Diffusion model based super-resolution method for remote sensing image with high texture fidelity and few sampling steps
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- 烟台哈尔滨工程大学研究院
- Filing Date
- 2026-07-09
- Publication Date
- 2026-08-07
AI Technical Summary
[0007]本发明提出了一种基于扩散模型的少采样步高纹理保真度遥感图像超分辨率方法，其目的是：解决现有扩散模型超分辨率方法采样效率低、训练稳定性差、纹理保真度不足的问题，实现仅需少量采样步即可完成高质量遥感图像超分辨率重建
[0030]1.本发明通过在前向过程中引入残差调度参数和由该参数推导得出的噪声调度参数，以残差为核心驱动、噪声为中间调节项，重构了扩散模型的变换逻辑。反向过程利用上述参数与预测模型输出的超分辨率重建图像进行迭代重构，能够在极少数迭代步内完成从低分辨率插值图像到高分辨率图像的重建。相较于传统扩散模型需要数百至上千个采样步的处理方式，本发明仅需5个采样步即可完成高质量超分辨率重建，有效降低了推理耗时，使采样效率得到实质改善。
Smart Images
Figure CN122529976A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of image data processing technology, and in particular to a method for super-resolution of remote sensing images with low sampling step and high texture fidelity based on a diffusion model. Background Technology
[0002] With the development of remote sensing technology, super-resolution reconstruction of remote sensing images has become an important research direction in the field of image data processing. Improving the resolution of remote sensing images through super-resolution processing can provide richer spatial information for downstream tasks such as ship target detection, biological change monitoring, vehicle tracking, and target classification, thereby significantly improving task performance.
[0003] In deep learning-driven super-resolution methods, regression model-based approaches were among the earliest techniques. These methods typically train the network using L1 or L2 loss functions, effectively preserving pixel accuracy by directly minimizing the pixel-level deviation between the reconstructed image and the true high-resolution image. However, the reconstruction results from these methods generally suffer from over-smoothing and lack high-frequency texture details, making it difficult to meet the texture fidelity requirements of remote sensing image interpretation.
[0004] Super-resolution methods based on Generative Adversarial Networks (GANs) have made progress in improving the visual quality of reconstructed images by introducing perceptual loss and adversarial training mechanisms. Compared with regression model methods, GAN methods can generate richer texture details. However, GAN methods themselves suffer from training instability and are prone to generating artifacts, which can even lead to training failure in severe cases. This limits their reliability in remote sensing image super-resolution tasks to some extent.
[0005] In recent years, diffusion models have shown potential in the field of image super-resolution. Diffusion models generate images through two processes: forward noise addition and backward denoising, resulting in more realistic texture details. However, existing diffusion model-driven super-resolution methods still have significant shortcomings: First, they have low sampling efficiency; traditional diffusion models typically require approximately 1000 sampling steps to complete reconstruction, leading to significant time consumption during training and inference. Second, they lack training stability, are highly dependent on the dataset, and have poor generalization ability on small-scale remote sensing image datasets. Third, it is difficult to balance texture fidelity and efficiency; some optimization methods reduce sampling steps at the expense of texture restoration quality.
[0006] Therefore, there is an urgent need for a remote sensing image super-resolution method that requires fewer sampling steps, has stable training, and maintains high texture fidelity. Summary of the Invention
[0007] This invention proposes a high-texture-fidelity remote sensing image super-resolution method based on a diffusion model with few sampling steps. Its purpose is to solve the problems of low sampling efficiency, poor training stability, and insufficient texture fidelity of existing diffusion model super-resolution methods, and to achieve high-quality remote sensing image super-resolution reconstruction with only a few sampling steps.
[0008] The technical solution of this invention is as follows:
[0009] A low-sampling-step high-texture-fidelity remote sensing image super-resolution method based on a diffusion model includes: defining a forward process and a corresponding reverse process, wherein the forward process is used to progressively convert a high-resolution image into a low-resolution image, and the reverse process is used to reconstruct a high-resolution image from a low-resolution image through iterative reconstruction; and using a prediction model to predict the super-resolution reconstructed image based on the current state.
[0010] In the forward process, a residual scheduling parameter and a noise scheduling parameter are introduced. The residual scheduling parameter is used to control the accumulation rate of the residual between the high-resolution image and the low-resolution image. The noise scheduling parameter is derived from the residual scheduling parameter, and the noise intensity is first enhanced and then attenuated in the forward process.
[0011] The reverse process takes a low-resolution interpolated image as the initial input and performs iterative reconstruction based on the residual scheduling parameters, the noise scheduling parameters, and the super-resolution reconstructed image output by the prediction model.
[0012] The prediction model uses the UNet architecture, in which the encoding layer, intermediate layer and decoding layer are all built on the noise residual attention module. The noise residual attention module performs preliminary feature extraction and current time step fusion on the input features, and then performs parallel processing through the first branch, the second branch and the third branch respectively. The first branch adopts multi-scale feature extraction and fusion self-attention mechanism, the second branch performs convolution processing, and the third branch enhances key feature extraction through the convolution block attention module. The outputs of the three branches are fused and then connected by residual to obtain the output features.
[0013] The low-resolution interpolated image to be super-resolution is used as the initial input of the reverse process. Starting from the maximum time step, the reverse iteration is performed sequentially to obtain the high-resolution reconstructed image.
[0014] As a further improvement to the diffusion model-based few-sampling-step high-texture-fidelity remote sensing image super-resolution method, the residual scheduling parameter The calculation formula is ,in For the current time step, The total number of time steps in the forward process. The parameter used to control the rate of change of the residual.
[0015] As a further improvement to the diffusion model-based few-sampling-step high-texture-fidelity remote sensing image super-resolution method, the noise scheduling parameter The expression is: when hour ,when hour ,in For the current time step, The residual scheduling parameters are... This is a noise switching node.
[0016] As a further improvement to the diffusion model-based few-sampling-step high-texture-fidelity remote sensing image super-resolution method, the expression for the forward process is: ,in For the forward process, the first Intermediate state images at each time step For high-resolution images, This is a low-resolution interpolated image obtained by interpolating a low-resolution image. The residual scheduling parameters are... The noise scheduling parameters are... This is the noise amplitude coefficient. This refers to random noise that follows a standard normal distribution.
[0017] As a further improvement to the diffusion model-based few-sampling-step high-texture-fidelity remote sensing image super-resolution method, the iterative formula for the reverse process is: ,in For the first step in the reverse process Intermediate state images at each time step For the first Intermediate state images at each time step, The prediction model is based on The predicted super-resolution reconstructed image, and The first and the The residual scheduling parameters for the time step, and The first and the The noise scheduling parameters for the time step, This is the noise amplitude coefficient. This refers to random noise that follows a standard normal distribution.
[0018] As a further improvement to the diffusion model-based few-sampling-step high-texture-fidelity remote sensing image super-resolution method, the processing flow of the noise residual attention module includes:
[0019] For input features Perform twice in sequence Features are obtained through convolution and ReLU activation function processing. ;
[0020] Set the current time step Through two parallel multilayer perceptrons and After processing and Fusion, resulting in features ,in This indicates element-wise multiplication. This indicates element-wise addition;
[0021] The first branch will through After adjusting the number of channels in the convolution, respectively through , and The convolution yields three sub-branch features. Each sub-branch feature is processed by the self-attention module and then multiplied element-wise with the feature of its sub-branch. The results of the three sub-branch features are then concatenated. Convolution yields the first branch output;
[0022] The second branch pair conduct Convolution yields the second branch output;
[0023] The third branch is processed by the convolutional block attention module. The third branch output is obtained;
[0024] The first branch output, the second branch output, and the third branch output are concatenated and then... Convolution adjustment, then combined with input features conduct The results of convolution and ReLU activation are added element by element, and after ReLU activation, the final output of the noise residual attention module is obtained.
[0025] As a further improvement to the diffusion model-based low-sampling-step high-texture-fidelity remote sensing image super-resolution method, the coding layer of the prediction model is composed of a combination of the noise residual attention module and the downsampling module, the intermediate layer is composed of the noise residual attention module, and the decoding layer is composed of a combination of transposed convolution and the noise residual attention module.
[0026] As a further improvement to the diffusion model-based few-sampling-step high-texture-fidelity remote sensing image super-resolution method, the training process of the prediction model includes: constructing training sample pairs comprising high-resolution and low-resolution images; for each training sample pair, from the total time step... A uniformly random sample of the current time step ,according to The residual scheduling parameters and noise scheduling parameters are calculated and substituted into the forward process to obtain the degraded image. The degraded image and the low-resolution interpolated image are concatenated along the channel dimension and then input into the prediction model to obtain the super-resolution reconstructed image. The weight parameters of the prediction model are updated based on the loss function through gradient backpropagation.
[0027] As a further improvement to the diffusion model-based few-sampling-step high-texture-fidelity remote sensing image super-resolution method, the loss function is a composite loss function of L1 loss and perceptual loss. ,in For L1 loss, In order to perceive loss, The weighting factor is used; the L1 loss is used to constrain the pixel-level deviation between the reconstructed image and the real high-resolution image, and the perceptual loss is used to constrain the semantic consistency between the two in the high-level feature space of the pre-trained convolutional neural network.
[0028] As a further improvement to the diffusion model-based few-sampling-step high-texture-fidelity remote sensing image super-resolution method, the calculation formula for the perceptual loss is as follows: ,in This represents a high-level feature map extracted by a pre-trained VGG16 convolutional neural network. The number of training samples, For the first High-resolution images of training sample pairs, For the first A training sample pair of super-resolution reconstructed images obtained through the forward process and the prediction model.
[0029] Compared with the prior art, the present invention has the following beneficial effects:
[0030] 1. This invention reconstructs the transformation logic of the diffusion model by introducing a residual scheduling parameter and a noise scheduling parameter derived from it into the forward process. The residual is the core driver, and noise is used as an intermediate adjustment term. The backward process uses these parameters and the super-resolution reconstructed image output by the prediction model for iterative reconstruction, enabling the reconstruction from a low-resolution interpolated image to a high-resolution image within a very small number of iterations. Compared to traditional diffusion models that require hundreds to thousands of sampling steps, this invention only requires 5 sampling steps to complete high-quality super-resolution reconstruction, effectively reducing inference time and substantially improving sampling efficiency.
[0031] 2. This invention constructs the diffusion process as a directional transformation between a high-resolution image and a low-resolution interpolated image, with the residual accumulation rate explicitly controlled by residual scheduling parameters during the forward pass. Within this framework, the noise intensity controlled by the noise scheduling parameters first increases and then decreases as the transformation progresses. Stronger noise is introduced in the later stages of the forward pass to cover uncertainties in the transformation process, while maintaining a lower noise level near the high-resolution and low-resolution images. This strategy, driven deterministically by residuals and with noise only playing a supplementary regulatory role in the later stages of the transformation, avoids the adverse effects of purely random noise-driven methods on training stability, resulting in smoother convergence during training and enhanced generalization ability for small-scale remote sensing image datasets.
[0032] 3. This invention embeds a noisy residual attention module into the UNet architecture of the prediction model. It extracts texture structure features from different receptive fields using convolutional kernels of various scales, and utilizes a convolutional block attention module to activate attention to key texture regions through channel attention and spatial attention mechanisms, while also strengthening global feature dependencies through a self-attention mechanism. The synergistic effect of multi-scale feature extraction and the attention mechanism enables the prediction model to continuously recover high-frequency texture details during time-step reconstruction.
[0033] 4. The composite loss function combines L1 loss with perceptual loss based on a pre-trained VGG16 network. It imposes additional constraints on the alignment of high-level semantic features while constraining pixel-level accuracy, effectively suppressing over-smoothing and artifact generation, and ensuring that the texture structure of the reconstructed image is highly consistent with that of the real high-resolution image. Attached Figure Description
[0034] Figure 1 This is a diagram of the predictive model framework.
[0035] Figure 2 This is a schematic diagram of the Noise Residual Attention (NRA) module.
[0036] Figure 3A visualization comparison of the 4x super-resolution results of this method and some comparative methods on a single image of the NWPU-RESISC45 dataset;
[0037] Figure 4 This is a visualization comparison of the super-resolution results using different residual change rate parameters and noise switching nodes in this method. Detailed Implementation
[0038] The technical solution of the present invention will now be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.
[0039] This embodiment uses the NWPU-RESISC45 remote sensing image dataset as an example to illustrate the method of the present invention. The NWPU-RESISC45 dataset contains 45 remote sensing scene categories. In this embodiment, the first 200 images of each category, totaling 9000 images, are selected for the super-resolution experiment. Specifically, 40 images are selected from each category as the test set, and 160 images are selected as the training set. Each image has a pixel size of 256×256, serving as the initial high-resolution image. For high-resolution images Downsampling was performed using bicubic interpolation to obtain a low-resolution image with a pixel size of 64×64. Thus constitutes the training materials Image pairs.
[0040] I. Forward and Reverse Processes
[0041] First, a forward process based on residual scheduling and dynamic noise adjustment mechanism is defined, as well as a corresponding reverse process. The forward process is used to progressively transform a high-resolution image into a low-resolution image; the reverse process is used to start from the low-resolution image, utilize the derived residual scheduling parameters, noise scheduling parameters, and the super-resolution reconstructed image output by the prediction model, and achieve high-resolution image reconstruction through iterative reconstruction, significantly reducing the number of sampling steps required.
[0042] Forward process with high-resolution images Using the initial input, through residual accumulation and dynamic noise adjustment, The image is gradually converted to a lower resolution image. The expression for the forward process is:
[0043]
[0044] In the above formula, Indicates the first step in the forward process Intermediate state images at each time step; For the initial high-resolution image; For low-resolution images The resolution of the image obtained by bicubic interpolation is the same as... Consistent; This is a residual scheduling parameter used to control the accumulation rate of residuals between high-resolution and low-resolution images; These are noise scheduling parameters used to control the intensity of the added noise; To follow a standard normal distribution Random noise; This is the noise amplitude coefficient.
[0045] Among them, residual scheduling parameters The formula for calculation is:
[0046]
[0047] Noise scheduling parameters Based on residual scheduling parameters The expression is derived as follows:
[0048]
[0049] In the above two formulas, This represents the total number of time steps in the forward process; Parameters used to control the rate of change of residuals; These are the switching node parameters for noise increase / decrease. Therefore, the noise scheduling parameters... Its characteristic lies in: in time step Less than switching nodes At that time, the noise intensity increased with The increase in time step enhances the effect; Reaching or exceeding At that time, the noise intensity increased with As the noise increases, it attenuates; the noise component exhibits a dynamic trend of first increasing and then decreasing throughout the forward propagation process. To switch nodes.
[0050] The reverse process uses low-resolution images with bicubic interpolation. Using the super-resolution reconstructed image output by the prediction model as initial input, and the aforementioned residual scheduling parameters and noise scheduling parameters, iterative reconstruction is achieved from... Super-resolution reconstruction of high-resolution images. The derivation of the reverse process is as follows.
[0051] First, according to the forward process Write out the expression. and The expression:
[0052]
[0053]
[0054] In the above formula Replace with a super-resolution reconstructed image predicted by the prediction model based on the current state. Then Substituting the expression From the expression, we obtain the iterative formula for the reverse process:
[0055]
[0056] In the above formula, Indicates the first step in the reverse process Intermediate state images at each time step; For prediction models based on The predicted super-resolution reconstructed image.
[0057] II. Prediction Model
[0058] Creating a predictive model ,in These are the network weight parameters. The input to the prediction model is a channel-stitched image of the intermediate state image at the current time step and the low-resolution image obtained through bicubic interpolation. The output is the predicted super-resolution reconstructed image. The prediction model uses the UNet architecture, whose encoding, intermediate, and decoding layers are all built based on the Noise Residual Attention (NRA) module. The NRA module employs a structure that combines multi-branch, multi-scale feature extraction with self-attention and convolutional block attention mechanisms to enhance the capture of key texture features and improve the fidelity of reconstruction details.
[0059] (a) NRA module
[0060] The structure of the NRA module is as follows: Figure 2 As shown, the processing procedure is as follows.
[0061] For input features Perform twice The convolution and ReLU activation functions are used to perform preliminary feature extraction and nonlinear transformation, resulting in the output. :
[0062]
[0063] In the above formula, represent Convolution operation, This represents the ReLU activation function.
[0064] Set the current time step Through two parallel multilayer perceptrons and After processing, with features The components are fused together to obtain the output. :
[0065]
[0066] In the above formula, and These represent the processing procedures of two multilayer perceptrons, respectively. This indicates element-wise multiplication; This indicates element-wise addition.
[0067] Subsequently, three branches are constructed to support the fused features. Perform parallel processing.
[0068] Branch 1: Pass through first Convolution adjusts the number of channels to obtain :
[0069]
[0070] In the above formula, represent Convolution operation.
[0071] After that, Divided into three multi-scale sub-branches, each employing... , , Convolution extracts multi-scale features, resulting in , and :
[0072]
[0073]
[0074]
[0075] In the above formula, and Represent Convolution and Convolution operation.
[0076] The output of each sub-branch is processed by the self-attention module and then multiplied element-wise with the sub-branch features to obtain... , and :
[0077]
[0078]
[0079]
[0080] In the above formula, This represents the self-attention module.
[0081] Finally, the results of the three sub-branches are concatenated along the channel dimension, and then... After convolution adjustment, the final output of branch 1 is obtained. :
[0082]
[0083] In the above formula, This indicates a splicing operation along the channel dimension.
[0084] Branch 2: Input features conduct Convolution processing yields :
[0085]
[0086] Branch 3: Enhance key feature extraction through convolutional block attention modules to obtain... :
[0087]
[0088] In the above formula, This represents the attention module of the convolutional block.
[0089] The outputs of the above three branches , and splicing along the channel dimension, via After convolution adjustment, we get :
[0090]
[0091] Then, by applying the ReLU activation function and introducing residual connections, the final output of the NRA module is obtained. :
[0092]
[0093] (II) UNet Architecture
[0094] Predictive Model The overall architecture is as follows Figure 1As shown, UNet is used as the backbone network. The encoding layer of the prediction model consists of an NRA module and a downsampling module. The intermediate layers of the prediction model also consist of NRA modules. The decoding layer of the prediction model consists of transposed convolutions and NRA modules, and the output of the decoding layer is the predicted super-resolution reconstructed image. Based on the conventional UNet architecture, the prediction model maintains the skip connection method between the encoding and decoding layers, using a combination of NRA modules and downsampling modules as the unit of the encoding layer, and a combination of transposed convolutions and NRA modules as the unit of the decoding layer. In this embodiment, the number of channels of the prediction model is configured as follows: .
[0095] III. Training Process
[0096] Constructing including high-resolution images and low-resolution images training samples The degradation image is obtained by random sampling time steps in the forward process. Then, the degradation image and the low-resolution image of bicubic interpolation are used to train the prediction model. A composite loss function is used to optimize the network weights, which takes into account both pixel-level accuracy and visual perception quality.
[0097] Specifically, a random diffusion step training strategy is adopted. For each training sample pair... From the beginning to the end of the total time step Uniformly randomized sampling at one time step Based on the current time step Calculate residual scheduling parameters and noise scheduling parameters And sample random noise Substituting into the forward process formula:
[0098]
[0099] Obtain the degradation image at the current time step. Then, the degraded image Low-resolution images with bicubic interpolation After being stitched along the channel dimension, the image is input into the prediction model, and the output predicted super-resolution reconstructed image is denoted as... Calculate the total loss function based on the prediction results. Update the prediction model through gradient backpropagation Weight parameters Repeat the above steps for each batch until the set number of iterations is reached.
[0100] During training, the prediction model Internal weight parameters Updates are performed using gradient descent and backpropagation to minimize the total loss function. The diffusion model framework itself—including the mathematical expressions for the forward and backward processes, the calculation methods for residual scheduling and noise scheduling—and its related hyperparameters—including the total number of diffusion steps—are crucial. residual rate of change parameter Noise switching node Noise intensity parameters —These parameters are pre-defined before training and remain fixed throughout the training process, not participating in gradient updates. The structural hyperparameters of the prediction model, including the number of channels in each layer, are also determined before training.
[0101] Based on the super-resolution results obtained from model training under different parameter settings, the optimal parameters were determined through experiments to be: total number of diffusion steps. residual rate of change parameter Noise switching node Noise intensity parameters The optimal number of channels for the prediction model is configured as follows: .
[0102] IV. Composite Loss Function
[0103] When training the prediction model, a composite loss function based on L1 loss and VGG perceptual loss is used. The L1 loss is used to constrain the pixel-level deviation between the reconstructed image and the real high-resolution image, while the perceptual loss is used to constrain the semantic consistency between the two in the high-level feature space of the pre-trained VGG network. The two are jointly optimized to balance pixel accuracy and texture fidelity, effectively suppressing over-smoothing and artifact generation.
[0104] Composite loss function The expression is:
[0105]
[0106] In the above formula, For L1 loss; To perceive loss; As a weighting factor, it is set to [value] in this embodiment. .
[0107] The formula for calculating perceived loss is:
[0108]
[0109] In the above formula, This represents a high-level feature map extracted by a pre-trained VGG16 convolutional neural network. represent The square of the norm; This represents the number of training samples in the current batch. For the first High-resolution images of each training sample pair; For the first The training samples are used to obtain super-resolution reconstructed images through a forward process and a prediction model.
[0110] V. Reasoning Structure
[0111] Based on the defined inverse process and the trained prediction model, a low-sampling-step super-resolution inference architecture is constructed. The low-resolution image to be super-resolution, which is obtained through bicubic interpolation, is input into this architecture. Starting from the maximum time step, the inverse iteration is performed sequentially, and finally a high-fidelity high-resolution reconstructed image is obtained, achieving high-quality super-resolution with only a few sampling steps.
[0112] The reasoning phase only requires the reverse process to be executed, and the forward process is not required. The specific process is as follows.
[0113] Given a low-resolution image to be super-resolution bicubic interpolation Let the initial state , .from arrive Perform reverse iterations sequentially. In each iteration, first view the current intermediate state image. and After concatenating along the channel dimension, the image is input into the trained prediction model to obtain the super-resolution reconstructed image predicted in this round. Then and Substitute into the reverse process iterative formula:
[0114]
[0115] get Then let Start the next iteration until you get... This image is then used as the final super-resolution reconstructed image.
[0116] VI. Experimental Verification
[0117] To verify the effectiveness of this method, a 4x super-resolution experiment was conducted on the NWPU-RESISC45 dataset, and Depth Image-Texture Similarity (DISTS) was used as the texture fidelity evaluation metric. The lower the DISTS value, the higher the consistency of the reconstructed image with the real high-resolution image in terms of texture structure.
[0118] Table 1: Comparison of our method with several existing super-resolution methods on the DISTS index.
[0119]
[0120] As shown in Table 1, our method achieved the best result (0.1358) on the DISTS index, which is significantly better than the traditional diffusion model method DDPM (0.2592) and the existing fast diffusion super-resolution method FastDiffSR (0.1547), demonstrating the superiority of our method in terms of texture fidelity. Figure 3 The results show a visual comparison of the 4x super-resolution of a single image on the NWPU-RESISC45 dataset between our method and some of the comparative methods.
[0121] Table 2: Different diffusion steps The results of the DISTS index of this method under the given conditions.
[0122]
[0123] As shown in Table 2, when The DISTS value was lowest at that time, and it increased with... As the value increases, the DISTS value shows an upward trend, indicating that only 5 sampling steps are needed to achieve optimal texture fidelity.
[0124] Table 3: Parameters for different rates of change of residuals and noise switching node Results of DISTS index under combined conditions.
[0125]
[0126] As shown in Table 3, when , The DISTS value is lowest at this time (0.1358), and this combination is the optimal parameter configuration. Figure 4 Showing different and Visual comparison of super-resolution results under different parameter combinations.
[0127] It should be noted that, as will be apparent to those skilled in the art, the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics thereof. The scope of the present invention is defined by the claims rather than the foregoing description.
Claims
1. A low-sampling-step high-texture-fidelity remote sensing image super-resolution method based on a diffusion model, comprising: Define a forward process and a corresponding reverse process. The forward process is used to progressively transform a high-resolution image into a low-resolution image, and the reverse process is used to reconstruct a high-resolution image from a low-resolution image through iterative reconstruction. A prediction model is used to predict the super-resolution reconstructed image based on the current state. Its features are, In the forward process, a residual scheduling parameter and a noise scheduling parameter are introduced. The residual scheduling parameter is used to control the accumulation rate of the residual between the high-resolution image and the low-resolution image. The noise scheduling parameter is derived from the residual scheduling parameter, and the noise intensity is first enhanced and then attenuated in the forward process. The reverse process takes a low-resolution interpolated image as the initial input and performs iterative reconstruction based on the residual scheduling parameters, the noise scheduling parameters, and the super-resolution reconstructed image output by the prediction model. The prediction model uses the UNet architecture, in which the encoding layer, intermediate layer and decoding layer are all built on the noise residual attention module. The noise residual attention module performs preliminary feature extraction and current time step fusion on the input features, and then performs parallel processing through the first branch, the second branch and the third branch respectively. The first branch adopts multi-scale feature extraction and fusion self-attention mechanism, the second branch performs convolution processing, and the third branch enhances key feature extraction through the convolution block attention module. The outputs of the three branches are fused and then connected by residual to obtain the output features. The low-resolution interpolated image to be super-resolution is used as the initial input of the reverse process. Starting from the maximum time step, the reverse iteration is performed sequentially to obtain the high-resolution reconstructed image.
2. The method for high-texture fidelity remote sensing image super-resolution based on a diffusion model with few sampling steps according to claim 1, characterized in that, The residual scheduling parameters The calculation formula is ,in For the current time step, The total number of time steps in the forward process. The parameter used to control the rate of change of the residual.
3. The method for high-fidelity remote sensing image super-resolution based on a diffusion model with few sampling steps according to claim 1, characterized in that, The noise scheduling parameters The expression is: when hour ,when hour ,in For the current time step, The residual scheduling parameters are... This is a noise switching node.
4. The method for high-texture fidelity remote sensing image super-resolution based on a diffusion model with few sampling steps according to claim 1, characterized in that, The expression for the forward process is: ,in For the forward process, the first Intermediate state images at each time step, For high-resolution images, This is a low-resolution interpolated image obtained by interpolating a low-resolution image. The residual scheduling parameters are... The noise scheduling parameters are... This is the noise amplitude coefficient. This refers to random noise that follows a standard normal distribution.
5. The method for high-texture fidelity remote sensing image super-resolution based on a diffusion model with few sampling steps according to claim 1, characterized in that, The iterative formula for the reverse process is: ,in For the first step in the reverse process Intermediate state images at each time step, For the first Intermediate state images at each time step, The prediction model is based on The predicted super-resolution reconstructed image, and The first and the The residual scheduling parameters for the time step, and The first and the The noise scheduling parameters for the time step, This is the noise amplitude coefficient. This refers to random noise that follows a standard normal distribution.
6. The method for high-texture fidelity remote sensing image super-resolution based on a diffusion model with few sampling steps according to claim 1, characterized in that, The processing flow of the noise residual attention module includes: For input features Perform twice in sequence Features are obtained through convolution and ReLU activation function processing. ; Set the current time step Through two parallel multilayer perceptrons and After processing and Fusion, resulting in features ,in This indicates element-wise multiplication. This indicates element-wise addition; The first branch will through After adjusting the number of channels in the convolution, respectively through , and The convolution yields three sub-branch features. Each sub-branch feature is processed by the self-attention module and then multiplied element-wise with the feature of its sub-branch. The results of the three sub-branch features are then concatenated. Convolution yields the first branch output; The second branch pair conduct Convolution yields the second branch output; The third branch is processed by the convolutional block attention module. The third branch output is obtained; The first branch output, the second branch output, and the third branch output are concatenated and then... Convolution adjustment, then combined with input features conduct The results of convolution and ReLU activation are added element by element, and after ReLU activation, the final output of the noise residual attention module is obtained.
7. The method for high-texture fidelity remote sensing image super-resolution based on a diffusion model with few sampling steps according to claim 1, characterized in that, The encoding layer of the prediction model is composed of a combination of the noise residual attention module and the downsampling module, the intermediate layer is composed of the noise residual attention module, and the decoding layer is composed of a combination of transposed convolution and the noise residual attention module.
8. The method for high-fidelity remote sensing image super-resolution based on a diffusion model with few sampling steps according to claim 1, characterized in that, The training process of the prediction model includes: constructing training sample pairs comprising high-resolution and low-resolution images; for each training sample pair, from the total time step... A uniformly random sample of the current time step ,according to The residual scheduling parameters and noise scheduling parameters are calculated and substituted into the forward process to obtain the degraded image. The degraded image and the low-resolution interpolated image are concatenated along the channel dimension and then input into the prediction model to obtain the super-resolution reconstructed image. The weight parameters of the prediction model are updated based on the loss function through gradient backpropagation.
9. The method for high-texture fidelity remote sensing image super-resolution based on a diffusion model with few sampling steps according to claim 8, characterized in that, The loss function is a composite loss function of L1 loss and perception loss. ,in For L1 loss, In order to perceive loss, The weighting factor is used; the L1 loss is used to constrain the pixel-level deviation between the reconstructed image and the real high-resolution image, and the perceptual loss is used to constrain the semantic consistency between the two in the high-level feature space of the pre-trained convolutional neural network.
10. The method for high-texture fidelity remote sensing image super-resolution based on a diffusion model with few sampling steps according to claim 9, characterized in that, The formula for calculating the perceived loss is as follows: ,in This represents a high-level feature map extracted by a pre-trained VGG16 convolutional neural network. The number of training samples, For the first High-resolution images of training sample pairs, For the first A training sample pair of super-resolution reconstructed images obtained through the forward process and the prediction model.