A fringe analysis method based on phase shift theory driving deep learning

Abstract

The present application relates to a kind of based on phase shift theory driving deep learning's stripe analysis method, comprising: based on random number sequence, construct data set;Wherein, data set includes: original stripe image containing noise, wrapped phase label, denominator stripe label and molecule stripe label;Based on convolutional neural network, construct multi-task coupling network model;Based on the preset composite loss function and data set, multi-task coupling network model is trained;Based on the multi-task coupling network model after training, the original stripe image to be processed is handled, obtains the wrapped phase image after stripe analysis.The present application introduces the basic principle of phase shift in model training, so that network model optimization is strictly carried out under the physical framework of stripe analysis, so as to ensure the interpretability of network model and the reliability of the results.Experimental results confirm that: the present application significantly improves the reliability, accuracy and robustness of stripe analysis based on deep learning.

Description

Technical Field

[0001] This invention relates to the fields of optical detection, image processing and artificial intelligence, and in particular to a fringe analysis method based on phase shift theory-driven deep learning. Background Technology

[0002] Fringe analysis is a crucial step in the engineering applications of modern optical measurement techniques (such as digital holographic speckle interferometry, structured light 3D imaging, and photoelasticity), and is widely used in defect detection, topography measurement, and mechanical property testing of structural components. Fringe analysis methods process interference fringe images to obtain wrapped phase results. Based on these wrapped phase results, conventional phase unwrapping methods can be used to obtain a continuous absolute phase, thus yielding the physical quantity to be measured. In actual measurements, interference fringe images are affected by speckle noise, inhomogeneous backgrounds, discontinuous phases, and varying fringe densities, leading to errors and loss of detail in the obtained wrapped phase results, thereby affecting the accuracy and robustness of phase recovery.

[0003] Conventional fringe analysis methods can be categorized into spatial phase-shifting methods, temporal phase-shifting methods, two-step phase-shifting methods, Fourier transform methods, and Hilbert transform methods based on their fundamental principles. Spatial phase-shifting methods offer strong noise resistance but have higher equipment costs and more complex structures. Temporal phase-shifting methods have lower equipment requirements but are sensitive to environmental interference. Both spatial and temporal phase-shifting methods generally require at least three images to obtain the wrapped phase. In contrast, two-step phase-shifting methods can obtain the wrapped phase using only two images, but their accuracy, runtime, and sensitivity to noise and interference need improvement. Fourier transform and Hilbert transform methods can obtain the wrapped phase using only one image, but their performance under high noise levels and complex fringe conditions requires further enhancement. Fringe analysis methods based on filtering preprocessing are another conventional approach. These methods can improve the accuracy of the results by processing the filtered fringe image; however, their efficiency is significantly reduced, and they also face parameter adjustment challenges.

[0004] In recent years, deep learning has demonstrated excellent generalization, optimization, and batch processing capabilities in image processing. However, commonly used network models still face challenges in fringe analysis. For example, U-Net-based models can directly obtain wrapped phase results, but the accuracy and reliability of the results need improvement. In contrast, Y-Net-based models generate denominator and numerator fringes and use the arctangent operator to obtain wrapped phase results, improving accuracy and reliability, but they lack the ability to directly obtain wrapped phase results. Both U-Net and Y-Net-based models rely on data-driven training in the absence of physical boundary constraints, requiring further enhancement of the interpretability of the results. Performance under conditions of high noise, non-uniform backgrounds, and complex fringe density remains unsatisfactory. Summary of the Invention

[0005] To address the aforementioned issues, this invention proposes a fringe analysis method based on phase-shift theory-driven deep learning. This method constructs a multi-task coupled network model based on a convolutional neural network to perform batch fully automatic fringe analysis on noisy raw interference fringe images, directly outputting noise-free wrapping phase results. It aims to achieve fringe analysis processing of low-quality, high-noise, and complex-shaped interference fringe images, thereby solving the problems existing in the aforementioned technologies.

[0006] To achieve the above objectives, the present invention provides the following solution: A fringe analysis method based on phase shift theory-driven deep learning includes: A dataset is constructed based on a random number sequence; wherein the dataset includes: the original stripe image with noise, wrapped phase labels, denominator stripe labels, and numerator stripe labels; Constructing a multi-task coupled network model based on convolutional neural networks; The multi-task coupled network model is trained based on a preset composite loss function and the dataset. The original stripe image is analyzed based on the trained multi-task coupled network model to obtain the wrapped phase image.

[0007] Optionally, the dataset can be constructed based on a sequence of random numbers, including: A random number sequence is generated. Based on the random number sequence, the absolute phase is obtained through two-dimensional interpolation. Based on the absolute phase and the principle of speckle interferometry, the original speckle image is obtained. Based on the original speckle image, a noisy original stripe image is obtained using a difference method; Based on the absolute phase, the phase shift method is used to obtain the wrapping phase label, denominator stripe label, and numerator stripe label; The dataset is constructed based on the noisy original stripe image, wrapped phase labels, denominator stripe labels, and numerator stripe labels.

[0008] Optionally, the multi-task coupled network model includes: a shared encoder (Encoder), a wrapping phase decoder (Decoder1), a denominator fringe decoder (Decoder2), a numerator fringe decoder (Decoder3), and wrapping phase reconstruction branches, denominator fringe reconstruction branches, and numerator fringe reconstruction branches. The shared encoder is used to downsample the input noisy raw stripe image step by step to obtain multi-scale feature maps; The wrapping phase decoder, denominator fringe decoder, and numerator fringe decoder are respectively used to upsample the multi-scale feature map step by step and fuse shallow information to obtain the corresponding wrapping phase features, denominator fringe features, and numerator fringe features. The wrapping phase reconstruction branch, denominator fringe reconstruction branch, and numerator fringe reconstruction branch output wrapping phase results, denominator fringe results, and numerator fringe results based on the wrapping phase features, denominator fringe features, and numerator fringe features, respectively.

[0009] Optionally, in the shared encoder, the noisy original stripe image is sequentially passed through a convolution module, a residual module, another convolution module, and a downsampling operation to obtain image features. At the same time, a stitching operation is used to improve the feature reuse capability, and finally a multi-scale feature map is obtained. The convolution module includes: convolution operation, batch normalization operation, and ReLU activation function.

[0010] Optionally, in the wrapping phase decoder, denominator fringe decoder, and numerator fringe decoder, the deep feature maps from the shared encoder are first subjected to a 2×2 bilinear interpolation operation, and then further fused with the shallow feature maps from the shared encoder using a concatenation operation. Subsequently, image features are obtained sequentially through a convolution module, a residual module, and another convolution module, while a concatenation operation is used to improve feature reuse capability. The convolution module includes convolution operations, batch normalization operations, and a ReLU activation function, ultimately restoring the deep features to the original input size, i.e., reconstructing feature 0. The reconstructed feature 0 includes: wrapping phase reconstruction feature 0, denominator fringe reconstruction feature 0, and numerator fringe reconstruction feature 0.

[0011] Optionally, the wrapping phase decoder introduces a convolutional block attention mechanism to improve the utilization of image features through channel attention and spatial attention mechanisms, thereby obtaining wrapping phase features; both the denominator stripe decoder and the numerator stripe decoder introduce a contextual self-attention mechanism to capture long-range dependencies and enhance the global feature representation capability, thereby obtaining denominator stripe features and numerator stripe features respectively.

[0012] Optionally, the wrapping phase reconstruction branch fuses wrapping phase features of different scales through upsampling and concatenation operations to obtain wrapping phase reconstruction feature 1; and fuses wrapping phase features of different scales through dilated convolution module, concatenation operation, and upsampling operation to obtain wrapping phase reconstruction feature 2. Subsequently, wrapping phase reconstruction feature 2 is concatenated and dilated convolution to obtain wrapping phase reconstruction feature 3. Finally, the concatenation operation is used to fuse wrapping phase reconstruction features 0, 1, 2, and 3, and convolution operation and Sigmoid activation function are used to obtain the wrapping phase result. The dilated convolution module includes dilated convolution operation, batch normalization operation, and ReLU activation function. The denominator fringe reconstruction branch fuses denominator fringe features at different scales through upsampling, concatenation, and dilated convolution to obtain denominator fringe reconstruction feature 1. Then, it uses a concatenation operation to fuse denominator fringe reconstruction feature 0 and denominator fringe reconstruction feature 1. Finally, it uses a dilated convolution module, skip connections, and a sigmoid activation function to obtain the denominator fringe result. The dilated convolution module includes dilated convolution, batch normalization, and ReLU activation function. The molecular stripe reconstruction branch adopts the same structure as the encapsulated phase reconstruction branch, and fuses molecular stripe features at different scales to obtain molecular stripe results; Optionally, the composite loss function includes: weighted average wrapping phase loss, denominator fringe loss, numerator fringe loss, and wrapping phase crossover loss; The wrapping phase loss, denominator fringe loss, numerator fringe loss, and wrapping phase crossover loss are all composed of MSE, MAE, and MS-SSIM.

[0013] Optionally, training the multi-task coupled network model includes: The difference between the output of the multi-task coupled network model and the true label is calculated based on the composite loss function. The network parameters are updated by backpropagation using gradient descent, and the weights and offsets are continuously optimized. Training stops when the training reaches the maximum number of iterations, the loss function value decreases to a stable value, or the network weights converge to the optimal value.

[0014] The beneficial effects of this invention are as follows: The proposed fringe analysis method based on phase shift theory-driven deep learning enables batch fully automatic analysis of noisy original fringe images by constructing a multi-task coupled network model based on convolutional neural networks, directly obtaining noise-free wrapped phase images.

[0015] The proposed method utilizes a shared encoder to progressively downsample noisy fringe images, obtaining multi-scale feature maps. Then, it employs a wrapper phase decoder, a denominator fringe decoder, and a numerator fringe decoder to perform upsampling and feature fusion on the feature maps, respectively. Finally, it obtains the wrapped phase, denominator fringe, and numerator fringe results through wrapper phase reconstruction branches, denominator fringe reconstruction branches, and numerator fringe reconstruction branches. Simultaneously, it utilizes the fundamental principle of phase shift to obtain the indirect wrapped phase result. During model training, the fundamental principles of fringe analysis are introduced into the optimization process of the network model using a composite loss function, enabling the network model to complete learning and training within a rigorous physical framework of fringe analysis. This improves the interpretability of the deep learning method and the reliability, accuracy, and robustness of the results.

[0016] Compared to conventional fringe analysis methods, the proposed method can automatically process multiple fringe images in batches without manual parameter adjustment, offering higher accuracy, stronger robustness, and better generalization. It maintains stable performance even in complex scenarios, demonstrating strong application value and promotion potential. This provides a technical path for the practical application and engineering implementation of modern optical measurement technologies (such as digital holographic speckle interferometry, structured light 3D imaging, and photoelasticity). Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a schematic diagram of the network model structure in an embodiment of the present invention; wherein, (a) is a schematic diagram of the overall structure of the network model; (b) is the basic module 1 of the encoder; (c) is the basic module 2 of the decoder; (d1)(d2)(d3)(d4)(d5) are reconstruction branches that enclose the phase and molecule fringes; (e1)(e2)(e3) are reconstruction branches that enclose the denominator fringes.

[0019] Figure 2 This is a schematic diagram of the loss function and model training in an embodiment of the present invention; Figure 3 The results of the method of this embodiment of the invention on a simulated non-uniform background stripe image are shown; where (a) is the original stripe image, (b) is the wrapped phase label, (c) is the absolute phase label, (d) is the wrapped phase result, (e) is the absolute phase corresponding to the wrapped phase result, and (f) is the accuracy distribution of the absolute phase. Figure 4The results of the method of this embodiment of the invention on a measured non-uniform background stripe image are shown; where (a1) and (a2) are the original stripe images, (b1) and (b2) are the wrapped phase results, and (c1) and (c2) are the absolute phases corresponding to the wrapped phase results. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0022] This embodiment proposes a fringe analysis method based on phase shift theory-driven deep learning, including: Step 1: Construct a dataset based on a random number sequence; wherein the dataset includes: the original stripe image with noise, wrapped phase labels, denominator stripe labels, and numerator stripe labels; Step 2: Construct a multi-task coupled network model based on a convolutional neural network; Step 3: Train the multi-task coupled network model based on the preset composite loss function and the dataset; Step 4: Analyze the original stripe image based on the trained multi-task coupled network model to obtain the wrapped phase image.

[0023] Specifically, in this embodiment, step 1: constructing the training dataset, the specific steps are as follows: Step 1-1: Generate a random number sequence using a computer. Based on the random number sequence, obtain the absolute phase through two-dimensional interpolation. Based on the absolute phase and the principle of speckle interferometry, obtain the original speckle image.

[0024] Step 1-2: Based on the original speckle image, a noisy original stripe image is obtained using the difference method; Steps 1-3: Based on the absolute phase, use the phase shift method to obtain the wrapping phase label, denominator stripe label, and numerator stripe label; Steps 1-4: Construct a dataset based on the noisy original stripe image, wrapped phase labels, denominator stripe labels, and numerator stripe labels. There are a total of 7000 sets of data, of which 500 sets are the validation set and 6500 sets are the training set. The size of each set is 256×256 pixels.

[0025] Specifically, in this embodiment, step 2: constructing the network model, the specific steps are as follows: This embodiment constructs a multi-task coupled network model based on a convolutional neural network to obtain the wrapping phase from a noisy raw stripe image. The overall architecture consists of a shared encoder (Encoder), a wrapping phase decoder (Decoder1), a denominator stripe decoder (Decoder2), a numerator stripe decoder (Decoder3), and wrapping phase reconstruction branches, denominator stripe reconstruction branches, and numerator stripe reconstruction branches. Figure 1 A schematic diagram of the network structure is shown. Among them, (a) is a schematic diagram of the overall structure of the network model; (b) is the basic module 1 of the encoder; (c) is the basic module 2 of the decoder; (d1)(d2)(d3)(d4)(d5) are the reconstruction branches of the phase and molecular stripes; (e1)(e2)(e3) are the reconstruction branches of the denominator stripes.

[0026] A shared encoder is used to downsample the input noisy raw stripe image step by step to obtain a multi-scale feature map. The wrapping phase decoder, denominator stripe decoder, and numerator stripe decoder are used to upsample the multi-scale feature map step by step and fuse shallow information to obtain the corresponding wrapping phase features, denominator stripe features, and numerator stripe features. The wrapping phase reconstruction branch, denominator stripe reconstruction branch, and numerator stripe reconstruction branch output the wrapping phase result, denominator stripe result, and numerator stripe result based on the wrapping phase feature, denominator stripe feature, and numerator stripe feature, respectively.

[0027] Step 2-1: Construct the shared encoder. Its basic module 1 obtains image features sequentially through a convolutional module, a residual module, another convolutional module, and downsampling operations, while employing a stitching operation (Cat) to improve feature reuse capabilities. The convolutional module consists of convolutional operations (Conv), batch normalization (BN), and the ReLU activation function (ReLU). The shared encoder processes the noisy original striped image input to obtain deep features at different scales.

[0028] Step 2-2: Construct the wrapper phase decoder (Decoder1), denominator stripe decoder (Decoder2), and numerator stripe decoder (Decoder3). All three use 2×2 bilinear interpolation for upsampling, progressively restoring deep features to the original input size. The basic module 2 first upsamples the deep feature map of the input shared encoder, then uses a concatenation operation (Cat) to further fuse the shallow feature map of the shared encoder. Subsequently, image features are obtained through a convolution module, a residual module, and another convolution module, while using a concatenation operation (Cat) to improve feature reuse. The convolution module consists of a convolution operation (Conv), a batch normalization operation (BN), and a ReLU activation function (ReLU), ultimately progressively restoring the deep features to the original input size. Among them, the wrapping phase decoder introduces a convolutional block attention mechanism, which improves the utilization of image features through channel attention and spatial attention mechanisms, and obtains the wrapping phase reconstruction feature 0; the denominator stripe decoder and the numerator stripe decoder both introduce a context self-attention mechanism to capture long-range dependencies and enhance the global feature representation capability, and obtain the denominator stripe reconstruction feature 0 and the numerator stripe reconstruction feature 0, respectively.

[0029] Steps 2-3: Construct the wrapping phase reconstruction branch, denominator fringe reconstruction branch, and numerator fringe reconstruction branch. The wrapping phase reconstruction branch fuses wrapping phase features at different scales through upsampling and stitching operations to obtain reconstructed feature 1; it also fuses wrapping phase features at different scales through a dilated convolution module, stitching, and upsampling operations to obtain reconstructed feature 2. Subsequently, reconstructed feature 2 is stitched together and dilated convolution to obtain reconstructed feature 3. Finally, reconstructed features 0, 1, 2, and 3 are fused using a stitching operation, and the wrapping phase result is obtained using convolution and the Sigmoid activation function. The dilated convolution module includes dilated convolution, batch normalization, and the ReLU activation function.

[0030] The denominator fringe reconstruction branch firstly fuses denominator fringe features at different scales through upsampling, concatenation, and dilated convolution to obtain reconstructed feature 1. Then, it uses a concatenation operation to fuse reconstructed feature 0 and reconstructed feature 1, followed by dilated convolution, skip connections, and a sigmoid activation function to obtain the denominator fringe result. The dilated convolution module includes dilated convolution, batch normalization, and ReLU activation.

[0031] The molecular stripe reconstruction branch uses the same structure as the encapsulated phase reconstruction branch to fuse molecular stripe features at different scales to obtain molecular stripe results.

[0032] Simultaneously, based on the mathematical relationship between the denominator and numerator fringes and the wrapping phase (i.e., the basic principle of phase shift analysis), the indirect wrapping phase result is calculated. Ultimately, this achieves end-to-end output from the noisy original fringe image to the wrapping phase, denominator fringes, and numerator fringes, without requiring additional post-processing steps.

[0033] Specifically, in this embodiment, step 3: train the model. Figure 2 The model training process is demonstrated, and the specific steps are as follows: Step 3-1: Initialize network parameters and load the training and validation datasets into the network model.

[0034] Step 3-2: Design a composite loss function consisting of wrapping phase loss, denominator fringe loss, numerator fringe loss, and wrapping phase crossover loss, used to train the network model and optimize network parameters. The wrapping phase loss, denominator fringe loss, numerator fringe loss, and wrapping phase crossover loss are all composed of MSE, MAE, and MS-SSIM, used to optimize the accuracy of the wrapping phase result, denominator fringe result, and numerator fringe result. The composite loss function is used to calculate the difference between the output result and the true label. Gradient descent is used for backpropagation to update the network parameters, continuously optimizing the weights and offsets, and fully utilizing the basic principles of fringe analysis to improve the performance of the network model.

[0035] Step 3-3: Stop training when the maximum number of iterations is reached, the loss function stabilizes, or the network weights converge to their optimal values. In training a convolutional neural network model, multiple iterations are required for the network to converge to the training set; continuing training will lead to varying degrees of overfitting. Since the network parameters are relatively small, data augmentation is used to prevent overfitting. Finally, the optimal weights and biases are determined and saved as the optimal network model parameters.

[0036] Specifically, this embodiment also includes step 4: model verification and testing. The specific steps are as follows: Step 4-1: After one iteration cycle, test the dataset with the currently trained model, and observe the fit and generalization of the network model based on the measurement results and performance metrics.

[0037] Step 4-2: After training, use the finally trained network model to test the noisy original stripe image to obtain noise-free wrapping phase results and complete the stripe analysis task.

[0038] Figure 3The results obtained by the proposed method in this embodiment on a simulated striped image are shown. (a) is the original striped image, (b) is the wrapped phase label, (c) is the absolute phase label, (d) is the wrapped phase result, (e) is the absolute phase corresponding to the wrapped phase result, and (f) is the accuracy distribution of the absolute phase.

[0039] Figure 4 The results obtained by the method proposed in this embodiment on the measured fringe image are shown. (a1) and (a2) are the original fringe images, (b1) and (b2) are the wrapped phase results, and (c1) and (c2) are the absolute phases corresponding to the wrapped phase results.

[0040] The results above show that the fringe analysis method proposed in this embodiment is well applicable to original fringe images with speckle noise, non-uniform background, and varying fringe density. It can directly obtain accurate and reliable wrapped phase results and has batch processing capabilities.

[0041] The fringe analysis method provided in this embodiment has the following advantages: (1) It can perform batch processing of multiple original fringe images, and the obtained results have high accuracy, good generalization, and strong robustness; (2) It does not require any complex parameter adjustment, nor any preprocessing or postprocessing operations. By processing the original fringe images containing noise, a noise-free wrapped phase image can be directly obtained; (3) The network model is interpretable, the obtained results are credible, and the proposed method has the potential for promotion and application value. The method provided in this embodiment can be widely applied to modern optical measurement technologies (such as digital holographic speckle interferometry, structured light three-dimensional imaging technology, photoelasticity, etc.), and is of great significance to optical detection, artificial intelligence and image processing technologies, as well as the fields of defect detection, morphology measurement, and mechanical property testing of structural components.

[0042] The network model in this embodiment is implemented using the PyTorch framework based on Python 3.9. The server hardware configuration used is as follows: CPU: Intel i5-13490F, RAM: 32GB, GPU: NVIDIA GeForce RTX 5070.

[0043] The experimental data used in this embodiment consists of 7000 sets of computer-simulated data, including: the original fringe image with noise, wrapped phase labels, denominator fringe labels, and numerator fringe labels, all with an image size of 256×256 pixels. After model training and testing, the method in this embodiment can obtain accurate wrapped phase results.

[0044] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A fringe analysis method based on phase shift theory-driven deep learning, characterized in that, include: A dataset is constructed based on a random number sequence; wherein the dataset includes: the original stripe image with noise, wrapped phase labels, denominator stripe labels, and numerator stripe labels; Constructing a multi-task coupled network model based on convolutional neural networks; The multi-task coupled network model is trained based on a preset composite loss function and the dataset. The original stripe image is analyzed based on the trained multi-task coupled network model to obtain the wrapped phase image.

2. The fringe analysis method based on phase shift theory-driven deep learning according to claim 1, characterized in that, The dataset, constructed based on a sequence of random numbers, includes: A random number sequence is generated. Based on the random number sequence, the absolute phase is obtained through two-dimensional interpolation. Based on the absolute phase and the principle of speckle interferometry, the original speckle image is obtained. Based on the original speckle image, a noisy original stripe image is obtained using a difference method; Based on the absolute phase, the phase shift method is used to obtain the wrapping phase label, denominator stripe label, and numerator stripe label; The dataset is constructed based on the noisy original stripe image, wrapped phase labels, denominator stripe labels, and numerator stripe labels.

3. The fringe analysis method based on phase shift theory-driven deep learning according to claim 1, characterized in that, The multi-task coupled network model includes: a shared encoder, a wrapped phase decoder, a denominator fringe decoder, a numerator fringe decoder, and a wrapped phase reconstruction branch, a denominator fringe reconstruction branch, and a numerator fringe reconstruction branch; The shared encoder is used to downsample the input noisy raw stripe image step by step to obtain multi-scale feature maps; The wrapping phase decoder, denominator fringe decoder, and numerator fringe decoder are respectively used to upsample the multi-scale feature map step by step and fuse shallow information to obtain the corresponding wrapping phase features, denominator fringe features, and numerator fringe features. The wrapping phase reconstruction branch, denominator fringe reconstruction branch, and numerator fringe reconstruction branch output wrapping phase results, denominator fringe results, and numerator fringe results based on the wrapping phase features, denominator fringe features, and numerator fringe features, respectively.

4. The fringe analysis method based on phase shift theory-driven deep learning according to claim 3, characterized in that, In the shared encoder, the noisy original stripe image is sequentially passed through a convolution module, a residual module, another convolution module, and a downsampling operation to obtain image features. At the same time, a stitching operation is used to improve the feature reuse capability, and finally a multi-scale feature map is obtained. The convolution module includes: convolution operation, batch normalization operation, and ReLU activation function.

5. The fringe analysis method based on phase shift theory-driven deep learning according to claim 3, characterized in that, In the wrapping phase decoder, denominator stripe decoder, and numerator stripe decoder, the deep feature maps from the shared encoder are first fused with the shallow feature maps from the shared encoder through a 2×2 bilinear interpolation operation and then through a stitching operation. Subsequently, image features are obtained by sequentially passing through a convolution module, a residual module, and another convolution module, while a stitching operation is used to improve feature reuse capability. The convolution module includes convolution operations, batch normalization operations, and the ReLU activation function, which ultimately restores the depth features to the original input size, i.e., reconstructed feature 0; wherein, the reconstructed feature 0 includes: wrapping phase reconstruction feature 0, denominator fringe reconstruction feature 0, and numerator fringe reconstruction feature 0.

6. The fringe analysis method based on phase shift theory-driven deep learning according to claim 5, characterized in that, The wrapped phase decoder introduces a convolutional block attention mechanism, which improves the utilization of image features through channel attention and spatial attention mechanisms to obtain wrapped phase features; Both the denominator stripe decoder and the numerator stripe decoder introduce a contextual self-attention mechanism to capture long-range dependencies and enhance the global feature representation capability, thereby obtaining denominator stripe features and numerator stripe features respectively.

7. The fringe analysis method based on phase shift theory-driven deep learning according to claim 5, characterized in that, The wrapping phase reconstruction branch fuses wrapping phase features at different scales through upsampling and concatenation operations to obtain wrapping phase reconstruction feature 1; and fuses wrapping phase features at different scales through dilated convolution, concatenation, and upsampling operations to obtain wrapping phase reconstruction feature 2. Subsequently, wrapping phase reconstruction feature 2 is concatenated and dilated convolution to obtain wrapping phase reconstruction feature 3. Finally, the concatenation operation is used to fuse wrapping phase reconstruction features 0, 1, 2, and 3, and convolution and sigmoid activation functions are used to obtain the wrapping phase result. The dilated convolution module includes dilated convolution, batch normalization, and ReLU activation function. The denominator fringe reconstruction branch fuses denominator fringe features at different scales through upsampling, concatenation, and dilated convolution to obtain denominator fringe reconstruction feature 1. Then, it uses a concatenation operation to fuse denominator fringe reconstruction feature 0 and denominator fringe reconstruction feature 1. Finally, it uses a dilated convolution module, skip connections, and a sigmoid activation function to obtain the denominator fringe result. The dilated convolution module includes dilated convolution, batch normalization, and ReLU activation function. The molecular stripe reconstruction branch adopts the same structure as the encapsulated phase reconstruction branch, and fuses molecular stripe features at different scales to obtain molecular stripe results; Meanwhile, based on the mathematical relationship between the denominator and numerator fringes and the wrapped phase, the indirect wrapped phase result is calculated.

8. The fringe analysis method based on phase shift theory-driven deep learning according to claim 1, characterized in that, The composite loss function includes: weighted average wrapping phase loss, denominator fringe loss, numerator fringe loss, and wrapping phase crossover loss; The wrapping phase loss, denominator fringe loss, numerator fringe loss, and wrapping phase crossover loss are all composed of MSE, MAE, and MS-SSIM.

9. The fringe analysis method based on phase shift theory-driven deep learning according to claim 1, characterized in that, Training the multi-task coupled network model includes: The difference between the output of the multi-task coupled network model and the true label is calculated based on the composite loss function. The network parameters are updated by backpropagation using gradient descent, and the weights and offsets are continuously optimized. Training stops when the training reaches the maximum number of iterations, the loss function value decreases to a stable value, or the network weights converge to the optimal value.

Images