A physical-prior-based encapsulated brief-hyperspectral image inversion method

Abstract

This invention discloses a method for inverting encapsulated bamboo and wooden slips hyperspectral images based on physical priors. The method inputs a hyperspectral observation image affected by glass encapsulation into a pre-trained hyperspectral image reconstruction network for encapsulated bamboo and wooden slips, recovering a clean hyperspectral reflectance image of the slip. This network includes parallel physical estimation and data reconstruction branches. The physical estimation branch predicts pixel-level degradation parameter fields based on an embedded two-color reflectance model using a differentiable physical prior estimation module. The data reconstruction branch performs global spectral morphology estimation and local ink detail recovery through a cascaded reconstruction module. This module also deploys a multi-view feature modeling module to extract multi-view features. The degradation parameter field is used to apply physical prior constraints to the clean hyperspectral reflectance image of the slip, ensuring consistency with the hyperspectral observation image. This invention solves the problems of poor ink readability and low spectral fidelity caused by non-uniform specular reflection and spectral nonlinear distortion due to curved glass tubes.

Description

Technical Field

[0001] This invention relates to the fields of hyperspectral image processing and computer vision, and in particular to a method for inverting encapsulated bamboo and wooden slips hyperspectral images based on physical priors. Background Technology

[0002] Bamboo and wooden slips, as the core document carriers of China from the Warring States period to the Wei and Jin dynasties, systematically recorded the political, economic, and cultural landscape of that era. However, these precious bamboo and wooden documents are highly susceptible to decay and ink fading over the long course of history, and the physical packaging and isolation for security reasons also make conventional academic reading extremely difficult. Therefore, digitization of bamboo and wooden slips is an important task in order to extract and preserve the historical information they carry to the greatest extent possible while protecting the artifacts themselves.

[0003] Hyperspectral imaging (HSI), with its "image-spectrum integration" characteristic, has become a core technology for the digital preservation of cultural heritage and the extraction of faint ink traces. However, precious bamboo and wooden slips are usually encapsulated in nitrogen-filled cylindrical glass tubes for preventative protection, such as... Figure 1 As shown, the existence of this curved glass medium presents a severe optical challenge to hyperspectral imaging: on the one hand, the curved geometry of the glass tube causes complex multiple refractions and reflections of the incident light, resulting in unpredictable ghosting and highly non-uniform specular spots in the image space; on the other hand, the transmittance of the glass medium exhibits nonlinear decay, leading to severe nonlinear distortion and shift in the acquired spectral curves, thereby destroying the inherent physical consistency of the hyperspectral data and severely limiting subsequent ink stain recognition and material identification.

[0004] To address the aforementioned problems, existing technologies can be broadly categorized into three types, but none offer truly effective solutions. The first type is based on single-image dereflection methods. These typically assume that the reflection components are spatially uniform or only vary at low frequencies, failing to handle the non-uniform specular reflection caused by the drastic changes in curvature and incident angle caused by cylindrical glass tubes. Furthermore, processing each band independently can lead to severe spectral shifts. The second type consists of hyperspectral image restoration algorithms, primarily focusing on additive noise removal, blur kernel estimation, or spatial super-resolution. Their degradation models are mostly linear and spatially invariant, unable to model the nonlinear transmission attenuation caused by curved glass and the position-dependent multiple reflections. The third type is based on physical models for dereflection. While theoretically interpretable, most assume known light source spectra or rely on polarization information, lacking adaptive physical parameter estimation capabilities and struggling to adapt to the spatially varying reflection parameters in the hyperspectral data of bamboo and wooden slips.

[0005] In recent years, deep learning algorithms have been attempted for artifact removal and hyperspectral restoration. However, in the specific scenario of glass-encased bamboo slips, the purely data-driven black-box model lacks explicit modeling of the optical degradation process, resulting in poor physical interpretability. Existing methods handle both spatial and spectral dimensions, making it difficult to decouple spatial texture and spectral dependencies. Furthermore, they lack focus on perceptualizing low-reflectivity areas such as ink stains, leading to insufficient spectral reconstruction accuracy. At the same time, direct end-to-end regression in high-dimensional spectral space has high computational overhead and slow convergence, while multi-stage progressive reconstruction strategies have not been fully explored, lacking an effective constraint mechanism for the monotonically decreasing reconstruction error between stages.

[0006] Therefore, there is an urgent need for a hyperspectral inversion method that can remove media interference from the physical source and accurately reconstruct the intrinsic spectra of bamboo and wooden slips. Summary of the Invention

[0007] To address the aforementioned shortcomings of existing technologies, this invention provides a method for inverting encapsulated bamboo and wooden slips hyperspectral images based on physical priors. By embedding a two-color reflectance model into a differentiable physical prior estimation module, and combining multi-view feature modeling and coarse-to-fine cascaded reconstruction, this method solves the problems of poor ink readability and low spectral fidelity caused by non-uniform specular reflection and spectral nonlinear distortion due to curved glass tubes in hyperspectral images of glass-encapsulated bamboo and wooden slips. This method achieves high-fidelity recovery of pure intrinsic reflectance images of bamboo and wooden slips from hyperspectral observations interfered with by glass.

[0008] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0009] A method for inverting hyperspectral images of encapsulated bamboo and wooden slips based on physical priors is proposed. This method acquires hyperspectral observation images that are interfered with by glass encapsulation. The hyperspectral observation images interfered with by glass encapsulation are then input into a pre-trained hyperspectral image reconstruction network for encapsulated bamboo and wooden slips to perform spectral inversion, thereby obtaining a pure hyperspectral reflectance image of the bamboo and wooden slips.

[0010] The encapsulated bamboo and wooden slips hyperspectral image reconstruction network includes parallel physical estimation and data reconstruction branches. The network transmits the input hyperspectral observation image to both the physical estimation and data reconstruction branches. The physical estimation branch predicts pixel-level degradation parameter fields based on an embedded two-color reflectance model using a differentiable physical prior estimation module. The data reconstruction branch performs coarse-to-fine global spectral morphology estimation and local ink detail restoration through a cascaded reconstruction module. At least one stage of the cascaded reconstruction module deploys a multi-view feature modeling module, which is used to extract spatial texture features, long-range spectral dependence features, and spectral gradient difference features in parallel, finally generating a clean hyperspectral reflectance image of the bamboo and wooden slips as output.

[0011] The degradation parameter field is used to apply physical prior constraints to the pure bamboo slip hyperspectral reflectance image output by the data reconstruction branch, so that the pure bamboo slip hyperspectral reflectance image is consistent with the hyperspectral observation image under the forward physical model simulation.

[0012] As a preferred embodiment, the forward physical model is based on an embedded two-color reflection model and a curved microcavity multiple reflection approximation, and generates a simulated glass observation image according to the following formula:

[0013] ;

[0014] In the formula: This is a simulated glass observation image obtained from the hyperspectral image of pure bamboo slips using a forward physical model; High spectrum of pure bamboo and wooden slips; This is element-wise multiplication.

[0015] As a preferred embodiment, the degradation parameter field predicted by the differentiable physics prior estimation module includes the bidirectional transmission attenuation coefficient. First-order reflection coupling coefficient Second-order nonlinear degradation coefficient and environment bias tensor The range of values ​​for the degradation parameter field is constrained by a nonlinear activation function: the Sigmoid function is used to... The constraint is within the interval (0,1), and the Tanh function is used to... Constraints Within the interval, These are preset positive numbers.

[0016] As a preferred embodiment, the differentiable physical prior estimation module further includes a physical inverse model analytical calculation module. This module derives the inverse analytical expression based on the degenerate parameter field and constrains the output of the data reconstruction branch using physical coupling loss. The inverse analytical expression is:

[0017] ;

[0018] in, ;

[0019] In the formula: The input is a hyperspectral observation image affected by glass encapsulation interference; The effective light intensity after removing environmental bias; This is the physical estimate of the hyperspectral reflectance image of the pure bamboo slips, calculated by inverse analytical formula.

[0020] As a preferred embodiment, the processing procedure of the multi-view feature modeling module is as follows:

[0021] The system performs spatial, spectral, and gradient branches in parallel on the input features. The spatial branch first undergoes 3D convolution and normalized activation, then uses large-kernel depthwise separable convolution to extract spatial texture features. The spectral branch first undergoes 3D convolution and normalized activation, then uses a spectral multi-head self-attention mechanism to extract long-range spectral dependency features, followed by another 3D convolution and normalized activation. The gradient branch extracts spectral gradient difference features through spectral gradient blocks. The output features of the spatial, spectral, and gradient branches are concatenated along the channel dimension and input into the global context module for adaptive feature recalibration and fusion to obtain fused multi-view features. Finally, the features are decoupled to output enhanced multi-view features.

[0022] As a preferred embodiment, the processing procedure for the spatial branch is expressed as follows:

[0023] ;

[0024] ;

[0025] In the formula: These are intermediate features of spatial branches; The input feature tensor; Normalize the group; This is a three-dimensional convolution with a kernel size of 1×3×3; For activation functions; Spatial texture features output by spatial branches; Large kernel depthwise convolution; Use the GELU activation function; For pointwise convolution;

[0026] The processing procedure for the spectral branch is expressed as follows:

[0027] ;

[0028] ;

[0029] In the formula: This represents an intermediate feature of the spectral branch; This represents the long-range spectral dependence characteristic of the spectral branch output; This is a three-dimensional convolution with a kernel size of 3×3×1; For spectral multi-head self-attention;

[0030] The gradient branching process is represented as follows:

[0031] ;

[0032] In the formula: This represents the spectral gradient difference feature output by the gradient branch.

[0033] As a preferred embodiment, the processing procedure of the global context module is expressed as follows:

[0034] ;

[0035] ;

[0036] In the formula: The resulting multi-view features; Indicates concatenation of channel dimensions; This is a three-dimensional global average pooling method. and All are linear mapping weights in the global context module; This indicates multiplication by channel; Use the Sigmoid activation function; Enhanced multi-view features output by the global context module.

[0037] As a preferred embodiment, the cascaded reconstruction module includes an encoder, a coarse-stage reconstruction, a feature fusion module, a fine-stage reconstruction, and a decoder. The coarse-stage reconstruction includes a cascaded first feature extraction block, a max-pooling layer, a second feature extraction block, and an upsampling layer. The fine-stage reconstruction includes a cascaded third feature extraction block and a fourth feature extraction block. The encoder is used to encode the input hyperspectral observation image step-by-step, extracting multi-level spatial-spectral features. The decoder receives the output of the fourth feature extraction block.

[0038] In the cascaded reconstruction module, the input hyperspectral observation image is reconstructed through a coarse stage to output a preliminary hyperspectral reflectance image of the bamboo and wooden slips. Then, the output of the coarse stage reconstruction is fused with the multi-level spatial spectral features extracted by the encoder in the feature fusion module and used as the input for the fine stage reconstruction to restore the high-frequency ink details at the original resolution. Finally, the decoder outputs a clean hyperspectral reflectance image of the bamboo and wooden slips.

[0039] Among them, at least one feature extraction module is equipped with the multi-view feature modeling module, which is used to perform parallel extraction and adaptive fusion of spatial texture features, long-range spectral dependence features and spectral gradient difference features of the current stage features, and to use the enhanced multi-view features to perform spectral morphology estimation or ink detail restoration of the current stage.

[0040] As a preferred embodiment, the training process of the encapsulated bamboo and wooden slips hyperspectral image reconstruction network is as follows:

[0041] A pre-constructed hyperspectral image dataset of glass-encased bamboo and wooden slips is used as a training dataset and input into the hyperspectral image reconstruction network of the encased bamboo and wooden slips. A joint multi-task loss function consisting of spatial reconstruction constraints, spectral fidelity constraints, phase progressive constraints, and physical prior constraints is constructed. The training loss of the hyperspectral image reconstruction network of the encased bamboo and wooden slips is calculated. The parameters of the network are updated with the goal of minimizing the joint multi-task loss function, thereby training the hyperspectral image reconstruction network of the encased bamboo and wooden slips.

[0042] The training dataset includes hyperspectral observation images affected by glass encapsulation and their corresponding labels for pure bamboo slip hyperspectral reflectance images.

[0043] As a preferred embodiment, the joint multi-task loss function is expressed as:

[0044] ;

[0045] In the formula: For joint multi-task loss function; Constraints for spatial reconstruction; For spectral fidelity constraints; As a phased, gradual constraint; Physical prior constraints;

[0046] The spatial reconstruction constraints include a global reconstruction loss and an ink-perception weighted loss, which are expressed as follows:

[0047] ;

[0048] ;

[0049] ;

[0050] In the formula: This represents a loss during the overall reconstruction. Weighted loss for ink perception; and These are the weighting coefficients for the global reconstruction loss and the ink perception weighted loss, respectively. A mask for the ink area; Adaptive convex penalty weights; To obtain a clean hyperspectral reflectance image of bamboo slips from a network that reconstructs hyperspectral images of bamboo slips, at the pixel level. The value at; This is a real label; It is an L1 norm; It is a small constant; For focusing parameters; For pixels Average reflectance in the spectral dimension;

[0051] The spectral fidelity constraint employs spectral angle mapping loss, which is expressed as:

[0052] ;

[0053] In the formula: For spectral angle mapping loss;

[0054] The stage-wise asymptotic constraints include coarse-stage supervision loss and inter-stage ranking loss, which are expressed as follows:

[0055] ;

[0056] ;

[0057] In the formula: Losses due to rough-stage monitoring; The loss is the inter-level ranking loss. and These are the weighting coefficients for the coarse-stage supervision loss and the inter-level ranking loss, respectively. Preliminary hyperspectral reflectance images of bamboo and wooden slips output from the coarse-stage reconstruction; These are boundary hyperparameters;

[0058] The physical prior constraints include forward physical consistency loss and reverse coupling loss, which are expressed as follows:

[0059] ;

[0060] ;

[0061] ;

[0062] In the formula: This represents a positive physical consistency loss. This is the inverse coupling loss; and These are the weighting coefficients for the forward physical consistency loss and the reverse coupling loss, respectively; This is a simulated glass observation image obtained from the hyperspectral image of pure bamboo slips using a forward physical model; The input is a hyperspectral observation image affected by glass encapsulation interference; The clean hyperspectral reflectance image of the bamboo slips output by the data reconstruction branch; This is the physical estimate of the hyperspectral reflectance image of the pure bamboo slips, calculated by inverse analytical formula.

[0063] Compared with the prior art, the present invention has the following technical effects:

[0064] 1. This invention achieves adaptive prediction of pixel-level degradation parameter fields (transmittance, reflection coupling coefficient, and environmental bias) by embedding a dual-color reflectance model and designing a differentiable physical prior estimation module. This module abandons the strong assumptions of "uniform reflection components" or "linear invariance of the degradation model" in traditional methods, and can accurately fit the spatial non-uniform specular reflection and nonlinear spectral distortion caused by the cylindrical glass tube. At the same time, due to the differentiability of this module, physical laws can be integrated into the end-to-end training of the deep learning network, which not only ensures the physical interpretability of the inverse solution, but also avoids the spectral shift and artifact residue that are prone to occur in pure black box models. On this basis, through the dual constraints of forward physical consistency loss and inverse coupling loss, the reconstruction results simultaneously satisfy the physical degradation model and data observation, significantly improving the visual discernibility and spectral fidelity of the encapsulated bamboo and wooden slip hyperspectral images.

[0065] 2. The multi-view feature modeling module of this invention extracts three types of features in parallel: spatial texture, long-range spectral dependence, and spectral gradient difference. These features are then adaptively fused through a global context module. This design effectively decouples the high coupling between spatial and spectral characteristics in hyperspectral data. Furthermore, this invention adopts a coarse-to-fine cascaded reconstruction architecture. In the coarse stage, the global spectral morphology is quickly estimated based on the fused features. In the fine stage, high-frequency ink blot details are restored at the original resolution, and an inter-stage ranking loss is introduced to force a monotonically decreasing error in the fine stage. Simultaneously, the synergistic effect of multi-view feature modeling and cascaded reconstruction enables the network to prioritize the accuracy of the macroscopic spectral morphology while suppressing glass interference, and then gradually refine the local ink blot texture. This avoids the problem of getting trapped in local optima and slow convergence in direct end-to-end regression in high-dimensional spectral space, achieving stable and high-quality hyperspectral image inversion.

[0066] 3. The joint multi-task loss function constructed in this invention supervises the network from four dimensions: spatial, spectral, stage-wise progressive, and physical prior. Among them, the ink-stain perception weighted loss guides the network to focus on low-reflectivity ink-stain areas through inverse brightness weights, avoiding overfitting on bamboo backgrounds and significantly improving the clarity of faded ink-stain recovery; the spectral angle loss directly constrains the shape of the reconstructed spectral curve, ensuring the high fidelity required for material spectral analysis; the inter-stage ranking loss ensures that the fine stage in the cascaded structure achieves a quantifiable performance improvement compared to the coarse stage, making progressive training more robust; and the physical consistency loss and reverse coupling loss ensure that the data-driven results never deviate from the physical degradation model.

[0067] 4. The SealGlassHyperNet (SGHNet) proposed in this invention is designed for the special scenario of glass-encapsulated bamboo and wooden slips. It removes the non-uniform reflection, ghosting and spectral distortion introduced by the curved glass tube from the physical source. The restored pure hyperspectral reflectance image of the bamboo and wooden slips not only has clear ink marks, but also achieves consistency between the spectral curve shape and the material of the real bamboo and wooden slips. Attached Figure Description

[0068] To make the objectives, technical solutions, and advantages of the invention clearer, the invention will now be described in further detail with reference to the accompanying drawings, wherein:

[0069] Figure 1 This is a comparison image of glass tube distortion in this invention;

[0070] Figure 2 This is a diagram of the encapsulated bamboo and wooden slips hyperspectral image reconstruction network framework used in the method disclosed in this invention;

[0071] Figure 3 This is a framework diagram of the multi-view feature modeling module in an embodiment of the present invention. Detailed Implementation

[0072] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0073] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of 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 a part of the embodiments of the present invention, not all of them. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to represent selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0074] It should be noted that similar reference numerals and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the figures, or the orientation or positional relationship commonly used when the product is in use. They are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Furthermore, the terms "first," "second," and "third," etc., are only used to distinguish descriptions and should not be construed as indicating or implying relative importance. In addition, the terms "horizontal," "vertical," etc., do not indicate that the component is required to be absolutely horizontal or suspended, but can be slightly tilted. For example, "horizontal" simply means that its direction is more horizontal than "vertical," and does not mean that the structure must be completely horizontal, but can be slightly tilted. In the description of this invention, it should also be noted that, unless otherwise explicitly specified and limited, the terms "set," "install," "connect," and "link" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0075] Example:

[0076] For hyperspectral imaging of glass-encased bamboo slips, the curved geometry of the cylindrical glass tubes severely impacts the acquired images due to non-uniform specular reflection, ghosting interference, and nonlinear spectral distortion, disrupting the physical consistency of the hyperspectral data and making subsequent ink mark recognition and material identification difficult. Current restoration models largely rely on simplified degradation assumptions, such as additive Gaussian noise, stripe noise, or spatially invariant linear fuzzy kernels. In real-world curved glass encapsulation scenarios, light undergoes cross-band nonlinear absorption and transmission attenuation as it passes through the curved medium. Existing hyperspectral restoration methods generally lack explicit modeling of this nonlinear, spatially varying reflection interference. This leads them to easily overlook the tight physical correlations between high-dimensional bands when processing encapsulated bamboo slip data, resulting in irreversible spectral shifts and failing to meet the stringent requirements of high-fidelity spectral inversion. To address this, the present invention proposes a method for hyperspectral image inversion of encapsulated bamboo and wooden slips based on physical priors. This method inputs the acquired hyperspectral observation image, which is interfered with by the glass encapsulation, into a pre-trained hyperspectral image reconstruction network for encapsulated bamboo and wooden slips to perform spectral inversion, obtaining a clean hyperspectral reflectance image of the bamboo and wooden slips. The hyperspectral image reconstruction network for encapsulated bamboo and wooden slips includes parallel physical estimation branches and data reconstruction branches. The network transmits the input hyperspectral observation image to the physical estimation branch and the data reconstruction branch respectively. The physical estimation branch predicts pixel-level degradation parameters based on an embedded two-color reflectance model using a differentiable physical prior estimation module. The data reconstruction branch performs coarse-to-fine global spectral morphology estimation and local ink detail restoration through a cascaded reconstruction module. At least one stage of the cascaded reconstruction module deploys a multi-view feature modeling module, which is used to extract spatial texture features, long-range spectral dependence features, and spectral gradient difference features in parallel, finally generating a clean bamboo slip hyperspectral reflectance image as output. The degradation parameter field is used to apply physical prior constraints to the clean bamboo slip hyperspectral reflectance image output by the data reconstruction branch, ensuring that the clean bamboo slip hyperspectral reflectance image remains consistent with the hyperspectral observation image under forward physical model simulation.

[0077] Specifically, this invention first constructs a differentiable physical prior estimation module, explicitly embedding a two-color reflectance model into the network for parameterizing the inversion of spatially non-uniform specular reflection and nonlinear spectral distortion, thereby avoiding purely data-driven black-box mapping. Second, in terms of feature extraction, a multi-view feature modeling module is proposed to extract spatial texture, long-range spectral dependence, and spectral gradient difference information respectively, achieving decoupling and enhancement of weak ink stain edges and continuous spectral features. In addition, in terms of loss function, an ink stain perception loss based on inverse brightness weights is constructed to guide the network to focus on low-reflectivity ink stain regions and suppress overfitting in bamboo background regions. At the same time, to address the stability problem of high-dimensional data reconstruction, a coarse-to-fine cascaded architecture is adopted, first restoring the global spectral morphology, then progressively refining high-frequency ink stain details, and introducing inter-level ranking loss to force the error of the fine stage to decrease monotonically, thereby ensuring the effectiveness of the progressive optimization strategy.

[0078] The following is a more detailed description of the present invention's method for inverting encapsulated bamboo and wooden slips using a physical prior.

[0079] 1. Physical degradation modeling

[0080] In this embodiment, a forward degradation model of the hyperspectral image of the glass-encapsulated bamboo slips is first established as the physical basis for the subsequent design of the hyperspectral image reconstruction network for the encapsulated bamboo slips.

[0081] 1.1 Two-color reflection model

[0082] This application is based on the Dichromatic Reflection Model (DRM). This model states that the intensity of light reflected from an object's surface and entering the camera... The diffuse reflection component ) and specular reflection It is formed by linear superposition, and can be expressed as:

[0083] ;

[0084] Among them, diffuse reflection component It carries the intrinsic material and spectral properties of the object, while the specular reflection component... The spectral composition is mainly dominated by the light source. In the case of glass-encapsulated bamboo and wooden slips, the diffuse reflection on the surface of the slips reflects the true optical characteristics of the bamboo and wooden substrate and the ink marks, while the specular reflection generated by the inner and outer walls of the glass tube appears as interfering highlights.

[0085] 1.2 Multiple Reflections and Nonlinear Degradation in Curved Microcavities

[0086] In a glass-encapsulated scenario, external light must first pass through a glass tube to reach the surface of the bamboo slip. After absorption, scattering, and reflection on the surface, it passes through the glass medium again to enter the camera. Let... The bidirectional transmission attenuation coefficient represents the spatial-spectral variation. This represents the hyperspectral image of a pure bamboo slip. Therefore, the observed component is a single reflection (one-time reflection) that did not undergo multiple bounces from the inner wall of the glass. It can be represented as:

[0087] ;

[0088] On the other hand, due to the air or nitrogen gaps between the inner and outer walls of the glass tube and the surface of the bamboo slip, a curved microcavity is formed, causing multiple position-dependent reflections of light at the interface between the media. Inspired by the multiple reflection attenuation law proposed by Nayar et al., the multiple reflections within the cavity are typically mathematically expressed as an infinite geometric series. This embodiment first introduces a first-order reflection coupling coefficient. This term describes the space-spectral correlated reflection coupling strength caused by variations in glass curvature, refractive index, surface roughness, and local incident angle. Each additional reflection or refracting of light within the cavity results in a different spatial-spectral correlation coupling strength. The energy decays, therefore, the first Secondary reflection component It can be derived from the previous reflection, and its formula is expressed as follows:

[0089] ;

[0090] Under weak coupling conditions Below, the total diffuse reflection inside an ideal cylindrical glass tube Contribution is everything Summation of an infinite series of secondary reflection components:

[0091] ;

[0092] Among them Expanding this geometric series, we get:

[0093] ;

[0094] When the series is truncated to the second-order terms, a second-order approximation of multiple reflections can be obtained:

[0095] ;

[0096] However, in the actual imaging process of a cylindrical glass tube, the second-order nonlinear degradation does not solely originate from the square of the first-order reflection coupling term. Changes in the refraction path caused by the curved glass, differences in the thickness of the inner and outer walls, local specular highlights, ambient stray light, and differences in transmittance across different wavelengths all introduce additional intensity-dependent nonlinear residuals. Just as Jensen et al. used an approximate model to handle complex subsurface light transmission, these complex factors make it difficult to establish a complete closed-loop analytical model. Therefore, inspired by the guiding principles of physics-informed machine learning, this embodiment does not assume a complete microscopic analytical model for the higher-order light transmission process. Instead, based on the first-order multiple reflection approximation, it introduces a learnable and effective second-order coupling coefficient. Used for absorption And other second-order residuals that are not explicitly modeled. From a series perspective, this can be expressed as:

[0097] ;

[0098] in, This represents squaring element by element. Therefore, in the denominator... This can be considered an effective approximation of the second-order nonlinear residuals in the degradation of real glass encapsulation, rather than an independently measured material constant. Since all hyperspectral cubes in this embodiment are normalized to... , and This can be considered as a dimensionless effective coupling parameter. Therefore, the second-order approximate degenerate model used in this embodiment can be obtained:

[0099] ;

[0100] in, This represents the degraded observation component formed after the reflected signal from the bamboo or wooden slip body is transmitted through the glass medium, undergoes multiple reflections, and is coupled by second-order nonlinearity. The bidirectional transmission attenuation coefficient represents the spatial-spectral variation. Represents the first-order reflection coupling coefficient. This represents the second-order nonlinear degradation coefficient. These are all pixel-level degradation parameter fields with spatial spectral variations. It should be noted that... These are not strict physical constants, but rather empirical approximations used to fit various unmodeled higher-order effects in real curved glass tubes, such as multiple reflections from non-parallel walls, asymptotic refraction, and scattering. This second-order truncated form does not pursue a complete physical closed-form solution, but rather serves as a learnable bias term in the data-driven network to enhance the model's ability to express complex nonlinear degradation. This idea of ​​absorbing unexplicitly modeled errors in real imaging systems through learnable approximations aligns with recent research trends in computational spectral imaging. For example, Gualdrón-Hurtado et al. introduced a regularization term from a deep learning nonlinear propagation model in compressed spectral imaging to characterize nonlinear propagation errors in complex optical systems. Inspired by this, this embodiment does not aim to establish a complete closed-form analysis of all higher-order light propagation processes inside the curved glass tube, but rather uses learnable second-order nonlinear terms to effectively approximate the residuals of local reflection, refraction, and multiple reflections.

[0101] As for the specular reflection component The ghosting is mainly caused by strong specular spots generated by the light source on the curved glass outer wall and stray ambient light. Combining this with the dark current noise of the sensor, this embodiment abstracts these spatially non-uniform additive interferences into a pixel-level environmental bias tensor. Based on the above deductions, the positive degradation model is established as follows:

[0102] ;

[0103] In the formula: This is a simulated glass observation image obtained from the hyperspectral image of pure bamboo slips using a forward physical model; High spectrum of pure bamboo and wooden slips; This is element-wise multiplication.

[0104] 1.3 Maximum A posteriori estimation

[0105] To transform the aforementioned physical degradation model into an optimizable algorithmic framework, this embodiment models the glass-encapsulated hyperspectral restoration problem from the perspective of Maximum A Posteriori (MAP) estimation. Given hyperspectral observations interfered with by the glass tube... (i.e., input image) The goal is to recover a pure hyperspectral reflectance image of the bamboo and wooden slips themselves. (i.e., output image) ),in Indicates the number of spectral bands. and Represents the spatial dimensions and the corresponding degenerate parameter field. According to Bayes' theorem, the problem of maximizing the posterior probability can be equivalently transformed into minimizing the energy functional problem:

[0106] ;

[0107] In the formula: To obtain through optimization and The optimal estimate; This represents the observation consistency term, requiring the reconstruction results to be reconstructed. After simulation using a physical degradation model, it can be compared with real glass observations. To maintain consistency. In this embodiment's architecture, this item consists of the physics estimation branch and the forward simulation loss ( This is implemented to constrain the network output to conform to the degradation laws described by glass transmission, multiple reflections, and environmental bias. This represents the target image prior, used to constrain the restoration result to have reasonable spatial texture, spectral continuity, and ink mark edge structure. In this embodiment, the implicit image prior is learned through the backbone reconstruction network, enabling the model to remove glass interference while preserving as much as possible the subtle reflection differences between the bamboo and wooden substrate and the ink marks on the bamboo slips. Represents prior physical parameters used for constraints. , , , The range of values ​​and physical rationality of the parameters are considered. In specific implementation, this embodiment restricts the parameter space through a nonlinear activation function: the Sigmoid function is used to optimize the equivalent transmission coefficient. First-order reflection coupling parameters and second-order empirical nonlinear parameters The constraint is within a finite positive interval, and the environment bias term is adjusted using Tanh. Limited to a finite scope. It should be noted that, although... It is an empirically valid parameter used to absorb unexplicitly modeled second-order residuals, rather than a directly measured material constant, but it is still subject to boundary constraints to avoid unbounded optimization and improve numerical stability.

[0108] The above correspond to the forward physics simulation loss, the structural prior of the data reconstruction branch, and the nonlinear activation function constraint at the output of the differentiable physics prior estimation module in the subsequent network design of this embodiment.

[0109] 2. Encapsulated Hyperspectral Image Reconstruction Network for Bamboo Slips

[0110] Based on the MAP estimation objective determined above, this embodiment encapsulates the Bamboo Slip Hyperspectral Image Reconstruction Network (SGHNet) with a data-physical dual-branch collaborative architecture, which learns degradation parameter estimation and hyperspectral image reconstruction simultaneously in an end-to-end joint training manner.

[0111] like Figure 2As shown, the encapsulated bamboo and wooden slip hyperspectral image reconstruction network includes parallel physical estimation branches and data reconstruction branches. The network transmits the input hyperspectral observation image to the physical estimation branch and the data reconstruction branch, respectively. The physical estimation branch predicts pixel-level degradation parameter fields based on an embedded two-color reflectance model using a differentiable physical prior estimation module. The data reconstruction branch performs coarse-to-fine global spectral morphology estimation and local ink detail restoration through a cascaded reconstruction module. At least one stage of the cascaded reconstruction module deploys a multi-view feature modeling module, which is used to extract spatial texture features, long-range spectral dependence features, and spectral gradient difference features in parallel, finally generating a clean bamboo and wooden slip hyperspectral reflectance image as the output. The degradation parameter field is used to apply physical prior constraints to the clean bamboo and wooden slip hyperspectral reflectance image output by the data reconstruction branch, so that the clean bamboo and wooden slip hyperspectral reflectance image is consistent with the hyperspectral observation image under forward physical model simulation.

[0112] The data-physical dual-branch collaborative architecture of this embodiment will be described in detail below.

[0113] 2.1 Physical Estimation Branch

[0114] In the physics estimation branch, pixel-level degradation parameter fields are adaptively predicted from hyperspectral observation images affected by the glass tube interference. These parameters characterize the spatially non-uniform transmission attenuation, multiple reflections, and nonlinear distortion caused by the curved glass tube. To achieve this goal, this embodiment designs a differentiable physics prior estimation module. This module can embed the physics degradation model into the network in a differentiable manner and constrain the output parameters to conform to preset physical boundaries through a nonlinear activation function, providing reliable parameter basis for subsequent forward physics consistency constraints and inverse coupling constraints.

[0115] 2.1.1 Differentiable Physics Prior Estimation Module

[0116] To ensure the rigorous physical validity of the predicted hidden layer tensors in the physics estimation branch, the differentiable physics prior estimation module imposes nonlinear boundary constraints on the network output. Indicates the bidirectional transmission attenuation coefficient. Represents the first-order reflection coupling coefficient. This represents the second-order nonlinear degradation coefficient, an empirical approximation parameter used to compensate for higher-order nonlinear residuals caused by curved glass surfaces. This represents an additive bias term indicating spatial variation. In this embodiment, we set... And use the Sigmoid function constraint The range of values ​​is constrained by the Tanh function. The amplitude range. This design can prevent physical parameters from degenerating into unbounded free variables, thereby reducing the feasible solution space of the ill-conditioned inverse problem.

[0117] This module also includes a built-in physical inverse model analytical calculation submodule. This module derives the inverse analytical expression based on the degenerate parameter field and constrains the output of the data reconstruction branch through physical coupling loss; its inverse analytical expression is:

[0118] ;

[0119] in, ;

[0120] In the formula: The input is a hyperspectral observation image affected by glass encapsulation interference; The effective light intensity after removing environmental bias; This is a physical estimate of the hyperspectral reflectance image of the pure bamboo slips calculated by inverse analytical derivation; in this embodiment, the discriminant inside the square root of the formula... Because the network activation function establishes The physical boundary, and Therefore, the inequality The condition holds true for all cases; therefore, the composite square root function... It is continuously differentiable within its valid domain. The denominator in the above formula... Similarly, it can be proven that, since all terms are non-negative, the denominator has a positive lower bound. The proof of this lower bound mathematically reduces division-by-zero overflow during forward computation. (and the risk of gradient explosion during backpropagation).

[0121] 2.2 Data Reconstruction Branch

[0122] In the data reconstruction branch, its core role is to extract data from hyperspectral observation images affected by glass tube interference. Recovering pure high-spectral reflectance images of bamboo slips This branch employs a coarse-to-fine cascaded reconstruction strategy. First, it rapidly estimates the global spectral morphology in the coarse stage, then gradually recovers high-frequency ink blot details in the fine stage. This mitigates the problem of slow convergence and local optima that can easily occur with direct end-to-end regression in high-dimensional spectral spaces. To achieve this, the data reconstruction branch mainly consists of a cascaded reconstruction module and a multi-view feature modeling module deployed within it. The cascaded reconstruction module provides the encoder-decoder macro-architecture and the coarse-to-fine progressive reconstruction logic, ensuring that the network focuses on global structure recovery and local detail refinement at different stages. The multi-view feature modeling module is embedded within each level of feature extraction blocks. By extracting spatial texture, long-range spectral dependence, and spectral gradient difference features in parallel, it achieves efficient decoupling and enhancement of spatial-spectral information, providing rich and less redundant feature representations for cascaded reconstruction. The two modules work together to support high-quality spectral inversion in the data reconstruction branch.

[0123] The cascaded reconstruction module and the multi-view feature modeling module will be explained in detail below.

[0124] 2.2.1 Cascaded Reconstruction Module

[0125] In this embodiment, the cascaded reconstruction module includes an encoder, a coarse-stage reconstruction, a feature fusion module, a fine-stage reconstruction, and a decoder. The coarse-stage reconstruction includes a cascaded first feature extraction block, a max-pooling layer, a second feature extraction block, and an upsampling layer. The fine-stage reconstruction includes a cascaded third feature extraction block and a fourth feature extraction block. The encoder is used to encode the input hyperspectral observation image step-by-step, extracting multi-level spatial-spectral features. The decoder receives the output of the fourth feature extraction block.

[0126] In the cascaded reconstruction module, the input hyperspectral observation image is reconstructed through a coarse stage to output a preliminary hyperspectral reflectance image of the bamboo and wooden slips. Then, the output of the coarse stage reconstruction is fused with the multi-level spatial spectral features extracted by the encoder in the feature fusion module and used as the input for the fine stage reconstruction to restore the high-frequency ink details at the original resolution. Finally, the decoder outputs a clean hyperspectral reflectance image of the bamboo and wooden slips.

[0127] Among them, at least one feature extraction module is equipped with the multi-view feature modeling module, which is used to perform parallel extraction and adaptive fusion of spatial texture features, long-range spectral dependence features and spectral gradient difference features of the current stage features, and to use the enhanced multi-view features to perform spectral morphology estimation or ink detail restoration of the current stage.

[0128] 2.2.2 Multi-view Feature Modeling Module

[0129] Hyperspectral data inherently possesses a high degree of coupling between spatial and spectral features. Standard 3D convolution, when processing hyperspectral data, can easily dilute fragile spatial textures (such as ink blot edges) with spectral dimensions, or disrupt continuous spectral features due to localized spatial receptive fields. To address this issue, this embodiment designs a multi-view feature modeling module that explicitly separates and enhances spatial, spectral, and gradient features through a parallel architecture.

[0130] In this embodiment, as Figure 3 As shown, the processing procedure of this multi-view feature modeling module is as follows: Input features... (in For batch size, The number of feature channels, This represents the number of spectral bands. The system executes spatial, spectral, and gradient branches in parallel (to achieve spatial resolution). The spatial branch first undergoes 3D convolution and normalized activation, then uses large-kernel depthwise separable convolution to extract spatial texture features. The spectral branch first undergoes 3D convolution and normalized activation, then uses a spectral multi-head self-attention mechanism to extract long-range spectral dependency features, followed by another 3D convolution and normalized activation. The gradient branch extracts spectral gradient difference features through spectral gradient blocks. The output features of the spatial, spectral, and gradient branches are concatenated along the channel dimension and input into the global context module for adaptive feature recalibration and fusion to obtain fused multi-view features. Finally, the features are decoupled to output enhanced multi-view features.

[0131] Spatial Branch: This branch aims to capture the topology and macroscopic spatial structure of large-area specular spots caused by curved surface geometry. To obtain a broad receptive field while controlling the number of parameters, this embodiment employs a 7×7 large-kernel depthwise separable convolution. Through a dimensionality recombination mechanism, the two-dimensional large-kernel convolution is independently applied to each spectral channel. This not only efficiently captures long-range dependencies in the spatial domain but also fundamentally avoids cross-contamination of features across bands. The calculation process can be represented as follows:

[0132] ;

[0133] ;

[0134] In the formula: These are intermediate features of spatial branches; The input feature tensor; Normalize the group; This is a three-dimensional convolution with a kernel size of 1×3×3; For activation functions; Spatial texture features output by spatial branches; Large kernel depthwise convolution; Use the GELU activation function; For pointwise convolution;

[0135] Spectral Branch: Nonlinear absorption in glass media induces global spectral shifts. To address this, this branch introduces a spectral multi-head self-attention mechanism (Spectral MHSA). Unlike conventional visual attention, this mechanism treats the flattened spatial dimension as a batch and the spectral band dimension S as the sequence length. By calculating the attention-scaled dot product matrix between bands, it adaptively aggregates long-range spectral dependencies. This design allows the network to utilize information from less disturbed bands to assist in reconstructing severely damaged bands. The process can be represented as:

[0136] ;

[0137] ;

[0138] In the formula: This represents an intermediate feature of the spectral branch; This represents the long-range spectral dependence characteristic of the spectral branch output; This is a three-dimensional convolution with a kernel size of 3×3×1; For spectral multi-head self-attention;

[0139] Gradient Branch: This branch is dedicated to extracting local differential signals between adjacent bands. In this embodiment, a 3D convolution operator with a kernel size of (3,1,1) is deployed, using asymmetric padding (0,1,1) to ensure that the operator's spatial receptive field is limited to a single pixel. This branch learns the variation patterns between adjacent bands through local 3D convolutions unfolded only along the spectral dimension, approximating local inter-spectral differences and thus enhancing the response to ink blot edges and weakly textured regions. The process can be represented as:

[0140] ;

[0141] In the formula: This represents the spectral gradient difference feature output by the gradient branch.

[0142] Finally, the three sets of features are concatenated along the channel dimension and input into the Global Context Block (GCB). The GCB first compresses the massive spatial-spectral features using 3D adaptive average pooling, then generates channel-level attention weights via a bottleneck multilayer perceptron and a sigmoid function. The output features are summed as residuals, completing the global recalibration and adaptive fusion of multi-source physical and data features. The process can be represented as:

[0143] ;

[0144] ;

[0145] In the formula: The resulting multi-view features; Indicates concatenation of channel dimensions; This is a three-dimensional global average pooling method. and All are linear mapping weights in the global context module; This indicates multiplication by channel; Use the Sigmoid activation function; This is the fused multi-view feature output by the global context module.

[0146] 3. Training of the network for reconstructing hyperspectral images of bamboo and wooden slips

[0147] The pre-constructed hyperspectral image dataset SealGlass-HSI, consisting of hyperspectral observation images affected by glass encapsulation and their corresponding labels of clean hyperspectral reflectance images of the bamboo and wooden slips, is used as the training dataset. This dataset is input into the hyperspectral image reconstruction network for the bamboo and wooden slips. A joint multi-task loss function, consisting of spatial reconstruction constraints, spectral fidelity constraints, phased progressive constraints, and physical prior constraints, is constructed. The training loss of the hyperspectral image reconstruction network for the bamboo and wooden slips is calculated, and the network parameters are updated with the goal of minimizing the joint multi-task loss function, thereby training the hyperspectral image reconstruction network for the bamboo and wooden slips.

[0148] In this embodiment, the training objective is divided into four parts: spatial reconstruction constraint, spectral fidelity constraint, phased progressive constraint, and physical prior constraint.

[0149] 3.1 Spatial Reconstruction Constraints

[0150] First, spatial reconstruction constraints are used to ensure the overall strength consistency of the reconstruction results in the spatial domain and to enhance the network's ability to recover low-reflectivity ink blot regions. This part consists of a global reconstruction loss and an ink blot perception weighted loss:

[0151] ;

[0152] In the formula: Constraints for spatial reconstruction; This represents a loss during the overall reconstruction. Weighted loss for ink perception; and These are the weighting coefficients for the global reconstruction loss and the ink-perception weighted loss, respectively. In specific implementation, However, in the hyperspectral images of bamboo and wooden slips, the bamboo and wood textures typically occupy a large area, while the ink stain area is small and has low reflectivity. If only global MSE loss is used for supervision, the network tends to focus its optimization efforts on large areas of bright background, thereby weakening its ability to recover low-reflectivity ink stain areas. To address this, this embodiment proposes a semantically guided ink stain perception weighted loss. By extracting real labels Mean intensity in the spectral dimension This embodiment constructs an adaptive convex penalty weight based on quadratic power decay. :

[0153] ;

[0154] In the formula: As the focus parameter, set to ; For pixels The average reflectance in the spectral dimension. This weighting function can assign a larger loss weight to low reflectance regions, thereby guiding the network to pay more attention to weak texture regions such as ink blots. Meanwhile, to avoid interference from pure background or sensor zero-response regions in the loss calculation, this embodiment introduces an effective pixel mask, i.e., a mask for the ink blot region. .when When the value is below the threshold of 0.05, the location is considered an invalid background area and is not included in the inkblot perception loss calculation. Finally, the inkblot perception weighted loss... Defined as:

[0155] ;

[0156] In the formula: To obtain a clean hyperspectral reflectance image of bamboo slips from a network that reconstructs hyperspectral images of bamboo slips, at the pixel level. The value at; This is a real label; It is an L1 norm; It is a small constant.

[0157] 3.2 Spectral Fidelity Constraints

[0158] Secondly, spectral fidelity constraints are used to suppress spectral shape shifts caused by transmission attenuation and nonlinear reflection in the glass medium. This embodiment uses spectral angle mapping loss to measure the angle between the predicted spectral vector and the true spectral vector, thereby constraining the overall shape consistency of the reconstructed spectral curve.

[0159] ;

[0160] In the formula: For spectral fidelity constraints; For spectral angle mapping loss;

[0161] 3.3 Phased Incremental Constraints

[0162] Furthermore, to ensure the effectiveness of the progressive reconstruction of the coarse-to-fine cascaded architecture, this embodiment establishes phase constraints. This part consists of coarse-stage supervision loss and inter-stage ranking loss:

[0163] ;

[0164] In the formula: As a phased, gradual constraint; Losses due to rough-stage monitoring; The loss is the inter-level ranking loss. and These are the weighting coefficients for the coarse-stage supervision loss and the inter-level ranking loss, respectively. In specific implementation... To prevent degradation of deep networks in the Coarse-to-Fine architecture (i.e., the fine-grained stage only performs identity mappings to the coarse-grained stage), this embodiment introduces a tolerance boundary. Inter-level ranking loss ( ):

[0165] ;

[0166] In the formula: Preliminary hyperspectral reflectance images of bamboo and wooden slips output from the coarse-stage reconstruction; The boundary hyperparameter is set to 0.01; this mechanism forces a gradient penalty to be applied to the network until the error in the fine stage is not only lower than that in the coarse stage, but also must exceed the preset boundary.

[0167] 3.4 Physical Prior Constraints

[0168] Finally, to explicitly incorporate the glass encapsulation degradation process into the network training, this embodiment constructs physical prior constraints, which consist of forward physical consistency loss and reverse coupling loss:

[0169] ;

[0170] ;

[0171] ;

[0172] In the formula: Physical prior constraints; This represents a positive physical consistency loss. This is the inverse coupling loss; and These are the weighting coefficients for the forward physical consistency loss and the reverse coupling loss, respectively. ; This is a simulated glass observation image obtained from the hyperspectral image of pure bamboo slips using a forward physical model; The input is a hyperspectral observation image affected by glass encapsulation interference; The clean hyperspectral reflectance image of the bamboo slips output by the data reconstruction branch; This is the physical estimate of the hyperspectral reflectance image of the pure bamboo slips, calculated through inverse analytical derivation. It is particularly important to note that, because the degradation parameters in the forward degradation model exhibit significant random uncertainty in the very early stages of model training, hastily introducing a physical closed loop can easily lead to systematic gradient oscillations or even collapse. Therefore, this embodiment deploys a dynamic physical weight annealing strategy: global weights that control the intensity of the physical loss. During the first 20 epochs of training, the threshold is smoothly and linearly increased from 0 to a preset threshold.

[0173] 3.5 Joint Multi-Task Loss Function

[0174] The joint multi-task loss function is expressed as follows:

[0175] ;

[0176] In the formula: This is a joint multi-task loss function.

[0177] 4. Overview

[0178] This embodiment addresses the challenges of non-uniform specular reflection and spectral distortion encountered in hyperspectral imaging of glass tube-encapsulated bamboo and wooden slips. It proposes a method for inverting hyperspectral images of encapsulated bamboo and wooden slips based on physical priors. This method proposes a hyperspectral image reconstruction network for encapsulated bamboo and wooden slips that combines physical interpretability with high-fidelity reconstruction capabilities.

[0179] Specifically, to mitigate the black-box nature of purely data-driven models, this embodiment explicitly models the complex optical degradation process as a differentiable physical prior estimation module, achieving pixel-level adaptive parameter prediction and physical consistency constraints. For the spatial-spectral feature coupling problem, a multi-view feature modeling module is designed to extract spatial texture, long-range spectral dependence, and gradient difference information in parallel. Furthermore, this embodiment proposes a coarse-to-fine cascaded reconstruction module and employs a joint multi-task optimization strategy including ink-spot perception loss and inter-level ranking loss to force the network to achieve progressive performance gains and enhance the recovery of weak ink spots.

[0180] Experiments on the SealGlass-HSI dataset constructed in this embodiment show that SGHNet achieves a PSNR of 32.83 dB, SSIM of 0.8863, SAM of 2.4244, and MAE of 0.0161 on the test set; compared to the best PSNR baseline Restormer, the PSNR is improved by 2.17 dB. SGHNet outperforms existing mainstream hyperspectral restoration algorithms in all quantitative metrics, including structural similarity, peak signal-to-noise ratio, and spectral angle mapping. Experimental results demonstrate that this method can effectively reduce the non-uniform reflection and spectral distortion introduced by the glass tube, and better preserve the structure and spectral curve morphology of the bamboo slip ink marks in the visual results.

[0181] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit the technical solutions. Those skilled in the art should understand that any modifications or equivalent substitutions to the technical solutions of the present invention without departing from the spirit and scope of the present invention should be covered within the scope of the claims of the present invention.

Claims

1. A method for inverting encapsulated bamboo and wooden slips hyperspectral images based on physical priors, characterized in that, Acquire hyperspectral observation images that are interfered with by glass encapsulation; input the hyperspectral observation images interfered with by glass encapsulation into a pre-trained hyperspectral image reconstruction network for encapsulated bamboo and wooden slips to perform spectral inversion and obtain a pure hyperspectral reflectance image of the bamboo and wooden slips; The encapsulated bamboo and wooden slips hyperspectral image reconstruction network includes parallel physical estimation and data reconstruction branches. The network transmits the input hyperspectral observation image to both the physical estimation and data reconstruction branches. The physical estimation branch predicts pixel-level degradation parameter fields based on an embedded two-color reflectance model using a differentiable physical prior estimation module. The data reconstruction branch performs coarse-to-fine global spectral morphology estimation and local ink detail restoration through a cascaded reconstruction module. At least one stage of the cascaded reconstruction module deploys a multi-view feature modeling module, which is used to extract spatial texture features, long-range spectral dependence features, and spectral gradient difference features in parallel, finally generating a clean hyperspectral reflectance image of the bamboo and wooden slips as output. The degradation parameter field is used to apply physical prior constraints to the pure bamboo slip hyperspectral reflectance image output by the data reconstruction branch, so that the pure bamboo slip hyperspectral reflectance image is consistent with the hyperspectral observation image under the forward physical model simulation.

2. The method for inverting encapsulated bamboo and wooden slips hyperspectral images based on physical priors according to claim 1, characterized in that, The forward physical model is based on an embedded two-color reflection model and a curved microcavity multiple reflection approximation, and generates simulated glass observation images according to the following formula: ； In the formula: This is a simulated glass observation image obtained from the hyperspectral image of pure bamboo slips using a forward physical model; High spectrum of pure bamboo and wooden slips; This is element-wise multiplication.

3. The method for inverting encapsulated bamboo and wooden slips hyperspectral images based on physical priors according to claim 2, characterized in that, The degradation parameter field predicted by the differentiable physics prior estimation module includes the bidirectional transmission attenuation coefficient. First-order reflection coupling coefficient Second-order nonlinear degradation coefficient and environment bias tensor The range of values ​​for the degradation parameter field is constrained by a nonlinear activation function: the Sigmoid function is used to... The constraint is within the interval (0,1), and the Tanh function is used to... Constraints Within the interval, These are preset positive numbers.

4. The method for inverting encapsulated bamboo and wooden slips hyperspectral images based on physical priors according to claim 3, characterized in that, The differentiable physical prior estimation module also includes a physical inverse model analytical calculation module, which derives the inverse analytical expression based on the degenerate parameter field and constrains the output of the data reconstruction branch through physical coupling loss. Its inverse analytical expression is: ； in, ; In the formula: The input is a hyperspectral observation image affected by glass encapsulation interference; The effective light intensity after removing environmental bias; This is the physical estimate of the hyperspectral reflectance image of the pure bamboo slips, calculated by inverse analytical formula.

5. The method for inverting encapsulated bamboo and wooden slips hyperspectral images based on physical priors according to claim 1, characterized in that, The processing procedure of the multi-view feature modeling module is as follows: The system performs spatial, spectral, and gradient branches in parallel on the input features. The spatial branch first undergoes 3D convolution and normalized activation, then uses large-kernel depthwise separable convolution to extract spatial texture features. The spectral branch first undergoes 3D convolution and normalized activation, then uses a spectral multi-head self-attention mechanism to extract long-range spectral dependency features, followed by another 3D convolution and normalized activation. The gradient branch extracts spectral gradient difference features through spectral gradient blocks. The output features of the spatial branch, spectral branch and gradient branch are concatenated along the channel dimension and then input into the global context module for adaptive feature recalibration and fusion to obtain fused multi-view features. Finally, the features are decoupled and the enhanced multi-view features are output.

6. The method for inverting encapsulated bamboo and wooden slips hyperspectral images based on physical priors according to claim 5, characterized in that, The processing procedure for the spatial branch is expressed as follows: ； ； In the formula: These are intermediate features of spatial branches; The input feature tensor; Normalize the group; This is a three-dimensional convolution with a kernel size of 1×3×3; For activation functions; Spatial texture features output by spatial branches; Large kernel depthwise convolution; Use the GELU activation function; For pointwise convolution; The processing procedure for the spectral branch is expressed as follows: ； ； In the formula: This represents an intermediate feature of the spectral branch; This represents the long-range spectral dependence characteristic of the spectral branch output; This is a three-dimensional convolution with a kernel size of 3×3×1; For spectral multi-head self-attention; The gradient branching process is represented as follows: ； In the formula: This represents the spectral gradient difference feature output by the gradient branch.

7. The method for inverting encapsulated bamboo and wooden slips hyperspectral images based on physical priors according to claim 6, characterized in that, The processing procedure of the global context module is represented as follows: ； ； In the formula: The resulting multi-view features; Indicates concatenation of channel dimensions; This is a three-dimensional global average pooling method. and All are linear mapping weights in the global context module; This indicates multiplication by channel; Use the Sigmoid activation function; Enhanced multi-view features output by the global context module.

8. The method for inverting encapsulated bamboo and wooden slips hyperspectral images based on physical priors according to claim 7, characterized in that, The cascaded reconstruction module includes an encoder, a coarse-stage reconstruction, a feature fusion module, a fine-stage reconstruction, and a decoder. The coarse-stage reconstruction includes a cascaded first feature extraction block, a max-pooling layer, a second feature extraction block, and an upsampling layer. The fine-stage reconstruction includes a cascaded third and fourth feature extraction blocks. The encoder is used to encode the input hyperspectral observation image step-by-step, extracting multi-level spatial-spectral features. The decoder receives the output of the fourth feature extraction block. In the cascaded reconstruction module, the input hyperspectral observation image is reconstructed through a coarse stage to output a preliminary hyperspectral reflectance image of the bamboo and wooden slips. Then, the output of the coarse stage reconstruction is fused with the multi-level spatial spectral features extracted by the encoder in the feature fusion module and used as the input for the fine stage reconstruction to restore the high-frequency ink details at the original resolution. Finally, the decoder outputs a clean hyperspectral reflectance image of the bamboo and wooden slips. Among them, at least one feature extraction module is equipped with the multi-view feature modeling module, which is used to perform parallel extraction and adaptive fusion of spatial texture features, long-range spectral dependence features and spectral gradient difference features of the current stage features, and to use the enhanced multi-view features to perform spectral morphology estimation or ink detail restoration of the current stage.

9. The method for inverting encapsulated bamboo and wooden slips hyperspectral images based on physical priors according to claim 1, characterized in that, The training process of the encapsulated bamboo slips hyperspectral image reconstruction network is as follows: A pre-constructed hyperspectral image dataset of glass-encased bamboo and wooden slips is used as a training dataset and input into the hyperspectral image reconstruction network of the encased bamboo and wooden slips. A joint multi-task loss function consisting of spatial reconstruction constraints, spectral fidelity constraints, phase progressive constraints, and physical prior constraints is constructed. The training loss of the hyperspectral image reconstruction network of the encased bamboo and wooden slips is calculated. The parameters of the network are updated with the goal of minimizing the joint multi-task loss function, thereby training the hyperspectral image reconstruction network of the encased bamboo and wooden slips. The training dataset includes hyperspectral observation images affected by glass encapsulation and their corresponding labels for pure bamboo slip hyperspectral reflectance images.

10. The method for inverting encapsulated bamboo and wooden slips hyperspectral images based on physical priors according to claim 9, characterized in that, The joint multi-task loss function is expressed as follows: ； In the formula: For joint multi-task loss function; Constraints for spatial reconstruction; For spectral fidelity constraints; As a phased, gradual constraint; Physical prior constraints; The spatial reconstruction constraints include a global reconstruction loss and an ink-perception weighted loss, which are expressed as follows: ； ； ； In the formula: This represents a loss during the overall reconstruction. Weighted loss for ink perception; and These are the weighting coefficients for the global reconstruction loss and the ink perception weighted loss, respectively. A mask for the ink area; Adaptive convex penalty weights; To obtain a clean hyperspectral reflectance image of bamboo slips from a network that reconstructs hyperspectral images of bamboo slips, at the pixel level. The value at; This is a real label; It is an L1 norm; It is a small constant; For focusing parameters; For pixels Average reflectance in the spectral dimension; The spectral fidelity constraint employs spectral angle mapping loss, which is expressed as: ； In the formula: For spectral angle mapping loss; The stage-wise asymptotic constraints include coarse-stage supervision loss and inter-stage ranking loss, which are expressed as follows: ； ； In the formula: Losses due to rough-stage monitoring; The loss is the inter-level ranking loss. and These are the weighting coefficients for the coarse-stage supervision loss and the inter-level ranking loss, respectively. Preliminary hyperspectral reflectance images of bamboo and wooden slips output from the coarse-stage reconstruction; These are boundary hyperparameters; The physical prior constraints include forward physical consistency loss and reverse coupling loss, which are expressed as follows: ； ； ； In the formula: This represents a positive physical consistency loss. This is the inverse coupling loss; and These are the weighting coefficients for the forward physical consistency loss and the reverse coupling loss, respectively; This is a simulated glass observation image obtained from the hyperspectral image of pure bamboo slips using a forward physical model; The input is a hyperspectral observation image affected by glass encapsulation interference; The clean hyperspectral reflectance image of the bamboo slips output by the data reconstruction branch; This is the physical estimate of the hyperspectral reflectance image of the pure bamboo slips, calculated by inverse analytical formula.

Images