A multi-scale gradient-guided snapshot compressive imaging reconstruction method and system

Abstract

A snapshot compressed imaging reconstruction method and system, belonging to the field of image processing technology, solves the problems of insufficient intra-stage interaction and inter-stage information attenuation in existing technologies. The method includes: acquiring a two-dimensional compressed measurement image; performing initial inversion to obtain an initial reconstruction result; calculating a gradient feature map based on the previous stage reconstruction result and mask parameters, performing gradient descent updates on the previous stage reconstruction result to obtain the intermediate reconstruction result of the current stage; performing multi-scale processing on the gradient feature map to extract gradient features at different scales and generate weight information; reconstructing the intermediate reconstruction result of the current stage, using the weight information to modulate the feature interactions in the near-end mapping process to generate the reconstruction result of the current stage; passing the reconstruction result of the current stage to the next stage, establishing inter-stage message passing; S7: repeating the iteration until a preset number of stages is completed, and outputting the final reconstruction result. This invention is applicable to snapshot compressed imaging scenarios.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology, specifically relating to a hyperspectral imaging reconstruction technique. Background Technology

[0002] Hyperspectral imaging technology can acquire spectral information of a scene across multiple consecutive bands, making it valuable for applications in remote sensing, medical imaging, and industrial inspection. However, traditional hyperspectral imaging systems typically rely on band-by-band scanning or mechanical scanning, which suffers from slow acquisition speeds, system complexity, and difficulty in adapting to dynamic scenes. To address these issues, Snapshot Compressive Imaging (SCI) technology uses optical encoding to compress three-dimensional hyperspectral data into two-dimensional measurement images, and then reconstructs the original hyperspectral image through computational reconstruction in the backend, significantly improving imaging efficiency. Existing SCI systems typically include an optical modulation unit, a detection unit, and a computational reconstruction unit, where the computational reconstruction algorithm directly determines the final image quality. Currently, SCI reconstruction algorithms mainly include model-based reconstruction methods and data-driven reconstruction methods. In recent years, reconstruction methods based on iterative optimization have gained increasing attention. These methods expand the optimization process into a multi-stage reconstruction workflow, balancing reconstruction performance and computational efficiency to a certain extent.

[0003] Although existing unfolding reconstruction methods have made some progress in snapshot compressed imaging, they still have the following shortcomings: (1) Insufficient information interaction within the stage: Existing methods typically update based only on reconstruction results or simple features within a single reconstruction stage, failing to fully utilize the gradient information generated by measurement consistency constraints, making it difficult to effectively characterize the detailed structure and spatial-spectral correlation of the scene. (2) Limited cross-stage information transmission capability: Most methods only transmit reconstruction results between different reconstruction stages, lack modeling of multi-stage feature evolution process, making it difficult to capture long-range dependencies between stages, which can easily lead to the gradual decay of effective information during iteration. (3) It is difficult to balance reconstruction quality and stability: In scenes with complex textures or strong noise interference, existing methods often cannot maintain spatial details and spectral consistency at the same time, which can easily produce artifacts or spectral distortion, thus affecting the stability and reliability of imaging results. Summary of the Invention

[0004] This invention provides a snapshot compressed imaging reconstruction method and system based on multi-scale gradient guidance and cross-stage message passing, which aims to solve the problems of insufficient intra-stage interaction and inter-stage information attenuation in the prior art.

[0005] Firstly, the present invention aims to provide a multi-scale gradient-guided snapshot compression imaging reconstruction method, comprising the following steps: S1: Modulate the hyperspectral scene to obtain a two-dimensional compressed measurement image; S2: Using an imaging model, the two-dimensional compressed measurement image is initially inverted to obtain the initial reconstruction result; S3: Based on the reconstruction results and mask parameters of the previous stage, calculate the gradient feature map, and perform gradient descent update on the reconstruction results of the previous stage based on the gradient feature map to obtain the intermediate reconstruction results of the current stage. S4: Perform multi-scale processing on the gradient feature map, extract gradient features at different scales, and generate weight information for modulating the near-end mapping reconstruction process; S5: Reconstruct the intermediate reconstruction results of the current stage in the near-end mapping network, and use the weight information to modulate the feature interactions in the near-end mapping process to generate the reconstruction results of the current stage. S6: Pass the reconstruction results of the current stage to the next stage, and establish message passing between multiple stages; S7: S3 to S6 are executed iteratively as a repeated reconstruction stage unit until the preset number of stages is completed, and the final reconstruction result is output.

[0006] Furthermore, a preferred embodiment is provided: the gradient feature map calculation step in S3 includes: Based on the reconstruction results and coding parameters of the previous stage, predictive measurements are generated, and measurement residuals are constructed with the two-dimensional compressed measurement images. The gradient feature map is obtained by back-projecting the measurement residual.

[0007] Furthermore, a preferred embodiment is provided: S4 includes: Using the gradient feature map as input, multi-scale gradient features are extracted through multiple parallel convolution operations with different receptive fields; The multi-scale gradient features are subjected to global average pooling and nonlinear activation to generate multi-scale gradient modulation weights for the near-end mapping reconstruction process.

[0008] Furthermore, a preferred embodiment is provided: the gradient descent update in S3 includes: Using the previous stage reconstruction result and gradient feature map as input, a lightweight neural network with residual connections is used to dynamically generate spatially adaptive step size parameters, and the previous stage reconstruction result is updated by gradient descent based on the step size parameters to obtain the intermediate reconstruction result of the current stage.

[0009] Furthermore, a preferred embodiment is provided: S6 includes: The reconstruction results of the current stage are used as the initial input for the next stage, and a cross-stage message passing mechanism is introduced in the near-end mapping network. The cross-stage message passing mechanism encodes and decodes multi-scale features through a hierarchical encoder-decoder structure.

[0010] Furthermore, a preferred embodiment is provided: embedding a state-space model-based network structure into the encoder-decoder structure to propagate cross-stage multi-scale long-range dependency information between adjacent iteration stages, thereby enhancing inter-stage information interaction.

[0011] Secondly, the object of this invention is to propose a multi-scale gradient-guided snapshot compression imaging reconstruction system, said system being implemented based on a multi-scale gradient-guided snapshot compression imaging reconstruction method as described in any one or more of the above-mentioned schemes, said system comprising: Image acquisition module: used to modulate hyperspectral scenes and acquire two-dimensional compressed measurement images; Initial Inversion Module: Used to perform initial inversion on the two-dimensional compressed measurement image using an imaging model to obtain initial reconstruction results; Gradient calculation module: used to calculate gradient feature map based on the reconstruction result of the previous stage and mask parameters, and perform gradient descent update on the reconstruction result of the previous stage based on the gradient feature map to obtain the intermediate reconstruction result of the current stage; Multi-scale gradient guidance module: used to perform multi-scale processing on the gradient feature map, extract gradient features at different scales, and generate weight information for modulating the near-end mapping reconstruction process; Reconstruction and Update Module: This module is used to reconstruct the intermediate reconstruction results of the current stage in the near-end mapping network, and to modulate the feature interactions in the near-end mapping process using the weight information to generate the reconstruction results of the current stage. Cross-stage transfer module: used to transfer the reconstruction results of the current stage to the next stage and establish message transfer between multiple stages; Iterative output module: used to iteratively execute the gradient calculation module to the cross-stage transfer module as a repeated reconstruction stage unit until the preset number of stages is completed, and output the final reconstruction result.

[0012] Thirdly, the present invention aims to provide a computer device, the computer device including a memory and a processor, the memory storing a computer program, and when the processor runs the computer program stored in the memory, the processor executes a multi-scale gradient-guided snapshot compression imaging reconstruction method according to any one or more of the above-described schemes.

[0013] Fourthly, the present invention aims to provide a computer-readable storage medium for storing a computer program that executes a multi-scale gradient-guided snapshot compression imaging reconstruction method as described in any one or more of the above-described schemes.

[0014] Compared with the prior art, the advantages of the present invention are: (1) Enhance information interaction during the phase.

[0015] This invention does not merely utilize gradients for conventional descent updates, but further constructs gradient features into learnable multi-scale gradient guidance information. This information is used simultaneously within the same reconstruction stage for: ① generating adaptive update intensity (dynamic step size / update mapping); and ② generating multi-scale modulation weights to guide feature interactions in subsequent reconstruction mappings. This achieves closed-loop information interaction between "update" and "reconstruction," resulting in stronger recovery capabilities for texture edges and fine-grained structures.

[0016] (2) Multi-scale gradient guidance performs "content-adaptive modulation" on key feature interactions, effectively suppressing artifacts and spectral crosstalk.

[0017] This invention utilizes multi-scale weights generated by a gradient-guided module to selectively enhance and suppress feature fusion / attention interaction in the reconstruction mapping: it strengthens effective information transmission in structurally significant regions and suppresses ineffective propagation in flat or noise-dominated regions, thereby reducing artifact diffusion caused by coding modulation and noise superposition, and improving the spectral consistency and detail fidelity of the reconstruction results.

[0018] (3) Cross-stage message passing introduces a multi-scale propagation mechanism with long-range dependence to alleviate information decay and error accumulation in multi-stage iteration.

[0019] To address the issues of existing multi-stage reconstruction methods that only transmit reconstruction results and have weak inter-stage correlations, this invention introduces a cross-stage information transmission path between multiple reconstruction stages. Through a hierarchical structure, it continuously propagates stage-specific features and state information at different scales, allowing effective structural information from earlier stages to be inherited and strengthened by subsequent stages. Specifically, the proximal mapping network recursively updates feature states based on a state-space modeling mechanism. This effectively models long-range dependencies without relying on explicit attention computation, thus better adapting to multi-stage iterative reconstruction processes and avoiding the phenomenon of "losing details with each iteration / accumulating and shifting," thereby improving the overall reconstruction stability and robustness.

[0020] (4) It takes into account both engineering deployability and universal adaptability, and is suitable for various types of snapshot compression imaging systems.

[0021] This invention takes imaging models and measurement data as inputs and adopts a modular "gradient guidance-reconstruction mapping-cross-stage transfer" framework, which can achieve stable reconstruction under different modulation methods, different noise levels and different snapshot compression imaging devices (such as filter array type, dispersive type, etc.). At the same time, it reduces the dependence on specific scenes through multi-scale and cross-stage mechanisms, and has good engineering application value.

[0022] This invention is applicable to snapshot compression imaging scenarios. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0024] Figure 1 This is a schematic diagram of the structure of a multi-scale gradient-guided snapshot compressed imaging reconstruction system according to a specific embodiment of the present invention; Figure 2 This is a schematic diagram of the internal structure of a single reconstruction stage and the connection relationship of its key sub-modules in the multi-stage reconstruction processing module described in a specific embodiment of the present invention. Figure 3 The figure shows the experimental results of the seventh specific embodiment of the present invention. Detailed Implementation

[0025] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application can also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods are omitted so as not to obscure the description of this application with unnecessary detail.

[0026] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0027] Many specific details are set forth in the following description in order to provide a full understanding of this application. However, this application may also be implemented in other ways different from those described herein. Those skilled in the art can make similar extensions without departing from the spirit of this application. Therefore, this application is not limited to the specific embodiments disclosed below.

[0028] Implementation Method 1 This embodiment proposes a snapshot compressed imaging reconstruction method. By introducing a gradient-guided mechanism and a cross-stage information transfer mechanism, it achieves high-quality reconstruction of hyperspectral images, overcoming the problems of insufficient intra-stage interaction and inter-stage information attenuation in existing technologies. Specifically, it includes the following steps: S1: Modulate the hyperspectral scene to obtain a two-dimensional compressed measurement image; S2: Using an imaging model, the two-dimensional compressed measurement image is initially inverted to obtain the initial reconstruction result; S3: Based on the reconstruction results and mask parameters of the previous stage, calculate the gradient feature map, and perform gradient descent update on the reconstruction results of the previous stage based on the gradient feature map to obtain the intermediate reconstruction results of the current stage. S4: Perform multi-scale processing on the gradient feature map, extract gradient features at different scales, and generate weight information for modulating the near-end mapping reconstruction process; S5: Reconstruct the intermediate reconstruction results of the current stage in the near-end mapping network, and use the weight information to modulate the feature interactions in the near-end mapping process to generate the reconstruction results of the current stage. S6: Pass the reconstruction results of the current stage to the next stage, and establish message passing between multiple stages; S7: S3 to S6 are performed iteratively as a repeatable reconstruction stage unit until the preset number of stages is completed, and the final reconstruction result is output.

[0029] The method described in this embodiment first performs an initial inversion of the two-dimensional compressed measurement signal based on the imaging model to obtain an initial reconstruction result; then, several cascaded optimization stages are executed sequentially, with each optimization stage having the same structure and its parameters being shared or set independently between different stages.

[0030] Each optimization stage includes a Mask-guided Descent Network (MDN) and a Proximal Mamba Network (PMN): the MDN module updates the reconstruction result of the previous stage by combining the mask parameters and gradient feature map to obtain the intermediate reconstruction result of the current stage; the PMN performs proximal mapping reconstruction on the intermediate reconstruction result of the current stage and outputs the reconstruction result of the current stage under the modulation of multi-scale gradient guidance information.

[0031] Meanwhile, an intra-stage information exchange mechanism is introduced within each optimization stage, and a cross-stage message passing mechanism is introduced between different optimization stages, so that gradient information, structural priors, and cross-stage features can flow and coordinate fully throughout the reconstruction process, thereby improving reconstruction quality and stability.

[0032] Implementation Method 2 This embodiment is a further illustrative example of a snapshot compression imaging reconstruction method described in Embodiment 1.

[0033] This implementation addresses the problems of underdetermined measurement height, unstable reconstruction, and easy loss of details in snapshot compressed hyperspectral imaging systems by modeling the compressed sensing reconstruction process as the following optimization problem: , in, Indicates a snapshot measurement signal. This represents the encoding and compression operators corresponding to the imaging system. For the hyperspectral data cube to be reconstructed, This represents the prior regularization term.

[0034] To efficiently solve the above optimization problem, this implementation follows the Proximal Gradient Descent (PGD) framework, and unfolds the overall solution process as follows: The process consists of sequentially executed iterative phases. Each phase comprises two steps: gradient descent update and proximal mapping update, thereby obtaining the intermediate reconstruction. With stage output Its update format is as follows: , in, For the first The dynamic step size parameter of the stage. For the first A staged parametric proximal mapping network is used to map intermediate reconstruction results after gradient updates. Further mapping to the current stage of reconstruction Therefore, this implementation achieves [something] at each stage. and Alternating iterative optimization, the final output As a result of the reconstruction.

[0035] In this embodiment, gradient descent updates are implemented using a Mask-Guided Descent Network (MDN) to generate... The near-end mapping update is implemented by the near-end Mamba network (PMN), used by... get And information is continued and enhanced through message passing mechanisms between each stage.

[0036] Implementation Method 3 This embodiment is a further illustrative example of the MDN described in the snapshot compressed imaging reconstruction method of Embodiment 1 or Embodiment 2.

[0037] To achieve adaptive modeling of the gradient of the data fidelity term during compressed sensing reconstruction and enhance intra-stage information interaction between gradient updates and subsequent near-end mappings, this implementation introduces a mask-guided dynamic gradient descent network in each iteration stage. This network utilizes the reconstruction results from the previous stage... Measurement signals and the encoding parameters of the imaging system As input, it outputs gradient feature maps and dynamic step size information to guide reconstruction updates. Its overall structure includes the following three functional sub-networks: 1. Dynamic gradient calculation subnetwork: This subnetwork incorporates imaging coding parameters explicitly. Related forward mapping networks and its corresponding reverse mapping network The measurement consistency constraints in the snapshot compressed imaging process are modeled to achieve adaptive calculation of gradient features. Both the forward mapping network and the backward mapping network are learnable neural network structures, specifically composed of multiple layers of convolutional operators, with residual connections introduced within the network to enhance feature representation and training stability. This network structure is used to perform data-driven approximate modeling of the imaging process and its backpropagation process.

[0038] Specifically, firstly, based on the results of the previous stage of reconstruction With encoding parameters Generate predictive measurements and compare them with actual measurement signals. Construct the measurement residuals, and use the inverse mapping network to map the measurement residuals back to the image domain to obtain the first... kThe gradient feature map of the stage is expressed as follows: .

[0039] Through the above design, the sub-network can achieve adaptive modeling of gradient information while preserving the consistency of physical imaging.

[0040] 2. Adaptive step-size generator subnetwork: To overcome the convergence instability problem caused by the fixed step size in traditional proximal gradient methods, this implementation further introduces an adaptive step-size generation subnetwork. This subnetwork reconstructs features and gradient features in the previous stage. As input, a spatially adaptive step size parameter is dynamically generated through a lightweight neural network structure with residual connections. This enables differentiated gradient update intensity control at different spatial locations.

[0041] Under the influence of the dynamic step size, the intermediate reconstruction results The update format is as follows:

[0042] in, This indicates element-wise multiplication.

[0043] 3. Multi-scale gradient-guided subnetwork: To enhance intra-stage synergy between gradient updates and proximal mapping learning, this implementation further introduces a multi-scale gradient-guided intra-stage message passing mechanism into the MDN.

[0044] This subnetwork uses gradient feature maps As input, gradient features at different scales are extracted through multiple parallel, separable convolutional operations with varying receptive fields. Subsequently, global average pooling is performed on the gradient features at each scale, followed by processing with a non-linear activation function to generate a set of multi-scale gradient modulation weights, the expression of which is: , in, This represents parallel convolution operations with different receptive fields. This indicates a global average pooling operation. This represents a non-linear activation function.

[0045] The multi-scale gradient modulation weight As guiding information within a stage, it is introduced into the proximal mapping network to adaptively modulate the feature fusion or attention interaction process, thereby effectively injecting the structural prior contained in the gradient information into the reconstruction mapping process within the current reconstruction stage, enhancing the ability to recover texture details and spatial structure.

[0046] The intermediate reconstruction results and multi-scale gradient guidance information generated by the mask-guided dynamic gradient descent network will be used as prior information for cross-stage message passing and further introduced into the subsequent near-end network to achieve information continuation and enhancement between multiple stages.

[0047] Implementation Method Four: This embodiment is a further illustrative example of the PMN described in the snapshot compression imaging reconstruction method described in Embodiment 1 or Embodiment 2.

[0048] To avoid the gradual decay of information during multi-stage unfolding and to enhance the modeling ability of long-range spectral dependence and cross-stage structural consistency of hyperspectral data, this implementation introduces a hierarchical encoder-decoder structure based on a selective state-space model in PMN to realize inter-stage message passing.

[0049] The core of PMN is the spatial-spectral Mamba block, which employs a dual-path structure: 1. Basic Feature Path: In the near-end mapping network, a basic feature path is set to extract stable low-order representations of the input features.

[0050] The basic feature path sequentially performs normalization, linear channel mapping, and nonlinear activation operations on the input features to obtain a basic feature representation, the calculation form of which is as follows: .

[0051] in, This indicates a normalization operation. Represents linear channel transformation. This represents a non-linear activation function.

[0052] This basic feature path can improve feature representation capabilities while ensuring feature stability.

[0053] 2. Enhance the interaction path: In the parallel branches of the basic feature path, an enhanced interaction path is set to improve the modeling capability for complex structural information.

[0054] The enhanced interaction path first performs linear projection and convolution enhancement on the input features, and then introduces a cross-attention mechanism to achieve interactive modeling between spatial features and spectral features. Its calculation form is as follows: .

[0055] in, This indicates a convolution enhancement operation. This represents the cross-attention operation.

[0056] The gradient-guided cross-attention mechanism is specifically as follows: 3. Gradient-guided cross-attention modulation mechanism: In cross-attention networks, features are reshaped as follows: Spatial view features and spectral view features And construct the corresponding query, key and value representation respectively.

[0057] Among them, multi-scale gradient modulation weights It is used to perform element-wise weighting of spatial keys and values ​​in cross-attention, and its calculation form is as follows: , .

[0058] in, and ...

[0059] Subsequently, a query vector is generated using spectral features. Perform cross-attention computation with the weighted spatial keys and values: , This enables the implementation of spatial-spectral feature interaction modeling guided by gradient priors within the current reconstruction phase.

[0060] Implementation Method 5: This embodiment is a further illustrative example of S7 in a snapshot compression imaging reconstruction method described in Embodiment 1 or Embodiment 2.

[0061] To simultaneously ensure the visual quality and physical consistency of the reconstruction results, this implementation adopts a composite training objective consisting of image domain reconstruction loss and measurement domain consistency loss. Its total loss function is defined as: , in, The reconstructed hyperspectral image output by the network. Represents the true hyperspectral image (ground truth). Indicates observation and measurement, These are encoding parameters.

[0062] This composite loss enhances image detail and spectral fidelity while constraining the reconstruction results to meet the imaging model, thereby improving reconstruction stability and reliability.

[0063] Implementation Method Six: This embodiment proposes a snapshot compression imaging reconstruction system, which is based on a snapshot compression imaging reconstruction method as described in Embodiments 1 to 5.

[0064] like Figure 1 The system includes: an optical encoding unit 1, a measurement acquisition module 2, an initial reconstruction module 3, a multi-stage reconstruction processing module 4, and a reconstruction result output module 5.

[0065] The optical coding unit 1 is used to optically modulate the hyperspectral scene and generate corresponding coding parameters; the measurement acquisition module 2 is connected to the optical coding unit 1 and is used to acquire two-dimensional compressed measurement signals under coding modulation conditions.

[0066] The initial reconstruction module 3 receives the two-dimensional compressed measurement signal output by the measurement acquisition module 2, performs a preliminary inversion on it according to the imaging model, and generates an initial reconstruction result.

[0067] The multi-stage reconstruction processing module 4 is connected to the initial reconstruction module 3 and is used to perform multi-stage iterative optimization processing on the initial reconstruction results. The multi-stage reconstruction processing module 4 includes multiple sequentially connected reconstruction stages and introduces a cross-stage information transmission mechanism between each reconstruction stage to enhance information interaction and message transmission between different stages.

[0068] The reconstruction result output module 5 is connected to the multi-stage reconstruction processing module 4 and is used to output the final hyperspectral reconstruction result.

[0069] like Figure 2 As shown, Figure 2 It shows Figure 1 The internal structure of a single reconstruction stage in the multi-stage reconstruction processing module 4 of the system shown, and the connection relationship of its key sub-modules, are illustrated. The single reconstruction stage includes a gradient descent module (MDN) 41 guided by mask parameters (MDN), a multi-scale gradient guidance module (MS-GGM) 43, and a near-end Mamba network module (PMN) 42.

[0070] The gradient descent module 41 guided by the mask parameters is used to generate intermediate reconstruction results. It internally includes a gradient calculation submodule 411, an adaptive step size generation submodule 412, and a gradient descent update unit. The gradient calculation submodule 411 receives the reconstruction result from the previous stage. Two-dimensional compressed measurement signal y and encoding parameters Under the constraints of the imaging model, predictive measurements are generated and measurement residuals are constructed. Gradient feature maps are then calculated through inverse mapping. The adaptive step size generation submodule 412 generates the gradient feature map based on the gradient feature map. Adaptive generation of update step size parameters The gradient descent update unit integrates the gradient feature map. With update step size parameter ,right Perform gradient descent updates and output intermediate reconstruction results. .

[0071] The multi-scale gradient guidance module 43 is connected to the gradient calculation submodule 411 and is used to process the gradient feature map. Perform multi-scale processing to generate multi-scale gradient modulation weights. The near-end Mamba network module 42 simultaneously receives the intermediate reconstruction results. and multi-scale gradient modulation weights Proximal mapping and feature reconstruction are performed on the intermediate reconstruction results, and the reconstruction result for the current stage is generated under multi-scale gradient-guided modulation. The reconstruction results at the current stage. This output is passed to the next reconstruction stage, thus achieving multi-stage iterative reconstruction.

[0072] Implementation Method Seven: To verify the effectiveness and stability of the multi-scale gradient-guided snapshot compression imaging reconstruction method proposed in this invention in practical applications, this embodiment conducted comparative experiments and performance evaluations of the method based on a typical snapshot compression hyperspectral imaging system.

[0073] 1. Experimental Setup In this embodiment, a snapshot compression imaging system is used to modulate the hyperspectral scene, compressing and mapping the original three-dimensional hyperspectral data into a two-dimensional measurement signal. The reconstruction method described in this invention is then used to perform hyperspectral reconstruction of the two-dimensional measurement signal. In the simulation experiment, publicly available standard hyperspectral simulation data is selected as the experimental object. This hyperspectral data contains 28 continuous spectral channels, covering the visible light band from 450 nm to 650 nm, which can effectively reflect the spatial structure information and spectral distribution characteristics of natural scenes. Specifically, multiple typical hyperspectral scenes from the CAVE and KAIST datasets are used as simulation experimental samples. During the specific experimental implementation, multiple hyperspectral images from the CAVE dataset are selected as training samples, and multiple typical hyperspectral scenes from the KAIST dataset are selected as test samples to evaluate the generalization performance of different reconstruction methods in unknown scenes. The hyperspectral images in the CAVE dataset are medium-resolution data with a spatial resolution of 512×512 pixels, while the hyperspectral images in the KAIST dataset are high-resolution data with a spatial resolution of approximately 2704×3376 pixels. They are used to simulate snapshot-compressed hyperspectral imaging scenarios at different spatial scales.

[0074] During the experiment, hyperspectral data was modulated based on a unified snapshot compressed imaging model to generate corresponding two-dimensional measurement signals. Simultaneously, noise interference consistent with that of the actual imaging system was introduced during the measurement process to simulate measurement uncertainties under real imaging conditions. All comparison methods were reconstructed under the same imaging model, spectral settings, and noise conditions to ensure the fairness and comparability of the experimental results. Peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) were used as quantitative evaluation indicators to comprehensively assess the performance differences of different reconstruction methods in terms of reconstruction accuracy, structure preservation capability, and spectral consistency. Regarding model structure settings, the number of iteration stages in the multi-stage reconstruction process was set to a preset number (e.g., 3), with consistent structures in each reconstruction stage, thereby controlling computational complexity while ensuring reconstruction performance. Furthermore, to further verify the applicability of the proposed method under real imaging conditions, multiple sets of actually acquired snapshot compressed hyperspectral data were used for reconstruction tests.

[0075] 2. Experimental Results and Data Explanation: Under the above experimental conditions, a quantitative comparative analysis was conducted between the method proposed in this invention and several existing snapshot compression imaging reconstruction methods. In this embodiment, for ease of description, the specific network structure of the multi-scale gradient-guided snapshot compression imaging reconstruction method in the experimental implementation is abbreviated as RM-DUN. This abbreviation is only used to illustrate the experimental results and does not constitute a limitation on the technical solution of this invention. The relevant experimental results are shown in Table 1, where the uplink values ​​for each test scenario in Table 1 represent the peak signal-to-noise ratio (PSNR), the downlink values ​​represent the structural similarity index (SSIM), and the average result is the statistical average of the corresponding indicators for each test scenario.

[0076] Table 1 Comparison of reconstruction performance of different snapshot compression imaging reconstruction methods in simulated experimental scenarios (1)

[0077] According to the experimental results shown in Table 1, under 10 simulated test scenarios, the RM-DUN method of this invention achieves an average reconstruction performance of 41.06 dB PSNR and 0.986 SSIM, which is superior to various snapshot compression imaging reconstruction methods compared to the previous method. Compared with the representative depth unfolding method MiJUN, the average PSNR of this invention is improved by approximately 0.46 dB, and the SSIM is improved by approximately 0.008. Compared with the Transformer-based method SPECAT, the average PSNR is improved by approximately 0.69 dB, while maintaining a comparable level of structural similarity. Furthermore, Table 2 also presents the comparison results of the RM-DUN* method, where "*" indicates a variant method implemented using parameter sharing while maintaining the network structure unchanged. The experimental results show that even under parameter sharing conditions, the method of this invention can still maintain high reconstruction performance, demonstrating the effectiveness and stability of the method's structural design.

[0078] To further verify the effectiveness and generalization ability of the method described in this invention under actual imaging conditions, a verification experiment was conducted on data acquired using a real Snapshot Compressed Hyperspectral Imaging System (CASSI system). To simulate noise interference in actual imaging, 11-bit shot noise was added to the measurement data. Figure 3 As shown, this method recovers clearer details and more accurate colors in complex structural regions (such as petals and stems), demonstrating its good adaptability and reconstruction stability in real-world imaging scenarios.

[0079] 3. Comparative analysis of experimental conclusions: Comparative experimental results show that the proposed method exhibits superior or more stable reconstruction performance across multiple test scenarios, balancing reconstruction accuracy, structure preservation, and spectral consistency under varying imaging conditions. Specifically, this invention introduces a multi-scale gradient guidance mechanism within a single reconstruction stage, enabling gradient information generated by measurement consistency constraints to participate in feature modulation during the reconstruction mapping process. This establishes effective information interaction between gradient updates and reconstruction mapping, enhancing the recovery of texture details and structural edges. Simultaneously, during multi-stage reconstruction, this invention continuously propagates and fuses stage-specific feature information through a cross-stage message passing mechanism, allowing effective structural information extracted in earlier stages to be inherited and strengthened in subsequent stages. This alleviates the information decay and error accumulation problems common in traditional multi-stage iterative methods. Experimental results demonstrate that this mechanism effectively improves the overall stability of the reconstruction process, maintaining consistent performance across different test scenarios. In summary, the experimental results show that this method, while ensuring high reconstruction quality, effectively suppresses reconstruction artifacts and maintains good spectral consistency. It is suitable for snapshot-compressed hyperspectral imaging reconstruction tasks in complex scenarios, demonstrating promising engineering applications and practical application value.

[0080] It is understood that the present invention has been described through some embodiments, and those skilled in the art will recognize that various changes or equivalent substitutions can be made to these features and embodiments without departing from the spirit and scope of the invention. Furthermore, under the teachings of the present invention, these features and embodiments can be modified to adapt to specific situations and materials without departing from the spirit and scope of the invention. Therefore, the present invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the protection scope of the present invention.

Claims

1. A multi-scale gradient-guided snapshot compression imaging reconstruction method, characterized in that, Includes the following steps: S1: Modulate the hyperspectral scene to obtain a two-dimensional compressed measurement image; S2: Using an imaging model, the two-dimensional compressed measurement image is initially inverted to obtain the initial reconstruction result; S3: Based on the reconstruction results and mask parameters of the previous stage, calculate the gradient feature map, and perform gradient descent update on the reconstruction results of the previous stage based on the gradient feature map to obtain the intermediate reconstruction results of the current stage. S4: Perform multi-scale processing on the gradient feature map, extract gradient features at different scales, and generate weight information for modulating the near-end mapping reconstruction process; S5: Reconstruct the intermediate reconstruction results of the current stage in the near-end mapping network, and use the weight information to modulate the feature interactions in the near-end mapping process to generate the reconstruction results of the current stage. S6: Pass the reconstruction results of the current stage to the next stage, and establish message passing between multiple stages; S7: S3 to S6 are executed iteratively as a repeated reconstruction stage unit until the preset number of stages is completed, and the final reconstruction result is output.

2. The multi-scale gradient-guided snapshot compression imaging reconstruction method according to claim 1, characterized in that, The gradient feature map calculation steps in S3 include: Based on the reconstruction results and coding parameters of the previous stage, predictive measurements are generated, and measurement residuals are constructed with the two-dimensional compressed measurement images. The gradient feature map is obtained by back-projecting the measurement residual.

3. The multi-scale gradient-guided snapshot compression imaging reconstruction method according to claim 1, characterized in that, S4 includes: Using the gradient feature map as input, multi-scale gradient features are extracted through multiple parallel convolution operations with different receptive fields; The multi-scale gradient features are subjected to global average pooling and nonlinear activation to generate multi-scale gradient modulation weights for the near-end mapping reconstruction process.

4. The multi-scale gradient-guided snapshot compression imaging reconstruction method according to claim 1, characterized in that, The gradient descent update in S3 includes: Using the previous stage reconstruction result and gradient feature map as input, a lightweight neural network with residual connections is used to dynamically generate spatially adaptive step size parameters, and the previous stage reconstruction result is updated by gradient descent based on the step size parameters to obtain the intermediate reconstruction result of the current stage.

5. The multi-scale gradient-guided snapshot compression imaging reconstruction method according to claim 1, characterized in that, S6 includes: The reconstruction results of the current stage are used as the initial input for the next stage, and a cross-stage message passing mechanism is introduced in the near-end mapping network. The cross-stage message passing mechanism encodes and decodes multi-scale features through a hierarchical encoder-decoder structure.

6. The multi-scale gradient-guided snapshot compression imaging reconstruction method according to claim 5, characterized in that, A state-space model-based network structure is embedded in the encoder-decoder structure to propagate cross-stage multi-scale long-range dependency information between adjacent iteration stages, thereby enhancing information interaction between stages.

7. A multi-scale gradient-guided snapshot compression imaging reconstruction system, characterized in that, The system is implemented based on a multi-scale gradient-guided snapshot compression imaging reconstruction method as described in any one of claims 1-6, and the system includes: Image acquisition module: used to modulate hyperspectral scenes and acquire two-dimensional compressed measurement images; Initial Inversion Module: Used to perform initial inversion on the two-dimensional compressed measurement image using an imaging model to obtain initial reconstruction results; Gradient calculation module: used to calculate gradient feature map based on the reconstruction result of the previous stage and mask parameters, and perform gradient descent update on the reconstruction result of the previous stage based on the gradient feature map to obtain the intermediate reconstruction result of the current stage; Multi-scale gradient guidance module: used to perform multi-scale processing on the gradient feature map, extract gradient features at different scales, and generate weight information for modulating the proximal mapping reconstruction process; Reconstruction and Update Module: This module is used to reconstruct the intermediate reconstruction results of the current stage in the near-end mapping network, and to modulate the feature interactions in the near-end mapping process using the weight information to generate the reconstruction results of the current stage. Cross-stage transfer module: used to transfer the reconstruction results of the current stage to the next stage and establish message transfer between multiple stages; Iterative output module: used to iteratively execute the gradient calculation module to the cross-stage transfer module as a repeated reconstruction stage unit until the preset number of stages is completed, and output the final reconstruction result.

8. A computer device, characterized in that, The computer device includes a memory and a processor. The memory stores a computer program. When the processor runs the computer program stored in the memory, the processor executes a multi-scale gradient-guided snapshot compression imaging reconstruction method according to any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium is used to store a computer program that executes a multi-scale gradient-guided snapshot compression imaging reconstruction method according to any one of claims 1-6.

Images