A Fringe Phase Retrieval Method Based on Untangling Theory and Deep Learning

Abstract

This invention discloses a fringe phase retrieval method based on untangling theory-driven deep learning, belonging to the field of optical detection. The method includes: generating absolute phase labels based on a random number sequence; obtaining the original speckle image, wrapped phase labels, and wrapped count labels based on the absolute phase labels; constructing training and validation datasets; constructing a dual-task coupled network model based on a convolutional neural network and a composite loss function; training the dual-task coupled network model using the training dataset, validation dataset, and composite loss function; inputting the original fringe image to be processed into the trained dual-task coupled network model, and outputting the absolute phase result after fringe phase retrieval. This invention achieves unified end-to-end processing of speckle removal, fringe analysis, phase unwrapping, and phase retrieval, significantly improving the accuracy, robustness, and interpretability of the phase retrieval results under scenarios with noise, uneven illumination, complex backgrounds, and varying fringe density.

Description

Technical Field

[0001] This invention belongs to the field of optical detection, and in particular relates to a fringe phase recovery method based on untangling theory-driven deep learning. Background Technology

[0002] Fringe phase retrieval is a crucial step in modern optical measurement technology, widely applied in fields such as digital holographic speckle interferometry, structured light 3D imaging, and photoelastic measurement. It obtains absolute phase information by analyzing interference fringe images, thereby enabling quantitative measurement of physical quantities such as object shape, deformation, and defects. Traditional methods mainly include those based on fringe skeleton interpolation and those based on fringe analysis and phase unwrapping. The former is suitable for dynamic measurements but struggles to obtain accurate phase across the entire field, while the latter offers higher accuracy but typically relies on multiple images and requires manual parameter adjustment, limiting its applicability in complex real-world scenarios. In recent years, deep learning-based phase retrieval methods have emerged, such as those using U-Net and Y-Net architectures to directly predict the absolute phase from a single fringe image. However, their training process is data-driven and lacks physical constraints, resulting in insufficient model interpretability. They also exhibit instability in scenarios with high noise, low illumination, non-uniform backgrounds, and significant variations in fringe density, and the reliability and robustness of the retrieval results still need improvement.

[0003] Current deep learning-based fringe phase retrieval methods often treat phase retrieval as a simple image mapping task, failing to effectively incorporate the physical prior of phase unwrapping. This results in a lack of clear physical boundary constraints during network training, making the model output prone to error accumulation, loss of detail, or structural distortion under complex disturbances. Furthermore, existing methods typically process speckle removal, fringe analysis, phase unwrapping, and phase retrieval step-by-step or in separate networks, leading to redundant processes that are prone to introducing errors. A unified, end-to-end, interpretable, and highly robust processing framework has yet to be achieved. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a fringe phase recovery method based on untangling theory-driven deep learning. This method constructs a dual-task coupled network model based on a convolutional neural network to achieve batch fully automatic phase recovery of noisy original interference fringe images, directly outputting noise-free absolute phase results. It aims to achieve fringe phase recovery for fringe images with high noise, low illumination, non-uniform backgrounds, and varying fringe density, thereby solving the problems existing in the aforementioned technologies. Specifically, it includes: An absolute phase label is generated based on a random number sequence. The original speckle image, the package phase label, and the package count label are obtained based on the absolute phase label. A training dataset and a validation dataset are then constructed. Construct a dual-task coupled network model based on a convolutional neural network and a composite loss function; The dual-task coupled network model is trained using the training dataset, the validation dataset, and the composite loss function. The original stripe image to be processed is input into the trained dual-task coupled network model, which outputs the absolute phase result after stripe phase recovery.

[0005] Preferably, the process of obtaining the original speckle image includes: The absolute phase label is generated based on a random number sequence and through two-dimensional interpolation. The original speckle image is generated based on the principle of speckle interferometry and the absolute phase label.

[0006] Preferably, the process of constructing the training dataset and the validation dataset includes: The original speckle image is subjected to differential processing to obtain a noisy original stripe image; The absolute phase tag is phase-shifted to obtain the wrapping phase tag; Based on the absolute phase tag and the package phase tag, the package count tag is calculated according to the phase unwrapping principle; The training dataset and the validation dataset are divided based on the noisy original stripe image, absolute phase label, wrap phase label, and wrap count label.

[0007] Preferably, the process of constructing the dual-task coupled network model includes: Construct a shared encoder, a wrapped phase decoder, an absolute phase decoder, a wrapped phase reconstruction branch, and an absolute phase reconstruction branch, respectively; The shared encoder is used to downsample the input noisy raw stripe image step by step to extract multi-scale feature maps; The wrapping phase decoder and the absolute phase decoder are used to upsample and fuse the multi-scale feature map to obtain wrapping phase features and absolute phase features, respectively. The wrapped phase reconstruction branch and the absolute phase reconstruction branch are used to output the wrapped phase result and the absolute phase result based on the wrapped phase feature and the absolute phase feature, respectively.

[0008] Preferably, the process of constructing the shared encoder includes: The coding structure is constructed using convolutional modules, batch normalization, activation functions, residual modules, and multi-scale fusion modules. The multi-scale feature map is obtained by performing multiple consecutive max pooling operations on the input image.

[0009] Preferably, the process of constructing the wrapper phase decoder includes: The decoding structure is constructed using convolutional modules, batch normalization, activation functions, residual modules, and feature concatenation operations. The encapsulated phase features are obtained by performing multiple consecutive upsampling operations on the multi-scale feature map.

[0010] Preferably, the process of constructing the absolute phase decoder includes: The decoding structure is constructed using convolutional modules, cross-branch attention modules, batch normalization, activation functions, residual modules, feature concatenation operations, and pixel-level multiplication operators. The absolute phase features are obtained by performing multiple consecutive upsampling operations on the multi-scale feature map and combining the cross-branch attention module to achieve feature interaction with the wrapping phase decoder.

[0011] Preferably, the wrap-around phase reconstruction branch and the absolute phase reconstruction branch are further used for: The package count result is calculated based on the phase untangling relationship between the package phase result and the absolute phase result.

[0012] Preferably, the composite loss function is obtained by combining the package phase loss, absolute phase loss, and package count loss in a weighted manner; The wrapping phase loss and absolute phase loss both include mean square error, mean absolute error and multi-scale structural similarity loss; The package counting loss includes mean square error, multi-class dice loss, and focus loss.

[0013] Preferably, the process of training the dual-task coupled network model includes: The error between the model output and the true label is calculated using the composite loss function, and the network parameters are updated using gradient descent. Training ends when the preset training stopping condition is met, and the model's fitting and generalization performance is evaluated using a validation dataset during the training process.

[0014] Compared with the prior art, the present invention has the following advantages and technical effects: This invention integrates the fundamental principle of phase unwrapping into the training process of a dual-task coupled network model, enabling hyperparameter optimization of the network model to be performed within a strict physical framework. This significantly improves the interpretability of deep learning methods and the reliability, accuracy, and robustness of the results. Simultaneously, it unifies speckle removal, fringe analysis, phase unrolling, and phase recovery within a single network framework, achieving end-to-end processing of images with high noise, low illumination, non-uniform backgrounds, and varying fringe density. Batch fully automated phase recovery can be completed without manual parameter adjustment, providing an effective technical path for the practical application and engineering implementation of modern optical measurement technology. Attached Figure Description

[0015] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the first network model structure of the method proposed in the embodiment of the present invention; Figure 3 This is a schematic diagram of the second network model structure of the method proposed in the embodiment of the present invention; Figure 4 This is a schematic diagram of the loss function and model training in an embodiment of the present invention; Figure 5 This is a result image on a simulated stripe image according to an embodiment of the present invention; Figure 6 This is a result image of a measured stripe image according to an embodiment of the present invention. Detailed Implementation

[0016] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0017] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0018] like Figure 1 As shown, this embodiment provides a fringe phase retrieval method based on untangling theory-driven deep learning, including: Absolute phase labels are generated based on random number sequences. The original speckle image, package phase labels, and package count labels are obtained based on the absolute phase labels. Training and validation datasets are then constructed. Construct a dual-task coupled network model based on a convolutional neural network and a composite loss function; The dual-task coupled network model was trained using a training dataset, a validation dataset, and a composite loss function. The original stripe image to be processed is input into the trained dual-task coupled network model, which outputs the absolute phase result after stripe phase recovery.

[0019] This embodiment constructs a training dataset containing absolute phase labels, wrapper phase labels, and wrapper count labels, establishes a dual-task coupled network model based on a convolutional neural network and a composite loss function, and trains the model using the training dataset and validation dataset. Finally, the image to be processed is input into the trained model to obtain the absolute phase result. This achieves end-to-end unified processing of speckle removal, stripe analysis, phase unwrapping, and phase restoration under a single network framework, and completes batch fully automatic phase restoration without manual parameter adjustment.

[0020] The method proposed in this embodiment integrates the basic principle of phase unwrapping into the training process of a dual-task coupled network model, enabling the hyperparameter optimization of the network model to be carried out within a strict physical framework. This improves the interpretability of the network model and the credibility, accuracy, and robustness of the results. At the same time, it unifies speckle removal, fringe analysis, phase unrolling, and phase retrieval under a separate network framework, enhancing the interpretability of phase retrieval based on deep learning. It has wide applications in various phase retrieval tasks and is of great significance to the interdisciplinary fields of optical measurement, artificial intelligence, and image processing.

[0021] Furthermore, the process of obtaining the original speckle image includes: Absolute phase labels are generated based on random number sequences and through two-dimensional interpolation. The original speckle image is generated based on the principle of speckle interferometry and the absolute phase label.

[0022] This embodiment generates absolute phase labels through two-dimensional interpolation based on random number sequences and generates original speckle images according to the principle of speckle interferometry. This ensures that the generation process of training data strictly follows physical laws, providing a highly interpretable physical basis for model training and improving the credibility and scientific rigor of deep learning models in phase retrieval tasks.

[0023] Furthermore, the process of constructing the training and validation datasets includes: The original speckle image is subjected to differential processing to obtain a noisy original stripe image; The absolute phase tag is phase-shifted to obtain the wrapped phase tag; Based on absolute phase tags and package phase tags, package count tags are calculated according to the phase unwrapping principle; The training and validation datasets are divided based on the noisy original stripe image, absolute phase label, wrap phase label, and wrap count label.

[0024] This embodiment obtains a noisy original stripe image by performing differential processing on the original speckle image, obtains a wrapped phase label by performing phase shift processing on the absolute phase label, and calculates the wrapped count label based on the phase unwrapping principle. This constructs supervision information that is completely consistent with the real physical process, enabling the model to learn feature representations that conform to the principles of optical measurement, which significantly improves the accuracy of training data and the reliability of model prediction results.

[0025] Furthermore, the process of constructing a dual-task coupled network model includes: Construct a shared encoder, a wrapped phase decoder, an absolute phase decoder, a wrapped phase reconstruction branch, and an absolute phase reconstruction branch, respectively; The shared encoder is used to downsample the input noisy raw stripe image step by step to extract multi-scale feature maps. The wrap-around phase decoder and the absolute phase decoder are used to upsample and fuse features on multi-scale feature maps, respectively, to obtain wrap-around phase features and absolute phase features. The wrap phase reconstruction branch and the absolute phase reconstruction branch are used to output the wrap phase result and the absolute phase result based on the wrap phase feature and the absolute phase feature, respectively.

[0026] This embodiment constructs a shared encoder, a wrapped phase decoder, an absolute phase decoder, a wrapped phase reconstruction branch, and an absolute phase reconstruction branch, respectively. The shared encoder extracts multi-scale feature maps, the two decoders perform upsampling and feature fusion respectively, and the two reconstruction branches output phase results respectively. This achieves joint learning and mutual constraints of wrapped phase and absolute phase, avoids the feature information loss problem caused by a single decoder structure, and enhances the model's expressive power and recovery accuracy.

[0027] Furthermore, the process of building a shared encoder includes: The coding structure is constructed using convolutional modules, batch normalization, activation functions, residual modules, and multi-scale fusion modules. By performing multiple consecutive max pooling operations on the input image, a multi-scale feature map is obtained.

[0028] Furthermore, this embodiment processes the input noisy original stripe image based on a shared encoder, and obtains the final multi-scale feature map through continuous max pooling operations.

[0029] This embodiment constructs a shared encoder structure through a convolution module, batch normalization, activation function, residual module, and multi-scale fusion module, and performs multiple consecutive max pooling operations on the input image, effectively extracting multi-level semantic features and detailed information of the noisy striped image, enhancing the model's adaptability and robustness to interference factors such as complex backgrounds, speckle noise, and uneven illumination.

[0030] Furthermore, the process of constructing the wrapper phase decoder includes: The decoding structure is constructed using convolutional modules, batch normalization, activation functions, residual modules, and feature concatenation operations. By performing multiple consecutive upsampling operations on the multi-scale feature map, the wrap-around phase features are obtained.

[0031] Furthermore, this embodiment constructs a wrapper phase decoder based on a convolution module, a multi-scale fusion module, batch normalization, activation functions, a residual module, and feature concatenation operations; The wrap-around phase encoder fuses multi-scale feature maps and obtains the final wrap-around phase features through continuous upsampling operations.

[0032] This embodiment constructs a wrapped phase decoder structure through a convolution module, batch normalization, activation function, residual module, and feature concatenation operation. It also performs multiple consecutive upsampling operations on multi-scale feature maps to achieve accurate reconstruction of wrapped phase features. This provides high-quality basic information for subsequent unwrapping of absolute phase and improves the accuracy of phase recovery.

[0033] Furthermore, the process of constructing an absolute phase decoder includes: The decoding structure is constructed using convolutional modules, cross-branch attention modules, batch normalization, activation functions, residual modules, feature concatenation operations, and pixel-level multiplication operators. By performing multiple consecutive upsampling operations on the multi-scale feature map and combining it with the cross-branch attention module to achieve feature interaction with the wrapper phase decoder, the absolute phase features are obtained.

[0034] Furthermore, this embodiment fuses multi-scale feature maps based on an absolute phase encoder, and obtains the final absolute phase features through continuous upsampling operations.

[0035] This embodiment constructs an absolute phase decoder structure through a convolution module, a cross-branch attention module, batch normalization, activation functions, a residual module, feature concatenation operations, and pixel-level multiplication operators. While performing multiple upsampling operations, the cross-branch attention module is used to realize feature interaction with the wrapper phase decoder, achieving collaborative cooperation and information sharing between the two decoders, thereby improving the detail preservation capability and recovery accuracy of the absolute phase features.

[0036] Furthermore, the wrap phase reconstruction branch and the absolute phase reconstruction branch are also used for: The package count is calculated based on the phase untangling relationship between the package phase result and the absolute phase result.

[0037] Furthermore, in this embodiment, the fused wrapping phase features are processed based on the wrapping phase reconstruction branch to output the wrapping phase result; The absolute phase reconstruction branch processes the fused absolute phase features and outputs the absolute phase result. At the same time, based on the mathematical relationship between the wrapped phase result and the absolute phase result (i.e., the basic principle of phase unwrapping), the wrapped count result is calculated.

[0038] This embodiment calculates the package count result based on the phase unwrapping relationship between the package phase result and the absolute phase result by using a package phase reconstruction branch and an absolute phase reconstruction branch. This achieves synchronous output of package counting and phase recovery tasks, and completes the end-to-end mapping from the original stripe image to the final phase result without additional post-processing steps, simplifying the processing flow and improving computational efficiency.

[0039] Furthermore, the composite loss function is obtained by combining the package phase loss, absolute phase loss, and package count loss in a weighted manner; Among them, the wrapping phase loss and the absolute phase loss both include mean square error, mean absolute error and multi-scale structural similarity loss, which are used to optimize the wrapping phase result and the absolute phase result; Package counting losses include mean squared error, multi-class dice loss, and focus loss, which are used to optimize package counting results.

[0040] This embodiment constructs a composite loss function by combining the wrapping phase loss, absolute phase loss, and wrapping count loss in a weighted manner. The wrapping phase loss and absolute phase loss include mean square error, mean absolute error, and multi-scale structural similarity loss, while the wrapping count loss includes mean square error, multi-class dice loss, and focus loss. This achieves collaborative optimization of multiple task objectives, fully utilizes the basic principle of phase unwrapping to guide network training, and improves the convergence stability and overall performance of the model.

[0041] Furthermore, the process of training the dual-task coupled network model includes: The error between the model output and the true label is calculated using a composite loss function, and the network parameters are updated using gradient descent. Training ends when the preset training stopping condition is met, and the model's fitting and generalization performance is evaluated using a validation dataset during the training process.

[0042] This embodiment calculates the error between the model output and the true label using a composite loss function and updates the network parameters using gradient descent. During training, a validation dataset is used to evaluate the model's fitting and generalization performance. Training ends when a preset training stopping condition is met, ensuring that the model has good generalization ability and anti-overfitting performance, obtaining optimal network weight parameters, and guaranteeing stability and reliability in practical applications.

[0043] Furthermore, this embodiment calculates the difference between the output of the multi-task coupled network model and the true label based on the composite loss function, uses gradient descent to backpropagate and update the network parameters, continuously optimizes the weights and offsets, and stops training when the training reaches the maximum number of iterations, the loss function value decreases to a stable level, or the network weights converge to the optimal level.

[0044] After a training iteration, the model is tested on the dataset and the fit and generalization of the network model are observed based on the measurement results and performance metrics.

[0045] After the model training is completed, the final trained network model is used to test the noisy original stripe image to obtain the absolute phase result and complete the stripe phase recovery task.

[0046] This embodiment integrates the basic principle of phase unwrapping into the training process of a dual-task coupled network, enabling network optimization to be performed within a strict physical framework. This achieves unified end-to-end processing of speckle removal, fringe analysis, phase unwrapping, and phase recovery, significantly improving the accuracy, robustness, and interpretability of phase recovery results in scenarios with noise, uneven illumination, complex backgrounds, and variations in fringe density.

[0047] More specifically, the fringe phase recovery method based on untangling theory-driven deep learning proposed in this embodiment realizes batch fully automatic phase recovery of noisy original fringe images by constructing a dual-task coupled network model based on convolutional neural network, and directly obtains noise-free absolute phase results.

[0048] The proposed method in this embodiment utilizes a shared encoder to progressively downsample noisy stripe images, obtaining multi-scale feature maps. Then, it performs upsampling and information fusion through a wrapper phase encoder and an absolute phase encoder. A cross-branch attention module enables feature interaction between the two decoders, and its collaborative cooperation with the wrapper phase decoder avoids the problem of missing feature information from a single decoder. Finally, wrapper phase and absolute phase results are obtained through wrapper phase and absolute phase branches, while the wrapper count result is obtained using the basic principle of phase unwrapping. During model training, the basic principle of phase unwrapping is introduced into the optimization process of the network model's hyperparameters using a composite loss function, allowing the network model to complete training within a strict physical framework. This improves the interpretability of the deep learning stripe phase retrieval method and the credibility, accuracy, and robustness of the results.

[0049] Compared to conventional fringe phase retrieval methods, the method proposed in this embodiment can automatically batch process multiple original fringe images without manual parameter adjustment, resulting in 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).

[0050] As a preferred implementation method, the specific steps of the fringe phase retrieval method based on untangling theory and driven by deep learning to perform phase retrieval on noisy original fringe images are as follows: Step 1: Construct the training and validation datasets. 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 label through two-dimensional interpolation. Based on the absolute phase label and the principle of speckle interferometry, obtain the original speckle image.

[0051] Step 1-2: Based on the original speckle image, use the difference method to obtain the original stripe image containing noise; Steps 1-3: Based on the absolute phase tag, the phase shift method is used to obtain the wrapping phase tag; Steps 1-4: Based on the absolute phase tag and the wrapped phase tag, obtain the wrapped count tag according to the basic principle of phase unwrapping; Steps 1-5: Construct a dataset based on the noisy original stripe image, absolute phase label, wrap phase label, and wrap count label. There are a total of 7000 sets of data, of which 500 sets are the validation set and 6500 sets are the training dataset. The size of each set is 256×256 pixels.

[0052] Step 2: Construct the network model. The specific steps are as follows: This embodiment constructs a dual-task coupled network model based on a convolutional neural network to obtain the absolute phase from a noisy raw stripe image. The overall architecture consists of a shared encoder, a wrapped phase decoder 1, an absolute phase decoder 2, and wrapped phase and absolute phase branches, as follows: Figure 2 The diagram shown illustrates the network structure.

[0053] Step 2-1: Construct a shared encoder. Its basic module 1 is as follows: Figure 3 As shown, it mainly consists of convolutional operations, a multi-scale fusion module, batch normalization operations, ReLU activation function, residual module, and feature concatenation operations. Based on the basic modules, the shared encoder processes the input noisy raw stripe image and obtains the final multi-scale feature map through five consecutive max pooling operations.

[0054] Step 2-2: Construct the wrapper phase decoder 1. Its basic module 2 is as follows: Figure 3 As shown, it mainly consists of convolution operations, batch normalization operations, ReLU activation function, residual module, and feature concatenation operations. Based on the basic module, the wrapper phase decoder processes multi-scale feature maps and obtains the final wrapper phase features through five consecutive upsampling operations.

[0055] Steps 2-3: Construct the absolute phase decoder 2. Its basic modules 3 and 4 are as follows: Figure 3 As shown, it mainly consists of convolution operations, batch normalization operations, ReLU activation function, Sigmoid activation function, residual module, feature concatenation operation, and pixel-level feature multiplication operator. Based on the basic module, the absolute phase decoder processes multi-scale feature maps and obtains the final absolute phase features through five consecutive upsampling operations.

[0056] Steps 2-4: Construct the wrapping phase branch and the absolute phase branch. Both the wrapping phase branch and the absolute phase branch use a 1×1 convolution operation and a sigmoid activation function to output the wrapping phase result and the absolute phase result, respectively. Simultaneously, based on the mathematical relationship between the wrapping phase result and the absolute phase result (i.e., the basic principle of phase unwrapping), the wrapping count result is calculated. Finally, an end-to-end output is achieved from the noisy original stripe image to the wrapping phase result, the absolute phase result, and the wrapping count result, without any additional post-processing steps.

[0057] Step 3: Train the model. For example... Figure 4 The image shows the model training process, with the specific steps as follows: Step 3-1: Initialize network parameters and load the training and validation datasets into the network model.

[0058] Step 3-2: Design a composite loss function consisting of wrapping phase loss, absolute phase loss, and wrapping count loss, and use it for hyperparameter optimization (i.e., model training). The wrapping phase loss and absolute phase loss are both composed of mean squared error, mean absolute error, and multi-scale structural similarity loss, used to optimize the accuracy of the wrapping phase and absolute phase results. The wrapping count loss is composed of mean squared error, multi-class dice loss, and focus loss, used to optimize the accuracy of the wrapping count results. The final composite loss function is obtained by weighting the wrapping phase loss, absolute phase loss, and wrapping count loss. The composite loss function is used to calculate the difference between the output result and the true label, and gradient descent is used for backpropagation to update the network parameters, continuously optimizing the weights and offsets, and fully utilizing the basic principle of phase unwrapping to improve the performance of the network model.

[0059] 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 to train the model will lead to varying degrees of overfitting. Since the network parameter size is relatively small, data augmentation is used to prevent overfitting. Finally, the optimal weights and biases are determined and saved as the optimal hyperparameters of the network model.

[0060] Step 4: Model Validation 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.

[0061] Step 4-2: After training, use the finally trained network model to test the noisy original stripe image to obtain noise-free absolute phase results and complete the stripe phase recovery task.

[0062] like Figure 5 As shown, the results obtained from the original simulated stripe image, absolute phase label, U-Net network model, and the proposed method are presented.

[0063] like Figure 6 As shown, the results obtained from the original measured stripe image, the U-Net network model, and the proposed method are presented.

[0064] The results above show that the fringe phase recovery method proposed in this embodiment is well applicable to original fringe images with low illumination, uneven background, and varying fringe density. It can directly obtain accurate and reliable absolute phase results and has batch processing capabilities.

[0065] The stripe analysis method provided in this embodiment has the following advantages: (1) It can perform batch processing of multiple original stripe images, and the results obtained have high accuracy, good generalization and strong robustness. (2) No complex parameter adjustment or preprocessing or postprocessing is required. Noise-free absolute phase results can be obtained directly by processing the noisy original stripe image. (3) Feature interaction between the two decoders is achieved through the cross-branch attention module, which avoids the problem of missing feature information in a single decoder; (4) The network model is interpretable, the results are credible, and the proposed method has the potential for promotion and application value.

[0066] 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, digital holographic microscopy, photoelasticity, etc.), and is of great significance to research fields such as optical detection, image processing, artificial intelligence technology, defect detection, morphology measurement, biomechanics, and experimental mechanics.

[0067] In the above implementation, the network model is implemented using the PyTorch framework based on Python 3.9. The server hardware configuration used is as follows: CPU is Intel i5-13490F, RAM is 32GB, and GPU is NVIDIA GeForce RTX 5070.

[0068] In the above implementation, the experimental data used in this embodiment consists of 7000 sets of computer-simulated data, including: noisy original stripe images, absolute phase labels, wrapper phase labels, and wrapper count labels, all with an image size of 256×256 pixels. After model training and testing, the method of this embodiment can obtain accurate absolute phase results.

[0069] This embodiment obtains absolute phase labels, wrapped phase labels, wrapped count labels, and a noisy original fringe image, and constructs training and validation datasets. Based on a convolutional neural network, a dual-task coupled network model is constructed with a shared encoder, wrapped phase decoder, and absolute phase decoder as the overall framework. A composite loss function is constructed based on wrapped phase loss, absolute phase loss, and wrapped count loss. The dual-task coupled network model is trained using the composite loss function and the training and validation datasets. The trained dual-task coupled network model is then used to process the original fringe image to obtain the absolute phase result of fringe phase recovery. This embodiment introduces the basic principle of phase unwrapping into the model training, ensuring that the hyperparameter optimization of the network model is performed within a strict physical framework, thus guaranteeing the interpretability of the method and the reliability of the results. Experimental results demonstrate that this embodiment utilizes prior information from phase unwrapping to unify speckle removal, fringe analysis, phase unwrapping, and phase recovery within a single network framework, improving the reliability, accuracy, and robustness of deep learning-based phase recovery. It can be widely applied to various phase recovery tasks and has significant implications for the interdisciplinary fields of optical measurement, artificial intelligence, and image processing.

[0070] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A fringe phase retrieval method based on untangling theory-driven deep learning, characterized in that, include: An absolute phase label is generated based on a random number sequence. The original speckle image, the package phase label, and the package count label are obtained based on the absolute phase label. A training dataset and a validation dataset are then constructed. Construct a dual-task coupled network model based on a convolutional neural network and a composite loss function; The dual-task coupled network model is trained using the training dataset, the validation dataset, and the composite loss function. The original stripe image to be processed is input into the trained dual-task coupled network model, which outputs the absolute phase result after stripe phase recovery.

2. The method according to claim 1, characterized in that, The process of obtaining the original speckle image includes: The absolute phase label is generated based on a random number sequence and through two-dimensional interpolation. The original speckle image is generated based on the principle of speckle interferometry and the absolute phase label.

3. The method according to claim 1, characterized in that, The process of building training and validation datasets includes: The original speckle image is subjected to differential processing to obtain a noisy original stripe image; The absolute phase tag is phase-shifted to obtain the wrapping phase tag; Based on the absolute phase tag and the package phase tag, the package count tag is calculated according to the phase unwrapping principle; The training dataset and the validation dataset are divided based on the noisy original stripe image, absolute phase label, wrap phase label, and wrap count label.

4. The method according to claim 1, characterized in that, The process of constructing the dual-task coupled network model includes: Construct a shared encoder, a wrapped phase decoder, an absolute phase decoder, a wrapped phase reconstruction branch, and an absolute phase reconstruction branch, respectively; The shared encoder is used to downsample the input noisy raw stripe image step by step to extract multi-scale feature maps; The wrapping phase decoder and the absolute phase decoder are used to upsample and fuse the multi-scale feature map to obtain wrapping phase features and absolute phase features, respectively. The wrapped phase reconstruction branch and the absolute phase reconstruction branch are used to output the wrapped phase result and the absolute phase result based on the wrapped phase feature and the absolute phase feature, respectively.

5. The method according to claim 4, characterized in that, The process of constructing the shared encoder includes: The coding structure is constructed using convolutional modules, batch normalization, activation functions, residual modules, and multi-scale fusion modules. The multi-scale feature map is obtained by performing multiple consecutive max pooling operations on the input image.

6. The method according to claim 4, characterized in that, The process of constructing the wrapped phase decoder includes: The decoding structure is constructed using convolutional modules, batch normalization, activation functions, residual modules, and feature concatenation operations. The encapsulated phase features are obtained by performing multiple consecutive upsampling operations on the multi-scale feature map.

7. The method according to claim 4, characterized in that, The process of constructing the absolute phase decoder includes: The decoding structure is constructed using convolutional modules, cross-branch attention modules, batch normalization, activation functions, residual modules, feature concatenation operations, and pixel-level multiplication operators. The absolute phase features are obtained by performing multiple consecutive upsampling operations on the multi-scale feature map and combining the cross-branch attention module to achieve feature interaction with the wrapping phase decoder.

8. The method according to claim 4, characterized in that, The wrap-around phase reconstruction branch and the absolute phase reconstruction branch are also used for: The package count result is calculated based on the phase untangling relationship between the package phase result and the absolute phase result.

9. The method according to claim 1, characterized in that, The composite loss function is obtained by combining the package phase loss, absolute phase loss, and package count loss in a weighted manner. The wrapping phase loss and absolute phase loss both include mean square error, mean absolute error and multi-scale structural similarity loss; The package counting loss includes mean square error, multi-class dice loss, and focus loss.

10. The method according to claim 1, characterized in that, The process of training the dual-task coupled network model includes: The error between the model output and the true label is calculated using the composite loss function, and the network parameters are updated using gradient descent. Training ends when the preset training stopping condition is met, and the model's fitting and generalization performance is evaluated using a validation dataset during the training process.

Images