High-dimensional light field detection method and device based on physical feature decoupling

Abstract

The application discloses a high-dimensional light field detection method based on physical feature decoupling, and belongs to the technical field of optical measurement. The method comprises the following steps: S1, collecting a speckle pattern of a to-be-detected high-dimensional light field encoded and output by a multimode interference medium; S2, performing light field physical feature separation and denoising processing on the speckle pattern, projecting a plurality of light field parameters of the to-be-detected high-dimensional light field into respective corresponding low-dimensional feature subspaces, and obtaining low-dimensional features corresponding to the plurality of light field parameters of the to-be-detected high-dimensional light field; and S3, inputting a feature vector composed of the low-dimensional features into a decoding module based on a physical feature subspace which is pre-established and trained, and obtaining the plurality of light field parameters of the to-be-detected high-dimensional light field. The application further discloses a high-dimensional light field detection device based on physical feature decoupling. The application can effectively improve the detection sensitivity and stability of subtle changes in the light field, reduce the demand for training data and the complexity of the model, and reduce the training and reasoning time.

Description

Technical Field

[0001] This invention relates to a high-dimensional light field detection method, belonging to the field of optical measurement technology. Background Technology

[0002] The inherent physical properties of light fields are naturally distributed across multiple independent dimensions, among which polarization state, wavelength, and intensity are the core parameters comprehensively characterizing the intrinsic properties of incident light fields. Joint detection of these multiple dimensions allows for the acquisition of more complete and complementary light field information in a single measurement, thereby improving the overall efficiency and information utilization of optical sensing. High-dimensional light field detection has significant application value in fields such as chemical composition and material structure analysis, biomedical imaging and endoscopic diagnosis, astronomical observation, multidimensional environmental sensing, and remote sensing monitoring.

[0003] Existing optical field detection technologies typically decompose the high-dimensional parameter space into single-dimensional or two-dimensional slices for separate measurements, such as measuring polarization state, wavelength, or spectral distribution. While single-dimensional detection technologies are relatively mature, they cannot handle multi-dimensional coupling characteristics. To jointly acquire complete optical field information, multiple independent-dimensional detection systems often need to collaborate or acquire data sequentially, relying on complex optical path switching and multiple measurements. This results in a complex and bulky overall system structure, limited in time efficiency and operational stability, making it difficult to meet the requirements for fast, compact, and high-throughput detection.

[0004] In recent years, one approach for joint measurement of high-dimensional optical fields has been to explicitly separate and map optical field information of different dimensions to different locations in space by introducing specific modulation or independent coding, and then achieve multi-parameter detection by measuring the spatial distribution and intensity of the signals. However, such methods usually rely on precise and time-consuming transmission matrix calibration and often involve multi-step measurements, making them unsuitable for rapidly measuring dynamically changing optical fields. At the same time, as the detection dimension and parameter range increase, crosstalk between signals from different channels intensifies, often forcing compromises in system resolution and accuracy, and placing high demands on the design and fabrication precision of optical components.

[0005] Another approach relies on nonlocally encoded speckle patterns generated by disordered media or nonlinear mappings in multi-input multi-output systems to reconstruct input optical field parameters from multi-channel intensity signals of a single measurement. This approach offers advantages such as compact structure, fast response, and ease of integration. However, due to the strong coupling and nonlinear mixing caused by scattering and multimode interference, the encoding relationships are highly complex and difficult to decouple effectively using traditional transfer matrices or simple statistical models. The reconstruction process often relies on deep neural networks, which are costly and time-consuming to train, to process large-scale pixel-level speckle or multi-channel signals. Furthermore, each dimension of the optical field typically only supports the detection of a few discrete states, and further improvements in system accuracy may be limited by inherent speckle correlation limits and model complexity. Simultaneously, the mapping mechanism from random speckle to high-dimensional parameters lacks interpretability. These factors, to some extent, restrict the widespread application of high-dimensional optical field detection technology.

[0006] Therefore, there is an urgent need for a measurement scheme that can efficiently and synchronously acquire high-dimensional optical field information in a single detection system. This scheme should ensure the accuracy of joint detection of multi-dimensional parameters such as polarization, wavelength, and intensity, while reducing the dependence on complex calibration processes, deep networks, and large-scale training data, and improving measurement efficiency, stability, and scalability. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of existing high-dimensional light field detection technologies and provide a high-dimensional light field detection method based on physical feature decoupling. This method can effectively improve the detection sensitivity and stability of subtle changes in the light field, while significantly reducing the training data requirements and model complexity, reducing training and inference time, and improving the system's practicality and rapid deployment capability.

[0008] The present invention specifically adopts the following technical solutions to achieve the above objectives: The high-dimensional light field detection method based on physical feature decoupling includes the following steps: S1. Acquire the speckle pattern output by encoding the high-dimensional light field to be measured through a multimode interference medium; S2. Perform optical field physical feature separation and noise reduction on the speckle pattern, and project the multiple optical field parameters of the high-dimensional optical field to be measured to their respective low-dimensional feature subspaces to obtain the low-dimensional features corresponding to the multiple optical field parameters of the high-dimensional optical field to be measured. S3. Input the feature vector formed by combining the low-dimensional features into a pre-established and trained decoding module based on physical feature subspace to obtain the complex number of light field parameters of the high-dimensional light field to be measured.

[0009] Based on the same inventive concept, the following technical solutions can also be obtained: A high-dimensional light field detection device based on physical feature decoupling includes: The multimode interferometer module is used to encode the high-dimensional optical field to be measured into a speckle pattern through a multimode interferometer medium; The speckle detection module is used to acquire the speckle pattern output by the multimode interferometer module; The optical field physical feature separation and denoising module is used to perform optical field physical feature separation and denoising processing on the speckle pattern, and to project the complex optical field parameters of the high-dimensional optical field to be measured to their respective low-dimensional feature subspaces to obtain the low-dimensional features corresponding to the complex optical field parameters of the high-dimensional optical field to be measured. A decoding module based on a physical feature subspace is pre-built and trained to take the feature vector composed of the low-dimensional features as input and output the complex number of light field parameters of the high-dimensional light field to be measured.

[0010] As a preferred embodiment of the above technical solution, principal component analysis is used to separate and denoise the optical field physical features of the speckle pattern. The low-dimensional feature corresponding to any optical field parameter of the high-dimensional optical field to be measured is at least one principal component score corresponding to the speckle pattern. The decoding module based on the physical feature subspace is established and trained by the following method: under a series of high-dimensional optical field conditions with known optical field parameters, speckle patterns of the high-dimensional optical field are collected and encoded by a multimode interference medium. Principal component analysis is performed on the speckle pattern, and the principal component scores most relevant to each optical field parameter are extracted. The known optical field parameters are used as training inputs, and the principal component scores most relevant to each optical field parameter are used as training outputs to train a lightweight multi-output regression model, thereby obtaining the decoding module based on the physical feature subspace.

[0011] More preferably, the plurality of optical field parameters include a full Stokes polarization parameter, an intensity parameter, and a wavelength parameter; the low-dimensional feature corresponding to the full Stokes polarization parameter is the score of the first three principal components, the low-dimensional feature corresponding to the intensity parameter is the score of the fourth principal component, and the low-dimensional feature corresponding to the wavelength parameter is the score of the fifth principal component.

[0012] More preferably, the lightweight multi-output regression model is a multilayer perceptron neural network.

[0013] Preferably, the multimode interference medium is one or a combination of two or more of the following media: multimode fiber, few-mode fiber, multimode waveguide, integrating sphere, and frosted glass.

[0014] Compared with the prior art, the technical solution of the present invention has the following beneficial effects: This invention separates and denoises speckle patterns formed by multimode interference, obtaining low-dimensional features of multiple high-dimensional optical field parameters in their respective feature subspaces. This enhances the separability of different optical field parameters in the feature space, structurally alleviating the strong coupling problem of optical field parameters such as polarization, wavelength, and intensity in speckle patterns, and improving the sensitivity and stability of detecting subtle changes in the optical field. Simultaneously, only a lightweight neural network is needed to complete the decoding in a low-dimensional feature space with separated parameter-related features. This significantly reduces the need for training data and model complexity while improving accuracy, reducing training and inference time, and enhancing the system's practicality and rapid deployment capabilities. Compared to existing optical field detection technologies, this invention can achieve single-synchronous reconstruction of optical field parameters such as full Stokes polarization, wavelength, and intensity without introducing additional optical modulation structures or complex calibration procedures. This provides a highly efficient and physically grounded approach for the application of high-dimensional optical fields in optical metrology, imaging sensing, and intelligent photonics systems. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the principle of the high-dimensional light field detection device based on physical feature decoupling of the present invention; Figure 2 This is a schematic diagram illustrating the principle of the high-dimensional light field detection method of the present invention; Figure 3 This is a structural diagram of a specific experimental device for the high-dimensional light field detection device of the present invention; Figure 4 A flowchart illustrating the workflow for reconstructing an unknown high-dimensional optical field from a single speckle measurement; Figure 5 This is an example of subspace load distribution when performing optical field feature separation on speckle; Figure 6 The effect of Stokes polarization measurement in high-dimensional light field detection; Figure 7 The effect of optical power measurement in high-dimensional optical field detection; Figure 8 The effect of wavelength shift measurement in high-dimensional optical field detection; Figure 9 This diagram illustrates the global importance of the scores of each principal component in the detection of high-dimensional light field parameters. Detailed Implementation

[0016] Existing high-dimensional light field detection technologies generally suffer from the following shortcomings: On the one hand, multiple light field parameters, such as polarization state, wavelength, and light intensity, are strongly coupled and nonlinearly encoded in multimode interference speckle. Existing methods often rely on complex calibration procedures or multi-step measurements, making them difficult to apply to single-detection scenarios with rapidly changing light fields. On the other hand, existing high-dimensional light field detection schemes based on speckle images and deep neural networks typically use pixel-level speckle images as model inputs directly. This requires the network to simultaneously learn physical features related to light field parameters as well as a large amount of irrelevant information related to random interference and environmental disturbances. This results in complex model structures, large amounts of training data, high training and deployment costs, and limited accuracy and stability under conditions of continuous wavelength changes or polarization disturbances.

[0017] To address the aforementioned issues, the inventors, through research on the formation mechanism of multimode interference speckle and the encoding law of high-dimensional optical field parameters, discovered that not all spatial degrees of freedom in the speckle pattern contribute equally to different optical field parameters. Different optical field parameters exhibit significant differences in characteristic distribution during the statistical process of speckle variation with parameter changes. That is, due to the different sensitivities of the propagation constants, polarization coupling relationships, and intermodal phase accumulation of different intrinsic modes to the incident optical field parameters, the speckle pattern does not respond uniformly across all spatial degrees of freedom when parameters change, but rather exhibits statistical structural evolution characteristics related to specific optical field parameters. Based on this, the inventors proposed the following technical solution: The high-dimensional light field detection method based on physical feature decoupling includes the following steps: S1. Acquire the speckle pattern output by encoding the high-dimensional light field to be measured through a multimode interference medium; S2. Perform optical field physical feature separation and noise reduction on the speckle pattern, and project the multiple optical field parameters of the high-dimensional optical field to be measured to their respective low-dimensional feature subspaces to obtain the low-dimensional features corresponding to the multiple optical field parameters of the high-dimensional optical field to be measured. S3. Input the feature vector formed by combining the low-dimensional features into a pre-established and trained decoding module based on physical feature subspace to obtain the complex number of light field parameters of the high-dimensional light field to be measured.

[0018] Based on the same inventive concept, the following technical solutions can also be obtained: A high-dimensional light field detection device based on physical feature decoupling includes: The multimode interferometer module is used to encode the high-dimensional optical field to be measured into a speckle pattern through a multimode interferometer medium; The speckle detection module is used to acquire the speckle pattern output by the multimode interferometer module; The optical field physical feature separation and denoising module is used to perform optical field physical feature separation and denoising processing on the speckle pattern, and to project the complex optical field parameters of the high-dimensional optical field to be measured to their respective low-dimensional feature subspaces to obtain the low-dimensional features corresponding to the complex optical field parameters of the high-dimensional optical field to be measured. A decoding module based on a physical feature subspace is pre-built and trained to take the feature vector composed of the low-dimensional features as input and output the complex number of light field parameters of the high-dimensional light field to be measured.

[0019] The speckle pattern can be subjected to optical field physical feature separation and denoising using methods such as Independent Component Analysis (ICA), Principal Component Analysis (PCA), and sparse decomposition. Preferably, PCA is used for optical field physical feature separation and denoising of the speckle pattern. The low-dimensional feature corresponding to any optical field parameter of the high-dimensional optical field to be measured is at least one principal component score corresponding to the speckle pattern. The decoding module based on the physical feature subspace is established and trained by the following method: under a series of high-dimensional optical field conditions with known optical field parameters, speckle patterns of the high-dimensional optical field are collected and encoded by a multimode interference medium. Principal component analysis is performed on the speckle pattern, and the principal component scores most relevant to each optical field parameter are extracted. The known optical field parameters are used as training inputs, and the principal component scores most relevant to each optical field parameter are used as training outputs to train a lightweight multi-output regression model, thereby obtaining the decoding module based on the physical feature subspace.

[0020] More preferably, the plurality of optical field parameters include a full Stokes polarization parameter, an intensity parameter, and a wavelength parameter; the low-dimensional feature corresponding to the full Stokes polarization parameter is the score of the first three principal components, the low-dimensional feature corresponding to the intensity parameter is the score of the fourth principal component, and the low-dimensional feature corresponding to the wavelength parameter is the score of the fifth principal component.

[0021] More preferably, the lightweight multi-output regression model is a multilayer perceptron neural network.

[0022] Preferably, the multimode interference medium is one or a combination of two or more of the following media: multimode fiber, few-mode fiber, multimode waveguide, integrating sphere, and frosted glass.

[0023] To facilitate public understanding, the technical solution of the present invention will be described in detail below through a preferred embodiment and in conjunction with the accompanying drawings: The high-dimensional light field detection device based on physical feature decoupling in this embodiment, such as Figure 1As shown, the system includes a multimode interferometry module, a speckle detection module, an optical field physical feature separation and denoising module, and a decoding module based on physical feature subspaces. The multimode interferometry module encodes the high-dimensional optical field to be measured into a speckle pattern using a multimode interferometer. The speckle detection module is used to acquire the speckle pattern output by the multimode interferometry module. The optical field physical feature separation and denoising module is used to perform optical field physical feature separation and denoising processing on the speckle pattern, projecting the complex optical field parameters of the high-dimensional optical field to be measured onto their respective low-dimensional feature subspaces to obtain the low-dimensional features corresponding to the complex optical field parameters of the high-dimensional optical field to be measured. The decoding module based on physical feature subspaces is pre-established and trained, and is used to output the complex optical field parameters of the high-dimensional optical field to be measured by taking the feature vector formed by the combination of the low-dimensional features as input.

[0024] The working principle of the above-mentioned high-dimensional light field detection device is as follows: Figure 2 As shown, the specific detection process includes the following steps: S1. Acquire the speckle pattern output by encoding the high-dimensional optical field to be measured through a multimode interferometer medium: like Figure 2 As shown, when the unknown light field to be measured propagates through a multimode interference medium (such as multimode fiber, few-mode fiber, multimode waveguide, integrating sphere, frosted glass, or a combination thereof), multimode interference and scattering will form a speckle pattern at the output end that appears random but has a deterministic mapping. This pattern originates from the coherent superposition of different degrees of freedom of the light field (such as polarization and wavelength) in space and time. Therefore, high-dimensional light field information is encoded as a whole in the intensity distribution of the speckle in different ways. The goal of high-dimensional light field detection is to decode and reconstruct the complete light field parameters from a single speckle measurement.

[0025] S2. Perform optical field physical feature separation and denoising processing on the speckle pattern, projecting the complex optical field parameters of the high-dimensional optical field to be measured onto their respective low-dimensional feature subspaces to obtain the low-dimensional features corresponding to the complex optical field parameters of the high-dimensional optical field to be measured: Since the propagation constants, polarization coupling relationships, and intermodal phase accumulation of different intrinsic modes have different sensitivities to the incident light field parameters, the speckle pattern does not respond uniformly in all spatial degrees of freedom when the parameters change, but exhibits statistical structural evolution characteristics related to specific light field parameters. Therefore, each light field parameter of the high-dimensional light field to be measured can be projected to its corresponding low-dimensional feature subspace.

[0026] In this embodiment, principal component analysis (PCA) is used to separate and denoise the optical field physical features of the speckle pattern. The speckle features are structurally separated, resulting in a series of principal component scores. The inventors discovered that statistical features highly correlated with polarization state changes are concentrated in the first three principal component directions most relevant to speckle structure evolution; features related to light intensity changes are concentrated in the fourth principal component direction, reflecting changes in overall brightness and contrast of the speckle; and features related to wavelength changes are more associated with the fifth principal component direction, which is sensitive to intermodal phase accumulation. Compared to these feature directions, higher-order principal components in the speckle feature space mainly carry random interference noise or redundant information, contributing little to the detection of high-dimensional optical fields. Therefore, a feature vector for optical field decoupling can be constructed by selecting low-order principal component scores, including the first five principal components, for subsequent joint reconstruction of high-dimensional parameters (polarization, light intensity, wavelength) of the incident light field.

[0027] S3. Input the feature vector formed by combining the low-dimensional features into a pre-established and trained decoding module based on the physical feature subspace to obtain the complex number of light field parameters of the high-dimensional light field to be measured: The decoding module based on the physical feature subspace is established and trained by the following method: under a series of high-dimensional light field conditions where the multiple light field parameters are known, speckle patterns of the high-dimensional light field are collected and encoded by a multimode interference medium; principal component analysis is performed on the speckle patterns, and the principal component scores most relevant to each light field parameter are extracted; the known multiple light field parameters are used as training inputs, and the principal component scores most relevant to each light field parameter are used as training outputs to train a lightweight multi-output regression model, thus obtaining the decoding module based on the physical feature subspace. The lightweight multi-output regression model in this embodiment employs a lightweight multilayer perceptron neural network. The model includes one input layer, six fully connected hidden layers, and output layers corresponding to each optical field parameter. L2 regularization and normalization mechanisms are introduced to improve training stability and generalization ability. Meanwhile, considering the differences in the dimensions and numerical scales of different output optical field parameters, to avoid one dimension parameter dominating training and causing a decrease in the estimation accuracy of other dimensions, a weighted mean square error is preferably used as the loss function to balance each output dimension. The weight coefficients for polarization, light intensity, and wavelength loss are 0.3, 0.1, and 0.6, respectively.

[0028] To verify the effectiveness of the technical solution of the present invention, the following was conducted: Figure 3The experimental setup shown was used to conduct a high-dimensional optical field detection verification experiment. In the experiment, the wavelength and polarization state of the light under test were tuned by an acousto-optic frequency shifter with a center frequency shift of 400 MHz and a high-speed electrically controlled polarization controller, respectively. An adjustable optical attenuator was used to set the system power, and the polarization controller was used to adjust the initial polarization state. The high-dimensional optical field detector consisted of a 200 μm core diameter step-index multimode fiber YOFC, FM2011-B, a lens, and a Xenics Bobcat-320-GigE CCD camera. Measurements from a commercial polarimeter, General Photonics, POD-201, were used as theoretical reference values ​​for the full Stokes polarization state and light intensity.

[0029] Figure 4 This paper demonstrates a workflow for reconstructing an unknown high-dimensional optical field from a single speckle measurement. Principal component analysis (PCA) is introduced to perform feature separation and denoising of the speckle, obtaining principal component scores in the polarization, intensity, and wavelength subspaces. PCA based on singular value decomposition re-represents the speckle intensity matrix as follows:

[0030] in, It includes p The speckle column vector is obtained by subtracting the mean from (256 × 256) pixels. n The sample size is indicated by a superscript. T This indicates the matrix transpose. U and V These are the left singular vector and the right singular vector, respectively. Λ It is a diagonal matrix containing singular values. For high-dimensional encoded light fields, the load matrix... V An orthogonal basis for each principal component subset space is defined, and the principal component scores... s = IV =( I - μ ) V The projection coordinates of speckle in these subspaces are represented. In the light field reconstruction of a single image, the loading matrix... V and mean μ The data is derived from the training set. After principal component analysis projection, the principal features highly correlated with changes in the optical field parameters are concentrated in a few low-order principal component directions, namely the polarization subspace, intensity subspace, and wavelength subspace, while higher-order components reflect more of the system's random noise or redundant changes. Therefore, by selecting the low-order principal component scores, including the first five principal components, to construct the feature vector for optical field decoupling, the discriminative features most sensitive to polarization, wavelength, and intensity are preserved, and the data dimensionality is significantly reduced. This not only improves the system's accuracy and noise resistance but also simplifies the subsequent nonlinear decoding structurally.

[0031] A multilayer perceptron neural network was used to construct a lightweight multi-output regression model to decode high-dimensional light field parameters from the parameter subspace of feature separation and denoising. This lightweight multi-output regression model consists of one input layer, six fully connected hidden layers, and an output layer corresponding to the high-dimensional light field parameters. L2 regularization and batch normalization are introduced to improve training stability and generalization ability. Each hidden layer uses the ReLU activation function, and parameter updates are performed using the AdamW optimizer. Given the scale differences in the model's outputs (polarization, wavelength, and light intensity), a weighted mean squared error is further designed as the loss function to avoid any single parameter dominating training. The total loss function (…) Loss ) is represented as:

[0032] in, n For the sample size, and These are the Stokes parameters and wavelength, respectively. and These are the corresponding estimated values. Weighting coefficients. α 1, α 2 and α 3 is used to balance the numerical differences of parameters in different dimensions and adjust their importance in training. In the specific implementation, they are taken as 0.3, 0.6 and 0.1 respectively.

[0033] To compare the different responses of the original speckle pattern and the parameter subspace score under high-dimensional optical field variations, the normalized root mean square error (NRMSE) of the original speckle pattern and principal component score relative to their respective initial reference states was calculated. An increase in NRMSE indicates that the features respond more sensitively to changes in optical field parameters. NRMSE is defined as:

[0034] in, x i To change the speckle pattern or principal component score vector after altering the high-dimensional light field parameters, x 0 is the reference value when the light field parameters remain unchanged. N The number of elements in the vector.

[0035] Furthermore, the mean Stokes angle error is used to evaluate the deviation between the estimated polarization state and the theoretical value on the Poincaré sphere, and is defined as:

[0036] in, and These are the theoretical reference value and the estimated value of the full Stokes vector, respectively.

[0037] To address high-dimensional optical field detection scenarios with mixed variations in polarization, wavelength, and power, an electrically controlled polarization controller is used to rapidly and arbitrarily tune the polarization state of the beam under test across the entire Poincaré sphere during optical field measurement. Simultaneously, continuous wavelength shift is achieved by tuning the radio frequency signal driving the acousto-optic frequency shifter. The system's optical power is adjusted via a tunable optical attenuator. Speckle patterns under different optical field parameters are acquired and divided into training, validation, and test sets. Principal component analysis is used to project the speckle data using the load matrix and mean vector from the training set, and the scores of the first five principal components are extracted as the input vector for the neural network. Figure 5 Examples of load distribution in the first five characteristic subspaces, including polarization subspace, intensity subspace, and wavelength subspace, during principal component analysis projection.

[0038] The multilayer perceptron model built with PyTorch, after 1000 iterations of training in a regression task, consistently reduced its joint loss on the validation set to 2.01 × 10⁻⁶. -5 The values ​​are close to those of the training set. This indicates that the network learns robust nonlinear mappings in the principal component score feature space and exhibits good generalization performance.

[0039] The trained neural network estimates the full Stokes polarization parameters from the principal component scores of the test set, such as... Figure 6 As shown, the normalized Stokes vector ( S 1 / S 0, S 2 / S 0, S 3 / S 0) Used to characterize the polarization state. The estimates on the Poincaré sphere are very close to the measurements of commercial polarimeters across the entire polarization state space. Specifically, the estimation errors for the three polarization components are concentrated within ±0.01, and... S 1 / S 0, S 2 / S 0, and S 3 / S The standard deviations of 0 were 0.0032, 0.0035 and 0.0028, respectively, indicating that the method can accurately detect arbitrary Stokes polarization states. Figure 7 The results demonstrate the effectiveness of optical power measurement in high-dimensional optical field detection, with the power estimated from the principal component scores of different samples showing a standard deviation of only 0.0636 μW compared to the theoretical value. Since the speckle pattern undergoes only slight structural evolution at an input wavelength shift of 800 fm, traditional methods struggle to directly extract wavelength-related discriminative features. This invention presents a high-dimensional optical field detection method based on physical feature decoupling, achieving effective wavelength estimation under different polarization states. Figure 8For the wavelength shift measurement results, the estimated value and the theoretical value show a good linear relationship over the entire wavelength variation range of 800 fm, and the coefficient of determination R of the fit is high. 2 The value is 0.99998, and the corresponding absolute error is always below 3.12 fm, which demonstrates the ability of this method to maintain stable measurement at the femtometer level wavelength under polarization perturbation.

[0040] Therefore, these results systematically verify the high-precision detection performance of the proposed high-dimensional optical field detection method based on physical feature decoupling in a three-dimensional continuous parameter space composed of polarization, wavelength and power, and demonstrate its intelligent sensing potential in complex optical field joint measurement scenarios.

[0041] Furthermore, the SHAP method is introduced to quantify the global contribution of each principal component score in the detection of different high-dimensional parameters, in order to analyze the physical meaning and feature separation mechanism of the high-dimensional light field detection framework based on physical feature decoupling in the high-dimensional light field parameter decoding process. Figure 9 The distribution of the mean absolute SHAP values ​​of the top 10 principal component scores (PC1~PC10) in the estimation of Stokes polarization component, optical power, and wavelength shift is presented. Polarization-related parameters. S 1, S 2 and S Figure 3 shows that the low-order principal components dominate, with the first three principal component subspaces (PC1~PC3) exhibiting a higher contribution to the Stokes component estimation. Therefore, the speckle features highly correlated with polarization state changes are mainly concentrated in the first three principal component subspaces, reflecting the strong consistency and dominance of polarization-induced speckle structure changes in the overall optical field evolution. In contrast, optical power... S The detection of 0 shows the strongest dependence on PC4, while the contributions of other principal components are weak. This indicates that the change in optical power is decoupled to a relatively independent characteristic direction, exhibiting good separability from polarization characteristics. This result is consistent with the fact that optical power mainly affects the overall brightness and contrast level of the speckle pattern, rather than finely altering the spatial interference structure. Wavelength shift λ The main contribution to the estimation comes from PC5, and this principal component has relatively low importance in polarization and power parameters. Intermodal phase accumulation caused by wavelength variation and the resulting speckle structure evolution tend to be distributed in relatively high-order characteristic directions. Although these features account for a small proportion of the overall variance and are more difficult to distinguish, they are crucial for achieving femtometer-level wavelength resolution.

[0042] SHAP analysis clearly reveals the distribution of optical field parameters in different physical dimensions within the principal component subspace. Polarization information is mainly concentrated in the first three principal components, optical power is characterized by the fourth principal component, and wavelength is more correlated with the fifth principal component. This parameter-dependent distribution not only explains why high-precision high-dimensional optical field reconstruction can be achieved with only a few principal component scores, but also verifies that the role of principal component analysis in complex speckle measurements is not simply data dimensionality reduction, but rather structurally promotes the separation and efficient decoupling of multidimensional optical field parameters. Furthermore, the results of this analysis can provide valuable physical basis for the selection and optimization of principal components in high-dimensional detection extension missions.

Claims

1. A high-dimensional light field detection method based on physical feature decoupling, characterized in that, Includes the following steps: S1. Acquire the speckle pattern output by encoding the high-dimensional light field to be measured through a multimode interference medium; S2. Perform optical field physical feature separation and noise reduction on the speckle pattern, and project the multiple optical field parameters of the high-dimensional optical field to be measured to their respective low-dimensional feature subspaces to obtain the low-dimensional features corresponding to the multiple optical field parameters of the high-dimensional optical field to be measured. S3. Input the feature vector formed by combining the low-dimensional features into a pre-established and trained decoding module based on physical feature subspace to obtain the complex number of light field parameters of the high-dimensional light field to be measured.

2. The high-dimensional light field detection method based on physical feature decoupling as described in claim 1, characterized in that, Principal component analysis is used to separate and denoise the optical field physical features of the speckle pattern; the low-dimensional feature corresponding to any optical field parameter of the high-dimensional optical field to be measured is at least one principal component score corresponding to the speckle pattern; the decoding module based on the physical feature subspace is established and trained by the following method: under a series of high-dimensional optical field conditions with known optical field parameters, the speckle pattern of the high-dimensional optical field is collected and encoded by a multimode interference medium. Principal component analysis is performed on the speckle pattern, and the principal component scores most relevant to each light field parameter are extracted. The lightweight multi-output regression model is trained using the known multiple light field parameters as training inputs and the principal component scores most relevant to each light field parameter as training outputs, thus obtaining the decoding module based on the physical feature subspace.

3. The high-dimensional light field detection method based on physical feature decoupling as described in claim 2, characterized in that, The plurality of optical field parameters include the full Stokes polarization parameter, the light intensity parameter, and the wavelength parameter; the low-dimensional feature corresponding to the full Stokes polarization parameter is the score of the first three principal components, the low-dimensional feature corresponding to the light intensity parameter is the score of the fourth principal component, and the low-dimensional feature corresponding to the wavelength parameter is the score of the fifth principal component.

4. The high-dimensional light field detection method based on physical feature decoupling as described in claim 2, characterized in that, The lightweight multi-output regression model is a multilayer perceptron neural network.

5. The high-dimensional light field detection method based on physical feature decoupling as described in claim 1, characterized in that, The multimode interference medium is one or a combination of two or more of the following media: multimode fiber, few-mode fiber, multimode waveguide, integrating sphere, and frosted glass.

6. A high-dimensional light field detection device based on physical feature decoupling, characterized in that, include: The multimode interferometer module is used to encode the high-dimensional optical field to be measured into a speckle pattern through a multimode interferometer medium; The speckle detection module is used to acquire the speckle pattern output by the multimode interferometer module; The optical field physical feature separation and denoising module is used to perform optical field physical feature separation and denoising processing on the speckle pattern, and to project the complex optical field parameters of the high-dimensional optical field to be measured to their respective low-dimensional feature subspaces to obtain the low-dimensional features corresponding to the complex optical field parameters of the high-dimensional optical field to be measured. A decoding module based on a physical feature subspace is pre-built and trained to take the feature vector composed of the low-dimensional features as input and output the complex number of light field parameters of the high-dimensional light field to be measured.

7. The high-dimensional light field detection device based on physical feature decoupling as described in claim 6, characterized in that, Principal component analysis is used to separate and denoise the optical field physical features of the speckle pattern; the low-dimensional feature corresponding to any optical field parameter of the high-dimensional optical field to be measured is at least one principal component score corresponding to the speckle pattern; the decoding module based on the physical feature subspace is established and trained by the following method: under a series of high-dimensional optical field conditions with known optical field parameters, the speckle pattern of the high-dimensional optical field is collected and encoded by a multimode interference medium. Principal component analysis is performed on the speckle pattern, and the principal component scores most relevant to each light field parameter are extracted. The lightweight multi-output regression model is trained using the known multiple light field parameters as training inputs and the principal component scores most relevant to each light field parameter as training outputs, thus obtaining the decoding module based on the physical feature subspace.

8. The high-dimensional light field detection device based on physical feature decoupling as described in claim 7, characterized in that, The plurality of optical field parameters include the full Stokes polarization parameter, the light intensity parameter, and the wavelength parameter; the low-dimensional feature corresponding to the full Stokes polarization parameter is the score of the first three principal components, the low-dimensional feature corresponding to the light intensity parameter is the score of the fourth principal component, and the low-dimensional feature corresponding to the wavelength parameter is the score of the fifth principal component.

9. The high-dimensional light field detection device based on physical feature decoupling as described in claim 7, characterized in that, The lightweight multi-output regression model is a multilayer perceptron neural network.

10. The high-dimensional light field detection device based on physical feature decoupling as described in claim 6, characterized in that, The multimode interference medium is one or a combination of two or more of the following media: multimode fiber, few-mode fiber, multimode waveguide, integrating sphere, and frosted glass.

Images