ISAR imaging method based on low-rank adaptive stable diffusion model

Abstract

The invention discloses an ISAR imaging method based on a low-rank adaptive stable diffusion model, and the method comprises the steps: carrying out the motion compensation, range compression and short-time Fourier transform processing of an ISAR radar echo signal of a rigid target, and generating a low-resolution TFR; inputting the low-resolution TFR into a low-rank adaptive stable diffusion model (LoRA-SD) to carry out TFR super-resolution processing, and improving the frequency resolution by refining the curve characteristics of the TFR to generate a high-resolution TFR; wherein the low-rank adaptive model (LoRA-SD) is obtained by performing fine adjustment on a pre-trained SD-Turbo large model by using a LoRA adapter; and finally, generating a high-resolution ISAR image by using a distance-instantaneous Doppler algorithm. The method breaks through the frequency resolution limit of traditional time-frequency analysis, significantly improves the resolution and quality of ISAR imaging, and has good generalization ability and robustness.

Description

Technical Field

[0001] The present invention belongs to the field of radar imaging technology, and in particular relates to an ISAR imaging method based on a low-rank adaptive stable diffusion model. Background Art

[0002] Inverse Synthetic Aperture Radar (ISAR) imaging has been widely used in both civilian and military applications due to its long-range, all-weather, and all-climate observation capabilities. When target motion is uniform, the Range-Doppler (RD) algorithm effectively achieves high azimuth resolution by using a Fourier transform (FT) after compensating for translational motion. However, in practical applications, target motion is often complex and non-uniform, resulting in temporal variations in the Doppler frequency, making it difficult for the FT-based RD algorithm to accurately estimate the Doppler frequency. To alleviate the imaging ambiguity caused by the time-varying nature of target motion, ISAR imaging methods based on time-frequency analysis (TFA) have been developed. These methods are often referred to as Range-Instantaneous Doppler (RID) methods.

[0003] TFA-based ISAR imaging of rigid targets aims to generate high-resolution time-frequency representations. Existing methods can be roughly divided into two categories. The first is traditional linear TFA methods, represented by the short-time Fourier transform (STFT). While these methods are computationally efficient, they are limited by the uncertainty principle and cannot simultaneously improve both temporal and frequency resolution. To improve TFR quality, advanced linear methods such as sparse Bayesian learning (SBL) and fast mean field (MF) have been introduced. However, their performance relies on the robustness of time-invariant coefficient estimates and is insufficiently robust in complex motion scenarios. The second category is nonlinear TFA methods based on deep learning, such as those based on conditional generative adversarial networks (CGANs) and U-Net architectures. These methods have shown promise in enhancing TFR. However, these methods often face generalization challenges when there is a large distribution gap between training and test radar data.

[0004] In summary, for ISAR imaging of rigid targets, the existing TFA-based ISAR imaging methods have a trade-off between temporal resolution and frequency resolution, resulting in limited resolution. In addition, the existing methods also have problems such as insufficient robustness in complex motion scenes and insufficient generalization on different data. Summary of the Invention

[0005] In order to solve the above problems existing in the prior art, the present invention provides an ISAR imaging method based on a low-rank adaptive stable diffusion model. The technical problem to be solved by the present invention is achieved through the following technical solutions:

[0006] In a first aspect, the present invention proposes an ISAR imaging method based on a low-rank adaptive stable diffusion model, comprising:

[0007] Perform motion compensation processing on the ISAR radar echo signal to obtain a motion compensated signal;

[0008] Perform range compression on the motion-compensated signal to generate an initial range image;

[0009] Perform short-time Fourier transform on the initial range image to generate a low-resolution TFR;

[0010] The low-resolution TFR is input into the low-rank adaptive stable diffusion model LoRA-SD (Low-Rank Adaptation Stable Diffusion) for TFR super-resolution processing. The frequency resolution is improved by refining the TFR curve features to obtain high-resolution TFR. Among them, the low-rank adaptive model LoRA-SD is obtained by fine-tuning the pre-trained SD-Turbo large model using the LoRA adapter.

[0011] The high-resolution TFR is imaged and processed using the range-instantaneous Doppler algorithm to generate the final ISAR image.

[0012] In a second aspect, the present invention proposes an ISAR imaging device based on a low-rank adaptive stable diffusion model, for implementing the method proposed in the first aspect of the present invention, the device comprising:

[0013] The motion compensation module is used to perform motion compensation processing on the ISAR radar echo signal of the rigid target to obtain the motion compensated signal;

[0014] The range compression module is used to perform range compression on the motion-compensated signal to generate an initial range image;

[0015] The Fourier transform module is used to perform short-time Fourier transform on the initial range image to generate a low-resolution TFR;

[0016] The super-resolution module is equipped with a low-rank adaptive stable diffusion model LoRA-SD. The low-rank adaptive stable diffusion model LoRA-SD is used to perform TFR super-resolution processing on low-resolution TFR. By refining the TFR curve features, the frequency resolution is improved to obtain high-resolution TFR. Among them, the low-rank adaptive model LoRA-SD is obtained by fine-tuning the pre-trained SD-Turbo large model using the LoRA adapter.

[0017] The imaging module is used to process the high-resolution TFR imaging using the range-instantaneous Doppler algorithm to generate the final ISAR image.

[0018] In a third aspect, the present invention provides an electronic device comprising a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other via the communication bus;

[0019] Memory for storing computer programs;

[0020] The processor is used to execute the program stored in the memory to implement the method provided by the first aspect of the present invention.

[0021] In a fourth aspect, the present invention provides a computer-readable storage medium, in which a computer program is stored. When the computer program is executed by a processor, the method provided in the first aspect of the present invention is implemented.

[0022] Beneficial effects of the present invention:

[0023] 1. The ISAR imaging method based on the low-rank adaptive stable diffusion model provided by the present invention first performs motion compensation, range compression and short-time Fourier transform processing on the ISAR radar echo signal of the rigid target in sequence to generate a low-resolution TFR; then introduces the low-rank adaptive stable diffusion model LoRA-SD for super-resolution processing, and improves the frequency resolution by refining the TFR curve features to obtain a high-resolution TFR; finally, the range-instantaneous Doppler algorithm is used for imaging to generate the final ISAR image. This method uses the LoRA adapter to fine-tune the pre-trained SD-Turbo large model to obtain the low-rank adaptive stable diffusion model LoRA-SD. Among them, the LoRA adapter effectively alleviates the distribution difference between the optical vision data of the pre-trained SD-Turbo and the ISAR radar data in the TFA task through low-rank decomposition technology, reducing the number of parameters and improving the generalization ability of the model. At the same time, the powerful texture representation capability of SD-Turbo can achieve efficient TFR super-resolution and has high robustness in complex motion scenes. In addition, this method is not constrained by the uncertainty principle, breaks through the frequency resolution limitation of traditional time-frequency analysis, improves the frequency resolution while maintaining the temporal resolution, and significantly improves the resolution and quality of ISAR imaging.

[0024] 2. The low-rank adaptive stable diffusion model LoRA-SD provided by the present invention enhances feature consistency through zero convolution and skip connection, further improving the quality of TFR super-resolution;

[0025] 3. In the process of fine-tuning the pre-trained SD-Turbo large model using the LoRA adapter, the present invention adopts the maximum-minimum game to optimize the model. By optimizing the objective function through adversarial training, the generated model has better performance in improving TFR quality.

[0026] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 1 is a flow chart of an ISAR imaging method based on a low-rank adaptive stable diffusion model provided by an embodiment of the present invention;

[0028] Figure 2 1 is a principle framework diagram of an ISAR imaging method based on a low-rank adaptive stable diffusion model provided by an embodiment of the present invention;

[0029] Figure 3 This is an architecture diagram of the low-rank adaptive stable diffusion model LoRA-SD provided by an embodiment of the present invention;

[0030] Figure 4The experimental results of the method of the present invention and the existing method on simulated radar data are as follows;

[0031] Figure 5 It is the TFR result of the method of the present invention and the existing method on the measured radar data;

[0032] Figure 6 It is the ISAR imaging result of rigid body target in different frames based on measured radar data;

[0033] Figure 7 This is a structural block diagram of an ISAR imaging device based on a low-rank adaptive stable diffusion model provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0034] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0035] The first aspect of the present invention provides an ISAR imaging method based on a low-rank adaptive stable diffusion model. Figure 1 and Figure 2 , Figure 1 is a flow chart of an ISAR imaging method based on a low-rank adaptive stable diffusion model provided by an embodiment of the present invention, Figure 2 : is a schematic diagram of a framework of an ISAR imaging method based on a low-rank adaptive stable diffusion model provided by an embodiment of the present invention. The method mainly includes the following steps:

[0036] Step 1: Perform motion compensation on the ISAR radar echo signal of the rigid target to obtain the motion compensated signal.

[0037] First, the ISAR radar echo signal needs to be collected.

[0038] Specifically, the radar transmits signals to the rigid target through the transmitting antenna and receives the echo signals through the receiving antenna.

[0039] Generally speaking, in the azimuth domain, the radar echo signal of a target at a specific range unit can be modeled as a linear combination of multiple linear frequency modulation (LFM) components, expressed as:

[0040]

[0041] Where s(t) represents the radar echo signal, t represents time, N represents the number of scattering points, and a n represents the complex amplitude of the nth scattering point, f n and k n They represent the center frequency and modulation frequency of the scattering point respectively.

[0042] Then, motion compensation is performed on the received echo signal.

[0043] Specifically, by processing the echo signal through translation and rotation compensation, the influence of target motion can be eliminated and a motion-compensated signal can be obtained.

[0044] The specific implementation of the translation and rotation compensation operations involved in this step can be referred to the existing related technologies, and this embodiment will not be introduced in detail here.

[0045] Step 2: Perform range compression on the motion-compensated signal to generate an initial range image.

[0046] Specifically, the signal obtained in step 1 may be processed according to an existing range compression method to obtain a one-dimensional range image, which is also referred to herein as an initial range image.

[0047] Step 3: Perform short-time Fourier transform on the initial range image to generate a low-resolution TFR.

[0048] The Short Time Fourier Transform (STFT) is a time-frequency analysis method used to analyze the frequency characteristics of a signal that varies over time. The STFT divides the signal into small segments and then performs a Fourier transform on each segment, obtaining the signal's frequency information at different time points.

[0049] Specifically, in this embodiment, in order to estimate the time-varying Doppler frequency, the STFT can be used to calculate the TFR of the signal, and the calculation formula is:

[0050]

[0051] Where STFT(t,ω) represents the short-time Fourier transform operation, t and ω represent time and frequency respectively, and h(·) represents the window function.

[0052] The calculated spectrum is given by S(t,ω)=|STFT(t,ω)| 2 Given, the final time-frequency spectrum is expressed as Where N and M represent the number of discrete times and frequencies, respectively.

[0053] Step 4: Input the low-resolution TFR into the low-rank adaptive stable diffusion model LoRA-SD for TFR super-resolution processing, and improve the frequency resolution by refining the TFR curve features to obtain high-resolution TFR; among them, the low-rank adaptive model LoRA-SD is obtained by fine-tuning the pre-trained SD-Turbo large model using the LoRA adapter.

[0054] Existing TFR methods suffer from a trade-off between temporal and frequency resolution, resulting in limited resolution. To address this issue, this paper proposes a low-rank adaptive stable diffusion model (LoRA-SD) to enhance low-resolution TFR, thereby improving the quality of the final ISAR image.

[0055] Optional, see Figure 3 , Figure 3 This is an architecture diagram of the low-rank adaptive stable diffusion model LoRA-SD provided by an embodiment of the present invention. The low-rank adaptive model LoRA-SD mainly consists of two parts: one is the pre-trained SD-Turbo large model, and the other is the LoRA adapter for fine-tuning.

[0056] TFA aims to meet the demand of upgrading from low-resolution TFR to high-resolution, which corresponds to pixel-level curve refinement. Large-scale visual models excel in perceiving texture features (curve features in TFR, as a special case of texture, are very suitable for this task). Therefore, the present invention uses the SD-Turbo large model to achieve efficient TFR super-resolution through SD-Turbo's powerful texture representation capabilities.

[0057] Specifically, such as Figure 3 As shown in the figure, the pre-trained SD-Turbo large model mainly includes an encoder network (Encoder), a U-Net network, and a decoder network (Decoder) connected in sequence;

[0058] The encoder network is used to extract features from the input low-resolution TFR, compress the information into the feature space, and obtain primary features;

[0059] The U-Net network is used to perform deep processing on the feature representation extracted by the encoder network to enhance the refinement of the TFR curve features;

[0060] The decoder network is used to perform high-resolution TFR reconstruction on the features processed by the U-Net network.

[0061] It should be noted that the U-Net network in this embodiment adopts a conventional encoder and decoder structure, and features are transferred through jump connections in the middle for deep feature extraction and reconstruction.

[0062] LoRA is a parameter-efficient fine-tuning technology that updates the weights of a pre-trained model through low-rank decomposition. In this embodiment, the process of using the LoRA adapter to fine-tune the pre-trained SD-Turbo large model to obtain the low-rank adaptive model LoRA-SD includes:

[0063] Use the pre-trained SD-Turbo large model as the base mapping function and freeze the parameters of the SD-Turbo large model;

[0064] The LoRA adapter is introduced into the encoder network, U-Net network and decoder network of the SD-Turbo large model for fine-tuning; at the same time, skip connections are introduced between the encoder network and the decoder network using zero-convolution. Through adversarial loss optimization training, the low-rank adaptive stable diffusion model LoRA-SD is obtained.

[0065] For more details, please see Figure 3 , where the LoRA adapter is set in each convolutional layer and Transformer layer of the encoder network, U-Net network, and decoder network, and the parameters of each layer of each network are optimized and updated to fine-tune the pre-trained SD-Turbo large model.

[0066] Furthermore, in this embodiment, the pre-trained weight matrix of SD-Turbo On the basis of , the low-rank decomposition update method is used to optimize the update parameters to reduce the amount of calculation and improve the fine-tuning efficiency. The update formula is:

[0067] W0+ΔW=W0+BA;

[0068] Where W0 represents the pre-trained weight matrix of the SD-Turbo large model, ΔW is the parameter update amount, A and B are low-rank matrices, d and k represent the input dimension and output dimension respectively, r represents low rank, and r<<min(d,k).

[0069] Correspondingly, in the adversarial optimization training process, the forward propagation process is expressed as:

[0070] h=W0x+ΔWx=W0x+BAx;

[0071] Where x represents the input of each layer of each network in the SD-Turbo large model, and h represents the output of each layer of each network in the SD-Turbo large model.

[0072] Furthermore, to ensure consistency between input and output and enhance feature transfer, this embodiment also uses a zero convolution method. A zero convolution is a convolution in which both weights and biases are initialized to zero. Using it as a soft skip connection can enhance feature transfer.

[0073] Optional, such as Figure 3 As shown, a soft skip connection is introduced between the encoder and decoder blocks of SD-Turbo. The skip connection transfers multi-scale features from the encoder to the decoder to reduce information loss and improve reconstruction accuracy. The encoder network includes Q encoder blocks, and the decoder network includes Q decoder blocks. Zero convolution connects the i-th encoder block and the Q-i+1-th decoder block to achieve multi-scale feature transfer from the encoder to the decoder. Preferably, Q can be 4.

[0074] To simplify the representation, this embodiment will uniformly record the SD-Turbo combined with the LoRA adapter and zero convolution as G θ,φ(·) , where φ represents the parameters of the LoRA adapter and zero convolution.

[0075] The low-rank adaptive stable diffusion model LoRA-SD provided by the present invention enhances feature consistency through zero convolution and jump connection, further improving the quality of TFR super-resolution.

[0076] Furthermore, during training, the goal of this invention is to optimize LoRA-SD so that it can generate high-resolution TFR as output when inputting low-resolution TFR. Inspired by CycleGAN in pixel-level image translation tasks, this invention optimizes the model through a max-min game.

[0077] Specifically, in the process of obtaining the low-rank adaptive stable diffusion model LoRA-SD through adversarial optimization training, for G φ and D φ The objective function for iterative optimization of trainable parameters is

[0078] Among them, G φ and D φ Denote the generator and adversarial discriminator respectively; discriminator D φ The CLIP model is used as the backbone network, but only its last fully connected layer is fine-tuned; Represents the loss function, which is expressed as:

[0079]

[0080] In the formula, the first term represents the reconstruction loss, which is used to minimize the mean square error between the super-resolution TFR and the true high-resolution TFR, Represents the square of the L2 norm; the second term and the third constitutes a counter-loss; Aims to maximize the discriminator's confidence in the true high-resolution TFR; By minimizing the probability that the generated TFR is misclassified as the true TFR, LoRA-SD generates a more realistic high-resolution TFR; S and Q represent the low-resolution TFR and the corresponding high-resolution TFR, respectively, and α and β represent the weight coefficients.

[0081] In the process of fine-tuning the pre-trained SD-Turbo large model using the LoRA adapter, the present invention adopts the maximum-minimum game to optimize the model, and optimizes the objective function through adversarial training, so that the generated model has better performance in improving TFR quality.

[0082] After the above-mentioned adversarial optimization training, an applicable low-rank adaptive stable diffusion model LoRA-SD was obtained.

[0083] The low-resolution TFR is input into the low-rank adaptive stable diffusion model LoRA-SD for TFR super-resolution processing to obtain high-resolution TFR.

[0084] Step 5: Use the range-instantaneous Doppler algorithm to process the high-resolution TFR imaging to generate the final ISAR image.

[0085] The specific implementation process of the range-instantaneous Doppler algorithm can refer to existing related technologies.

[0086] The ISAR imaging method based on the low-rank adaptive stable diffusion model provided by the present invention first performs motion compensation, range compression and short-time Fourier transform processing on the ISAR radar echo signal of the rigid target in sequence to generate a low-resolution TFR; then introduces the low-rank adaptive stable diffusion model LoRA-SD for super-resolution processing to obtain a high-resolution TFR; finally, uses the range-instantaneous Doppler algorithm for imaging to generate the final ISAR image. This method uses the LoRA adapter to fine-tune the pre-trained SD-Turbo large model to obtain the low-rank adaptive stable diffusion model LoRA-SD; among them, the LoRA adapter effectively alleviates the distribution difference between the optical vision data of the pre-trained SD-Turbo and the ISAR radar data in the TFA task through low-rank decomposition technology, reducing the number of parameters and improving the generalization ability of the model; at the same time, the powerful texture representation capability of SD-Turbo can achieve efficient TFR super-resolution and has high robustness in complex motion scenes; in addition, this method is not constrained by the uncertainty principle, breaks through the frequency resolution limitation of traditional time-frequency analysis, improves the frequency resolution while maintaining the time resolution, and significantly improves the resolution and quality of ISAR imaging.

[0087] The following simulation experiments are used to verify the beneficial effects of the ISAR imaging method based on the low-rank adaptive stable diffusion model proposed in the present invention.

[0088] Experiment 1

[0089] The method of the present invention and the existing STFT, SBL and MF methods were respectively used to conduct experiments on simulated radar data. The results are shown in Fig. Figure 4 shown.

[0090] in, Figure 4 The left figure in the figure shows the frequency estimation error of different methods evaluated on simulated radar data. It can be seen that when the signal-to-noise ratio (SNR) increases from 0dB to 16dB (at intervals of 2dB), the frequency estimation error of the method of the present invention is significantly lower than that of the existing STFT, SBL, and MF methods. This shows that the LoRA-SD architecture proposed in the present invention can effectively enhance noise resistance due to its stronger perception of texture features such as curves. In contrast, SBL and MF rely on solving the TVAR model to calculate TFR, and their performance is greatly affected by the sensitivity of hyperparameter selection.

[0091] Figure 4 The upper right figure in the figure shows the average inference time for each data; it can be seen that the method of the present invention performs well in terms of computational efficiency and can achieve a good balance between accuracy and computational speed.

[0092] Figure 4The lower right figure in the figure shows the number of trainable parameters and memory usage of the method of the present invention. As can be seen, since the LoRA-SD of the present invention has fewer trainable parameters, its memory usage is more controllable, allowing the method to run efficiently even with limited computing resources.

[0093] Experiment 2

[0094] The method of the present invention and the existing STFT, SBL and MF methods were respectively used to conduct experiments on the measured radar data. The results are as follows: Figure 5 and Figure 6 shown.

[0095] Figure 5 is the TFR result of the method of the present invention and the existing method on the measured radar data; wherein, Figure 5 The left figure in the figure shows the TFR of different methods when performing spin and precession analysis on measured radar data at a signal-to-noise ratio (SNR) of 5dB. It can be seen that although the proposed LoRA-SD is only fine-tuned on simulated data, it can still generate high-resolution TFR on measured data of different bandwidths, demonstrating strong generalization capabilities. Figure 5 The right figure in the figure shows the TFR results of the model of the present invention without adversarial training. It can be seen that without adversarial training (i.e., removing the adversarial loss from the loss function), the TFR quality generated by LoRA-SD is significantly reduced, with a larger error. This demonstrates that the objective function based on adversarial training proposed in this paper is crucial for improving TFR quality.

[0096] Figure 6 It is the ISAR imaging result of a rigid target in different frames on the measured radar data; the target motion is spin and precession. Figure 6 The left picture in the figure is the ISAR imaging result of the existing STFT method. Figure 6 The right figure in the figure shows the ISAR imaging results of the proposed method. Comparing ISAR images generated from multiple frames under different motion states, it can be seen that the proposed method has a sharper focus than STFT, with clearer outlines of scattering points and accurate counting. Furthermore, the proposed method can accurately capture the occlusion and reappearance effects of scattering points, demonstrating its potential for application in 3D pose estimation of rigid targets.

[0097] Based on the same inventive concept, the second aspect of the present invention further provides an ISAR imaging device based on a low-rank adaptive stable diffusion model. Figure 7 , Figure 7 : is a structural block diagram of an ISAR imaging device based on a low-rank adaptive stable diffusion model provided by an embodiment of the present invention, the device comprising:

[0098] The motion compensation module is used to perform motion compensation processing on the ISAR radar echo signal of the rigid target to obtain the motion compensated signal;

[0099] The range compression module is used to perform range compression on the motion-compensated signal to generate an initial range image;

[0100] The Fourier transform module is used to perform short-time Fourier transform on the initial range image to generate a low-resolution TFR;

[0101] The spectral super-resolution module is equipped with a low-rank adaptive stable diffusion model LoRA-SD. The low-rank adaptive stable diffusion model LoRA-SD is used to perform TFR super-resolution processing on low-resolution TFR. By refining the TFR curve features, the frequency resolution is improved to obtain high-resolution TFR. Among them, the low-rank adaptive model LoRA-SD is obtained by fine-tuning the pre-trained SD-Turbo large model using the LoRA adapter.

[0102] The imaging module is used to process the high-resolution TFR imaging using the range-instantaneous Doppler algorithm to generate the final ISAR image.

[0103] Based on the same inventive concept, the third aspect of the present invention further provides an electronic device, which includes a processor, a communication interface, a memory, and a communication bus; wherein the processor, the communication interface, and the memory communicate with each other via the communication bus.

[0104] Memory is used to store computer programs;

[0105] When the processor is used to execute the program stored in the memory, the method steps provided in the first aspect of the present invention are implemented.

[0106] Based on the same inventive concept, the fourth aspect of the present invention further proposes a computer-readable storage medium, in which a computer program is stored. When the computer program is executed by a processor, the method steps provided in the first aspect of the present invention are implemented.

[0107] It should be noted that, for the device, electronic device and storage medium embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiments.

[0108] The device, electronic device and storage medium of the embodiments of the present invention are respectively a device, electronic device and storage medium for applying the ISAR imaging method based on the low-rank adaptive stable diffusion model. All embodiments of the above-mentioned ISAR imaging method based on the low-rank adaptive stable diffusion model are applicable to the device, electronic device and storage medium, and can achieve the same or similar beneficial effects.

[0109] The above is a further detailed description of the present invention in conjunction with specific preferred embodiments, and the specific implementation of the present invention should not be considered to be limited to these descriptions. For those skilled in the art of the present invention, without departing from the concept of the present invention, several simple deductions or substitutions can be made, which should be considered to fall within the scope of protection of the present invention.

Claims

1. An ISAR imaging method based on a low-rank adaptive stable diffusion model, characterized in that: include: Perform motion compensation processing on the ISAR radar echo signal of the rigid target to obtain the motion compensated signal; performing range compression on the motion-compensated signal to generate an initial range image; Performing short-time Fourier transform on the initial range image to generate a low-resolution TFR; The low-resolution TFR is input into the low-rank adaptive stable diffusion model LoRA-SD for TFR super-resolution processing, and the frequency resolution is improved by refining the TFR curve features to generate a high-resolution TFR; wherein the low-rank adaptive model LoRA-SD is obtained by fine-tuning the pre-trained SD-Turbo large model using the LoRA adapter; The high-resolution TFR is imaged using a range-instantaneous Doppler algorithm to generate a final ISAR image.

2. The ISAR imaging method based on a low-rank adaptive stable diffusion model according to claim 1, characterized in that: The pre-trained SD-Turbo large model includes an encoder network, a U-Net network, and a decoder network connected in sequence; The encoder network is used to extract curve features from the input low-resolution TFR and generate feature representation; The U-Net network is used to perform in-depth processing on the feature representation extracted by the encoder network to enhance the refinement of the TFR curve features; The decoder network is used to perform high-resolution TFR reconstruction on the features processed by the U-Net network.

3. The ISAR imaging method based on a low-rank adaptive stable diffusion model according to claim 2, characterized in that: The process of using the LoRA adapter to fine-tune the pre-trained SD-Turbo large model to obtain the low-rank adaptive model LoRA-SD includes: Using a pre-trained SD-Turbo large model as a base mapping function, freezing the parameters of the SD-Turbo large model; The LoRA adapter is introduced into the encoder network, U-Net network and decoder network of the SD-Turbo large model for fine-tuning; at the same time, a jump connection is introduced between the encoder network and the decoder network using zero convolution, and the low-rank adaptive stable diffusion model LoRA-SD is obtained through adversarial loss optimization training.

4. The ISAR imaging method based on a low-rank adaptive stable diffusion model according to claim 3, characterized in that: The LoRA adapter is set in each convolutional layer and Transformer layer of the encoder network, the U-Net network, and the decoder network, and the parameters of each layer of each network are optimized and updated to fine-tune the pre-trained SD-Turbo large model.

5. The ISAR imaging method based on a low-rank adaptive stable diffusion model according to claim 4, characterized in that: The parameters of each layer of each network are optimized and updated using the low-rank decomposition update method, and the update formula is: W0+ΔW=W0+BA; Where W0 represents the pre-trained weight matrix of the SD-Turbo large model, ΔW is the parameter update amount, A and B are low-rank matrices, d and k represent the input dimension and output dimension respectively, r represents low rank, and r<<min(d,k); Correspondingly, in the adversarial optimization training process, the forward propagation process is expressed as: h=W0x+ΔWx=W0x+BAx; Where x represents the input of each layer of each network in the SD-Turbo large model, and h represents the output of each layer of each network in the SD-Turbo large model.

6. The ISAR imaging method based on a low-rank adaptive stable diffusion model according to claim 3, characterized in that: The encoder network includes Q encoder blocks, the decoder network includes Q decoder blocks, and the zero convolution connects the i-th encoder block and the Q-i+1-th decoder block to achieve multi-scale feature transfer from the encoder to the decoder, thereby enhancing the refinement of the TFR curve features to improve the frequency resolution.

7. The ISAR imaging method based on a low-rank adaptive stable diffusion model according to claim 3, characterized in that: In the process of obtaining the low-rank adaptive stable diffusion model LoRA-SD through adversarial optimization training, the optimization objective function is Among them, G φ and D φ Represent the generator and adversarial discriminator respectively; Represents the loss function, which is expressed as: In the formula, the first term represents the reconstruction loss, which is used to minimize the mean square error between the super-resolution TFR and the true high-resolution TFR, Represents the square of the L2 norm; the second term and the third constitutes a counter-loss; Aims to maximize the discriminator's confidence in the true high-resolution TFR; By minimizing the probability that the generated TFR is misclassified as the true TFR, LoRA-SD generates a more realistic high-resolution TFR; S and Q represent the low-resolution TFR and the corresponding high-resolution TFR, respectively, and α and β represent the weight coefficients.

8. An ISAR imaging device based on a low-rank adaptive stable diffusion model, used to implement the method according to any one of claims 1 to 7, characterized in that: The device includes: The motion compensation module is used to perform motion compensation processing on the ISAR radar echo signal of the rigid target to obtain the motion compensated signal; A range compression module, configured to perform range compression on the motion-compensated signal to generate an initial range image; A Fourier transform module, configured to perform a short-time Fourier transform on the initial range image to generate a low-resolution TFR; A super-resolution module is configured with a low-rank adaptive stable diffusion model LoRA-SD, which is used to perform TFR super-resolution processing on the low-resolution TFR, improve the frequency resolution by refining the TFR curve features, and generate a high-resolution TFR; wherein the low-rank adaptive model LoRA-SD is obtained by fine-tuning the pre-trained SD-Turbo large model using the LoRA adapter; The imaging module is used to perform imaging processing on the high-resolution TFR using a range-instantaneous Doppler algorithm to generate a final ISAR image.

9. An electronic device, characterized in that: It includes a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other via the communication bus; Memory for storing computer programs; A processor, configured to execute a program stored in a memory to implement the method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, and when the computer program is executed, the method according to any one of claims 1 to 7 can be implemented.

Images