Hyperspectral fusion imaging method based on topology perception and differential homeomorphic mechanism

Abstract

The invention relates to a hyperspectral calculation imaging technology, and aims to solve the problems that in an actual remote sensing observation environment, a single imaging mode is limited by a physical sensor, high space resolution and hyperspectral resolution are difficult to obtain at the same time, and geometry and spectrum complementarity exists among multi-source images. The invention provides a hyperspectral fusion imaging method based on topology perception and a differential homeomorphic mechanism. According to the method, differential homeomorphic transformation is introduced to construct spatial degradation priori, a dual learning framework is adopted to jointly optimize high-resolution generation and bidirectional reconstruction tasks, meanwhile, a topology sensing Transform module is utilized to explicitly model a spectrum-space double-domain manifold structure, complementary correlation among multi-source data is fully mined, and the spectrum-space double-domain manifold structure is constructed. The spatial structure consistency and the spectral fidelity of the fused image are effectively improved, the problem of fusion quality reduction caused by geometric deformation and spectral variation in a complex ground feature scene is significantly improved, and a more reliable data support is provided for applications such as remote sensing monitoring and environmental perception.

Description

Technical Field

[0001] This invention belongs to the field of computational imaging and remote sensing image processing, and specifically relates to a hyperspectral fusion imaging method based on topological sensing and differential homeomorphism mechanism. Background Technology

[0002] Hyperspectral imaging technology aims to construct high-dimensional spectral-spatial joint representations of ground features. However, due to the physical characteristics of imaging sensors, it is difficult to achieve high spatial and spectral resolution simultaneously with a single mode. Therefore, information fusion has become a mainstream reconstruction paradigm, using auxiliary modes to compensate for missing information in the main mode, thereby achieving cross-modal enhancement.

[0003] As a typical multi-source reconstruction method, hyperspectral image fusion leverages the complementarity between hyperspectral and multispectral images to generate images with both rich spectral fidelity and fine spatial details, making it a core technology in computational imaging. However, existing hyperspectral image fusion methods have significant shortcomings in geometric constraint modeling. In the spatial domain, spatial misalignment and structural distortion easily occur, leading to a decrease in the spatial consistency of the fusion result. In the spectral domain, the lack of explicit modeling of local spectral neighborhoods makes it difficult to effectively reconstruct fine-grained spectral details.

[0004] Traditional deep learning-based models and prior-based methods often struggle to maintain structural integrity in complex terrain scenes, limiting their generalization ability in high-dimensional feature spaces. Hyperspectral images typically exhibit complex geometric constraints and nonlinear manifold structures, where spatial deformation of object boundaries reflects pixel-level correspondences across modalities, while fine-grained spectral variations convey rich information about terrain textures. Existing methods struggle to dynamically adapt to geometric deformations and spectral variations in diverse scenes, lacking a unified framework for integrating spatial deformation modeling with high-dimensional spectral manifold learning.

[0005] Therefore, there is an urgent need for a general fusion method that can integrate spatial deformation modeling and high-dimensional spectral manifold learning, so as to effectively model multi-scale geometric features and nonlinear mapping relationships while maintaining spectral fidelity and spatial structure consistency. Summary of the Invention

[0006] The technical problem this invention aims to solve is the deficiency of existing hyperspectral image fusion methods in geometric constraint modeling, particularly the problems of spatial misalignment and structural distortion in the spatial domain and the difficulty in effectively recovering fine-grained spectral details in the spectral domain. This invention provides a hyperspectral fusion imaging method based on topological awareness and differential homeomorphism. This invention introduces differential homeomorphism to construct a structure-preserving spatial degradation prior, and utilizes a dual learning framework to jointly optimize high-resolution generation and bidirectional low-resolution reconstruction tasks. Simultaneously, it employs a topologically aware Transformer module to explicitly model the spectral-spatial dual-domain manifold structure, thereby improving the spatial consistency and spectral fidelity of the fusion results and effectively overcoming the problem of fusion quality degradation caused by geometric deformation and spectral variation in complex scenes.

[0007] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0008] A hyperspectral fusion imaging method based on topological sensing and differential homeomorphism includes:

[0009] 1) Input a low-resolution hyperspectral image (LR-HSI) and a high-resolution multispectral image (HR-MSI), and perform spatial and spectral dimension alignment preprocessing.

[0010] 2) Through the spatial-spectral degradation-reconstruction prior module, the spatial degradation process is modeled using differential homeomorphism transformation, and forward and backward operations are performed through the dual learning framework to establish cyclic consistency constraints for spatial and spectral data.

[0011] 3) Using the Topology Aware Transformer (TPFomer) module, shallow features of LR-HSI and HR-MSI are extracted, and spectral and spatial domain graph structures are constructed. Fine-grained feature representations are captured by combining graph convolution and attention mechanisms.

[0012] 4) Through the spectral-spatial collaborative fusion module, features from TPFomer are integrated, and feature enhancement and fusion are performed using spectral-driven attention fusion, discrete cosine transform and channel-spatial attention to generate a high-resolution hyperspectral image (HR-HSI).

[0013] Optionally, the detailed steps in step 2) include:

[0014] 2.1) Spatial Degradation-Reconstruction Prior: The spatial degradation process of LR-HSI is simulated using a Gaussian blur kernel and downsampling, and reconstruction is performed using upsampling and an anti-blur kernel, forming forward and backward operations. Its mathematical expression is:

[0015]

[0016] In the above formula, The value of the Gaussian convolution kernel at position (x, y), where Z is the normalization constant, σ is the standard deviation of the Gaussian distribution, and k is the kernel size. It is an exponential function.

[0017] 2.2) Spectral Degradation-Reconstruction Prior: The HR-HSI is projected into the multispectral space using a learnable or predefined spectral response kernel, and reconstruction is performed using a pseudo-inverse matrix to ensure spectral fidelity. The mathematical expression is:

[0018]

[0019] In the above formula, For spectral transformation operations, For the spectral response kernel tensor, Here, X is the actual transformation matrix used, X is the input image tensor, and inverse indicates whether forward spectral degradation or inverse spectral reconstruction is performed. Channel index of the output data.

[0020] Optionally, the detailed steps in step 3) include:

[0021] 3.1) Graph Structure Construction: Treating the feature maps of LR-HSI and HR-MSI as graph nodes, a spectral domain K-nearest neighbor graph and a spatial domain neighborhood graph are constructed based on feature similarity. The mathematical expression for this is:

[0022]

[0023] SpatialEdges

[0024] In the above formula, It is a constant. This represents the set of edge indices corresponding to the spectral domain graph. The k nearest neighbors of each pixel are determined based on feature similarity, thereby establishing the edge topology of the graph. The SpatialEdges(.) function is used to construct edges based on spatial adjacency relationships. F is a node feature matrix, H represents the spatial height of the feature graph, and W represents the spatial width of the feature graph.

[0025] 3.1) Graph convolution feature extraction uses Chebyshev polynomial-enhanced graph representation and integrates graph attention networks, graph convolutional networks, and super graph attention networks for multi-branch feature extraction. Its mathematical expression is:

[0026]

[0027] In the above formula, For the diffusion adjacency matrix, The coefficients of the k-th order Chebyshev polynomial are... For the k-th order Chebyshev polynomial, This is the normalized graph Laplace matrix.

[0028] 3.3) Manifold Attention Unit: A bidirectional attention mechanism is designed, combining forward and backward attention to enhance the discriminative power of feature representations. Its mathematical expression is:

[0029]

[0030] In the above formula, The final attention output is represented by α, which is a learnable attention parameter. Forward attention function, The fused graph features These are the original node features. This is the backward attention function.

[0031] Optionally, the detailed steps in step 4) include:

[0032] 4.1) Spectrum-driven attention fusion: Features of HR-MSI and LR-HSI are fused through a cross-modal attention mechanism, the expression of which is:

[0033]

[0034] In the above formula, Features resulting from enhanced attention For spectral multi-head attention, The features of the normalized and permuted multispectral image (MSI) are... Normalized and permuted features of the hyperspectral image (HSI);

[0035] 4.2) Discrete Cosine Transform: Enhances the high-frequency components of features, improving detail and edge clarity.

[0036] 4.3) Channel-Spatial Attention: This method optimizes feature representation through an attention mechanism, emphasizing important channels and spatial locations. Its expression is:

[0037]

[0038] In the above formula, The output of the low-rank attention function. Here, represents the convolution kernel weights, and Att represents the attention function. The input feature tensor.

[0039] 4.4) The overall loss of the construction is as follows:

[0040]

[0041] In the above formula, To rebuild the losses, and For dual loss, For cycle consistency loss, For Jacobi regularization loss, These are learnable weights.

[0042] Compared with existing technologies, this invention has the following advantages: First, by introducing differential homeomorphism to construct spatial degradation priors, this invention effectively models the geometric deformation constraints in hyperspectral image fusion, significantly improving the spatial structural consistency and geometric fidelity of the fusion results, overcoming the problem of spatial misalignment and structural distortion easily generated by traditional methods in complex terrain scenes. Second, this invention employs a dual learning mechanism to jointly optimize high-resolution generation and bidirectional low-resolution reconstruction tasks. Through cyclic consistency constraints of space and spectrum, prior knowledge of the physical imaging process is embedded into network learning, enhancing the physical credibility of the model and the realism of the reconstruction results. Third, this invention designs a topology-aware Transformer module, which explicitly models the spatial-spectral dual-domain manifold structure of hyperspectral data by constructing a spectral domain K-nearest neighbor graph and a spatial domain adjacency graph. This effectively captures fine-grained feature dependencies, significantly improving the ability to recover spectral details. Fourth, this invention integrates multiple mechanisms such as spectral-driven attention, discrete cosine transform, and channel-spatial attention through a spectral-spatial collaborative fusion module, achieving effective interaction and complementarity of cross-modal features, fully utilizing the correlation between different modalities, and improving the model's feature fusion capability. Comparative experimental results of this invention with several advanced methods show that this method can effectively recover spatial details while maintaining spectral fidelity. It exhibits strong robustness and practicality in scenarios with complex geometric deformation and spectral variation, providing reliable technical support for practical applications such as remote sensing monitoring, environmental perception, and urban planning. Attached Figure Description

[0043] Figure 1 This is a schematic diagram of the basic process of the method in an embodiment of the present invention.

[0044] Figure 2This is a schematic diagram illustrating the working principle of the spatial-spectral degradation-reconstruction prior module in an embodiment of the present invention.

[0045] Figure 3 This is a schematic diagram of the topology-aware Transformer module in an embodiment of the present invention.

[0046] Figure 4 This is a schematic diagram of the composition of the spectral-spatial collaborative fusion module in an embodiment of the present invention.

[0047] Figure 5 This is a schematic diagram of the diagram structure construction in an embodiment of the present invention. Detailed Implementation

[0048] Figure 1 and Figure 2 As shown, a hyperspectral fusion imaging method based on topological sensing and differential homeomorphism mechanism is characterized by comprising:

[0049] 1) Input low-resolution hyperspectral image LR-HSI and high-resolution multispectral image HR-MSI, and perform spatial and spectral dimension alignment preprocessing.

[0050] 2) Through the spatial-spectral degradation-reconstruction prior module, the spatial degradation process is modeled using differential homeomorphism transformation, and forward and backward operations are performed through the dual learning framework to establish cyclic consistency constraints for spatial and spectral data.

[0051] 3) Through the topology-aware Transformer module, shallow features of LR-HSI and HR-MSI are extracted, and spectral and spatial domain graph structures are constructed. Fine-grained feature representations are captured by combining graph convolution and attention mechanisms.

[0052] 4) Through the spectral-spatial collaborative fusion module, features from TPFomer are integrated, and feature enhancement and fusion are performed using spectral-driven attention fusion, discrete cosine transform and channel-spatial attention to generate high-resolution hyperspectral images.

[0053] See Figure 2 It can be seen that step 1) is the data preprocessing process; step 2) is the degradation-reconstruction prior modeling, which uses differential homeomorphism transformation and dual learning framework to establish spatial-spectral cyclic consistency constraints; step 3) is topology-aware feature extraction, which constructs spectral and spatial graph structures through TPFomer, and captures fine-grained features by combining graph convolution and attention; step 4) is collaborative fusion and reconstruction, which integrates features and generates high-resolution hyperspectral images through spectral-driven attention, discrete cosine transform and dual attention mechanisms.

[0054] like Figure 2As shown, step 1) involves inputting a low-resolution hyperspectral image (LR-HSI) and a high-resolution multispectral image (HR-MSI), and performing spatial and spectral dimension alignment preprocessing; in this embodiment, the detailed steps in step 2) include:

[0055] like Figure 2 As shown, step 2) utilizes the spatial-spectral degradation-reconstruction prior module to model the spatial degradation process using differential homeomorphism, and performs forward and backward operations through a dual learning framework to establish cyclic consistency constraints for spatial and spectral data. In this embodiment, the detailed steps in step 2) include:

[0056] 2.1) Spatial Degradation-Reconstruction Prior: The input image is convolved using a Gaussian blur kernel to simulate the point spread function of an optical system. Its mathematical expression is:

[0057]

[0058] In the above formula, The value of the Gaussian convolution kernel at position (x, y), where Z is the normalization constant, σ is the standard deviation of the Gaussian distribution, and k is the kernel size. It is an exponential function. A downsampling operation is then performed to complete the forward process. The reverse process reconstructs the image through upsampling and deconvolution operations.

[0059] 2.2) Spectral Degradation-Reconstruction Prior: Hyperspectral data is projected onto the multispectral space using a learnable spectral response matrix R, the mathematical expression of which is:

[0060]

[0061] In the above formula, For spectral transformation operations, For the spectral response kernel tensor, Here, X is the actual transformation matrix used, X is the input image tensor, and inverse indicates whether forward spectral degradation or inverse spectral reconstruction is performed. Channel index of the output data.

[0062] like Figure 3 As shown, step 3) addresses the limitations of existing fusion methods in modeling geometric structures and spectral neighborhood relationships by designing a topology-aware Transformer module. This module explicitly captures the intrinsic manifold geometry of hyperspectral data in both the spatial and spectral domains and achieves perceptual modeling of neighborhood dependencies through a topology-guided attention mechanism. The extracted shallow features are then injected into the TPFomer module. In this embodiment, the detailed steps in step 3) include:

[0063] 3.1) Treating the feature maps of LR-HSI and HR-MSI as graph nodes, a spectral domain K-nearest neighbor graph and a spatial domain 4-neighbor graph are constructed based on feature similarity. In the spectral domain graph, each node is connected to its K most similar nodes in its feature space; in the spatial domain graph, each node is connected to its four spatial neighbors (up, down, left, and right), thus constructing the spectral and spatial graph structures.

[0064]

[0065] SpatialEdges

[0066] In the above formula, It is a constant. This represents the set of edge indices corresponding to the spectral domain graph. The k nearest neighbors of each pixel are determined based on feature similarity, thereby establishing the edge topology of the graph. SpatialEdges represents a 4-neighborhood graph. It is a function that constructs edges based on spatial adjacency relationships, where F is a node feature matrix, H represents the spatial height of the feature graph, and W represents the spatial width of the feature graph;

[0067] 3.2) To effectively capture long-range dependencies, Chebyshev polynomial expansion is used to enhance graph representation. To expand the receptive field, a three-branch graph convolutional architecture integrating graph attention networks, graph convolutional networks, and super graph attention networks is used for feature extraction. A bidirectional attention mechanism is designed, combining forward and backward attention to enhance the discriminative power of feature representation. Chebyshev polynomials are used for graph representation enhancement, the mathematical expression of which is:

[0068]

[0069] In the above formula, Let k be the diffusion adjacency matrix, and k be the highest order of the Chebyshev polynomial expansion. The coefficients of the k-th order Chebyshev polynomial are... For the k-th order Chebyshev polynomial, This is the normalized graph Laplace matrix.

[0070] 3.3) Manifold Attention Unit: A bidirectional attention mechanism is designed, combining forward and backward attention. Its mathematical expression is:

[0071]

[0072] In the above formula, The final attention output is represented by α, which is a learnable attention parameter. Forward attention function, The fused graph features These are the original node features. This is the backward attention function.

[0073] like Figure 4 As shown, step 4) introduces a low-rank feature fusion module to enhance channel and spatial representation capabilities; in this embodiment, the detailed steps in step 4) include:

[0074] 4.1) Spectrum-driven attention fusion: Features of HR-MSI and LR-HSI are fused through a cross-modal attention mechanism. The mathematical expression is as follows:

[0075]

[0076] In the above formula, Features resulting from enhanced attention For spectral multi-head attention, The normalized and permuted MSI features, Normalized and permuted HSI features.

[0077] 4.2) Discrete Cosine Transform: Performs frequency domain transformation on features to enhance high-frequency components and improve detail and edge clarity.

[0078] 4.3) Channel-Spatial Attention: This optimizes feature representation through an attention mechanism, emphasizing important channels and spatial locations. The mathematical expression is:

[0079]

[0080] In the above formula, The output of the low-rank attention function. Here, represents the convolution kernel weights, and Att represents the attention function. The input feature tensor.

[0081] 4.4) A high-resolution hyperspectral image is generated by fusing low-resolution hyperspectral images with high-resolution multispectral images. Reconstruction loss is used as one of the main components. The loss function is constructed, and its expression is:

[0082]

[0083] In the above formula, This represents a true high-resolution hyperspectral image, while This represents the reconstructed high-resolution hyperspectral image. This represents the L1 norm, used to calculate the absolute error between the reconstructed image and the original output.

[0084] By leveraging the propagation properties of differential homeomorphisms and introducing cyclic consistency constraints for joint optimization, the physical interpretability and structural fidelity of the reconstruction results are improved. HSI branch dual loss is employed. The mathematical formula is expressed as follows:

[0085]

[0086] In the above formula, This represents a low-resolution hyperspectral image. Represents the downsampling Dual loss of the MSI branch Specifically defined as:

[0087]

[0088] In the above formula, Represents high-resolution multispectral images. This represents an image generated through a multispectral degradation process.

[0089] 4.5) To improve the spatial consistency and structural fidelity of the reconstruction results, a pullback operation is applied simultaneously in both the spatial and spectral domains to construct a cyclic consistency loss. Cyclic consistency loss of the LR-HSI branch. The definition is as follows:

[0090]

[0091] In the above formula, This represents the reconstructed image. Similar to the LR-HSI branch, the HR-MSI branch exhibits cycle consistency loss. Given by the following formula:

[0092]

[0093] In the above formula, This represents the reconstructed image. The overall cycle consistency loss is calculated as follows:

[0094]

[0095] 4.6) To improve the stability of the model during the degradation process and suppress the influence of anomalous perturbations, Jacobi regularization is introduced. As a constraint term, it is defined as follows:

[0096]

[0097] In the above formula, Network output to input Jacobian matrix, It is the Frobenius norm.

[0098] Finally, by integrating all the aforementioned loss terms, the overall loss function is constructed as follows:

[0099]

[0100] In the above formula, To rebuild the losses, and For dual loss, For cycle consistency loss, For Jacobi regularization loss, These are learnable weights.

[0101] To verify the effectiveness of this invention, experimental verification was conducted on the CAVE public dataset. The CAVE dataset contains high-resolution multispectral images covering a variety of natural and synthetic materials, encompassing 31 spectral bands within the 400-700 nm wavelength range, with a spatial resolution of 512×512 pixels for each scene. Four quantitative evaluation metrics were used to assess the fusion performance: Peak Signal-to-Noise Ratio (PSNR), Spectral Angle Mapper (SAM), Root Mean Square Error (RMSE), and Dimensionless Global Error (ERGAS). As a comparison with the method in this embodiment: the comparison method is SSRNet (see Zhang, X., Huang, W., Wang, Q., Li, X.: Ssr-net: Spatial–spectral reconstruction network for hyperspectral and multispectral image fusion. IEEE Transactions on Geoscience and Remote Sensing 59(7), 5953–5965(2021)); the comparison method is PSRT (see Deng, S.-Q., Deng, L.-J., Wu, X., Ran, R., Hong, D., Vivone, G.: Psrt: Pyramid shuffle-and-reshuffle transformer for multispectral and hyperspectral image fusion. IEEE Transactions on Geoscience and Remote Sensing 61, 1–15 (2023)); the comparison method is DSPNet (see Sun, Y., Xu, H., Ma, Y., Wu, M, Mei, X., Huang, J., Ma, J.: Dual spa-tial-spectral pyramid network with transformer for hyperspectral image fusion.IEEE Transactions on Geoscience and Remote Sensing 61, 1-16 (2023)); Comparison method MIMO (see Fang, J., Yang, J., Khader, A., Xiao, L.: Mimo-sst: Multi-input multi-output spatial-spectral transformer for hyperspectral and multispectral image fusion. IEEE Transactions on Geoscience and Remote Sensing 62, 1-20 (2024)); Comparison method LRTN (see Liu, Y., Dian, R., Li, S.: Low-rank transformer for high-resolution hyperspectral computational imaging. International Journal of Computer Vision 133(2), 809–824 (2025)); Comparison method SMGU-Net (see Yan, J., Zhang, K., Sun, Q., Ge, C., Wan, W., Sun, J., Zhang, H.: Spatial-spectral unfolding network). With mutual guidance for multispectral and hyperspectral image fusion. PatternRecognition 161, 111277(2025)); contrasting method CSGAV (see Chi, B., Lu, H., Liu, R., Yang, Y., Xu, L., Wan, W.: Multispectral-hyperspectral image fusion via similarity-guided graph attention and vae-transformer. IEEE Transactions on Geoscience and Remote Sensing 63, 1–16 (2025)); contrasting method OTIAS (see Deng, S., Ma, J., Deng, L.-J., Wei, P.).Otias: Octree implicit adaptive sampling for multispectral and hyperspectral image fusion. In: Proceedings of the AAAIConference on Artificial Intelligence, vol. 39, pp.2708-2716(2025)). Table 1 shows the quantitative comparison results between the method of this invention and eight advanced algorithms.

[0102] Table 1 compares the experimental results of the method in this embodiment with those of eight advanced algorithms on the CAVE dataset.

[0103] method PSNR SAM ERGAS RMSE SSRNet 43.6021 4.4768 0.9048 2.0789 PSRT 45.9228 3.1022 0.7010 1.6614 DSPNet 47.6970 2.4971 0.5800 1.3282 MIMO 48.2850 2.5157 0.5477 1.2588 LRTN 48.1070 2.3358 0.5517 1.2524 SMGU-Net 46.2577 2.8722 0.7831 1.7826 CSGAV 46.8428 2.5950 0.6600 1.4939 OTIAS 48.0643 2.3984 0.5555 1.2919 Method of this embodiment 48.6787 2.2026 0.5146 1.1674

[0104] As shown in Table 1, the method of the present invention outperforms the comparative method in all four evaluation indicators: PSNR, SAM, ERGAS, and RMSE, proving that it can effectively recover spatial details while maintaining spectral fidelity.

[0105] In summary, this invention constructs a spatial degradation prior by introducing differential homeomorphism, achieves image reconstruction under physical constraints using a dual learning mechanism, and explicitly models the spectral-spatial dual-domain manifold structure using a topology-aware Transformer module. Finally, high-fidelity, high-resolution hyperspectral images are generated through spectral-spatial collaborative fusion. Extensive experimental verification demonstrates that this method exhibits excellent performance under complex geometric deformation and spectral variation scenarios, providing reliable technical support for practical applications such as remote sensing monitoring and environmental perception.

[0106] Meanwhile, this embodiment provides a computer-readable storage medium storing a computer program or instructions. When the computer program or instructions are executed by a processor, the hyperspectral fusion imaging method based on topological sensing and differential homeomorphism mechanism as described above is implemented.

[0107] Furthermore, this embodiment also provides a computer program product comprising a computer program or instructions which, when executed by a processor, are used to implement the hyperspectral fusion imaging method based on topological sensing and differential homeomorphism mechanism.

[0108] Those skilled in the art will understand that embodiments of the present invention can be provided in the form of a method, system, or computer program product. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can also be implemented as a computer program product comprising computer-usable program code contained on one or more computer-readable storage media, including but not limited to disk storage, CD-ROM, optical storage, etc.

[0109] This invention is described with reference to flowchart illustrations and / or structural diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It should be understood that each flow, block, and combination thereof in the flowchart illustrations and / or structural diagrams can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, or other programmable data processing device to form machine instructions that, when executed by a computer or other programmable device, perform the functions defined by one or more flows and / or blocks in the flowchart illustrations and / or structural diagrams.

[0110] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium form an article of manufacture containing instruction means that perform the functions specified by one or more flowcharts and / or blocks in a flowchart and / or structural diagram.

[0111] In addition, these computer program instructions may be loaded onto a computer or other programmable device to cause a series of operational steps to be performed on the computer or other programmable device, producing a computer-implemented process, such that the instructions, which execute on the computer or other programmable device, provide steps to implement the function described in one or more processes or blocks in the flowchart and / or structural diagram.

[0112] The above description is merely a preferred embodiment of the present invention, and the scope of protection of the present invention is not limited to the above embodiments. Any technical solution proposed under the design concept of the present invention falls within the protection scope of the present invention. It should be noted that for those skilled in the art, any improvements, equivalent substitutions, or modifications made without departing from the basic principles of the present invention should be considered as included within the protection scope of the present invention.

Claims

1. A hyperspectral fusion imaging method based on topological sensing and differential homeomorphism, characterized in that, include: 1) Perform spatial and spectral dimension alignment preprocessing on the input low-resolution hyperspectral image (LR-HSI) and high-resolution multispectral image (HR-MSI); 2) Model the spatial degradation process using differential homeomorphism transformation, and perform forward and backward operations through a dual learning framework to establish cyclic consistency constraints for space and spectrum; 3) Extract shallow features of LR-HSI and HR-MSI through Topology Aware Transformer (TPFomer), and construct spectral and spatial domain graph structures, and capture fine-grained feature representations by combining graph convolution and attention mechanisms; 4) The spectral-spatial collaborative fusion module integrates features from the topology-aware Transformer and performs multi-level feature enhancement and fusion through spectral-driven attention fusion, discrete cosine transform and channel-spatial attention mechanism to finally generate a high-resolution hyperspectral image (HR-HSI).

2. The hyperspectral fusion imaging method according to claim 1, characterized in that, The preprocessing described in step 1) includes: upsampling the LR-HSI to align it spatially with the HR-MSI, and extracting shallow features of the LR-HSI and HR-MSI using convolutional layers.

3. The hyperspectral fusion imaging method according to claim 1, characterized in that, Step 2) includes the following detailed steps: 1) Spatial-spectral degradation-reconstruction priors: The spatial degradation process is modeled using differential homeomorphism and forward and backward operations are performed through a dual learning framework to establish cyclic consistency constraints for spatial and spectral data. 2) Order Represents the image domain, if the mapping A function is called a differential homeomorphism if it meets the following conditions. It is a one-to-one mapping and is surjective. And its inverse mapping Where C¹ denotes the first-order continuous differentiability of the function and Ω' denotes the image region after the transformation by mapping φ, and the Jacobian determinant is positive definite at position. Place The determinant of the Jacobian matrix satisfies 0; 3) Under this mapping, the image Transformation is achieved through function composition. In image modeling, differential homeomorphism is widely used to construct invertible spatial transformations. 4) Although It exhibits nonlinear coupling characteristics overall, but its generation process is affected by the local velocity field. This regulation makes the deformation process clearly physically explainable; 5) Due to It has reversibility, and its inverse mapping It exists and is itself a differential homeomorphism; 6) Spatial Degradation-Reconstruction Prior: The spatial degradation process of LR-HSI is simulated using Gaussian blur kernel and downsampling, and reconstruction is performed by upsampling and anti-blur kernel to form forward and backward operations; 7) Spectral Degradation-Reconstruction Prior: HR-HSI is projected into the multispectral space using a learnable or predefined spectral response kernel, and spectral fidelity is ensured by using a pseudo-inverse matrix for reconstruction.

4. The hyperspectral fusion imaging method according to claim 1, characterized in that, Step 3) includes the following detailed steps: 1) Assume that in the hyperspectral and multispectral image fusion task, the ground truth of the target high-resolution hyperspectral image can be interpreted as being jointly generated by the low-resolution hyperspectral image through a spatial degradation process and the high-resolution multispectral image through a spectral selective response; 2) Observed low-resolution hyperspectral images and observed high-resolution multispectral images Spatial alignment is performed, and shallow features are extracted through a cross-modal fusion module to promote cross-modal consistency. 3) TPFomer is used to explicitly capture the intrinsic manifold geometry of hyperspectral data in the spatial and spectral domains and realize perceptual modeling of neighborhood dependencies through a topology-guided attention mechanism. Shallow features are then extracted and injected into the TPFomer module. 4) The TPFomer module consists of a learnable spectral convolution operator and a manifold attention unit, which performs graph construction mapping on the LR-HSI branch; 5) Simultaneously construct the spectral domain and spatial graph structure: ， SpatialEdges , In the above formula, It is a constant. This represents the set of edge indices corresponding to the spectral domain graph. The k nearest neighbors of each pixel are determined based on feature similarity, thereby establishing the edge topology of the graph. The SpatialEdges(.) function is used to construct edges based on spatial adjacency relationships. F is a node feature matrix, H represents the spatial height of the feature graph, and W represents the spatial width of the feature graph. To effectively capture long-range dependencies, Chebyshev multinomial expansion is used to enhance graph representations. To expand the receptive field, a three-branch graph convolutional architecture integrating graph attention networks, graph convolutional networks, and super graph attention networks is employed for feature extraction. A bidirectional attention mechanism is designed, combining forward and backward attention to enhance the discriminative power of feature representations.

5. The hyperspectral fusion imaging method according to claim 4, characterized in that, The specific implementation of the manifold attention unit in step 3) includes: designing a bidirectional attention mechanism that combines forward and backward attention, the mathematical expression of which is: ， In the above formula, The final attention output is represented by α, which is a learnable attention parameter. Forward attention function, The merged graph features These are the original node features. This is the backward attention function.

6. The hyperspectral fusion imaging method according to claim 1, characterized in that, Step 4) Spectrum-driven attention fusion: Fuse HR-MSI and LR-HSI features through a cross-modal attention mechanism; Step 4) Discrete cosine transform: Enhance the high-frequency components of the features to improve detail and edge clarity; Step 4) Channel-spatial attention: Optimize feature representation through an attention mechanism to emphasize important channels and spatial locations.

7. A hyperspectral fusion imaging method based on topological sensing and differential homeomorphism, comprising a microprocessor and a memory, wherein the microprocessor and the memory are communicatively connected; characterized in that, The microprocessor is configured to invoke and execute a computer program stored in the memory to implement the hyperspectral fusion imaging method based on topological sensing and differential homeomorphism mechanism as described in any one of claims 1 to 6.

8. A computer-readable storage medium having a computer program or instructions stored thereon, characterized in that, When the computer program or instructions are executed by the processor, they implement the hyperspectral fusion imaging method based on topological sensing and differential homeomorphism mechanism as described in any one of claims 1 to 6.

9. A computer program product, comprising a computer program or instructions, characterized in that, When the computer program or instructions are executed by the processor, they implement the hyperspectral fusion imaging method based on topological sensing and differential homeomorphism mechanism as described in any one of claims 1 to 6.

Images