A gradual chain type multi-modal remote sensing image super-resolution reconstruction method and device
By employing a progressively chained multimodal remote sensing image super-resolution reconstruction method, and utilizing an adaptive multimodal alignment module and a progressively chained gating fusion mechanism, the problems of modal alignment and noise suppression in multimodal remote sensing image super-resolution reconstruction are solved, achieving efficient detail restoration and semantic consistency improvement.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- XIAN UNIV OF POSTS & TELECOMM
- Filing Date
- 2026-01-23
- Publication Date
- 2026-05-26
AI Technical Summary
Existing multimodal remote sensing image super-resolution reconstruction methods have shortcomings in modal alignment and noise suppression, making it difficult to fully utilize multi-source heterogeneous information, resulting in insufficient detail restoration and semantic consistency.
A progressive multimodal remote sensing image super-resolution reconstruction method is adopted. The potential correlation between the main mode and the auxiliary mode is mined by a context-aware adaptive multimodal alignment module. The progressive gating fusion mechanism is used to fuse features step by step. The modal contribution ratio is adaptively adjusted by combining the reliability score to suppress noise interference.
It achieves effective interaction and alignment of multimodal features, improves the detail restoration effect and semantic consistency of images, and enhances the realism and fidelity of reconstruction.
Smart Images
Figure CN122089569A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of image processing technology, and in particular to a method and apparatus for super-resolution reconstruction of multimodal remote sensing images using a progressive chain approach. Background Technology
[0002] Image super-resolution reconstruction aims to recover high-resolution images from low-resolution images to enhance spatial details and visual quality, and is an important research direction in the field of computer vision. This technology plays a crucial role in areas such as ground object recognition, environmental monitoring, and urban planning. Traditional super-resolution methods mainly rely on interpolation, reconstruction constraints, or prior models, and their ability to recover high-frequency details in complex scenes is limited. With the development of deep learning technology, super-resolution methods based on convolutional neural networks (CNNs) learn the nonlinear mapping between low-resolution and high-resolution images end-to-end, significantly outperforming traditional methods. However, limited by the local receptive field of convolutional operations, CNN methods have shortcomings in modeling long-range dependencies. To overcome these limitations, the Transformer architecture, with its self-attention mechanism, has been introduced into image super-resolution tasks, providing a new technical approach for image super-resolution reconstruction. Although Transformer-based methods have surpassed traditional CNNs in reconstruction performance, they generally rely on single-modal information and struggle to fully utilize the multi-source heterogeneous information contained in remote sensing images, thus still falling short in detail restoration and semantic consistency.
[0003] To further improve the performance of super-resolution reconstruction, research has begun to incorporate multimodal information into image super-resolution tasks. The core idea of multimodal super-resolution is to utilize complementary features between different modalities to compensate for the shortcomings of single modalities in terms of texture details, semantic structure, and spatial consistency. Early methods mainly relied on guided filtering or feature concatenation to achieve simple fusion, but these methods struggled to effectively address issues such as modal differences and spatial shifts. Subsequently, the rise of the Transformer architecture propelled multimodal super-resolution to higher levels. MTrans, through an asymmetric multi-scale cross-modal Transformer design, injects fine-grained structural information into the target modality in small blocks and provides high-level global context in large blocks, thus improving reconstruction performance. MMSR, through token-by-token multimodal encoding and efficient multimodal latent connectors, seamlessly injects multi-source context into the pre-trained model. Combined with a novel multimodal classifier-free guided mechanism, it effectively suppresses illusions while maintaining linear computational complexity, thereby significantly improving the realism and reconstruction fidelity of image super-resolution.
[0004] Despite the positive progress made by these methods, multimodal image super-resolution tasks still face unresolved challenges. Firstly, significant differences in feature distributions among different modalities make modal alignment difficult. Secondly, the inconsistent quality of auxiliary modalities easily introduces noise interference. Finally, existing fusion strategies often employ simple stitching, failing to fully reflect the complementary relationships between multimodal features at the hierarchical semantic level. Summary of the Invention
[0005] The main objective of this invention is to provide a method and apparatus for super-resolution reconstruction of multimodal remote sensing images using a progressive chain, which solves the problem of poor structural detail restoration in existing progressive chain multimodal remote sensing image super-resolution reconstruction methods.
[0006] To achieve the above objectives, the present invention provides a progressively chained multimodal remote sensing image super-resolution reconstruction method, the method comprising: S10, Input a low-resolution image; S20, extract the main modality features and auxiliary modality features from the low-resolution image; S30, the main modal features and auxiliary modal features are input into the context-aware adaptive multimodal alignment module, and the input modal features are processed to obtain the processed auxiliary modal features; S40, the processed auxiliary modal features are fused step by step at different levels through a progressively chained gating fusion mechanism to obtain the final fused features; S50, the final fusion features are input into a preset reconstruction module to obtain the final super-resolution reconstructed image; S60 outputs the final super-resolution reconstructed image.
[0007] Optionally, step S20 includes the following steps: S210: Input the low-resolution image LR into the pre-trained SegFormer segmentation model to obtain a semantic map, CLIP to generate text descriptions, and a texture map to generate a small CNN network. S220, the low-resolution image, the texture map, and the semantic map are divided into several non-overlapping image blocks, and mapped to a unified feature space through linear projection to obtain the main modality features. and auxiliary modal features .
[0008] Optionally, step S30 includes the following steps: S310, based on the main modal features Calculate the self-attention matrix Based on the auxiliary modal features Calculate the self-attention matrix and ; S320, calculate each auxiliary mode With dominant modal features Differences between characteristics Modal features , and Perform concatenation and then global average pooling; S330, based on the modal features after global average pooling, obtains the reliability score of the current modality. ; S340, based on the reliability score The self-attention matrix V is weighted and modified to obtain the modified self-attention matrix. ; S350, in , and Attention calculations are performed during the process; S360, acquires a variety of texture features.
[0009] Optionally, step S330 includes the following steps: S3310 performs non-linear mapping on the modal features after global average pooling through a fully connected layer; S3320, based on the Sigmoid activation function, maps the output to an interval Obtain the reliability score of the current mode. .
[0010] Optionally, the reliability score of the current mode Obtained based on the following formula:
[0011] in, It is the Sigmoid activation function. This indicates a fully connected layer mapping operation. This indicates feature splicing.
[0012] Optionally, step S40 includes the following steps: S410 will process the texture features pass , , Three convolutional kernels are used to extract texture features at different scales, and then the multi-scale features are concatenated and processed. Convolution, batch normalization, and ReLU activation function are fused to obtain texture enhancement features. ; S420, texture enhancement features With dominant modal features Concatenated input gating generator network generates weights The fused features are obtained by adaptive weighted fusion based on the weights. ; S430, the processed texture features Channel attention weights are generated using global average pooling and two multilayer perceptrons. right Weighting and through Semantic enhancement features are obtained by fusing convolution, batch normalization, and ReLU activation functions. ; S440, and Concatenated input gating generator network generates weights And obtain by fusion. ; S450, the processed texture features After re-plastic surgery Convolutional GELU activation function and Convolution enhances text features. ; S460, and Concatenated input gating generator network generates weights And based on this, the final fusion feature is obtained by fusion. .
[0013] Optionally, the fusion feature Obtained based on the following formula:
[0014]
[0015] in, Indicates a gating network, by Convolution, ReLU activation function It consists of stacked convolutional and sigmoid activation functions; Indicates the weights generated by the network; Indicates the dominant modal characteristics; This indicates a texture enhancement feature.
[0016] Furthermore, to achieve the above objectives, embodiments of this application also provide a progressively chained multimodal remote sensing image super-resolution reconstruction device, the device comprising: Image input module, used for inputting low-resolution images; The feature extraction module is used to extract primary modality features and secondary modality features from low-resolution images; The feature processing module is used to input the main modal features and auxiliary modal features into the context-aware adaptive multimodal alignment module, and process the input modal features to obtain the processed auxiliary modal features; The feature fusion module is used to fuse the processed auxiliary modal features at different levels through a progressively chained gating fusion mechanism to obtain the final fused features. An image reconstruction module is used to input the final fusion features into a preset reconstruction module to obtain the final super-resolution reconstructed image; The image output module is used to output the final super-resolution reconstructed image.
[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: The progressive multimodal remote sensing image super-resolution reconstruction method provided in this application employs a context-aware adaptive multimodal alignment module. By mining the potential correlations between the primary and auxiliary modalities, it achieves interaction and alignment of multimodal features. It also employs a progressively chained gated fusion mechanism, integrating modal features at different levels step by step to achieve progressive reconstruction from local detail enhancement to high-level semantic guidance. Furthermore, this application adaptively adjusts the contribution ratio of each modality by obtaining the reliability score of each modality feature, suppressing noise interference and improving the stability of feature alignment. Attached Figure Description
[0018] Figure 1 A flowchart of the progressive multimodal remote sensing image super-resolution reconstruction method provided in the embodiments of this application; Figure 2 A structural block diagram of the progressively multimodal remote sensing image super-resolution reconstruction device provided in the embodiments of this application; Figure 3 This is a schematic diagram of the structure of the context-aware adaptive multimodal alignment module provided in an embodiment of this application.
[0019] The realization of the purpose, functional features and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation
[0020] It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of the application. Rather, these embodiments are provided to make the disclosure more thorough and complete, and to fully convey the scope of the disclosure to those skilled in the art.
[0021] To address the aforementioned technical problems, embodiments of this application provide a progressively chained multimodal remote sensing image super-resolution reconstruction method, which can be executed by a computer, such as... Figure 1 As shown, the method may include the following steps: S10, Input a low-resolution image.
[0022] S20 extracts primary modality features and secondary modality features from low-resolution images.
[0023] Among them, auxiliary modal features include texture, semantic and text features.
[0024] In an exemplary embodiment, step S20 includes the following steps: S210: Input the low-resolution image LR into the pre-trained SegFormer segmentation model to obtain a semantic map and CLIP to generate text descriptions. Then, generate a texture map through a small CNN (convolutional neural network) network (3×3 convolution, ReLU activation function and 1×1 convolution). S220, the low-resolution image, the texture map, and the semantic map are divided into several non-overlapping image blocks, and mapped to a unified feature space through linear projection to obtain the main modality features. and auxiliary modal features Auxiliary modal features Including texture features semantic features and text features .
[0025] S30, the main modal features and auxiliary modal features are input into the context-aware adaptive multimodal alignment module, and the input modal features are processed to obtain the processed auxiliary modal features.
[0026] like Figure 3 As shown, to fully explore the potential correlations between multimodal features, a context-aware adaptive multimodal alignment module utilizes the Transformer architecture to design an adaptive alignment mechanism. This mechanism performs feature alignment in parallel at different levels: the texture layer focuses on matching high-frequency details, the semantic layer focuses on the correspondence between spatial regions, and the text layer establishes semantic associations at the content level, thereby generating aligned multimodal feature representations. (See figure.) and Indicates output features, After passing through a module called MRE, the corrected feature R is generated. t Next, R t The attention matrix V is weighted and modified to obtain the modified self-attention matrix. The following examples are... Figure 3A detailed description of the steps.
[0027] In an exemplary embodiment, step S30 includes the following steps: S310, based on the main modal features Calculate the self-attention matrix Based on the auxiliary modal features Calculate the self-attention matrix and ; S320, calculate each auxiliary mode With dominant modal features Differences between characteristics Modal features , and Perform concatenation and then global average pooling; S330, based on the modal features after global average pooling, obtains the reliability score of the current modality. ; S340, based on the reliability score The self-attention matrix V is weighted and modified to obtain the modified self-attention matrix. ; S350, in , and Attention calculations are performed during the process; S360, acquires a variety of texture features.
[0028] Figure 3 The example demonstrates multiple parallel attention heads, each independently computing attention weights and applying them to... The matrix, the output of each attention head is normalized using the softmax function to obtain the final attention weights, these weights are... Matrix multiplication yields a weighted feature representation for each head. The outputs of all heads are concatenated or otherwise fused to form the final output feature X. t .
[0029] Among them, the corrected self-attention matrix Obtained based on the following formula:
[0030]
[0031] in, This is the corrected value matrix. For reliability score, For value matrices, This represents element-wise multiplication; Let be the modal feature dimension, and softmax represent the activation function. This represents the attention function.
[0032] This context-aware adaptive multimodal alignment module, guided by three layers of auxiliary information—texture, semantics, and text—follows a process from local to global and from low to high levels to progressively extract, match, and align modal features. This module adaptively enhances the feature representation of low-resolution images, promoting consistency in structure, semantics, and content, thereby generating higher-resolution reconstruction results with greater detail and semantic integrity.
[0033] In an exemplary embodiment, step S330 includes the following steps: S3310 performs non-linear mapping on the modal features after global average pooling through a fully connected layer; S3320, based on the Sigmoid activation function, maps the output to an interval Obtain the reliability score of the current mode. .
[0034] In an exemplary embodiment, the reliability score of the current mode Obtained based on the following formula:
[0035] in, For reliability score, It is the Sigmoid activation function. This represents the fully connected layer mapping operation, and GAP represents global average pooling. Indicates feature splicing, Indicates auxiliary modal features, This represents the dominant modality feature.
[0036] S40, the processed auxiliary modal features are fused step by step at different levels through a progressively chained gating fusion mechanism to obtain the final fused features.
[0037] In an exemplary embodiment, step S40 includes the following steps: S410 will process the texture features pass , , Three convolutional kernels are used to extract texture features at different scales, and then the multi-scale features are concatenated and processed. Convolution, batch normalization, and ReLU activation function are fused to obtain texture enhancement features. The expression is as follows:
[0038] in, Represents texture enhancement features, ReLU represents the activation function, and BN represents batch normalization. Indicates the kernel size as Convolution operation, Indicates feature splicing, Represents the aligned texture features; S420, texture enhancement features With dominant modal features Concatenated input gating generator network generates weights The fused features are obtained by adaptive weighted fusion based on the weights. ; S430, the processed texture features Channel attention weights are generated using global average pooling and two multilayer perceptrons. right Weighting and through Semantic enhancement features are obtained by fusing convolution, batch normalization, and ReLU activation functions. ; S440, and Concatenated input gating generator network generates weights And obtain by fusion. ; S450, the processed texture features After re-plastic surgery Convolutional GELU activation function and Convolution enhances text features. ; S460, and Concatenated input gating generator network generates weights And based on this, the final fusion feature is obtained by fusion. .
[0039] In this exemplary embodiment, semantic features focus on global relationships and region structure, compared to texture features. First, we extract high-level semantic dependencies and generate channel attention weights using GAP (Global Average Pooling) and two multilayer perceptrons. :
[0040] in, The channel attention weights are represented by MLP, MLP stands for Multilayer Perceptron, and GAP stands for Global Average Pooling. This represents the semantic features after alignment.
[0041] Then, use the obtained Weighting the semantic features yields enhanced semantic features. :
[0042] in, ReLU represents enhanced semantic features, BN represents batch normalization, and ReLU represents activation function. This represents a convolution operation with a kernel size of 1. This indicates the features of the aligned text.
[0043] In an exemplary embodiment, the fusion feature Obtained based on the following formula:
[0044]
[0045] in, Indicates a gating network, by Convolution, ReLU activation function It consists of stacked convolutional and sigmoid activation functions; Indicates the weights generated by the network; Indicates the dominant modal characteristics; This indicates a texture enhancement feature.
[0046] This exemplary embodiment employs a progressive chain structure for multimodal feature fusion under the guidance of the primary modality, and incorporates a gating mechanism to adaptively adjust the contribution of each modality feature. This fusion mechanism achieves synergistic complementarity of texture, semantic, and textual features at the hierarchical level, effectively alleviating the feature imbalance problem in traditional methods, thereby balancing low-level detail enhancement with high-level semantic consistency.
[0047] S50, the final fused features are input into a preset reconstruction module to obtain the final super-resolution reconstructed image. This reconstruction module consists of a 3×3 convolutional layer, a sub-pixel convolutional upsampling layer, and a terminal 3×3 convolutional layer for feature upsampling. The final super-resolution reconstructed image is obtained based on the following formula:
[0048] in, This represents the reconstructed image. Y represents the reconstruction module, and Y represents the final fusion feature.
[0049] S60 outputs the final super-resolution reconstructed image.
[0050] In the above embodiments, a context-aware adaptive multimodal alignment module is employed to achieve interaction and alignment of multimodal features by mining the potential correlations between the primary and secondary modalities. A progressive chain gating fusion mechanism is used to achieve progressive reconstruction from local detail enhancement to high-level semantic guidance by integrating modal features at different levels. Furthermore, this application adaptively adjusts the contribution ratio of each modality by obtaining the reliability score of each modality feature, suppressing noise interference and improving the stability of feature alignment.
[0051] Based on the above embodiments, refer to Figure 2 Another embodiment of this application also provides a progressively chained multimodal remote sensing image super-resolution reconstruction device, which may include the following modules: Image input module 210 is used to input low-resolution images; Feature extraction module 220 is used to extract main modality features and auxiliary modality features from low-resolution images; The feature processing module 230 is used to input the main modal features and auxiliary modal features into the context-aware adaptive multimodal alignment module, and process the input modal features to obtain the processed auxiliary modal features; The feature fusion module 240 is used to fuse the processed auxiliary modal features at different levels through a progressively chained gating fusion mechanism to obtain the final fused features. Image reconstruction module 250 is used to input the final fusion features into a preset reconstruction module to obtain the final super-resolution reconstructed image; Image output module 260 is used to output the final super-resolution reconstructed image.
[0052] In the description of this application, it should be noted that the terms "first", "second", and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.
[0053] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.
[0054] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Additionally, the coupling or direct coupling or communication connection shown or discussed may be through some communication interface; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.
[0055] 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.
[0056] In addition, the functional units in the various embodiments of this application 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.
[0057] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a processor-executable, non-volatile, computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion 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 several 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 this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.
[0058] Finally, it should be noted that the above-described embodiments are merely specific implementations of this application, used to illustrate the technical solutions of this application, and not to limit them. The protection scope of this application is not limited thereto. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some of the technical features, within the technical scope disclosed in this application. Such modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be covered within the protection scope of this application. Therefore, the protection scope of this application should be determined by the protection scope of the claims.
[0059] Furthermore, although the operations of the method of this application are described in a specific order in the accompanying drawings, this does not require or imply that these operations must be performed in that specific order, or that all the operations shown must be performed to achieve the desired result. Additionally or alternatively, certain steps may be omitted, multiple steps may be combined into one step, and / or one step may be broken down into multiple steps.
Claims
1. A progressively chained multimodal remote sensing image super-resolution reconstruction method, characterized in that, The method includes the following steps: S10, Input a low-resolution image; S20, extract the main modality features and auxiliary modality features from the low-resolution image; S30, the main modal features and auxiliary modal features are input into the context-aware adaptive multimodal alignment module, and the input modal features are processed to obtain the processed auxiliary modal features; S40, the processed auxiliary modal features are fused step by step at different levels through a progressively chained gating fusion mechanism to obtain the final fused features; S50, the final fusion features are input into a preset reconstruction module to obtain the final super-resolution reconstructed image; S60 outputs the final super-resolution reconstructed image.
2. The progressive multimodal remote sensing image super-resolution reconstruction method according to claim 1, characterized in that, Step S20 includes the following steps: S210: Input the low-resolution image LR into the pre-trained SegFormer segmentation model to obtain a semantic map, CLIP to generate text descriptions, and a texture map to generate a small CNN network. S220, the low-resolution image, the texture map, and the semantic map are divided into several non-overlapping image blocks, and mapped to a unified feature space through linear projection to obtain the main modality features. and auxiliary modal features .
3. The progressive multimodal remote sensing image super-resolution reconstruction method according to claim 2, characterized in that, Step S30 includes the following steps: S310, based on the main modal features Calculate the self-attention matrix Based on the auxiliary modal features Calculate the self-attention matrix and ; S320, calculate each auxiliary mode With dominant modal features Differences between characteristics Modal features , and Perform concatenation and then global average pooling; S330, based on the modal features after global average pooling, obtains the reliability score of the current modality. ; S340, based on the reliability score The self-attention matrix V is weighted and modified to obtain the modified self-attention matrix. ; S350, in , and Attention calculations are performed during the process; S360, acquires a variety of texture features.
4. The progressive multimodal remote sensing image super-resolution reconstruction method according to claim 3, characterized in that, Step S330 includes the following steps: S3310 performs non-linear mapping on the modal features after global average pooling through a fully connected layer; S3320, based on the Sigmoid activation function, maps the output to an interval Obtain the reliability score of the current mode. .
5. The progressive multimodal remote sensing image super-resolution reconstruction method according to claim 4, characterized in that, The reliability score of the current mode Obtained based on the following formula: in, For reliability score, It is the Sigmoid activation function. This represents the fully connected layer mapping operation, and GAP represents global average pooling. Indicates feature splicing, Indicates auxiliary modal features, This represents the dominant modality feature.
6. The progressive multimodal remote sensing image super-resolution reconstruction method according to claim 1, characterized in that, Step S40 includes the following steps: S410 will process the texture features pass , , Three convolutional kernels are used to extract texture features at different scales, and then the multi-scale features are concatenated and processed. Convolution, batch normalization, and ReLU activation function are fused to obtain texture enhancement features. ; S420, texture enhancement features With dominant modal features Concatenated input gating generator network generates weights The fused features are obtained by adaptive weighted fusion based on the weights. ; S430, the processed texture features Channel attention weights are generated using global average pooling and two multilayer perceptrons. right Weighting and through Semantic enhancement features are obtained by fusing convolution, batch normalization, and ReLU activation functions. ; S440, and Concatenated input gating generator network generates weights And obtain by fusion. ; S450, the processed texture features After re-plastic surgery Convolutional GELU activation function and Convolution enhances text features. ; S460, and Concatenated input gating generator network generates weights And based on this, the final fusion feature is obtained by fusion. .
7. The progressive multimodal remote sensing image super-resolution reconstruction method according to claim 6, characterized in that, The fusion feature Obtained based on the following formula: in, Indicates a gating network, by Convolution, ReLU activation function It consists of stacked convolutional and sigmoid activation functions; Indicates the weights generated by the network; Indicates the dominant modal characteristics; This indicates a texture enhancement feature.
8. A progressively chained multimodal remote sensing image super-resolution reconstruction device, characterized in that, The device includes: Image input module, used for inputting low-resolution images; The feature extraction module is used to extract primary modality features and secondary modality features from low-resolution images; The feature processing module is used to input the main modal features and auxiliary modal features into the context-aware adaptive multimodal alignment module, and process the input modal features to obtain the processed auxiliary modal features; The feature fusion module is used to fuse the processed auxiliary modal features at different levels through a progressively chained gating fusion mechanism to obtain the final fused features. An image reconstruction module is used to input the final fusion features into a preset reconstruction module to obtain the final super-resolution reconstructed image; The image output module is used to output the final super-resolution reconstructed image.