Nonlinear depth-aware microwave imaging method and system
By employing a nonlinear depth-sensing microwave imaging method, combined with deep learning and physical embedding reconstruction, the contradiction between resolution and complexity in traditional microwave imaging methods is resolved, achieving high-resolution microwave imaging with physical constraints.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- GANDONG UNIV
- Filing Date
- 2026-04-17
- Publication Date
- 2026-06-09
AI Technical Summary
Traditional microwave imaging methods struggle to balance imaging resolution and system complexity, and existing deep learning models lack physical constraints, resulting in poor generalization ability.
A nonlinear depth-sensing microwave imaging method is adopted. The high-dimensional echo signal is adaptively compressed into low-dimensional nonlinear measurement values through a nonlinear depth-sensing subnetwork, and the nonlinear inverse scattering solution is performed by a physical embedding reconstruction subnetwork. By combining the advantages of physical laws and deep learning, high-resolution imaging is achieved.
It achieves high-resolution imaging in a variety of practical application scenarios, has strong generalization ability, is suitable for complex scenarios, and the imaging results meet physical constraints.
Smart Images
Figure CN122172184A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of microwave imaging technology, and relates to, but is not limited to, a nonlinear depth-sensing microwave imaging method and system. Background Technology
[0002] Microwave imaging, due to its advantages of penetrating non-metallic materials, producing no ionizing radiation, and being naturally compatible with communication signals, has significant application value in fields such as public safety, industrial non-destructive testing, biomedical diagnostics, and the integration of sensing and communication in 6G mobile communications. However, traditional microwave imaging methods face a fundamental contradiction: it is difficult to achieve both high imaging resolution and system complexity simultaneously.
[0003] From an electromagnetic theory perspective, the interaction between microwaves and targets is inherently nonlinear. While precise electromagnetic inverse scattering models (such as contrast source inversion CSI) describe this nonlinear relationship, existing solution algorithms typically employ an alternating direction iterative strategy, decomposing the nonlinear problem into a series of linear subproblems for approximation. This approach is prone to getting trapped in local optima and suffers from high computational complexity. In recent years, deep learning has been used for image reconstruction to achieve microwave imaging. However, existing deep learning methods are mostly purely data-driven black-box models, lacking physical constraints, resulting in poor generalization ability and a sharp decline in reconstruction performance for scenes outside the training set. Summary of the Invention
[0004] In view of this, embodiments of the present invention provide a nonlinear depth-sensing microwave imaging method and system, which not only has strong generalization ability, but also can achieve high-resolution imaging.
[0005] The specific technical solutions of this invention are as follows: A first aspect of this application provides a nonlinear depth-sensing microwave imaging method, comprising: Microwave signals are transmitted to the area to be imaged via a transmitting antenna, and the corresponding high-dimensional echo signals are received via a receiving antenna array. The high-dimensional echo signal is adaptively compressed into the optimal low-dimensional nonlinear measurement value through a pre-defined target nonlinear deep sensing subnetwork; wherein the dimension of the low-dimensional nonlinear measurement value is smaller than the dimension of the high-dimensional echo signal. By using a pre-defined target physical embedding reconstruction subnetwork to perform nonlinear inverse scattering on low-dimensional nonlinear measurements, a high-resolution image corresponding to the region to be imaged is obtained.
[0006] In some embodiments, the target nonlinear deep perception subnetwork includes a nonlinear convolutional block, a global residual fusion module, and an adaptive nonlinear combination layer; The target nonlinear deep sensing subnetwork is specifically used to perform the following operations: Multi-scale feature extraction is performed on high-dimensional echo signals using nonlinear convolutional blocks to obtain multi-scale feature maps; these multi-scale feature maps include shallow feature maps, mid-level feature maps, and deep feature maps. The global residual fusion module performs cross-layer fusion of the mid-layer feature map and the deep feature map to obtain the fused feature. The fused features are flattened by an adaptive nonlinear combination layer and then input into a fully connected layer. A learnable parameterized activation function is used for nonlinear dynamic weighting to obtain low-dimensional nonlinear measurement values.
[0007] In some embodiments, a nonlinear convolutional block is composed of multiple convolutional layers and multiple nonlinear activation function layers stacked alternately, with a batch normalization layer connected after each nonlinear activation function; wherein, the convolutional layers are used to perform convolution operations on the input signal to extract local features at different scales; the nonlinear activation function layers are used to perform nonlinear mapping on the output of the convolutional layers; and the batch normalization layer is used to normalize the activated feature map output by the nonlinear activation function.
[0008] In some embodiments, the target physical embedding reconstruction subnetwork includes an initialization subnetwork and a deep unrolling iteration module; The target physical embedding reconstruction subnetwork specifically performs the following operations: Low-dimensional nonlinear measurements are mapped to contrast functions and contrast sources by initializing subnetworks; The contrast function and the contrast source are iteratively processed by the depth expansion iteration module until the preset number of iterations is reached, so as to obtain a high-resolution image that satisfies the electromagnetic state equation.
[0009] In some embodiments, the depth unrolling iteration module is composed of multiple alternating stacks of differentiable conjugate gradient layers and data consistency layers; The depth unrolling iteration module performs the following operations: By iteratively solving the contrast function and the contrast source through a differentiable conjugate gradient layer, a new contrast source satisfying the electromagnetic state equation is obtained; where the electromagnetic state equation is constructed based on the in-domain Green's function matrix and the incident field. The contrast function is updated by the data consistency layer based on the scattering data equation and the new contrast source to obtain a new contrast function; Using the new contrast source and the new contrast function as input for the next iteration, the operations corresponding to the differentiable conjugate gradient layer and the data consistency layer are repeatedly executed until the preset number of iterations is reached, resulting in a high-resolution image.
[0010] In some embodiments, the differentiable conjugate gradient layer specifically performs the following operations: Based on the current contrast function, and using the current contrast source as the initial value for iteration, the conjugate gradient algorithm is expanded into a network layer with a fixed number of iteration steps; each layer of the network layer includes vector inner product, matrix-vector multiplication, and linear combination operations, and all operations are differentiable; By using network layers to solve the electromagnetic state equations, a new contrast source is obtained.
[0011] In some embodiments, the data consistency layer updates the contrast function according to the following formula: in, Let i be the contrast function of the i-th pixel. To compare the components of the source vector J at the i-th pixel, For the incident electric field, For the predefined domain Green's function matrix, As a new source of comparison, It is a constant.
[0012] In some embodiments, the method further includes: Obtain the training dataset; the training dataset includes high-dimensional echo signal samples and the corresponding real image samples; Construct an initial nonlinear deep sensing subnetwork and an initial physical embedding reconstruction subnetwork, and then cascade the initial nonlinear deep sensing subnetwork and the initial physical embedding reconstruction subnetwork. The parameters of the cascaded initial nonlinear deep sensing subnetwork and the initial physical embedding reconstruction subnetwork are iteratively updated using a joint loss function until the cascaded initial nonlinear deep sensing subnetwork and the initial physical embedding reconstruction subnetwork converge, thus obtaining the target nonlinear deep sensing subnetwork and the target physical embedding reconstruction subnetwork; the joint loss function includes the mean squared error loss function and the physical consistency loss function.
[0013] A second aspect of this application provides a nonlinear depth-sensing microwave imaging system, comprising: The transmitting module is used to transmit microwave signals to the area to be imaged via a transmitting antenna; The receiving module is used to receive the corresponding high-dimensional echo signal through the receiving antenna array; The processing module is used to adaptively compress high-dimensional echo signals into optimal low-dimensional nonlinear measurement values through a preset target nonlinear depth sensing subnetwork; wherein the dimension of the low-dimensional nonlinear measurement value is smaller than the dimension of the high-dimensional echo signal; and to obtain a high-resolution image by performing nonlinear inverse scattering on the low-dimensional nonlinear measurement values through a preset target physical embedding reconstruction subnetwork.
[0014] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: In this embodiment of the invention, a microwave signal is first transmitted to the area to be imaged via a transmitting antenna, and the corresponding high-dimensional echo signal is received via a receiving antenna array. Then, a preset target nonlinear depth sensing subnetwork adaptively compresses the high-dimensional echo signal into the optimal low-dimensional nonlinear measurement value. The dimension of the low-dimensional nonlinear measurement value is smaller than the dimension of the high-dimensional echo signal. Finally, a preset target physical embedding reconstruction subnetwork performs nonlinear inverse scattering on the low-dimensional nonlinear measurement value to obtain a high-resolution image corresponding to the area to be imaged. In this way, by combining the advantages of physical laws and deep learning through the target nonlinear depth sensing subnetwork and the target physical embedding reconstruction subnetwork, it not only has strong generalization ability and is suitable for a variety of practical application scenarios, but also can handle microwave imaging tasks in complex scenarios and achieve high-resolution imaging. Attached Figure Description
[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort, wherein: Figure 1 A flowchart illustrating the nonlinear depth-sensing microwave imaging method provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of a target nonlinear deep sensing subnetwork provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of a target physical embedding reconstruction subnetwork provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the structure of a nonlinear depth-sensing microwave imaging system provided in an embodiment of the present invention. Detailed Implementation
[0016] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. The following embodiments are used to illustrate the present invention, but are not intended to limit the scope of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0017] In the following description, references are made to “some embodiments,” which describe a subset of all possible embodiments. However, it is understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.
[0018] It should be noted that the terms "first, second, and third" used in the embodiments of the present invention are only used to distinguish similar objects and do not represent a specific ordering of objects. It is understood that "first, second, and third" can be interchanged in a specific order or sequence where permitted, so that the embodiments of the present invention described herein can be implemented in an order other than that illustrated or described herein.
[0019] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which these embodiments of the invention pertain. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art and should not be interpreted in an idealized or overly formal sense unless specifically defined as herein.
[0020] Figure 1 This is a schematic flowchart illustrating a nonlinear depth-sensing microwave imaging method according to an embodiment of the present invention. The nonlinear depth-sensing microwave imaging method can be executed by a control device, which includes at least one of a personal computer, laptop computer, smartphone, tablet computer, and portable wearable device; this embodiment does not limit the type of control device.
[0021] like Figure 1 As shown, the nonlinear depth sensing microwave imaging method provided in this embodiment of the invention may include steps S101-S103.
[0022] S101. Transmit microwave signals to the area to be imaged through a transmitting antenna, and receive the corresponding high-dimensional echo signals through a receiving antenna array.
[0023] In some embodiments, the control device controls the transmitting antenna to transmit a microwave signal of a preset waveform toward the area to be imaged; after the microwave signal is scattered by the target object in the area to be imaged, the receiving antenna array synchronously collects the scattered echo signal and processes the scattered echo signal (such as frequency conversion, filtering and digitization) to obtain a high-dimensional echo signal.
[0024] For example, in emergency rescue scenarios such as building collapses, in order to quickly detect whether there are living beings behind the wall and their location and posture information, a microwave signal with a preset waveform can be transmitted to the collapsed area (i.e., the area to be imaged) by a transmitting antenna. When the microwave signal encounters a wall or other obstacle during propagation, it will be scattered, thereby generating a scattered echo signal. The receiving antenna array collects the scattered echo signal and performs down-conversion, filtering, and digitization processing on the scattered echo signal to obtain a high-dimensional echo signal.
[0025] In some embodiments, the transmitting antenna can be a horn antenna or an array antenna, etc. The receiving antenna array consists of multiple (e.g., 8) physical receiving channels, and the array elements can be arranged as a uniform linear array, a planar array, or a conformal array. This application does not limit the types of transmitting and receiving antenna arrays.
[0026] In some embodiments, the microwave signal with the preset waveform can be a continuous wave, a frequency-modulated continuous wave, or a pulse signal; the waveform of the microwave signal in this application embodiment is not limited. In some application scenarios, the waveform of the microwave signal can be S-band or C-band (e.g., 2GHz or 5.8GHz). In this way, both penetration capability and resolution can be considered.
[0027] S102. The high-dimensional echo signal is adaptively compressed into the optimal low-dimensional nonlinear measurement value through a preset target nonlinear deep sensing subnetwork.
[0028] In some embodiments, a high-dimensional echo signal is input into a preset target nonlinear deep sensing subnetwork. The preset target nonlinear deep sensing subnetwork performs layer-by-layer nonlinear transformation and feature compression on the high-dimensional echo signal to obtain a low-dimensional nonlinear measurement value, thereby completing the nonlinear sensing encoding from high-dimensional space to low-dimensional space. The dimension of the low-dimensional nonlinear measurement value is smaller than the dimension of the high-dimensional echo signal.
[0029] In some embodiments, Figure 2 This is a schematic diagram of the structure of a nonlinear depth-sensing subnetwork provided in an embodiment of this application. Figure 2 As shown, the target nonlinear deep sensing subnetwork includes nonlinear convolutional blocks, a global residual fusion module, and an adaptive nonlinear combination layer. Specifically, the target nonlinear deep sensing subnetwork performs the following operations: multi-scale feature extraction of the high-dimensional echo signal is performed through nonlinear convolutional blocks to obtain multi-scale feature maps; these multi-scale feature maps include shallow, mid-level, and deep feature maps; the mid-level and deep feature maps are fused across layers through the global residual fusion module to obtain fused features; the fused features are flattened and input into a fully connected layer through the adaptive nonlinear combination layer, where a learnable parameterized activation function is used for nonlinear dynamic weighting to obtain low-dimensional nonlinear measurement values.
[0030] In some embodiments, such as Figure 2 As shown, the nonlinear convolutional block is composed of multiple convolutional layers and multiple nonlinear activation function layers stacked alternately, with a batch normalization layer connected after each nonlinear activation function. The convolutional layers are used to perform convolution operations on the input signal to extract local features at different scales; the nonlinear activation function layers are used to perform nonlinear mapping on the output of the convolutional layers; and the batch normalization layer is used to normalize the activated feature map output by the nonlinear activation function.
[0031] For example, a nonlinear convolutional block can be composed of four convolutional layers (conv) and four nonlinear activation function layers (ReLE) stacked alternately. Each nonlinear activation function is followed by a batch normalization (BN) layer. That is, a nonlinear convolutional block can be composed of four stacked nonlinear convolutional layers, each consisting of a convolutional layer, a nonlinear activation function layer, and a batch normalization layer stacked sequentially. The convolutional kernel size can be 3×3, with a stride of 1 and padding of 1. The nonlinear activation function layer can be a ReLU activation function. The global residual fusion module is equivalent to residual connections, used to perform element-wise addition of the mid-level feature maps and the deep feature maps.
[0032] In some embodiments, such as Figure 2 As shown, the adaptive nonlinear combination layer consists of a flattening layer and multiple fully connected layers; each fully connected layer includes a learnable parameterized activation function (such as the PReLU activation function). The flattening layer flattens the fused features into a one-dimensional vector; the fully connected layers use the learnable parameterized activation function to nonlinearly and dynamically weight the flattened features, generating low-dimensional nonlinear measurements.
[0033] For example, the adaptive nonlinear combination layer consists of two fully connected layers and two nonlinear activation layers (such as the PReLU activation function) stacked alternately. This layer is used to dynamically weight the flattened fused features nonlinearly, generating low-dimensional nonlinear measurements. The first fully connected layer maps the input features (i.e., the fused features) to a high-dimensional space, and the second fully connected layer maps the features mapped to the high-dimensional space back to the low-dimensional space. The nonlinear activation layers introduce nonlinear transformation capabilities after each fully connected layer, thereby enhancing the network's expressive power.
[0034] In practical applications, nonlinear convolutional blocks extract multi-scale deep features related to the imaging region from high-dimensional echo signals through alternating convolution operations and nonlinear activation functions (such as ReLU activation function). This achieves preliminary nonlinear transformation and dimensionality reduction, resulting in a multi-scale feature map. Subsequently, the global residual fusion module adds the mid-level feature map (output of the third nonlinear convolutional layer) and the deep feature map (output of the fourth nonlinear convolutional layer) element-wise from the multi-scale feature map to fuse features of different depths, obtaining a fused feature. This avoids the loss of effective information in deep networks, ensuring the preservation of key scene information during compression. Finally, the fused features are flattened into a one-dimensional vector by a flattening layer in the adaptive nonlinear combination layer and input into a fully connected layer. The fully connected layer uses a learnable parameterized activation function (such as PReLU activation function) to perform nonlinear dynamic weighting on the flattened features, thereby generating low-dimensional nonlinear measurements. In this way, not only can the data dimensionality be reduced, but key information in high-dimensional echo signals can also be preserved through nonlinear transformation.
[0035] S103. The low-dimensional nonlinear measurement values are solved by nonlinear inverse scattering through a preset target physical embedding reconstruction subnetwork to obtain a high-resolution image corresponding to the region to be imaged.
[0036] In some embodiments, the target nonlinear depth sensing subnetwork outputs low-dimensional nonlinear measurements to the target physical embedding reconstruction subnetwork, which then performs nonlinear inverse scattering on the low-dimensional nonlinear measurements using a depth unfolding architecture to recover a high-resolution image of the region to be imaged.
[0037] In some embodiments, Figure 3 A schematic diagram of the structure of a target physical embedding reconstruction subnetwork is shown. For example... Figure 3 As shown, the target physical embedding reconstruction subnetwork includes an initialization subnetwork and a deep unfolding iteration module. The target physical embedding reconstruction subnetwork specifically performs the following operations: the low-dimensional nonlinear measurement values are mapped to a contrast function and a contrast source through the initialization subnetwork; the contrast function and the contrast source are iteratively processed by the deep unfolding iteration module until the number of iterations reaches a preset number of iterations (e.g., k times, where k is a positive integer), to obtain a high-resolution image that satisfies the electromagnetic state equation.
[0038] In some embodiments, the initialization subnetwork may include two stacked fully connected networks for mapping the input low-dimensional nonlinear measurements to obtain an initial contrast function and contrast source. The initial contrast function and contrast source are then input as initial inputs to a depth unrolling iteration module for iterative processing until a preset number of iterations is reached, resulting in a high-resolution image that satisfies the electromagnetic state equations. The electromagnetic state equations are the fundamental equations in electromagnetic field theory, used to describe the propagation and interaction characteristics of electromagnetic fields. The electromagnetic state equations can be expressed by the following Equations 1 and 2: Formula 1; Formula 2; in, To compare with the source vector, For contrast function, For the incident electric field, The scattered electric field in the region to be imaged. The scattered electric field at the receiving antenna array, To receive the Green's function matrix, This is a point-by-point multiplication operation for vectors.
[0039] In some embodiments, such as Figure 3 As shown, the depth unrolling iterative module consists of alternating stacks of differentiable conjugate gradient layers and data consistency layers. Specifically, the depth unrolling iterative module performs the following operations: Iteratively solving the contrast function and contrast source using the differentiable conjugate gradient layer yields a new contrast source that satisfies the electromagnetic state equation; the electromagnetic state equation is constructed based on the in-domain Green's function matrix and the incident field; updating the contrast function using the data consistency layer based on the scattering data equation and the new contrast source yields a new contrast function; using the new contrast source and the new contrast function as input for the next iteration, the operations corresponding to the differentiable conjugate gradient layer and the data consistency layer are repeatedly executed until the preset number of iterations (e.g., k times, where k is a positive integer) is reached, resulting in a high-resolution image.
[0040] In some embodiments, the depth unrolling iterative module can be composed of eight alternating stacks of differentiable conjugate gradient layers and data consistency layers. That is, the depth unrolling iterative module consists of eight iterative units, each consisting of a stacked differentiable conjugate gradient layer and a data consistency layer. The depth unrolling iterative module performs eight iterations (k=8) on the contrast function and the contrast source through these eight iterative units to obtain a high-resolution image. In other words, the preset number of iterations is equal to the number of iterative units.
[0041] In some embodiments, during each iteration, the differentiable conjugate gradient layer solves for a new contrast source that satisfies the electromagnetic state equation (Equation 1 above) based on the current contrast function and the contrast source, combined with the in-domain Green's function matrix and the incident electric field. Thus, by introducing the conjugate gradient method, the solution accuracy can be improved while maintaining computational efficiency, thereby approximating the scattering characteristics in real physical scenarios. Subsequently, the data consistency layer processes the new contrast source using the scattering data equation (Equation 2 above) and updates the contrast function to obtain a new contrast function.
[0042] In some embodiments, the differentiable conjugate gradient layer specifically performs the following operations: based on the current contrast function and with the current contrast source as the initial value for iteration, the conjugate gradient algorithm is expanded into a network layer with a fixed number of iteration steps; wherein each layer of the network layer includes vector inner product, matrix-vector multiplication and linear combination operations, and all operations are differentiable; the electromagnetic state equation is solved using the network layer to obtain a new contrast source.
[0043] In some embodiments, the data consistency layer updates the contrast function according to the following formula 3: Formula 3; in, Let i be the contrast function of the i-th pixel. To compare the components of the source vector J at the i-th pixel, For the incident electric field, For the predefined domain Green's function matrix, As a new source of comparison, It is a constant.
[0044] In some embodiments, to enable end-to-end training of the differentiable conjugate gradient layer, all operations in the differentiable conjugate gradient layer are continuously differentiable elementary operations. Specifically, each iteration of the conjugate gradient algorithm includes only the following types of operations: vector inner product operations (such as dot product), matrix-vector multiplication operations (such as matrix-vector multiplication), and linear combination operations (such as vector addition and scalar multiplication). Since these operations are all basic algebraic operations, their composite functions remain differentiable throughout the iteration process. Therefore, when the conjugate gradient algorithm is expanded into a network layer with a fixed number of iterations, the entire computation graph will be completely differentiable. In practical applications, the automatic differentiation mechanism of deep learning frameworks (such as PyTorch or TensorFlow) can be used. This allows the deep learning framework to automatically track all operations performed during forward propagation and construct a computational graph. During backpropagation, the deep learning framework can automatically calculate gradients according to the chain rule, eliminating the need for manual derivation or writing of gradient code. The gradients are propagated back from the loss function to the network input (i.e., the contrast function), thus achieving end-to-end optimization of the network parameters. This allows physical equations (i.e., electromagnetic state equations) to be seamlessly embedded as differentiable modules into deep learning networks, preserving the rigor of physical constraints while gaining the flexibility of data-driven training.
[0045] In some embodiments, the target physical embedding reconstruction subnetwork can utilize a deep unfolding architecture to progressively restore the compressed low-dimensional nonlinear measurements into a high-resolution image corresponding to the area to be imaged. For example, in a building collapse rescue scenario, the high-resolution image obtained after multiple iterations can clearly display the position and orientation information of living beings behind the wall, while satisfying the physical constraints of electromagnetic wave propagation. This not only improves imaging resolution but also enhances adaptability to complex scenarios.
[0046] In some embodiments, the nonlinear depth-sensing microwave imaging method provided in this application further includes: acquiring a training dataset; wherein the training dataset includes high-dimensional echo signal samples and real image samples corresponding to the high-dimensional echo signal samples; constructing an initial nonlinear depth-sensing subnetwork and an initial physical embedding reconstruction subnetwork, and cascading the initial nonlinear depth-sensing subnetwork and the initial physical embedding reconstruction subnetwork; iteratively updating the parameters of the cascaded initial nonlinear depth-sensing subnetwork and the initial physical embedding reconstruction subnetwork through a joint loss function until the cascaded initial nonlinear depth-sensing subnetwork and the initial physical embedding reconstruction subnetwork converge, thereby obtaining a target nonlinear depth-sensing subnetwork and a target physical embedding reconstruction subnetwork; wherein the joint loss function includes a mean squared error loss function and a physical consistency loss function.
[0047] In some embodiments, various imaging scenarios can be constructed using electromagnetic simulation software. Then, by setting different target shapes, materials, and distribution locations, corresponding high-dimensional echo signal samples and their corresponding real image samples can be generated. Alternatively, high-dimensional echo signals from different scenarios can be directly acquired in a real environment, and the corresponding real images can be recorded simultaneously. This application does not limit the method for obtaining the training dataset.
[0048] In some embodiments, the initial nonlinear depth sensing subnetwork and the target nonlinear depth sensing subnetwork have the same structure, and the initial physical embedding reconstruction subnetwork and the target physical embedding reconstruction subnetwork have the same structure.
[0049] In some embodiments, the joint loss function is expressed by the following formula 4: Formula 4; in, For the joint loss function, Let the mean squared error loss function be . Let physical consistency loss function be used. This is the balance coefficient.
[0050] In some embodiments, the mean squared error loss function is prior art and will not be described further here. The physical consistency loss function can be expressed by the following formula 5: Formula 5 in, Let physical consistency loss function be used. For comparison source, For contrast function, For the incident electric field, The scattered electric field in the region to be imaged. This is a point-by-point multiplication operation for vectors.
[0051] In some embodiments, high-dimensional echo signal samples are input into a cascaded initial nonlinear depth sensing subnetwork and an initial physical embedding reconstruction subnetwork to obtain a predicted contrast source and a reconstructed image. Then, a first loss value between the reconstructed image and the real image sample is determined using a mean squared error loss function, and a second loss value between the predicted contrast source and the theoretical contrast source is determined using a physical consistency loss function. The theoretical contrast source can be derived from the electromagnetic state equation. The second loss value is then multiplied by a preset balance coefficient, and the product is added to the first loss value to obtain the total loss value. The parameters of the cascaded initial nonlinear depth sensing subnetwork and the initial physical embedding reconstruction subnetwork are adjusted based on the total loss value to gradually optimize the network performance until the cascaded initial nonlinear depth sensing subnetwork and the initial physical embedding reconstruction subnetwork converge, resulting in the target nonlinear depth sensing subnetwork and the target physical embedding reconstruction subnetwork.
[0052] In some embodiments, the training process described above may include a pre-training phase and a joint fine-tuning phase; wherein, in the pre-training phase, the parameters of the initial physical embedding reconstruction subnetwork are fixed, and the initial nonlinear deep perception subnetwork is trained separately; in the joint fine-tuning phase, the nonlinear deep perception subnetwork obtained in the pre-training phase is cascaded with the initial physical embedding reconstruction subnetwork for end-to-end joint training.
[0053] It is understood that the training using the joint loss function in this embodiment can simultaneously take into account the quality and physical consistency of the reconstructed image. This not only improves the generalization ability of the target nonlinear deep perception subnetwork and the target physical embedding reconstruction subnetwork, but also ensures that the generated high-resolution image has higher resolution and accuracy while satisfying physical laws.
[0054] Based on the same inventive concept, this application also provides a nonlinear depth-sensing microwave imaging system for implementing a nonlinear depth-sensing microwave imaging method. The solution provided by this system is similar to the solution described in the above method; therefore, please refer to the limitations of the nonlinear depth-sensing microwave imaging method described above, and will not be repeated here. Specifically, Figure 4 This is a schematic diagram of the structure of a nonlinear depth-sensing microwave imaging system according to an embodiment of this application. Figure 4 As shown, the nonlinear depth-sensing microwave imaging system includes: Transmitting module 410 is used to transmit microwave signals to the area to be imaged via a transmitting antenna; The receiving module 420 is used to receive the corresponding high-dimensional echo signal through the receiving antenna array; The processing module 430 is used to adaptively compress the high-dimensional echo signal into the optimal low-dimensional nonlinear measurement value through a preset target nonlinear depth sensing subnetwork; wherein the dimension of the low-dimensional nonlinear measurement value is smaller than the dimension of the high-dimensional echo signal; and to obtain a high-resolution image by performing nonlinear inverse scattering on the low-dimensional nonlinear measurement value through a preset target physical embedding reconstruction subnetwork.
[0055] In some embodiments, the target nonlinear deep sensing subnetwork includes a nonlinear convolutional block, a global residual fusion module, and an adaptive nonlinear combination layer; the processing module 430 is further configured to extract multi-scale features from the high-dimensional echo signal through the nonlinear convolutional block to obtain a multi-scale feature map; wherein the multi-scale feature map includes a shallow feature map, a mid-level feature map, and a deep feature map; the mid-level feature map and the deep feature map are fused across layers through the global residual fusion module to obtain fused features; the fused features are flattened through the adaptive nonlinear combination layer and input into a fully connected layer, and nonlinear dynamic weighting is performed using a learnable parameterized activation function to obtain a low-dimensional nonlinear measurement value.
[0056] In some embodiments, a nonlinear convolutional block is composed of multiple convolutional layers and multiple nonlinear activation function layers stacked alternately, with a batch normalization layer connected after each nonlinear activation function; wherein, the convolutional layers are used to perform convolution operations on the input signal to extract local features at different scales; the nonlinear activation function layers are used to perform nonlinear mapping on the output of the convolutional layers; and the batch normalization layer is used to normalize the activated feature map output by the nonlinear activation function.
[0057] In some embodiments, the target physical embedding reconstruction subnetwork includes an initialization subnetwork and a depth unfolding iteration module; the processing module 430 is further configured to map low-dimensional nonlinear measurement values to a contrast function and a contrast source through the initialization subnetwork; and to perform iterative processing on the contrast function and the contrast source through the depth unfolding iteration module until the number of iterations reaches a preset number of iterations, thereby obtaining a high-resolution image that satisfies the electromagnetic state equation.
[0058] In some embodiments, the depth unfolding iteration module is composed of multiple alternating stacks of differentiable conjugate gradient layers and data consistency layers; the processing module 430 is further configured to iteratively solve the contrast function and the contrast source through the differentiable conjugate gradient layer to obtain a new contrast source that satisfies the electromagnetic state equation; wherein the electromagnetic state equation is constructed based on the in-domain Green's function matrix and the incident field; the contrast function is updated through the data consistency layer according to the scattering data equation and the new contrast source to obtain a new contrast function; the new contrast source and the new contrast function are used as inputs for the next iteration, and the operations corresponding to the differentiable conjugate gradient layer and the data consistency layer are repeatedly executed until the number of iterations reaches the preset number of iterations to obtain a high-resolution image.
[0059] In some embodiments, the processing module 430 is further configured to expand the conjugate gradient algorithm into a network layer with a fixed number of iteration steps, based on the current contrast function and with the current contrast source as the initial value for iteration; wherein each layer of the network layer includes vector inner product, matrix-vector multiplication and linear combination operations, and all operations are differentiable; and to use the network layer to solve the electromagnetic state equation to obtain a new contrast source.
[0060] In some embodiments, the data consistency layer updates the contrast function according to the following formula: Formula 3 in, Let i be the contrast function of the i-th pixel. To compare the components of the source vector J at the i-th pixel, For the incident electric field, For the predefined domain Green's function matrix, As a new source of comparison, It is a constant.
[0061] In some embodiments, the nonlinear depth-sensing microwave imaging system further includes an acquisition module for acquiring a training dataset; wherein the training dataset includes high-dimensional echo signal samples and real image samples corresponding to the high-dimensional echo signal samples; The building module is used to construct the initial nonlinear depth perception subnetwork and the initial physical embedding reconstruction subnetwork, and to cascade the initial nonlinear depth perception subnetwork and the initial physical embedding reconstruction subnetwork. The training module is used to iteratively update the parameters of the cascaded initial nonlinear deep sensing subnetwork and the initial physical embedding reconstruction subnetwork using a joint loss function until the cascaded initial nonlinear deep sensing subnetwork and the initial physical embedding reconstruction subnetwork converge, thus obtaining the target nonlinear deep sensing subnetwork and the target physical embedding reconstruction subnetwork; wherein the joint loss function includes the mean squared error loss function and the physical consistency loss function.
[0062] It should be understood that the phrase "one embodiment" or "an embodiment" throughout the specification means that a specific feature, structure, or characteristic related to the embodiment is included in at least one embodiment of the invention. Therefore, "in one embodiment" or "in an embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. It should be understood that in the various embodiments of the invention, the sequence numbers of the above-described processes do not imply a sequential order of execution; the execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the invention. The sequence numbers of the above-described embodiments of the invention are merely descriptive and do not represent the superiority or inferiority of the embodiments.
[0063] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.
[0064] In the several embodiments provided by this invention, it should be understood that the disclosed methods can be implemented in other ways. The methods disclosed in the several method embodiments provided by this invention can be arbitrarily combined without conflict to obtain new method embodiments. The features disclosed in the several method embodiments provided by this invention can be arbitrarily combined without conflict to obtain new method embodiments.
[0065] The above description is merely an embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.
Claims
1. A nonlinear depth-sensing microwave imaging method, characterized in that, include: Microwave signals are transmitted to the area to be imaged via a transmitting antenna, and the corresponding high-dimensional echo signals are received via a receiving antenna array. The high-dimensional echo signal is adaptively compressed into the optimal low-dimensional nonlinear measurement value through a pre-defined target nonlinear deep sensing subnetwork; wherein the dimension of the optimal low-dimensional nonlinear measurement value is smaller than the dimension of the high-dimensional echo signal. By using a pre-defined target physical embedding reconstruction subnetwork to perform nonlinear inverse scattering on low-dimensional nonlinear measurements, a high-resolution image corresponding to the region to be imaged is obtained.
2. The method according to claim 1, characterized in that, The target nonlinear deep perception subnetwork includes nonlinear convolutional blocks, a global residual fusion module, and an adaptive nonlinear combination layer; The target nonlinear deep sensing subnetwork is specifically used to perform the following operations: Multi-scale feature extraction is performed on high-dimensional echo signals using nonlinear convolutional blocks to obtain multi-scale feature maps; these multi-scale feature maps include shallow feature maps, mid-level feature maps, and deep feature maps. The global residual fusion module performs cross-layer fusion of the mid-layer feature map and the deep feature map to obtain the fused feature. The fused features are flattened by an adaptive nonlinear combination layer and then input into a fully connected layer. A learnable parameterized activation function is used for nonlinear dynamic weighting to obtain low-dimensional nonlinear measurement values.
3. The method according to claim 2, characterized in that, A nonlinear convolutional block consists of multiple convolutional layers and multiple nonlinear activation function layers stacked alternately, with a batch normalization layer connected after each nonlinear activation function. The convolutional layers are used to perform convolution operations on the input signal to extract local features at different scales; the nonlinear activation function layers are used to perform nonlinear mapping on the output of the convolutional layers; and the batch normalization layer is used to normalize the activated feature map output by the nonlinear activation function.
4. The method according to claim 1, characterized in that, The target physical embedding reconstruction subnetwork includes an initialization subnetwork and a deep unrolling iteration module; The target physical embedding reconstruction subnetwork specifically performs the following operations: Low-dimensional nonlinear measurements are mapped to contrast functions and contrast sources by initializing subnetworks; The contrast function and the contrast source are iteratively processed by the depth expansion iteration module until the preset number of iterations is reached, so as to obtain a high-resolution image that satisfies the electromagnetic state equation.
5. The method according to claim 4, characterized in that, The deep unrolling iterative module consists of multiple alternating stacks of differentiable conjugate gradient layers and data consistency layers; The depth unrolling iteration module performs the following operations: By iteratively solving the contrast function and the contrast source through a differentiable conjugate gradient layer, a new contrast source satisfying the electromagnetic state equation is obtained; where the electromagnetic state equation is constructed based on the in-domain Green's function matrix and the incident field. The contrast function is updated by the data consistency layer based on the scattering data equation and the new contrast source to obtain a new contrast function; Using the new contrast source and the new contrast function as input for the next iteration, the operations corresponding to the differentiable conjugate gradient layer and the data consistency layer are repeatedly executed until the preset number of iterations is reached, resulting in a high-resolution image.
6. The method according to claim 5, characterized in that, Differentiable conjugate gradient layers specifically perform the following operations: Based on the current contrast function, and using the current contrast source as the initial value for iteration, the conjugate gradient algorithm is expanded into a network layer with a fixed number of iteration steps; each layer of the network layer includes vector inner product, matrix-vector multiplication, and linear combination operations, and all operations are differentiable; By using network layers to solve the electromagnetic state equations, a new contrast source is obtained.
7. The method according to claim 5, characterized in that, The data consistency layer updates the contrast function according to the following formula: in, Let i be the contrast function of the i-th pixel. To compare the components of the source vector J at the i-th pixel, For the incident electric field, For the predefined domain Green's function matrix, As a new source of comparison, It is a constant.
8. The method according to claim 1, characterized in that, The method further includes: Obtain the training dataset; the training dataset includes high-dimensional echo signal samples and the corresponding real image samples; Construct an initial nonlinear deep sensing subnetwork and an initial physical embedding reconstruction subnetwork, and then cascade the initial nonlinear deep sensing subnetwork and the initial physical embedding reconstruction subnetwork. The parameters of the cascaded initial nonlinear deep sensing subnetwork and the initial physical embedding reconstruction subnetwork are iteratively updated using a joint loss function until the cascaded initial nonlinear deep sensing subnetwork and the initial physical embedding reconstruction subnetwork converge, thus obtaining the target nonlinear deep sensing subnetwork and the target physical embedding reconstruction subnetwork; the joint loss function includes the mean squared error loss function and the physical consistency loss function.
9. A nonlinear depth-sensing microwave imaging system, characterized in that, include: The transmitting module is used to transmit microwave signals to the area to be imaged via a transmitting antenna; The receiving module is used to receive the corresponding high-dimensional echo signal through the receiving antenna array; The processing module is used to adaptively compress high-dimensional echo signals into optimal low-dimensional nonlinear measurement values through a preset target nonlinear depth sensing subnetwork. The low-dimensional nonlinear measurement values are solved by nonlinear inverse scattering through a pre-defined target physical embedding reconstruction sub-network to obtain a high-resolution image corresponding to the region to be imaged; wherein, the dimension of the optimal low-dimensional nonlinear measurement value is smaller than the dimension of the high-dimensional echo signal.