered echoes from the target scatterer, processes the scattered echoes to obtain echo measurements and constructs an observation matrix. Furthermore, the true backscatter coefficients are obtained based on the target scatterer processing, and a microwave imaging dataset is constructed based on the echo measurements, the true backscatter coefficients, and the observation matrix.

[0046] Target scatterers are objects that can generate scattered echoes under microwave irradiation, including but not limited to human bodies, aircraft, vehicles, metal components, etc.

[0047] The echo measurement value is the amplitude value of the scattered echo, that is, g int .

[0048] Methods Random irradiation microwave imaging system was used. Figure 2 As shown, the random irradiation microwave imaging system includes a transmitting antenna array and a receiving antenna. The transmitting antenna array transmits electromagnetic waves of preset frequency and random phase modulation and irradiates the target scatterer. The electromagnetic waves interact with the target scatterer and generate scattered echoes, which are received by the receiving antenna.

[0049] S1 is specifically:

[0050] 1) Divide the preset imaging area into several regions, place the same scatterer unit in each region, and transmit electromagnetic waves with a preset frequency f1 and different random phase modulation to the scatterer unit M times in sequence. The receiving antenna receives M groups of scattered echoes, and constructs the echo measurement value of the scatterer unit in the current region based on the M groups of scattered echoes. Then, the echo measurement value corresponding to each region is obtained to construct the observation matrix;

[0051] Specifically, the observation matrix is ​​constructed according to the following steps: first, the imaging area is discretized into a grid of the system resolution size, and then the same scatterer units are placed in sequence on a single grid. At the same time, the transmitting antenna array is used to sequentially transmit M electromagnetic waves with a preset frequency and random phase modulation to the scatterer unit. The receiving antenna receives M groups of scattered echoes. The column vector of the observation matrix at the corresponding position is constructed based on the M groups of scattered echoes. After completing the point-by-point measurement of all grids and subtracting the background noise interference, the complete observation matrix is ​​obtained. For the assembled imaging system, its observation matrix is ​​a complex-valued matrix with phase information and is unique.

[0052] 2) Using a transmitting antenna array, an electromagnetic wave with a preset frequency f1 and a different random phase modulation is transmitted M times to target scatterers of different shapes. The receiving antenna receives M groups of scattered echoes, and the echo measurement value of the target scatterer is constructed based on the M groups of scattered echoes. The target scatterer is processed to obtain the true backscattering coefficient of the target scatterer;

[0053] 3) The observation matrix is ​​combined with the echo measurement values ​​and true backscatter coefficients obtained from target scatterers of different shapes to form several samples and then form a microwave imaging data set.

[0054] The observation matrix plus the echo measurement and true backscatter coefficient corresponding to a target scatterer is considered a sample. The microwave imaging dataset includes several samples, all of which are randomly divided into training and test sets in a ratio of 19:1 to complete the construction of the neural network dataset.

[0055] The target scatterer in this embodiment is a random irradiation microwave imaging system with a preset frequency of f1. The model uses an image read from a handwritten digit set as the target scatterer and models it. The model simulates P×P transmitting antennas that have undergone M times of random phase modulation to transmit M groups of electromagnetic waves to the same target scatterer. The receiving antenna receives and stores the M groups of echo measurements as g int .

[0056] Specifically, the preset frequency of the electromagnetic wave in this embodiment is 10GHz. In the example, the picture read from the handwritten digit set is used as the target scatterer and modeled. For target scatterers with different resolutions, corresponding data sets need to be constructed: For target scatterers with an image resolution of 16×16: all 7291 pictures are read from the USPS handwritten digit set and used as target scatterers. Simulate 15×15 transmitting antennas that have undergone 225 different random phase modulations to transmit 225 groups of electromagnetic waves to the same target scatterer, and the receiving antenna receives and stores 225 groups of echo measurement values ​​g; the sample consists of the echo measurement value g corresponding to the same target scatterer int , the observation matrix H, and the true backscattering coefficient f of the target scatterer are paired to construct the neural network data set.

[0057] For a target scatterer with a resolution of 32×32: First, read 12,000 images from the MNIST handwritten digit library. Then, add a zero pixel to each of the four edges of the image, expanding the 28×28 image to a 32×32 image and using it as the target scatterer. Simulate 15×15 transmitting antennas that have been subjected to 225 different random phase modulations to transmit 225 sets of electromagnetic waves to the same target scatterer. The receiving antenna receives and stores the 225 sets of echo measurements as g. int ; The echo measurement value g corresponding to the same target scattererint , the observation matrix H, and the true backscattering coefficient f of the target scatterer are paired to construct the neural network data set.

[0058] S2. Construct a random irradiation microwave imaging network that includes a generative adversarial network and a gated recurrent neural network, input the microwave imaging dataset into the random irradiation microwave imaging network for training, and obtain a trained random irradiation microwave imaging network.

[0059] S2 is specifically:

[0060] 1) The random irradiation microwave imaging network includes a generative adversarial network and a gated recurrent neural network. The microwave imaging dataset is input into the generative adversarial network for phase retrieval processing to obtain the initial value of the backscattering coefficient. The initial value of the backscattering coefficient is then input into the gated recurrent neural network for processing to obtain the estimated backscattering coefficient of the target scatterer, which serves as the output of the random irradiation microwave imaging network.

[0061] 2) Generative adversarial network includes generator model G and discriminator model D. Generator model G receives echo measurement value g int As input, it outputs the initial value of the backscatter coefficient; the discriminator model D receives the true backscatter coefficient and the initial value of the backscatter coefficient output by the generator model G, and outputs the true and false label probability value of the initial value of the backscatter coefficient;

[0062] The loss functions L1 and L2 of the generative adversarial network are set according to the following formula:

[0063]

[0064]

[0065]

[0066] Among them, E represents the expected value of the variable in the sample set, D represents the discriminator model, G represents the generator model, and P data Represents the sample distribution of the training data set, f, g int and H represent the true backscatter coefficient, echo measurement value and observation matrix respectively, λ GP Represents the gradient penalty strength, GP represents the gradient penalty term, are samples interpolated between f and the output of the generator, is the discriminator in the generator pair Output The gradient at , ‖‖2 is the Euclidean norm, and || represents the absolute value.

[0067] The training data set is used as a training sample to perform unsupervised training on the generative adversarial network. The L1 loss function is first used to update the parameters so that the network can learn the physical characteristics of the random irradiation model, and then the L2 loss function is used to make the network further adjust the parameters through adversarial operation. After completing a parameter update, the generator G is generated according to the echo measurement value g. int Regenerate a new initial value of the backscatter coefficient and pass it into Figure 5 The gated recurrent neural network shown in the figure is iteratively solved. The training process of the neural network is as follows Figure 3 shown.

[0068] The generator model G is mainly composed of several generation layers and an output layer connected in sequence. Each generation layer is mainly composed of a transposed convolution layer, a normalization layer and a Relu activation function connected in sequence. The output layer is mainly composed of a transposed convolution layer and a Tanh activation function connected in sequence.

[0069] The input of the transposed convolution layer in the first generation layer is used as the input of the generator model G. The transposed convolution layers in the second generation layer to the last generation layer all take the output of the Relu activation function of the previous generation layer as input. The output of the Relu activation function of the last generation layer is input to the transposed convolution layer of the output layer, and the output of the Tanh activation function of the output layer is used as the output of the generator model G.

[0070] The generator model G needs to change the number of generation layers when processing target scatterers of different resolutions. The number of layers can be reduced to process low-resolution target scatterers, and the number of layers can be increased to process high-resolution target scatterers.

[0071] The discriminator model D is mainly composed of an import layer, several discrimination layers and an output layer connected in sequence. The import layer is mainly composed of a convolutional layer and a LeakyRelu activation function connected in sequence. The discrimination layer is mainly composed of a convolutional layer, a normalization layer and a LeakyRelu activation function connected in sequence. The output layer is mainly composed of a convolutional layer and a Tanh activation function connected in sequence.

[0072] The input of the convolutional layer in the import layer is the input of the discriminator model D. The output of the LeakyRelu activation function in the import layer is input to the convolutional layer in the first discrimination layer. The convolutional layers in the second to last discrimination layers all use the output of the LeakyRelu activation function of the previous discrimination layer as input. The output of the LeakyRelu activation function in the last discrimination layer is input to the convolutional layer of the output layer, and the output of the Tanh activation function of the output layer is used as the output of the discriminator model D.

[0073] The structure of the discriminator model D needs to be adapted to the generator model G to perform its discrimination function, reducing the number of discrimination layers to process low-resolution target scatterers and increasing the number of discrimination layers to process high-resolution target scatterers.

[0074] Specifically, using Figure 4 The conditional adversarial network shown generates reliable iterative initialization values. For processing a 16×16 resolution target scatterer: G contains three connected transposed convolutional layers, with a normalization layer and a Relu activation function added to the tail of the first and second layers, and a Tanh activation function added to the tail of the third layer; D contains three connected convolutional layers, with a LeakyRelu activation function added to the tail of the first convolutional layer, a normalization layer and a LeakyRelu activation function added to the tail of the second layer, and the third layer directly outputting.

[0075] For processing target scatterers with a resolution of 32×32: G contains four connected transposed convolutional layers, with normalization layers and Relu activation functions added to the tails of the first to third layers, and a Tanh activation function added to the tail of the fourth layer; D contains four connected convolutional layers, with a LeakyRelu activation function added to the tail of the first convolutional layer, normalization layers and LeakyRelu activation functions added to the tails of the second and third layers, and the fourth layer directly outputs.

[0076] 3) The gated recurrent neural network is mainly composed of L iterative layers of the same structure connected in sequence. Each iterative layer is composed of an encoder, a gated recurrent unit (GRU) module, and a decoder connected in sequence.

[0077] The input of the first iterative layer is the initial value of the backscattering coefficient, the echo measurement value and the observation matrix. The input of the second to the Lth iterative layer are the echo measurement value, the observation matrix and the estimated backscattering coefficient output by the previous iterative layer. The encoder of each iterative layer splices its own input and the gradient calculated according to its own input in the same dimension to obtain a high-dimensional tensor after feature enhancement, that is, two M*1 tensors are spliced ​​in the dimension of M to form a new 2M*1 tensor, which is input as the output of the encoder of the current iterative layer to the gated recurrent unit of the current iterative layer. The output of the encoder of the current iterative layer and the hidden state of the gated recurrent unit in the previous iterative layer are used as the input of the gated recurrent unit in the current iterative layer. The output of the gated recurrent unit in the current iterative layer is input to the decoder of the current iterative layer. The output of the decoder of the current iterative layer is used as part of the input of the next iterative layer or the output of the gated recurrent neural network.

[0078] That is, the encoder of the first layer receives the iterative initial value generated by G and starts the calculation. The input of each iterative layer is the output of the previous layer. The encoder of the kth layer converts the input f of this layer into k-1 and according to fk-1 The calculated gradients are concatenated at the same latitude to obtain a tensor with enhanced features. The gradient calculation formula is as follows:

[0079]

[0080] Among them, h represents the column vector of the observation matrix H, g int is the amplitude value obtained by processing the echo measurement value, the symbol * refers to the conjugate transpose, M is the number of groups of electromagnetic waves modulated with different random phases, and f k-1 is the estimated backscattering coefficient output by the k-1th iterative layer.

[0081] The tensor output by the k-th layer encoder is input to the gated recurrent unit of the k-th layer; the gated recurrent unit uses the linear layer to operate on the tensor, and controls the degree of influence of the hidden state of the k-1-th layer gated recurrent unit on the current linear layer operation through the update gate and reset gate; the k-th layer decoder receives the output of the gated recurrent unit of this layer, and obtains a more accurate backscatter coefficient estimate after this iteration through two nonlinear neural network layers with ReLU activation function; the output f of the L-th layer L The estimated backscatter coefficient obtained by the entire neural network iteration.

[0082] The loss function L3 of the gated recurrent neural network is set according to the following formula:

[0083]

[0084] Where, f represents the true backscattering coefficient, f L represents the estimated backscatter coefficient output by the gated recurrent neural network, Represents the square of the Euclidean norm.

[0085] S3. Collect the scattered echoes of the target scatterer to be measured and process them to obtain the measured echo measurement value, input the measured echo measurement value and the observation matrix into the trained random irradiation microwave imaging network to obtain the backscattering coefficient of the target scatterer to be measured, and then construct the image of the target scatterer to be measured based on the backscattering coefficient of the target scatterer to be measured.

[0086] The training dataset is used as training samples for supervised training of a gated recurrent neural network. During training, the L3 loss function is used to update the network parameters by minimizing the normalized root mean square error between the backscatter coefficient estimate and the true value. Furthermore, the trained deep learning neural network can reconstruct the image of the target scatterer using only the echo measurements and the observation matrix. The network output is the backscatter coefficient estimate of the target scatterer. Figure 6 Visualization of the inversion results for the test set of target scatterers with a resolution of 16×16. Figure 7Visualization of the inversion results for the test set of target scatterers with a resolution of 32×32.

[0087] The innovation of the present invention is:

[0088] 1. Traditional imaging methods usually rely on multiple information (such as phase and amplitude) for image reconstruction, which is usually greatly affected by noise and environmental factors. The present invention reconstructs images based on the amplitude value (intensity information) of the echo signal, reducing imaging requirements, improving adaptability, and having higher robustness.

[0089] 2. A generative adversarial network (WGAN) replaces the traditional spectral method to obtain the iterative initial value, that is, the initial value of the backscattering coefficient. Without relying on manual feature extraction, the adversarial network can generate more realistic and reliable initial values ​​of the backscattering coefficient by learning the physical properties of the electromagnetic imaging model.

[0090] 3. The present invention does not use neural networks to directly reconstruct images end-to-end, but instead simulates traditional algorithms to build a mapping network to achieve efficient iterative solution of electromagnetic inverse problems. It uses neural networks to learn "physical properties" and then uses neural networks to reconstruct the electromagnetic imaging process, retaining physical interpretability.

[0091] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in the present invention should be included in the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

Claims

1. A generative adversarial network augmented unfolded random irradiation phaseless microwave imaging method for acquiring images of target scatterers, characterized in that: The method comprises the following steps: S1. The random irradiation microwave imaging system collects scattered echoes from the target scatterer, processes the scattered echoes to obtain echo measurements and constructs an observation matrix. Furthermore, the true backscatter coefficients are obtained based on the target scatterer processing, and a microwave imaging dataset is constructed based on the echo measurements, the true backscatter coefficients, and the observation matrix. S2. Construct a random irradiation microwave imaging network including a generative adversarial network and a gated recurrent neural network, input the microwave imaging dataset into the random irradiation microwave imaging network for training, and obtain a trained random irradiation microwave imaging network; S3. Collect the scattered echoes of the target scatterer to be measured and process them to obtain the measured echo measurement value, input the measured echo measurement value and the observation matrix into the trained random irradiation microwave imaging network to obtain the backscattering coefficient of the target scatterer to be measured, and then construct the image of the target scatterer to be measured based on the backscattering coefficient of the target scatterer to be measured.

2. The method of claim 1, wherein: The echo measurement value is the amplitude value of the scattered echo.

3. The method of claim 1, wherein: The method adopts a random irradiation microwave imaging system, which includes a transmitting antenna array and a receiving antenna. The transmitting antenna array transmits electromagnetic waves with a preset frequency and random phase modulation to irradiate the target scatterer. The electromagnetic waves interact with the target scatterer and generate scattered echoes, which are received by the receiving antenna.

4. The method of claim 3, wherein: The S1 is specifically: 1) Divide the preset imaging area into several regions, place the same scatterer unit in each region, and transmit electromagnetic waves with a preset frequency f1 and different random phase modulation to the scatterer unit M times in sequence. The receiving antenna receives M groups of scattered echoes, and constructs the echo measurement value of the scatterer unit in the current region based on the M groups of scattered echoes. Then, the echo measurement value corresponding to each region is obtained to construct the observation matrix; 2) Using a transmitting antenna array, an electromagnetic wave with a preset frequency f1 and a different random phase modulation is transmitted M times to target scatterers of different shapes. The receiving antenna receives M groups of scattered echoes, and the echo measurement value of the target scatterer is constructed based on the M groups of scattered echoes. The target scatterer is processed to obtain the true backscattering coefficient of the target scatterer; 3) The observation matrix is ​​combined with the echo measurement values ​​and true backscatter coefficients obtained from target scatterers of different shapes to form several samples and then form a microwave imaging data set.

5. The method of claim 1, wherein: The random irradiation microwave imaging network includes a generative adversarial network and a gated recurrent neural network. The microwave imaging data set is input into the generative adversarial network for phase retrieval processing to obtain an initial value of the backscattering coefficient. The initial value of the backscattering coefficient is then input into the gated recurrent neural network for processing to obtain an estimated value of the backscattering coefficient of the target scatterer, which serves as the output of the random irradiation microwave imaging network.

6. The method of claim 1, wherein: The generative adversarial network includes a generator model G and a discriminator model D, wherein the generator model G receives the echo measurement value as input and outputs an initial value of the backscatter coefficient, and the discriminator model D receives the true backscatter coefficient and the initial value of the backscatter coefficient output by the generator model G, and outputs a true or false label probability value of the initial value of the backscatter coefficient; The loss function of the generative adversarial network is set according to the following formula: Among them, E represents the expected value of the variable in the sample set, D represents the discriminator model, G represents the generator model, and P data Represents the sample distribution of the training data set, f, g int and H represent the true backscatter coefficient, echo measurement value and observation matrix respectively, λ GP Represents the gradient penalty strength, GP represents the gradient penalty term, are samples interpolated between f and the output of the generator, is the discriminator in the generator pair Output The gradient at , ‖‖2 is the Euclidean norm.

7. The method of claim 5, wherein: The generator model G is mainly composed of several generating layers and an output layer connected in sequence, each generating layer is mainly composed of a transposed convolution layer, a normalization layer and a Relu activation function connected in sequence, and the output layer is mainly composed of a transposed convolution layer and a Tanh activation function connected in sequence; The input of the transposed convolution layer in the first generation layer is used as the input of the generator model G. The transposed convolution layers in the second generation layer to the last generation layer all take the output of the Relu activation function of the previous generation layer as input. The output of the Relu activation function of the last generation layer is input to the transposed convolution layer of the output layer, and the output of the Tanh activation function of the output layer is used as the output of the generator model G. The discriminator model D is mainly composed of an import layer, several discrimination layers and an output layer connected in sequence. The import layer is mainly composed of a convolutional layer and a LeakyRelu activation function connected in sequence. The discrimination layer is mainly composed of a convolutional layer, a normalization layer and a LeakyRelu activation function connected in sequence. The output layer is mainly composed of a convolutional layer and a Tanh activation function connected in sequence. The input of the convolutional layer in the import layer is used as the input of the discriminator model D. The output of the LeakyRelu activation function in the import layer is input to the convolutional layer in the first discrimination layer. The convolutional layers in the second discrimination layer to the last discrimination layer all use the output of the LeakyRelu activation function of the previous discrimination layer as input. The output of the LeakyRelu activation function in the last discrimination layer is input to the convolutional layer of the output layer, and the output of the Tanh activation function of the output layer is used as the output of the discriminator model D.

8. The method of claim 1, wherein: The gated recurrent neural network is mainly composed of L iterative layers connected in sequence, and each iterative layer is composed of an encoder, a gated recurrent unit and a decoder connected in sequence; The input of the first iterative layer is the initial value of the backscattering coefficient, the echo measurement value and the observation matrix. The input of the second to the Lth iterative layer are all echo measurement values, the observation matrix and the estimated backscattering coefficient output by the previous iterative layer. The encoder of each iterative layer concatenates its own input and the gradient calculated based on its own input to obtain a high-dimensional tensor with feature enhancement, which is input as the output of the encoder of the current iterative layer to the gated recurrent unit of the current iterative layer. The output of the encoder of the current iterative layer and the hidden state of the gated recurrent unit in the previous iterative layer are used as the input of the gated recurrent unit in the current iterative layer. The output of the gated recurrent unit in the current iterative layer is input to the decoder of the current iterative layer. The output of the decoder of the current iterative layer is used as part of the input of the next iterative layer or the output of the gated recurrent neural network. The loss function L3 of the gated recurrent neural network is set according to the following formula: Where, f represents the true backscattering coefficient, f L represents the estimated backscatter coefficient output by the gated recurrent neural network, Represents the square of the Euclidean norm.

Images