Material hyperspectral BRDF measuring method based on diffractive optical element
By combining diffraction optical components and deep learning methods, optimizing the DOE surface height and using a CCD camera, the problems of high spectral BRDF measurement cost and inaccurate data are solved, and efficient and economical BRDF measurement is achieved.
Patent Information
- Application Number
- CN202510165679.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-02-14
- Publication Date
- 2025-07-04
AI Technical Summary
The existing spectral BRDF measurement is costly, the process is complicated and the data is not accurate enough.
Using a method of combining diffraction optical elements and deep learning, the BRDF hyperspectral network parameters are learned by optimizing the DOE surface highly synchronously, and the diffraction optical elements and CCD cameras are used to replace the hyperspectral detector to achieve BRDF hyperspectral inversion of the material.
Significantly reduce the cost of measuring equipment, simplify the measurement process, improve measurement accuracy and efficiency, and ensure data accuracy and interpretability.
Smart Images
Figure CN120253761A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a hyperspectral BRDF measurement method, and particularly to a hyperspectral BRDF measurement method for materials based on diffractive optical elements. Background Art
[0002] The bidirectional reflectance distribution function (BRDF) is an important function for describing the scattering characteristics of an object's surface, which represents the reflection characteristics of the object's surface at any observation angle under different incident angles. It is determined by factors such as the surface roughness, dielectric constant, surface texture, polarization, and incident wavelength of the target, and has extensive applications in the fields of optics, remote sensing, computer graphics, etc. The measurement of BRDF is of great significance for understanding the reflection characteristics of an object's surface, performing optical simulations and renderings, etc.
[0003] In order to model the spectral BRDF of a material, it is necessary to collect the data of the spectral BRDF varying with wavelength at arbitrary incident angles and scattering angles; however, its computational complexity is high, and sufficient spectral BRDF measurements need to be carried out for each material; therefore, the measurement process of spectral BRDF requires a large amount of time, resulting in a high measurement cost, and the measurement process is complex. In practical applications, it is often unrealistic to measure the spectral BRDF of each material, and the measured data is not accurate enough. Summary of the Invention
[0004] In order to solve the technical problems of the high measurement cost, complex measurement process, and inaccurate measured data of the existing spectral BRDF, the present invention provides a hyperspectral BRDF measurement method for materials based on diffractive optical elements.
[0005] The inventive concept of the present invention:
[0006] A diffractive optical element is a kind of micro-nano structure optical element based on the diffraction principle, which has the advantages of light weight, small volume, and arbitrary phase function regulation. Therefore, it is often used as an optical optimization element combined with a neural network to achieve simple optical structure multi-dimensional high-quality imaging. In addition, the diffractive optical element also has the characteristic of negative dispersion, and it usually only clearly images a single wavelength at a set focal length. Based on this concept, the present invention adopts an end-to-end network jointly optimized by a diffractive optical element and deep learning for the hyperspectral BRDF data of the target material, that is, while optimizing the surface height of the DOE, synchronously learning the network parameters for BRDF hyperspectrum to realize the inversion of the material BRDF hyperspectrum. Using the optimized diffractive optical element and CCD camera to replace the spectral detector to collect hyperspectral BRDF data can greatly reduce the volume of the measurement equipment, lower the overall cost, and improve the accuracy and precision.
[0007] To achieve the above object and complete the above inventive concept, the present invention adopts the following technical solutions:
[0008] A method for measuring the hyperspectral BRDF of materials based on diffractive optical elements, characterized in that it includes the following steps:
[0009] Step 1: Perform multi-band measurement on the material to be measured through a BRDF hyperspectral tester to obtain the original BRDF data;
[0010] Step 2: Perform data fitting processing on the original BRDF data through a BRDF model to obtain the BRDF simulation data set of the material to be measured; the BRDF simulation data set includes BRDF parameters corresponding to different wavelengths;
[0011] Step 3: Establish a convolutional network model, introduce a random DOE phase distribution and model weight parameters into the convolutional network model, and perform initialization of the convolutional network model, and set the main wavelength and wavelength interval threshold of the convolutional network model; the main wavelength is any wavelength in the BRDF simulation data set; the wavelength interval threshold is less than the mean square deviation of the DOE phases corresponding to different BRDF parameters in the BRDF simulation data set;
[0012] Step 4: Input the BRDF simulation data set into the set convolutional network model, take the BRDF parameter corresponding to the main wavelength as a reference, map the BRDF parameters corresponding to the remaining different wavelengths to the corresponding phases in the DOE phase distribution, optimize the DOE phase distribution, so that the difference between any remaining wavelength and the main wavelength is greater than or equal to the wavelength interval threshold, and separate its corresponding BRDF parameter from the BRDF parameter of the main wavelength; while optimizing, the model weight parameters are continuously updated, and finally the optimal DOE phase distribution and the optimal model weight parameters are obtained, completing the optimization of the neural network model;
[0013] Step 5: Process a diffractive optical element based on the optimal DOE phase distribution;
[0014] Step 6: Perform BRDF measurement of any wavelength on the material to be measured through the processed diffractive optical element combined with a CCD camera to obtain the BRDF parameters of any wavelength;
[0015] Step 7: Input the BRDF parameters of any wavelength into the updated and optimized convolutional network model, and perform data inversion through the optimal model weight parameters to obtain the BRDF parameters of the material to be measured at different wavelengths.
[0016] Further, in step 4, the optimizing the DOE phase distribution so that the difference between any remaining wavelength and the main wavelength is greater than or equal to the wavelength interval threshold, and separating its corresponding BRDF parameter from the BRDF parameter of the main wavelength; specifically:
[0017] Compare the remaining wavelengths with the main wavelength respectively;
[0018] If the interval between the remaining wavelength and the main wavelength is greater than or equal to the wavelength interval threshold, separate the corresponding BRDF parameter from the BRDF parameter of the main wavelength;
[0019] If the interval between the remaining wavelength and the main wavelength is less than the wavelength interval threshold, optimize the DOE phase distribution to make its interval from the main wavelength greater than or equal to the wavelength interval threshold, and then separate the corresponding BRDF parameter from the BRDF parameter of the main wavelength.
[0020] Further, in step 4:
[0021] When optimizing the DOE phase distribution, a random noise needs to be introduced into the DOE phase distribution; the magnitude of the random noise is 0-5% of the maximum phase in the randomly introduced DOE phase distribution.
[0022] Further, in step 3:
[0023] The convolutional network model is a symmetric convolutional network model.
[0024] Further, in step 3:
[0025] The established symmetric convolutional network model includes a fully connected layer and a plurality of convolutional blocks connected in sequence from the input end to the output end; the number of convolutional kernels of the plurality of convolutional blocks is symmetrically arranged, and the larger the number of convolutional kernels, the closer to the axis of symmetry. The output end of one convolutional block with a smaller number of convolutional kernels is connected to the input end of the fully connected layer, the output end of the fully connected layer is the output end of the symmetric convolutional network model, and the input end of the other convolutional block with a smaller number of convolutional kernels is the input end of the symmetric convolutional network model.
[0026] Further, in step 3:
[0027] Each convolutional block respectively includes a convolutional layer, a regularization layer and an activation layer arranged in sequence from the input end to the output end.
[0028] Further, in step 3:
[0029] The activation function adopted by the activation layer is the ReLU function.
[0030] Further, step 1 is specifically:
[0031] Light of different wavelengths is emitted by a composite-color light source and irradiated onto the material to be measured at different angles. After diffraction by the material to be measured, the diffracted light is measured in multiple bands by a BRDF hyperspectral tester to obtain the original BRDF data; the original BRDF data includes wavelength, incident zenith angle, observation zenith angle, observation azimuth angle, and BRDF parameters.
[0032] Further, in step 2:
[0033] The BRDF model is a five-parameter semi-empirical statistical model;
[0034] The optimizer in the five-parameter semi-empirical statistical model is the Adam optimizer.
[0035] Further, in step 5:
[0036] The diffractive optical element is a Fresnel lens or a Fresnel zone plate.
[0037] Advantages of the present invention:
[0038] 1. After data fitting processing of the original BRDF data of the five-parameter semi-empirical statistical model in the present invention, it is input into the established convolutional network model and corresponding settings are made to process the diffractive optical element by optimizing the obtained DOE phase distribution. The innovative combination of the processed diffractive optical element and the CCD camera as a detector successfully replaces the costly hyperspectral detector. This solution not only benefits from the optimized design of the diffractive optical element but also because the manufacturing process of processing the diffractive optical element based on the optimal DOE phase distribution is mature, thus significantly reducing the cost; in addition, the procurement and replacement process of the diffractive optical element is simple, further improving the overall economy.
[0039] 2. The present invention can obtain the BRDF parameters of the material to be measured at different wavelengths by single-wavelength inversion, greatly simplifying the measurement process; during the measurement process, there is no need to frequently replace the light source, and the backend data acquisition is also more simple and fast, thus significantly improving the measurement efficiency.
[0040] 3. The present invention uses a five-parameter semi-empirical statistical model to perform data fitting processing on the original BRDF data at different wavelengths, and combines the dual advantages of the physical model and the empirical model to ensure that the obtained BRDF simulation data set is not only accurate and reliable but also highly interpretable; and greatly improves the accuracy of the measurement results; this comprehensive data processing method provides a solid foundation for in-depth analysis and application of the data. Description of the Drawings
[0041] Figure 1 is a flowchart of an embodiment of a method for measuring the hyperspectral BRDF of a material based on a diffractive optical element according to the present invention;
[0042] Figure 2 It is a schematic structural diagram of the symmetric convolutional network model in the embodiment of the present invention;
[0043] Figure 3 It is the phase distribution diagram of the diffractive optical element in the embodiment of the present invention; among them, (a) is the initial phase distribution diagram, and (b) is the optimized phase distribution diagram;
[0044] Figure 4 It is a schematic diagram of the normal dispersion (c) of the ordinary refractive lens and the negative dispersion (d) of the diffractive optical element in the embodiment of the present invention;
[0045] Figure 5 It is a schematic diagram of the hyperspectral micro-angle BRDF measurement of the diffractive optical element in the embodiment of the present invention; among them, (e) is the schematic diagram of the BRDF spectrum measurement of the diffractive optical element, and (f) is the schematic diagram of the BRDF micro-angle measurement of the diffractive optical element;
[0046] Figure 6 It is a curve graph of the fitting result of the three-wavelength 0-180° BRDF in the third embodiment of the present invention;
[0047] Figure 7 It is a curve graph of the fitting result of the three-wavelength -85-90° BRDF in the embodiment of the present invention;
[0048] Figure 8 It is a curve graph of the inversion result of the three-wavelength 0-180° BRDF in the embodiment of the present invention;
[0049] Figure 9 It is a curve graph of the inversion result of the three-wavelength -85-90° BRDF in the embodiment of the present invention.
[0050] Reference numerals in the drawings:
[0051] 1-diffractive optical element, 2-CCD camera, 3-light source, 4-first focus, 5-second focus, 6-third focus. Detailed implementation manners
[0052] Next, the technical solutions of the present invention will be clearly and completely described in conjunction with the drawings and embodiments. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.
[0053] A method for measuring the hyperspectral BRDF of materials based on a diffractive optical element provided by an embodiment of the present invention, as Figure 1 shown, the method includes the following steps:
[0054] Step 1: Lights of different wavelengths are emitted by a composite-color light source and irradiated onto the material to be measured at different angles. After diffraction by the material to be measured, the diffracted light is measured in multiple bands by a BRDF hyperspectral tester to obtain the original BRDF data. The original BRDF data includes wavelength, incident zenith angle, observation zenith angle, observation azimuth angle, and BRDF parameters.
[0055] Z306 extinction paint is a spraying material widely used in the aerospace field. In order to understand the suppression of stray light on the surface of the light shield after spraying treatment, it is necessary to understand the scattering characteristics of the surface of the light shield sprayed with Z306 on stray light. In this embodiment, the measured BRDF spectral data of Z306 extinction paint on carbon fiber is selected as the original BRDF data.
[0056] Step 2: The original BRDF data is processed by data fitting through a five-parameter semi-empirical statistical model to obtain the BRDF simulation data set of the material to be measured. The BRDF simulation data set includes BRDF parameters corresponding to different wavelengths. The optimizer in the five-parameter semi-empirical statistical model is the Adam optimizer.
[0057] The gradient descent algorithm of the five-parameter semi-empirical statistical model is used to fit the BRDF parameters of the material corresponding to different wavelengths, that is, the hyperspectral BRDF values representing real information.
[0058] The formula of the five-parameter semi-empirical statistical model is:
[0059]
[0060] Among them, k b , k r , k d , a, and b are unknowns, k b is the specular reflection coefficient, and k d is the diffuse reflection coefficient.
[0061] Using the gradient descent optimization algorithm, the optimizer is set as the Adam optimizer, the number of batches is 1500, the number of training rounds is 3000 rounds, the loss function is set to 0.05, and the incident angles and azimuth angles θ i , θ r of the incident light corresponding to three wavelengths, and the reflection angles and azimuth angles of the reflected light are 0° and -85°, -180°, -85° to 85°, -90° to 90° respectively. The reflection angle sampling interval is 1° and the azimuth angle sampling interval is 10°. After optimization, k b = 0.5587, k r = 0.2385, b = 0.4837, a = 0.0654, k d = -2.7409. The gradient descent fitting results of the five-parameter semi-empirical statistical model are as shown in Figure 6 andFigure 7 as shown
[0062] The gradient descent optimization algorithm of the five-parameter semi-empirical statistical model is used to fit the incident angles and azimuth angles θ i 、θ r of the incident light corresponding to the wavelengths of 632.8 nm, 355 nm, and 932 nm, and the reflection angles and azimuth angles of the reflected light are 0°, -180°, -85° to 85°, -90° to 90° respectively. The minimum mean square errors (Mean Square Error, MSE) are 0.0005372, 0.0004683, and 0.0003201 respectively, and the fitting accuracy is 0.001. The results of another group of data are basically the same as the above, and the minimum mean square error and fitting accuracy are better than other algorithms.
[0063] Step 3: Establish a convolutional network model. Introduce a random DOE phase distribution and model weight parameters into the convolutional network model, and initialize the convolutional network model. Set the main wavelength and wavelength interval threshold of the convolutional network model; the main wavelength is any wavelength in the BRDF simulation dataset; the wavelength interval threshold is less than the mean square deviation of the DOE phases corresponding to different BRDF parameters in the BRDF simulation dataset.
[0064] As Figure 2 shown, the established convolutional network model is a symmetric convolutional network model. The symmetric convolutional neural network includes a fully connected layer and multiple convolutional blocks connected in sequence from the input end to the output end; the number of convolutional kernels of the multiple convolutional blocks is symmetrically set, and the larger the number of convolutional kernels, the closer it is to the axis of symmetry. The output end of one convolutional block with a smaller number of convolutional kernels is connected to the input end of the fully connected layer, and the output end of the fully connected layer is the output end of the symmetric multi-processing neural network model. The input end of another convolutional block with a smaller number of convolutional kernels is the input end of the symmetric multi-processing neural network model. Each convolutional block includes a convolutional layer, a regularization layer, and an activation layer arranged in sequence from the input end to the output end. The activation function used in the activation layer is the ReLU function, which is a function applied to the output of each neuron in the neural network model. It converts the linearly combined input signal into a non-linear output, thereby introducing non-linearity into the neural network.
[0065] The model structure in the neural network model refers to the arrangement, connection method, and hierarchical structure of neurons (nodes) in the network. These structural characteristics jointly determine how the network processes input data, extracts features, and makes predictions or classifications. The selection and design of the model structure have a decisive impact on the performance of the neural network. The current main role of the symmetric convolutional network model is wavelength inversion and improving fitting accuracy. It has few model parameters, so the model architecture is relatively simple, and the function can be extended by replacing the model later.
[0066] In an application, a symmetric convolutional network model can be regarded as a "black box" with an input end and an output end, and multiple computing units inside. By giving an input, the "black box" completes the mapping from the input to the output by extracting the features of the input data. At the same time, the BRDF measurement data used consists of five parts, namely wavelength λ, incident zenith angle θ i , observation zenith angle θ r , observation azimuth angle and the measured BRDF value f0.
[0067] The model input is wavelength λ, incident zenith angle θ i , observation zenith angle θ r and observation azimuth angle The output is the measured BRDF value f0. The optimizer selected is Adam. The main role of the optimizer is to adjust the parameters of the model according to the loss function (0.01) of the model to minimize the loss function and improve the performance and generalization ability of the model; Adam adjusts the learning rate of each parameter by calculating the first moment estimate (i.e., the mean) and the second moment estimate (i.e., the uncentered variance) of the gradient, and has strong robustness.
[0068] The replacement of the BRDF fitting model and the parameter optimization algorithm will not have too much impact on the final fitting result. Using other models (such as Torrance-Sparrow, Cook) and optimization algorithms (simulated annealing algorithm, particle swarm optimization algorithm, etc.) can still obtain similar results, and the difference is only the difference in fitting accuracy. This embodiment preferably selects a five-parameter semi-empirical statistical model.
[0069] Step 4: Input the BRDF simulation data set into the set convolutional network model, and map the BRDF parameters corresponding to the remaining different wavelengths to the corresponding phases in the DOE phase distribution based on the BRDF parameters corresponding to the main wavelength;
[0070] Compare the remaining wavelengths with the main wavelength respectively; if the interval between the remaining wavelength and the main wavelength is greater than or equal to the set wavelength interval threshold, separate the BRDF parameters corresponding to it from the BRDF parameters of the main wavelength; if the interval between the remaining wavelength and the main wavelength is less than the set wavelength interval threshold, optimize the DOE phase distribution to make its interval with the main wavelength greater than or equal to the set wavelength interval threshold, and then separate the BRDF parameters corresponding to it from the BRDF parameters of the main wavelength.
[0071] While optimizing, the model weight parameters are continuously updated, and finally the optimal DOE phase distribution and the optimal model weight parameters are obtained to complete the optimization of the neural network model; when optimizing the DOE phase distribution, a random noise also needs to be introduced into the DOE phase distribution; the size of the random noise is 0-5% of the maximum phase in the randomly introduced DOE phase distribution.
[0072] According to the negative dispersion property of the diffractive optical element, a specific wavelength is selected as the main acquisition wavelength of the diffractive optical element (in this paper, the main wavelength is selected as 632.8 nm), and the results of mapping the BRDF parameters corresponding to the main acquisition wavelength and the fitting of the five-parameter semi-empirical statistical model are mutually mapped. The wavelength interval thresholds for different wavelengths are set, and the DOE phase distribution is optimized so that the mean square error of the phases of the BRDF parameters corresponding to different wavelengths of the diffractive optical element is greater than the interval threshold, that is, after optimization, the diffractive optical element separates the BRDF parameters of different wavelengths.
[0073] Step 5: Fabricate the diffractive optical element based on the optimal DOE phase distribution. The phase distribution of the fabricated diffractive optical element is as Figure 3 shown; the diffractive optical element can be any diffractive lens that works based on the diffraction principle, such as a Fresnel lens or a Fresnel zone plate.
[0074] In this embodiment, a parallel plate is selected as the diffractive optical element to be optimized. Considering that the maximum processing height in the current mature processing technology is 1 μm and the minimum processing unit of the lithography equipment is 0.544 μm, the initial height of the parallel plate is set to 0.5 μm, the size is 5.99 mm, the focal length f = 60 mm, the propagation distance z = 50 mm, the sampling main wavelength is 632.8 nm, the sampling interval is 2.7225 μm, the refractive index of the corresponding photoresist is 1.604, the resolution size is 2204×2204, and the extended wavelengths are 355 nm and 932 nm.
[0075] After the diffractive optical element is optimized, it is combined with a CCD camera and used as a detector in the actual measurement device to collect the BRDF parameters corresponding to different incident and scattering angles of the diffractive optical element at any wavelength.
[0076] Step 6: Use the fabricated diffractive optical element combined with a CCD camera to measure the BRDF of the material to be measured at any wavelength, and obtain the BRDF parameters at any wavelength.
[0077] As Figure 4 shown, the DOE adopted in this embodiment is the core element of the detection scheme. The dispersion of a common refractive lens mainly stems from the difference in the propagation speeds of light with different wavelengths in the medium, that is, the refractive index of the material changes with the wavelength, usually generating positive dispersion, that is, the focal length of long-wavelength light is longer than that of short-wavelength light. The dispersion of the diffractive optical element, that is, the DOE, mainly depends on the phase modulation of the light wave by the microstructure on its surface and has negative dispersion characteristics. The direction of the change of its focal length with the wavelength is opposite to that of a common refractive lens.
[0078] As Figure 5As shown in the figure, in this embodiment, the diffractive optical element 1 and the CCD camera 2 are combined as a detector. With the optical center of the CCD camera as the center of the YZ axis of the detection coordinate system, hyperspectral micro-angle measurement is constituted. In BRDF measurement, a polychromatic light can be used as the measurement light source 3. The polychromatic light source emits lights of different wavelengths, which are converged at different positions on the optical axis (the first focus 4, the second focus 5, and the third focus 6) through the diffractive optical element. During the optimization process, with the main wavelength as the center, the convergence positions of the remaining wavelengths are "pulled away". The symmetric convolutional network model can inversely deduce the BRDF parameters of the remaining wavelengths at different incident angles and detection angles according to the deviation degree between the main wavelength and the remaining wavelengths. In the backward propagation of the diffractive optical element, with the single measurement benchmark at the center of the optical axis of the diffractive optical element, the angle between a single pixel and the optical axis center is θ, Then the measurable angle range of the sensor image plane is: to dx and dy are the pixel sizes in the x and y directions of the sensor, and m and n are the numbers of pixels in the x and y directions of the sensor, that is, the BRDF micro-angle measurement of the diffractive optical element.
[0079] Step 7: Input the BRDF parameters of any wavelength into the updated and optimized symmetric convolutional network model, and perform data inversion through the optimal model weight parameters to obtain the BRDF parameters of the material to be measured at different wavelengths.
[0080] Use the above optimized diffractive optical element and CCD camera to form a detector assembly, collect the incident zenith angle and azimuth angle corresponding to different incident angles and reflection angles at the wavelength of 632.8 nm, and input them into the network for inversion. Taking the above three wavelengths as an example, the inversion results are as shown in 8 and Figure 9 as shown.
[0081] In actual measurement, for the Z306 extinction paint on carbon fiber, when the wavelength is 632.8 nm, the incident zenith angle θ i = 0°, and the observation zenith angle θ r = -180°, the maximum value = 0.0037, the maximum value of the simulation data = 0.0032. Without considering measurement errors, the measured value is 15.625% higher than the true value; when the wavelength is 632.8 nm, the maximum value = 0.0040, the maximum value of the simulation data = 0.00376, and the measured value is 6.38% higher than the true value; when the wavelength is 932 nm, the minimum value = 0.00087, the minimum value of the simulation data = 0.000683, and the measured value is 27.3% higher than the true value. This may be because in actual measurement, the environment, electronic noise, and processing errors of the diffractive optical element cause the overall value to be larger, but the trend coincidence degree of the three groups of wavelengths is excellent, and the overall data detection accuracy is high.
[0082] In Figure 1Among them, the yellow box is the diffraction optical element optimization link. The original data set of actual measurement or equivalent BRDF is simulated and imaged through the diffraction optical element. The result of the simulation imaging is not the optical image of the material, but the BRDF simulation data set corresponding to the incident angle and scattering angle of the material. Taking the optical center of the CCD camera as the axis of symmetry, the corresponding micro-angle BRDF matrix in the CCD image plane can be obtained.
[0083] As described above, it is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention. Therefore, the protection scope of the present invention should be subject to the protection scope of the claims.
Claims
1. A method for measuring the hyperspectral BRDF of materials based on diffractive optical elements, characterized in that, It includes the following steps: Step 1: Perform multi-band measurement on the material to be measured by a BRDF hyperspectral tester to obtain the original BRDF data; Step 2: Perform data fitting on the original BRDF data through a BRDF model to obtain the BRDF simulation data set of the material to be measured; the BRDF simulation data set includes BRDF parameters corresponding to different wavelengths; Step 3: Establish a convolutional network model, introduce a random DOE phase distribution and model weight parameters into the convolutional network model, and perform initialization of the convolutional network model, and set the main wavelength and wavelength interval threshold of the convolutional network model; The main wavelength is any wavelength in the BRDF simulation data set; the wavelength interval threshold is less than the mean square error of the DOE phases corresponding to different BRDF parameters in the BRDF simulation data set; Step 4: Input the BRDF simulation data set into the set convolutional network model, and based on the BRDF parameters corresponding to the main wavelength, map the BRDF parameters corresponding to the remaining different wavelengths to the corresponding phases in the DOE phase distribution, optimize the DOE phase distribution, so that the difference between any remaining wavelength and the main wavelength is greater than or equal to the wavelength interval threshold, and separate the corresponding BRDF parameters from the BRDF parameters of the main wavelength; while optimizing, the model weight parameters are continuously updated, and finally the optimal DOE phase distribution and the optimal model weight parameters are obtained to complete the optimization of the neural network model; Step 5: Process the diffractive optical element based on the optimal DOE phase distribution; Step 6: Perform BRDF measurement of any wavelength on the material to be measured by combining the processed diffractive optical element with a CCD camera to obtain the BRDF parameters of any wavelength; Step 7: Input the BRDF parameters of any wavelength into the updated and optimized convolutional network model, and perform data inversion through the optimal model weight parameters to obtain the BRDF parameters of the material to be measured at different wavelengths.
2. The method for measuring the material hyperspectral BRDF based on a diffractive optical element according to claim 1, wherein In Step 4, the optimization of the DOE phase distribution makes the difference between any remaining wavelength and the main wavelength greater than or equal to the wavelength interval threshold, and separates the corresponding BRDF parameters from the BRDF parameters of the main wavelength; specifically: Compare the remaining wavelengths with the main wavelength respectively; If the interval between the remaining wavelength and the main wavelength is greater than or equal to the wavelength interval threshold, separate the corresponding BRDF parameters from the BRDF parameters of the main wavelength; If the interval between the remaining wavelength and the main wavelength is less than the wavelength interval threshold, optimize the DOE phase distribution to make its interval with the main wavelength greater than or equal to the wavelength interval threshold, and then separate the corresponding BRDF parameters from the BRDF parameters of the main wavelength.
3. The method for measuring the material hyperspectral BRDF based on a diffractive optical element according to claim 2, wherein In Step 4: When optimizing the DOE phase distribution, a random noise needs to be introduced into the DOE phase distribution; the size of the random noise is 0-5% of the maximum phase in the randomly introduced DOE phase distribution.
4. The method for measuring the material hyperspectral BRDF based on a diffractive optical element according to any one of claims 1-3, characterized in that In Step 3: The convolutional network model is a symmetric convolutional network model.
5. The method for measuring the material hyperspectral BRDF based on a diffractive optical element according to claim 4, wherein In Step 3: The established symmetric convolutional network model includes a fully connected layer and multiple convolutional blocks connected in sequence from the input end to the output end; the number of convolutional kernels of the multiple convolutional blocks is symmetrically set, and the larger the number of convolutional kernels, the closer it is to the axis of symmetry. The output end of one convolutional block with a smaller number of convolutional kernels is connected to the input end of the fully connected layer, the output end of the fully connected layer is the output end of the symmetric convolutional network model, and the input end of another convolutional block with a smaller number of convolutional kernels is the input end of the symmetric convolutional network model.
6. The method for measuring the material hyperspectral BRDF based on a diffractive optical element according to claim 5, characterized in that In step 3: Each convolutional block respectively includes a convolutional layer, a regularization layer, and an activation layer arranged in sequence from the input end to the output end.
7. The method for measuring the material hyperspectral BRDF based on a diffractive optical element according to claim 6, wherein In step 3: The activation function used in the activation layer is the ReLU function.
8. The method for measuring the material hyperspectral BRDF based on a diffractive optical element according to claim 7, characterized in that, Step 1 is specifically: Light of different wavelengths is emitted by a polychromatic light source and irradiated onto the material to be measured at different angles. After being diffracted by the material to be measured, the diffracted light is measured in multiple bands by a BRDF hyperspectral tester to obtain the original BRDF data; the original BRDF data includes wavelength, incident zenith angle, observation zenith angle, observation azimuth angle, and BRDF parameters.
9. The method for measuring the material hyperspectral BRDF based on a diffractive optical element according to claim 8, wherein In step 2: The BRDF model is a five-parameter semi-empirical statistical model; The optimizer in the five-parameter semi-empirical statistical model is the Adam optimizer.
10. The method for measuring the material hyperspectral BRDF based on a diffractive optical element according to claim 9, wherein In step 5: The diffractive optical element is a Fresnel lens or a Fresnel zone plate.