A priori perception adaptive reconstruction method for remote sensing image unknown degradation fusion

Abstract

The application proposes a prior-aware adaptive reconstruction method for remote sensing image unknown degradation fusion. The method first acquires the main source degraded image and the auxiliary guide image, extracts features through the dual-source feature extraction module, and performs cross-source interaction modeling; then, based on the interaction features, a structured degradation prior is generated, which includes global degradation embedding, multi-scale spatial degradation response, channel restoration control factor and fusion confidence constraint map; under the guidance of the prior, the fusion branch weight and the local modulation map are dynamically calculated, the prior-aware multi-strategy adaptive fusion and channel-level restoration modulation are performed, and finally the high-quality fusion reconstruction result is output. In addition, the method introduces a reference guide correction mechanism, which uses reference images to correct the prior generation process during the training stage, and eliminates the dependence on reference images during the inference stage. The method effectively enhances the fusion robustness in complex degradation scenarios, and has good task universality and engineering application value.

Description

Technical Field

[0001] The embodiments of this application relate to the field of remote sensing image processing technology, and in particular to a priori perception adaptive reconstruction method for fusion of unknown degradation in remote sensing images. Background Technology

[0002] Remote sensing image fusion technology aims to integrate complementary information from different source images to obtain high-quality fused images (such as panchromatic and multispectral fusion, optical and SAR fusion, etc.). In real remote sensing scenarios, due to the limitations of sensor hardware, atmospheric turbulence, complex meteorological conditions, and transmission interference, the acquired main source images often suffer from unknown degradation. These degradations may include noise, blurring, striping, missing elements, compression distortion, local contamination, or combinations thereof, and the intensity of degradation is extremely unevenly distributed in both spatial and channel dimensions.

[0003] Existing remote sensing image fusion methods typically rely on predefined single degradation models (such as ideal Gaussian blur or specific noise distributions) or separate degradation processing from fusion processing. These methods often struggle to accurately model complex, mixed, and unknown degradation in real-world scenarios. Furthermore, many deep learning-based fusion methods require high-quality reference images for strong supervision during the inference phase, which are often unavailable in practical applications. Traditional methods and some deep models employ a uniform fusion strategy for all regions and channels, making it difficult to strike a balance between "degradation restoration" in severely degraded areas and "information fidelity and detail enhancement" in well-degraded areas.

[0004] Based on the above analysis, the existing technology has the following two urgent technical problems that need to be solved.

[0005] First, existing methods struggle to explicitly and effectively model mixed degradation in real-world scenarios where the type is unknown and the spatial and channel distributions are uneven, resulting in poor fusion robustness and generalization ability in complex degradation scenarios.

[0006] Second, most existing methods adopt fixed global fusion rules and lack the ability to adaptively modulate different degradation regions, different channel information and different scale details, making it difficult to balance degradation repair, information fidelity and auxiliary detail enhancement. Summary of the Invention

[0007] To address the aforementioned technical issues, embodiments of this application propose a priori-aware adaptive reconstruction method for fusion of unknown degradation in remote sensing images. The method aims to first perform structured modeling of unknown degradation in the main source image to generate transferable degradation priors, and then use the degradation priors to perform global, local, and channel-level modulation on the subsequent fusion and reconstruction process, ultimately achieving adaptive processing of different degradation regions, different channel information, and different scale details.

[0008] To achieve the above objectives, embodiments of this application propose a prior-aware adaptive reconstruction method for fusion of unknown degradation in remote sensing images. This method is based on a reconstruction model composed of a dual-source feature extraction module, a lightweight cross-source interaction module, a reference-guided degradation prior correction module, a structured degradation prior generation module, a multi-strategy adaptive fusion module, and a reconstruction decoder. The method includes the following steps: S1, acquiring a primary source degradation image and an auxiliary guidance image; S2, using the primary encoder and auxiliary encoder of the dual-source feature extraction module to extract features from the primary source degradation image and the auxiliary guidance image, respectively, to obtain primary source image features and auxiliary image features; S3, establishing a correlation representation between degradation information and guidance information through the lightweight cross-source interaction module, performing cross-source interaction on the primary source image features and auxiliary image features to obtain primary source guidance features and auxiliary perception features; S4, using the structured degradation prior... The proof generation module infers the main source guiding features and auxiliary perceptual features to obtain the inference path representation, and generates a structured degradation prior based on the inference path representation, including global degradation embedding, multi-scale spatial degradation response, channel recovery control factor, and fusion credibility constraint graph; S5, the multi-strategy adaptive fusion module performs multi-dimensional adaptive modulation of global fusion strategy, local region processing, and channel recovery behavior based on the structured degradation prior to obtain multi-layer fusion features; S6, the multi-layer fusion features are sent to the reconstruction decoder for image reconstruction to obtain a high-quality reconstructed image; In the training phase, a high-quality reference image is additionally acquired, and the reference-guided degradation prior correction module corrects the degradation prior generation process based on the reference features extracted from the high-quality reference image, and through consistency constraints, the reconstruction model can accurately extract degradation semantics without a high-quality reference image during the inference phase.

[0009] Optionally, the main encoder and auxiliary encoder of the dual-source feature extraction module are used to extract features from the main source degraded image and the auxiliary guide image, respectively, to obtain main source image features and auxiliary image features, including: Let the main source degraded image be , The auxiliary guidance image is , , The number of channels in the primary degraded image. and The spatial dimensions of the primary source degraded image. To assist in the number of channels of the guiding image, This indicates the spatial scale relationship between the primary degraded image and the auxiliary guiding image; Main encoder using dual-source feature extraction module and auxiliary encoder The main source degradation image was analyzed separately. and auxiliary guidance images Feature extraction is performed to obtain the main source image features. and auxiliary image features ; , .

[0010] Optionally, a lightweight cross-source interaction module is used to establish a correlation representation between degradation information and guidance information, and cross-source interaction is performed on the main source image features and auxiliary image features to obtain main source guidance features and auxiliary perception features, including: To avoid excessive substitution of the main source degraded image by the structure of the auxiliary guiding image, a cross-source interaction operator is designed. and Main source image features and auxiliary image features Lightweight coupling is performed to obtain the main source guidance features after interaction. and auxiliary perceptual features ; ; ; in, and All are composed of gated convolutional or dynamic mapping units, used to stably extract degradation representations from the main source degradation image with auxiliary structural reference.

[0011] Optionally, a high-quality reference image is set as follows: , , The number of channels for a high-quality reference image. via reference encoder Perform feature extraction to obtain reference features ; During the training phase, a correction operator is used. Main source guidance features Auxiliary perceptual features and reference features Processing is performed to generate a reference correction characterization. , ; At the same time, use the inference path operator Directly guide the main source features and auxiliary perceptual features Processing to generate inference path representations , .

[0012] Optionally, a structured degradation prior is generated based on the inference path representation, including a global degradation embedding, a multi-scale spatial degradation response, a channel recovery control factor, and a fusion credibility constraint graph, comprising: Based on reasoning path representation A structured degenerate prior set is constructed using multiple generator heads. ; ; ; ; ; ; in, This is a global degradation embedding used to describe the overall degradation state, information loss trend, and restoration preference of the main source degradation image. Generate a global degradation embedding header. For the first Multi-scale spatial degradation responses at various scales are used to reflect the intensity and type mixing of degradation in different regions. For the first A degradation response generation head at each scale, For the total number of scales, These are channel recovery control factors, used to describe the reliability, recovery difficulty, and recovery priority of each channel. Generate a header for channel control. To integrate the credibility constraint graph, which is used to characterize the degree of acceptance of auxiliary guidance information in different regions of the output space, Generate a header for the credibility constraint graph.

[0013] Optionally, the multi-strategy adaptive fusion module performs multi-dimensional adaptive modulation of the global fusion strategy, local region processing, and channel recovery behavior based on structured degradation priors to obtain multi-layer fusion features, including: Suppose that the multi-layer basic fusion feature representation extracted from the main source degraded image and the auxiliary guiding image is... and build There are three fusion branches with different processing preferences, denoted as the first. The fusion branches are , ; By global degradation embedding Sample-level fusion branch weights are generated using linear mapping and the Softmax function. ; ; in, and All are learnable parameters. For the Softmax function; In the Layers will respond to multi-scale spatial degradation. With fusion credibility constraint graph The joint mapping is a local modulation map. , , This is a modulation mapping function used to control the fusion ratio between the branch output at the current position and the basic features; Integrating global branch control and local region modulation, the first Layer fusion features Represented as: ; in, This indicates that element-wise multiplication is performed. For the first The weights of each fusion branch; Using channel recovery control factor Intermediate representation during the fusion and reconstruction process Modulation is represented as: ; in, and They are respectively from The generated scaling and offset functions, For the first The middle representation of each channel, For the first Features after modulation of each channel.

[0014] Optionally, the reconstruction decoder is denoted as Multi-layer fusion features Send to reconstruction decoder Perform image reconstruction to obtain a high-quality reconstructed image. , .

[0015] Optionally, the reconstruction model can be optimized and trained in stages, with the overall training objective being... Represented as: ; in, This indicates a priori consistency constraint. This indicates channel reliability constraints. This represents a multi-scale degenerate distribution constraint. This represents the total variation constraint. This represents the fusion reconstruction constraint, used to compute a high-quality reconstructed image. With high-quality reference images Reconstruction error between , , , , These are the weighting coefficients for each item.

[0016] Optionally, prior consistency constraints Used for constraint reasoning path generation Approximation reference correction embedding The degenerate semantics it carries are represented as: ; in, This indicates the search for the L2 norm. Channel credibility constraints Based on the difference measurement between the main source degraded image and the high-quality reference image in each channel. Construct reference confidence level To impose constraints, it is represented as: ; ; in, For adjustment coefficients, The number of channels for high-quality reconstructed images; Multiscale Degeneracy Distribution Constraints Multiscale residual maps constructed based on main source degraded images and high-quality reference images Monitoring spatial degradation response The distribution of is expressed as: ; in, For channel compression or mapping functions, This indicates the search for the L1 norm. Integration and Reconstruction Constraints High-quality reconstructed images With high-quality reference images Direct supervision is represented as: ;、 ; in, For any supplementary structural consistency or channel consistency item, The corresponding weights are given.

[0017] Optionally, after obtaining high-quality reconstructed images, they can be applied to tasks such as fusion of unknown degraded remote sensing images, multi-source remote sensing fusion, and related remote sensing reconstruction.

[0018] This application proposes a priori-aware adaptive reconstruction method for fusion of unknown degradation in remote sensing images. It employs a novel technical framework that decouples and organically combines degradation modeling and fusion reconstruction, achieving efficient modeling and high-quality reconstruction of unknown complex degradation. Existing techniques often conflate degradation processing and fusion processing, while this application first performs structured modeling of unknown degradation in the source image to generate transferable degradation priors. Then, it utilizes these priors to perform global, local, and channel-level modulation on subsequent fusion and reconstruction processes, ultimately achieving adaptive processing of different degradation regions, different channel information, and different scale details. This design completely eliminates the strong assumptions of traditional methods regarding specific degradation types (such as single Gaussian blur or fixed additive noise), enabling it to accurately fit complex degradation in the real world that is spatially non-uniform, of unknown type, and highly mixed. It minimizes the interference of degradation components on complementary information extraction, thus producing clear and high-fidelity reconstruction results even under extremely challenging observation conditions. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies of this application will be briefly introduced below. The following drawings are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. The drawings described herein are only used to explain this application and are not intended to limit this application.

[0020] Figure 1 This is a flowchart of a priori perception adaptive reconstruction method for unknown degradation fusion of remote sensing images provided in one embodiment of this application; Figure 2 This is a schematic diagram of the prior perception adaptive reconstruction framework for unknown degradation fusion of remote sensing images provided in one embodiment of this application; Figure 3 This is a schematic diagram of the structure of a multi-strategy adaptive fusion reconstruction module provided in one embodiment of this application; Figure 4 This is a schematic diagram of a phased training strategy provided in one embodiment of this application. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the various embodiments of this application will be described in detail below with reference to the accompanying drawings. Those skilled in the art will understand that many technical details have been presented in the embodiments of this application to facilitate better understanding. However, the technical solutions claimed in this application can be implemented even without these technical details and various variations and modifications based on the following embodiments. The division of the following embodiments is for ease of description and should not constitute any limitation on the specific implementation of this application. The following embodiments can be combined with and referenced by each other without contradiction.

[0022] One embodiment of this application proposes a priori-aware adaptive reconstruction method for fusion of unknown degradation in remote sensing images. This method is based on a reconstruction model consisting of a dual-source feature extraction module, a lightweight cross-source interaction module, a reference-guided degradation prior correction module, a structured degradation prior generation module, a multi-strategy adaptive fusion module, and a reconstruction decoder. The implementation details of this priori-aware adaptive reconstruction method for fusion of unknown degradation in remote sensing images are described below. These details are provided for ease of understanding and are not essential for implementing this solution.

[0023] The specific process of the prior-aware adaptive reconstruction method for fusion of unknown degradation in remote sensing images proposed in this embodiment can be described as follows: Figure 1 As shown, the specific architecture of the reconstruction model is as follows: Figure 2 As shown, the method includes: S1, acquire the primary source degradation image and the secondary guide image.

[0024] S2, using the main source encoder and auxiliary encoder of the dual-source feature extraction module, features are extracted from the main source degraded image and the auxiliary guide image respectively, to obtain the main source image features and the auxiliary image features.

[0025] Specifically, the basis for image reconstruction is to obtain the main source degraded image and the auxiliary guide image. Then, the main source encoder and the auxiliary encoder of the dual-source feature extraction module are used to extract features from the main source degraded image and the auxiliary guide image respectively, so as to obtain the main source image features and the auxiliary image features.

[0026] Figure 2 This is a schematic diagram of the prior perception adaptive reconstruction framework for fusion of unknown degradation of remote sensing images provided in this embodiment, which shows the overall network topology from input image, feature extraction, prior generation to multi-strategy fusion reconstruction.

[0027] In the part of acquiring input data, it is necessary to acquire the main source degraded image and the auxiliary guide image. In the training phase, it is also necessary to acquire high-quality reference images.

[0028] Let the main source degraded image be , The auxiliary guidance image is , , The number of channels in the primary degraded image. and The spatial dimensions of the primary source degraded image. To assist in the number of channels of the guiding image, This indicates the spatial scale relationship between the primary degraded image and the auxiliary guiding image.

[0029] Main encoder using dual-source feature extraction module and auxiliary encoder The main source degradation image was analyzed separately. and auxiliary guidance images Feature extraction is performed to obtain the main source image features. and auxiliary image features .

[0030] , .

[0031] S3 establishes a correlation representation between degradation information and guidance information through a lightweight cross-source interaction module, and performs cross-source interaction on the main source image features and auxiliary image features to obtain the main source guidance features and auxiliary perception features.

[0032] Specifically, in the cross-source interaction modeling stage, it is necessary to establish the association between degradation information and guidance information through a lightweight cross-source interaction module, and to perform cross-source interaction on the main source image features and auxiliary image features to obtain the main source guidance features and auxiliary perception features.

[0033] To avoid excessive substitution of the structure of the auxiliary guiding image by the main source degraded image, this embodiment designs a cross-source interaction operator. and Main source image features and auxiliary image features Lightweight coupling is performed to obtain the main source guidance features after interaction. and auxiliary perceptual features .

[0034] ; ; in, and All are composed of gated convolutional or dynamic mapping units, used to stably extract degradation representations from the main source degradation image with auxiliary structural reference.

[0035] S4 uses a structured degradation prior generation module to reason about the main source guiding features and auxiliary perception features to obtain a reasoning path representation, and generates a structured degradation prior based on the reasoning path representation, including global degradation embedding, multi-scale spatial degradation response, channel recovery control factor and fusion credibility constraint graph.

[0036] Specifically, after obtaining the main source guiding features and auxiliary sensing features, the structured degradation prior generation module can be used to reason about the main source guiding features and auxiliary sensing features to obtain the reasoning path representation, and generate a structured degradation prior including global degradation embedding, multi-scale spatial degradation response, channel recovery control factor and fusion credibility constraint graph based on the reasoning path representation.

[0037] During the training phase, a correction operator is used. Main source guidance features Auxiliary perceptual features and reference features Processing is performed to generate a reference correction characterization. , .

[0038] At the same time, use the inference path operator Directly guide the main source features and auxiliary perceptual features Processing to generate inference path representations , .

[0039] Based on reasoning path representation A structured degenerate prior set is constructed using multiple generator heads. , represented as ; , ; , ; , ; , ; in, This is a global degradation embedding used to describe the overall degradation state, information loss trend, and restoration preference of the main source degradation image. Generate a global degradation embedding header. For the first Multi-scale spatial degradation responses at various scales are used to reflect the intensity and type mixing of degradation in different regions. For the first A degradation response generation head at each scale, For the total number of scales, These are channel recovery control factors, used to describe the reliability, recovery difficulty, and recovery priority of each channel. Generate a header for channel control. To integrate the credibility constraint graph, which is used to characterize and constrain the degree of acceptance of auxiliary guidance information in different regions of the output space, Generate a header for the credibility constraint graph.

[0040] S5, based on the structured degradation prior, performs multi-dimensional adaptive modulation of the global fusion strategy, local region processing and channel recovery behavior by the multi-strategy adaptive fusion module to obtain multi-layer fusion features.

[0041] Specifically, after obtaining the structured degradation prior, the multi-strategy adaptive fusion module can perform multi-dimensional adaptive modulation of the global fusion strategy, local region processing, and channel recovery behavior based on the structured degradation prior to obtain multi-layer fusion features.

[0042] like Figure 3 As shown, the fusion reconstruction process of the multi-strategy adaptive fusion module includes global strategy control, local region modulation, and channel-level recovery modulation.

[0043] Suppose that the multi-layer basic fusion feature representation extracted from the main source degraded image and the auxiliary guiding image is... and build A fusion branch with different processing preferences (such as fidelity, detail enhancement, and degradation repair).

[0044] Record No. The fusion branches are , .

[0045] By global degradation embedding Sample-level fusion branch weights are generated using linear mapping and the Softmax function. .

[0046] ; in, and All are learnable parameters. This refers to the Softmax function.

[0047] In the Layers will respond to multi-scale spatial degradation. With fusion credibility constraint graph The joint mapping is a local modulation map. , , This is a modulation mapping function used to control the fusion ratio between the branch output at the current position and the base feature.

[0048] Integrating global branch control and local region modulation, the first Layer fusion features Represented as: ; in, This indicates that element-wise multiplication is performed. For the first The weights of each fusion branch.

[0049] Using channel recovery control factor Intermediate representation during the fusion and reconstruction process Modulation is represented as: ; in, and They are respectively from The generated scaling and offset functions, For the first The middle representation of each channel, For the first The features modulated by each channel are used to achieve multi-dimensional adaptive modulation.

[0050] S6 feeds the multi-layer fusion features into the reconstruction decoder for image reconstruction, resulting in a high-quality reconstructed image.

[0051] Specifically, let the reconstruction decoder be represented as Multi-layer fusion features Send to reconstruction decoder Perform image reconstruction to obtain a high-quality reconstructed image. , .

[0052] Figure 4 This is a schematic diagram of a phased training strategy. In this embodiment, a multi-constraint joint optimization objective is used to perform phased training on the reconstruction of the fusion results.

[0053] The first phase focuses on training modules related to degenerate prior generation (i.e., dual-source feature extraction, cross-source interaction, and prior generation head), utilizing reference correction representations. Guide the reasoning path and extract transferable degenerate semantics.

[0054] The second-stage fixed or partially fixed prior generation module feeds the multi-layer fusion features, which have undergone multi-dimensional adaptive modulation, into the reconstruction decoder. Output the final high-quality fused image. .

[0055] The reconstruction model is optimized and trained in stages, with the overall training objective being... Represented as: ; in, This indicates a priori consistency constraint. This indicates channel reliability constraints. This represents a multi-scale degenerate distribution constraint. This represents the total variation constraint. This represents the fusion reconstruction constraint, used to compute a high-quality reconstructed image. With high-quality reference images Reconstruction error between , , , , These are the weighting coefficients for each item.

[0056] Prior consistency constraints Used for constraint reasoning path generation Approximation reference correction embedding The degenerate semantics it carries are represented as: ; in, This indicates the calculation of the L2 norm.

[0057] Channel credibility constraints Based on the difference measurement between the main source degraded image and the high-quality reference image in each channel. Construct reference confidence level To impose constraints, it is represented as: ; ; in, For adjustment coefficients, The number of channels for high-quality reconstructed images.

[0058] Multiscale Degeneracy Distribution Constraints Multiscale residual maps constructed based on main source degraded images and high-quality reference images Monitoring spatial degradation response The distribution of is expressed as: ; in, For channel compression or mapping functions, This indicates the calculation of the L1 norm.

[0059] Integration and Reconstruction Constraints High-quality reconstructed images With high-quality reference images Direct supervision is represented as: ;、 ; in, For any supplementary structural consistency or channel consistency item, The corresponding weights are given.

[0060] Once high-quality reconstructed images are obtained, they can be applied to the fusion of remote sensing images with unknown degradation (such as multispectral and panchromatic image fusion), multi-source remote sensing fusion (such as optical and SAR fusion), and related remote sensing reconstruction tasks, demonstrating strong engineering application value in practical deployments. Furthermore, since this embodiment does not rely on high-quality reference images during the inference phase, its high performance and low latency characteristics enable it to be seamlessly deployed in satellite vehicle-mounted processing terminals, UAV payload platforms, and large-scale ground-based remote sensing data centers, which has significant practical application value for promoting the automation and intelligence of remote sensing image processing.

[0061] This embodiment proposes a priori-aware adaptive reconstruction method for fusion of unknown degradation in remote sensing images. It employs a novel technical framework that decouples and organically combines degradation modeling and fusion reconstruction, achieving efficient modeling and high-quality reconstruction of unknown complex degradation. Existing technologies often conflate degradation processing and fusion processing. This embodiment first performs structured modeling of unknown degradation in the source image to generate transferable degradation priors. Then, it utilizes these priors to perform global, local, and channel-level modulation on subsequent fusion and reconstruction processes, ultimately achieving adaptive processing of different degradation regions, different channel information, and different scale details. This design completely eliminates the strong assumptions of traditional methods regarding specific degradation types (such as single Gaussian blur or fixed additive noise), enabling it to accurately fit complex degradation in the real world—spatially non-uniform, of unknown type, and highly mixed—and minimizing the interference of degradation components on complementary information extraction. This results in clear, high-fidelity reconstruction even under extremely challenging observation conditions.

[0062] The expected benefits and commercial value of the technical solution in this embodiment are extremely significant. This embodiment can recover high-quality texture and spectral features from main-source degraded images that have been severely affected by degradation (such as cloud cover, sensor stripe noise, atmospheric turbulence interference, etc.). It can also be widely applied in scenarios with harsh weather conditions or limited by sensor hardware bottlenecks, such as disaster emergency assessment, wide-area natural resource surveys, and high-precision detection and reconnaissance. Furthermore, it significantly improves the availability of remote sensing data under adverse observation conditions, greatly reducing the cost of satellite or UAV re-enhancing due to substandard data. Simultaneously, the high-quality reconstructed images output provide a more reliable data foundation for subsequent downstream advanced vision tasks such as remote sensing target detection and land cover classification, indirectly improving the accuracy of the entire remote sensing information processing chain.

[0063] The technical solution of this embodiment fills the technical gap in the domestic and international industry regarding "blind adaptive fusion without predefined degradation models and reference images." Existing high-performance fusion networks based on deep learning often rely heavily on high-quality reference images as guidance during the inference phase in practical deployments. This embodiment innovatively introduces a reference-guided correction mechanism, enabling the reconstruction model to learn transferable representations of degradation semantics during the training phase. In the actual inference (practical application) phase, it completely eliminates the dependence on high-quality reference images and pre-known degradation parameters, successfully achieving true reference-free adaptive fusion. This greatly expands the deployment capability and engineering versatility of the reconstruction model in real-world open scenarios.

[0064] The technical solution of this embodiment solves a long-standing but unresolved technical problem: the amplification of local errors in the fusion process caused by extremely uneven spatial and channel degradation. Traditional whole-image unified fusion strategies often suffer from trade-offs when dealing with mixed degradation (e.g., blurring clear details of the entire image while repairing severely degraded areas). This embodiment achieves precise modulation at the pixel and channel levels through multi-scale spatial degradation response and channel restoration control factors. This not only effectively suppresses the spread and amplification of local degradation errors to the entire image, but also perfectly balances "powerful repair of severely degraded areas" with "high-fidelity preservation of original information in intact areas," truly achieving precise image restoration "tailored to local conditions," and further improving the rationality of high-quality reconstructed images.

[0065] The technical solution in this embodiment overcomes technical bias. For a long time, there has been a prevalent technical bias in the industry, namely, the belief that using a unified, fixed fusion network structure and parameter rules is sufficient to handle all image inputs. This embodiment breaks this dependence on inertial network structures and innovatively designs multiple fusion branches with different processing preferences (such as fidelity, detail enhancement, and degradation repair). Combined with four types of structured degradation priors, the reconstruction model can perform multi-level dynamic adaptive modulation at the sample level according to the actual degradation state of the input, fundamentally improving the flexibility and performance ceiling of fusion and image reconstruction.

[0066] The steps described above are merely for clarity in describing the technical solution. In actual implementation, they can be combined into one step, or certain steps can be broken down into multiple steps, as long as they involve the same logical relationship, they are all within the scope of protection of this application. Any insignificant modifications or designs added to the algorithm or process, as long as they do not change the core of the algorithm or process, are also within the scope of protection of this application.

[0067] In the above embodiments, each network module can be implemented entirely or partially through convolutional networks, Transformers, or other equivalent network structures. When implemented entirely or partially as a computer program product, the computer program product includes one or more computer instructions. When the computer program instructions are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device.

[0068] It will be understood by those skilled in the art that the above embodiments are specific implementations of this application, and in practical applications, various changes can be made in form, detail, and description without departing from the spirit and scope of this application. For those skilled in the art, several improvements and modifications can be made without departing from the principles of this application, and these improvements and modifications are also considered to be within the scope of protection of this application.

Claims

1. A prior-aware adaptive reconstruction method for fusion of unknown degradation in remote sensing images, based on a reconstruction model consisting of a dual-source feature extraction module, a lightweight cross-source interaction module, a reference-guided degradation prior correction module, a structured degradation prior generation module, a multi-strategy adaptive fusion module, and a reconstruction decoder, characterized in that... The method includes: S1, acquire the primary source degradation image and the auxiliary guidance image; S2, using the main source encoder and auxiliary encoder of the dual-source feature extraction module, features are extracted from the main source degraded image and the auxiliary guide image respectively to obtain the main source image features and the auxiliary image features; S3 establishes a correlation representation between degradation information and guidance information through a lightweight cross-source interaction module, performs cross-source interaction on main source image features and auxiliary image features, and obtains main source guidance features and auxiliary perception features. S4. The structured degradation prior generation module is used to reason about the main source guiding features and auxiliary perception features to obtain the reasoning path representation. Based on the reasoning path representation, a structured degradation prior including global degradation embedding, multi-scale spatial degradation response, channel recovery control factor and fusion credibility constraint graph is generated. S5, based on the structured degradation prior, the multi-strategy adaptive fusion module performs multi-dimensional adaptive modulation of the global fusion strategy, local region processing and channel recovery behavior to obtain multi-layer fusion features; S6, the multi-layer fusion features are fed into the reconstruction decoder to reconstruct the image and obtain a high-quality reconstructed image; During the training phase, high-quality reference images are additionally acquired. The degradation prior correction module guided by the reference images corrects the degradation prior generation process based on the reference features extracted from the high-quality reference images. Through consistency constraints, the reconstruction model can accurately extract degradation semantics without high-quality reference images during the inference phase.

2. The prior-aware adaptive reconstruction method for fusion of unknown degradation in remote sensing images according to claim 1, characterized in that, The dual-source feature extraction module uses a primary encoder and an auxiliary encoder to extract features from the primary degraded image and the auxiliary guiding image, respectively, to obtain primary image features and auxiliary image features, including: Let the main source degraded image be , The auxiliary guidance image is , , The number of channels in the primary degraded image. and The spatial dimensions of the primary source degraded image. To assist in the number of channels of the guiding image, This indicates the spatial scale relationship between the primary degraded image and the auxiliary guiding image; Main encoder using dual-source feature extraction module and auxiliary encoder The main source degradation image was analyzed separately. and auxiliary guidance images Feature extraction is performed to obtain the main source image features. and auxiliary image features ; ， 。 3. The prior-perception adaptive reconstruction method for unknown degradation fusion of remote sensing images according to claim 2, characterized in that, A lightweight cross-source interaction module is used to establish a correlation between degradation information and guidance information. Cross-source interaction is performed on the main source image features and auxiliary image features to obtain main source guidance features and auxiliary perception features, including: To avoid excessive substitution of the main source degraded image by the structure of the auxiliary guiding image, a cross-source interaction operator is designed. and Main source image features and auxiliary image features Lightweight coupling is performed to obtain the main source guidance features after interaction. and auxiliary perceptual features ; ； ； in, and All are composed of gated convolutional or dynamic mapping units, used to stably extract degradation representations from the main source degradation image with auxiliary structural reference.

4. The prior-perception adaptive reconstruction method for unknown degradation fusion of remote sensing images according to claim 3, characterized in that, let... High-quality reference image , , The number of channels for a high-quality reference image. via reference encoder Perform feature extraction to obtain reference features ; During the training phase, a correction operator is used. Main source guidance features Auxiliary perceptual features and reference features Processing is performed to generate a reference correction characterization. , ; At the same time, use the inference path operator Directly guide the main source features and auxiliary perceptual features Processing to generate inference path representations , .

5. The prior-aware adaptive reconstruction method for unknown degradation fusion of remote sensing images according to claim 4, characterized in that, Structured degradation priors, generated based on inference path representations, include global degradation embeddings, multi-scale spatial degradation responses, channel recovery control factors, and fusion credibility constraint graphs. Based on reasoning path representation A structured degenerate prior set is constructed using multiple generator heads. ; ； ； ； ； ； in, This is a global degradation embedding used to describe the overall degradation state, information loss trend, and restoration preference of the main source degradation image. Generate a global degradation embedding header. For the first Multi-scale spatial degradation responses at various scales are used to reflect the intensity and type mixing of degradation in different regions. For the first A degradation response generation head at each scale, For the total number of scales, These are channel recovery control factors, used to describe the reliability, recovery difficulty, and recovery priority of each channel. Generate a header for channel control. To integrate the credibility constraint graph, which is used to characterize the degree of acceptance of auxiliary guidance information in different regions of the output space, Generate a header for the credibility constraint graph.

6. The prior-aware adaptive reconstruction method for unknown degradation fusion of remote sensing images according to claim 5, characterized in that, The multi-strategy adaptive fusion module performs multi-dimensional adaptive modulation of the global fusion strategy, local region processing, and channel recovery behavior based on structured degradation priors, resulting in multi-layer fusion features, including: Suppose that the multi-layer basic fusion feature representation extracted from the main source degraded image and the auxiliary guiding image is... and build There are three fusion branches with different processing preferences, denoted as the first. The fusion branches are , ; By global degradation embedding Sample-level fusion branch weights are generated using linear mapping and the Softmax function. ; ； in, and All are learnable parameters. For the Softmax function; In the Layers will respond to multi-scale spatial degradation. With fusion credibility constraint graph The joint mapping is a local modulation map. , , This is a modulation mapping function used to control the fusion ratio between the branch output at the current position and the basic features; Integrating global branch control and local region modulation, the first Layer fusion features Represented as: ； in, This indicates that element-wise multiplication is performed. For the first The weights of each fusion branch; Using channel recovery control factor Intermediate representation during the fusion and reconstruction process Modulation is represented as: ； in, and They are respectively from The generated scaling and offset functions, For the first The middle representation of each channel, For the first Features after modulation of each channel.

7. The prior-aware adaptive reconstruction method for unknown degradation fusion of remote sensing images according to claim 6, characterized in that, Let the reconstruction decoder be denoted as Multi-layer fusion features Send to reconstruction decoder Perform image reconstruction to obtain a high-quality reconstructed image. , .

8. The prior-aware adaptive reconstruction method for unknown degradation fusion of remote sensing images according to claim 7, characterized in that, The reconstruction model is optimized and trained in stages, with the overall training objective being... Represented as: ； in, This indicates a priori consistency constraint. This indicates channel reliability constraints. This represents a multi-scale degenerate distribution constraint. This represents the total variation constraint. This represents the fusion reconstruction constraint, used to compute a high-quality reconstructed image. With high-quality reference images Reconstruction error between , , , , These are the weighting coefficients for each item.

9. A priori perceptual adaptive reconstruction method for fusion of unknown degradation in remote sensing images according to claim 8, characterized in that, Prior consistency constraints Used for constraint reasoning path generation Approximation reference correction embedding The degenerate semantics it carries are represented as: ； in, This indicates the search for the L2 norm. Channel credibility constraints Based on the difference measurement between the main source degraded image and the high-quality reference image in each channel. Construct reference confidence level To impose constraints, it is represented as: ； ； in, For adjustment coefficients, The number of channels for high-quality reconstructed images; Multiscale Degeneracy Distribution Constraints Multiscale residual maps constructed based on main source degraded images and high-quality reference images Monitoring spatial degradation response The distribution of is expressed as: ； in, For channel compression or mapping functions, This indicates the search for the L1 norm. Integration and Reconstruction Constraints High-quality reconstructed images With high-quality reference images Direct supervision is represented as: ；、 ； in, For any supplementary structural consistency or channel consistency item, The corresponding weights are given.

10. A priori-aware adaptive reconstruction method for fusion of unknown degradation in remote sensing images according to any one of claims 1 to 9, characterized in that, After obtaining high-quality reconstructed images, they are applied to tasks such as fusion of unknown degraded remote sensing images, multi-source remote sensing fusion, and related remote sensing reconstruction.

Images