Calculation type polarization hyperspectral reconstruction method

Abstract

The invention discloses a calculation type polarization hyperspectral reconstruction method, and relates to the technical field of polarization hyperspectral imaging, and the method comprises the steps: obtaining a nine-channel mixed measurement signal of the combination of three polarization angles and three spectral bands through the single exposure of a nine-channel polarization wide-spectrum image collection system; and inputting the signal into a pre-trained image reconstruction model for processing. The image reconstruction model comprises a feature coding module and a calculation reconstruction module, the calculation reconstruction module adopts a plurality of SP modules based on U-Net architecture, a core SPG module of the calculation reconstruction module can extract and fuse spectral features and polarization features respectively, and finally, through calculation under physical constraints, the spectral features and the polarization features are obtained. And reconstructing a three-dimensional data cube of a complete hyperspectral image at a plurality of polarization angles from nine input channels. The method overcomes the defects that a traditional method is complex in light path, needs multiple times of collection and is low in reconstruction quality, achieves single-time snapshot and high-fidelity polarization hyperspectral imaging, and remarkably improves the imaging efficiency and the system integration degree.

Description

Technical Field

[0001] This invention belongs to the field of polarization hyperspectral imaging technology, specifically relating to a computational polarization hyperspectral reconstruction method. Background Technology

[0002] Hyperspectral imaging technology acquires a "three-dimensional data cube" containing two-dimensional spatial information and one-dimensional spectral information by imaging the same scene in dozens to hundreds of consecutive narrow bands. It has shown great advantages in fields such as material identification, anomaly detection, and fine classification. Polarization imaging technology can capture physical characteristics of objects such as surface roughness, texture orientation, and material type, and can effectively suppress strong reflections and enhance contrast, thereby solving the problem of "different objects with the same spectrum".

[0003] Polarization hyperspectral imaging technology integrates the material identification capabilities of spectroscopy with the physical property perception capabilities of polarization, providing high-dimensional data support for fields such as fine classification and anomaly detection. However, acquiring multi-dimensional data cubes of polarization hyperspectral imaging has always been a technical challenge in this field. Existing technical solutions mainly revolve around two approaches: one is the direct acquisition method, which uses a mechanical combination of polarization and spectral elements to acquire complete information through time-sharing or spatial scanning. Although the data quality is high, the system is bulky and the imaging speed is slow, making it unsuitable for dynamic scenes; the other is the indirect acquisition method, namely the computational reconstruction method, which uses optical encoding to compress high-dimensional information into a small number of two-dimensional images, and then reconstructs them using algorithms, aiming to achieve snapshot-like imaging.

[0004] In indirect acquisition methods, mainstream techniques often employ components such as liquid crystal tunable filters, quarter-wave plates, and linear polarizers to construct the encoded optical path. Specifically, some studies have proposed a compressed sensing reconstruction method based on particle swarm optimization sparse basis. This method utilizes the tunability of liquid crystal tunable filters, switching their center wavelength multiple times in conjunction with waveplate angle modulation to achieve compressed measurement and reconstruction of Stokes parameters. While this method reduces the number of observation frames, the multiple wavelength switching introduces significant time delays. Subsequent research has attempted to simplify this process, for example, by introducing linear polarizers with fixed angles, reducing the required compressed observations to two. However, the increased number of optical path components also makes system assembly and adjustment more complex. Modeling errors of the Mueller matrix and alignment errors of the devices accumulate and propagate, reducing the system's robustness.

[0005] In recent years, deep learning has been introduced to further optimize the reconstruction process, resulting in methods that require only a single compressed exposure and use convolutional neural networks to reconstruct Stokes parameters end-to-end. This approach has significantly improved efficiency, but its core network architecture is still based on traditional convolutional neural networks. Its inherent assumptions of local connectivity and translation invariance make it difficult to effectively model the long-range dependence of the spectral dimension and the complex coupling between the polarization-spectral dimension, leading to problems such as spectral distortion, excessive smoothing of details, or inaccurate physical quantities in the reconstruction results. Summary of the Invention

[0006] To address the aforementioned problems in the existing technology, this invention provides a computational polarization hyperspectral reconstruction method. The technical problem to be solved by this invention is achieved through the following technical solution: This invention provides a computational polarization hyperspectral reconstruction method, comprising: Step 1: Use a nine-channel polarization broadband image acquisition system to acquire a nine-channel mixed measurement signal combining three polarization angles and three spectral bands; Step 2: Input the nine-channel mixed measurement signal into the trained image reconstruction model to obtain polarization hyperspectral images at different polarization angles; The image reconstruction model includes a feature encoding module and a computational reconstruction module. The feature encoding module is used to perform preliminary feature extraction and encoding on the input nine-channel mixed measurement signal to obtain initial encoded features. The computational reconstruction module is used to perform computational reconstruction on the initial encoded features to obtain polarized hyperspectral images at different polarization angles. The computational reconstruction module includes multiple SP modules using a U-Net network architecture. The SP modules obtain the spectral features and polarization features of the hyperspectral images at different polarization angles based on the SPG module, and calculate the polarized hyperspectral images at different polarization angles.

[0007] Compared with the prior art, the beneficial effects of the present invention are as follows: The computational polarization hyperspectral reconstruction method of this invention acquires nine-channel mixed measurement signals through a nine-channel polarization broadband image acquisition system. Based on these nine-channel mixed measurement signals, a single-frame polarization hyperspectral image can be acquired and reconstructed using an image reconstruction model, eliminating the need for complex optical path design and time-consuming wavelength tuning. The main network of the image reconstruction model uses U-Net as its backbone network, where the SPG module fully represents and effectively fuses information from both the spectral and polarization dimensions, ensuring accurate reconstruction of the output results at different wavelengths and polarization angles. This not only significantly shortens the reconstruction latency and reduces the system implementation complexity, but also enables the scheme to adapt to more and more complex application scenarios, providing greater flexibility and support for the engineering applications of the detection end.

[0008] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described in detail below with reference to the accompanying drawings. Attached Figure Description

[0009] Figure 1This is a flowchart of a computational polarization hyperspectral reconstruction method provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of the image reconstruction model provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the SP module provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the structure of the SPG module provided in an embodiment of the present invention. Detailed Implementation

[0010] To further illustrate the technical means and effects adopted by the present invention to achieve the intended purpose, the following describes in detail a computational polarization hyperspectral reconstruction method proposed according to the present invention, in conjunction with the accompanying drawings and specific embodiments.

[0011] The foregoing and other technical contents, features, and effects of the present invention will be clearly presented in the following detailed description of specific embodiments in conjunction with the accompanying drawings. Through the description of the specific embodiments, a more in-depth and concrete understanding can be gained of the technical means and effects adopted by the present invention to achieve its intended purpose. However, the accompanying drawings are for reference and illustration only and are not intended to limit the technical solutions of the present invention.

[0012] This invention provides a computational polarization hyperspectral reconstruction method; please refer to [link to relevant documentation]. Figure 1 , Figure 1 This is a flowchart of a computational polarization hyperspectral reconstruction method provided in an embodiment of the present invention. Figure 1 As shown, the computational polarization hyperspectral reconstruction method of this embodiment includes the following steps: Step 1: Use a nine-channel polarization broadband image acquisition system to acquire a nine-channel mixed measurement signal combining three polarization angles and three spectral bands.

[0013] In this embodiment, the nine-channel polarization broadband image acquisition system includes an objective lens and a polarization broadband detector.

[0014] The objective lens is used to collect scene radiation and image it onto the detection surface of the polarization broadband detector. The polarization broadband detector includes several macropixels of equal size. Each macropixel contains several subpixels, which are periodically arranged according to at least three different polarization angles and at least three linearly separable spectral bands.

[0015] It is understandable that if the sub-pixels are periodically arranged according to three different polarization analysis angles θ1, θ2, and θ3 and three linearly separable spectral bands (the center wavelengths of the spectral bands are λ1, λ2, and λ3, and the spectral band range and center wavelength can be set according to the application), then each macro-pixel can simultaneously acquire a mixed measurement signal of nine channels, consisting of three polarization angles and three spectral bands, in a single exposure.

[0016] Step 2: Input the nine-channel mixed measurement signal into the trained image reconstruction model to obtain polarization hyperspectral images at different polarization angles.

[0017] Image reconstruction is performed on a macro-pixel basis, recovering linear Stokes parameters (S0, S1, S2) and their derivatives from the three polarization intensities within each spectral band. The final output is a spatially co-located nine-channel polarized broadband image, namely I(x,y,λ1,θ1), I(x,y,λ1,θ2), I(x,y,λ1,θ3), I(x,y,λ2,θ1), I(x,y,λ2,θ2), I(x,y,λ2,θ3), I(x,y,λ3,θ1), I(x,y,λ3,θ2), I(x,y,λ3,θ3).

[0018] In this embodiment, the polarization hyperspectral reconstruction process is as follows: a nine-channel polarization broadband image acquisition system is used to acquire a mixed measurement signal of nine channels (three polarizations × three spectral bands). This signal is then fed into a trained image reconstruction model to obtain the spectral bands λ1~λ3 with intervals of λ at three polarization angles. Δ Subgraph I(x,y,λ) * m1 ,θ1),I(x,y,λ * m2 ,θ2),I(x,y,λ * m3 The system generates m images, i.e., m1 images at polarization angle θ1, m2 images at polarization angle θ2, and m3 images at polarization angle θ3, for a total of m images. This completes the reconstruction of a three-dimensional data cube from a nine-channel polarized broadband image to a complete polarized hyperspectral image. The final output is I(x,y,λ). * m1-m3 ,θ1),I(x,y,λ * m1-m3 ,θ2),I(x,y,λ * m1-m3 A three-dimensional data cube of a complete polarization hyperspectral image consisting of m images (θ3).

[0019] Furthermore, the structure of the image reconstruction model and the specific process of the reconstruction workflow in this embodiment will be described.

[0020] In this embodiment, the image reconstruction model includes a feature encoding module and a computational reconstruction module. The feature encoding module is used to perform preliminary feature extraction and encoding on the input nine-channel mixed measurement signal to obtain initial encoded features. The computational reconstruction module is used to perform computational reconstruction on the initial encoded features to obtain polarization hyperspectral images at different polarization angles. The computational reconstruction module includes multiple SP modules using a U-Net network architecture.

[0021] Please see Figure 2 , Figure 2 This is a schematic diagram of the image reconstruction model provided in an embodiment of the present invention, as shown below. Figure 2 As shown, the nine-channel mixed measurement signal undergoes feature extraction and encoding by the feature encoding module, and the encoded result is input into the SP module. The SP module has an independent feature extraction and restoration section, so it ultimately restores the polarization features to the image and outputs it as an image, resulting in polarization hyperspectral images at three polarization angles, each with a resolution of m. n Zhang, m sheets in total.

[0022] Understandably, the effect of a single SP module is limited, and multiple SP modules are typically stacked to increase model complexity. For example, a U-Net network architecture can be used to connect three SP modules. Figure 2 As shown, the output of the first SP module undergoes further processing. Part of it is downsampled and then passed to the next SP module. In the diagram, C means concatenation along the channel dimension. Another part is combined with the channel number symmetrical to it in U-Net. The next SP module is processed similarly. After being passed to the third SP module, the output is upsampled and combined with the previously downsampled output along the C channel dimension. Then, a 1×1 convolutional block is applied to ensure that the number of channels is consistent. The next SP module is processed similarly. The output of this process is repeated once, and then combined with the output of the network after passing through the feature encoding module at the beginning. Finally, they are used as the output of the image reconstruction model, which is the reconstruction result.

[0023] In this embodiment, the SP module obtains the spectral and polarization features of hyperspectral images at different polarization angles based on the SPG module, and calculates polarized hyperspectral images at different polarization angles.

[0024] Furthermore, the specific structures of the SP module and the SPG module will be explained.

[0025] Please see Figure 3 , Figure 3 This is a structural diagram of the SP module provided in an embodiment of the present invention, as shown below. Figure 3As shown, the SP module in this embodiment adopts a residual network structure. The SP module includes an SPG module, a layer normalization module, and a first convolution module. The SPG module calculates the spectral and polarization features of the polarization hyperspectral image and obtains the polarization hyperspectral image based on these features. The polarization hyperspectral image is then processed sequentially by the layer normalization module and the convolution module, and finally added to the output of the SPG module to obtain the output of the SP module.

[0026] In this embodiment, the first convolutional module includes a first 1×1 convolutional layer, a first GELU activation function layer, a 3×3 convolutional layer, a second GELU activation function layer, and a second 1×1 convolutional layer connected together.

[0027] Specifically, the input of the SP module is processed by the SPG module to obtain m. n The SPG module output consists of m images at the three polarization angles. After passing through a layer normalization module, a 1×1 convolutional layer transforms the m channels into 2m channels. Then, it passes through a 3×3 convolutional layer and then another 1×1 convolutional layer to transform the 2m channels back into m channels. A GELU activation function is added between the convolutional layers to introduce nonlinear features. The output of the second 1×1 convolutional layer is added to the output after passing through the SPG module and used as the output of the SP module.

[0028] Please see Figure 4 , Figure 4 This is a schematic diagram of the SPG module provided in an embodiment of the present invention. Figure 4 As shown, the SPG module includes a spectral information submodule, a polarization information submodule, and a physical fusion submodule. The spectral information submodule employs a self-attention Transformer architecture, using the Transformer self-attention mechanism to process the Stokes parameters of different input bands. Learn and output the spectral features of polarization hyperspectral images The polarization information submodule uses residual networks and physical coding to obtain Stokes parameters for different input bands. and Stokes parameters Optimize and output the first polarization feature of the polarization hyperspectral image. Second polarization characteristics The physical fusion submodule is used based on spectral characteristics. First polarization characteristic Second polarization characteristics The polarization hyperspectral image was calculated.

[0029] Specifically, for the spectral information submodule, its input is the Stokes parameters for different wavelength bands. Total m n(The number of outputs here is consistent with the final number of channels required, which is also consistent with the number of output images of the image reconstruction model.) Figure 4 To make the content easier to see, the drawing is relatively small. After encoding, it becomes a two-dimensional vector, which is then multiplied by matrix W. Q W K W V We obtain vectors Q, K, and V. The self-attention matrix is ​​obtained by multiplying K and Q, then multiplied by vector V. Finally, the position code PE is added as the output of the spectral information submodule, i.e., m. n Stokes parameters across the entire spectral range In the spectral information submodule, the input to the spectral head is... ,in It is the height of the input image. It is the width of the input image. It is the number of input channels. The Stokes parameters represent the polarization components at different wavelengths. To facilitate self-attention calculation, it is transformed into Then, the input X is mapped to vectors Q, K, and V, each with dimensions HW×C, where each component can be written as... , and ,in This represents the number of heads in the bulls' self-attention. .

[0030] Therefore, the spectral head for each dimension under multi-head attention is: ; in, These are learnable parameters. It is K's j The transpose of the component matrices of each dimension, then... The attention points are pieced together and combined with the location code. Add them together to get the output: MSA= ; Where W is a learnable parameter, This is the part used to generate positional encoding. Its structure is the same as in a typical Transformer. In multi-head attention, the actual learnable parameter W... K W Q W V And W, finally through two DWconv3×3 (depth 3×3 convolutions) and an activation layer GELU, and after dimensionality transformation, the final output size is .

[0031] In this embodiment, the polarization information submodule in the SPG module adopts a residual network architecture. The polarization information submodule includes a polarization coding module, a second convolutional module, and a Tanh activation function layer. The Stokes parameters for different wavebands are also specified. and Stokes parameters After dimensionality compression by the polarization encoding module and convolution by the second convolution module, the result is added to the output of the polarization encoding module and then passed through a Tanh activation function layer to output the first polarization feature of the polarization hyperspectral image. Second polarization characteristics The second convolutional module includes a first 3×3 convolutional layer, a third GELU activation function layer, a first pixel 1×1 convolutional layer, a second 3×3 convolutional layer, a fourth GELU activation function layer, and a second pixel 1×1 convolutional layer.

[0032] Specifically, if it is the first SPG module in the first SP module of the calculation reconstruction module, then its input is the three Stokes parameters obtained from the initial nine-channel mixed measurement signal. and Stokes parameters The matrix is ​​then extended to m using linear interpolation. n Zhang Stokes parameters and Stokes parameters The image is then compressed in dimensions from 2×m by a polarization encoding module. n Become 2m n Each channel is then passed through a 3×3 convolution (DWconv3x3) and the GELU activation function, with the channel dimension remaining unchanged. Finally, it undergoes a 1×1 convolution (PWconv1x1) for each pixel, increasing the channel count from 2m. n Change to m n Stokes parameters for each band and Stokes parameters The mixture is then passed through a depthwise 3×3 convolution (DWconv3x3) and the GELU activation function, followed by a pixel-wise 1×1 convolution (PWconv1x1), which further reduces the number of channels from m. n Become 2m n The number of input channels is kept consistent with the total number of input channels. The result is then added to the input and passed through the Tanh activation function to produce the output, which is the first polarization feature. Second polarization characteristics The matrix.

[0033] In this embodiment, the two inputs of the polarization encoding module are obtained from the Malus formula: ,in, .

[0034] To facilitate integration with the output of the spectrometer, another deformation is performed. ;in, To represent a small quantity, then... and Reorganization into Then, the residual is connected to the output of the last pixel's 1×1 convolution (PWconv1x1), and then deformed back to the initial dimension. ,in, This is a learnable parameter, which can be directly set to 1. It is a shape transformation of the matrix, the final result is... to be given respectively and It was later restored to .

[0035] Finally, the physical fusion submodule multiplies the outputs of the polarization information submodule and the spectral information submodule to obtain the final output of the SPG module. The polarization hyperspectral image is calculated according to the following formula: ; In the formula, Indicates the polarization angle as Polarization hyperspectral image, Represents the pixel coordinates of a polarization hyperspectral image. Indicates the center wavelength of the spectral band. Indicates the polarization angle. Spectral characteristics, This is the first polarization feature. This is the second polarization feature.

[0036] It should be noted that, because physical information is introduced into the image reconstruction model, the network loss function used during the training process has been redesigned as follows: ; in, ; .

[0037] In the formula, Let be the network loss function. To reconstruct the loss function, For physical loss function, The weighting coefficients of the physical loss function can be actively selected as hyperparameters. It is a three-dimensional data cube of the polarization hyperspectral image reconstructed by the image reconstruction model. The true value of the three-dimensional data cube of the polarization hyperspectral image. This represents the number of heads in the bulls' self-attention. This is the first polarization feature. This is the second polarization characteristic. This represents a function that takes the maximum value.

[0038] Understandably, the reason for designing the physical loss function in this way is... and These are variations of the Stokes parameters, whose root mean square should be less than 1, so they are designed as penalty terms.

[0039] The computational polarization hyperspectral reconstruction method of this invention acquires a nine-channel mixed measurement signal through a nine-channel polarization broadband image acquisition system. Based on this nine-channel mixed measurement signal, an image reconstruction model can be used to acquire and reconstruct a single-frame polarization hyperspectral image. This method only requires three polarization images acquired in three bands to reconstruct a complete hyperspectral image across all polarization angles. It eliminates the need for complex optical path design and time-consuming wavelength tuning, requires no additional measurement steps, has low requirements for detector specifications, does not require complex optical path configuration, and has a friendly model size, making it easy to deploy on mobile devices / edge devices.

[0040] The image reconstruction model of this invention uses U-Net as its backbone network and incorporates SP and SPG modules to address the polarization hyperspectral reconstruction problem. The SP module employs a residual network architecture to deepen the transmission of input and output information. The SPG module, through a multi-head self-attention Transformer structure for spectral information and a polarization information submodule combining physical encoding and residual networks, fully represents and effectively fuses information from both the spectral and polarization dimensions, ensuring accurate reconstruction at different wavelengths and polarization angles. This significantly shortens reconstruction latency and reduces system complexity, enabling the scheme to adapt to more complex application scenarios and providing greater flexibility and support for engineering applications at the detection end.

[0041] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations are intended to cover non-exclusive inclusion, such that an article or device comprising a list of elements includes not only those elements but also other elements not expressly listed. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the article or device comprising said element. Terms such as "connected" or "linked" are not limited to physical or mechanical connections but can include electrical connections, whether direct or indirect. The orientations or positional relationships indicated by terms such as "upper," "lower," "left," and "right" are based on the orientations or positional relationships shown in the accompanying drawings and are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be construed as limiting the invention.

[0042] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features or characteristics described may be combined in any suitable manner in one or more embodiments or examples. In addition, those skilled in the art can combine and integrate the different embodiments or examples described in this specification.

[0043] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A computational polarimetric hyperspectral reconstruction method, characterized in that, The application relates to a method for reconstructing a polarized hyperspectral image, and belongs to the technical field of image processing. Step 1: obtaining a nine-channel mixed measurement signal of three polarization angles and three spectral band combinations by using a nine-channel polarization wide spectrum image acquisition system; Step 2: inputting the nine-channel mixed measurement signal into a trained image reconstruction model to obtain a polarized hyperspectral image under different polarization angles; The image reconstruction model comprises a feature coding module and a calculation reconstruction module, wherein the feature coding module is used for performing preliminary feature extraction and coding on the input nine-channel mixed measurement signal to obtain initial coding features; the calculation reconstruction module is used for performing calculation reconstruction on the initial coding features to obtain a polarized hyperspectral image under different polarization angles; the calculation reconstruction module comprises a plurality of SP modules adopting a U-Net network architecture, and the SP modules obtain spectral features and polarization features of the hyperspectral image under different polarization angles based on an SPG module and calculate a polarized hyperspectral image under different polarization angles.

2. The computational polarized hyperspectral reconstruction method of claim 1, wherein, The nine-channel polarization wide spectrum image acquisition system comprises an objective lens and a polarization wide spectrum detector; The objective lens is used for collecting scene radiation and imaging on a detection surface of the polarization wide spectrum detector; the polarization wide spectrum detector comprises a plurality of macro-pixels of the same specification, each macro-pixel is arranged with a plurality of sub-pixels, and the sub-pixels are periodically arranged according to at least three different polarization angles and at least three linearly separable spectral bands.

3. The computational polarized hyperspectral reconstruction method of claim 1, wherein, The SP module comprises an SPG module, a layer normalization module and a first convolution module; wherein, The SPG module is used for calculating spectral features and polarization features of a polarized hyperspectral image and calculating a polarized hyperspectral image according to the spectral features and the polarization features; The polarized hyperspectral image is sequentially processed by the layer normalization module and the convolution module and then added to the output of the SPG module as the output of the SP module.

4. The computational polarizing hyperspectral reconstruction method of claim 3, wherein, The first convolution module comprises a connected first 1x1 convolution layer, a first GELU activation function layer, a 3x3 convolution layer, a second GELU activation function layer and a second 1x1 convolution layer.

5. The computational polarized hyperspectral reconstruction method of claim 1, wherein, The SPG module comprises a spectral information submodule, a polarization information submodule and a physical fusion submodule, wherein, The spectrum information submodule adopts a self-attention Transformer architecture, and the Stokes parameters of different wave bands are input based on a Transformer self-attention mechanism Learning and outputting the spectral characteristics of the polarized hyperspectral image ; The polarization information submodule is based on a residual network and physical encoding of different waveband Stokes parameters of the input and Stokes parameters Optimizes and outputs first polarization features of the polarization hyperspectral image and second polarization features ; The physical fusion submodule calculates the polarized hyperspectral image according to the spectral characteristics , the first polarization characteristics , and the second polarization characteristics .

6. The computational polarized hyperspectral reconstruction method of claim 5, wherein, The polarization information submodule comprises a polarization coding module, a second convolution module and a Tanh activation function layer; Wherein, the stokes parameters of different bands and the stokes parameters After dimension compression by the polarization encoding module and convolution processing by the second convolution module in turn, and after adding the output of the polarization encoding module, the Tanh activation function layer is output to output the first polarization feature of the polarized hyperspectral image and the second polarization feature .

7. The computational polarized hyperspectral reconstruction method of claim 6, wherein, The second convolution module comprises a connected first deep 3x3 convolution layer, a third GELU activation function layer, a first pixel 1x1 convolution layer, a second deep 3x3 convolution layer, a fourth GELU activation function layer and a second pixel 1x1 convolution layer.

8. The computational polarized hyperspectral reconstruction method of claim 5, wherein, The polarized hyperspectral image is calculated according to the following formula: ； In the formula, represents a polarized hyperspectral image with a polarization angle of represents a pixel point coordinate of the polarized hyperspectral image, represents a spectral band center wavelength, represents a polarization angle, is a spectral feature, is a first polarization feature, is a second polarization feature.​ 9. The computational polarized hyperspectral reconstruction method of claim 1, wherein, The network loss function adopted by the image reconstruction model in the training process is: ； ； ； In the formula, is a network loss function, is a reconstruction loss function, is a physical loss function, is a weight coefficient of the physical loss function, is a three-dimensional data cube of the polarized hyperspectral image reconstructed by the image reconstruction model, is a three-dimensional data cube true value of the polarized hyperspectral image, is the number of heads in the multi-head self-attention, is a first polarization feature, is a second polarization feature, represents a function of taking the maximum value.

Images