An image restoration method, device, system and storage medium

By constructing a hyperbolic degradation representation and a cross-scale progressive restoration framework, the problem of image restoration under complex and unknown degradation conditions is solved. This enables hierarchical representation of degradation features and explicit separation of information, thereby improving the accuracy and stability of image restoration.

CN122841185BActive Publication Date: 2026-11-10SUN YAT SEN UNIV
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202611327548.6
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2026-08-31
Publication Date
2026-11-10
Estimated Expiration
2046-08-31

AI Technical Summary

Technical Problem

Existing image restoration methods are difficult to adapt effectively to complex and unknown degradation conditions, and suffer from problems such as limited ability to express degradation features, coupling of degradation information and image content information, insufficient transmission of multi-scale restoration information, and low efficiency in utilizing restoration knowledge.

Method used

A novel restoration framework combining hyperbolic degradation representation, structured key-value memory, and cross-scale state transfer is constructed. Through multi-scale decomposition, deep feature extraction, hyperbolic degradation representation, degradation-content orthogonal decoupling, structured memory, and cross-scale progressive restoration, adaptive restoration of complex and unknown degradation images is achieved.

Benefits of technology

It improves the accuracy and stability of image restoration, enables high-quality image restoration in complex environments, and enhances the adaptability and restoration efficiency to unknown degradation conditions.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122841185B_ABST
    Figure CN122841185B_ABST
Patent Text Reader

Abstract

The present application relates to the technical field of computer vision and image processing, in particular to an image restoration method, device, system and storage medium, the method comprising: obtaining a to-be-restored degraded image and performing multi-scale decomposition to generate a multi-scale degraded image sequence; performing deep feature extraction on each scale image to generate encoded features and dividing the encoded features into degraded features and content features; generating hyperbolic degradation representation for the degraded features and calculating hyperbolic distances between the degraded features and each degradation key; calculating matching weights according to the hyperbolic distances and reading restoration content values to generate restoration prior features; performing gated fusion and feature reconstruction on the restoration prior features and the content features to generate a restored image; and transferring the restoration state information of the current scale to the next scale for progressive restoration and outputting a final restored image. The present application realizes adaptive restoration of complex unknown degraded images and significantly improves the accuracy, stability and generalization ability of image restoration.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of computer vision and image processing technology, and in particular to an image restoration method, apparatus, system and storage medium. Background Technology

[0002] With the widespread application of intelligent vision systems in fields such as autonomous driving, remote sensing monitoring, industrial inspection, medical imaging, and smart terminals, the images actually acquired are often affected by complex environmental factors, such as sensor noise, low light, rain and snow, atmospheric scattering, motion-induced blurring, lens defocus, and compression distortion, resulting in varying degrees of information loss. Therefore, how to recover clear, high-quality images from complex degradation conditions has become an important research direction in the field of computer vision.

[0003] Traditional image restoration methods typically design dedicated models for a single degradation type. For example, denoising models are mainly for restoring random noise, deraining models are mainly for restoring raindrop occlusion, and dehazing models are mainly for restoring atmospheric scattering degradation. Although these methods can achieve good restoration results under known degradation types and fixed data distribution conditions, image degradation in real-world applications is often unknown, diverse, and combinatorial. For instance, the same image may simultaneously contain multiple degradation factors such as low illumination, noise, and blur, making it difficult for a single degradation model to effectively adapt to complex scenes. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention proposes an image restoration method, apparatus, system, and storage medium, aiming to achieve adaptive restoration of complex and unknown degraded images, effectively improving the accuracy, stability, and generalization ability of image restoration.

[0005] On one hand, embodiments of the present invention provide an image restoration method, the method comprising the following steps: Obtain the degraded image to be restored; The degraded image to be restored is decomposed into multiple scales according to a preset scale factor to generate a multi-scale degraded image sequence. The images at each scale in the multi-scale degraded image sequence are sequentially input into the image restoration encoder. The image restoration encoder performs deep feature extraction on each scale image to generate coded features, and the coded features are divided into degradation features and content features. The degradation features are mapped to hyperbolic space to generate a hyperbolic degradation representation, and the hyperbolic distance between the hyperbolic degradation representation and each degradation key in the degradation key memory module is calculated. The matching weight is calculated based on the hyperbolic distance, and the corresponding restored content value is read from the degenerate key-value memory module based on the matching weight to generate the restored prior features; The prior recovery features and the content features are gated and fused to generate fused recovery features; The fused restoration features are input into the image restoration decoder for feature reconstruction to generate a restored image at the corresponding scale; The restoration status information at the current scale is transferred to the next scale after being scaled by the cross-scale memory transfer module. The next scale then uses the restoration status information to perform progressive image restoration and outputs the final restored image.

[0006] Optionally, the step of performing deep feature extraction on images at various scales using the image restoration encoder to generate encoded features, and dividing the encoded features into degradation features and content features, includes: The image restoration encoder performs convolution operations, residual feature learning, and nonlinear feature enhancement on the images at each scale to generate deep features. The deep features are input into the degradation encoder, and the spatial dimension is compressed through feature aggregation operation to generate degradation features. The degradation features include noise intensity, blur degree, brightness change and degradation combination relationship information. The deep features are input into the content encoder, which preserves the spatial structure information of the image and generates content features. The content features include edge information, texture information, target structure information, and semantic content information.

[0007] Optionally, the step of mapping the degenerate features to a hyperbolic space to generate a hyperbolic degenerate representation, and calculating the hyperbolic distance between the hyperbolic degenerate representation and each degenerate key in the degenerate key memory module, includes: The degenerate features are mapped from Euclidean space to hyperbolic space, and the hierarchical expressive power of the degenerate features is enhanced by the negative curvature structure in hyperbolic space. In the hyperbolic space, calculate the hyperbolic distance between the hyperbolic degradation representation of the current input image and each degradation key in the degradation key memory module; The distribution position of the degradation features in the hyperbolic space is adjusted according to the hyperbolic distance, so that features with similar degradation properties are close to each other in the hyperbolic space, and features with different degradation modes are kept apart in the hyperbolic space.

[0008] Optionally, the degradation key-value memory module includes multiple structured memory units, each of which consists of a degradation key and a restored content value; the method further includes: The degradation key represents the feature prototype information corresponding to the degradation mode, and the feature prototype information includes degradation type, degradation intensity, and degradation combination relationship; The recovery experience corresponding to the degradation mode is stored through the recovery content value. The recovery experience includes structural compensation features, high-frequency texture recovery information, and recovery priors learned during the historical recovery process.

[0009] Optionally, the step of gating and fusing the prior recovery features with the content features to generate fused recovery features includes: The recovery prior features and the content features are received through the gating fusion module; Based on the importance of the recovered prior features and the content features, the fusion weights are adaptively calculated. The prior recovery features and the content features are weighted and fused according to the fusion weights to generate fused recovery features.

[0010] Optionally, the step of transferring the recovery state information at the current scale to the next scale after scale transformation via the cross-scale memory transfer module includes: After feature recovery is completed at the current scale, the low-frequency structural information extracted at the current scale, the recovered feature representation, the recovered prior features, and the recovered residual information are written into the cross-scale memory transfer module. The cross-scale memory transfer module performs spatial size adjustment, channel mapping, and feature normalization on the written information to generate scale-adaptive information. The scale adaptation information is passed to the next scale recovery stage, so that the next scale recovery stage can optimize texture, edges and local details based on the existing recovery state.

[0011] Optionally, the method further includes: During training, orthogonal decoupling constraints are applied to the degradation features and the content features. These orthogonal decoupling constraints maintain the relative independence of the degradation feature space and the content feature space by restricting the correlation between the feature vectors of the degradation features and the feature vectors of the content features. The degradation feature space focuses on noise intensity, blur level, brightness variation, and environmental interference factors. The content feature space focuses on image structure, edge information, texture information, and semantic information.

[0012] On the other hand, embodiments of the present invention provide an image restoration apparatus, comprising: The first module is used to acquire the degraded image to be restored; The second module is used to perform multi-scale decomposition on the degraded image to be restored according to a preset scale factor, and generate a multi-scale degraded image sequence. The third module is used to input the images at each scale in the multi-scale degraded image sequence into the image restoration encoder in sequence, and to extract deep features from the images at each scale through the image restoration encoder to generate coded features, and to divide the coded features into degraded features and content features. The fourth module is used to map the degradation features to hyperbolic space, generate a hyperbolic degradation representation, and calculate the hyperbolic distance between the hyperbolic degradation representation and each degradation key in the degradation key memory module. The fifth module is used to calculate the matching weight based on the hyperbolic distance, and read the corresponding restored content value from the degenerate key value memory module based on the matching weight to generate the restored prior features; The sixth module is used to perform gated fusion of the prior recovery features and the content features to generate fused recovery features; The seventh module is used to perform feature reconstruction on the fused recovery feature input image recovery decoder to generate a recovery image of the corresponding scale; The eighth module is used to transfer the restoration status information of the current scale to the next scale after scale conversion via the cross-scale memory transfer module, so that the next scale can use the restoration status information to perform progressive image restoration and output the final restored image.

[0013] On the other hand, embodiments of the present invention provide an image restoration system, comprising: At least one processor; At least one memory for storing at least one program; When the at least one program is executed by the at least one processor, the at least one processor performs the method described above.

[0014] On the other hand, embodiments of the present invention provide a computer-readable storage medium storing a processor-executable program, which, when executed by a processor, is used to perform the above-described method.

[0015] The embodiments of the present invention have the following beneficial effects: This invention acquires the degraded image to be restored and performs multi-scale decomposition, enabling the low-resolution scale to obtain the overall image structure and global degradation information, and the high-resolution scale to restore edge texture and local high-frequency information, thus achieving targeted processing of multi-level degradation information. Through an image restoration encoder, deep feature extraction is performed on images at each scale, and the encoded features are divided into degradation features and content features. Degradation features describe noise intensity, ambiguity, and degradation combination relationships, while content features retain edge, texture, and semantic information, achieving explicit separation of degradation factors and image content factors. By mapping degradation features to hyperbolic space to generate hyperbolic degradation representations and calculating the hyperbolic distance between degradation features and degradation bonds, the negative curvature structure of hyperbolic space enhances the hierarchical expressive power of degradation features, making similar degradation attribute features close to each other and different degradation mode features spaced apart, achieving accurate differentiation of degradation modes. The invention further utilizes the hyperbolic distance to calculate... By matching weights and reading restored content values ​​from the degradation key-value memory module to generate restoration prior features, an explicit association between degradation patterns and restoration knowledge is established, enabling the model to directly utilize existing restoration experience. By gating and fusing restoration prior features and content features to generate fused restoration features, the ratio of restoration prior information to content features is adaptively adjusted, avoiding information redundancy caused by a fixed fusion method. By inputting the fused restoration features into the image restoration decoder for feature reconstruction, an effective conversion of high-dimensional restoration features to image space is achieved. By transferring the restoration state information at the current scale to the next scale through the cross-scale memory transfer module after scale transformation, the global structural understanding and degradation estimation results of the previous scale are continuously transferred to the next scale, realizing a progressive restoration from coarse to fine and from global to local, comprehensively improving the image restoration system's adaptability, restoration accuracy, and stability in complex degradation environments. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a schematic flowchart of the steps of an image restoration method provided in an embodiment of the present invention; Figure 2 This is a flowchart of multi-scale image decomposition and coding feature extraction provided in an embodiment of the present invention; Figure 3 This is a structural block diagram of an image restoration device provided in an embodiment of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0019] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein is for the purpose of describing embodiments of the invention only and is not intended to limit the invention.

[0020] Furthermore, the described features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. Numerous specific details are provided in the following description to give a full understanding of embodiments of the invention. However, those skilled in the art will recognize that the technical solutions of the invention can be practiced without one or more of the specific details, or other methods, components, apparatuses, steps, etc., can be employed. In other instances, well-known methods, apparatuses, implementations, or operations are not shown or described in detail to avoid obscuring various aspects of the invention.

[0021] The block diagrams shown in the accompanying drawings are merely functional entities and do not necessarily correspond to physically independent entities. That is, these functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.

[0022] The flowcharts shown in the accompanying drawings are merely illustrative and do not necessarily include all content and operations / steps, nor do they necessarily have to be performed in the described order. For example, some operations / steps can be broken down, while others can be combined or partially combined; therefore, the actual execution order may change depending on the specific circumstances.

[0023] In related technologies, deep learning-based methods have achieved joint restoration of multiple degradation types by constructing a unified network structure. However, the all-around image restoration methods in related technologies still have the following technical shortcomings: First, existing methods have limited ability to represent degradation features. Most methods directly learn the distinguishing relationships between different degradation patterns in the traditional Euclidean feature space. However, complex image degradation often exhibits hierarchical and combinatorial relationships. For example, real noise may include multiple factors such as sensor noise, exposure variations, and compression perturbations, and different degradation patterns share some features. If only Euclidean distance is used to measure degradation similarity, it can easily lead to overlap between different degradation patterns in the feature space, making it difficult for the model to accurately determine the degradation type of the input image, thus affecting the subsequent restoration process.

[0024] Second, existing methods fail to effectively address the coupling problem between degradation information and image content information. In actual image restoration, input image features simultaneously contain two types of information: one is degradation-related information, such as noise intensity, blurriness, and brightness variations; the other is image content-related information, such as target structure, texture details, and semantic information. However, existing networks typically use hybrid features directly for degradation analysis, leading to models that may match based on image content rather than actual degradation factors. For example, sky regions and building regions may have different feature representations due to texture differences, and even if affected by the same degradation, the network may still make different judgments, thus reducing the model's generalization ability in unknown scenes.

[0025] Third, while existing encoder-decoder structures can fuse features from different levels through skip connections, they lack explicit information memory and transfer mechanisms between different restoration scales. In multi-scale restoration, the low-resolution stage is typically responsible for learning the overall image structure, brightness trends, and global degradation information, while the high-resolution stage is mainly responsible for texture and edge detail restoration. However, existing methods usually only transfer information through fixed feature connections, failing to preserve the important restoration experience gained in the low-scale stage. This leads to the need for repeated inference of global structural information at subsequent scales, reducing restoration efficiency and potentially causing inconsistencies in restoration results across different scales.

[0026] Fourth, existing image restoration methods based on memory mechanisms typically use simple feature databases to store historical sample information, lacking a structured organization of degradation patterns and restoration knowledge. As memory capacity increases, it becomes susceptible to the distribution of training data and struggles to establish a stable priori mapping between degradation patterns and restoration. Therefore, existing methods still suffer from performance degradation when faced with unknown degradation types or complex combinations of degradation.

[0027] To address the aforementioned technical problems, this invention proposes an image restoration method, apparatus, system, and storage medium. By constructing a novel restoration framework that combines hyperbolic degradation representation, structured key-value memory, and cross-scale state transfer, the accuracy and stability of image restoration under complex degradation conditions are improved.

[0028] like Figure 1 As shown, Figure 1 An image restoration method provided in this embodiment of the invention includes the following steps: S100, acquire the degraded image to be restored; S200, perform multi-scale decomposition on the degraded image to be restored according to a preset scale factor to generate a multi-scale degraded image sequence; S300, the images at each scale in the multi-scale degraded image sequence are sequentially input into the image restoration encoder, and the image restoration encoder performs deep feature extraction on the images at each scale to generate coded features, and divides the coded features into degradation features and content features; S400, the degradation features are mapped to hyperbolic space to generate a hyperbolic degradation representation, and the hyperbolic distance between the hyperbolic degradation representation and each degradation key in the degradation key memory module is calculated; S500, calculate the matching weight based on the hyperbolic distance, and read the corresponding restored content value from the degenerate key value memory module based on the matching weight to generate the restored prior features; S600, the prior recovery features and the content features are gated and fused to generate fused recovery features; S700, the fused recovery features are input to the image recovery decoder for feature reconstruction to generate a recovery image of the corresponding scale; S800: The restoration status information at the current scale is transferred to the next scale after being scaled by the cross-scale memory transfer module, so that the next scale can use the restoration status information to perform progressive image restoration and output the final restored image.

[0029] This invention proposes an image restoration method, apparatus, system, and storage medium. It achieves hierarchical representation of degradation features through a hyperbolic degradation representation mechanism, effectively separates degradation information from content information through a degradation-content orthogonal decoupling mechanism, explicitly stores and retrieves restoration knowledge through a structured key-value memory module, and effectively transfers restoration experience across different resolution stages through a cross-scale progressive memory mechanism. Based on this, the system can perform adaptive analysis, restoration knowledge retrieval, and multi-scale progressive reconstruction of complex, unknown degraded images, achieving high-quality image restoration under both conventional and unknown combined degradation conditions.

[0030] In this embodiment, the degraded image to be restored is first acquired. Based on the characteristics of image restoration, a deep learning-based image restoration system is constructed. This system consists of an input image processing module, a multi-scale degraded image construction module, an image restoration encoding module, a degradation feature extraction module, a content feature extraction module, a hyperbolic degradation mapping module, a degradation key-value memory module, a cross-scale memory transfer module, a gated fusion module, and an image restoration decoding module.

[0031] The system comprises the following modules: an input image processing module for normalizing the degraded image to be restored; a multi-scale degraded image construction module for constructing a multi-scale degraded image sequence based on a preset scale factor; an image restoration coding module for extracting deep features from images at each scale; a degraded feature extraction module and a content feature extraction module for dividing the encoded features into degraded features and content features; a hyperbolic degradation mapping module for mapping degraded features to hyperbolic space; a degradation key-value memory module for storing and retrieving the correspondence between degradation patterns and restoration priors; a cross-scale memory transfer module for transferring restoration states between different scales; a gated fusion module for adaptively fusing restoration prior features and content features; and an image restoration decoding module for reconstructing the restored image from the fused features.

[0032] In some embodiments, the image restoration encoder performs deep feature extraction on images at various scales to generate encoded features, and divides the encoded features into degradation features and content features, including: S310, The image restoration encoder performs convolution operations, residual feature learning, and nonlinear feature enhancement on the images at each scale to generate deep features; S320, The deep features are input into the degradation encoder, and the spatial dimension is compressed through feature aggregation operation to generate degradation features. The degradation features include noise intensity, blur degree, brightness change and degradation combination relationship information. S330, the deep features are input into the content encoder, and the content encoder maintains the spatial structure information of the image to generate content features, which include edge information, texture information, target structure information and semantic content information.

[0033] refer to Figure 2 The multi-scale image decomposition and coding feature extraction process is as follows: For the input degraded image, it is first normalized by the input image processing module, and a multi-scale degraded image sequence is constructed according to the preset scale factor. Specifically, the system decomposes the original input image according to different spatial resolutions, forming a multi-scale input structure from low resolution to high resolution. Different scale images undertake different levels of information recovery tasks. The low-resolution scale, due to its larger effective receptive field, can more fully acquire information on the overall image structure, brightness variation trend, and global degradation distribution; the high-resolution scale is mainly used to recover the edge structure, texture details, and local high-frequency information in the image.

[0034] After constructing multi-scale images, the degraded images at each scale are sequentially input into the image restoration encoder. The image restoration encoder performs deep feature extraction on the input images, transforming the original image space into a high-dimensional feature space containing rich semantic information through convolution operations, residual feature learning, and nonlinear feature enhancement operations. Unlike traditional image restoration networks that directly utilize single-path features for restoration, this invention further divides the extracted deep features into degradation information branches and content information branches during the encoding stage to achieve explicit separation of degradation factors and image content factors.

[0035] The degradation encoder extracts feature information related to the image degradation state. This branch compresses the spatial dimension through feature aggregation operations and generates a degradation feature vector that describes the degradation attributes of the current input image. These degradation features mainly include information such as noise intensity, blur level, brightness variation, rain / fog effects, and various degradation combinations, which are used for subsequent degradation type analysis and prior retrieval of restoration data.

[0036] Meanwhile, the content encoder is used to extract information related to the image's own structure. This branch preserves the image's spatial structure information and obtains edge, texture, target structure, and semantic content features through feature encoding. By modeling degradation features and content features separately, the problem of degradation judgment bias caused by differences in image content in traditional methods can be avoided, enabling the network to more accurately focus on real degradation factors.

[0037] In some embodiments, mapping the degenerate features to a hyperbolic space to generate a hyperbolic degenerate representation, and calculating the hyperbolic distance between the hyperbolic degenerate representation and each degenerate key in the degenerate key-value memory module, includes: S410, the degenerate feature is mapped from Euclidean space to hyperbolic space, and the hierarchical expressive power of the degenerate feature is enhanced by the negative curvature structure in hyperbolic space; S420, in the hyperbolic space, calculate the hyperbolic distance between the hyperbolic degradation representation of the current input image and each degradation key in the degradation key memory module; S430, adjust the distribution position of the degradation features in the hyperbolic space according to the hyperbolic distance, so that features with similar degradation properties are close to each other in the hyperbolic space, and features with different degradation modes are kept apart in the hyperbolic space.

[0038] Specifically, to further improve the expressive power of degradation features, this invention introduces a hyperbolic degradation mapping module after degradation feature extraction. Traditional methods typically represent degradation features in Euclidean space, but real-world image degradation exhibits significant hierarchy and combinatoriality. For example, there are inclusion relationships between different noise types and gradual changes between degradations of different intensities. Simply using Euclidean distance is insufficient to effectively describe such complex relationships.

[0039] Therefore, this invention first maps the degradation features output by the degradation encoder to a hyperbolic space, enhancing the hierarchical expressive power of the degradation features through the negative curvature structure in the hyperbolic space. In the hyperbolic space, the system calculates the distance relationship between the degradation features of the current input image and different degradation prototypes in the memory bank, making features with similar degradation attributes closer together in space, while maintaining a larger gap between features with different degradation modes. This approach results in a clearer distribution of degradation features, improving recognition accuracy under complex degradation conditions and providing a reliable basis for subsequent knowledge retrieval.

[0040] Hyperbolic space is a geometric space with constant negative curvature. Compared to Euclidean space, hyperbolic space has a natural advantage in representing hierarchical structural relationships. In hyperbolic space, the distance from a point to the origin increases exponentially, making it more effective in representing hierarchical data. For image degradation features, there are inclusion relationships between different degradation types (e.g., noise degradation includes subtypes such as Gaussian noise and salt-and-pepper noise), and there are asymptotic relationships between different degradation intensities (e.g., a continuous change from mild blur to severe blur). These hierarchical structural characteristics make hyperbolic space an ideal choice for representing degradation features.

[0041] In some embodiments, the degradation key-value memory module includes a plurality of structured memory units, each of the structured memory units consisting of a degradation key and a restored content value; the method further includes: The degradation key represents the feature prototype information corresponding to the degradation mode, and the feature prototype information includes degradation type, degradation intensity, and degradation combination relationship; The recovery experience corresponding to the degradation mode is stored through the recovery content value. The recovery experience includes structural compensation features, high-frequency texture recovery information, and recovery priors learned during the historical recovery process.

[0042] Specifically, after completing the hyperbolic space degradation matching, the system further accesses the degradation key-value memory module. This module is used to store the degradation patterns learned during the historical recovery process and the corresponding recovery prior information. It contains multiple structured memory units, each consisting of a degradation key and a recovery content value.

[0043] The degradation key (Key) represents the feature prototype information corresponding to a certain degradation mode, including degradation type, degradation intensity, and degradation combination relationship, which describes the degradation effect on the current image. The restoration content value (Value) stores the restoration experience corresponding to the degradation mode, including structural compensation features, high-frequency texture restoration information, and effective restoration priors learned in the historical restoration process, which describes how to restore the image for this degradation.

[0044] When a new degraded image is input into the system, the matching weights between the current degraded features and each key are first calculated based on the degradation similarity relationships in hyperbolic space. Then, the corresponding value information is retrieved from memory based on the matching results. Multiple relevant restoration priors are weighted and fused to form the restoration guidance features corresponding to the current input image. These features are then fused with the current scale-encoded features to participate in the subsequent image restoration process.

[0045] Compared to traditional methods that implicitly learn recovery capabilities through network parameters, this invention establishes an explicit association between degradation patterns and recovery knowledge through structured memory. This allows the model to directly utilize existing recovery experience, improving recovery efficiency under complex degradation conditions. The memory module update mechanism employs an end-to-end training approach, automatically adjusting the key and value representations based on reconstruction loss and recovery quality feedback during training, thus continuously optimizing the memory content.

[0046] In some embodiments, the step of gating and fusing the prior recovery features with the content features to generate fused recovery features includes: S610, the recovery prior features and the content features are received through the gating fusion module; S620, adaptively calculate the fusion weights based on the importance of the recovered prior features and the content features; S630, the prior recovery features and the content features are weighted and fused according to the fusion weight to generate fused recovery features.

[0047] Specifically, in the cross-scale information fusion stage, this invention further introduces a gated fusion module. This module adaptively calculates the fusion weights based on the importance of the current-scale encoded features and the memory-read features, enabling the network to dynamically adjust the ratio between the original input features and the recovered prior information. When the input image has a high degree of degradation, the system can increase the utilization ratio of the memory-recovered prior; when the input image contains rich structural information, it retains more content features extracted at the current scale.

[0048] The gating fusion module can be implemented using an attention mechanism or a gating recurrent unit structure. Taking the attention mechanism as an example, the module first calculates the attention scores of the recovered prior features and content features, then normalizes them using the Softmax function to obtain the fusion weights, and finally performs a weighted summation of the two types of features based on the weights. The gating mechanism avoids information redundancy caused by a fixed fusion method, enabling features from different sources to work together more effectively, thus improving the accuracy and stability of the recovery results.

[0049] In some embodiments, the step of transferring the recovery state information of the current scale to the next scale after scale transformation via the cross-scale memory transfer module includes: S810, after feature recovery is completed at the current scale, the low-frequency structural information extracted at the current scale, the recovered feature representation, the recovered prior features, and the recovered residual information are written into the cross-scale memory transfer module. S820, through the cross-scale memory transfer module, spatial size adjustment, channel mapping and feature normalization are performed on the written information to generate scale adaptation information; S830, the scale adaptation information is passed to the next scale recovery stage, so that the next scale recovery stage optimizes the texture, edge and local details based on the existing recovery state.

[0050] Specifically, to fully utilize the information generated at different scale recovery stages, this invention further designs a cross-scale memory transfer mechanism. After feature recovery is completed at the current scale, the system does not directly discard the information generated at the current scale, but instead writes the important recovery states obtained at that stage into the cross-scale memory module.

[0051] The written information includes low-frequency structural information extracted at the current scale, the restored feature representation, the restoration prior information obtained from memory reading, and the restoration residual information. Subsequently, the cross-scale memory module performs spatial size adjustment, channel mapping, and feature normalization on the stored information through scale transformation operations, enabling it to adapt to the next scale restoration network.

[0052] Spatial size adjustment typically employs upsampling to enlarge low-resolution feature maps to the target resolution; channel mapping is achieved through convolutional layers, adjusting the number of feature channels to match the input requirements of the next-scale network; feature normalization uses batch normalization or instance normalization to stabilize the feature distribution. Through this mechanism, the global structural understanding and degradation estimation results obtained at the previous scale can be continuously propagated to the next scale, allowing the high-resolution restoration stage to further optimize texture, edges, and local details based on the existing restoration state, rather than relearning the already acquired low-frequency information, thus achieving a progressive restoration process from coarse to fine and from global to local.

[0053] In some embodiments, the method further includes degradation-content orthogonality decoupling constraints: During training, orthogonal decoupling constraints are applied to the degradation features and the content features. These orthogonal decoupling constraints maintain the relative independence of the degradation feature space and the content feature space by restricting the correlation between the feature vectors of the degradation features and the feature vectors of the content features. The degradation feature space focuses on noise intensity, blur level, brightness variation, and environmental interference factors. The content feature space focuses on image structure, edge information, texture information, and semantic information.

[0054] Specifically, to further reduce the mutual interference between degradation information and content information, this invention introduces degradation-content orthogonal decoupling constraints during the training process. Specifically, by constraining the correlation between degradation feature vectors and content feature vectors, the two feature spaces remain relatively independent.

[0055] Orthogonal decoupling constraints can be implemented using cosine similarity loss or mutual information minimization strategies. Taking cosine similarity loss as an example, the system calculates the cosine similarity between degraded features and content features, and adds it as a regularization term to the total loss function. Minimizing this regularization term makes the two types of features tend to be orthogonal. The degradation branch focuses more on degradation factors such as noise, blur, brightness changes, and environmental influences, while the content branch focuses more on image structure, texture, and semantic information.

[0056] This orthogonal decoupling mechanism reduces the impact of image content on degradation judgment and improves the model's generalization ability when facing different scenes, targets, and unknown combinations of degradation conditions. For example, sky areas and building areas, both subject to noise pollution, may produce different responses in traditional methods due to their different texture complexities, thus reducing the accuracy of degradation estimation. However, under the orthogonal decoupling mechanism of this invention, degradation features mainly describe degradation attributes such as noise intensity, and are not affected by differences in image content, thereby improving the accuracy and consistency of degradation judgment.

[0057] Finally, after completing multi-scale feature fusion, memory restoration prior enhancement, and cross-scale information transfer, the fused features are input to the image for image reconstruction. The decoder transforms the high-dimensional restored features back into the image space through progressive upsampling, feature reconstruction, and detail enhancement operations, ultimately outputting a high-quality restored image after removing degradation factors.

[0058] Through the above process, this invention establishes a unified image restoration framework that combines degradation perception, knowledge retrieval, cross-scale transfer, and progressive restoration, achieving adaptive restoration of complex and unknown degraded images. It effectively solves the problems of insufficient degradation representation, low efficiency of restoration knowledge utilization, and insufficient multi-scale information transfer in the prior art.

[0059] Compared with related technologies, the present invention has the following advantages: This invention constructs a hyperbolic degradation representation mechanism, mapping degradation features from traditional Euclidean space to hyperbolic space. Leveraging the inherent suitability of hyperbolic space for expressing hierarchical relationships, different degradation patterns can form a clearer spatial distribution structure. By calculating the hyperbolic distance between the input degradation feature and the degradation prototype in the memory bank, degradation pattern matching can be performed more accurately, effectively solving the problem of difficulty in distinguishing complex degradation features in traditional Euclidean space and improving the accuracy of degradation recognition.

[0060] This invention achieves effective separation of degradation information and image content information by constructing degradation coding branches and content coding branches. Through degradation-content orthogonal decoupling constraints, degradation features primarily describe noise intensity, blurriness, brightness variations, and environmental interference factors, while content features primarily preserve image structure, edge information, and semantic details. By reducing the correlation between the two types of features, the restoration network can more accurately select the restoration method based on the degradation state, while reducing the impact of image content differences on degradation judgment. This significantly improves the model's generalization ability when facing different scenes, different targets, and unknown combinations of degradation conditions.

[0061] This invention employs a structured key-value memory module to explicitly organize and store existing restoration experience. A key-value structure establishes a correspondence between degradation patterns and restoration knowledge, where the key describes different degradation types, intensities, and combinations of degradation, and the value stores structural compensation information, texture restoration features, and enhancement priors for the corresponding restoration process. When a new degraded image is input, the system first determines the degradation state corresponding to the current input through a degradation matching mechanism in hyperbolic space. Then, it reads the restoration knowledge most relevant to the current state from the memory module and integrates the results into the current restoration network. This approach fully utilizes existing restoration experience, improves information utilization efficiency under complex degradation conditions, and avoids the network repeatedly learning similar degradation patterns.

[0062] This invention establishes a cross-scale progressive memory mechanism, enabling the preservation of crucial structural information, degradation estimation information, and prior restoration information generated in the low-scale restoration stage. This information is then transferred to the higher-resolution restoration stage after scale transformation. In the low-resolution stage, due to the network's larger receptive field, global structural information, degradation distribution information, and overall restoration trends can be extracted more accurately, and this information is written into the cross-scale memory module. Subsequently, the low-scale memory information is mapped to the high-resolution restoration stage through a scale adaptation module, allowing the high-scale network to directly utilize existing structural restoration experience and further optimize edges, textures, and local details. Compared to traditional skip connection methods that only perform feature splicing or direct fusion, this invention achieves progressive transfer of restoration knowledge across different scales through a memory mechanism. This allows the restoration process to gradually transition from global structural restoration to local detail enhancement, improving the consistency and visual quality of the final output image.

[0063] This invention achieves adaptive fusion of prior features and content features through a gated fusion module. This module adaptively calculates fusion weights based on the importance of features encoded at the current scale and features retrieved from the memory module, enabling the network to dynamically adjust the ratio between the original input features and the recovered prior information. When the input image has a high degree of degradation, the system increases the utilization of the memory module's prior information; when the input image contains rich structural information, it retains more content features extracted at the current scale. The gating mechanism avoids information redundancy caused by a fixed fusion method, allowing features from different sources to work more effectively together, improving the accuracy and stability of the recovery results.

[0064] This invention leverages the synergistic effects of hyperbolic degradation representation, orthogonal decoupling, structured memory, and a cross-scale progressive restoration mechanism to enable stable and high-quality image restoration in various complex visual environments. It maximizes restoration quality under stable degradation conditions and ensures restoration stability under unknown degradation combinations, comprehensively enhancing the image restoration system's adaptability to complex degradation environments. Through multi-scale progressive restoration, each restoration stage optimizes its strategy based on the restoration experience of the previous scale, achieving a progressive image reconstruction from coarse to fine and from global to local, contributing to improved accuracy and visual quality of image restoration.

[0065] Experimental verification: To verify the effectiveness of the method of the present invention, experiments were conducted on several publicly available image restoration benchmark datasets, covering joint restoration tasks with three degradations, unified restoration tasks with five degradations, and restoration tasks with a single degradation.

[0066] In a combined degradation recovery experiment involving defogging, deraining, and noise reduction, the present invention achieved superior recovery performance compared to existing methods in multiple degradation scenarios. In the defogging task, this invention achieved a recovery result of 23.42 dB PSNR and 0.905 SSIM, which is a significant improvement over existing methods AirNet (23.18 dB, 0.900 SSIM) and PromptIR (22.92 dB, 0.900 SSIM). In the rain removal task, this invention achieved 35.12 dB PSNR and 0.907 SSIM, further improving the recovery accuracy compared to PromptIR (34.54 dB, 0.900 SSIM) and AirNet (34.05 dB, 0.977 SSIM). In the denoising tasks at different noise levels, this invention achieved 34.22 dB / 0.937 SSIM, 31.58 dB / 0.895 SSIM, and 28.36 dB / 0.808 SSIM under the three test conditions of BSD68σ=15, σ=25, and σ=50, respectively, all of which exceeded the performance of existing representative methods. In summary, this invention achieves an average PSNR of 31.24 dB and a SSIM of 0.905 across the three degradation tasks, outperforming existing unified recovery frameworks.

[0067] Table 1: Comparison of Data Recovery Results for Three Degradations

[0068] In the unified restoration task encompassing five degradation categories—defogging, deraining, noise reduction, deblurring, and low-light enhancement—this invention achieves optimal or near-optimal performance across all five restoration tasks. In the defogging task, this invention achieved a PSNR of 28.32 dB and a SSIM of 0.966, significantly outperforming existing methods PromptIR (26.54 dB, 0.949 SSIM) and Restormer (24.07 dB, 0.927 SSIM). In the rain removal task, this invention achieved a PSNR of 37.38 dB and a SSIM of 0.978. In the BSD68 denoising task, this invention achieved PSNRs of 31.52 dB / 0.889 SSIM, 31.52 dB / 0.889 SSIM, and 31.52 dB / 0.889 SSIM at noise levels of σ=15, 25, and 50, respectively. In the GoPro deblurring task, this invention achieved a PSNR of 29.58 dB and a SSIM of 0.889. In the LOL-v1 low-light enhancement task, this invention achieved a PSNR of 23.02 dB and a SSIM of 0.837. Ultimately, in the comprehensive evaluation of five types of complex image restoration tasks, this invention achieved an average performance of 30.24 dB PSNR and 0.908 SSIM, significantly outperforming existing unified image restoration networks.

[0069] Table 2: Comparison of data recovery effects in five degraded states

[0070] Tests in a single degradation scenario demonstrate that this invention also achieves significantly better performance than existing methods. In the standalone dehazing task, this invention achieves 23.42 dB PSNR and 0.905 SSIM; in the standalone deraining task, it achieves 35.12 dB PSNR and 0.907 SSIM; in BSD68 tests at different noise levels, this invention achieves 34.22 dB, 31.58 dB, and 28.36 dB PSNR at σ=15, 25, and 50, respectively. These results further illustrate that the hyperbolic degradation modeling, orthogonal feature decoupling, and memory enhancement mechanisms proposed in this invention are not only designed for multi-task joint restoration, but can effectively improve feature representation capabilities and restoration quality in various image restoration tasks.

[0071] Below is a comparison table of the recovery effects of single degenerate data: Table 3: Image Dehazing (Dataset: SOTS)

[0072] Table 4: Image rain removal (Dataset: Rain100L)

[0073] Table 5: Image Denoising (Dataset: CBSD68)

[0074] The experimental results above demonstrate that the hyperbolic degradation representation mechanism proposed in this invention can effectively solve the problem of difficulty in distinguishing complex degradation features in traditional Euclidean space. Since real image degradation typically exhibits a clear hierarchical structure and combinational relationships—for example, there are continuous variations between mild, moderate, and strong noise, and different degradation types may have overlapping relationships—traditional Euclidean feature space easily leads to feature confusion between different degradation patterns. This invention maps degradation features to hyperbolic space, leveraging the inherent suitability of hyperbolic space for expressing hierarchical structures, enabling different degradation patterns to form a clearer spatial distribution relationship. By calculating the hyperbolic distance between the input degradation features and the degradation prototypes in memory, degradation pattern matching can be performed more accurately, thus providing more reliable prior information for the subsequent recovery process. The experimental results for three- and five-degradation degradations show that even when faced with unknown degradation combinations, this invention maintains high recovery performance, indicating that hyperbolic degradation modeling effectively enhances the model's ability to express complex degradation distributions.

[0075] Meanwhile, this invention achieves effective separation of degradation information and image content information by constructing degradation coding branches and content coding branches. Existing image restoration methods typically use mixed features directly for restoration, which can easily lead to the target region, texture structure, and semantic information affecting the degradation judgment process. This invention utilizes degradation-content orthogonal decoupling constraints, allowing degradation features to primarily describe noise intensity, blurriness, brightness changes, and environmental interference factors, while content features primarily preserve image structure, edge information, and semantic details. By reducing the correlation between the two types of features, the restoration network can more accurately select the restoration method based on the degradation state, while reducing the impact of image content differences on degradation judgment.

[0076] Furthermore, this invention designs a structured key-value memory module to explicitly organize and store existing restoration experience. Unlike traditional methods that rely on network parameters to implicitly learn the mapping relationship between degradation and restoration, this invention establishes a correspondence between degradation patterns and restoration knowledge through a key-value structure. When a new degraded image is input, the system first determines the degradation state corresponding to the current input through a degradation matching mechanism in hyperbolic space, then reads the restoration knowledge most relevant to the current state from the memory, and integrates the reading results into the current restoration network. This approach can fully utilize existing restoration experience, improve information utilization efficiency under complex degradation conditions, and avoid the network repeatedly learning similar degradation patterns.

[0077] To address the issue of insufficient information transfer between different scales in traditional Encoder-Decoder reconstruction networks, this invention further proposes a cross-scale progressive Memory reconstruction mechanism. In the low-resolution stage, due to the network's larger receptive field, it can more accurately extract global structural information, degradation distribution information, and overall reconstruction trends, and write this information into the cross-scale memory. Subsequently, a scale adaptation module maps the low-scale memory information to the high-resolution reconstruction stage, enabling the high-scale network to directly utilize existing structural reconstruction experience and further optimize edges, textures, and local details. Compared to traditional skip connection methods that only perform feature splicing or direct fusion, this invention achieves progressive transfer of reconstruction knowledge between different scales through the memory mechanism, allowing the reconstruction process to gradually transition from global structural reconstruction to local detail enhancement, thus improving the consistency and visual quality of the final output image.

[0078] See Figure 3 This invention provides an image restoration device, comprising: The first module is used to acquire the degraded image to be restored; The second module is used to perform multi-scale decomposition on the degraded image to be restored according to a preset scale factor, and generate a multi-scale degraded image sequence. The third module is used to input the images at each scale in the multi-scale degraded image sequence into the image restoration encoder in sequence, and to extract deep features from the images at each scale through the image restoration encoder to generate coded features, and to divide the coded features into degraded features and content features. The fourth module is used to map the degradation features to hyperbolic space, generate a hyperbolic degradation representation, and calculate the hyperbolic distance between the hyperbolic degradation representation and each degradation key in the degradation key memory module. The fifth module is used to calculate the matching weight based on the hyperbolic distance, and read the corresponding restored content value from the degenerate key value memory module based on the matching weight to generate the restored prior features; The sixth module is used to perform gated fusion of the prior recovery features and the content features to generate fused recovery features; The seventh module is used to perform feature reconstruction on the fused recovery feature input image recovery decoder to generate a recovery image of the corresponding scale; The eighth module is used to transfer the restoration status information of the current scale to the next scale after scale conversion via the cross-scale memory transfer module, so that the next scale can use the restoration status information to perform progressive image restoration and output the final restored image.

[0079] It is evident that the content of the above method embodiments is applicable to this device embodiment. The specific functions implemented in this device embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0080] This invention provides an image restoration system, comprising: At least one processor; At least one memory for storing at least one program; When the at least one program is executed by the at least one processor, the at least one processor performs the method described above.

[0081] It is evident that the content of the above method embodiments is applicable to this system embodiment. The specific functions implemented in this system embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0082] Furthermore, embodiments of the present invention also disclose a computer program product or computer program stored in a computer-readable storage medium. A processor of a computer device can read the computer program from the computer-readable storage medium, and the processor executes the computer program, causing the computer device to perform the described method. Similarly, the content of the above method embodiments is applicable to this storage medium embodiment. The specific functions implemented in this storage medium embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0083] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0084] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0085] In the several embodiments provided by this invention, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0086] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0087] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0088] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes multiple instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing programs, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0089] The preferred embodiments of the present invention have been described above with reference to the accompanying drawings, but this does not limit the scope of the claims of the present invention. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and spirit of the present invention should be within the scope of the claims of the present invention.

Claims

1. An image restoration method, characterized in that, The method includes the following steps: Obtain the degraded image to be restored; The degraded image to be restored is decomposed into multiple scales according to a preset scale factor to generate a multi-scale degraded image sequence. The images at each scale in the multi-scale degraded image sequence are sequentially input into the image restoration encoder. The image restoration encoder performs deep feature extraction on each scale image to generate coded features, and the coded features are divided into degradation features and content features. The degradation features are mapped to hyperbolic space to generate a hyperbolic degradation representation, and the hyperbolic distance between the hyperbolic degradation representation and each degradation key in the degradation key memory module is calculated. The matching weight is calculated based on the hyperbolic distance, and the corresponding restored content value is read from the degenerate key value memory module based on the matching weight to generate the restoration prior features; The prior recovery features and the content features are gated and fused to generate fused recovery features; The fused restoration features are input into the image restoration decoder for feature reconstruction to generate a restored image at the corresponding scale; The restoration status information at the current scale is transferred to the next scale after being scaled by the cross-scale memory transfer module. The next scale then uses the restoration status information to perform progressive image restoration and outputs the final restored image.

2. The method according to claim 1, characterized in that, The process involves extracting deep features from images at various scales using the image restoration encoder to generate encoded features, and then dividing these encoded features into degradation features and content features, including: The image restoration encoder performs convolution operations, residual feature learning, and nonlinear feature enhancement on the images at each scale to generate deep features. The deep features are input into the degradation encoder, and the spatial dimension is compressed through feature aggregation operation to generate degradation features. The degradation features include noise intensity, blur degree, brightness change and degradation combination relationship information. The deep features are input into the content encoder, which preserves the spatial structure information of the image and generates content features. The content features include edge information, texture information, target structure information, and semantic content information.

3. The method according to claim 1, characterized in that, The step of mapping the degenerate features to hyperbolic space to generate a hyperbolic degenerate representation, and calculating the hyperbolic distance between the hyperbolic degenerate representation and each degenerate key in the degenerate key-value memory module, includes: The degenerate features are mapped from Euclidean space to hyperbolic space, and the hierarchical expressive power of the degenerate features is enhanced by the negative curvature structure in hyperbolic space. In the hyperbolic space, calculate the hyperbolic distance between the hyperbolic degradation representation of the current input image and each degradation key in the degradation key memory module; The distribution position of the degradation features in the hyperbolic space is adjusted according to the hyperbolic distance, so that features with similar degradation properties are close to each other in the hyperbolic space, and features with different degradation modes are kept apart in the hyperbolic space.

4. The method according to claim 1, characterized in that, The degenerate key-value memory module comprises multiple structured memory units, each of which consists of a degenerate key and a restored content value; the method further includes: The degradation key represents the feature prototype information corresponding to the degradation mode, and the feature prototype information includes degradation type, degradation intensity, and degradation combination relationship; The recovery experience corresponding to the degradation mode is stored through the recovery content value. The recovery experience includes structural compensation features, high-frequency texture recovery information, and recovery priors learned during the historical recovery process.

5. The method according to claim 1, characterized in that, The step of gating and fusing the prior recovery features with the content features to generate fused recovery features includes: The recovery prior features and the content features are received through the gating fusion module; Based on the importance of the recovered prior features and the content features, the fusion weights are adaptively calculated. The prior recovery features and the content features are weighted and fused according to the fusion weights to generate fused recovery features.

6. The method according to claim 1, characterized in that, The step of transferring the recovery state information at the current scale to the next scale after scale conversion via the cross-scale memory transfer module includes: After feature recovery is completed at the current scale, the low-frequency structural information extracted at the current scale, the recovered feature representation, the recovered prior features, and the recovered residual information are written into the cross-scale memory transfer module. The cross-scale memory transfer module performs spatial size adjustment, channel mapping, and feature normalization on the written information to generate scale-adaptive information. The scale adaptation information is passed to the next scale recovery stage, so that the next scale recovery stage can optimize texture, edges and local details based on the existing recovery state.

7. The method according to claim 1, characterized in that, The method further includes: During training, orthogonal decoupling constraints are applied to the degradation features and the content features. These orthogonal decoupling constraints maintain the relative independence of the degradation feature space and the content feature space by restricting the correlation between the feature vectors of the degradation features and the feature vectors of the content features. The degradation feature space focuses on noise intensity, blur level, brightness variation, and environmental interference factors. The content feature space focuses on image structure, edge information, texture information, and semantic information.

8. An image restoration device, characterized in that, include: The first module is used to acquire the degraded image to be restored; The second module is used to perform multi-scale decomposition on the degraded image to be restored according to a preset scale factor, and generate a multi-scale degraded image sequence. The third module is used to input the images at each scale in the multi-scale degraded image sequence into the image restoration encoder in sequence, and to extract deep features from the images at each scale through the image restoration encoder to generate coded features, and to divide the coded features into degraded features and content features. The fourth module is used to map the degradation features to hyperbolic space, generate a hyperbolic degradation representation, and calculate the hyperbolic distance between the hyperbolic degradation representation and each degradation key in the degradation key memory module. The fifth module is used to calculate the matching weight based on the hyperbolic distance, and read the corresponding restored content value from the degenerate key value memory module based on the matching weight to generate the restored prior features; The sixth module is used to perform gated fusion of the prior recovery features and the content features to generate fused recovery features; The seventh module is used to perform feature reconstruction on the fused recovery feature input image recovery decoder to generate a recovery image of the corresponding scale; The eighth module is used to transfer the restoration status information of the current scale to the next scale after scale conversion via the cross-scale memory transfer module, so that the next scale can use the restoration status information to perform progressive image restoration and output the final restored image.

9. An image restoration system, characterized in that, include: At least one processor; At least one memory for storing at least one program; When the at least one program is executed by the at least one processor, the at least one processor performs the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a processor-executable program, characterized in that, The processor-executable program, when executed by the processor, is used to perform the method as described in any one of claims 1 to 7.

Citation Information

Patent Citations

  • Degenerated image restoration method and device, equipment, medium and product

    CN122347507A

  • Image restoration method and system and storage medium

    CN122492500A