A hyperspectral image reconstruction system and method based on adaptive perceptual coding and space spectrum enhancement network

Abstract

The application discloses a hyperspectral image reconstruction system and method based on adaptive perceptual coding and space-spectrum enhancement network. A two-dimensional compressed measurement image modulated by space-spectrum aliasing is collected through a photosensitive image sensor; a rough spectral estimation value is extracted by using a space-spectrum feature decoupling network; wherein the front end of the space-spectrum feature decoupling network comprises a learnable perceptual encoder formed by a 1*1 convolution layer, which is used for simulating the light transmittance distribution of a physical coding mask and is updated in cooperation with the rear-end network during training; in order to reduce the reconstruction deviation caused by the sampling-reconstruction split, a forward learnable perceptual encoder is constructed to simulate the physical degradation projection; and the iterative denoising prior block of the reconstruction network adopts a three-dimensional residual convolution structure.

Description

Technical Field

[0001] This invention belongs to the field of computational optical imaging and image processing optimization technology, specifically relating to a hyperspectral image reconstruction system and method based on adaptive perceptual coding and spatial spectrum enhancement network. Background Technology

[0002] Single-snapshot computational spectral imaging techniques (such as coded aperture snapshot spectral imaging (CASSI) add a coded mask and dispersive devices to the optical path. This allows for the capture of two-dimensional aliased data after spatial-spectral mixing modulation with only a single exposure of the image sensor, overcoming the limitation of multi-frame spatiotemporal displacement in traditional pushbroom hyperspectral cameras. Currently, deep unfolding networks are commonly used to achieve spectral inversion calculations.

[0003] However, in actual system operation and data processing, the existing solution faces the following challenges: The disconnect between the sampling end and the reconstruction algorithm: In existing compressed sensing hyperspectral reconstruction, hardware masking is typically performed at the physical device end using randomly designed or fixed coded matrix masks, followed by a separate algorithm for discretization and solution. Because the sampling and back-end reconstruction processes do not form an end-to-end closed-loop joint design, the coded mask fails to include any useful prior information specific to the reconstruction task. Under high compression ratios, crucial spatial-spectral information is lost, limiting the upper limit of image restoration accuracy.

[0004] Spectral distortion and high-frequency texture blurring: Hyperspectral cubes simultaneously encompass rich spatial two-dimensional scene shapes and continuous reflectance curves in the spectral dimension. Existing two-dimensional convolutional neural networks or self-attention-based architectures are prone to fragmenting the spectral curves along the channel dimension when recovering high-dimensional spatial spectral features, resulting in jagged distortion or even distortion of the reconstructed reflectance intensity curves, which limits the application of hyperspectral images in subsequent accurate recognition and classification. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a hyperspectral image reconstruction system and method based on adaptive perceptual coding and spatial-spectral enhancement networks. It solves the problem of severe loss of important detail information under high compression ratios due to the disconnect between traditional random physical coding matrices and backend neural network algorithm modules. Furthermore, this solution addresses the drawbacks of spectral distortion and edge alignment blurring that easily occur in high-dimensional data restoration by constructing a spectral enhancement network based on three-dimensional residual convolution and channel adaptive learning characteristics, thereby improving spatial-spectral fidelity performance.

[0006] To achieve the above objectives, the present invention specifically adopts the following technical solution: A hyperspectral image reconstruction method based on adaptive perceptual coding and spatial-spectral enhancement network includes the following steps: S1: Simultaneously acquire a two-dimensional compressed observation image of the target object formed on the photosensitive chip after physical aliasing modulation via spatial-spectral mapping. ; S2: Compress the two-dimensional observation image Input a spatial-spectral feature decoupling network to obtain multi-channel stretched alignment, and output a preliminarily estimated three-dimensional spectral data cube; S3: Construct a learnable physical forward degradation operator that includes random masking perturbations and small biases of spectral channels to simulate the differential coding sampling process of a computational spectral imaging system, and synchronously inject mixed additive physical noise during the forward propagation of model training to achieve robustness enhancement. S4: In the cascaded iteration, the learnable physical forward degradation operator is used to perform forward projection on the currently estimated three-dimensional spectral data cube to obtain a simulated two-dimensional projection estimation image, and the physical residual between it and the actual captured image is calculated. S5: The physical residual is back-projected to the hyperspectral data space using the adjoint transpose back-projection operator of the learnable physical forward degradation operator, and the spatial-spectral consistency gradient is calculated to correct the physical projection bias. S6: Input the three-dimensional spectral data cube and the uniformity gradient into the prior iterative denoising network. The prior iterative denoising network includes a three-dimensional residual convolutional denoising branch and a spectral enhancement attention branch. The three-dimensional spectral data cube is adaptively updated by three-dimensional spatial-spectral joint convolution, channel attention, and frequency domain difference weighting. S7: Repeat steps S4 to S6 until the preset cascade iteration is completed, and output the final three-dimensional spectral data cube after reconstruction.

[0007] Furthermore, in S1, the spatial modulation coding aperture is placed on the imaging optical path to perform adaptive spatial coding modulation on the intensity of the converged target reflected light; the dispersive prism assembly is used to physically deflect and refract the modulated mixed light in the spectral dimension and map it to the spatial-spectral scale; the photosensitive image sensor captures a single frame and directly converts the aliased light intensity signal into the corresponding two-dimensional compressed observation image.

[0008] Furthermore, in S3, the learnable physical forward degradation operator simulates the physical sampling masking process of the system, which is discretized and parameterized into a trainable network degradation layer in the forward computation. During the model training stage, the operator introduces a robustness enhancement strategy against physical disturbances: channel-level numerical biases are injected on the spectral channel axis of the input hyperspectral image, and density-adaptive Gaussian and Poisson mixed noise is superimposed on the encoded two-dimensional compressed observation image to simulate the dispersion distortion error generated by physical assembly and the dark current noise of the optoelectronic chip.

[0009] Furthermore, in S4, the constructed forward two-dimensional projection estimation formula is specifically as follows: In the formula, (u,v) represents the two-dimensional pixel plane coordinates on the photosensitive sensor; Let be the brightness value of the estimated intensity of the three-dimensional spectral feature plane in the c-th spectral band; The transmission coefficient of the learnable spatial coding mask at the corresponding pixel coordinates in band c; The zero-mean mixed physical additive noise generated in the system background.

[0010] Furthermore, in S5, the gradient mapping formula for single-stage iterative physical correction is: In the formula, For learnable physical forward degeneracy operators, the adjoint transpose backprojection operator is used. For the first Learnable step size parameters in cascaded iterations; the data inconsistency difference under the current iteration is obtained by calculation, and the physical backpropagation convergence of the spatial-spectral gradient is guided by the accompanying transpose backprojection to eliminate edge alignment ghosting or high-frequency ringing artifacts generated during inversion.

[0011] Furthermore, in S6, the prior iterative denoising network includes a three-dimensional residual convolutional denoising branch and a spectral enhancement attention branch: The three-dimensional residual convolutional noise reduction branch uses a size of The three-dimensional spatial-spectral joint convolution kernel extracts spatial-spectral collaborative local details through a sliding window spanning two-dimensional space and one-dimensional spectrum, and establishes a residual direct channel composed of identity jump connections between the input and output of the forward computation to maintain the integrity of the spatial contour structure. The spectral enhancement attention branch keeps the spatial dimension coordinates unchanged and performs global average pooling on the input three-dimensional feature map in the dimensions of spatial height H and width W to extract channel attribution information with global receptive field, resulting in a one-dimensional channel vector. Then, the band-dependent correlation weights are adaptively estimated through a bottleneck subnet with a fully connected layer to calculate the adaptive gain of each spectral channel. Finally, the feature map of the corresponding channel is multiplied back by the Hadamard product to correct the jump distortion of the continuous spectral reflectance curve caused by the physical deviation of the filter transmittance.

[0012] Furthermore, in S7, the model introduces manual ground truth labels for supervision during the training phase, and the constraint function formula used in the hybrid joint loss network is: In the formula, α and β are positive weighting constants that adjust the reuse ratio among various optimization objectives. In an optimal scheme, these coefficients are set to 0.1 and 0.05 respectively, but are not limited to these values. The discretization formula for the spectral angle loss is: This term is used to calculate the multidimensional geometric cosine angle between the predicted spectral impedance vector and the true reference reflection vector for each coordinate pixel in space, so as to prevent nonlinear feature jumps or curve jagged artifacts from being introduced into multi-band data during independent deconvolution restoration, and maintain the physical smoothness and fidelity of the hyperspectral reflectance curve.

[0013] Furthermore, after the reconstruction method is deployed on the edge chip, the convolutional computation acceleration unit inside the edge computing processor configures the hardware logic of memory alignment blocks when performing forward inversion calculation. This enables the three-dimensional feature tensor to be continuously addressed in the on-chip high-speed cache memory, skipping the multidimensional large matrix transpose and multiplication overhead of the channel attention matrix, and transforming the one-dimensional multiplication of spectral enhancement attention into a direct dynamic weight register product. This eliminates the high-frequency physical reading of large data blocks from the main external memory in continuous computation, improves the reconstruction frame rate, and reduces hardware power consumption.

[0014] Compared with the prior art, the present invention has the following beneficial effects: An integrated adaptive design for sampling and its inversion algorithm. This invention constructs learnable perceptual parameters in the forward path, equating the discrete matrix of the physical compression coding process to a high-parameter coding convolutional layer. Under this structure, the sampling structure output by the front-end hardware micromirror mask can adaptively match the hyperspectral probability distribution of a specific task, breaking the bottleneck of separation between hardware design and software computation, and ensuring feature retention under high compression ratios.

[0015] Multidimensional spatial spectral feature cascaded 3D denoising and spectral fidelity preservation. This invention introduces a combination of a 3D-CNN denoising network and a dual-branch spectral enhancement module. The 3D residual structure effectively protects the boundary shape and physical smoothness of each band slice, saving computational memory overhead for large correlation models; the dual-branch spectral enhancement path corrects spectral reflection energy level distortion caused by optical lens and hardware deviations through adaptive amplitude multiplication of channel axes and Fourier frequency domain spatial residual repair.

[0016] Geometric angle and pixel-level 3D hybrid optimization supervision. By introducing an angle loss function that integrates structural similarity, pixel absolute error and multidimensional spectral reflectance vector into the network, the parameter change rate is constrained from dimensions such as structure, pixel distribution and band continuity physical logic. This eliminates the jagged edges or spectral envelope drift generated by the reconstructed feature map under large-span dispersion, and achieves deep hardware and software consistency of the physical imaging mechanism. Attached Figure Description

[0017] Figure 1 The overall algorithm flowchart for the adaptive perceptual coding and spatial spectrum enhancement reconstruction method; Figure 2 This is a schematic diagram of the component layout of the hyperspectral snapshot hardware system provided by the present invention; Figure 3 The diagram shows the topology of the spectral enhancement module (SEM) and the internal spatial spectral channel weighting calculation of the 3D convolutional network. Detailed Implementation

[0018] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] The technical solutions in the embodiments of the present invention are clearly and completely described in the figures. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0020] Example 1: like Figures 1 to 3 As shown, the implementation process of the hyperspectral image reconstruction method based on adaptive perceptual coding and spatial spectrum enhancement network proposed in this invention includes the following steps: S1: Initialization Configuration and Adaptive Data Acquisition: The hyperspectral physical sensor camera is equipped with a dispersive unit with 28 spectral channels, and the center wavelength of the calibrated imaging reflection falls within the 450nm-650nm range. The externally incident light field is first converged by an achromatic front condenser lens, and then modulated with a specific spatial structure at a digital reflection mask (transmittance is discretized by a learnable sensor network and represented by the chromium layer thickness). The aliased coded light is refracted laterally along the positive direction of the image plane by the dispersive prism assembly, and finally, a single 20ms exposure is performed by the shutter CMOS grayscale sensor.

[0021] On the computing side, the main control module reads the dimension generated by the photosensitive chip, which has a size of [missing value]. A two-dimensional compressed observation image y with a depth channel of 1; the data scale corresponding to the real input hyperspectral scene is set to... (Spatial size 256×256, spectral channels 28).

[0022] S2: Spatial-spectral feature decoupling and preliminary 3D modeling: The input two-dimensional compressed data is fed into a spatial-spectral decoupling network.

[0023] The network head uses a convolutional kernel with a stride of 1 and kernel parameters of 1. The transposed 2D convolutional layer directly copies and maps the input single-channel first-order tensor features along the depth direction into 28 multi-dimensional layers; then, the second layer introduces a decoupling pipeline consisting of 3D convolution operations (3D Conv). Convolution parameter settings: the 3D convolution kernel dimension is... It is used to extract information about adjacent spatial networks and channel transitions.

[0024] After the data is forwarded in this branch, the output is a preliminary three-dimensional spectral cube after dimensional expansion, with a dimension of [missing value]. .

[0025] S3: Forward optimization of end-to-end learnable perceptual encoder: To overcome the accuracy loss problem caused by the separation of sampling and demodulation, this invention deploys a learnable perceptual encoder at the front end of the forward computation process: This expression is abstractly represented in the algorithm as a The two-dimensional convolutional layer has 28 input channels and 1 output channel, without any added bias components. To prevent overfitting of this sampling convolutional kernel, its network parameters are constrained to the range [0, 1] by the Sigmoid function during backpropagation.

[0026] During the forward propagation phase, physical degradation perturbations are introduced into the output of the sampling operator: additive white Gaussian noise and Poisson noise with variable signal-to-noise ratio are injected into the two-dimensional compressed data y in real time; and random channel-level numerical micro-shift perturbations (Spectral Shift) are applied to each band image in the spectral dimension of the original hyperspectral input X, and random masking (Mask Dropout) is introduced in the spatial dimension to prevent the perception matrix from getting overfitted during backpropagation.

[0027] S4: Forward reprojection simulation and error calculation: Discretized simulation of the real sampling mechanism is performed using the coarse spectral results output by S2: A forward estimation simulation image was obtained by superimposing and multiplying the data from 28 spectral bands according to their corresponding transmittance. Then, the input image acquired by the actual camera was used. With the forward estimation graph Perform differential calculations to obtain the pixel energy fluctuations corresponding to the two-dimensional topological plane, and obtain the physical residual feature map. .

[0028] S5: Deep Unfolding Physical Gradient Cascade Solution: The iterative demodulation network constructed in this invention has a total of Each node is an independent cascaded expansion node. In the computation of any iteration level k, the first part receives the output features of the previous level by the Data Consistency Physical Constraint Module (DC), and the transpose of the learnable degenerate operator is used for the inverse adjoint mapping operator. Dominant correction gradient: In the formula, For the first Learnable step size parameters in cascading. Updated output features are backpropagated through data consistency. Input into the 3D convolutional decoding prior denoising network: This denoising network contains 3D convolutional layers constrained by cascaded 3D residual blocks. The size of each 3D convolutional kernel is [size missing]. The step size is 1, the padding is 1, and Group Normalization (grouping normalization, with the number of groups set to 8) and PReLU activation layers are cascaded in the middle. Identifier direct channels are set at the input and output of the entire 3D-CNN so that the shallow high-frequency detail information calculated during the forward physical residual iteration can bypass the computational reduction of deep networks and avoid gradient vanishing or spectral gradient diffusion in deep networks.

[0029] S6: Channel attention and frequency domain residual spectral enhancement tuning: The denoising features are fed into a dual-branch parallel spectral enhancement module for processing to correct reflectance: Channel spectral correction branch: First, The feature map is compressed using global average pooling along the spatial dimensions (H, W) to... The one-dimensional feature vector is then passed through a bottleneck multilayer perceptron (MLP) consisting of two fully connected layers (with 28 and 7 hidden nodes respectively, and a channel attenuation compression ratio of r=4) and an activation function to calculate the activation vector for each spectral channel. The corresponding weight is multiplied by the corresponding channel of the feature map using the Hadamard product to complete the amplitude correction.

[0030] Frequency domain residual correction branch: The adjusted feature map is transformed to the frequency domain using fast Fourier transform, and its imaginary and real residual high-frequency coefficients are predicted. The high-frequency residuals are then restored by inverse Fourier transform in the frequency domain and spatially added together.

[0031] The output characteristics of the spectral enhancement module ensure that the 28-dimensional data can smoothly match the physical refractive constant of the material on the continuous reflection curve.

[0032] S7: Training based on hybrid loss for comprehensive physical consistency: The network utilizes a multispectral database labeled with scientific truth values ​​for parameter optimization. The Adam optimizer guides end-to-end backpropagation during training. The final total loss metric during training is determined by... Composition, including pixel loss Maintaining basic stability of brightness, structural loss Constrained geometric profile, spectral angle loss This is used to measure the angle between two multidimensional vectors, penalizing nonlinear jump artifacts introduced into the spectral reflectance curve due to physical sampling mismatch, and ensuring the physical authenticity of the output hyperspectral land cover attributes. Its calculation model is as follows: The optimal scheme is adopted for setting the adjustment parameters: weight ratio adjustment parameters. By constraining the angle variation of the spectral reflectance curve in multiple spectral dimensions, the undesirable local abrupt value jump phenomenon caused by independent deconvolution is eliminated.

[0033] S8: Edge hardware solidification and model acceleration: After training, it will include the perception matrix. The weights of the joint network, consisting of the backend 3D-CNN and the spectral enhancement module, are exported as a generic ONNX file.

[0034] It is loaded and deployed in an image processor chip equipped with a tensor computing accelerator. The inference stage uses half-precision floating-point (FP16) computing power precision conversion and divides the large 3D feature input matrix into several parallel high-dimensional sliding window blocks at the edge, avoiding a large number of redundant 3D convolutional layer activations occupying the on-chip cache, thereby achieving sub-second response time for hyperspectral video stream reconstruction performance on the edge platform.

[0035] Example 2: This embodiment provides details of the component layout and computational collaborative architecture of a hyperspectral image reconstruction system: The system's hardware components and installation structure are shown in the attached figure. Figure 2 As shown, a lens objective is positioned on the outer side of the optical path to converge the reflected field of view from a distance. A two-dimensional adjustable DMD (Digital Micromirror Device), a collimating lens, and a prism with a stable refractive index are arranged coaxially along the optical axis to form the dispersion components. A CMOS grayscale camera with wide dynamic range is positioned close to the light output end of the system to capture and output the final aliasing and compression characteristics after the dispersion and diffraction superposition.

[0036] The reconstruction computing core is based on a high-efficiency edge computing SoC device, which directly obtains the two-dimensional observation data stream from the CMOS image sensor through a high-speed video interface.

[0037] When the hyperspectral fast camera is powered on and running in real-time streaming video mode, the edge computing chip loads the network compression static parameter file, which has been compiled and pruned with discrete precision, into its on-chip memory and directly decomputes each frame of the sensor video stream's two-dimensional compressed original image.

[0038] When performing back-projection operations, the sparse operation acceleration unit performs parallel shift indexing for the fixed adaptive coding spatial positions. Instead of allocating redundant high-order large matrices for overlapping positioning, the acceleration unit writes offset physical words to the pointer offset registers corresponding to the features of each channel. This allows for one-step alignment and translation during hardware memory access, ultimately recovering a full-resolution 3D spectral image at the very end of the cascaded network.

[0039] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A hyperspectral image reconstruction method based on adaptive perceptual coding and spatial-spectral enhancement network, characterized in that, The method runs on a computing terminal and specifically includes the following steps: S1: Adaptive compressed sampling step: Construct a learnable perceptual encoder composed of a convolutional network, input high-dimensional hyperspectral data, simulate the spatial-spectral joint degradation projection process in the actual optical imaging path through the parameter perceptual matrix in the forward propagation, and output a two-dimensional compressed observation image. S2: Spatial-spectral feature decoupling three-dimensional reconstruction steps: The two-dimensional compressed observation image is input into a preset cascaded three-dimensional convolutional decoding and reconstruction network. The joint neighborhood and channel correlation features of the spatial dimension and the spectral dimension are extracted through the three-dimensional convolution kernel. The preliminary estimated three-dimensional spectral data cube is reconstructed by the cyclical alternation of the data consistency module and the denoising prior module. S3: Spectral Feature Correction and Enhancement Steps: The initially estimated 3D spectral data cube is input into a dual-branch parallel spectral enhancement module; wherein, the first branch is a channel attention branch, used to perform average convergence on the input feature map along the spatial dimension to extract the global specificity of the spectral dimension, and to perform adaptive amplitude weighting adjustment on the spectral band energy; the second branch is a frequency domain residual branch, used to transform the spatial feature map to the frequency domain using a two-dimensional fast Fourier transform, use frequency domain convolution to estimate and inject the missing high-frequency spectral residuals, and then perform an inverse transformation and concatenate it with the output of the first branch in the channel dimension to obtain a concatenation matrix, and then use a convolutional layer with a convolutional kernel size of 1×1 to reduce the channel dimensionality to the total number of spectral channels, and output the enhanced 3D hyperspectral image; S4: Collaborative Iterative Optimization Steps: Construct an end-to-end joint training framework, utilize a hybrid loss network based on real labels to calculate the multidimensional joint difference between the final reconstructed hyperspectral image and the artificial ground truth label, backpropagate the error gradient and synchronously adjust the perceptual matrix weights in the learnable perceptual encoder to optimize physical sampling, and adjust all global weight parameters of the backend 3D convolutional decoding reconstruction network and spectral enhancement module.

2. The hyperspectral image reconstruction method based on adaptive perceptual coding and spatial-spectral enhancement network according to claim 1, characterized in that: The discretized physical-mathematical model of the learnable perceptual encoder in step S1 is expressed as follows: In the formula, Represents a two-dimensional compressed observation image; represents the monochromatic image component of the original hyperspectral image data in the c-th channel, where C is the preset total number of spectral channels; This represents the adaptive parameter sensing weights corresponding to the c-th spectral band. The matrix is ​​the Hadamard product, where n is the external photoelectric additive physical noise generated on the sampling link; the parameter sensing weights In a deep network, a network has one input channel with C channels, one output channel, a bias of 0, and weights that have undergone activation constraints. within the range The two-dimensional convolutional layer mapping represents that the weights of the convolutional layers are collaboratively optimized and updated during backpropagation.

3. The hyperspectral image reconstruction method based on adaptive perceptual coding and spatial-spectral enhancement network according to claim 2, characterized in that: In step S1, a robust enhancement interference mechanism is introduced during the training phase. The interference mechanism is as follows: during forward propagation, Gaussian white noise and photon Poisson noise are injected in real time into the two-dimensional compressed observation image y output by the learnable sensing encoder at a ratio; and at the input end of the training samples, random channel-level small numerical offset perturbations are applied to the amplitude of the original hyperspectral image X on the spectral channel axis, so as to improve the robustness of the adaptive parameter sensing matrix to the actual lens imaging color shift and position assembly error.

4. The hyperspectral image reconstruction method based on adaptive perceptual coding and spatial-spectral enhancement network according to claim 1, characterized in that: The cascaded 3D convolutional decoding and reconstruction network in step S2 adopts a depth-unfolded network framework, and the nearest neighbor operator iteration update rule in its k-th cascade satisfies: In the formula, This represents the adjoint backprojection operator corresponding to the parameter sensing matrix \Phi, used to reconstruct and map the two-dimensional physical estimation difference to the three-dimensional spectral space. Let be the learnable iterative gradient step size in the k-th cascade. The hyperspectral cube estimate is the input for the k-th cascade. This represents a cascaded denoising prior mapping network composed of 3D convolution and residual jump paths.

5. The hyperspectral image reconstruction method based on adaptive perceptual coding and spatial-spectral enhancement network according to claim 4, characterized in that: The cascaded noise reduction prior mapping network employs three-dimensional residual convolutional blocks, each with two symmetrical three-dimensional convolutional layers. The kernel size of each three-dimensional convolutional layer is [size missing]. The stride is 1 and the padding is 1. Each 3D convolutional layer is followed by a grouping normalization layer with 8 groups and a PReLU nonlinear activation layer. Furthermore, an identity skip connection is established between the input and output of each 3D residual convolutional block to ensure that the high-density edge information and spectral features in the shallow spatial domain are not diluted by overfitting during multi-stage forward propagation.

6. The hyperspectral image reconstruction method based on adaptive perceptual coding and spatial-spectral enhancement network according to claim 1, characterized in that: In step S3, the spectral enhancement module performs channel attention branch processing for feature extraction and distortion removal of the three-dimensional feature map. Includes the following steps: Global compression of spatial features: The size of the input feature map is denoted as... Where H and W represent the height and width of the image, respectively, three-dimensional global average pooling is performed along the spatial dimension to flatten it into a one-dimensional channel feature vector containing the global receptive field. The calculation formula is: In the formula, F(i,j,c) represents the coordinates of the feature map on the c-th channel. Pixel feature values; Bottleneck Neural Network Mapping: The one-dimensional channel feature vector z is input into a bottleneck subnetwork containing two fully connected layers. A nonlinear activation function is used to map the correlation between different spectral bands, and an adaptive channel activation weight vector is output. Its mapping calculation satisfies: In the formula, Let r be the channel weight dimensionality reduction transformation matrix. This is a channel amplification matrix. This represents the Sigmoid activation function, used to map the output activation values ​​to... ; Channel-weighted product correction: The activation weight vector is adjusted using the Hadamard product. The first in Each component Corresponding to the first of the input feature maps Multiplying the bands adjusts the reflectivity energy of each band.

7. The hyperspectral image reconstruction method based on adaptive perceptual coding and spatial-spectral enhancement network according to claim 6, characterized in that: In step S3, the spectral enhancement module performs the following steps for feature extraction and distortion removal of the three-dimensional feature map in the frequency domain residual branch processing: Two-dimensional frequency domain mapping: Using two-dimensional fast Fourier transform, the channel-weighted corrected features are transformed into the frequency domain along the spatial direction to obtain the two-dimensional frequency domain amplitude-frequency features. Frequency domain residual operator adjustment: The obtained feature matrix is ​​used to calculate the high-frequency repair residual in multiple frequency domain residual convolutional layers, and the high-frequency spectral interpolation prediction is output. ; Frequency domain inverse transform output: The frequency domain feature matrices before and after the convolution operation are summed and sent to a two-dimensional fast inverse Fourier transform to complete the inverse fast Fourier transform mapping back to the real number space.

8. The hyperspectral image reconstruction method based on adaptive perceptual coding and spatial-spectral enhancement network according to claim 1, characterized in that: In step S4, the hybrid loss function for joint training consists of pixel-level absolute error loss. Spatial structure similarity loss And the spectral angle fidelity loss due to the limitation of the absolute angle of spectral reflectance curvature. It is a weighted sum, and its specific expression is: In the formula, and These are the positive bias parameters for adjusting the multi-task optimization balance; the spectral angle fidelity loss. Defined as: Using the inverse cosine function to reconstruct the hyperspectral pixel vector With the true pixel vector The penalty term is calculated based on the geometric angle between the two sides to constrain the physical reflection continuity of the spectrum.

9. A computational hyperspectral image reconstruction system for performing the method according to any one of claims 1 to 8, characterized in that, include: The sampling physical optical path includes a front-end converging lens arranged coaxially, a physical digital micromirror matrix with learnable and perceptual parameter mapping, a collimation mechanism, and a dispersive prism assembly that performs physical segregation and spatial unfolding of the transmitted light spot in the spectral dimension. An image sensor chip is used to perform pixel-level photoelectric conversion integration on the dispersion-stretched aliased optical path and output a two-dimensional compressed observation image. An edge image processor is communicatively connected to the photosensitive chip circuit and includes a high-bandwidth on-chip cache memory and a sparse network unit. It loads a deep inference application based on a portable format statically compiled and stored with the optimal adaptive perception matrix and network weights; when running reconstruction operations, the edge image processor introduces a multi-level physical offset register (Pointer Offset Register) for the three-dimensional spatial addressing of each channel, and maps the attention matrix multiplication to an asynchronous pointer offset shift accumulation, skipping the read and write of DRAM external memory and directly outputting the three-dimensional spectral data cube.

10. A computer-readable storage medium storing computer instructions thereon, characterized in that: When the instruction is executed by the processor, the computational spectral reconstruction method as described in any one of claims 1 to 8 is implemented.

Images