Distance image super-resolution method for large dynamic range target based on transformer-u-net fusion network

Abstract

This invention discloses a range super-resolution method for large dynamic range targets based on a Transformer-U-Net fusion network. It addresses the inherent limitations of existing technologies in handling long-range dependencies and global context modeling, making it difficult to meet the global information integration requirements of high-precision one-dimensional spectral estimation tasks. This method achieves range super-resolution for one-dimensional range images of large dynamic range targets. The method includes: separating the original complex echo signal into real and imaginary parts and concatenating them, while simultaneously performing digital beamforming to obtain initialization data; combining the concatenated data with the initialization data and inputting it into a pre-trained neural network model, which includes upper and lower layers that output first and second complex results, respectively; finally, multiplying the two results element-wise to obtain the one-dimensional range super-resolution result, where the upper layer simulates the numerator of the IAA algorithm, and the lower layer simulates its denominator.

Description

Technical Field

[0001] This invention relates to the field of radar imaging technology, and in particular to a range image super-resolution method for large dynamic range targets based on a Transformer-U-Net fusion network. Background Technology

[0002] Good range resolution allows for the differentiation of different targets at similar distances, enabling further assessment of their attributes and types based on the intensity and characteristics of the radar echoes received. In certain scenarios, the energy difference between two targets to be distinguished can be significant. Traditional resolution methods can easily misidentify weaker targets as noise or allow them to be masked by the sidelobes of stronger targets. If the range resolution fails to meet practical requirements, preventing the radar system from accurately assessing complex real-world scenarios, it could lead to unpredictable consequences. While increasing system bandwidth can directly improve radar range resolution, this method is difficult and costly to implement in hardware, and may even be physically impossible. Therefore, improving radar range resolution without increasing system bandwidth has significant practical application value.

[0003] In the field of radar signal processing, neural networks, especially deep learning models, are triggering a profound paradigm shift. Their core advantage lies in their ability to directly learn highly abstract feature representations from massive and complex radar data, thereby effectively addressing the challenges of nonlinear, non-stationary, and low signal-to-noise ratio environments that are difficult for traditional methods to handle.

[0004] Traditional radar processing workflows, such as pulse compression, moving target detection, and constant false alarm rate (CFAR) detection, heavily rely on signal processing chains and feature engineering designed with expert knowledge. Neural networks, however, can automatically learn robust features highly relevant to target tasks, such as target classification and identification, from raw or shallowly processed radar data, such as time-frequency maps, range-Doppler spectra, and raw echo sequences, through multi-layer nonlinear transformations. This avoids the tediousness and subjectivity of manual feature design, achieving end-to-end optimization from data to decision, and significantly improving the representational power of features and the automation level of the system.

[0005] Radar targets are often obscured by strong ground clutter, sea clutter, and electronic interference. Neural networks, especially deep learning models, trained on massive amounts of data, can implicitly learn the complex statistical characteristics of clutter and interference, and on this basis, achieve effective signal separation and target enhancement. In constant false alarm rate detection and weak target identification tasks, they often demonstrate better detection probabilities and lower false alarm rates than algorithms based on traditional statistical models (such as Gaussian distribution and K-distribution).

[0006] The Iterative Adaptive Approach (IAA) is a high-resolution, non-parametric spectral estimation method widely used in radar, sonar, medical imaging, and astronomical observation for direction-of-arrival estimation and spectral analysis. In signal processing, recovering multiple signal sources with extremely similar frequencies or spatial locations from limited observation data is a common and challenging task. The resolution of the classical periodogram method is limited by the Rayleigh limit, making it difficult to effectively distinguish neighboring signal sources. Parametric methods, such as Multiple Signal Classification (MUSIC) and Estimation of Signal Parameters via Rotational Invariance Techniques (ESPRIT), while offering higher resolution, require prior knowledge of the number of signal sources and are highly sensitive to signal models, such as coherence, limiting their practical applicability.

[0007] As a non-parametric method, IAA does not rely on prior information about the number of signal sources and has no strict requirements on signal coherence. Its core lies in its iterative and adaptive mechanism: it gradually optimizes the signal power spectrum estimation through multiple iterations, and adaptively calculates the weighting matrix based on the current estimation result in each iteration. Its function is similar to the covariance inverse matrix in the Capon method, thereby effectively suppressing interference and noise and significantly improving resolution.

[0008] Previous studies have attempted to combine multilayer perceptrons (MLPs) with U-Net to simulate the IAA algorithm. However, U-Net has inherent limitations in handling long-range dependencies and global context modeling, and is gradually becoming unable to meet the requirements of high-precision one-dimensional spectral estimation tasks for global information integration. Summary of the Invention

[0009] This invention provides a range super-resolution method for large dynamic range targets based on the Transformer-U-Net fusion network. It solves the problem that existing technologies have inherent limitations in handling long-range dependencies and global context modeling, making it difficult to meet the global information integration requirements of high-precision one-dimensional spectral estimation tasks. This invention achieves range super-resolution for one-dimensional range images of large dynamic range targets.

[0010] This invention provides a range super-resolution method for large dynamic range targets based on a Transformer-U-Net fusion network, comprising: The input original complex echo signal is separated into real and imaginary data and spliced ​​together to obtain spliced ​​data. Digital beamforming processing is then performed on the original complex echo signal to obtain initialization data. The spliced ​​data is combined with the initialization data to obtain the combined input data; The combined input data is input into a pre-trained neural network model to obtain a first complex result and a second complex result. The first complex result and the second complex result are multiplied element-wise to obtain a one-dimensional distance super-resolution result. The neural network model includes an upper layer network and a lower layer network. The upper-layer network is used to process the combined input data and output a first complex number result to simulate the molecule of the iterative adaptive algorithm IAA. The lower-level network is used to process the initialization data and output a second complex number result to simulate the denominator of the IAA algorithm.

[0011] In one possible implementation, the training process of the neural network model includes: Construct a training dataset that includes multiple samples with different source distances and source energy differences, where the source distance is between 1 and 2 times the resolution unit and the source energy difference is within 20 dB; The upper and lower layers of the network are iteratively trained using the training dataset. During the training process, a weighted loss function with amplitude and phase separation is used, and the iterative process of IAA is simulated to gradually approach the target output, so as to finally obtain the parameters of the trained network model.

[0012] In one possible implementation, the calculation process of the amplitude- and phase-separated weighted loss function includes: Calculate the mean square error of the amplitude component and the mean square error of the phase component between the network prediction results and the true labels, respectively. Using preset non-negative weighting coefficients and By linearly combining the amplitude error and the phase error, the overall loss function is obtained as follows: ; in, For mean square error of amplitude, This represents the phase mean square error.

[0013] In one possible implementation, the iterative process of simulating IAA during training to gradually approximate the target output and finally obtain the parameters of the trained network model includes: During the training process, training objectives are set in stages, and the results of one iteration of the IAA algorithm are used as labels to train the network and obtain the first set of weight parameters. Based on the first set of weight parameters, the results of the IAA algorithm after two iterations are used as labels to continue training and obtain updated weight parameters. Repeat this process to gradually approximate the result after the IAA algorithm has iterated a preset number of times.

[0014] In one possible implementation, the upper-layer network includes an MLP network with an input layer, a first hidden layer, a second hidden layer, and an output layer; The input layer is used to receive the combined input data; The first hidden layer is used to perform a nonlinear transformation on the combined input data to obtain the first hidden layer output; The second hidden layer is used to perform a nonlinear transformation on the output of the first hidden layer to obtain the output of the second hidden layer; The output layer is used to perform a linear transformation on the output of the second hidden layer to output a real matrix; the first half of the real matrix is ​​used as the real part of the complex number and the second half is used as the imaginary part of the complex number, and the two parts are combined to form an intermediate complex result; the intermediate complex result is processed by digital beamforming to obtain the first complex result.

[0015] In one possible implementation, the first hidden layer has 1024 nodes, the second hidden layer has 512 nodes, and the output layer has 256 nodes.

[0016] In one possible implementation, the lower-layer network includes: An encoder is formed by cascading encoders and A decoder is formed by cascading several decoder layers; among them, and All are integers greater than 1; The encoder and the decoder perform multi-scale feature fusion through symmetric skip connections in the U-Net architecture; The input to the first encoder layer is the initialization data, and the input to the second encoder layer is the initialization data. The output of the encoder layer is the first The input of each encoder layer, From to Integers; No. The output of the decoder layer is the... The input of each decoder layer From to Integers; The skip connection will be the first The output features of the encoder layer are connected to the first encoder layer. The input of the decoder layer, and the... The input features of each decoder layer are concatenated, where From to Integers.

[0017] In one possible implementation, the encoder layer comprises, in sequence: The first multi-head self-attention module is used to model the global dependencies of the input data and obtain feature representations enhanced by global context. The first residual connection and normalization module is used to perform residual connection and normalization on the feature representation of the global context enhancement to obtain intermediate features with stable gradients and numerical normalization. The first feedforward network layer, implemented by a one-dimensional convolutional layer, is used to perform a non-linear transformation on the intermediate features to obtain a non-linear feature mapping. The second residual connection and normalization module is used to perform residual connection and normalization on the nonlinear feature map to obtain the final output features of a single encoder layer.

[0018] In one possible implementation, the decoder layer comprises, in sequence: The second multi-head self-attention module is used to perform self-attention calculation on the output features of the previous layer decoder to obtain the context representation of the decoded sequence itself. The third residual connection and normalization module is used to perform residual connection and normalization on the context representation of the decoded sequence itself to obtain stable self-attention features; The encoder-decoder multi-head attention module is used to perform cross-attention calculation on the stabilized self-attention features and the output features of the corresponding encoder layer to obtain an enhanced feature representation that integrates global and local information. The fourth residual connection and normalization module is used to perform residual connection and normalization on the enhanced feature representation that integrates global and local information to obtain stable cross-attention features; The second feedforward network layer is used to perform a nonlinear transformation on the stabilized cross-attention features to obtain the reconstructed high-resolution feature representation. The fifth residual connection and normalization module is used to perform residual connection and normalization on the reconstructed high-resolution feature representation to obtain the final output features of a single decoder layer. The lower-layer network adopts a symmetrical skip connection structure of U-Net, which skips the output of each encoder layer to the decoder layer of the corresponding resolution, and concatenates it with the input of the encoder-decoder multi-head attention module to serve as the input of the next part.

[0019] One or more technical solutions provided in this invention have at least the following technical effects or advantages: This invention separates the input raw complex echo signal into real and imaginary data and concatenates them to obtain concatenated data. This preserves the complete phase and amplitude information of the complex signal, avoiding the complexity of direct complex number processing, and providing structured real number input for the neural network. Digital beamforming is then applied to the raw complex echo signal to obtain initialization data, achieving preliminary signal resolution and providing stable initialization conditions for subsequent simulation of the IAA algorithm, similar to matrix initialization in traditional IAA, thus improving the algorithm's convergence. Combining the concatenated data with the initialization data yields combined input data. This combination operation is beneficial for fusing the detailed features of the raw echo with the contextual information after preliminary processing, providing rich input for upper-layer networks, enhancing the model's ability to extract signal features, and thus more accurately simulating the molecular part of the IAA algorithm. The combined input data is fed into a pre-trained neural network model to obtain a first complex result and a second complex result. The upper layer of the network processes the combined input data and outputs the first complex result to simulate the numerator of the Iterative Adaptive Algorithm (IAA). This design helps avoid the complex large matrix iterative inversion calculations in IAA, and the forward propagation of the neural network efficiently approximates the numerator result, significantly improving computational efficiency. The lower layer of the network processes the initialization data and outputs the second complex result to simulate the denominator of the IAA algorithm. This step helps to overcome the limitations of traditional U-Net in long-range dependency modeling by utilizing the global attention mechanism and U-shaped jump connections of the Transformer-U-Net fusion architecture, achieving multi-scale feature fusion and improving the ability to distinguish targets with large dynamic ranges. Finally, the first and second complex results are multiplied element-wise to obtain a one-dimensional distance super-resolution result. This multiplication operation helps to fully simulate the mathematical form of the IAA algorithm, directly generating a high-resolution distance image through the neural network output, ensuring the physical consistency of the result in amplitude and phase, and ultimately achieving efficient and accurate super-resolution performance. The overall approach combines the advantages of data-driven and model-driven approaches, effectively solving the problem of global information integration in high-precision one-dimensional spectrum estimation. Attached Figure Description

[0020] Figure 1 A flowchart illustrating the steps of the large dynamic range target range super-resolution method based on Transformer-U-Net fusion network provided in this embodiment of the invention; Figure 2 This is a diagram of the lower-layer network structure in the neural network model provided in this embodiment of the invention; Figure 3 This is a flowchart illustrating the implementation steps of a specific embodiment of the present invention. Detailed Implementation

[0021] 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, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0022] This invention provides a range super-resolution method for large dynamic range targets based on a Transformer-U-Net fusion network, comprising the following steps S101 to S105.

[0023] S101, the input original complex echo signal is separated into real part data and imaginary part data and spliced ​​together to obtain spliced ​​data, and the original complex echo signal is processed by digital beamforming to obtain initialization data; For example, the original complex echo signal Perform two different processing steps. Original complex echo signal. Perform the separation of real and imaginary components and digital beamforming (DBF) operations separately. Separating the real and imaginary components refers to separating the original complex echo signal... The data is separated into real and imaginary parts, with the imaginary part appended to the real part in matrix form. Digital beamforming performs preliminary signal resolution on the original signal, yielding the result... Similar to matrix initialization in the Iterative Adaptive Algorithm (IAA).

[0024] S102, combine the spliced ​​data with the initialization data to obtain the combined input data; For example, since the upper-layer network is a key part of simulating the IAA algorithm, it is necessary to process the original complex echo signal. With initialization data All data is fed into the upper-layer MLP network. Since the network cannot handle complex data, the processed original data, separating the real and virtual components, is appended to the initialization data. It is then sent to the upper-layer MLP network.

[0025] Here, the IAA algorithm formula ;in, Indicates the first The incident angle of the signal; Indicates about direction The guiding vector matrix; Represents the covariance estimation matrix; Indicates the number of iterations; Indicates the first In the next iteration, from the direction The estimated complex amplitude of the signal.

[0026] S103, the combined input data is input into a pre-trained neural network model to obtain a first complex result and a second complex result. The first and second complex results are then multiplied element-wise to obtain a one-dimensional distance super-resolution result. The neural network model includes an upper layer and a lower layer. The upper layer processes the combined input data and outputs the first complex result to simulate the numerator of the Iterative Adaptive Algorithm (IAA). The lower layer processes the initialization data and outputs the second complex result to simulate the denominator of the IAA algorithm.

[0027] Specifically, the training process of a neural network model includes: (1) Construct a training dataset that includes multiple samples with different source distances and source energy differences, wherein the source distance is between 1 and 2 times the resolution unit and the source energy difference is within 20dB; (2) Use the training dataset to iteratively train the upper and lower network layers. During the training process, a weighted loss function with amplitude and phase separation is used, and the iterative process of IAA is simulated during training to gradually approach the target output, and finally the parameters of the trained network model are obtained.

[0028] Here, the calculation process of the weighted loss function with amplitude and phase separation includes: (2.1) Calculate the mean square error of the amplitude component and the mean square error of the phase component of the network prediction result and the true label respectively; (2.2) Using preset non-negative weighting coefficients and By linearly combining the amplitude error and the phase error, the overall loss function is obtained as follows: ; in, For mean square error of amplitude, This represents the phase mean square error.

[0029] Here, during training, the iterative process of IAA is simulated to gradually approximate the target output, ultimately obtaining the parameters of the trained network model, including: (1) During the training process, training objectives are set in stages, and the results after one iteration of the IAA algorithm are used as labels to train the network and obtain the first set of weight parameters. (2) Based on the first set of weight parameters, the results of the IAA algorithm after two iterations are used as labels to continue training and obtain updated weight parameters; Repeat this process to gradually approximate the result after the IAA algorithm has iterated a preset number of times.

[0030] For example, in training the neural network model, this invention employs a weighted loss function designed specifically for the characteristics of complex domain data. This loss function separates the amplitude and phase components of the complex output, calculates their mean square errors separately, and then performs weighted fusion. The specific calculation process is as follows: First, calculate the mean square error between the magnitude components of the network prediction results and the true labels, denoted as . Simultaneously, the mean square error of the phase components of both is calculated separately and denoted as... Subsequently, using two configurable non-negative weighting coefficients... and The two losses are linearly combined into a single loss function, where the weighting coefficients are... and This design adjusts the relative importance of amplitude and phase during the optimization process, tailored to specific task requirements. This allows neural network models to perform differentiated optimization for amplitude and phase errors during training, making it particularly suitable for fields like signal processing that demand high fidelity in complex data, thereby improving the overall reconstruction or regression accuracy of the model.

[0031] The training method of this invention is to first use The initial learning rate is set and flexibly adjusted during the training of the neural network model, with each training round consisting of 100 rounds. During weight training, the IAA iteration process is simulated. First, the network is trained using the result of one IAA iteration as the label to obtain weight 1. Then, based on weight 1, the network is trained again using the result of two IAA iterations as the label to obtain weight 2, and so on, ultimately obtaining weights that approximate the result of 10 IAA iterations. To keep the data load of a single training session within a reasonable range, each round of training is iterated with a fixed batch size of 64.

[0032] Specifically, datasets with different distances and source differences are created to ensure that the distance between sources is between 2 and 1 times, and the energy difference between sources is within 20dB.

[0033] See Figure 2 The upper layers of the neural network model include an MLP network with an input layer, a first hidden layer, a second hidden layer, and an output layer. The input layer receives combined input data. The first hidden layer performs a nonlinear transformation on the combined input data to obtain the first hidden layer output. The second hidden layer performs a nonlinear transformation on the first hidden layer output to obtain the second hidden layer output. The output layer performs a linear transformation on the second hidden layer output to output a real matrix. The first half of the real matrix is ​​used as the real part of the complex number, and the second half is used as the imaginary part of the complex number. These parts are combined to form an intermediate complex result. The intermediate complex result is then processed by digital beamforming to obtain the first complex result.

[0034] Here, the number of nodes in the first hidden layer is 1024, the number of nodes in the second hidden layer is 512, and the number of nodes in the output layer is 256.

[0035] For example, an upper-layer network, MLP, is constructed. This simulates the molecular part of the IAA algorithm, implementing the function of iterative inversion of large matrices. The MLP network consists of an input layer, two hidden layers, and an output layer. The input layer size is twice the original input size. The first hidden layer has a size of 1024, the second hidden layer has a size of 512, and the output layer has a size of 256. The MLP output is a batchsize × 256 matrix. The first 128 columns are used as the real part, and the last 128 columns are used as the imaginary part to synthesize a new 128-column complex matrix. This output is then subjected to digital beamforming to expand the dimension to batchsize × 512, yielding the first complex result.

[0036] Specifically, the lower-layer network includes: An encoder is formed by cascading encoders and A decoder is formed by cascading several decoder layers; among them, and All are integers greater than 1; (1) Multi-scale feature fusion is performed between the encoder and decoder through symmetric skip connections in the U-Net architecture; wherein, the input of the first encoder layer is the initialization data, and the input of the second encoder layer is the initialization data. The output of the encoder layer is the first The input of each encoder layer, From to Integers; (2) No. The output of the decoder layer is the... The input of each decoder layer From to Integers; (3) The skip connection will be the first The output features of the encoder layer are connected to the first encoder layer. The input of the decoder layer, and the... The input features of each decoder layer are concatenated, where From to Integers.

[0037] See Figure 2Here, the encoder layer sequentially includes: a first multi-head self-attention module, used to model the global dependencies of the input data to obtain a feature representation enhanced by the global context; a first residual connection and normalization module, used to perform residual connection and normalization on the feature representation enhanced by the global context to obtain intermediate features with stable gradients and normalized values; a first feedforward network layer, implemented by a one-dimensional convolutional layer, used to perform nonlinear transformation on the intermediate features to obtain a nonlinear feature map; and a second residual connection and normalization module, used to perform residual connection and normalization on the nonlinear feature map to obtain the final output feature of a single encoder layer.

[0038] Here, the decoder layer sequentially includes: a second multi-head self-attention module, used to perform self-attention calculation on the output features of the previous decoder layer to obtain the context representation of the decoded sequence itself; a third residual connection and normalization module, used to perform residual connection and normalization on the context representation of the decoded sequence itself to obtain stable self-attention features; an encoder-decoder multi-head attention module, used to perform cross-attention calculation on the stable self-attention features and the output features of the corresponding encoder layer to obtain an enhanced feature representation that integrates global and local information; and a fourth residual connection and normalization module, used to process the enhanced feature that integrates global and local information. The first layer performs residual connections and normalization to obtain stable cross-attention features; the second feedforward network layer performs nonlinear transformation on the stable cross-attention features to obtain the reconstructed high-resolution feature representation; the fifth residual connection and normalization module performs residual connections and normalization on the reconstructed high-resolution feature representation to obtain the final output features of a single decoder layer; the lower network adopts the symmetric skip connection structure of U-Net, which skips the output of each encoder layer to the corresponding resolution decoder layer, and concatenates it with the input of the encoder-decoder multi-head attention module to serve as the input of the next part.

[0039] For example, a lower-layer network is constructed, which is a fusion of U-Net and Transformer. This network is used to simulate the denominator of the IAA algorithm. The lower-layer network uses the Transformer's encoder and decoder, employing its self-attention mechanism, but still uses the U-shaped structure of the U-Net network. The fully connected layers of the feedback layer in the Transformer are replaced with convolutional layers to save resources.

[0040] The data processing flow of the lower-layer network is as follows: Initialized data is input in the form of a complex matrix with a dimension of batchsize × 512. This data first enters the first-level encoder of the lower-layer network, namely the Transformer encoder layer. Each encoder layer sequentially contains a multi-head self-attention module, a residual connection, and a feedforward network, where the feedforward network is implemented by a one-dimensional convolutional layer. Finally, it passes through a residual connection once more to complete the processing of a single encoder. The above structure is repeated through four encoders, and the final encoded output data still maintains a complex matrix with a dimension of batchsize × 512.

[0041] The data then enters the decoder. The decoder consists of two multi-head self-attention layers and one feedforward network. During decoding, a U-Net skip connection structure is introduced, whereby the output of the corresponding encoder layer is concatenated to the output of the previous decoder layer, and together they serve as the input to the next decoder layer.

[0042] The final output of the lower layer network is a complex matrix with a dimension of batchsize×512.

[0043] Specifically, in step S103, the first complex result and the second complex result are multiplied element-wise to obtain the one-dimensional distance super-resolution result. Here, the upper-layer network result and the lower-layer network result are multiplied element-wise to combine the simulation effects and obtain the final IAA simulation result, thus achieving one-dimensional distance super-resolution.

[0044] See Figure 3 In one specific embodiment of the present invention, the radar system receives a one-dimensional raw complex echo signal as input. Preprocessing is performed on this signal to separate it into real and imaginary data, and the imaginary data is concatenated with the real data to form concatenated data. Simultaneously, digital beamforming processing is performed on the raw complex echo signal to generate initialization data.

[0045] The concatenated data is combined with the initial data to form the combined input data. This combined input data is then fed into the upper layer of a pre-trained neural network, which is a multilayer perceptron. The multilayer perceptron performs a nonlinear transformation on the input data, outputting a real matrix. The first half of this real matrix is ​​analyzed into its real part, and the second half into its imaginary part. These are combined to form a complex intermediate result, which is then processed by digital beamforming to finally generate the first complex result.

[0046] Initialized data is fed in parallel into the lower layers of the pre-trained neural network. The lower layers employ a fusion of Transformer and U-Net architecture, containing an encoder and a decoder. The encoder consists of multiple cascaded encoder layers, each containing a multi-head self-attention module and a feedforward network. The decoder consists of multiple cascaded decoder layers, fusing feature information from each encoder layer via skip connections. The lower layers ultimately output the second complex result.

[0047] The first complex result and the second complex result are multiplied element-wise, and the product is the final one-dimensional distance super-resolution result. This process simulates the calculation and fusion of the numerator and denominator in an iterative adaptive algorithm.

[0048] The neural network model is trained using a specific strategy. When constructing the training dataset, it is ensured that the source distance in the samples is between 1 and 2 times the resolution unit, and the source energy difference is within 20 dB. During training, a weighted loss function with amplitude and phase separation is applied, and the iterative process of an iterative adaptive algorithm is simulated to gradually approach the target output. Training is performed with a fixed batch size, and the initial learning rate is set to 1×10⁻⁶. - ³, and dynamically adjusted based on the performance on the validation set, with a total of 100 training rounds.

[0049] This method effectively simulates the function of traditional iterative adaptive algorithms through a neural network architecture, avoids complex matrix inversion operations, significantly improves computational efficiency, and at the same time ensures high resolution for targets with a large dynamic range.

[0050] To address the issues of complex and computationally intensive large matrix inversion calculations in traditional IAA algorithms, and the inherent limitations of existing algorithms like U-Net in handling long-range dependencies and global context modeling, which are increasingly failing to meet the global information integration requirements of high-precision one-dimensional spectral estimation tasks, this invention proposes a method combining a neural network integrating Transformer and U-Net with traditional digital beamforming. This method technically simulates the IAA algorithm, achieving one-dimensional range super-resolution. The fusion of U-Net and Transformer captures global semantics and provides local details. The U-Net structure also alleviates the gradient vanishing problem during Transformer training, thus overcoming the inherent limitations of U-Net in handling long-range dependencies and global context modeling, which are increasingly failing to meet the global information integration requirements of high-precision one-dimensional spectral estimation tasks. The loss function used is a weighted loss function based on amplitude and phase separation, ensuring the physical meaning of the signal and allowing for weight adjustment according to application requirements. Replacing the fully connected layers of the Transformer with convolutional details significantly improves model flexibility, enhances the model's ability to handle variable-length inputs, and strengthens the flexibility of multi-scale inference.

[0051] The various embodiments described in this specification are presented in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on its differences from other embodiments. All or part of this invention can be used in numerous general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, mobile communication terminals, multiprocessor systems, microprocessor-based systems, programmable electronic devices, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices, etc.

[0052] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the present invention.

Claims

1. A super-resolution method for range images of large dynamic range targets based on a Transformer-U-Net fusion network, characterized in that, include: The input original complex echo signal is separated into real and imaginary data and spliced ​​together to obtain spliced ​​data. Digital beamforming processing is then performed on the original complex echo signal to obtain initialization data. The spliced ​​data is combined with the initialization data to obtain the combined input data; The combined input data is input into a pre-trained neural network model to obtain a first complex result and a second complex result. The first complex result and the second complex result are multiplied element-wise to obtain a one-dimensional distance super-resolution result. The neural network model includes an upper layer network and a lower layer network. The upper-layer network is used to process the combined input data and output a first complex number result to simulate the molecule of the iterative adaptive algorithm IAA. The lower-level network is used to process the initialization data and output a second complex number result to simulate the denominator of the IAA algorithm.

2. The large dynamic range target range super-resolution method based on Transformer-U-Net fusion network according to claim 1, characterized in that, The training process of the neural network model includes: Construct a training dataset that includes multiple samples with different source distances and source energy differences, where the source distance is between 1 and 2 times the resolution unit and the source energy difference is within 20 dB; The upper and lower layers of the network are iteratively trained using the training dataset. During the training process, a weighted loss function with amplitude and phase separation is used, and the iterative process of IAA is simulated to gradually approach the target output, so as to finally obtain the parameters of the trained network model.

3. The range image super-resolution method for large dynamic range targets based on Transformer-U-Net fusion network according to claim 2, characterized in that, The calculation process of the amplitude and phase separated weighted loss function includes: Calculate the mean square error of the amplitude component and the mean square error of the phase component between the network prediction results and the true labels, respectively. Using preset non-negative weighting coefficients and By linearly combining the amplitude error and the phase error, the overall loss function is obtained as follows: ； in, For mean square error of amplitude, This represents the phase mean square error.

4. The large dynamic range target range super-resolution method based on Transformer-U-Net fusion network according to claim 2, characterized in that, The iterative process of simulating IAA during training to gradually approximate the target output, ultimately obtaining the parameters of the trained network model, includes: During the training process, training objectives are set in stages, and the results of one iteration of the IAA algorithm are used as labels to train the network and obtain the first set of weight parameters. Based on the first set of weight parameters, the results of the IAA algorithm after two iterations are used as labels to continue training and obtain updated weight parameters. Repeat this process to gradually approximate the result after the IAA algorithm has iterated a preset number of times.

5. The range image super-resolution method for large dynamic range targets based on Transformer-U-Net fusion network according to claim 1, characterized in that, The upper-layer network includes an MLP network with an input layer, a first hidden layer, a second hidden layer, and an output layer; The input layer is used to receive the combined input data; The first hidden layer is used to perform a nonlinear transformation on the combined input data to obtain the first hidden layer output; The second hidden layer is used to perform a nonlinear transformation on the output of the first hidden layer to obtain the output of the second hidden layer; The output layer is used to perform a linear transformation on the output of the second hidden layer to output a real matrix; The first half of the real number matrix is ​​used as the real part of the complex number, and the second half is used as the imaginary part of the complex number. These are combined to form an intermediate complex number result. The intermediate complex number result is then subjected to digital beamforming processing to obtain the first complex number result.

6. The range image super-resolution method for large dynamic range targets based on the Transformer-U-Net fusion network according to claim 5, characterized in that, The first hidden layer has 1024 nodes, the second hidden layer has 512 nodes, and the output layer has 256 nodes.

7. The range image super-resolution method for large dynamic range targets based on Transformer-U-Net fusion network according to claim 1, characterized in that, The lower-layer network includes: An encoder is formed by cascading encoders and A decoder is formed by cascading several decoder layers; among them, and All are integers greater than 1; The encoder and the decoder perform multi-scale feature fusion through symmetric skip connections in the U-Net architecture; The input to the first encoder layer is the initialization data, and the input to the second encoder layer is the initialization data. The output of the encoder layer is the first The input of each encoder layer, From to Integers; No. The output of the decoder layer is the... The input of each decoder layer From to Integers; The skip connection will be the first The output features of the encoder layer are connected to the first encoder layer. The input of the decoder layer, and the... The input features of each decoder layer are concatenated, where From to Integers.

8. The range image super-resolution method for large dynamic range targets based on the Transformer-U-Net fusion network according to claim 7, characterized in that, The encoder layer comprises, in sequence: The first multi-head self-attention module is used to model the global dependencies of the input data and obtain feature representations enhanced by global context. The first residual connection and normalization module is used to perform residual connection and normalization on the feature representation of the global context enhancement to obtain intermediate features with stable gradients and numerical normalization. The first feedforward network layer, implemented by a one-dimensional convolutional layer, is used to perform a non-linear transformation on the intermediate features to obtain a non-linear feature mapping. The second residual connection and normalization module is used to perform residual connection and normalization on the nonlinear feature map to obtain the final output features of a single encoder layer.

9. The range image super-resolution method for large dynamic range targets based on the Transformer-U-Net fusion network according to claim 7, characterized in that, The decoder layer comprises, in sequence: The second multi-head self-attention module is used to perform self-attention calculation on the output features of the previous layer decoder to obtain the context representation of the decoded sequence itself. The third residual connection and normalization module is used to perform residual connection and normalization on the context representation of the decoded sequence itself to obtain stable self-attention features; The encoder-decoder multi-head attention module is used to perform cross-attention calculation on the stabilized self-attention features and the output features of the corresponding encoder layer to obtain an enhanced feature representation that integrates global and local information. The fourth residual connection and normalization module is used to perform residual connection and normalization on the enhanced feature representation that integrates global and local information to obtain stable cross-attention features; The second feedforward network layer is used to perform a nonlinear transformation on the stabilized cross-attention features to obtain the reconstructed high-resolution feature representation. The fifth residual connection and normalization module is used to perform residual connection and normalization on the reconstructed high-resolution feature representation to obtain the final output features of a single decoder layer. The lower-layer network adopts a symmetrical skip connection structure of U-Net, which skips the output of each encoder layer to the decoder layer of the corresponding resolution, and concatenates it with the input of the encoder-decoder multi-head attention module to serve as the input of the next part.

Images