A remote sensing image blind super-resolution reconstruction method

Abstract

This invention discloses a blind super-resolution reconstruction method for remote sensing images, comprising: inputting a low-resolution remote sensing image into a pre-trained visual base model to obtain shallow degradation features and deep semantic features; using the shallow degradation features and deep semantic features as conditions to guide K iterations of reconstruction to obtain a high-resolution remote sensing image; wherein, the k-th iteration of reconstruction includes: generating a momentum state for the k-th iteration using a spatial adaptive convolution operator based on the output image features of the (k-1)-th and (k-2)-th iterations of reconstruction; inputting the momentum state and shallow degradation features into a gradient estimator to output intermediate hidden features; and inputting the intermediate hidden features and deep semantic features into a proximal mapping network to obtain the output image features of the k-th iteration of reconstruction. This invention can resolve the feature entanglement between physical degradation and semantic structure, achieving efficient propagation of stride features and high-fidelity remote sensing image reconstruction while maintaining strict mathematical interpretability.

Description

Technical Field

[0001] This invention relates to the fields of image processing and computer vision, and more specifically to a method for blind super-resolution reconstruction of remote sensing images. Background Technology

[0002] High-resolution (HR) remote sensing imagery plays a crucial role in Earth observation applications such as urban planning, environmental monitoring, and disaster management. However, due to sensor performance limitations and atmospheric interference, acquired low-resolution (LR) remote sensing images often suffer from complex and unknown degradation processes. Blind super-resolution (BSR) technology aims to reconstruct high-frequency details from severely degraded low-resolution inputs in the absence of prior knowledge about the degradation process.

[0003] To address the complex and varied image degradation problems in real-world scenarios, existing blind super-resolution methods mainly fall into two categories: explicit degradation estimation and implicit degradation representation. Explicit degradation estimation attempts to parameterize the degradation process by predicting spatially varying physical parameters such as the blur kernel, downsampling rate, and noise, guiding subsequent non-blind restoration networks. This method is effective under ideal assumptions, but real-world remote sensing degradation scenarios are complex and difficult to fully fit using simplified mathematical models. Deviating from the pre-defined simplified mathematical model can lead to biases in the prediction of physical parameters such as the blur kernel or noise, resulting in distortion and artifacts in the reconstructed image. To circumvent the limitations of explicit parameterization, implicit degradation representation transforms the estimation of degradation information into latent representation learning, encoding degradation features into a latent space. It then utilizes mechanisms such as contrastive learning or learnable cues to improve the model's generalization ability to unknown degradation. While this implicit degradation representation avoids parameterized modeling, it lacks strict physical constraints. In latent representation learning, the network struggles to maintain effective discriminative power in the latent space, which can lead to feature entanglement between degradation factors and key semantic elements of the image, making it impossible to achieve effective separation of physical degradation and image content.

[0004] In the field of deep unfolded networks, existing architectures combine physical interpretability with the representational power of deep learning, attracting widespread attention in recent years. This method unfolds classic iterative optimization algorithms such as Proximal Gradient Descent (PGD), alternating direction multiplier method, or semi-quadratic splitting into cascaded network layers. Through this structured decomposition, the highly ill-posed super-resolution problem can be transformed into two alternating optimization subproblems: a data fidelity step, primarily used to handle physical degradation processes and ensure that the generated image conforms to the distribution of observed data; and a proximal mapping step, serving as an implicit image prior or denoising module, mapping intermediate states to the manifold space of a clear, high-resolution image. This structurally decoupled design of degradation estimation and image content restoration significantly enhances the network's interpretability, and its iterative optimization logic is well-suited to the solution requirements of blind super-resolution tasks. However, existing architectures typically rely on data-driven, bottom-up internal prior designs when handling the proximal mapping subproblem. When remote sensing images suffer severe degradation and high-frequency textures are largely lost, relying solely on the input image's internal priors is insufficient to provide adequate structural constraints, easily leading to overly smooth reconstruction results or artifacts that do not conform to the real terrain. Furthermore, conventional unfolded networks often employ a sequential iterative strategy, limiting the receptive field across stages and hindering effective information transfer between iterative steps, which can cause slow model convergence and information transfer bottlenecks.

[0005] Unlike natural images, remote sensing images contain specific physical meanings and exhibit significant scale variations and dense spatial distribution features. This necessitates that blind super-resolution algorithms for remote sensing images restore high-frequency textures while maintaining accurate semantic information. Recently, to further achieve effective estimation and separation of physical degradation and image semantics, the underlying vision field has begun to explore the introduction of generative priors or multimodal knowledge (such as large vision-language models or diffusion models) to provide external guidance for restoration networks. These methods primarily use text-driven language cues to provide degradation feature guidance or semantic structure constraints when dealing with complex spatial distributions. Furthermore, when processing highly degraded images, the semantic priors of large models assist the network in completing texture synthesis and detail reconstruction. However, the applicability of introducing generative priors or multimodal knowledge in fine-grained remote sensing scenarios is limited. Remote sensing images typically contain dense and heterogeneous geographic entities, while language text descriptions provide relatively coarse-grained spatial constraints at the pixel level. This cross-modal guidance approach is prone to causing modal alignment bias in practical applications. When a network relies heavily on text prompts for texture synthesis, it may generate artifacts that do not match the actual geographic structure (such as distorting straight building edges), thus affecting the accuracy and usability of remote sensing images.

[0006] Therefore, how to resolve the feature entanglement between physical degradation and semantic structure, and achieve efficient propagation of step features and high-fidelity remote sensing image reconstruction while maintaining strict mathematical interpretability, is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0007] In view of the above problems, the present invention proposes a blind super-resolution reconstruction method for remote sensing images, which aims to overcome or at least partially solve the above problems.

[0008] To achieve the above objectives, the present invention adopts the following technical solution:

[0009] This invention provides a method for blind super-resolution reconstruction of remote sensing images, comprising:

[0010] Low-resolution remote sensing images are input into a pre-trained visual base model to obtain shallow degradation features and deep semantic features.

[0011] A deep unfolded network based on the proximal gradient descent algorithm is used, with the shallow degradation features and deep semantic features as conditions to guide K iterations of reconstruction, resulting in a high-resolution remote sensing image; wherein the k-th iteration of reconstruction includes:

[0012] Momentum extrapolation: Based on the output image features reconstructed in the (k-1)th iteration and the output image features reconstructed in the (k-2)th iteration, the momentum state of the kth iteration is generated using the spatial adaptive convolution operator;

[0013] Gradient descent: The momentum state of the kth iteration and the shallow degradation feature are input into the gradient estimator. The shallow degradation feature guides the gradient estimator to perform data fidelity constraints and outputs intermediate hidden features.

[0014] Proximal mapping: The intermediate hidden features and the deep semantic features are input into the proximal mapping network. The deep semantic features are used as prior information to modulate and fuse the intermediate hidden features to obtain the output image features reconstructed in the kth iteration.

[0015] Furthermore, the visual underlying model is the DINOv3 model.

[0016] Furthermore, the shallow degradation features are extracted from the shallow network of the visual base model, and the deep semantic features are extracted from the deep network of the visual base model.

[0017] Furthermore, before performing the K-th iteration reconstruction, the low-resolution remote sensing image is subjected to initial projection processing to obtain initial observation features;

[0018] The initial observation features are used as the output image features of the 0th iteration reconstruction, and the output image features of the -1st iteration reconstruction are set to be equal to the output image features of the 0th iteration reconstruction.

[0019] Furthermore, the momentum state of the k-th iteration is represented as:

[0020]

[0021] in, This represents the momentum state in the k-th iteration; This represents the output image features reconstructed in the (k-1)th iteration; This represents the output image features reconstructed in the (k-2)th iteration; Represents a spatially adaptive convolution operator.

[0022] Furthermore, the gradient estimator performs data fidelity constraints in the following manner:

[0023] The momentum state of the input k-th iteration is cross-attentioned with the shallow degradation feature, and the momentum state is guided by the shallow degradation feature to generate intermediate features containing degradation priors.

[0024] The difference between the intermediate features and the initial observation features of the low-resolution remote sensing image is calculated to obtain the data fidelity residual;

[0025] The self-attention mechanism is used to extract global features from the data fidelity residuals to obtain updated intermediate hidden features.

[0026] Furthermore, the gradient estimator employs a multi-head depthwise convolutional transposed attention mechanism to model spatial features in the cross-attention interaction and the self-attention mechanism, respectively.

[0027] Furthermore, the modulation and fusion of the intermediate hidden features using deep semantic features as prior information specifically includes:

[0028] The deep semantic features are input into a preset shared convolutional layer, and conditional guided vectors are obtained through feature mapping.

[0029] The conditional guiding vector is processed by two parallel convolutional branches to generate a spatial scale scaling factor and a spatial offset factor, respectively.

[0030] The intermediate hidden features are affinely modulated using the spatial scale scaling factor and spatial offset factor to obtain semantically modulated features.

[0031] The intermediate hidden features are used as the query matrix, and the semantic modulation features are used as the key matrix and value matrix. They are fused based on the cross-attention mechanism to obtain the output image features reconstructed in the kth iteration.

[0032] As can be seen from the above technical solution, compared with the prior art, the present invention discloses a blind super-resolution reconstruction method for remote sensing images, which has the following beneficial effects:

[0033] This invention utilizes hierarchical representations of a visual fundamental model to replace coarse-grained text prompts, alleviating the modal misalignment problem between text descriptions and pixel space, providing effective spatial structure constraints, thereby suppressing artifact generation and improving image reconstruction quality.

[0034] This invention proposes to transform traditional linear extrapolation weights into learnable spatially adaptive convolution operators, realizing the transformation of momentum terms from spatially invariant to spatially adaptive, and changing the information transmission method from linear combination to nonlinear mapping, enabling the network to extract features that are more conducive to subsequent gradient descent from previous iteration states.

[0035] This invention constructs a learnable gradient estimator based on a visual fundamental model and utilizes a multi-head deep convolution transpose attention mechanism to equivalently simulate the physical operations in traditional mathematical formulas. This not only preserves the strict physical interpretability of traditional optimization algorithms, but also achieves more accurate data fidelity than traditional black-box models under complex and unknown degradation conditions in the real world.

[0036] This invention proposes a near-end mapping network that combines external deep semantic modulation, effectively avoiding dependence on a single weak internal prior. Attached Figure Description

[0037] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0038] Figure 1 This is a schematic diagram of the blind super-resolution reconstruction method for remote sensing images provided in an embodiment of the present invention;

[0039] Figure 2 This is a schematic diagram of the overall framework of the remote sensing image blind super-resolution reconstruction network of the present invention;

[0040] Figure 3 This is a visualization diagram of the DINOv3 feature provided in an embodiment of the present invention;

[0041] Figure 4This is a schematic diagram of a degenerate kernel provided in an embodiment of the present invention;

[0042] Figure 5 This is a schematic diagram showing the visual effect comparison of blind super-resolution results provided in an embodiment of the present invention. Detailed Implementation

[0043] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0044] Existing super-resolution networks often rely on implicit feature decoupling or the introduction of textual cues. When handling complex blind super-resolution tasks of remote sensing images, these networks are prone to feature entanglement and artifacts caused by semantic misalignment. To address these issues, this invention provides a blind super-resolution reconstruction method for remote sensing images, such as... Figure 1 As shown, it includes the following steps:

[0045] Low-resolution remote sensing images are input into a pre-trained visual base model to obtain shallow degradation features and deep semantic features.

[0046] A deep unfolded network based on the proximal gradient descent algorithm is employed, using shallow degradation features and deep semantic features as conditions to guide K iterations of reconstruction, resulting in a high-resolution remote sensing image; wherein the k-th iteration of reconstruction includes:

[0047] Momentum extrapolation: Based on the output image features reconstructed in the (k-1)th iteration and the output image features reconstructed in the (k-2)th iteration, the momentum state of the kth iteration is generated using the spatial adaptive convolution operator;

[0048] Gradient descent: The momentum state of the kth iteration and the shallow degradation features are input into the gradient estimator. The shallow degradation features guide the gradient estimator to perform data fidelity constraints and output intermediate hidden features.

[0049] Proximal mapping: The intermediate hidden features and deep semantic features are input into the proximal mapping network. The deep semantic features are used as prior information to modulate and fuse the intermediate hidden features to obtain the output image features reconstructed in the kth iteration.

[0050] After K iterations, the final output image features are the reconstructed high-resolution remote sensing image.

[0051] Next, the overall framework of the designed method and each of the above-mentioned parts will be explained in detail.

[0052] I. Overall Methodological Framework:

[0053] Figure 2 The network architecture constructed in this embodiment is illustrated. Specifically, the input low-resolution observed image is known to be y, and this image has undergone an unknown degradation process. For example... Figure 2 As shown in Figure (a), the low-resolution observation image y is first input into the frozen DINOv3 model to extract multi-scale, multi-level feature representations. Based on the physical properties of multi-level features, the network constructed in this invention defines shallow features as degenerate representations, denoted as . Deep features are defined as semantic priors, denoted as... .

[0054] In the constructed visual prior-guided deep unfolding network, the input low-resolution image y is first projected onto the feature space through a standard convolutional layer to obtain initial observation features. Subsequently, the initial observed features are fed into a processing chain consisting of K cascaded iteration stages for sequential processing. To improve the network's training stability and convergence speed, a global residual skip connection is also employed.

[0055] In terms of specific module design, each of the K iteration stages strictly follows the optimization steps of the proximal gradient descent algorithm and the fast iterative shrinkage threshold algorithm for topology construction. Structurally, each iteration stage sequentially contains three cascaded sub-modules: momentum extrapolation module, gradient descent module, and proximal mapping module. Figure 2 Figures (b) and (c) in the diagram illustrate the detailed structures of the gradient descent module and the proximal mapping module, respectively. The hierarchical feature extraction process of the aforementioned visual basic model and the specific implementation mechanisms of the three modules are described below.

[0056] II. Hierarchical Feature Extraction from Visual Basic Models:

[0057] This invention uses a pre-trained and parameter-frozen DINOv3 model as the visual foundation model. DINOv3 obtains rich spatial topological representations through self-supervised learning on a massive dataset. Furthermore, because its pre-training data includes a large number of aerial and satellite remote sensing images, the model exhibits good domain adaptability to geographic entities in remote sensing scenes and is able to extract more refined texture features.

[0058] This invention utilizes the "hierarchical feature differentiation" characteristic within DINOv3 to divide the network's internal features into shallow degradation representations and deep semantic priors. Specifically, the extraction method involves freezing the input parameters of the low-resolution remote sensing image y in a pre-trained visual base model DINOv3, and extracting shallow degradation features from the model's shallow network. It is used to characterize physical degradation patterns; deep semantic features are extracted from the deep network of the model. Used to provide macroscopic semantic structure priors that are not affected by degradation.

[0059] To verify the effectiveness of this hierarchical approach, this invention employs Principal Component Analysis (PCA) to visualize the feature responses of high-resolution clear images and low-resolution degraded images at different network depths through dimensionality reduction. For example... Figure 3 As shown.

[0060] By comparing the response distribution of shallow features, it can be seen that (e.g.) Figure 3 The shallow feature maps (layers 5 and 11) can extract local texture details of the target when given a clear image input. However, when the input is a degraded image contaminated by noise and blur, the shallow feature maps exhibit disordered responses and high-frequency clutter, blurring the boundary between the target and the background. This phenomenon indicates that shallow networks are sensitive to pixel-level spatial perturbations, and their internal features retain the degrading components in the input image. Therefore, extracting the output of this layer as a degradation representation can capture the underlying physical degradation pattern of the image, providing a characterization basis for the degradation matrix estimation of the subsequent gradient descent module.

[0061] Observing the visualization heatmap of deep features can reveal (such as) Figure 3 The 17th and 23rd layer features in the super-resolution network exhibit response patterns different from those of the shallow features. Regardless of whether the input is a high-resolution, clear image or a low-resolution image severely degraded, the deep features maintain a relatively consistent activation pattern. Specifically, the deep features of the degraded image weaken the interference of high-frequency noise, enabling the separation of target entities from complex backgrounds, and the extracted contours are basically consistent with the deep responses of the high-resolution input. This indicates that the network's deep features focus on the macroscopic semantic topology and entity attributes of remote sensing targets, exhibiting strong robustness to complex physical degradation. Using these features as deep semantic priors for images can anchor the geometric structure of ground objects under severe degradation conditions, providing degrade-free semantic guidance for the near-end mapping process of the super-resolution network.

[0062] Existing technologies often implicitly mix degradation estimation and texture restoration within the same backbone network, easily leading to feature entanglement between degradation information and image semantics. This invention, however, leverages the multi-level feature representation characteristics of DINOv3 to explicitly decouple shallow degradation features and deep semantic features, thus improving the accuracy of degradation estimation and the effectiveness of semantic guidance. This avoids the feature entanglement problem caused by implicitly mixing degradation estimation and texture restoration within the same backbone network in existing technologies. Furthermore, compared to guidance schemes based on vision-language models (such as CLIP), DINOv3, as a purely visual foundational model, alleviates the modal misalignment problem between text description and pixel space. In blind super-resolution reconstruction, this method can extract visual features from aerial images based on the DINOv3 model, providing effective spatial structure constraints, thereby suppressing artifact generation and improving image reconstruction quality.

[0063] III. Conduct in-depth network initialization:

[0064] like Figure 2 As shown in (a), before performing K iterations of reconstruction, the low-resolution remote sensing image undergoes initial projection processing. Specifically, it passes through a learnable convolutional layer to map the input low-resolution remote sensing image from the image space to the feature space, thus obtaining the initial observation features. The initial observation features are used as the output image features of the 0th iteration reconstruction; at the same time, the output image features of the -1st iteration reconstruction are set to be equal to the output image features of the 0th iteration reconstruction.

[0065] Based on this, during the first iteration of reconstruction, the momentum increment calculated from the features of the output images from the 0th and -1st iterations is zero, and the algorithm performs the standard proximal gradient descent step; starting from the second iteration of reconstruction, the momentum extrapolation is performed normally using the difference between the features of the previous iterations.

[0066] IV. Momentum extrapolation module based on fast iterative shrinkage threshold:

[0067] To accelerate convergence and alleviate the information transfer bottleneck between cascaded layers in deep unfolded networks, this method introduces an extrapolation module at the initial position of each iteration stage. The classic Fast Iterative Shrinkage-Thresholding Algorithm (FISTA) typically uses an inertial extrapolation step to accelerate the proximal gradient descent process. Taking the k-th iteration as an example, its standard mathematical form can be expressed as:

[0068]

[0069] in, This represents the momentum state at the k-th iteration and can be used as an auxiliary starting point for inputting the k-th iteration stage; This represents the output image features reconstructed in the (k-1)th iteration; This represents the output image features reconstructed in the (k-2)th iteration; This represents a predefined scalar weight, i.e., a linear scalar combination coefficient, used to control the impact of the difference between the states of the previous two iterations (i.e., the momentum term) on the current update.

[0070] However, traditional scalar combination methods have limitations when dealing with complex blind super-resolution tasks of remote sensing images. This is due to the limitations of scalar weights. It exhibits spatial invariance, meaning it applies a uniform fusion weight to all pixels within an image, thus failing to effectively distinguish between smooth backgrounds and the edges of complex geographic entities. This linear combination method limits the network's ability to represent nonlinear complex features, making it difficult to adaptively preserve crucial momentum information while suppressing noise.

[0071] To address the limitation of flexibility in scalar combination mentioned above, this invention proposes to modify the traditional linear extrapolation weights. This is transformed into a learnable spatially adaptive convolution operator. By parameterizing this scalar weight as a convolutional layer, the extrapolation step is reconstructed as follows:

[0072]

[0073] in, This invention introduces a spatially adaptive convolution operator in a standard convolutional layer to replace traditional scalar weights. .

[0074] The improvements in the above design are mainly reflected in two aspects. First, the momentum term achieves a transformation from spatial invariance to spatial adaptation. By introducing convolution operations, the network can dynamically adjust the momentum fusion intensity based on differences in local image features (such as edges, textures, or flat regions). This is particularly effective in regions with drastic changes in image structure, where the spatially adaptive convolution operator... It can extract deep nonlinear residuals, thereby generating more effective auxiliary starting points. Secondly, the information transmission method changes from linear combination to nonlinear mapping. Convolutional layers combined with activation mechanisms introduce nonlinear modeling capabilities for cross-stage information transmission, enabling the network to extract features from previous iteration states that are more conducive to subsequent gradient descent.

[0075] V. Gradient Descent Module Based on Shallow Degradation Priors:

[0076] After feature extrapolation, the generated momentum state is fed into the gradient descent module. The core task of this module is to solve the "data fidelity" problem in image reconstruction, i.e., minimizing the error between the real low-resolution image and the degraded reconstructed image. In classical mathematical optimization, the data fidelity term is typically modeled as a least-squares energy functional. :

[0077]

[0078] in, Let be the known physical degradation matrix; x represent the high-resolution image updated in the iteration; and y represent the observation data of the low-resolution image. Taking the derivative of this energy functional yields its gradient: Therefore, based on the prior momentum state Using this as an auxiliary starting point, the standard gradient descent feature update process can be rigorously described as follows:

[0079]

[0080] in, This represents the iteration step size for gradient descent. This is the gradient term, used to guide the update of the current state.

[0081] However, in blind super-resolution tasks of remote sensing images, the actual degradation process is extremely complex and completely unknown, which means that the physical degradation matrix in the above classical formula is... and its transpose These problems are difficult to solve directly. Traditional deep unfolded networks often use conventional convolutional layers for unconstrained black-box fitting, which is prone to directional deviation when facing complex remote sensing degradation scenarios. This invention abandons the direct fitting of unknown physical matrices and instead constructs a learnable gradient estimator based on a visual fundamental model. It extracts shallow degradation features that are sensitive to degradation information from stage one. As a condition-guided feature injection, the gradient descent feature update formula of this invention evolves into:

[0082]

[0083] in, The initial observation features are obtained after initial projection of low-resolution observation images in the network, corresponding to the observation value y in the mathematical formula. To complete the intermediate hidden features after the data fidelity constraints of this round.

[0084] The specific network implementation architecture is as follows: Figure 2 As shown in (b), the gradient estimator of this invention utilizes a multi-head depthwise convolution transposed attention mechanism to equivalently simulate the physical operations in traditional mathematical formulas. The specific execution flow is as follows: First, the momentum state of the kth iteration is input... With shallow degradation characteristics Cross-attention interaction is performed, using shallow degradation features to guide the momentum state. This operation dynamically assembles the degradation matrix using prior knowledge, which is physically equivalent to performing [the following action / operation]. The degraded projection is used to generate intermediate features containing the degradation prior; subsequently, the difference between the intermediate features and the initial observed features of the low-resolution remote sensing image is calculated. Specifically, the estimated intermediate features containing the degradation prior can be derived from the initial observed features. Subtracting from the middle yields an accurate data-fidelity residual; finally, a self-attention mechanism is used to extract global deep features from the data-fidelity residual, thereby equivalently simulating the transpose matrix in the traditional formula. The mapping operation obtains the updated intermediate hidden features. This one-to-one correspondence network design not only preserves the strict physical interpretability of traditional optimization algorithms, but also achieves more accurate data fidelity than traditional black-box models under complex and unknown degradation conditions in the real world.

[0085] VI. Proximal mapping module combined with deep semantic modulation:

[0086] After completing the data fidelity step of gradient descent, the output intermediate hidden features While it achieves observational consistency in the low-frequency space, it inevitably suffers from severe random noise, ringing artifacts, and structural distortion. Therefore, this intermediate hidden feature... It needs to be further processed by the near-end mapping module.

[0087] In the classic proximal gradient descent algorithm, the proximal mapping step essentially solves a subproblem of maximum a posteriori inference. Its purpose is to introduce a prior regularization term regarding the image structure, pulling the corrupted intermediate states back to the true manifold space of the clear, high-resolution image. Mathematically, this is equivalent to solving the following energy minimization problem:

[0088]

[0089] in, It constrains the approximation between the input and the output; It serves as a structural prior for the image, or a regularization term; This represents the equilibrium coefficient. Physically, solving this equation is equivalent to solving for the intermediate hidden features. Apply a denoiser or feature optimizer.

[0090] However, traditional deep unfolded networks have some limitations in solving this subproblem. They typically parameterize the proximal operator as a purely data-driven bottom-up neural network, such as conventional ResNet or U-Net. This means that the network can only utilize the "internal prior" extracted from the degraded image itself for restoration. When remote sensing images suffer severe degradation and high-frequency textures are completely erased, this internal prior, which relies on damaged input, struggles to provide accurate structural guidance, making it difficult for the network to simultaneously "suppress complex artifacts" and "restore high-frequency details," ultimately resulting in overly smooth or artifact-laden reconstructions.

[0091] To avoid dependence on a single weak internal prior, this invention proposes a proximal mapping network that combines external deep semantic modulation. The deep semantic features extracted from the pure visual basic model in Phase 1 have high robustness against quality degradation. This is introduced as a macroscopic topological constraint into the prior distribution optimization. Therefore, the near-end mapping process of this invention is reconstructed as a conditional distribution mapping:

[0092]

[0093] The specific implementation architecture is as follows: Figure 2 As shown in (c), firstly, deep semantic features are extracted. The input is fed into a pre-defined shared convolutional layer, where a conditional guided vector is obtained through feature mapping. Then, two parallel convolutional branches further process this conditional guided vector. The first branch is used to generate a spatial scale scaling factor. (Scaling factor), the second branch is used to generate the spatial offset factor. (Shifting factor). These two sets of parameters characterize the strength and offset of the influence of semantic information on image features in the spatial dimension. Finally, the spatial scale scaling factor is used... and spatial offset factor Hidden features in the middle Pixel-by-pixel affine modulation is performed to explicitly inject semantic information, thereby obtaining semantic modulation features. The modulation process is shown in the following equation:

[0094]

[0095] in, This represents element-wise multiplication. This step uses affine transformation to initially couple semantic priors with spatial structural features, enhancing the network's ability to perceive ground object attributes.

[0096] To correct semantic misalignment and suppress structural artifacts that may result from a single affine transformation, this invention modulates semantic features. With intermediate hidden features The data is fed into a cross-attention-based fusion module. In this module, intermediate hidden features are... Construct a query matrix to preserve the original topological structure of the image; use semantic modulation features. A key matrix and a value matrix are constructed to provide semantically enhanced texture cues. Through a cross-attention mechanism, the network can adaptively aggregate high-frequency details from semantic features based on the index of the original structure, thereby improving the reconstruction effect and obtaining the output image features of the k-th iteration reconstruction.

[0097] The momentum extrapolation module, gradient descent module, and proximal mapping module are sequentially connected to form a complete unfolded iterative cascade. After K iterations of optimization, the low-resolution image is finally output as a restored high-quality super-resolution image.

[0098] VI. Technical Effects and Experimental Verification:

[0099] To verify the effectiveness of the method provided in this invention, a systematic quantitative experiment was conducted on the standard remote sensing dataset UCMerced. The experiment simulated the complex and varied degradation process in real-world scenes and was comprehensively compared with current mainstream implicit degradation representation methods, deep unfolded network methods, and generative blind super-resolution methods based on diffusion models. The comparison methods included DASR, DRSR, BSRSNF, DSAT, CDFormer, LBSR, and DiffBSR. Among them, DRSR, BSRSNF, and DiffBSR are specifically designed for remote sensing imagery, with the latter two primarily utilizing diffusion models to improve the perceived quality of the images. To ensure fairness in the comparison, all models were retrained under the same degradation settings and remote sensing dataset.

[0100] As shown in Table 1, Table 2 and Figure 4 As shown in Tables 1 and 2, the quantitative evaluation results of each method on the UCMerced test set are presented. Eight degradation kernels and two noise intensities were selected to simulate different types and degrees of degradation during remote sensing image acquisition and transmission. Specifically, degradation kernels 1 to 3 are isotropic Gaussian kernels of different sizes, while degradation kernels 4 to 8 are anisotropic Gaussian kernels. These anisotropic kernels exhibit non-uniform diffusion in different directions, mainly controlled by the standard deviation of the two principal axes and the rotation angle. Evaluation metrics include Peak Signal-to-Noise Ratio (PSNR), Structural Similarity (SSIM), and Learned Perceived Patch Similarity (LPIPS). Higher PSNR and SSIM values ​​indicate higher pixel-level fidelity; lower LPIPS values ​​indicate better subjective perceived quality.

[0101] Table 1. Quantitative comparison of the performance of different methods on different kernel functions at a noise level of 10.0.

[0102]

[0103] Table 2. Quantitative comparison of the performance of different methods on different kernel functions at a noise level of 20.0.

[0104]

[0105] As shown in Tables 1 and 2, our proposed method achieved the best PSNR and SSIM across all eight degenerate kernels. Specifically, at a noise level of 10.0, our method achieved a PSNR of 25.7014 dB on degenerate kernel 1, outperforming the second-best method, CDFormer (25.6416 dB). At a noise level of 20.0, our method achieved a PSNR of 24.4402 dB on degenerate kernel 1, outperforming the second-best method, DSAT (24.3257 dB), under the same conditions. Furthermore, comparing the results with different degenerate kernels reveals that the performance degradation of most baseline models is closely related to the specific parameters of the Gaussian kernel. For example, a Gaussian kernel with a larger standard deviation along the principal axis corresponds to more severe and widespread blurring, leading to a significant decrease in model performance. This phenomenon indicates that traditional methods have limitations in robustness when dealing with severe and variable blur degradation. In contrast, our proposed method maintained stable performance improvements across different levels of degenerate kernels, validating its effectiveness in maintaining accurate spatial structure and recovering high-frequency details under complex degradation modes.

[0106] Although our method performs well in pixel-level fidelity metrics, it still lags behind diffusion-based methods in perceptual quality, measured by LPIPS. For example, while the PSNR and SSIM values ​​of the BSRSNF method are significantly lower than other comparative methods, its LPIPS score is the lowest across all degenerate kernels and noise levels. Specifically, on degenerate kernel 1 with a noise level of 10.0, BSRSNF achieves an LPIPS of 0.2083 and a PSNR of 24.0619 dB, while our method achieves 0.3081 and a PSNR of 25.7014 dB. This is because diffusion-based generative methods (such as BSRSNF and DiffBSR) optimize perceptual quality through synthetic textures but sacrifice structural fidelity and pixel-level alignment, which are features that our method focuses on preserving. Notably, our method's LPIPS score remains competitive among similar methods that optimize pixel-level alignment, validating its superior overall performance.

[0107] Figure 5Qualitative comparisons of various methods under complex degradation conditions are presented. The experiment selected an anisotropic degradation kernel (degradation kernel 8) with an added noise intensity of 10, and the core region within the red rectangle is magnified to compare local reconstruction details. Observation Figure 5 As can be seen from the triangular roof region in the building image in (a), most contrast methods exhibit significant distortion and structural falsification when restoring edges, failing to accurately reconstruct the original geometric contours. Furthermore, while the generative method based on the diffusion model synthesizes some texture, it fails to effectively constrain physical edges. In contrast, our proposed method effectively suppresses artifacts caused by complex degradation, resulting in smoother and more continuous roof edges. The overall geometric contour largely matches the real high-resolution image, indicating that this model has certain advantages in constraining spatial structure and maintaining edge consistency.

[0108] Figure 5 Figure (b) shows the comparison results of the playground image. After being severely blurred by a degradation kernel of 2 and degraded by noise of intensity 20, the reconstruction results of most comparison methods showed varying degrees of topological destruction or feature loss. Specifically, the boundary line reconstructed by the DSAT method showed obvious breaks on the right side, failing to maintain continuity; methods such as CDFormer, DiffBSR, and BSRSNF produced structural misalignment and deviation, distorting the straight tennis court boundary into an irregular curve. In contrast, our method, guided by explicit semantic features, avoided line distortion, and the restored boundary contour was closer to the real rectangular structure. The reconstructed boundary lines were continuous and conformed to the real physical distribution, further verifying the fidelity and effectiveness of our method in high-frequency detail reconstruction of complex remote sensing scenes.

[0109] In summary, this invention provides a blind super-resolution reconstruction method for remote sensing images, aiming to resolve the feature entanglement between physical degradation and semantic structure. This method utilizes hierarchical representations of a pure visual base model to replace coarse-grained textual cues, extracting shallow features sensitive to degradation information and deep features robust to degradation and rich in semantics. These two types of features are decoupled and injected into the gradient descent module and proximal mapping module of the unfolded network, respectively. Furthermore, addressing the iterative information bottleneck of the unfolded network, this invention introduces an inertial extrapolation mechanism based on a fast iterative shrinkage threshold algorithm, achieving efficient propagation of step features and high-fidelity remote sensing image reconstruction while maintaining strict mathematical interpretability.

[0110] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0111] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for blind super-resolution reconstruction of remote sensing images, characterized in that, include: Low-resolution remote sensing images are input into a pre-trained visual base model to obtain shallow degradation features and deep semantic features. A deep unfolded network based on the proximal gradient descent algorithm is used, with the shallow degradation features and deep semantic features as conditions to guide K iterations of reconstruction, resulting in a high-resolution remote sensing image; wherein the k-th iteration of reconstruction includes: Momentum extrapolation: Based on the output image features reconstructed in the (k-1)th iteration and the output image features reconstructed in the (k-2)th iteration, the momentum state of the kth iteration is generated using the spatial adaptive convolution operator; Gradient descent: The momentum state of the kth iteration and the shallow degradation feature are input into the gradient estimator. The shallow degradation feature guides the gradient estimator to perform data fidelity constraints and outputs intermediate hidden features. Proximal mapping: The intermediate hidden features and the deep semantic features are input into the proximal mapping network. The deep semantic features are used as prior information to modulate and fuse the intermediate hidden features to obtain the output image features reconstructed in the kth iteration. 2.The blind super-resolution reconstruction method of remote sensing image of claim 1, wherein, The visual underlying model is the DINOv3 model. 3.The blind super-resolution reconstruction method of remote sensing image of claim 1, wherein, The shallow degradation features are extracted from the shallow network of the visual base model, and the deep semantic features are extracted from the deep network of the visual base model. 4.The blind super-resolution reconstruction method of remote sensing image of claim 1, wherein, Before performing the K iterations of reconstruction, the low-resolution remote sensing image is subjected to initial projection processing to obtain initial observation features; The initial observation features are used as the output image features of the 0th iteration reconstruction, and the output image features of the -1st iteration reconstruction are set to be equal to the output image features of the 0th iteration reconstruction. 5.The blind super-resolution reconstruction method of remote sensing image of claim 1, wherein, The momentum state in the k-th iteration is represented as: ； wherein, denotes the momentum state of the kth iteration; denotes the output image feature of the k-1th iteration reconstruction; denotes the output image feature of the k-2th iteration reconstruction; denotes a spatially adaptive convolution operator.

6. The blind super-resolution reconstruction method for remote sensing images as described in claim 4, characterized in that, The gradient estimator performs data fidelity constraints in the following manner: The momentum state of the input k-th iteration is cross-attentioned with the shallow degradation feature, and the momentum state is guided by the shallow degradation feature to generate intermediate features containing degradation priors. The difference between the intermediate features and the initial observation features of the low-resolution remote sensing image is calculated to obtain the data fidelity residual; The self-attention mechanism is used to extract global features from the data fidelity residuals to obtain updated intermediate hidden features.

7. The blind super-resolution reconstruction method for remote sensing images as described in claim 6, characterized in that, The gradient estimator internally employs a multi-head depthwise convolutional transposed attention mechanism to realize spatial feature modeling in the cross-attention interaction and the self-attention mechanism, respectively.

8. The blind super-resolution reconstruction method for remote sensing images as described in claim 1, characterized in that, The modulation and fusion of the intermediate hidden features using deep semantic features as prior information specifically includes: The deep semantic features are input into a preset shared convolutional layer, and conditional guided vectors are obtained through feature mapping. The conditional guiding vector is processed by two parallel convolutional branches to generate a spatial scale scaling factor and a spatial offset factor, respectively. The intermediate hidden features are affinely modulated using the spatial scale scaling factor and spatial offset factor to obtain semantically modulated features. The intermediate hidden features are used as the query matrix, and the semantic modulation features are used as the key matrix and value matrix. They are fused based on the cross-attention mechanism to obtain the output image features reconstructed in the kth iteration.

Images