Space target radar image data generation method based on low-light enhancement correction

Abstract

The application relates to a space target radar image data generation method based on low-light enhancement correction. The method comprises the following steps: designing a content transfer decomposition network, decomposing low-light features into illumination-independent reflection components and adaptive illumination components through Hadamard product constraints in a latent space, proposing a learnable intensity compression function, dynamically generating a latent space mask through gradient consistency constraints and sparse regularization, embedding the mask into a reverse denoising process to realize directional repair of degradation areas, combining the generation ability of a diffusion model and the guidance of a physical prior, reconstructing a diffusion path based on a bidirectional diffusion principle, generating the physical consistency of the process through endpoint binding constraints, adaptively adjusting the diffusion step, introducing a cyclic implicit iteration mechanism, gradually refining the generation result through a recurrent neural network module, and realizing accurate optical-ISAR image translation.

Description

Technical Field

[0001] This application relates to the fields of computer vision and image processing technology, and in particular to a method for generating radar image data of aerospace targets based on low-light enhancement correction. Background Technology

[0002] With the rapid development of computer vision technology and deep learning methods, image processing has gradually become a cutting-edge research hotspot in aerospace target monitoring. Especially under low-light conditions, optical images of aerospace targets often suffer from overexposure, extreme darkness, and insufficient contrast due to environmental factors, leading to degraded image quality and affecting target identification, analysis, and monitoring. On the other hand, ISAR, as an advanced radar imaging technology, can provide high-resolution images, especially with significant advantages in complex environments. However, the feature differences between optical and ISAR images, as well as the impact of low light on image quality, increase the challenges of image generation. While existing low-light image enhancement and restoration methods can improve image quality to some extent, few methods can simultaneously achieve both low-light image enhancement and efficient conversion from optical to ISAR images. Summary of the Invention

[0003] Therefore, it is necessary to provide a method for generating aerospace target radar image data based on low-light enhancement correction that can achieve accurate optical-ISAR image translation, addressing the aforementioned technical problems.

[0004] A method for generating radar image data of aerospace targets based on low-light enhancement correction, the method comprising:

[0005] Optical images of aerospace targets under low-light conditions are acquired, and the optical images are input into a preset encoder. The encoder maps the optical images to the latent space to obtain the low-light features of the latent space. An image data generation model is constructed. The image data generation model includes a low-light image enhancement module and an image generation module. The low-light image enhancement module includes a content transfer decomposition network, a learnable intensity compression function, a mask-guided diffusion module, and an adaptive enhancement submodule.

[0006] In the low-light image enhancement module, a content delivery decomposition network is designed based on Retinex theory. The latent space low-light features are input into the content delivery decomposition network, and the latent space low-light features are decomposed into illumination-independent reflection components and adaptive illumination components through Hadamard product constraints.

[0007] The adaptive illumination component is input into a learnable intensity compression function. Combined with gradient consistency constraints and sparsity regularization, a latent space mask is dynamically generated. The latent space mask is used to locate the low-illumination degradation region in the optical image.

[0008] The reflection component, adaptive illumination component, and latent space mask are input into the mask-guided diffusion module. The latent space mask is embedded into the inverse denoising process to perform targeted repair on the low-illuminance degradation region, and the repaired latent space features are obtained.

[0009] The repaired latent space features are decoded into a repaired optical image and input into the adaptive enhancement submodule. Through spatial adaptive brightness curve adjustment and residual illumination estimation correction, the enhanced optical image under non-uniform brightness scene is obtained and re-encoded into enhanced latent space features.

[0010] An image generation module is constructed based on the principle of bidirectional diffusion. The enhanced latent space features are input into the image generation module. The physical consistency of the generation process is ensured by endpoint binding constraints. The diffusion step size is adaptively adjusted, and a cyclic implicit iteration mechanism is introduced. The generation result is gradually refined through a recurrent neural network module to obtain ISAR radar images of aerospace targets.

[0011] The aforementioned method for generating aerospace target radar image data based on low-light enhancement correction addresses the aforementioned challenges by innovatively integrating Retinex theory, adaptive masking mechanisms, and diffusion models through a low-light image enhancement module. This approach addresses the problem from both mathematical constraints and generative learning perspectives. First, a content transitive decomposition network is designed to decompose low-light features into illumination-independent reflection and adaptive illumination components in the latent space using Hadamard product constraints, avoiding noise interference from pixel-level decomposition. Second, a learnable intensity compression function is proposed, dynamically generating latent space masks through gradient consistency constraints and sparse regularization to accurately locate degraded regions and suppress redundant computation. Furthermore, a mask-guided diffusion module is constructed, embedding the mask into the inverse denoising process to achieve targeted restoration of degraded regions. Combining the generative capabilities of the diffusion model with the guidance of physical priors significantly improves the quality of detail restoration. Finally, an adaptive enhancement scheme for non-uniform brightness scenes enhances the model's generalization ability. Furthermore, the image generation module is designed to reconstruct the diffusion path based on the bidirectional diffusion principle. By binding endpoints to constrain the physical consistency of the generation process, the diffusion step size is adaptively adjusted. A cyclic implicit iteration mechanism is introduced, and the generation result is gradually refined through a recurrent neural network module, which significantly enhances the determinism of the generation path. This effectively solves the quality improvement problem in the image generation process under low light conditions and enables accurate optical-ISAR image translation. Attached Figure Description

[0012] Figure 1 This is a flowchart illustrating a method for generating radar image data of aerospace targets based on low-light enhancement correction in one embodiment.

[0013] Figure 2 This is a translation result of a target optical-ISAR image from one embodiment;

[0014] Figure 3 This is a diagram showing the results of a quantitative evaluation index for target optical-ISAR image translation in one embodiment. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0016] In one embodiment, such as Figure 1 As shown, a method for generating air and space target radar image data based on low-light enhancement correction is provided, including the following steps:

[0017] Step 102: Acquire optical images of aerospace targets under low-light conditions, input the optical images into a preset encoder, and map the optical images to the latent space through the encoder to obtain the low-light features of the latent space; construct an image data generation model; the image data generation model includes a low-light image enhancement module and an image generation module; the low-light image enhancement module includes a content transfer decomposition network, a learnable intensity compression function, a mask-guided diffusion module, and an adaptive enhancement submodule.

[0018] The low-light image enhancement module combines Retinex theory, diffusion model and adaptive masking mechanism to achieve precise suppression of low-light degradation regions through rigorous mathematical derivation.

[0019] This application innovatively proposes a low-light image enhancement module, which achieves precise noise suppression and detail restoration in low-light regions through latent space decomposition and dynamic mask generation mechanisms. The mask-guided diffusion denoising mechanism enables differentiated processing for regions with varying illumination intensities, significantly improving image restoration and noise suppression capabilities. This module's low-light region processing method effectively reduces redundant computation while preserving high-frequency details, resulting in more refined image restoration and applicability to various complex degradation scenarios.

[0020] Step 104: In the low-light image enhancement module, a content delivery decomposition network is designed based on Retinex theory. The latent space low-light features are input into the content delivery decomposition network, and the latent space low-light features are decomposed into illumination-independent reflection components and adaptive illumination components through Hadamard product constraints.

[0021] Retinex theory applies to images Decomposed into reflectivity (Inherent attributes of the scene) and lighting Hadamard product of (ambient light):

[0022]

[0023] Implementing this decomposition in latent space avoids pixel-level noise interference. Let the encoder... Input image Mapping to latent space features Latent space decomposition is generated through content delivery decomposition networks:

[0024]

[0025] in, , For learnable convolutional kernels, For the Sigmoid function, ensure ， .

[0026] Step 106: The adaptive illumination component is input into the learnable intensity compression function. Combined with gradient consistency constraints and sparsity regularization, a latent space mask is dynamically generated. The low-illumination degradation region in the optical image is located by the latent space mask.

[0027] The mask design goal is to suppress noise in low-light areas while preserving highlight details. A learnable intensity compression function is introduced. :

[0028]

[0029] In the formula, Control the compression steepness, As the intensity threshold, both are obtained through a neural network. Dynamic adjustment. This function is suitable for low-light areas. Nonlinear compression is performed to suppress noise; for areas with normal illumination Maintaining a linear response and avoiding overexposure. Mask optimization constraints include gradient consistency and sparsity regularization. Mask optimization constraints can maintain the consistency of the local structure of the masked region with that of the original illumination.

[0030]

[0031] Sparsity regularity can force mask sparsity, focusing on key degradation regions:

[0032]

[0033] Where TV is the total variation regularization term.

[0034] Step 108: Input the reflection component, adaptive illumination component, and latent space mask into the mask-guided diffusion module, embed the latent space mask into the inverse denoising process, perform targeted repair on the low-illuminance degradation region, and obtain the repaired latent space features.

[0035] The mask-guided diffusion module decouples the generative capability of the diffusion model from the physical degradation features of low-light regions through dynamic guidance of the latent space mask, achieving targeted noise suppression and accurate detail recovery. During training, the model progressively adds noise to the latent space features, simulating the degradation process from a clear image to one contaminated with noise. Unlike traditional diffusion models that operate in pixel space, this method restricts diffusion to the latent space features extracted by the encoder, significantly reducing computational complexity while preserving the structural consistency of high-frequency details. The diffusion process is defined as performing forward diffusion (noise addition) and reverse denoising (signal recovery) in the latent space. The diffusion step size is defined. noise modulation { } Then the forward process:

[0036]

[0037] in, The latent space features are represented in the first The diffusion state of the step.

[0038] In the inverse denoising process, the model assigns spatial weights to low-light regions using pre-generated latent space masks. These masks serve as prior guidance for the attention mechanism, causing the convolutional layers of the denoising network to prioritize high-noise regions (such as shadows) while neglecting areas with high-light illumination.

[0039] The background region remains low-response. This dynamic focusing mechanism not only reduces redundant computation but also effectively mitigates the oversmoothing or artifact problems that may be introduced by diffusion models during global denoising.

[0040]

[0041] in, This represents the denoising network. Furthermore, the model employs differentiated denoising step-size strategies for regions with varying illumination intensities. For severely degraded regions marked by masks, fine-grained multi-step denoising is used to progressively restore texture; for regions with normal illumination, large-step skip-style denoising is employed to accelerate inference. The loss function for the diffusion process is defined as:

[0042]

[0043] By combining Retinex theory, adaptive masking mechanisms, and diffusion models, the optical image quality of aerospace targets under low-light conditions can be effectively improved, and problems such as noise, blurring, and insufficient contrast caused by low light can be solved. This technology not only enhances the detail reproduction capability of low-light images, but also ensures that image quality meets the needs of target recognition and monitoring in complex environments.

[0044] Step 110: Decode the repaired latent space features into a repaired optical image, input it into the adaptive enhancement submodule, and obtain the enhanced optical image under non-uniform brightness scene through spatial adaptive brightness curve adjustment and residual illumination estimation correction, and re-encode it into enhanced latent space features.

[0045] In non-uniform lighting scenes, traditional overall enhancement methods often result in overexposure in local areas or continued blinding in dark areas. To address this issue, this paper proposes a scheme combining spatial adaptive brightness mapping and residual illumination estimation—the basic framework of adaptive enhancement technology for non-uniform lighting scenes. This framework includes: First, in terms of brightness mapping, a spatial adaptive brightness curve is introduced for normalized brightness input. The enhancement result is:

[0046]

[0047] Among them, parameters As location changes, the output is predicted by a lightweight prediction network and then... The constraint that the function is non-negative ensures that the overall curve is monotonically increasing and avoids brightness reversal. This form can be viewed as a power-law transformation. The expansion of the local enhancement index makes the local enhancement index It can adaptively adjust according to the regional lighting conditions: smaller values ​​are used in dark areas to enhance the image, and larger values ​​are used in bright areas to suppress overexposure. To ensure naturalness, the parameter field is adjusted. Introducing smoothing regularization:

[0048]

[0049] At the same time, regularization is reduced in areas with strong image gradients, allowing for more flexible adjustments at edges. This avoids stripe artifacts in smooth areas and maintains structural clarity in detailed areas.

[0050] For illumination estimation, a residual correction strategy is adopted. The initial illumination map is obtained by filtering and histogram equalization. Then, the residual term is predicted through the learning module. The final illumination estimate is obtained as follows:

[0051]

[0052] This design leverages fast initial values ​​to ensure efficiency and compensates for deficiencies under complex lighting conditions through residual learning. Therefore, images can be recovered using a Retinex model.

[0053]

[0054] in To prevent stable terms with a denominator of zero, another enhancement result is obtained using a spatially adaptive brightness curve. The final result was obtained through weighted fusion:

[0055]

[0056] Among them, weight The illumination is determined by local brightness and confidence maps. Curve mapping is relied upon more in noisy or extremely dark areas, while the Retinex path is used more frequently in areas where illumination estimation is reliable. Through the aforementioned design of local adaptive curves + residual illumination estimation + dual-path fusion, this application maintains the physical interpretability of the Retinex model while possessing flexible regional adaptability, effectively mitigating overexposure, under-enhancement, and structural distortion problems in non-uniform illumination scenes.

[0057] Step 112: Based on the principle of bidirectional diffusion, an image generation module is constructed. The enhanced latent space features are input into the image generation module. The physical consistency of the generation process is ensured by endpoint binding constraints. The diffusion step size is adaptively adjusted, and a cyclic implicit iteration mechanism is introduced. The generation result is gradually refined through a recurrent neural network module to obtain the ISAR radar image of the aerospace target.

[0058] The core advantage of the image generation module lies in ensuring the physical plausibility of the generation process through endpoint binding (path constraints from ISAR domain features to optical domain features). Specifically, the state update formula for the forward diffusion process is:

[0059]

[0060] definition , , Let be the total number of diffusion steps. This design strictly constrains the diffusion path within the linear interpolation interval between and , avoiding the self-label domain shift problem caused by random walks in traditional diffusion models. To enhance the determinism of generation, a deep loop implicit inference module is introduced into the inverse denoising process. The specific implementation is as follows: ,in, It is a cyclic residual network, whose internal structure includes hidden state propagation and residual prediction. Representing the control noise term, the hidden state is propagated through a gated loop unit, transmitting the hidden state of the time step. Capture long-term dependencies:

[0061]

[0062] The output layer of residual prediction predicts the residual correction amount for the current step. :

[0063]

[0064] in, The linear weighting parameter represents the residual prediction layer.

[0065]

[0066] Traditional diffusion models employ linear time-synchronous scheduling ( This leads to insufficient noise injection in the early stages of diffusion and excessively slow convergence speed in the later stages. Therefore, an exponentially decaying scheduling function is designed:

[0067]

[0068] in, The attenuation coefficient controls the nonlinearity of the diffusion rate. At this time, it is approximately linear scheduling, suitable for simple generation tasks; 2 It can rapidly approach the target domain in the early stage of diffusion and make fine adjustments in the later stage, making it suitable for complex cross-mode conversion.

[0069] This application also introduces dynamic variance adjustment. Combined with the scheduling function, the variance of diffused noise is adjusted as follows:

[0070]

[0071] in, The oscillation amplitude coefficient is introduced, and a periodic modulation term is used to enhance the model's ability to generate high-frequency details. To balance computational efficiency and generation quality, this application proposes a gradient-sensitive time-stepping strategy: initial coarse-grained partitioning, dynamic refinement, and iterative convergence. The initial coarse-grained partitioning reduces the total number of steps... Average score Segment, calculate the loss gradient magnitude within each segment; dynamically refine the intervals with larger gradient magnitudes by further subdividing them and increasing the number of diffusion steps; iteratively converge and repeat the above process until the gradient magnitude of all intervals is below a threshold. .

[0072] The image generation module achieves high-precision feature mapping and generation between optical and ISAR images by introducing a bidirectional diffusion principle and a deep cyclic implicit inference mechanism. Endpoint binding constraints ensure physical consistency in the generation process and effectively solve the self-labeling domain offset problem in traditional diffusion models. By adjusting the diffusion step size and introducing dynamic variance adjustment, the module can adaptively adjust the noise injection and denoising speed during the generation process, ensuring high efficiency and high accuracy in image generation. Figure 2 and Figure 3The figure shows the target optical-ISAR image translation results and the target optical-ISAR image translation quantitative evaluation index. This application, through innovative cross-domain conversion and denoising strategies, significantly improves the image quality and identifiability of targets in complex electromagnetic environments. The robustness and efficiency of this technology can be widely applied in many high-demand fields such as civil aviation and satellite monitoring, providing strong technical support for related applications.

[0073] The aforementioned method for generating aerospace target radar image data based on low-light enhancement correction addresses the aforementioned challenges by innovatively integrating Retinex theory, adaptive masking mechanisms, and diffusion models through a low-light image enhancement module. This approach addresses the problem from both mathematical constraints and generative learning perspectives. First, a content transitive decomposition network is designed to decompose low-light features into illumination-independent reflection and adaptive illumination components in the latent space using Hadamard product constraints, avoiding noise interference from pixel-level decomposition. Second, a learnable intensity compression function is proposed, dynamically generating latent space masks through gradient consistency constraints and sparse regularization to accurately locate degraded regions and suppress redundant computation. Furthermore, a mask-guided diffusion module is constructed, embedding the mask into the inverse denoising process to achieve targeted restoration of degraded regions. Combining the generative capabilities of the diffusion model with the guidance of physical priors significantly improves the quality of detail restoration. Finally, an adaptive enhancement scheme for non-uniform brightness scenes enhances the model's generalization ability. Furthermore, the image generation module is designed to reconstruct the diffusion path based on the bidirectional diffusion principle. By binding endpoints to constrain the physical consistency of the generation process, the diffusion step size is adaptively adjusted. A cyclic implicit iteration mechanism is introduced, and the generation result is gradually refined through a recurrent neural network module, which significantly enhances the determinism of the generation path. This effectively solves the quality improvement problem in the image generation process under low light conditions and enables accurate optical-ISAR image translation.

[0074] In one embodiment, a content delivery decomposition network is designed based on Retinex theory. The latent space low-light features are input into the content delivery decomposition network, and the latent space low-light features are decomposed into illumination-independent reflection components and adaptive illumination components through Hadamard product constraints. This includes:

[0075] A content delivery decomposition network is designed based on Retinex theory. Latent space low-light features are input into the network, and the Hadamard product constraint decomposes these features into illumination-independent reflection and adaptive illumination components.

[0076]

[0077] in, , For learnable convolutional kernels, For the Sigmoid function, ensure ， ,

[0078] For the reflection component, For the learnable convolution kernel corresponding to the reflection component decomposition, Low-light characteristics of latent space, These are the bias parameters corresponding to the decomposition of the reflection components; To adapt to the illumination component, It is a linear rectified function. For the learnable convolutional kernel corresponding to adaptive illumination component decomposition, These are the bias parameters corresponding to adaptive illumination component decomposition.

[0079] In one embodiment, the learnable intensity compression function is:

[0080]

[0081] in, To control the compression steepness, The intensity threshold, Represents the parameters of the neural network. To adapt to the illumination component, It is a learnable intensity compression function;

[0082] Gradient consistency constraint and sparsity regularization are respectively

[0083]

[0084]

[0085] in, For gradient consistency constraints, For sparse regularization, TV is the total variation regularization term.

[0086] In one embodiment, the mask-guided diffusion module performs a forward diffusion process and a reverse denoising process in the latent space; the forward diffusion process is as follows:

[0087]

[0088] in, The latent space features are represented in the first The diffusion state of the step, Represents the diffusion step size. For noise modulation;

[0089] The inverse denoising process uses a latent space mask to assign spatial weights to low-light regions. This mask serves as a priori guide for the attention mechanism, causing the convolutional layers of the denoising network to prioritize processing high-noise regions while maintaining a low response to well-lit background regions. The process is as follows:

[0090]

[0091] in, Represents a denoising network. This is a learnable intensity function.

[0092] In one embodiment, the loss function of the low-light image enhancement module is:

[0093]

[0094]

[0095]

[0096] in, For reflectivity and illumination reconstruction loss, Let be the loss function of the diffusion process. For the reflection component, To adapt to the illumination component, For gradient consistency constraints, For sparsity regularization, , , and Representing different weights, This represents a normally lit image. For the expectation, This represents actual noise.

[0097] In a specific implementation, a multi-level loss function and dynamic weight adjustment mechanism are designed to coordinate the joint training of the three sub-tasks—image decomposition, mask generation, and diffusion denoising—ensuring the model's robustness in complex degradation scenarios. The model's overall loss function integrates decomposition, masking, and diffusion constraints. The decomposition loss constraint, based on Retinex theory, applies a reconstruction loss to the product of the reflection and illumination components, ensuring that the latent variables after decomposition can accurately reconstruct the normally illuminated image. The mask constraint aims to force the edge structure of the mask to be consistent with the original illumination components through gradient alignment loss, preventing mask generation from destroying the image topology. The diffusion denoising loss uses the standard noise prediction loss of the diffusion model, but its calculation range is limited to the degradation region defined by the mask. This region-limited strategy allows the model to focus on modeling the noise distribution in key regions while reducing the impact of interference signals in normal regions on the training process.

[0098] In one embodiment, the repaired latent space features are decoded into a repaired optical image, input to an adaptive enhancement submodule, and then, through spatial adaptive brightness curve adjustment and residual illumination estimation correction, an enhanced optical image for a non-uniform brightness scene is obtained. This enhanced optical image is then re-encoded into enhanced latent space features, including:

[0099] The repaired latent space features are decoded into a repaired optical image, which is then input into the adaptive enhancement submodule. The image is then adjusted using a spatial adaptive brightness curve to obtain a brightness-enhanced image.

[0100] In the illumination estimation, a residual correction strategy is adopted. The initial illumination map is obtained by filtering and histogram equalization. Then, the residual term is predicted by the learning module to obtain the final illumination estimate. The deficiencies under complex illumination are compensated by residual learning to obtain the illumination-enhanced image.

[0101] The enhanced optical image is obtained by fusing the brightness-enhanced image and the illumination-enhanced image.

[0102] In one embodiment, the enhanced latent space features are input to the image generation module. Endpoint binding constraints ensure the physical consistency of the generation process, the diffusion step size is adaptively adjusted, and a recurrent implicit iteration mechanism is introduced. The generated results are gradually refined through a recurrent neural network module to obtain an ISAR radar image of the aerospace target, including:

[0103] Initialize the diffusion parameters of the image generation module and determine the total number of diffusion steps. T The enhanced latent space features are used as the diffusion start endpoint features, and the standard features of the ISAR radar image of the aerospace target are used as the diffusion end endpoint features. The diffusion path is limited to the linear interpolation interval between the diffusion start endpoint features and the diffusion end endpoint features through endpoint binding constraints.

[0104] Starting from the initial endpoint features, the diffusion step size is adjusted according to a preset exponential decay scheduling function until completion. T Step-by-step positive diffusion;

[0105] Using the final forward diffusion feature as input, a recurrent implicit iteration mechanism is introduced. The hidden state of each denoising step is passed through the gated recurrent unit. In each step of the reverse denoising, the residual correction amount of the current diffusion feature is predicted based on the hidden state output by the recurrent neural network module. The feature correction is completed through the residual correction amount until T steps of reverse denoising are completed, and the target latent space feature with the same physical properties as the starting endpoint feature is obtained.

[0106] The target's latent space characteristics will be decoded to output ISAR radar images of the aerospace target.

[0107] In one embodiment, the diffusion step size is adjusted according to a preset exponential decay scheduling function.

[0108]

[0109] in, For diffusion initiation endpoint characteristics, exponential decay scheduling function , The attenuation coefficient is... Characteristics of a diffusion intermediate state. The total number of diffusion steps, for, This represents a Gaussian distribution with a mean of 0 and a covariance of the identity matrix.

[0110] In one embodiment, the hidden state of each denoising step is transmitted through a gated loop unit.

[0111]

[0112] in, It is in a hidden state. For the first t -1 step hidden state For the first t Characteristics after diffusion step This is a characteristic of the diffusion initiation endpoint.

[0113] In one embodiment, predicting the residual correction amount of the current diffusion feature based on the hidden state output by the recurrent neural network module includes:

[0114] Based on the hidden state output by the recurrent neural network module, the residual correction amount for predicting the current diffusion feature is:

[0115]

[0116] in, The linear weight parameters representing the residual prediction layer, It is in a hidden state.

[0117] It should be understood that, although Figure 1 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order in which these steps are executed, and they can be performed in other orders. Figure 1At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0118] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0119] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these modifications and improvements all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for generating radar image data of aerospace targets based on low-light enhancement correction, characterized in that, The method includes: Optical images of aerospace targets under low-light conditions are acquired, and the optical images are input into a preset encoder. The encoder maps the optical images to the latent space to obtain the low-light features of the latent space. An image data generation model is constructed. The image data generation model includes a low-light image enhancement module and an image generation module. The low-light image enhancement module includes a content transitive decomposition network, a learnable intensity compression function, a mask-guided diffusion module, and an adaptive enhancement submodule. In the low-light image enhancement module, a content delivery decomposition network is designed based on Retinex theory. The latent space low-light features are input into the content delivery decomposition network, and the latent space low-light features are decomposed into illumination-independent reflection components and adaptive illumination components through Hadamard product constraints. The adaptive illumination component is input into a learnable intensity compression function. Combined with gradient consistency constraints and sparsity regularization, a latent space mask is dynamically generated. The low-illumination degradation region in the optical image is located by the latent space mask. The reflection component, adaptive illumination component, and latent space mask are input into the mask-guided diffusion module. The latent space mask is embedded into the inverse denoising process to perform targeted repair on the low-illuminance degradation region, thereby obtaining the repaired latent space features. The repaired latent space features are decoded into a repaired optical image and input into the adaptive enhancement submodule. Through spatial adaptive brightness curve adjustment and residual illumination estimation correction, the enhanced optical image under non-uniform brightness scene is obtained and re-encoded into enhanced latent space features. An image generation module is constructed based on the principle of bidirectional diffusion. The enhanced latent space features are input into the image generation module. The physical consistency of the generation process is ensured by endpoint binding constraints. The diffusion step size is adaptively adjusted, and a cyclic implicit iteration mechanism is introduced. The generation result is gradually refined through a recurrent neural network module to obtain ISAR radar images of aerospace targets.

2. The method according to claim 1, characterized in that, A content delivery decomposition network is designed based on Retinex theory. The latent space low-light features are input into the content delivery decomposition network, and the latent space low-light features are decomposed into illumination-independent reflection components and adaptive illumination components through Hadamard product constraints, including: A content delivery decomposition network is designed based on Retinex theory. The latent space low-light features are input into the content delivery decomposition network, and the latent space low-light features are decomposed into illumination-independent reflection components and adaptive illumination components through Hadamard product constraints. in, , For learnable convolutional kernels, For the Sigmoid function, ensure ， , For the reflection component, For the learnable convolution kernel corresponding to the reflection component decomposition, Low-light characteristics of latent space, These are the bias parameters corresponding to the decomposition of the reflection components; To adapt to the illumination component, It is a linear rectified function. For the learnable convolutional kernel corresponding to adaptive illumination component decomposition, These are the bias parameters corresponding to adaptive illumination component decomposition.

3. The method according to claim 1, characterized in that, The learnable intensity compression function is: in, To control the compression steepness, The intensity threshold, Represents the parameters of the neural network. To adapt to the illumination component, It is a learnable intensity compression function; The gradient consistency constraint and sparsity regularization are respectively in, For gradient consistency constraints, For sparse regularization, TV is the total variation regularization term.

4. The method according to claim 1, characterized in that, The mask-guided diffusion module performs a forward diffusion process and a reverse denoising process in the latent space; the forward diffusion process is as follows: in, The latent space features are represented in the first The diffusion state of the step, Represents the diffusion step size. For noise modulation; The inverse denoising process assigns spatial weights to low-light regions using the latent space mask. The mask serves as a priori guide for the attention mechanism, causing the convolutional layers of the denoising network to prioritize high-noise regions while maintaining a low response to well-lit background regions. The process is as follows: in, Represents a denoising network. This is a learnable intensity function.

5. The method according to claim 4, characterized in that, The loss function of the low-light image enhancement module is: in, For reflectivity and illumination reconstruction loss, Let be the loss function of the diffusion process. For the reflection component, To adapt to the illumination component, For gradient consistency constraints, For sparsity regularization, , , and Representing different weights, This represents a normally lit image. As expected, This represents actual noise.

6. The method according to claim 1, characterized in that, The repaired latent space features are decoded into a repaired optical image, input into the adaptive enhancement submodule, and then, through spatial adaptive brightness curve adjustment and residual illumination estimation correction, an enhanced optical image for non-uniform brightness scenes is obtained. This enhanced optical image is then re-encoded into enhanced latent space features, including: The repaired latent space features are decoded into a repaired optical image, which is then input into the adaptive enhancement submodule. The image is then adjusted using a spatial adaptive brightness curve to obtain a brightness-enhanced image. In the illumination estimation, a residual correction strategy is adopted. The initial illumination map is obtained by filtering and histogram equalization. Then, the residual term is predicted by the learning module to obtain the final illumination estimate. The deficiencies under complex illumination are compensated by residual learning to obtain the illumination-enhanced image. The enhanced brightness image and the enhanced illumination image are fused together to obtain an enhanced optical image.

7. The method according to claim 1, characterized in that, The enhanced latent space features are input into the image generation module. Endpoint binding constraints ensure physical consistency in the generation process. The diffusion step size is adaptively adjusted, and a recurrent implicit iteration mechanism is introduced. The generated results are gradually refined through a recurrent neural network module to obtain ISAR radar images of aerospace targets, including: Initialize the diffusion parameters of the image generation module and determine the total number of diffusion steps. T The enhanced latent space features are used as the diffusion start endpoint features, and the standard features of the ISAR radar image of the aerospace target are used as the diffusion end endpoint features. The diffusion path is limited to the linear interpolation interval between the diffusion start endpoint features and the diffusion end endpoint features by endpoint binding constraints. Starting from the initial endpoint features, the diffusion step size is adjusted according to a preset exponential decay scheduling function until completion. T Step-by-step positive diffusion; Using the final forward diffusion feature as input, a recurrent implicit iteration mechanism is introduced. The hidden state of each denoising step is passed through the gated recurrent unit. In each step of the reverse denoising, the residual correction amount of the current diffusion feature is predicted based on the hidden state output by the recurrent neural network module. The feature correction is completed through the residual correction amount until T steps of reverse denoising are completed, and the target latent space feature with the same physical properties as the starting endpoint feature is obtained. The target's latent space characteristics will be decoded to output an ISAR radar image of the aerospace target.

8. The method according to claim 7, characterized in that, The diffusion step size is adjusted according to a preset exponential decay scheduling function. in, For diffusion initiation endpoint characteristics, the exponentially decaying scheduling function , The attenuation coefficient is... Characteristics of a diffusion intermediate state. The total number of diffusion steps, for, This represents a Gaussian distribution with a mean of 0 and a covariance of the identity matrix.

9. The method according to claim 7, characterized in that, The hidden state of each denoising step is transmitted through the gated loop unit. in, It is in a hidden state. For the first t -1 step hidden state For the first t Characteristics after diffusion step This is a characteristic of the diffusion initiation endpoint.

10. The method according to claim 7, characterized in that, Based on the hidden state output by the recurrent neural network module, predict the residual correction amount for the current diffusion feature, including: Based on the hidden state output by the recurrent neural network module, the residual correction amount for predicting the current diffusion feature is: in, The linear weight parameters representing the residual prediction layer, It is in a hidden state.

Images