Diffusion model image super-resolution method based on degradation prior and dynamic feature adjustment
By using a diffusion model based on degradation priors and dynamic feature adjustment, the problems of feature alignment and network structure optimization in image super-resolution processing of diffusion models are solved, achieving higher quality image reconstruction results and adapting to complex degradation scenarios.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-24
- Publication Date
- 2026-03-13
AI Technical Summary
Existing diffusion models suffer from problems such as feature representation alignment, insufficient network structure optimization, and inadequate adaptability to complex degradation scenarios in image super-resolution processing, resulting in inaccurate and unstable reconstruction results.
A multimodal large language model is used to extract degenerate prior information. Through feature alignment module and dynamic U-Net network optimization, combined with an end-to-end training framework, the coordinated optimization of feature alignment and denoising process is achieved.
It improves the accuracy and robustness of image super-resolution reconstruction, better adapts to complex degradation scenes, suppresses artifacts and structural distortion, and enhances image visual quality.
Smart Images
Figure CN121660891A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of image super-resolution technology, specifically to an image super-resolution method based on a diffusion model with degradation priors and dynamic feature adjustment. Background Technology
[0002] Diffusion models have demonstrated strong potential in image generation, achieving the transformation from simple to complex data distributions through forward noise addition and inverse denoising. Their progressive generation method effectively maintains the coherence of image content while efficiently recovering detail information. In image super-resolution research, diffusion model-based methods have formed diverse technical approaches. Supervised learning methods, through end-to-end training, utilize large amounts of pairwise data to learn the mapping relationship from low to high resolution. These methods perform well on certain specific datasets, generating visually appealing high-resolution images. However, they require high-quality and abundant training data and perform poorly when dealing with degradation types not covered by the training set. Another type of zero-shot learning method attempts to leverage the intrinsic learning priors of pre-trained diffusion models to achieve super-resolution reconstruction without additional training. While these methods offer greater flexibility, they often suffer from insufficient stability when handling complex real-world scenes, especially when the input image contains severe degradation, easily leading to structural distortion or detail errors.
[0003] Existing technical solutions exhibit their own characteristics in specific implementation. Some methods employ a two-stage processing strategy, first performing degradation removal and then using a diffusion model to enhance details. These methods improve reconstruction results to some extent, but still have limitations in feature alignment. Other methods attempt to improve performance by modifying the network architecture, such as introducing new connection mechanisms or attention modules into the U-Net structure. While these improvements bring some performance gains, they are still insufficient when handling complex degradation scenarios. In addition, some studies focus on improving the inference efficiency of diffusion models by reducing the number of sampling steps or simplifying the network structure to accelerate the processing, but these optimizations often come at the cost of image quality.
[0004] Although existing methods have made some progress, they still have some shortcomings in practical applications: (1) Feature Representation Alignment Problem: In the super-resolution processing of the diffusion model, there is a significant difference between the feature representation of the low-resolution input and the feature space inside the diffusion model. This feature mismatch makes it difficult for the model to fully utilize the information of the input image, affecting the accuracy and reliability of the reconstruction results. This problem is particularly prominent when processing real-world images with complex degradation patterns.
[0005] (2) Insufficient network structure optimization: Existing methods do not fully utilize the core components of the diffusion model. In particular, when processing features in different regions, there is a lack of effective adaptive mechanisms to balance the relationship between detail recovery and noise suppression. This leads to the easy loss of details in regions with complex textures, while artifacts may be generated in flat regions.
[0006] (3) Adaptability to complex degradation scenes: Real-world images often contain the combined effects of multiple degradation factors, such as noise, blur, and compression artifacts. Existing methods perform poorly when dealing with such complex degradation scenes, making it difficult to accurately identify and adapt to different degradation features, resulting in unstable reconstruction results. Summary of the Invention
[0007] The purpose of this invention is to address the shortcomings of existing technologies by providing a diffusion model-based image super-resolution method based on degradation priors and dynamic feature adjustment.
[0008] The technical solution of this invention to solve the above problems is: an image super-resolution method based on a diffusion model with degradation prior and dynamic feature adjustment, comprising the following steps: Step 1: Low-resolution image input and degradation prior extraction; The input low-resolution image is acquired, and a low-resolution encoder with a pre-trained variational autoencoder is used to extract the low-resolution features of the low-resolution image. At the same time, a multimodal large language model is used to analyze the content of the low-resolution image, generate natural language text describing the degradation type, and convert the text into semantic vectors through CLIP text encoder as degradation priors for the diffusion process. Step 2: Alignment processing of degraded sensing features; The extracted low-resolution features and the noise latent variable features of the diffusion model are input into the feature alignment module. The feature alignment module first processes the two types of features separately through convolutional layers and concatenates the processing results. Then, the concatenated results are processed sequentially through two Transform blocks to enhance the interaction between features. The key information in the noise latent variable features is preserved through residual connections. Finally, the features are fused through fully connected layers, and the feature map dimension is aligned with the input of the U-Net network. Step 3: Dynamic U-Net network optimization processing; The network structure of the diffusion model is improved by introducing a dynamic feature adjustment strategy in the decoder part; a scaling factor based on channel mean is applied to the backbone features to enhance the denoising capability; frequency domain transformation is performed on the skip connection features to suppress low-frequency components and preserve high-frequency details; the scaling factor is dynamically generated according to the image content to achieve a balance between denoising capability and detail preservation. Step 4, Joint training and image reconstruction; The low-resolution encoder, feature alignment module, and diffusion model are trained end-to-end in a unified framework. After training, the optimized diffusion model is used for iterative denoising, and finally the high-resolution image is reconstructed through the pre-trained decoder.
[0009] Furthermore, in step 1, the specific implementation of degradation prior extraction is as follows: a multimodal large language model is used to analyze the low-resolution image and generate a text description containing degradation attributes; then, the text description is mapped into a semantic vector through the CLIP text encoder; the semantic vector is injected into the denoising process of the diffusion model through a cross-attention mechanism to achieve adaptive guidance for specific degradation types.
[0010] Furthermore, the feature alignment module in step 2 adopts a Transformer architecture with residual connections, and its specific calculation process is as follows: First, the low-resolution image features extracted in step 1 and the noise latent variables of the diffusion model are convolved and then concatenated to obtain the concatenated features. :
[0011] in, The low-resolution image features extracted in step 1 In order to achieve the first step in the diffusion process The noise latent variable corresponding to each time step express Convolution operation, Represents the concatenation operation of features; Then the spliced features The processed features are obtained by sequentially processing the data through two Transformer blocks. :
[0012] in , These represent two separate Transformer block processes; Finally, alignment features are obtained through residual connections and fully connected layers. :
[0013] in, Indicates a fully connected layer. In the diffusion process The noise latent variable corresponding to each time step has passed through Features extracted by convolution operations.
[0014] Furthermore, the dynamic feature adjustment strategy in step 3 specifically includes two parts: adding a trunk scaling factor and adding a jump scaling factor. The specific process is as follows: First, a backbone scaling factor is applied to each layer of the U-Net network decoder. This backbone scaling factor enhances the denoising capability of the backbone network by amplifying the backbone features. The specific application process of the backbone scaling factor for each layer is as follows: The first part of the calculation of the U-Net network decoder Channel average value of backbone network features in the layer :
[0015] in, For the U-Net network decoder section, the first Layer 1 backbone network Characteristics of each channel For the U-Net network decoder section, the first The number of channels characteristic of the layered backbone network; The U-Net network decoder part Channel average value of backbone network features in the layer Generate the backbone scaling factor corresponding to this layer. :
[0016] in, An adjustable scaling factor is used to control the overall scaling intensity. Indicates the U-Net network decoder section. Channel average value of backbone network features in the layer The minimum value of all elements in the set. Indicates the U-Net network decoder section. Channel average value of backbone network features in the layer The maximum value of all elements in the set; Subsequently, a backbone scaling factor is applied to the first half of the backbone network features, while the second half remains unchanged, to balance denoising and detail preservation.
[0017] in, For the U-Net network decoder section, the first Layer 1 backbone network The features of each channel are scaled by the backbone scaling factor. The scaling factor is the main scaling factor. For the U-Net network decoder section, the first Layer 1 backbone network Characteristics of each channel This represents element-wise multiplication. For the U-Net network decoder section, the first The number of channels characteristic of the layered backbone network; Secondly, a skip scaling factor is applied to the features transmitted through the skip connections of the U-Net network in each backbone layer. This skip scaling factor suppresses low-frequency components in the features, thus mitigating the loss of high-frequency details caused by denoising enhancement. The specific application process of the skip scaling factor in each layer is as follows: For the jump connection at the 1st The features transmitted by the layer backbone network are subjected to Fast Fourier Transform to obtain the skip connections in the 1st layer. Frequency domain characteristics of features transmitted by the layered backbone network :
[0018] in, For the jump connection in the 1st Characteristics of layered backbone network transmission This indicates the Fast Fourier Transform operation; Construct a frequency mask as a jump scaling factor This allows it to be applied only to the low-frequency components of the feature, while keeping the high-frequency components of the feature unchanged:
[0019] in, For the jump connection in the 1st The frequency radius corresponding to the frequency domain features of the features transmitted by the layer backbone network. It is a hyperparameter representing the low-frequency threshold. An adjustable scaling factor to control the scaling intensity; Subsequently, the skip scaling factor is applied to the frequency domain features of the skip connection features, and then the features are restored to the spatial domain features using an inverse Fourier transform. :
[0020] in, This indicates the inverse Fourier transform operation. For the jump connection in the 1st Frequency domain characteristics of features transmitted in a layered backbone network. This is the jump scaling factor. This indicates element-wise multiplication.
[0021] Furthermore, the joint training in step S4 employs a multi-objective loss function:
[0022] in, Standard Gaussian noise, This indicates that the diffusion model predicts the noise process. For noise scheduling parameters, Image features extracted from high-resolution images, The alignment feature obtained in step S2, The semantic vector is obtained by the CLIP text encoder from the text description mapping. This is the time step for the diffusion model.
[0023] The present invention has the following beneficial effects: This invention provides an image super-resolution method based on degradation priors and dynamic feature adjustment using a diffusion model. It fully considers the correspondence between degradation characteristics in low-resolution images and the internal feature space of the diffusion model, achieving higher-quality image super-resolution reconstruction. First, a multimodal large language model is used to extract degradation prior information from low-resolution images, transforming it into semantic vectors as guiding conditions for the diffusion process, enabling the model to adapt to different degradation types (such as motion blur, noise interference, etc.). Second, a feature alignment module is designed, effectively fusing low-resolution features and noise latent variables through a Transformer architecture and residual connection mechanism, solving the problem of insufficient information utilization caused by feature space mismatch. Third, a dynamic U-Net optimization strategy is introduced, using backbone feature scaling and skip connection frequency domain filtering to enhance denoising capabilities while preserving high-frequency details, achieving a balance between denoising and detail recovery. Finally, an end-to-end joint training framework is used to collaboratively optimize degradation prior extraction, feature alignment, and diffusion denoising processes, significantly improving the accuracy and robustness of super-resolution reconstruction. This approach enables the diffusion model to better adapt to complex degradation scenarios, effectively suppressing artifacts and structural distortions while improving image visual quality, thus providing an efficient and reliable solution for image super-resolution tasks. Attached Figure Description
[0024] Figure 1 This is a flowchart of the diffusion model image super-resolution method based on degradation prior and dynamic feature adjustment described in this invention. Detailed Implementation
[0025] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.
[0026] like Figure 1 As shown, a diffusion model-based image super-resolution method based on degradation priors and dynamic feature adjustment includes the following steps: Step 1: Low-resolution image input and degradation prior extraction. A low-resolution input image is acquired, and a pre-trained variational autoencoder is used to extract the latent space features of the image. Simultaneously, a multimodal large language model is used to analyze the image content, generating natural language text describing the degradation type. This text is then converted into semantic vectors using the CLIP text encoder, serving as a degradation prior for the diffusion process.
[0027] When extracting latent space features from low-resolution images, since different layers of features from the VAE encoder contain information of different granularities, the last layer of features, although highly compressible, may lose details. In this embodiment, the penultimate layer of features from the VAE encoder is preferred. This layer contains 512 channels and can retain more degradation information and structural details.
[0028] The text description generation process guides the multimodal large language model to focus on image degradation features through prompts. For example, the prompt "As an image quality analysis expert, please carefully examine the following images for the following low-level visual feature problems: Gaussian noise, motion blur, JPEG compression block effect, scaling. Please determine and describe their presence and degree one by one" is input into the multimodal large language model to generate natural language text describing the image degradation type. Subsequently, the CLIP text encoder is used to map the text into a semantic vector, and this semantic vector is injected into the denoising process of the diffusion model through a cross-attention mechanism to achieve adaptive guidance for specific degradation types.
[0029] Step 2: Degradation-Aware Feature Alignment Processing. The extracted low-resolution image features and the noise latent variables from the diffusion model are input into the feature alignment module. This module first processes the two types of features separately through 3×3 convolutional layers and concatenates the results. Then, the concatenated result is processed sequentially through two Transformer blocks to enhance the interaction between features. Key information in the noise latent variables is preserved through residual connections. Finally, the features are fused through fully connected layers, and the feature map dimensions are aligned with the input of the U-Net network.
[0030] In the feature alignment module of this embodiment, a Transformer architecture with residual connections is adopted. Although the self-attention mechanism of Transformer is good at capturing long-range dependencies, it suffers from feature space mismatch when aligning low-resolution features with noisy latent variables. Low-resolution features contain rich degradation information, while noisy latent variables gradually become clear during diffusion. Direct fusion will cause degradation information to interfere with the generation process. In order to effectively align the two types of features, this embodiment implements progressive fusion through a feature alignment module. The specific calculation process of the alignment module includes the following sub-steps: First, 3×3 convolution operations are performed on the extracted low-resolution image features and noise latent variables respectively to extract local features. The convolution results are then concatenated along the channel dimension to obtain the concatenated features. :
[0031] in, For the extracted low-resolution image features, In order to achieve the first step in the diffusion process The noise latent variable corresponding to each time step express Convolution operation, This indicates a splicing operation along the channel dimension.
[0032] Next, the concatenated features are processed through two Transformer blocks to enhance the interaction between features. Each Transformer block contains a multi-head self-attention mechanism and a feedforward network, ultimately yielding the processed features. :
[0033] in, , For the same Transformer block, each Transformer block contains a multi-head self-attention mechanism (with 8 heads) and a feedforward network (with 512 hidden layers) to enhance the interaction between features.
[0034] Finally, key information in the noisy latent variables is preserved through residual connections, and the features are fused through fully connected layers to obtain aligned features. :
[0035] in, This indicates a fully connected layer, with the output dimension aligned to the U-Net input. In the diffusion process The noise latent variable corresponding to each time step has passed through Features extracted by convolution operations.
[0036] Step 3: Dynamic U-Net Network Optimization. The U-Net network structure of the diffusion model is improved by introducing a dynamic feature adjustment strategy in the decoder. A scaling factor based on channel mean is applied to the backbone features to enhance denoising capability; frequency domain transformation is performed on skip connection features to suppress low-frequency components while preserving high-frequency details. The scaling factor is dynamically generated based on the image content, achieving a balance between denoising capability and detail preservation.
[0037] The dynamic feature adjustment strategy aims to balance denoising capability and detail preservation. Since the U-Net backbone primarily handles denoising, while skip connections inject high-frequency features, these skip connections may weaken the backbone's denoising ability, leading to abnormal generated details. This embodiment adaptively adjusts feature contributions by applying dynamic scaling factors. These dynamic scaling factors include a backbone scaling factor based on channel mean applied to the backbone features and a skip scaling factor that performs frequency domain transformation on the skip connection features. By applying the backbone scaling factor to each layer of the U-Net decoder, the denoising capability of the U-Net network is enhanced. By applying the skip scaling factor to the features transmitted through the skip connections in each layer of the backbone network, low-frequency components are suppressed, thereby preserving high-frequency details of the image.
[0038] In each layer of the U-Net network decoder, the application of the backbone scaling factor is as follows: First, calculate the channel average of the backbone features of the nth layer of the U-Net decoder. To capture overall structural information:
[0039] in, For the U-Net network decoder section, the first Layer 1 backbone network Characteristics of each channel For the U-Net network decoder section, the first The number of channels characteristic of the layer backbone network.
[0040] Then, the backbone scaling factor is dynamically generated based on the calculated channel average. This adapts it to the dynamic range of the input features.
[0041] in, To control the overall scaling intensity, the adjustable scaling factor is initially set to 1.2. Indicates the U-Net network decoder section. Channel average value of backbone network features in the layer The minimum value of all elements in the set. Indicates the U-Net network decoder section. Channel average value of backbone network features in the layer The maximum value of all elements in the set.
[0042] Finally, the backbone scaling factor is applied to the first half of the backbone features, while the second half remains unchanged, to balance denoising and detail preservation.
[0043] in, For the U-Net network decoder section, the first Layer 1 backbone network The features of each channel are scaled by the backbone scaling factor. The scaling factor is the main scaling factor. For the U-Net network decoder section, the first Layer 1 backbone network Characteristics of each channel This represents element-wise multiplication. For the U-Net network decoder section, the first The number of channels characteristic of the layer backbone network.
[0044] The application of the skip scaling factor in the features transmitted through skip connections in each layer of the U-Net backbone network is as follows: First, a Fast Fourier Transform is performed to transform the features to the frequency domain, obtaining the frequency domain features. :
[0045] in, This indicates the inverse Fourier transform operation. For the jump connection in the 1st Characteristics of layered backbone network transmission This is the jump scaling factor. This indicates element-wise multiplication.
[0046] Then, a frequency mask is constructed as a scaling factor. Used to suppress low-frequency components:
[0047] in, For the jump connection in the 1st The frequency radius corresponding to the frequency domain features of the features transmitted by the layer backbone network. This is a hyperparameter representing the low-frequency threshold, initially set to 30% of the frequency range. The adjustable scaling factor, which controls the scaling intensity, is initially set to 0.8.
[0048] Finally, the jump scaling factor is applied to the frequency domain features, and the adjusted frequency domain features are restored to spatial domain features through inverse Fourier transform. :
[0049] in, This indicates the inverse Fourier transform operation. For the jump connection in the 1st Frequency domain characteristics of features transmitted in a layered backbone network. This is the jump scaling factor. This indicates element-wise multiplication.
[0050] Step 4: Joint Training and Image Reconstruction. The low-resolution encoder, feature alignment module, and diffusion model are trained end-to-end within a unified framework. After training, the optimized diffusion model is used for iterative denoising, and finally, a high-resolution image is reconstructed using a pre-trained decoder.
[0051] The specific implementation process of joint training is as follows: The training dataset uses publicly available super-resolution datasets, including DIV2K, LSDIR, Flickr2K, DIV8K, and FFHQ, where the high-resolution image size is 1024×1024, and the low-resolution images are generated through a degradation model. The degradation model includes blurring, scaling, noise addition, and JPEG compression operations: the blurring operation uses a Gaussian blur kernel, with the kernel width sampled from a uniform distribution (range 3-21) and the standard deviation sampled from a uniform distribution (range 0.2-3.0); the scaling operation is performed twice, with the first scaling factor sampled from a uniform distribution (range 1-4), and the second scaling factor determined by the target size; the noise addition operation randomly selects from Gaussian or Poisson noise, with the noise intensity sampled from a uniform distribution (range 0-50); the JPEG compression quality factor is sampled from a uniform distribution (range 30-95). Training uses the AdamW optimizer with a weight decay of 0.01 and a batch size of 256. The learning rate uses a cosine annealing strategy, with an initial value of 5×10⁻. 5 After 5000 warm-up steps, the temperature decays to 1×10⁻ 5 The total number of training iterations was 40,000. The loss function adopted a multi-objective form, combining L1 loss and LPIPS loss:
[0052] in, Standard Gaussian noise, This indicates that the diffusion model predicts the noise process. For noise scheduling parameters, Image features extracted from high-resolution images, For alignment features, The semantic vector is obtained by the CLIP text encoder from the text description mapping. This is the time step for the diffusion model.
[0053] The inference phase employs the Euler sampler with 20 sampling steps and a classifier-independent guidance ratio of 5. Finally, a high-resolution image is reconstructed using a pre-trained VAE decoder.
[0054] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments for application in other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.
Claims
1. A diffusion model-based image super-resolution method based on degradation priors and dynamic feature adjustment, characterized in that: Includes the following steps: Step 1: Low-resolution image input and degradation prior extraction; The input low-resolution image is acquired, and a low-resolution encoder with a pre-trained variational autoencoder is used to extract the low-resolution features of the low-resolution image. At the same time, a multimodal large language model is used to analyze the content of the low-resolution image, generate natural language text describing the degradation type, and convert the text into semantic vectors through CLIP text encoder as degradation priors for the diffusion process. Step 2: Alignment processing of degraded sensing features; The extracted low-resolution features and the noise latent variable features of the diffusion model are input into the feature alignment module. The feature alignment module first processes the two types of features separately through convolutional layers and concatenates the processing results. Then, the concatenated results are processed sequentially through two Transform blocks to enhance the interaction between features. The key information in the noise latent variable features is preserved through residual connections. Finally, the features are fused through fully connected layers, and the feature map dimension is aligned with the input of the U-Net network. Step 3: Dynamic U-Net network optimization processing; The network structure of the diffusion model is improved by introducing a dynamic feature adjustment strategy in the decoder part; A scaling factor based on channel mean is applied to the backbone features to enhance denoising capabilities; frequency domain transformation is performed on the skip connection features to suppress low-frequency components and preserve high-frequency details. The scaling factor is dynamically generated based on the image content, achieving a balance between noise reduction capability and detail preservation; Step 4, Joint training and image reconstruction; The low-resolution encoder, feature alignment module, and diffusion model are placed in a unified framework for end-to-end joint training. After training, the optimized diffusion model is used for iterative denoising, and finally the high-resolution image is reconstructed through the pre-trained decoder.
2. The image super-resolution method based on degradation prior and dynamic feature adjustment according to claim 1, characterized in that: In step 1, the specific implementation of degradation prior extraction is as follows: a multimodal large language model is used to analyze the low-resolution image and generate a text description containing degradation attributes; then, the text description is mapped into a semantic vector through the CLIP text encoder; the semantic vector is injected into the denoising process of the diffusion model through a cross-attention mechanism to achieve adaptive guidance for specific degradation types.
3. The image super-resolution method based on degradation prior and dynamic feature adjustment based on a diffusion model as described in claim 1, characterized in that: The feature alignment module in step 2 adopts a Transformer architecture with residual connections, and its specific calculation process is as follows: First, the low-resolution image features extracted in step 1 and the noise latent variables of the diffusion model are convolved and then concatenated to obtain the concatenated features. : ； in, The low-resolution image features extracted in step 1 In order to achieve the first step in the diffusion process The noise latent variable corresponding to each time step express Convolution operation, Represents the concatenation operation of features; Then the spliced features The processed features are obtained by sequentially processing the data through two Transformer blocks. : ； in , These represent two separate Transformer block processes; Finally, alignment features are obtained through residual connections and fully connected layers. : ； in, Indicates a fully connected layer. In the diffusion process The noise latent variable corresponding to each time step has passed through Features extracted by convolution operations.
4. The image super-resolution method based on a diffusion model with degradation priors and dynamic feature adjustment as described in claim 1, characterized in that: The dynamic feature adjustment strategy in step 3 specifically includes two parts: adding a backbone scaling factor and adding a jump scaling factor. The specific process is as follows: First, a backbone scaling factor is applied to each layer of the U-Net network decoder. This backbone scaling factor enhances the denoising capability of the backbone network by amplifying the backbone features. The specific application process of the backbone scaling factor for each layer is as follows: The first part of the calculation of the U-Net network decoder Channel average value of backbone network features in the layer : ； in, For the U-Net network decoder section, the first Layer 1 backbone network Characteristics of each channel For the U-Net network decoder section, the first The number of channels characteristic of the layered backbone network; The U-Net network decoder part Channel average value of backbone network features in the layer Generate the backbone scaling factor corresponding to this layer. : ； in, An adjustable scaling factor is used to control the overall scaling intensity. Indicates the U-Net network decoder section. Channel average value of backbone network features in the layer The minimum value of all elements in the set. Indicates the U-Net network decoder section. Channel average value of backbone network features in the layer The maximum value of all elements in the set; Subsequently, a backbone scaling factor is applied to the first half of the backbone network features, while the second half remains unchanged, to balance denoising and detail preservation. ； in, For the U-Net network decoder section, the first Layer 1 backbone network The features of each channel are scaled by the backbone scaling factor. The scaling factor is the main scaling factor. For the U-Net network decoder section, the first Layer 1 backbone network Characteristics of each channel This represents element-wise multiplication. For the U-Net network decoder section, the first The number of channels characteristic of the layered backbone network; Secondly, a skip scaling factor is applied to the features transmitted through the skip connections of the U-Net network in each backbone layer. This skip scaling factor suppresses low-frequency components in the features, thus mitigating the loss of high-frequency details caused by denoising enhancement. The specific application process of the skip scaling factor in each layer is as follows: For the jump connection at the 1st The features transmitted by the layer backbone network are subjected to Fast Fourier Transform to obtain the skip connections in the 1st layer. Frequency domain characteristics of features transmitted by the layered backbone network : ； in, For the jump connection in the 1st Characteristics of layered backbone network transmission This indicates the Fast Fourier Transform operation; Construct a frequency mask as a jump scaling factor This allows it to be applied only to the low-frequency components of the feature, while keeping the high-frequency components of the feature unchanged: ； in, For the jump connection in the 1st The frequency radius corresponding to the frequency domain features of the features transmitted by the layer backbone network. It is a hyperparameter representing the low-frequency threshold. An adjustable scaling factor to control the scaling intensity; Subsequently, the skip scaling factor is applied to the frequency domain features of the skip connection features, and then the features are restored to the spatial domain features using an inverse Fourier transform. : ； in, This indicates the inverse Fourier transform operation. For the jump connection in the 1st Frequency domain characteristics of features transmitted in a layered backbone network. This is the jump scaling factor. This indicates element-wise multiplication.
5. The image super-resolution method based on degradation prior and dynamic feature adjustment according to claim 1, characterized in that: The joint training in step S4 employs a multi-objective loss function: ； in, Standard Gaussian noise, This indicates that the diffusion model predicts the noise process. For noise scheduling parameters, Image features extracted from high-resolution images, The alignment feature obtained in step S2, The semantic vector is obtained by the CLIP text encoder from the text description mapping. This is the time step for the diffusion model.