A high-precision phase extraction method for orthogonal lateral shearing interferograms based on generative multi-modal AI

Abstract

This invention relates to the fields of optical interferometry and artificial intelligence image processing, specifically to a high-precision phase extraction method for orthogonal transverse shearing interferograms based on generative multimodal AI. This method integrates the spatial texture of the interferogram with the constraints of the orthogonal shearing mathematical model through a multimodal feature fusion module, constructing a hybrid architecture of Generative Adversarial Network (GAN) and U-Net with a physical prior loss function. This achieves an end-to-end nonlinear mapping from the interferogram to the wrapped phase, and combines differential Zernike polynomial (DZF) to complete high-precision reconstruction of the absolute phase. Experimental results show that in low signal-to-noise ratio orthogonal transverse shearing interferograms, the proposed method reduces the root mean square error (RMSE) of phase extraction by 42.7% compared to the traditional DZF method, and improves the spatial resolution to λ / 100, providing an efficient and high-precision phase extraction scheme for optical wavefront detection.

Description

Technical Field

[0001] This invention relates to the fields of optical interferometry and artificial intelligence image processing technology, specifically to a high-precision phase extraction method for orthogonal transverse shearing interferograms based on generative multimodal AI. Background Technology

[0002] Orthogonal transverse shearing interferometry is widely used in optical system wavefront detection and laser beam quality assessment due to its non-contact and high dynamic range characteristics. Its core principle is to extract the fringe phase information from the interferogram to invert the distortion characteristics of the wavefront under test.

[0003] The essence of orthogonal transverse shearing interference is to superimpose the wavefront to be measured with its own wavefront sheared in orthogonal directions to form interference fringes. Let the shearing amounts in the x and y directions be respectively... , Then the interference intensity distribution can be expressed as: ,in Background light intensity, To detect the wavelength of light, The phase of the wavefront to be measured.

[0004] Traditional phase extraction methods, such as differential Zernike polynomial fitting and Fourier transform, rely on strict physical model assumptions. When there is noise interference or uneven fringe contrast, they are prone to problems such as phase unwrapping error propagation and loss of details, making it difficult to meet the requirements of high-precision wavefront detection.

[0005] With the development of artificial intelligence technology, phase extraction methods based on convolutional neural networks have been gradually applied. Currently, phase extraction methods based on convolutional neural networks (CNNs) have achieved initial applications.

[0006] The patent publication number "CN121300025A" discloses "A Deep Learning-Based Method for Phase Recovery of Transverse Shearing Interferometry Wavefronts," which directly learns the mapping relationship from the interferogram to the phase through a convolutional neural network. However, this method suffers from the following problems: 1. It is a purely data-driven model, lacking deep integration with the physical mechanisms of orthogonal shearing interferometry, resulting in weak generalization ability. 2. Because the network relies solely on statistical data to learn the mapping relationship, it does not incorporate physical constraints and high-frequency feature perception mechanisms for wavefront distortion, making it difficult to capture the high-frequency components and local details of complex distorted wavefronts. Furthermore, it lacks adaptability to low signal-to-noise ratio and large-amplitude distortion scenarios, thus limiting accuracy in complex distorted wavefront detection scenarios. 3. Due to the lack of optimization of the network structure and inference process, accuracy is limited in complex distorted wavefront detection scenarios, and processing efficiency is low, making it difficult to meet real-time requirements. Therefore, there is an urgent need for an intelligent phase extraction method that integrates physical priors and multimodal features to improve the accuracy and robustness of phase extraction from orthogonal transverse shearing interferograms.

[0007] The patent publication number "CN121280959A" discloses a "multimodal interferometric phase extraction method" that integrates image features and some physical parameters. However, this method suffers from the following problems: 1. The physical parameter fusion is superficial, lacking a systematic physical constraint mechanism. Only some physical parameters are incorporated as auxiliary features, without deeply embedding the orthogonal shearing interferometric light propagation physical model into the network architecture or loss function. This results in a lack of physical consistency guidance for multimodal fusion, making the model mapping relationship prone to deviating from actual interferometric laws. Performance fluctuates significantly when generalized to solar observation scenarios with different shearing amounts and turbulence intensities. 2. The multimodal data processing mechanism is incomplete, lacking a specific suppression strategy designed for the noise characteristics of solar observations. Noise between modes is easily coupled and amplified, leading to phase jump misjudgments in low signal-to-noise ratio scenarios. 3. It fails to enhance the perception of high-frequency aberrations in solar observations, making it difficult to balance global wavefronts and local details. Phase recovery details are severely lost in complex distortion scenarios, and 2π jump errors accumulate during unwrapping, failing to meet the correction requirements of solar adaptive optics systems.

[0008] Therefore, developing a wavefront phase recovery method that deeply integrates physical constraints, adapts to complex solar observation scenarios, and combines high precision with strong robustness has become an urgent technical problem to be solved. Summary of the Invention

[0009] To address the shortcomings of existing technologies, such as weak generalization ability of pure data-driven models, limited accuracy and weak robustness in complex distortion wavefront detection scenarios, this invention provides a high-precision phase extraction method for orthogonal transverse shearing interferograms based on generative multimodal AI.

[0010] To achieve the objectives of this invention, the technical solution provided by this invention is: a high-precision phase extraction method for orthogonal transverse shearing interferograms based on generative multimodal AI, comprising the following steps:

[0011] Step 1: The orthogonal transverse shearing interferogram is preprocessed using image multimodal data, and through physical prior mode construction and multimodal feature fusion, a fused feature tensor is obtained:

[0012] The feature fusion employs an attention mechanism fusion method, resulting in the following fused feature tensor:

[0013] ,

[0014] in: , , Attention weights for adaptive learning of the model;

[0015] To extract the image feature tensor, it can be expressed as: ,in For the number of channels, , Image size;

[0016] The physical characteristic tensor, constructed based on a mathematical model of orthogonal transverse shearing interference, is expressed as: ,in For physical parameter dimensions;

[0017] Step 2: Construct a generative AI model that combines U-Net and generative adversarial networks, design the overall loss function, pre-train the model using a simulation dataset, and then fine-tune it using an experimental dataset.

[0018] Total loss function for adversarial loss: ,

[0019] Among them: reconstruction loss is The physical constraint loss is , Using WGAN-GP loss, , , These are the loss weighting coefficients;

[0020] Step 3: Embed a lightweight untangling module at the generator output of the generative AI model constructed in Step 2 to output the absolute phase. Combined with the differential Zernike polynomial algorithm, high-precision reconstruction of the wavefront phase is achieved;

[0021] The difference Zernike polynomial algorithm is as follows:

[0022] ,

[0023] in: These are the coefficients of the Zernike polynomial. These are n-order Zernike basis functions;

[0024] Furthermore, the preprocessing described in step one above includes normalization, pre-denoising, and multi-scale convolutional feature extraction. The normalization algorithm is as follows:

[0025] ,

[0026] in: , These represent the minimum and maximum interference light intensities, respectively.

[0027] Furthermore, in step one above, the pre-denoising is implemented using a generator from a generative adversarial network.

[0028] Furthermore, in step one above, the physical feature tensor is constructed based on the shearing quantities in the x and y directions and the difference Zernike polynomials to encode the orthogonal shearing physical prior, wherein the Zernike polynomial encoding formula is: .

[0029] Furthermore, in step three above, the lightweight untangling module is constructed based on the least squares method, and the simplified least squares untangling formula is as follows:

[0030] ,

[0031] in: It is an identity matrix.

[0032] Compared with the prior art, the beneficial effects of the present invention are:

[0033] 1. This invention integrates interferogram image features with orthogonal shearing physical priors through an attention mechanism, enabling the generative model learning process to possess interpretable physical meaning. This solves the problem of weak generalization ability in purely data-driven models. The specific core points are as follows:

[0034] (1) Compared with the pure data-driven model, this invention dynamically couples image spatial texture features with physical priors through an attention mechanism; it adopts a dual-guided learning approach of "data features + physical logic" and follows clear physical meanings; it uses physical priors as general constraints to reduce dependence on training data for specific scenarios. Therefore, the generalization ability of this invention is significantly improved, and the performance fluctuation is small under different shearing, distortion, or noise scenarios, eliminating the need for retraining; the learning process is highly interpretable, and the phase results not only conform to data patterns but also satisfy optical propagation logic, resulting in higher reliability; it has better anti-interference capabilities, with the attention mechanism adaptively allocating weights and strengthening physical prior guidance in noisy scenarios.

[0035] (2) Compared with the traditional DZF method, this invention integrates image features and physical priors to achieve multi-dimensional information complementarity; it dynamically optimizes the fusion weights through an attention mechanism to adapt to different scenarios; it strengthens effective features and weakens noise through an attention mechanism, forming dual noise resistance in combination with physical priors. Therefore, this invention has higher accuracy in low signal-to-noise ratio scenarios, with RMSE reduced by 42.7% compared to the traditional DZF method, effectively solving the problems of stripe misjudgment and phase shift; it is more adaptable, requiring no redesign of fitting parameters, and has a wider range of applications; it restores details more accurately, taking into account both global accuracy and local fineness, and capturing subtle wavefront distortions.

[0036] 2. The multi-constraint loss function designed in this invention balances phase reconstruction accuracy and physical rationality, effectively suppressing the generation of phase artifacts and improving the ability to restore details of complex distorted wavefronts. The specific core points are as follows:

[0037] (1) This invention designs a multi-constraint loss function that simultaneously considers phase reconstruction accuracy and physical rationality; it embeds orthogonal shearing interference physical constraints into the loss function to ensure that the phase conforms to the laws of light propagation; and through the synergistic effect of multiple constraints, it directionally suppresses the generation of phase artifacts. Therefore, this invention significantly reduces phase artifacts and greatly improves the reliability of the results; it restores details more accurately in complex distortion scenarios and has a stronger ability to capture subtle wavefront distortions; and it strengthens physical consistency to avoid the problem of "data fitting but not conforming to physical laws".

[0038] (2) Compared with traditional methods (such as DZF and Fourier transform), this invention optimizes the accuracy and physical rationality of phase reconstruction through a multi-constraint loss function; enhances detail perception with the help of the loss function, achieving a wavefront reconstruction spatial resolution of λ / 100; and achieves adaptive optimization through multiple constraints, actively improving the ability to restore details in complex distortion scenes. Therefore, this invention doubles the spatial resolution, accurately capturing subtle wavefront distortions; it eliminates the need for manual parameter adjustment, making it more adaptable to complex scenes; and it balances global accuracy with local detail restoration, resulting in superior overall wavefront reconstruction quality.

[0039] 3. The U-Net+GAN hybrid model constructed in this invention achieves end-to-end phase extraction and unwrapping without complex manual preprocessing steps, resulting in high processing efficiency. It can meet the real-time requirements of optical detection and is suitable for scenarios such as optical wavefront detection and laser beam quality assessment. The specific core points are as follows:

[0040] (1) Compared with the pure data-driven model, this invention constructs a U-Net+GAN hybrid architecture, where the generator is responsible for phase extraction and unwrapping in one step, and the discriminator optimizes the physical rationality of the phase; a lightweight unwrapping module is embedded at the generator output to achieve end-to-end closed-loop processing; and complex manual preprocessing is eliminated through built-in pre-denoising and multi-scale feature extraction. Therefore, this invention has higher processing efficiency, with a total processing time of only 9.2ms for a single interferogram, meeting the requirements for real-time detection; the process is simpler, avoiding error propagation in the independent unwrapping stage and improving the stability of the results; no manual intervention is required for preprocessing parameters, reducing the operational threshold and making it more versatile for different scenarios.

[0041] (2) Compared with traditional methods (such as DZF and Fourier transform), this invention achieves end-to-end processing through a hybrid model, directly outputting the absolute phase from the interferogram without step-by-step operations; traditional methods are based on iterative fitting of fixed mathematical formulas, and the processing speed is limited by the number of iterations; parallel computing through deep learning models significantly shortens the processing time; the model of this invention adaptively processes different noise and distortion scenarios without manual parameter tuning. Therefore, the processing speed of this invention is greatly improved, with a single frame processing time of 9.2ms, which is far superior to traditional methods (usually >10ms), meeting real-time requirements; the operation is simpler, without the need for professional personnel to design preprocessing and unwrapping processes, reducing application costs; the scenario adaptability is wider, and it can be directly applied to multiple scenarios such as optical wavefront detection and laser beam quality assessment without customized adjustments.

[0042] 4. This invention first obtains image feature tensors through multimodal data preprocessing (normalization, pre-denoising, and multi-scale feature extraction) and physical prior mode construction. With physical characteristic tensor Such bimodal features can be further dynamically weighted and fused through an attention mechanism to enhance the coupling between image texture details and orthogonal shearing physical laws. This enhanced coupling is then fed into a designed U-Net+GAN hybrid generative model, trained using a multi-constraint loss function to output a high-precision wrapped phase. Finally, the absolute wavefront phase is obtained through a lightweight unwrapping module built into the generator and differential Zernike polynomial fitting. Using root mean square error (RMSE), peak signal-to-noise ratio (PSNR), and structural similarity (SSIM) as evaluation metrics, the proposed method achieves a comprehensive 38.5% improvement in phase extraction accuracy compared to the current state-of-the-art pure data-driven models. In low SNR orthogonal transverse shearing interferograms, the proposed method reduces the RMSE of phase extraction by 42.7% compared to the traditional DZF method, and improves the spatial resolution to λ / 100 (λ is the wavelength of the probe light), providing an efficient and high-precision phase extraction method for optical wavefront detection. Attached Figure Description

[0043] Figure 1 This is a schematic diagram of the optical path structure of an orthogonal transverse shearing interference system;

[0044] Figure 2 A flowchart illustrating the changes in the loss function during model training;

[0045] Figure 3 Line graphs comparing the RMSE of phase extraction using three methods at different noise levels (0%-20%);

[0046] Figure 4 The image shows an optical plate interferogram and a comparison of the phase extraction results from three methods.

[0047] Figure 5 This is a magnified comparison image of the phase extraction of a microlens array. Detailed Implementation

[0048] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0049] See Figure 1 This paper demonstrates the signal acquisition process of "Hl-neon laser → beam expansion and collimation system → sample under test → shearing interferometry module → CMOS camera," clarifying the source of the processing object (interferogram) generated by this method. The system includes the following components in a known arrangement: a hel-neon laser (wavelength 632.8 nm), a beam expansion and collimation system (beam expansion ratio 10×), a sample under test stage, a shearing interferometry module (shearing amount adjustable from 1-10 pixels), and a CMOS camera (resolution 512×512, frame rate 100fps).

[0050] This invention provides a high-precision phase extraction method for orthogonal transverse shearing interferograms based on generative multimodal AI, which specifically includes the following steps:

[0051] Step 1: The orthogonal transverse shearing interferogram is preprocessed with multimodal image data, and the physical prior modes are constructed and multimodal features are fused to obtain the fused feature tensor. The specific sub-steps are as follows:

[0052] 1.1 Image multimodal data preprocessing:

[0053] Obtain an orthogonal transverse shearing interferogram, normalize it, and map the light intensity values ​​to the [0,1] interval. The normalization algorithm is as follows:

[0054] ,

[0055] in: , These are the minimum and maximum values ​​of the interference light intensity, respectively;

[0056] A generator pre-denoising module using a generative adversarial network is employed to filter out shot noise and background stray light in the interferogram. Low-frequency global fringe distribution features and high-frequency local detail features of the interferogram are extracted through multi-scale convolutional layers, and the image feature tensor is extracted, represented as follows:

[0057] ,

[0058] in For the number of channels, , Image size;

[0059] 1.2 Construction of Physical Prior Modes:

[0060] Based on the mathematical model of orthogonal transverse shear interference, the shear amounts in the x and y directions are calculated. , By combining the difference Zernike polynomial, a set of physical constraint parameters is constructed; the physical constraint parameters are then encoded into physical feature tensors. Where D is the physical parameter dimension, this physical feature tensor is constructed based on the shearing amounts in the x and y directions and the difference Zernike polynomials to encode the orthogonal shearing physical prior. The Zernike polynomial encoding formula is as follows:

[0061] ,

[0062] In the formula: The coefficients are Zernike polynomials of orders 1 to 36, corresponding to typical wavefront distortion amplitudes such as defocus and astigmatism. d is the dimension index of the physical feature tensor, with each index corresponding to an orthogonal shearing physical constraint parameter (such as the shearing amount in the x-direction). y-direction shear amount Zernike polynomial coefficients (etc.). In this embodiment, D=38 is taken, where the first two dimensions correspond to the shearing amounts in the x and y directions, and the last 36 dimensions correspond to the coefficients of Zernike polynomials of orders 1 to 36, thereby encoding the physical laws of orthogonal shearing and wavefront distortion characteristics into a tensor form that can be recognized by the model.

[0063] 1.3 Multimodal Feature Fusion:

[0064] An attention mechanism fusion module is used to process the image feature tensor. With physical characteristic tensor Perform weighted fusion to generate a fusion feature tensor. This enables deep coupling between image texture features and physical priors.

[0065] Step 2, Generative AI Model Construction and Training, see [link / reference] Figure 2 The horizontal axis represents the training epochs, and the vertical axis represents the loss value (logarithmic scale), showing the "total loss". Reconstruction losses Physical constraint loss Combating losses The change in the training curve reflects the process of the model moving from fluctuation to convergence, specifically including the following sub-steps:

[0066] 2.1 Model Architecture Design:

[0067] A generative AI model combining U-Net and Generative Adversarial Network (GAN) is constructed, consisting of a generator G and a discriminator D. The generator adopts a U-Net encoder-decoder structure, and the encoder processes the fused feature tensor. Downsampling is performed to extract multi-scale features. The decoder restores the phase space resolution through upsampling and skip connections, and outputs a wrapped phase. The discriminator D uses a convolutional neural network structure, and the input generates the phase. With true phase To determine the physical rationality of the phase.

[0068] 2.2 Loss Function Design:

[0069] The overall loss function, which integrates reconstruction loss, physical constraint loss, and adversarial loss, is designed as follows: ,in To reconstruct the loss, a cosine generation relationship between the phase and the interference light intensity is forced to be generated; The physical constraint loss is used to constrain the generation of a mathematical model whose phase satisfies orthogonal shearing; LGAN uses WGAN-GP loss to improve the model training stability. , , The loss weighting coefficients are set to 1.0, 0.8, and 0.2 in this embodiment.

[0070] 2.3 Model Training:

[0071] A simulation dataset was constructed to generate orthogonal transverse shearing interferograms with different Zernike distortion and noise levels. These interferograms were then labeled with the actual phase data. The model was pre-trained using the simulation data and fine-tuned using a small number of experimental interferograms. The Adam optimizer was selected, with a learning rate set to 1e. -4 The training iterations are performed for 200 epochs until the loss on the validation set stabilizes. The fused feature tensor from step one is then... The generative AI model from step two is input and trained to obtain a high-precision wrapped phase. .

[0072] See Figure 3 The horizontal axis represents the noise level (0%-20%), and the vertical axis represents RMSE (unit: rad). The three curves represent: the traditional DZF method (square markers), the pure CNN method (circle markers), and the method of this invention (triangle markers), respectively. When the noise level is 10%, the RMSE of the method of this invention is 0.023 rad, which is 42.5% lower than that of the DZF method (0.040 rad).

[0073] By integrating the spatial texture of the interferogram with the constraints of the orthogonal shearing mathematical model through the multimodal feature fusion module, a hybrid architecture of generative adversarial network (GAN) and U-Net with physical prior loss function is constructed to achieve end-to-end nonlinear mapping from the interferogram to the wrapped phase, and high-precision reconstruction of the absolute phase is achieved by combining differential Zernike polynomial (DZF).

[0074] Figure 3 The comparison of phase extraction RMSE results for three methods under different noise levels is presented: the horizontal axis represents the noise level from 0% to 20%, and the vertical axis represents RMSE (unit: rad). During data preparation, 100 repeated experiments were conducted for the traditional DZF method, the pure CNN method, and the proposed method. After statistically calculating the mean and standard deviation of each data set, three comparative line graphs were plotted and error bars were added to characterize the data dispersion. A legend was also added to distinguish the three methods. The results show that as the noise level increases, the RMSE of the traditional DZF method increases exponentially, the pure CNN method increases linearly, and the proposed method shows the most gradual increase. When the noise level > 10%, the error of the DZF method rises sharply, while the proposed method maintains high accuracy, verifying its robustness advantage in low signal-to-noise ratio scenarios.

[0075] Figure 4 This image presents a comparison of optical plane interferograms and phase extraction results, including the original interferogram, the DZF method, the CNN method, and the proposed method. The first column shows the original orthogonal transverse sheared interferogram; the second column shows the absolute phase image extracted by the traditional DZF method; the third column shows the absolute phase image extracted by the pure CNN method; and the fourth column shows the absolute phase image extracted by the proposed method. The color scale uses a pseudo-color encoded Jet color map, with a phase range of [0, 2π] rad. The results show that the traditional DZF method exhibits edge phase transitions, the pure CNN method suffers from detail loss, and the proposed method shows a high degree of agreement. The workflow demonstrates the process from interferogram preprocessing, feature fusion, model output to obtaining the phase image, and finally, the output of the phase comparison image.

[0076] Figure 5 This paper presents a magnified comparison of phase extraction from a microlens array, including original region markings, local phase distributions for the three methods, and profile analysis curves. In the magnified region, a 200×200 pixel area at the center of the microlens interferogram is selected and marked with a dashed box in the original image. In the phase comparison section, local phase distributions for the DZF method, CNN method, and the proposed method are generated, followed by annotation of the radius of curvature deviation. In the profile analysis section, phase values ​​are extracted along the x-axis, and curves for the three methods are plotted and superimposed with the theoretical curve. In the accuracy calculation section, the overlap rate is calculated: 99.2% for the proposed method, 87.5% for the DZF method, and 94.3% for the CNN method, completing the local phase comparison and verification.

[0077] Step 3: Phase unwrapping and high-precision wavefront reconstruction: A lightweight unwrapping module is embedded in the generator output of the generative AI model constructed in Step 2. The wrapped phase obtained in Step 2... After processing by a lightweight unwrapping module, the absolute phase of the wavefront is output; combined with the differential Zernike polynomial algorithm, high-precision phase reconstruction and unwrapping of the wavefront phase are achieved.

[0078] 3.1 End-to-end phase unwrapping:

[0079] A lightweight untangling module based on the least squares method is embedded at the generator output to directly output the absolute phase, skipping the error propagation stage of the traditional untangling algorithm. The lightweight untangling module is constructed based on the least squares method, and the simplified least squares untangling formula is as follows:

[0080] ,

[0081] in: Using an identity matrix, end-to-end phase unwrapping is achieved, avoiding error propagation in traditional unwrapping algorithms.

[0082] 3.2 Wavefront Reconstruction Optimization:

[0083] The generated absolute phase gradient is fused with the differential Zernike polynomial method to achieve high-precision reconstruction of the wavefront phase and correct the global distortion error of the wavefront.

[0084] Specific application example 1:

[0085] This embodiment uses the orthogonal transverse shearing interferogram of a conventional distorted wavefront of an optical flat plate to verify the high-precision extraction capability of the method of the present invention in the basic scenario of phase extraction of a conventional distorted wavefront.

[0086] Step 1: Multimodal data preprocessing, physical prior mode construction, and feature fusion:

[0087] 1.1 Orthogonal transverse shearing interferograms generated by a helium-neon laser with a wavelength of λ=632.8nm were acquired, with an image resolution of 512×512 pixels. The interferograms were normalized to map the light intensity values ​​to the [0,1] interval. A GAN generator pre-denoising module was used to filter out 10% of shot noise. Image features were extracted through three multi-scale convolutional layers to obtain an image feature tensor with C=64 channels. .

[0088] Orthogonal transverse shear interference light intensity distribution: .

[0089] Multi-scale convolution feature extraction: , This refers to the convolutional layer number (layers 1-4). , For the first Layer convolution operation.

[0090] 1.2 The shear amounts in the x and y directions were measured.

[0091] Pixels, based on the difference Zernike polynomial, construct a set of physical constraint parameters, which are encoded as a physical feature tensor of dimension D=32. .

[0092] 1.3 An attention mechanism fusion module is adopted to... and Perform weighted fusion to generate a fusion feature tensor. .

[0093] Attention fusion weight learning: , MLP stands for Multilayer Perceptron.

[0094] Step 2: Generative AI Model Construction and Training

[0095] 2.1 The generator adopts a U-Net structure, with the encoder containing 4 convolutional layers and the decoder containing 4 transposed convolutional layers. Skip connections enable multi-scale feature fusion. The discriminator contains 5 convolutional layers and outputs a physical rationality judgment value for the phase.

[0096] U-Net generator output: .

[0097] 2.2 Weighting coefficients in the total loss function , , The adversarial training process is optimized using WGAN-GP loss.

[0098] 2.3 Construct a dataset of 10,000 simulated interferograms, including 8,000 for training, 1,000 for validation, and 1,000 for testing. Pre-train the model for 150 epochs, then fine-tune it for 50 epochs using 50 sets of experimental interferograms. The learning rate of the optimizer Adam is set to 1 e^(-1 / 2). -4 .

[0099] Step 3: Phase unwrapping and wavefront reconstruction:

[0100] 3.1 Embed a least-squares untangling module at the generator output to directly output the absolute phase;

[0101] 3.2 Combining the differential Zernike polynomial method, wavefront phase reconstruction is completed, and global distortion error is corrected.

[0102] Specific application example 2

[0103] This embodiment focuses on the orthogonal transverse shearing interferogram of a microlens array with large distortion wavefronts to verify the high-precision phase extraction capability of the method of the present invention in complex phase extraction scenarios with large distortion wavefronts.

[0104] Step 1: Multimodal data preprocessing, physical prior mode construction, and feature fusion:

[0105] 1.1 Orthogonal transverse shearing interferograms were acquired by irradiating a microlens array (focal length 5mm) with a helium-neon laser at a wavelength of λ=632.8nm. The image resolution was 512×512 pixels, and the shearing amounts in the x and y directions were [not specified]. =4 pixels =4 pixels, the noise level of the interferogram is 8%. The interferogram is normalized to map the light intensity values ​​to the [0,1] interval; shot noise is filtered out using a GAN generator pre-denoising module; high-frequency detail features are extracted through 4 multi-scale convolutional layers to obtain an image feature tensor with C=128 channels. .

[0106] 1.2 Based on the spherical phase distortion characteristics of microlens arrays, a set of physical constraint parameters is constructed using differential Zernike polynomials (orders 1-36), with a focus on strengthening the curvature constraint of the phase gradient, which is encoded as a physical feature tensor of dimension D=64. .

[0107] 1.3 An attention mechanism fusion module is adopted to... and Weighted fusion is performed, assigning higher physical prior weights to local regions of the microlens unit to generate a fused feature tensor. .

[0108] Step 2: Generative AI Model Construction and Training

[0109] 2.1 The generator adopts a deep U-Net structure, with the encoder and decoder each containing 5 convolutional / transposed convolutional layers. Skip connections add cross-layer feature fusion branches to enhance the ability to extract local phase details. The discriminator introduces a physical discrimination branch for phase curvature and additionally outputs a score for the physical rationality of the phase distribution.

[0110] 2.2 The weighting coefficients in the total loss function are adjusted to... , , Increase the weight of physical constraint loss to ensure that the generated phase conforms to the spherical distortion law of the microlens; We will still use WGAN-GP loss.

[0111] 2.3 A simulation dataset with large phase distortion (distortion amplitude 0.1λ~10λ) was constructed, consisting of 12,000 interferograms: 9,000 for training, 1,500 for validation, and 1,500 for testing. The model was pre-trained for 180 epochs, and then fine-tuned for 70 epochs using 80 sets of microlens array interferograms. The learning rate of the optimizer Adam was dynamically adjusted to 1 e^(-λ / λ). −4 ~5e −5 .

[0112] Step 3: Phase unwrapping and wavefront reconstruction

[0113] 3.1 A least-squares untangling module is embedded at the generator output to add phase continuity constraints to the edge region of the microlens unit and directly output the absolute phase.

[0114] 3.2 Combining the differential Zernike polynomial partitioning fitting method, the interferogram of the microlens array is divided into multiple independent units, and wavefront reconstruction is performed separately and then stitched together to correct the phase stitching error between units.

[0115] Partition fitting wavefront reconstruction: ,in For partition number, Let be the order of the Zernike polynomial in the k-th region. For the corresponding coefficient.

[0116] Specific application example 3

[0117] This embodiment uses the orthogonal transverse shearing interferogram of the dynamic wavefront in laser processing to verify the real-time performance and dynamic adaptability of the method of the present invention in real-time phase extraction of the dynamic wavefront.

[0118] Step 1: Multimodal data preprocessing, physical prior mode construction, and feature fusion

[0119] 1.1 Acquire orthogonal transverse shearing interferogram sequence of dynamic wavefronts during laser processing (frame rate 100fps), with a single frame resolution of 256×256 pixels, and shearing amount... =5 pixels =5 pixels, noise level is 12%. Lightweight normalization and pre-denoising processing are performed on the interferogram. Image features are extracted using two convolutional layers to obtain a lightweight image feature tensor with C=32 channels. .

[0120] 1.2 Based on the phase change characteristics of dynamic wavefronts, a time-varying physical constraint parameter set is constructed, and the constraint weights of shearing and phase gradient are updated in real time, encoded as a lightweight physical feature tensor of dimension D=16. .

[0121] 1.3 A lightweight attention fusion module is adopted, retaining only the weighted branches of key features to reduce computational complexity and generate a fused feature tensor. .

[0122] Step 2: Generative AI Model Construction and Training

[0123] 2.1 A lightweight U-Net+GAN model is adopted to compress the number of convolutional kernels in the encoder and decoder, remove redundant branches, and keep the total number of model parameters within 3.2M; the discriminator adopts a 3-layer convolutional structure to improve inference speed.

[0124] 2.2 The weighting coefficients of the total loss function are set to... While ensuring physical plausibility, priority should be given to improving the inference efficiency of the model.

[0125] 2.3 Construct a simulation interferogram sequence dataset of dynamic wavefronts, including dynamic scenarios such as phase abrupt changes and gradual changes; adopt a transfer learning strategy, fine-tune the pre-trained model based on Example 1, and the training iterations only require 50 epochs, which greatly shortens the training cycle.

[0126] Step 3: Phase unwrapping and wavefront reconstruction:

[0127] 3.1 A lightweight least squares unwrapping module is adopted, which is combined with the inter-frame phase continuity constraint of dynamic wavefront. The phase result of the previous frame is used to assist the unwrapping of the current frame, thereby reducing the error caused by dynamic phase abrupt change.

[0128] Dynamic inter-frame phase constraints: Where t is the frame number, This represents the inter-frame phase difference.

[0129] 3.2 A real-time wavefront reconstruction algorithm is adopted to directly input the phase extraction results into the wavefront inversion model to realize real-time monitoring of dynamic wavefronts.

[0130] Tests showed that the method of this invention has a single-frame interferogram processing time of only 7.8 ms, which meets the real-time detection requirement of 100 fps; the phase extraction RMSE=0.032 rad, PSNR=35.8 dB, SSIM=0.978, which verified the high accuracy and real-time performance of this method in dynamic wavefront detection scenarios.

[0131] The foregoing descriptions are merely preferred embodiments of the present invention and do not impose any limitations on the structure of the present invention. It should be understood by those skilled in the art that any modifications to the present invention are variations made within the scope of knowledge possessed by those skilled in the art, without departing from the spirit of the present invention, and all such modifications fall within the protection and disclosure scope of the present invention.

Claims

1. A high-precision phase extraction method for orthogonal lateral shearing interferograms based on generative multi-modal AI, characterized in that, Includes the following steps: Step 1: The orthogonal transverse shearing interferogram is preprocessed using image multimodal data, and through physical prior mode construction and multimodal feature fusion, a fused feature tensor is obtained: The feature fusion employs an attention mechanism fusion method, resulting in the following fused feature tensor: ， in: , , Attention weights for adaptive learning of the model; To extract the image feature tensor, it can be expressed as: ,in For the number of channels, , Image size; The physical characteristic tensor, constructed based on a mathematical model of orthogonal transverse shearing interference, is expressed as: ,in For physical parameter dimensions; Step 2: Construct a generative AI model that combines U-Net and generative adversarial networks, design the overall loss function, pre-train the model using a simulation dataset, and then fine-tune it using an experimental dataset. Total loss function for adversarial loss: , Among them: reconstruction loss is The physical constraint loss is , Using WGAN-GP loss, , , These are the loss weighting coefficients; Step 3: Embed a lightweight untangling module at the generator output of the generative AI model constructed in Step 2 to output the absolute phase. Combined with the differential Zernike polynomial algorithm, high-precision reconstruction of the wavefront phase is achieved; The difference Zernike polynomial algorithm is as follows: ， in: These are the coefficients of the Zernike polynomial. These are n-order Zernike basis functions.

2. The high-precision phase extraction method for orthogonal transverse shearing interferograms based on generative multimodal AI according to claim 1, characterized in that, The preprocessing described in step one includes normalization, pre-denoising, and multi-scale convolutional feature extraction. The normalization algorithm is as follows: ， in: , These represent the minimum and maximum interference light intensities, respectively.

3. The high-precision phase extraction method for orthogonal transverse shearing interferograms based on generative multimodal AI according to claim 1, characterized in that, The pre-denoising described in step 1 is implemented using a generator from a generative adversarial network.

4. The high-precision phase extraction method for orthogonal transverse shearing interferograms based on generative multimodal AI according to claim 1, characterized in that, The physical feature tensor described in step one is constructed based on the shearing quantities in the x and y directions and the difference Zernike polynomials, thereby encoding the orthogonal shearing physical prior. The Zernike polynomial encoding formula is as follows: .

5. The high-precision phase extraction method for orthogonal transverse shearing interferograms based on generative multimodal AI according to claim 1, characterized in that, The lightweight untangling module described in step three is constructed based on the least squares method. The simplified least squares untangling formula is as follows: ， in: It is an identity matrix.

Images