Scene-level SAR image super-resolution methods, apparatus and computer equipment

Abstract

This application relates to a scene-level SAR image super-resolution method, apparatus, and computer device. The method designs a diffusion model for spatial degradation perception, which includes a sparse degradation estimation module, a conditional diffusion U-Net network, and a VQ-VAE decoder. The sparse degradation estimation module includes a keypoint prediction network, a degradation kernel estimation network, and an interpolation weight estimation network. The sparse degradation estimation framework estimates the degradation kernel of keypoints through keypoint selection, degradation kernel estimation, and adaptive local interpolation. A query mechanism injects sparse degradation priors into the conditional diffusion U-Net network, achieving pixel-by-pixel adaptive linear feature modulation. This method effectively alleviates the problems of excessive parameter quantity and high memory consumption associated with pixel-level prior estimation. It can adaptively handle degraded SAR images under various coupled degradation conditions, including spatially varied degradation, without requiring model retraining for different degradation scenarios.

Description

Technical Field

[0001] This invention belongs to the field of synthetic aperture radar image processing technology, and relates to a scene-level SAR image super-resolution method, apparatus and computer equipment. Background Technology

[0002] Synthetic Aperture Radar (SAR) has become an indispensable technology in modern remote sensing systems[1] due to its unique all-weather and all-sky imaging capabilities, playing an irreplaceable role in disaster assessment, resource exploration, and environmental monitoring. Among the key factors that determine the performance of these applications, image quality plays a crucial role, as higher image quality can lead to stronger interpretability and better performance of downstream tasks. However, in the process of acquiring SAR images, they are usually affected by the following three types of degradation and their coupling effects: limited spatial resolution, coherent speckle noise, and defocus. These degradations are mainly due to the inherent hardware limitations of the imaging system, coherent imaging mechanisms, and non-ideal platform / target motion. In addition, since scene-level SAR images have a wide coverage area and a rich variety of land cover categories, they exhibit obvious spatial variation degradation characteristics (such as Figure 1 (As shown). The degradation patterns and severity vary significantly across different spatial regions of the same image, which seriously challenges the uniform degradation assumption commonly used in existing methods. These complex degradations severely impair the visual quality of scene-level SAR images and their performance in downstream tasks, including but not limited to target detection, target recognition, and semantic segmentation.

[0003] As a key technology for improving SAR image quality, super-resolution (SR) methods aim to recover high-resolution (HR) images from low-resolution (LR) observations. However, existing SR methods are mainly based on the idealized assumption of uniform degradation and construct LR-HR image pairs for training based on this assumption. This approach is fundamentally different from the spatial variation coupled degradation encountered in real-world SAR image acquisition.

[0004] More importantly, existing methods face two fundamental limitations when processing scene-level SAR images: (1) Insufficient modeling of spatial degradation: Existing methods typically assume uniform spatial degradation and use a single degradation kernel for modeling. However, due to the wide observation range and rich land cover types of scene-level SAR, the degradation patterns and degrees of degradation in different spatial regions within the same image can vary significantly. A uniform degradation kernel cannot accurately characterize this spatial heterogeneity, resulting in unsatisfactory reconstruction performance. (2) Inefficient pixel-by-pixel degradation prior estimation: To address the problem of spatial degradation, existing methods primarily estimate the degradation prior for each pixel. While this method can estimate an independent degradation kernel for each pixel, it requires maintaining a high-dimensional pixel-level degradation tensor when processing large-scale scene-level SAR images, leading to significant memory overhead. These issues severely limit the effectiveness of existing methods in scene-level SAR image super-resolution tasks and further restrict their practical application in downstream tasks. Summary of the Invention

[0005] To address the problems existing in the above-mentioned traditional methods, this invention proposes a scene-level SAR image super-resolution method, apparatus, and computer equipment, which can adaptively process degraded SAR images under various coupled degradation conditions, including spatially varied degradation, without the need for model retraining for different degradation scenarios.

[0006] To achieve the above objectives, the embodiments of the present invention adopt the following technical solutions: On the one hand, a scene-level SAR image super-resolution method is provided, including the following steps: Step 1: Acquire low-resolution scene-level SAR images.

[0007] Step 2: Input the scene-level SAR image into the trained spatial degradation perception diffusion model to obtain a super-resolution SAR image; the diffusion model includes a sparse degradation estimation module, a conditional diffusion U-Net network, and a VQ-VAE decoder; the sparse degradation estimation module includes a key point prediction network, a degradation kernel estimation network, and an interpolation weight estimation network.

[0008] Step 2 specifically includes: The scene-level SAR image is input into the keypoint prediction network to generate a keypoint mask; the scene-level SAR image and the keypoint mask are input into the degradation kernel estimation network to obtain the keypoint degradation descriptor; the scene-level SAR image and the keypoint mask are input into the interpolation weight estimation network to obtain the local interpolation weights; based on the keypoint mask and the local interpolation weights, the keypoint degradation descriptor is interpolated and reconstructed to obtain the sparse degradation prior.

[0009] The sparse degradation prior, latent features corresponding to the scene-level SAR image, and initial latent noise are input into the conditional diffusion U-Net network with an integrated query mechanism. Backdiffusion iteration is performed under the guidance of the sparse degradation prior to obtain the final latent features. The final latent features are decoded by the VQ-VAE decoder to obtain the super-resolution SAR image. The query mechanism is used to treat the features at each location in the U-Net as queries, retrieve the corresponding degradation information from the sparse degradation prior, and generate adaptive modulation parameters.

[0010] On the other hand, a scene-level SAR image super-resolution device is also provided, comprising: Scene-level SAR image acquisition module, used to acquire low-resolution scene-level SAR images.

[0011] A super-resolution SAR image generation module is used to input scene-level SAR images into a trained spatial degradation perception diffusion model to obtain super-resolution SAR images. The diffusion model includes a sparse degradation estimation module, a conditional diffusion U-Net network, and a VQ-VAE decoder. The sparse degradation estimation module includes a keypoint prediction network, a degradation kernel estimation network, and an interpolation weight estimation network. Specifically, it includes: inputting the scene-level SAR image into the keypoint prediction network to generate a keypoint mask; inputting the scene-level SAR image and the keypoint mask into the degradation kernel estimation network to obtain keypoint degradation descriptors; and inputting the scene-level SAR image and the keypoint mask into the interpolation weight estimation network. In the value estimation network, local interpolation weights are obtained. Based on the keypoint mask and local interpolation weights, the keypoint degradation descriptor is reconstructed by interpolation to obtain a sparse degradation prior. The sparse degradation prior, the latent features corresponding to the scene-level SAR image, and the initial latent noise are input into the conditional diffusion U-Net network with an integrated query mechanism. Backdiffusion iteration is performed guided by the sparse degradation prior to obtain the final latent features. The final latent features are decoded by the VQ-VAE decoder to obtain the super-resolution SAR image. The query mechanism is used to treat the features at each location in the U-Net as queries, retrieve the corresponding degradation information from the sparse degradation prior, and generate adaptive modulation parameters.

[0012] In another aspect, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of any of the above-described scene-level SAR image super-resolution methods.

[0013] One of the above technical solutions has the following advantages and beneficial effects: The aforementioned scene-level SAR image super-resolution method, apparatus, and computer equipment include a spatial degradation sensing diffusion model. This model comprises a sparse degradation estimation module, a conditional diffusion U-Net network, and a VQ-VAE decoder. The sparse degradation estimation module includes a keypoint prediction network, a degradation kernel estimation network, and an interpolation weight estimation network. The sparse degradation estimation framework estimates the degradation kernel of keypoints through keypoint selection, degradation kernel estimation, and adaptive local interpolation. A query mechanism injects sparse degradation priors into the conditional diffusion U-Net network, achieving pixel-by-pixel adaptive linear feature modulation. This method effectively alleviates the problems of excessive parameter quantity and high memory consumption associated with pixel-level prior estimation. It can adaptively handle degraded SAR images under various coupled degradation conditions, including spatially varied degradation, without requiring model retraining for different degradation scenarios. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of this application or the conventional technology, the drawings used in the description of the embodiments or the conventional technology will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 This is a flowchart illustrating a scene-level SAR image super-resolution method in one embodiment; Figure 2 This is a schematic diagram of a sparse degradation estimation framework in one embodiment; Figure 3 A visualization of keypoint masks and estimated degradation descriptors in one embodiment; Figure 4 This is a schematic diagram of the network structure of a conditional diffusion U-Net network in one embodiment; Figure 5 This is a schematic diagram illustrating the specific training process of a diffusion model for spatial change degradation perception in one embodiment. Figure 6 This is a schematic diagram of the key point prediction network structure in one embodiment; Figure 7 This is a schematic diagram of the structure of a degenerate kernel estimation network in one embodiment; Figure 8 This is a schematic diagram of the interpolation weight estimation network structure in one embodiment; Figure 9 This is a schematic diagram comparing an HR image and a defocused image in one embodiment, wherein... Figure 9 Image (a) is the HR image. Figure 9 (b) is a defocused image; Figure 10 This is a schematic diagram illustrating the causes, mechanisms, and kernel models of three types of coupling degradation in a SAR image in one embodiment. Figure 10 (a) Schematic diagram of defocus degradation. Figure 10 Schematic diagram of medium (b) resolution degradation. Figure 10 (c) Schematic diagram of coherent speckle noise; Figure 11 This is a schematic diagram of the query mechanism framework in one embodiment; Figure 12 This is a schematic diagram of the attention mechanism for calculating modulation parameters in one embodiment; Figure 13 This is one embodiment. A schematic diagram showing the comparison results of grayscale values ​​generated under different degradation kernel models with the original LR image, wherein... Figure 13 (a) is a schematic diagram showing the comparison between the grayscale values ​​generated under the linear degradation kernel and the original LR image. Figure 13 (b) is a schematic diagram showing the comparison between the grayscale values ​​generated under the full quadratic degradation kernel and the original LR image. Figure 13 (c) is a schematic diagram showing the comparison between the gray values ​​generated by the approximate second-order degenerate kernel and the original LR image; Figure 14 This is a schematic diagram showing partial results of different SR algorithms under different degradation conditions in one embodiment; Figure 15 This is a schematic diagram of the detection results of various super-resolution algorithms under different degradation conditions in one embodiment. Detailed Implementation

[0016] 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.

[0017] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to be limiting of the application.

[0018] It should be noted that, in this document, the reference to "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The presentation of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. Those skilled in the art will understand that the embodiments described herein can be combined with other embodiments. The term "and / or" as used herein refers to any combination of one or more of the associated listed items, and all possible combinations, including such combinations.

[0019] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0020] In one embodiment, such as Figure 1 As shown, a scene-level SAR image super-resolution method is provided, which may include the following processing steps 1 to 7: Step 1: Acquire low-resolution scene-level SAR images.

[0021] Specifically, SAR images suffer from spatially varying coupled degradation, including resolution loss, speckle noise, and defocusing effects, which renders the uniform degradation assumption commonly used in existing methods invalid.

[0022] Step 2: Input the scene-level SAR image into the trained spatial degradation perception diffusion model to obtain a super-resolution SAR image; the diffusion model includes a sparse degradation estimation module, a conditional diffusion U-Net network, and a VQ-VAE decoder; the sparse degradation estimation module includes a key point prediction network, a degradation kernel estimation network, and an interpolation weight estimation network.

[0023] Step 2 specifically includes: Step 2.1: Input the scene-level SAR image into the keypoint prediction network to generate a keypoint mask; input the scene-level SAR image and the keypoint mask into the degradation kernel estimation network to obtain the keypoint degradation descriptor; input the scene-level SAR image and the keypoint mask into the interpolation weight estimation network to obtain the local interpolation weights; based on the keypoint mask and the local interpolation weights, perform interpolation reconstruction on the keypoint degradation descriptor to obtain the sparse degradation prior.

[0024] Specifically, the keypoint prediction network, the degenerate kernel estimation network, and the interpolation weight estimation network are constructed as follows: Figure 2The sparse degradation estimation module shown comprises: (1) a degradation kernel estimation network that estimates degradation descriptors of keypoints from the LR image; (2) a keypoint prediction network that predicts a binary keypoint mask representing the set of locations where the degradation descriptors will be directly estimated; and (3) an interpolation weight estimation network that estimates the local interpolation weights of non-keypoint pixels. The keypoint prediction network and the degradation kernel estimation network output the keypoint mask and degradation descriptors, respectively, while the interpolation weight estimation network outputs the local interpolation weights. The degradation descriptors of all pixels are reconstructed using the interpolation weights, and then the degradation operator applies these degradation priors to the HR image to generate a synthetic LR image. By minimizing the reconstruction loss and perceptual loss between the generated synthetic LR image and the real LR image, the entire framework learns accurate degradation estimation in a self-supervised manner. Figure 2 middle, Indicates the size of the defocusing kernel. Indicates the size of the fuzzy kernel. This represents the variance of the coherent speckle noise (with a constant dimension of 1). Let represent the coefficients of the linear and quadratic terms (with a constant dimension of 2). t This indicates the ratio of keypoints to total pixels. Indicates the image height and width. This indicates the size of the interpolation window.

[0025] Visualization of keypoint masks and estimated degradation descriptors from a sparse space variational degradation estimation framework, as shown below. Figure 3 As shown.

[0026] Figure 3 This shows the keypoint masks and estimated degradation descriptors generated at different locations using the sparse degradation estimation framework of this method. Figure 3 As shown, the keypoint prediction network identifies approximately 10% of the pixels as keypoints (represented as white dots in the keypoint mask). Then, the degradation kernel estimation network estimates degradation descriptors for these keypoints. The top of the degradation descriptor visualization shows the degradation kernels of adjacent pixels, which exhibit similar features due to spatial continuity; the bottom part shows the degradation descriptors of non-adjacent pixels, which exhibit distinctly different patterns. This visualization demonstrates that the sparse estimation framework proposed in this application can effectively capture the spatial variation characteristics of degradation.

[0027] As a preferred option, the network structure of the conditional diffusion U-Net network is as follows: Figure 4 As shown, the conditional diffusion U-Net network employs a U-Net network structure with degenerate prior injection, serving as a noise prediction or denoising network in the conditional diffusion model.

[0028] The following is about Figure 4The conditional diffusion U-Net network is illustrated below. The conditional diffusion U-Net network may include a temporal coding module, a low-resolution conditional coding module, a degradation prior generation module, a U-Net downsampling path, an intermediate bottleneck path, a U-Net upsampling path, and an output convolution module.

[0029] The time encoding module is used to encode the diffusion time step (time). Specifically, the time step is first normalized, and then the temporal embedding feature (t_emb) is generated through PositionalEncoding and the multilayer perceptron time_mlp, and injected into the network backbone as a modulation signal in the residual block.

[0030] The low-resolution conditional coding module can use TimeAwareLREncoder to encode the low-resolution conditional image cond_lr to obtain the conditional feature z_lr that matches the diffusion latent spatial resolution. In this code implementation, when z_lr is not explicitly provided externally, the network automatically calculates the conditional feature based on cond_lr and t_emb.

[0031] The degradation prior generation module is used to extract degradation-related prior information from low-resolution conditional images. In a specific implementation, the degradation prior generation module includes at least: a MaskPredictor, used to predict key location masks; a KernelEstimator, used to estimate degradation kernels or degradation parameter representations at the key locations and output kernels; and ALIW, used to generate an adaptive local interpolation weight map aliw_cond_map. After processing, the above degradation priors can form prior conditions for modulating U-Net feature maps, including a list of keypoint coordinates, degradation kernel embeddings, mask maps, and local weight maps.

[0032] The U-Net downsampling path consists of several ResNetBlock residual blocks and several DownsampleWithPriors downsampling modules with prior injection. At the input, the latent noise z_noisy is concatenated with the low-resolution conditional feature z_lr in the channel dimension to form the backbone input feature. Subsequently, the feature is sequentially processed through multiple residual blocks to extract multi-scale features, and at several downsampling positions, degenerate priors are injected into the feature map via DownsampleWithPriors to implement conditional constraints on the diffusion denoising process.

[0033] The intermediate bottleneck path can consist of one or more residual blocks, at least some of which may have an attention structure to enhance global context modeling capabilities. In this code implementation, the intermediate path includes two residual blocks: the first intermediate residual block has an attention mechanism, while the second intermediate residual block does not.

[0034] The U-Net upsampling path consists of several ResNetBlock residual blocks and several UpsampleWithPriors upsampling modules with prior injection. During upsampling, the current features are concatenated with the skip connection features saved in the downsampling stage, and then fed into the residual blocks for feature fusion. At the same time, at several upsampling positions, degradation priors are injected again through UpsampleWithPriors, thereby improving the adaptability of the reconstruction results to degradation patterns.

[0035] The output convolutional module can use a final_conv convolutional block to map the features output by the upsampling path back to the target latent channel number, obtaining the network output `out` required by the diffusion model. This output can be used as a noisy prediction result, or as a clean latent estimate obtained by further transforming the noisy prediction result, depending on the prediction method used by the diffusion model.

[0036] During the training or sampling phase, GaussianDiffusion can input the noisy latent z_noisy, time step time, low-resolution conditional image cond_lr, and optional low-resolution latent condition z_lr into UNetWithPriors (Conditional Diffusion U-net model) to output the corresponding diffusion denoising result. Therefore, from an implementation perspective, UNetWithPriors is a specific structural example of the Conditional Diffusion U-Net network in this embodiment.

[0037] It should be noted that the number of network layers, channels, attention heads, prior injection positions, degenerate priors, and conditional coding methods mentioned above can all be adjusted according to task requirements. This invention is not limited to the specific network parameters and layer configurations shown in this embodiment.

[0038] Step 2.2: Input the sparse degradation prior, the latent features corresponding to the scene-level SAR image, and the initial latent noise into the conditional diffusion U-Net network with an integrated query mechanism. Perform backdiffusion iteration guided by the sparse degradation prior to obtain the final latent features. Decode the final latent features through the VQ-VAE decoder to obtain the super-resolution SAR image. The query mechanism is used to treat the features at each location in the U-Net as queries, retrieve the corresponding degradation information from the sparse degradation prior, and generate adaptive modulation parameters.

[0039] Specifically, a low-resolution feature encoder is used to encode low-resolution scene-level SAR images into low-resolution latent features; initial latent noise is obtained by sampling from a standard Gaussian distribution; sparse degradation prior, low-resolution latent features and initial latent noise state are input into a conditional diffusion U-Net network with an integrated query mechanism, and back-diffusion iteration is performed guided by the sparse degradation prior to update the latent features until the iteration ends and the final latent features are obtained.

[0040] The query mechanism treats each feature location in the conditional diffusion U-Net network as a query, retrieving the corresponding degradation parameters from the estimated sparse degradation priors, thereby achieving pixel-level adaptive linear feature modulation. This mechanism enables spatially varied feature modulation, allowing the model to adaptively handle complex spatially varied coupled degradation in scene-level SAR images.

[0041] In one specific embodiment, the training process of the spatial change degradation sensing diffusion model can be implemented using the standard conditional diffusion model training framework. The specific training process of the spatial change degradation sensing diffusion model is as follows: Figure 5 As shown.

[0042] First, high-resolution image samples and their corresponding low-resolution image samples are acquired. The low-resolution image samples can be generated by applying degradation processes such as downsampling, defocusing blurring, and speckle noise to the original SAR data, or by generating them in real time through online degradation construction.

[0043] Furthermore, the high-resolution image can be input into the pre-trained encoder and mapped to the latent space to obtain the latent variable representation corresponding to the high-resolution image, denoted as the target latent. Subsequently, in each training iteration, a time step is randomly sampled from a preset diffusion time step range, and Gaussian noise is added to the target latent according to the time step to obtain the noisy latent at the current time step.

[0044] Simultaneously, the low-resolution image is fed as a conditional input into the Conditional Diffusion U-Net network. The Conditional Diffusion U-Net network can further extract spatial degradation prior information based on the low-resolution image. This degradation prior information includes, but is not limited to, key location masks, degradation kernel representations, adaptive local interpolation weights, and key point coordinate information. Subsequently, the aforementioned degradation prior is injected into the upsampling and downsampling feature processing of the denoising U-Net to adaptively modulate features at different spatial locations.

[0045] The denoising network receives noisy latency, time step information, and a low-resolution conditional image as input, and outputs predictions for the noise or velocity terms. During training, the network's predictions can be compared with the standard diffusion model at the corresponding time step to calculate the error and obtain the diffusion principal loss.

[0046] To further improve image reconstruction quality, during the training phase, the clean latent estimate recovered from the network output can be decoded and subjected to image domain constraints with the high-resolution image to construct one or more auxiliary losses among reconstruction loss, perceptual loss, and structural texture similarity loss. Finally, the diffusion main loss and one or more auxiliary losses are weighted and summed to obtain the total model loss, and the diffusion model parameters are updated through backpropagation.

[0047] The low-resolution conditional coding module and the spatial degradation prior estimation module can use pre-trained parameters. The degradation prior estimation module can be frozen during the diffusion training phase, with only the main body of the diffusion denoising network and its related conditional modulation parameters being trained. In this way, while maintaining the stability of standard diffusion training, the denoising network can possess adaptive perception and recovery capabilities for spatial degradation.

[0048] It should be noted that this method is not limited to a specific number of time steps, loss function form, conditional encoding method, or prior injection location.

[0049] The aforementioned scene-level SAR image super-resolution method employs a diffusion model for spatial degradation awareness. This model comprises a sparse degradation estimation module, a conditional diffusion U-Net network, and a VQ-VAE decoder. The sparse degradation estimation module includes a keypoint prediction network, a degradation kernel estimation network, and an interpolation weight estimation network. The sparse degradation estimation framework estimates the degradation kernel of keypoints through keypoint selection, degradation kernel estimation, and adaptive local interpolation. A query mechanism injects sparse degradation priors into the conditional diffusion U-Net network, achieving pixel-wise adaptive linear feature modulation. This method effectively alleviates the problems of excessive parameter quantity and high memory consumption associated with pixel-level prior estimation. It can adaptively handle degraded SAR images under various coupled degradation conditions, including spatially varied degradation, without requiring model retraining for different degradation scenarios.

[0050] In one embodiment, the keypoint prediction network adopts a lightweight encoder-decoder architecture consisting of a feature extraction backbone network and a scoring network. The process of inputting a scene-level SAR image into the keypoint prediction network to generate a keypoint mask includes: inputting a low-resolution scene-level SAR image into the backbone network and extracting multi-scale features using convolutional layers; inputting the multi-scale features into the scoring network, outputting a scalar score for each pixel location, and selecting a preset proportion of pixels as keypoints through a unified sampling strategy and a Top-K selection mechanism to generate a keypoint mask.

[0051] Specifically, the keypoint prediction network employs a lightweight encoder-decoder architecture consisting of a feature extraction backbone network and a scoring network. The backbone network extracts multi-scale features from the LR image through convolutional layers, while the scoring network outputs a scalar score for each pixel location, representing the probability that the location is selected as a keypoint. Through a unified sampling strategy and a Top-K selection mechanism, the network selects approximately 10% of pixels as keypoints, ensuring both spatial uniformity and prioritizing locations with significant degradation features.

[0052] In one specific embodiment, a keypoint prediction network can be used to predict keypoint location masks from low-resolution images. This network can employ a lightweight convolutional network structure as a keypoint selection module within a sparse spatial variation degradation estimation framework.

[0053] Key point prediction network structure such as Figure 6 As shown, the keypoint prediction network may include a feature extraction backbone network, a scoring network, and an auxiliary feature embedding branch. The feature extraction backbone network receives the low-resolution input image and extracts the basic feature map; the scoring network outputs keypoint scores for each pixel position based on the basic feature map; and the auxiliary feature embedding branch generates low-dimensional feature representations for use in additional constraints during the training phase.

[0054] The feature extraction backbone network can consist of a single convolutional layer, such as a 3×3 convolutional layer, with the number of input channels corresponding to the number of channels in the low-resolution image, and the number of output channels set to 64. The scoring network receives a single-channel feature map from the backbone feature map after channel mean compression, and passes it sequentially through a first convolutional layer, a group normalization layer, an activation function, and a second convolutional layer, outputting pixel-level scores (logits). The auxiliary feature embedding branch can use a 1×1 convolutional layer to compress the backbone features into a lower-dimensional feature representation, such as a 16-channel feature representation.

[0055] In the keypoint selection process, a sigmoid transformation can be applied to the score logits to obtain a pixel-level probability map. Then, the number of target keypoints is determined based on a preset keypoint ratio. Simultaneously, a basic sampling mask with a uniform spatial distribution can be constructed first, and then Top-K selection can be performed from the remaining locations based on the score probability, thus balancing spatial uniformity and prioritizing the preservation of significantly degraded regions. The resulting binary keypoint mask can be used as input for subsequent degradation kernel estimation and interpolation weight estimation.

[0056] During the training phase, the keypoint prediction network can output not only the final keypoint mask, but also low-dimensional features and differentiable keypoint mask representations to support diversity loss or other auxiliary training constraints. During the inference phase, it can output only the final keypoint mask.

[0057] It should be noted that this method is not limited to a specific number of convolutional layers, number of channels, normalization method, activation function form, embedding dimension, or keypoint selection strategy.

[0058] In one embodiment, the degenerate kernel estimation network employs a U-Net-based encoder-decoder architecture.

[0059] Specifically, the degradation kernel estimation network employs a U-Net-based encoder-decoder architecture to extract multi-scale feature maps from LR images. This network samples local features from the feature map at each keypoint location, combines them with location encoding information, then inputs them into a multilayer perceptron regression head, and finally outputs the corresponding degradation descriptor.

[0060] In one specific embodiment, the structure of the degenerate kernel estimation network is as follows: Figure 7 As shown, the degradation kernel estimation network can be used to estimate the degradation descriptors corresponding to keypoints based on low-resolution images and keypoint locations. This network can employ a U-Net-based encoder-decoder backbone, combined with location encoding and parametric regression heads, to explicitly model spatial variation degradation.

[0061] The degradation kernel estimation network consists of a U-Net backbone network, a location encoding module, and a regression head module. The U-Net backbone network extracts multi-scale feature maps from low-resolution images; the location encoding module encodes the spatial coordinates of keypoints; and the regression head module maps local keypoint features to degradation descriptor vectors.

[0062] The U-Net backbone network comprises a two-stage encoder, a one-stage bottleneck layer, and a two-stage decoder. Each encoder stage includes an input convolutional layer, several residual blocks, and a downsampling layer; each decoder stage includes an upsampling layer, a skip connection fusion layer corresponding to the encoder features, and several residual blocks. Through the collaborative action of the encoder and decoder, the network outputs a feature map that preserves spatial resolution.

[0063] In the keypoint degradation estimation process, the keypoint locations are first determined based on the keypoint mask, and then local features are sampled from the feature map that preserves spatial resolution according to the keypoint locations. Simultaneously, the two-dimensional coordinates of the keypoints are normalized and fed into the location encoding module to obtain the corresponding location encoding vector. Subsequently, the local features and the location encoding vector are fused and input into the regression head module to output the degradation descriptor corresponding to the keypoint.

[0064] The degradation descriptor may include one or more of the following: blur kernel parameters, speckle noise parameters, and other modulation parameters. The network's final output may be a set of degradation descriptors corresponding to a batch of keypoints and information on the number of keypoints in each image, for use in subsequent interpolation, degradation generation, or diffusion condition injection.

[0065] It should be noted that this method is not limited to a specific number of U-Net layers, number of channels, number of residual blocks, positional encoding format, local feature sampling method, or regression head structure.

[0066] In one embodiment, the interpolation weight estimation network employs a U-Net-based encoder-decoder architecture.

[0067] Specifically, the interpolation weight estimation network also employs the lightweight U-Net architecture, taking the LR image and keypoint mask as input. This network predicts the interpolation weight distribution within a local window for each pixel location. Within this local window at each pixel location, the window is centered on the current pixel and has a size of [missing information]. The network extracts multi-scale features through an encoder, upsamples the features to the original resolution through a decoder, and finally outputs the result for each pixel location. Window-dimensional weight vector. This network uses a mask-constrained softmax operation, so that only keypoints within the window have non-zero weights, while the weights of non-keypoints are forced to zero. For the keypoints themselves, the local interpolation weights are forced to be one-hot vectors, meaning the weight at the center is 1, and the weights at other locations are 0.

[0068] In one specific embodiment, the interpolation weight estimation network structure is as follows: Figure 8As shown, the interpolation weight estimation network can be used to estimate the interpolation weight distribution of each pixel location within a local window based on a low-resolution image and a keypoint mask. This network can employ a U-Net-based encoder-decoder backbone, combined with keypoint mask constraints and a normalization mechanism, to model local interpolation relationships.

[0069] The interpolation weight estimation network may include a U-Net backbone network and a weight constraint output module. The U-Net backbone network is used to extract multi-scale features from the joint input of the low-resolution image and the keypoint mask, and outputs the local window weight logits for each pixel location. The weight constraint output module is used to constrain the selectable keypoints in the local window according to the keypoint mask, and obtain the final interpolation weights through normalization.

[0070] The U-Net backbone network can include a shallow encoding block, two downsampling modules, an intermediate feature block, and two upsampling modules. Each convolutional block can include two convolutional layers, normalization, and activation functions; the upsampling modules can include deconvolutional upsampling and skip connections fused with encoder features. With this structure, the network can output a local window weight logits graph with the same spatial resolution as the original image.

[0071] During the local interpolation weight generation process, the keypoint mask can be expanded within a local window corresponding to each pixel position to form an allowable matrix, allowing only keypoint positions within the window to have non-zero weights. For cases where no keypoints exist within the window, a one-hot weight distribution with the current pixel center position set to 1 can be used as a fallback. For the keypoint positions themselves, their interpolation weights can also be forced to a centered one-hot distribution. Subsequently, the weights are normalized using constrained softmax to obtain the final local interpolation weights.

[0072] It should be noted that this method is not limited to a specific number of U-Net layers, number of channels, window size, normalization method, upsampling method, or constraint rules.

[0073] The keypoint prediction network, the degradation kernel estimation network, and the interpolation weight estimation network are jointly trained in an end-to-end manner, and the network parameters are optimized through forward consistency constraints.

[0074] In one embodiment, the three types of coupling degradation in SAR images include: resolution degradation, coherent speckle noise, and defocus degradation; before step 1, an explicit coupling degradation kernel modeling step is also included, specifically including: mathematically representing the degradation kernel as a convolution operation in the spatial domain, introducing a quadratic term into the degradation kernel representation, and representing the degradation kernel as a weighted sum of linear and quadratic terms, with the degradation process expression being: (1) (2) in, This is a degraded, low-resolution image. and These represent the weight parameters for the linear and quadratic terms, respectively. The output of the linear convolution kernel. Indicates a degenerate nucleus. i , j These represent the pixel rows and columns of the image, respectively. For the original high-resolution image, k The kernel size is the convolution kernel size. This represents the position coordinates of one pixel.

[0075] Specifically, the process of explicit coupled degradation kernel modeling includes: The inherent hardware constraints of SAR imaging systems, coupled with coherent imaging mechanisms and non-ideal platform / target motion, inevitably lead to three types of coupling degradation in SAR images: limited spatial resolution, multiplicative speckle noise, and defocus. This relates to reference 1 (Zilong Chen, Caiguang Zhang, Chenyu Wan, Siqian Zhang, and Boli Xiong. Dadsr: Degradation-aware diffusion super-resolution model for object-level SAR image). IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing Consistent with the degradation modeling framework proposed in (2025), the degradation process is described as: (3) in Represents an LR image. Represents a complex-valued HR image. This represents the defocus operator. This indicates k-th spatial downsampling, and |·| provides the amplitude calculation. This represents element-wise multiplication. This represents multiplicative speckle noise.

[0076] For ground-level SAR images, the spatial uniformity of degradation features is approximated due to the limited spatial extent and uniform terrain type, which restricts the spatial variation of degradation characteristics. However, for scene-level SAR images with wide coverage and diverse ground features, the spatial uniformity assumption becomes less valid. Even with globally consistent degradation parameters, the spatial representation of the degradation process exhibits inherent non-uniformity, mainly due to two factors: (1) the large spatial coverage amplifies the accumulation of spatial variation effects; and (2) the huge spatial scale inevitably includes a variety of ground feature categories with highly diverse scattering characteristics, resulting in significantly different degradation performances between different regions. Specifically, since resolution reduction operations are in the complex domain, defocusing operations are in the range-Doppler domain, speckle noise is multiplicative and tightly coupled with image content, the spatial representation of these degradations is naturally non-uniform. Figure 9 (a) and Figure 9 As shown in (b), even with globally consistent defocus degradation parameters, there are still significant differences in degradation performance between the target and background regions. In scene-level images, this non-uniformity is even more pronounced due to the greater diversity of scattering characteristics.

[0077] Given the inherent spatial non-uniformity of degradation in scene-level SAR images discussed above, it is necessary to construct spatially varying degradation descriptors to accurately characterize the degradation process at each spatial location. To this end, this method employs an explicit degradation modeling approach, applying it to each pixel... Allocate a degenerate descriptor It is used to characterize local degradation properties.

[0078] like Figure 10 As shown, this embodiment models the three types of coupling degradation as follows: (1) Resolution degradation: such as Figure 10 As shown in (b), the resolution degradation stems from two main causes: hardware limitations that lead to broadening of the point spread function, and multi-view processing that results in downsampling. Since the resolution reduction is achieved through complex domain downsampling, its spatial representation can be modeled as a function of the spatial domain. Degenerate nucleus, among which Indicates the effective fuzzy kernel size. (2) Defocus: such as Figure 10 As shown in (a), defocus degradation originates from phase errors introduced by the motion of a non-ideal platform / target. Linear phase errors only cause local or global image translation without affecting image quality, while higher-order phase errors (order ≥ 2) and sinusoidal phase errors cause energy to diffuse along the azimuth direction in the spatial domain. Since defocusing acts in the range-Doppler domain, it mainly manifests as energy diffusion along the azimuth direction, thus affecting only pixels in the same column. Therefore, defocusing is modeled as... Degenerate nucleus, among which This indicates the size of the defocusing nucleus along the azimuth direction. (3) Coherent speckle noise: such as Figure 10 As shown in (c), coherent speckle noise is generated by the coherent superposition of multipath echoes. The coherent speckle noise follows a pattern with a mean of 1 and a variance of... The gamma distribution can be obtained by directly estimating the variance parameter at each pixel location. To characterize.

[0079] However, the aforementioned degradation modeling suffers from a fundamental limitation: the three coupled degradation processes operate in different domains, and their interactions introduce complex nonlinear effects that cannot be fully captured by simple linear convolution kernels. Specifically, resolution reduction is achieved through complex domain downsampling, defocus degradation is simulated by introducing phase errors in the range-Doppler domain, and only coherent speckle noise directly operates in the spatial domain. Furthermore, the formation of the amplitude image involves amplitude calculation operations (|·|), which introduce additional nonlinearity. These domain transformations and amplitude operations generate complex nonlinear interactions between the HR amplitude image and the degraded LR amplitude image.

[0080] The degenerate kernel can be mathematically represented as a convolution operation in the spatial domain, as shown in Equation (2). However, this linear convolution model cannot capture the complex nonlinear effects generated by domain transformations (complex domain, distance-Doppler domain) and amplitude calculation operations. To address this issue, this application extends the degenerate kernel representation by introducing quadratic terms to approximate the nonlinear interactions between different degenerate types and those generated by amplitude operations.

[0081] In other words, the degenerate kernel is expressed as a combination of linear and quadratic terms: (4) in and These represent the linear and quadratic terms of the degradation nucleus, respectively.

[0082] However, this approach introduces a significant computational burden: estimating the linear and quadratic degenerate kernels doubles the number of parameters compared to a standalone linear model, severely impacting computational speed and memory consumption. To address this issue, this application employs a more efficient approximation scheme. For each degenerate descriptor, only two additional parameters are introduced to approximate the nonlinear effects. Therefore, the degeneracy process can be reformulated as shown in Equation (2). This approximation scheme introduces nonlinearity while adding only two extra parameters per degenerate descriptor, significantly reducing computational overhead compared to directly estimating the degenerate kernel using linear and quadratic terms.

[0083] Figure 10In (a) Defocus degradation: Non-ideal platform / target motion introduces phase errors, among which higher-order phase errors (order ≥ 2) and sinusoidal phase errors cause energy to diffuse along the azimuth direction, modeled as Degenerate nucleus. Figure 10 Medium (b) resolution degradation: Hardware limitations lead to point spread function expansion, while multi-view processing leads to downsampling, both of which contribute to... Space-degenerate nucleus. Figure 10 (c) Coherent speckle noise: Multiplicative coherent speckle noise is generated by the coherent superposition of multipath echoes, characterized by its variance parameter. It follows a gamma distribution.

[0084] In one embodiment, in the query mechanism: sparse degenerate priors are sorted using singular value decomposition, and then the top-ranked priors are retrieved within the corresponding window to obtain candidate degenerate priors; modulation parameters are calculated using an attention mechanism based on the candidate degenerate priors; specifically, this includes: mapping the feature vector and degenerate kernel descriptor at each position in the U-Net to the query space and key-value space respectively through linear projection, where the query comes from the U-Net feature mapping, and the keys and values ​​come from the candidate degenerate priors; the interpolation weight priors are processed using depthwise separable pyramid convolutions and 1×1 convolutions, and injected as attention biases into the attention calculation process; query and keywords The attention weights between them are: (5) in, For query and keywords Attention weights between them Attention bias derived from local interpolation weights. Let be the feature dimension.

[0085] Scaling parameters are generated using modulation attention branch. To control the intensity of feature scaling, translation parameters are generated using a translation attention branch combined with interpolation weight bias. To adjust the feature offset; through affine transformation The features are adaptively modulated so that each feature location receives different modulation parameters based on the degenerate prior of its query.

[0086] In one embodiment, sparse degenerate priors are sorted using singular value decomposition (SVD) and then the top-ranked priors are retrieved from the corresponding window to obtain candidate degenerate priors. This includes: sorting the sparse degenerate priors by global importance using SVD to obtain a global ranking list; for each feature location in the U-Net, a local search window is determined based on its corresponding HR spatial coordinates, and within the local search window, the top-ranked prior is selected from the global ranking list. Degenerate priors serve as a candidate set of degenerate priors; where It is adaptively determined based on the feature map scale and window size.

[0087] Specifically, after obtaining sparse degradation priors, effectively injecting these priors into the diffuse U-Net to achieve pixel-by-pixel adaptive feature modulation is crucial for improving the quality of super-resolution reconstruction. Traditional global adjustment methods cannot capture the degradation process of spatially varying properties, while pixel-level degradation parameter estimation faces high computational complexity. To address this issue, this application designs a query mechanism that treats the feature vector at each location in the U-Net as a query, retrieving the most relevant degradation information from the sparse degradation priors and generating adaptive modulation parameters. Unlike traditional feature modulation methods, this mechanism fully utilizes the spatial distribution features of the sparse priors, achieving efficient and accurate adaptive feature enhancement through window pre-filtering and attention mechanisms. The query mechanism architecture is as follows: Figure 11 As shown.

[0088] In order to effectively retrieve relevant information from a large number of sparse degenerate priors, this method sorts the sparse degenerate priors using singular value decomposition and then retrieves the top-ranked priors in the corresponding window to obtain candidate degenerate priors. Specifically: (1) Use singular value decomposition (SVD) to sort all sparse degenerate kernels by global importance. Specifically, first, all degenerate kernels are centered by subtracting the mean, and then SVD decomposition is performed on the centered kernel matrix to extract the first right singular vector as the main direction. Subsequently, the projected energy of each degenerate kernel is calculated along this main direction, and all kernels are sorted in descending order according to the energy. This global sorting process extracts the most representative degenerate patterns, providing an important basis for subsequent local queries. (2) For each feature location in U-Net, a local search window is determined according to its corresponding HR space coordinates; within the local search window, the top-ranked priors are selected from the global ranking list. Degenerate priors are used as a candidate set of degenerate priors, among which It is adaptively determined based on the feature map scale and window size.

[0089] After obtaining the candidate degradation prior set, modulation parameters are calculated through an attention mechanism. Specifically, feature vectors and degradation kernel descriptors are mapped to the query space and key-value space respectively through linear projection, where the query comes from the U-Net feature map, and the keys and values ​​come from the candidate degradation priors. To enhance spatial relevance, the interpolation weight priors are processed using depthwise separable pyramid convolutions and 1×1 convolutions, and injected as attention biases into the attention calculation process. The attention mechanism for calculating modulation parameters adopts a dual-path design: the modulation attention branch generates scaling parameters. This controls the intensity of feature scaling, while the translation attention branch, combined with interpolation weight bias, generates translation parameters. To adjust the feature offset. The query is calculated using formula (5). and keywords Attention weights are assigned between individuals. The attention weights are calculated using a scaled dot product attention mechanism, with softmax normalization ensuring a reasonable weight distribution. Finally, an affine transformation is applied. By adaptively modulating the features, each feature location receives different modulation parameters based on its queried degradation prior, thereby achieving spatially varied feature enhancement. This design enables the network to adaptively adjust the feature representation according to local degradation features, thus better handling spatially varied degradation processes.

[0090] In one specific embodiment, the attention mechanism for calculating modulation parameters is as follows: Figure 12 As shown, the attention mechanism for calculating modulation parameters can employ a dual-path attention modulation mechanism, injecting keypoint degradation priors into the feature map modulation process. The dual-path attention modulation mechanism can include a modulation attention branch and a translation attention branch, where the modulation attention branch generates scaling parameters and the translation attention branch generates translation parameters.

[0091] For any given feature layer, the feature vector at each spatial location of that layer can be used as the query vector, and the result of linear transformation of the keypoint degradation descriptor can be used as the candidate set of key vectors and value vectors. Subsequently, independent key projection and value projection layers are set for the modulation attention branch and the translation attention branch, respectively, so as to learn the parameter representations for scaling modulation and translation modulation respectively.

[0092] In the modulation attention branch, scaling dot product attention is performed based on the attention relationship between the query vector and the candidate set of keypoint degradation descriptors to obtain a scaling parameter tensor γ with the same spatial size as the current feature map. In the translation attention branch, the same attention calculation method as above is used, further combined with the attention bias formed by local interpolation weights to obtain a translation parameter tensor β with the same spatial size as the current feature map.

[0093] Both modulation attention branches and translation attention branches can be implemented based on the ordinary scaling dot product attention mechanism. The difference lies in using different key-value projection parameters and different attention biases. After obtaining the scaling parameter γ and the translation parameter β, the features can be modulated in an affine modulation manner. For example, the feature map h can be modulated position-by-position and channel-by-channel by (1 + γ) * h + β.

[0094] It should be noted that this method is not limited to a specific number of attention heads, projection dimension, bias construction method, or modulation position.

[0095] This method integrates this query mechanism into the encoder and decoder of the conditional diffusion U-Net network, performing modulation operations at key downsampling and upsampling layers. Specifically, in the downsampling and upsampling layers, feature modulation is applied after convolution operations, enabling the network to adaptively adjust feature representations based on degradation priors at different scales. During diffusion denoising, this query mechanism queries the degradation prior based on the current feature state at each time step, generating corresponding modulation parameters, thereby continuously guiding the network to focus on degradation-related features throughout the denoising process. This design naturally incorporates degradation priors as conditional information into the diffusion process, achieving degradation-aware super-resolution reconstruction.

[0096] In one embodiment, the degradation descriptor of keypoints is reconstructed by interpolation based on keypoint masks and local interpolation weights to obtain a sparse degradation prior, including: based on keypoint masks... Check the position of each pixel Is it a key point; if the pixel position If it is a key point, then the pixel position The degradation descriptor is extracted directly from the estimated keypoint degradation descriptor; if the pixel location If it is not a key point, then define a... The size of the center is A local window is defined, within which only the locations corresponding to keypoints have non-zero weights, while the weights of non-keypoint locations are zero. The non-keypoint pixel locations are reconstructed by calculating the weighted sum of the degenerate descriptors of the keypoints within the local window. Degenerate descriptor.

[0097] Specifically, the interpolation reconstruction process of the degraded descriptor employs an adaptive local interpolation strategy. For each pixel location... The framework first uses keypoint masks. Check if it is a keypoint. If (x, y) is a keypoint (i.e. If ), then the degradation descriptor at that location will be extracted directly from the already estimated keypoint degradation descriptors. If Not the key point (i.e.) ), then the frame query location Local interpolation weights at the location It defines a Centered A local window is defined within the window, where only the locations corresponding to keypoints have non-zero weights, while the weights of non-keypoint locations are zero. Subsequently, by calculating the weighted sum of the degenerate descriptors of the keypoints within the local window, the degenerate descriptor of the non-keypoint pixel (x, y) can be reconstructed. Mathematically, a degenerate descriptor... For pixels The calculation is as follows: (6) in express by The local neighborhood centered on, Indicates key point location Degenerate descriptor, Indicate key points Relative to pixels The interpolation weights are determined by the interpolation process. It's important to note that the interpolation weights are estimated so that the weights for non-keypoint locations are always zero, ensuring that only keypoint descriptors participate in the interpolation process.

[0098] Comparison of grayscale values ​​of generated images with the original LR image under different degenerate kernel models, as follows: Figure 13 As shown, where Figure 13 (a) shows the comparison results of the grayscale values ​​generated under the linear degradation kernel with the original LR image. Figure 13 (b) shows the comparison results of the grayscale values ​​generated under the full quadratic degradation kernel with the original LR image. Figure 13 (c) shows the comparison results of the gray values ​​generated under the approximate quadratic degenerate kernel with the original LR image. Figure 7 This visually demonstrates the comparison of grayscale distribution between the generated LR images and the original LR images, obtained using three different degradation kernel modeling strategies to extract degradation priors. The horizontal axis represents the grayscale values ​​of the original image, the vertical axis represents the grayscale values ​​of the generated image, and the red dashed line... The ideal unbiased estimate is one where the generated pixel intensity perfectly matches the original pixel intensity. By comparing the convergence of the scatter points to the unbiased estimation line and the mean squared error (MSE), it can be observed that the linear degradation kernel produces the most dispersed distribution and the largest error. Conversely, when the parameter overhead is negligible, the approximate quadratic degradation kernel proposed in this application exhibits a more balanced distribution along the line. The most densely clustered scatter distribution yielded the lowest error. Its reconstruction accuracy not only significantly outperformed the linear model but also surpassed the fully quadratic degenerate kernel, which requires higher computational costs, further validating the effectiveness and superiority of the proposed nonlinear approximation scheme.

[0099] After interpolation, a complete degradation prior for all pixels is obtained, i.e., a sparse degradation prior. This sparse degradation prior is only used for forward consistency constraints. The estimated degradation is applied to the HR image to generate a degraded LR image that is as close as possible to the original LR image. In practical applications, only the degradation descriptor, keypoint mask, and local interpolation weights (such as...) are used. Figure 2(As shown) is used for subsequent feature modulation in the diffusion U-Net. Compared with pixel-level prior estimation, the sparse degradation estimation module of this application significantly reduces the number of parameters and memory consumption. Specifically, pixel-level estimation requires maintaining the degradation descriptors of all pixels, and its parameter count is... The terms represent the blur kernel parameter, defocus kernel parameter, speckle noise variance, and linear / quadratic coefficients, respectively. In contrast, this method only requires... ,in t The first term represents the keypoint ratio (approximately 10%), the second term represents the local interpolation weight, and the third term represents the keypoint mask. Because... and The parameters are typically small (e.g., 5 or 7), thus achieving a significant reduction in parameter count and memory footprint compared to pixel estimation.

[0100] The pseudocode for the reasoning process of the diffusion model of spatial change degradation perception is shown in Table 1.

[0101] Table 1. Pseudocode of the reasoning process of the diffusion model for spatial change degradation perception.

[0102] It should be understood that, although the above Figure 1 The steps in the code are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated otherwise in this method, there is no strict order requirement for the execution of these steps; they can be executed in other orders. Furthermore, the above... Figure 1 At least some of the steps may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed 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.

[0103] In a validation implementation, the dataset was constructed using the FARAD dataset. Three degradation types were defined: resolution reduction, speckle noise, and defocus blur. Each degradation type included four severity levels: no degradation, mild, moderate, and severe, which combined to produce 64 different degradation patterns.

[0104] For the training set, this embodiment employs an online data construction strategy, dynamically generating training samples during training. Specifically, it starts with approximately 45 raw SAR images, each approximately 4000 × 1600 pixels in size. In each training iteration, image patches are randomly sampled from these raw images, and a randomly selected degradation level is applied to generate degraded low-resolution samples. All image patches are cropped to a fixed size of 512 × 512 pixels. For the test set, this embodiment extracts 240 image patches from another set of 9 images, ensuring they do not overlap with the training images. For each degradation level, this embodiment randomly selects 30 patches from this set, resulting in a total of 1920 test samples. All test patches are also cropped to 512 × 512 pixels. Furthermore, the degradation parameters for the training set remain fixed, while controlled perturbations are introduced into the test set parameters to rigorously evaluate model generalization beyond predefined degradation conditions.

[0105] Model Configuration: For the sparse degradation prior estimation framework described in Section III-B, this embodiment uses an AdamW optimizer with a batch size of 4. The initial learning rate is set to 1×10⁻⁴ and decays to 0.1 times its initial value via cosine annealing. The blur kernel size kblur is set to 17, the defocus kernel size kdefocus is set to 33, and the interpolation weight window size is set to 5×5. The framework estimates approximately 10% of the degraded descriptors as keypoints.

[0106] For the diffusion model, this embodiment adopts the architecture of Latent Diffusion-On Models (LDM) and integrates the query mechanism into the output features of the upsampling and downsampling layers. The model is optimized using the AdamW optimizer with a batch size of 4 and an initial learning rate of 1×10⁻⁴, which is then decayed to 0.1 times the initial value using cosine annealing. During this training phase, the degenerate prior estimation framework remains frozen. The network is trained for 200,000 iterations on an NVIDIA RTX 4090 GPU. Furthermore, the time step is set to T = 1000.

[0107] (1) Evaluation indicators In SR tasks, structural similarity index (SSIM) and peak signal-to-noise ratio (PSNR)

[35] are commonly used to evaluate image quality. However, PSNR relies on the mean square error calculation per pixel, and because it is insensitive to structural distortion, it ignores the importance of local features and therefore cannot accurately reflect human visual perception. SSIM is too sensitive to geometric deformation and small perturbations, which affects the reliability of its quantification of perceptual quality. Some results of different SR algorithms under different degradation conditions are as follows Figure 14As shown, although RCAN achieved higher PSNR and SSIM scores, its visual results were perceptually inferior to competing methods. Given these limitations, the evaluation framework of this embodiment only uses the above conventional metrics as supplementary references. To establish a perceptually meaningful evaluation system, this embodiment introduces complementary metrics, including Learned Perceptual Image Patch Similarity (LPIPS), Fr'cheetInception Distance (FID), and Depth Image Structure and Texture Similarity (DISTS).

[0108] Given that deep learning has been widely applied to object detection tasks, this embodiment further introduces a downstream task evaluation metric based on object detection to more comprehensively evaluate the performance of the super-resolution (SR) algorithm. Specifically, this embodiment uses the YOLOv12 object detection model and trains it on a subset of the training dataset with ground truth annotations. Subsequently, the SR-processed images are input into the network for this specific task, and their object detection accuracy is used as a quantitative metric to measure the effectiveness of the SR algorithm.

[0109] (2) Comparison with other state-of-the-art methods This embodiment evaluates the SR results of the proposed method on coupled degraded SAR images and compares them with several typical SR methods. SR3, RCAN, SWINIR, and RealESRGAN are selected as baseline models for comparative analysis. This embodiment comprehensively evaluates different super-resolution algorithms based on traditional metrics and downstream tasks. The specific parameter configurations of these models are detailed in their respective original publications using the same training dataset.

[0110] To clearly demonstrate the differences in visual effects among various SR methods, this embodiment presents the reconstruction results of SR images. Figure 14Compared to other methods, this approach better preserves texture realism under various degradation conditions while effectively avoiding unnatural texture artifacts and oversmoothing effects frequently observed in other methods. Models optimized by MSE, such as RCAN and SWINIR, exhibit significant oversmoothing artifacts during super-resolution reconstruction. This phenomenon arises primarily because in super-resolution tasks, a low-resolution observation typically corresponds to multiple possible high-resolution solutions. The inherent mechanism of MSE optimization drives these models to generate a weighted average of potential solutions, resulting in suppression of high-frequency components and damage to texture details. Given that SAR images have a higher dynamic range and richer high-frequency content than natural optical images, this averaging effect is particularly detrimental to super-resolution images. While generative methods, including perceptual optimization (RealESRGAN) and the diffusion model (SR3), have successfully circumvented the averaging effect and generated high-resolution images with enhanced details, they often produce unnatural texture artifacts due to the lack of degradation prior guidance.

[0111] Table 2 presents the metrics for different super-resolution algorithms used in metric-based SR image processing and a comparison of their performance on downstream tasks. Our proposed method achieves the best performance on perceptual consistency metrics (LPIPS, FID, DISTS). It also demonstrates significant advantages in downstream tasks, as shown in the specific results below. Figure 15 As shown in Table 2. However, its PSNR and SSIM scores are lower than those of the MSE-based method. Although Figure 14 Visual evidence shows that reconstructions based on MSE exhibit significant oversmoothing. This discrepancy stems from the limitations of traditional metrics: PSNR primarily relies on pixel-by-pixel mean square error calculation.

[0112] Table 2. Performance Comparison of Different Super-Resolution Algorithms and Downstream Tasks

[0113] Because it is insensitive to structural distortion and ignores the importance of local features, it cannot fully reflect the quality of human visual perception. Furthermore, SSIM is overly sensitive to geometric deformation and minute perturbations, affecting the reliability of its quantification of perceptual quality. Given these inherent limitations, the PSNR and SSIM measurements provided in this embodiment are for reference only.

[0114] (3) Ablation research To evaluate the effectiveness of this method, this embodiment conducted an ablation study from three aspects: prior form, keypoint ratio, and interpolation weight window size. Experimental results are quantitatively demonstrated in Tables 3, 4, 5, and 6. Table 3 compares the performance of sparse degradation priors and pixel-wise degradation priors. For pixel-wise baselines, the extraction of dense priors uses the exact same degradation kernel estimator as our method; however, its injection process employs a classic feature pyramid network to map the prior to multiple scales.

[0115] Table 3 Comparison of reconstruction performance for different prior forms

[0116] Table 4 Comparison of computational efficiency for pixelation and sparse degradation prior modulation processes

[0117] Table 5 Evaluation results of different key point proportions

[0118] Table 6 Evaluation results for different interpolation window sizes

[0119] Feature maps are combined with spatial feature transformation for feature modulation. Table 4 compares the efficiency parameters between pixel-level degradation prior modulation and sparse degradation prior modulation. The efficiency comparison in Table 4 is based on the forward inference process, with an interpolation window size of 5×5, a keypoint ratio of 10%, and a batch size of 4. The results show that the proposed sparse degradation prior modulation significantly improves computational efficiency, requiring only a negligible trade-off in reconstruction accuracy. Tables 5 and 6 are used to verify the impact of keypoint ratio and interpolation window size. Experimental results show that keypoint ratio and interpolation window size have minimal impact on super-resolution performance, indicating that achieving efficient spatially variable degradation sensing super-resolution of scene-level SAR images with a small degradation prior is feasible.

[0120] Quantitative evaluations of the examples revealed inherent differences between traditional pixel-level metrics (such as PSNR and SSIM) and perceptual or downstream task metrics. MSE-optimized models (RCAN and SWINIR) achieved higher PSNR and SSIM scores but suffered from severe oversmoothing artifacts. This is because MSE-based optimization essentially averages multiple potential high-resolution solutions, discarding critical high-frequency structural details and complex texture features inherent in SAR images. In contrast, SVDA leverages the generative capabilities of latent spatial diffusion models and is guided by spatial variation degradation priors to generate highly realistic textures, achieving significant improvements in perceptual fidelity metrics (LPIPS, FID, DISTS) and downstream target detection performance (0.8465 mAP@0.5). This paradigm shift highlights the limitations of relying solely on PSNR / SSIM for evaluating scene-level SAR super-resolution tasks and validates the necessity of introducing task-driven and human-visual-aligned evaluation criteria.

[0121] Addressing spatially varying coupled degradation typically requires pixel-level degradation kernel estimation. When processing large-scale scene-level SAR images, this approach inevitably faces extreme memory consumption and computational bottlenecks. The ablation experiments in this embodiment demonstrate that the sparse degradation estimation framework proposed in this application achieves an excellent trade-off between efficiency and accuracy. Switching from pixel-level priors to a sparse prior framework reduces computational overhead by over 50% and the number of parameters by approximately 85%, while the impact on reconstruction accuracy is negligible. Furthermore, the model exhibits high stability across different keypoint ratios and local interpolation window sizes. This robustness confirms the core assumption of this method: the degradation process of scene-level SAR images possesses strong local spatial continuity. Therefore, dense pixel-level estimation is highly redundant, and sampling approximately 10% of the keypoints is sufficient to generate accurate pixel-level adaptive feature modulation parameters.

[0122] Experimental results on the FARAD dataset demonstrate that this method not only achieves excellent perceptual fidelity in synthesizing realistic high-frequency SAR textures, but also achieves considerable performance improvement in downstream target detection tasks, while significantly reducing computational and memory overhead.

[0123] In one embodiment, a scene-level SAR image super-resolution device is also provided, comprising: Scene-level SAR image acquisition module, used to acquire low-resolution scene-level SAR images.

[0124] A super-resolution SAR image generation module is used to input scene-level SAR images into a trained spatial degradation perception diffusion model to obtain super-resolution SAR images. The diffusion model includes a sparse degradation estimation module, a conditional diffusion U-Net network, and a VQ-VAE decoder. The sparse degradation estimation module includes a keypoint prediction network, a degradation kernel estimation network, and an interpolation weight estimation network. Specifically, it includes: inputting the scene-level SAR image into the keypoint prediction network to generate a keypoint mask; inputting the scene-level SAR image and the keypoint mask into the degradation kernel estimation network to obtain keypoint degradation descriptors; and inputting the scene-level SAR image and the keypoint mask into the interpolation weight estimation network. In the value estimation network, local interpolation weights are obtained. Based on the keypoint mask and local interpolation weights, the keypoint degradation descriptor is reconstructed by interpolation to obtain a sparse degradation prior. The sparse degradation prior, the latent features corresponding to the scene-level SAR image, and the initial latent noise are input into the conditional diffusion U-Net network with an integrated query mechanism. Backdiffusion iteration is performed guided by the sparse degradation prior to obtain the final latent features. The final latent features are decoded by the VQ-VAE decoder to obtain the super-resolution SAR image. The query mechanism is used to treat the features at each location in the U-Net as queries, retrieve the corresponding degradation information from the sparse degradation prior, and generate adaptive modulation parameters.

[0125] In one embodiment, the keypoint prediction network adopts a lightweight encoder-decoder architecture consisting of a feature extraction backbone network and a scoring network. The super-resolution SAR image generation module is also used to input low-resolution scene-level SAR images into the backbone network and extract multi-scale features using convolutional layers. The multi-scale features are input into the scoring network, which outputs a scalar score for each pixel location. Through a unified sampling strategy and Top-K selection mechanism, a preset proportion of pixels are selected as keypoints to generate a keypoint mask.

[0126] In one embodiment, the degradation kernel estimation network in the super-resolution SAR image generation module adopts a U-Net-based encoder-decoder architecture.

[0127] In one embodiment, the interpolation weight estimation network in the super-resolution SAR image generation module adopts a U-Net-based encoder-decoder architecture.

[0128] In one embodiment, the three types of coupled degradation of SAR images include: resolution degradation, coherent speckle noise and defocus degradation; before the sparse spatial variation degradation estimation module, there is also an explicit coupled degradation kernel modeling module, which is used to mathematically represent the degradation kernel as a convolution operation in the spatial domain, and introduce a quadratic term in the degradation kernel representation, representing the degradation kernel as a weighted sum of linear and quadratic terms, and the degradation process expression is shown in Equation (1) and Equation (2).

[0129] In one embodiment, the super-resolution SAR image generation module is also used in the query mechanism to: sort the sparse degradation priors using singular value decomposition and then retrieve the top-ranked ones in the corresponding window to obtain candidate degradation priors; calculate the modulation parameters using an attention mechanism based on the candidate degradation priors; specifically including: mapping the feature vector and degradation kernel descriptor of each position in U-Net to the query space and key-value space respectively through linear projection, wherein the query comes from the U-Net feature mapping, and the key and value come from the candidate degradation priors; process the interpolation weight priors using depthwise separable pyramid convolution and 1×1 convolution, and inject them as attention biases into the attention calculation process; calculate the query using formula (5). and keywords Attention weights between them; scaling parameters are generated using modulated attention branches. To control the intensity of feature scaling, translation parameters are generated using a translation attention branch combined with interpolation weight bias. To adjust the feature offset; through affine transformation The features are adaptively modulated so that each feature location receives different modulation parameters based on the degenerate prior of its query.

[0130] In one embodiment, the super-resolution SAR image generation module is further configured to perform global importance ranking on the sparse degradation prior using singular value decomposition to obtain a global ranking list; for each feature location in the U-Net, a local search window is determined based on its corresponding HR spatial coordinates, and within the local search window, the top-ranked feature is selected from the global ranking list. Degenerate priors serve as a candidate set of degenerate priors; where It is adaptively determined based on the feature map scale and window size.

[0131] In one embodiment, the super-resolution SAR image generation module is also used to generate images based on keypoint masks. Check the position of each pixel Is it a key point; if the pixel position If it is a key point, then the pixel position The degradation descriptor is extracted directly from the estimated keypoint degradation descriptor; if the pixel location If it is not a key point, then define a... The size of the center is A local window is defined, within which only the locations corresponding to keypoints have non-zero weights, while the weights of non-keypoint locations are zero. The non-keypoint pixel locations are reconstructed by calculating the weighted sum of the degenerate descriptors of the keypoints within the local window. Degenerate descriptor.

[0132] It is understood that for detailed explanations of the scene-level SAR image super-resolution device, please refer to the corresponding explanations of the various embodiments of the scene-level SAR image super-resolution method above, and will not be repeated here. Each module in the above-mentioned scene-level SAR image super-resolution device can be implemented entirely or partially through software, hardware, or a combination thereof. Each module can be embedded in hardware or independently of a device with data processing capabilities, or stored in software in the memory of the aforementioned device, so that the processor can call and execute the operations corresponding to each module. The aforementioned device can be, but is not limited to, various types of data processing computer devices already existing in the art.

[0133] In one embodiment, a computer device is also provided, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described scene-level SAR image super-resolution method embodiment.

[0134] It is understood that, in addition to the memory and processor mentioned above, the computer equipment described above also includes other hardware and software components not listed in this specification. The specific components can be determined according to the model of the image processing computer in different application scenarios, and will not be listed and described in detail in this specification.

[0135] 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.

[0136] The above embodiments are merely illustrative of several implementation methods of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of protection 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 all such modifications and improvements fall within the scope of protection of this application.

Claims

1. A scene-level SAR image super-resolution method, characterized in that, Including the following steps: Step 1: Acquire low-resolution scene-level SAR images; Step 2: Input the scene-level SAR image into the trained spatial degradation perception diffusion model to obtain a super-resolution SAR image; the diffusion model includes a sparse degradation estimation module, a conditional diffusion U-Net network, and a VQ-VAE decoder. The sparse degradation estimation module includes a keypoint prediction network, a degradation kernel estimation network, and an interpolation weight estimation network; Step 2 specifically includes: A scene-level SAR image is input into a keypoint prediction network to generate a keypoint mask; the scene-level SAR image and the keypoint mask are input into a degradation kernel estimation network to obtain a keypoint degradation descriptor; the scene-level SAR image and the keypoint mask are input into an interpolation weight estimation network to obtain local interpolation weights; based on the keypoint mask and the local interpolation weights, the keypoint degradation descriptor is interpolated and reconstructed to obtain a sparse degradation prior. The sparse degradation prior, latent features corresponding to the scene-level SAR image, and initial latent noise are input into the conditional diffusion U-Net network with an integrated query mechanism. Backdiffusion iteration is performed under the guidance of the sparse degradation prior to obtain the final latent features. The final latent features are then decoded by the VQ-VAE decoder to obtain the super-resolution SAR image. The query mechanism is used to treat the features at each location in the U-Net as queries, retrieve the corresponding degradation information from the sparse degradation prior, and generate adaptive modulation parameters.

2. The scene-level SAR image super-resolution method according to claim 1, characterized in that, The keypoint prediction network adopts a lightweight encoder-decoder architecture consisting of a feature extraction backbone network and a scoring network; The scene-level SAR image is input into the keypoint prediction network to generate a keypoint mask, including: Scene-level SAR images are input into the backbone network, and multi-scale features are extracted using convolutional layers; Multi-scale features are input into the scoring network, and a scalar score is output for each pixel location. Through a unified sampling strategy and Top-K selection mechanism, a preset proportion of pixels are selected as key points to generate a key point mask.

3. The scene-level SAR image super-resolution method according to claim 1, characterized in that, The degenerate kernel estimation network adopts a U-Net-based encoder-decoder architecture.

4. The scene-level SAR image super-resolution method according to claim 1, characterized in that, The interpolation weight estimation network adopts an encoder-decoder architecture based on U-Net.

5. The scene-level SAR image super-resolution method according to claim 1, characterized in that, The three types of coupling degradation in SAR images include: resolution degradation, coherent speckle noise, and defocus degradation; The step of inputting low-resolution scene-level SAR images into a keypoint prediction network to generate keypoint masks includes, before the step of explicitly coupling a degradation kernel modeling step, specifically including: The degenerate kernel is mathematically represented as a convolution operation in the spatial domain. A quadratic term is introduced into the degenerate kernel representation, making it a weighted sum of linear and quadratic terms. The degenerate process is expressed as: in, This is a degraded, low-resolution image. and These represent the weight parameters for the linear and quadratic terms, respectively. The output of the linear convolution kernel. Indicates a degenerate nucleus. i , j These represent the pixel rows and columns of the image, respectively. For the original high-resolution image, k The kernel size is the convolution kernel size. This represents the position coordinates of one pixel.

6. The scene-level SAR image super-resolution method according to claim 1, characterized in that, In the query mechanism: The sparse degenerate priors are sorted using singular value decomposition and then the top-ranked priors are retrieved from the corresponding window to obtain candidate degenerate priors; The modulation parameters are calculated using an attention mechanism based on the candidate degradation priors. Specifically, this includes: mapping the feature vector and degradation kernel descriptor at each location in the U-Net to the query space and key-value space respectively via linear projection, where the query comes from the U-Net feature mapping, and the keys and values ​​come from the candidate degradation priors; processing the interpolation weight priors using depthwise separable pyramid convolutions and 1×1 convolutions, and injecting them as attention biases into the attention calculation process; the query... and keywords The attention weights between them are: in, For query and keywords Attention weights between them Attention bias derived from local interpolation weights. Let be the feature dimension. Scaling parameters are generated using modulation attention branch. To control the intensity of feature scaling, translation parameters are generated using a translation attention branch combined with interpolation weight bias. To adjust the feature offset; Through affine transformation The features are adaptively modulated so that each feature location receives different modulation parameters based on the degenerate prior of its query.

7. The scene-level SAR image super-resolution method according to claim 1, characterized in that, The sparse degenerate priors are sorted using singular value decomposition, and then the top-ranked priors are retrieved within the corresponding window to obtain candidate degenerate priors, including: The sparse degenerate priors are ranked globally by singular value decomposition to obtain a global ranking list; For each feature location in U-Net, a local search window is determined based on its corresponding HR spatial coordinates. Within the local search window, the top-ranked feature is selected from the global ranking list. Degenerate priors serve as a candidate set of degenerate priors; where It is adaptively determined based on the feature map scale and window size.

8. The scene-level SAR image super-resolution method according to claim 1, characterized in that, Based on the keypoint mask and the local interpolation weights, the keypoint degradation descriptor is reconstructed by interpolation to obtain a sparse degradation prior, including: Based on keypoint mask Check the position of each pixel Whether or not is the key point; If pixel position If it is a key point, then the pixel position The degenerate descriptors are extracted directly from the estimated keypoint degenerate descriptors; If pixel position If it is not a key point, then define a... The size of the center is A local window is defined, within which only the positions corresponding to keypoints have non-zero weights, while the weights of non-keypoint positions are zero. The non-keypoint pixel positions are reconstructed by calculating the weighted sum of the degenerate descriptors of the keypoints within the local window. Degenerate descriptor.

9. A scene-level SAR image super-resolution device, characterized in that, include: The sparse spatial variation degradation estimation module is used to input low-resolution scene-level SAR images into the key point prediction network to generate key point masks. The low-resolution scene-level SAR image and the keypoint mask are input into the degradation kernel estimation network to obtain the degradation descriptor of the keypoint; the low-resolution scene-level SAR image and the keypoint mask are input into the interpolation weight estimation network to obtain the local interpolation weight; based on the keypoint mask and the local interpolation weight, the degradation descriptor of the keypoint is interpolated and reconstructed to obtain the sparse degradation prior; The back-diffusion initialization module is used to encode low-resolution scene-level SAR images into low-resolution latent features using a low-resolution feature encoder. The initial latent noise state is obtained by sampling from a standard Gaussian distribution; The backdiffusion iteration module is used to input the sparse degenerate prior, low-resolution latent features and initial latent noise state into the conditional diffusion U-Net network with integrated query mechanism, and perform backdiffusion iteration guided by the sparse degenerate prior to update the latent features until the iteration ends and the final latent features are obtained. The query mechanism is used to treat the feature vector at each location in the U-Net as a query, retrieve the most relevant degradation information from the sparse degradation prior, and generate adaptive modulation parameters; The super-resolution SAR image generation module is used to decode the final latent features through the VQ-VAE decoder to obtain a super-resolution SAR image.

10. A computer device, comprising a memory and a processor, characterized in that, The memory stores a computer program, and when the processor executes the computer program, it implements the steps of the scene-level SAR image super-resolution method according to any one of claims 1 to 8.

Images