Deep learning-based fringe pattern phase recovery method

Abstract

The invention discloses a deep learning-based fringe pattern phase recovery method, and relates to the technical field of fringe pattern phase recovery, and the method comprises the steps: obtaining a to-be-processed fringe image sequence; inputting the fringe image sequence into a pre-trained phase recovery neural network model; and obtaining and outputting a recovered phase diagram generated by the phase recovery neural network model, and evaluating the phase recovery precision by calculating a root-mean-square error between the recovered phase diagram and a real phase diagram. According to the invention, a SlowFast double-branch encoder is adopted, a decoder adopts depth separable convolution and nearest neighbor upsampling to replace traditional convolution and transpose convolution, and group normalization is matched to reduce batch dependence and reduce the model calculation amount and parameter amount, so that the model can be directly deployed in a mobile terminal or embedded equipment, the real-time measurement requirement is met, and the real-time measurement efficiency is improved. Microscopic geometric global features, edge regions, local deformation and other microscopic detail features are focused, multi-scale information complementation is achieved after channel splicing and fusion, and phase recovery errors are reduced.

Description

Technical Field

[0001] This invention relates to the field of stripe pattern phase recovery technology, specifically a stripe pattern phase recovery method based on deep learning. Background Technology

[0002] Fringe pattern phase retrieval is a key technology in optical measurement, industrial inspection, and 3D scanning. Its core objective is to extract the enclosed phase from a fringe pattern with a periodic or directional structure and unfold it into an absolute phase, thereby obtaining spatially varying physical quantities such as surface height, deformation, and terrain elevation. Fringe patterns are diverse, including grayscale fringe patterns in image processing, physical interference fringe patterns, and grating images. Their phase distribution encodes crucial spatial variation data, and phase differences directly correspond to differences in physical quantities.

[0003] Traditional phase retrieval methods primarily rely on physical models. For example, the phase-shifting method calculates the wrapped phase using multiple phase-shifted fringe patterns, but requires additional phase unwrapping to eliminate 2π periodic tortuosity and demands extremely high system calibration accuracy. The Fourier transform method extracts the phase through frequency domain transformation but is susceptible to noise interference. While the Hilbert transform method can achieve single-frame phase retrieval and avoid unwrapping errors, it is sensitive to shadows and discontinuities. These traditional methods generally suffer from problems such as reliance on multiple frames, low processing efficiency, and insufficient adaptability to complex scenes, making it difficult to meet the real-time and robustness requirements of dynamic measurements and embedded device deployments.

[0004] In recent years, deep learning technology, with its data-driven advantages, has provided a new path for stripe pattern phase retrieval, giving rise to methods such as numerator-denominator prediction networks, dynamic stripe convolutional networks (DFC-Net), and deep neural network stripe processors (DNNFP), which significantly outperform traditional methods in single-frame phase retrieval efficiency and noise resistance. However, existing deep learning models still face some challenges: on the one hand, the models have high computational overhead and redundant parameters, making them difficult to deploy on resource-constrained platforms such as mobile terminals or embedded systems; on the other hand, they are insufficient in capturing detailed features such as edge regions and local deformations of stripe patterns, resulting in phase retrieval accuracy that needs improvement, and the models have limited generalization ability, exhibiting poor adaptability to different stripe frequencies, orientations, and low signal-to-noise ratio scenarios. These problems limit the widespread application of deep learning-based stripe pattern phase retrieval methods in practical engineering scenarios.

[0005] Based on this, a deep learning-based stripe pattern phase recovery method is now provided, which can eliminate the drawbacks of existing technical solutions. Summary of the Invention

[0006] The purpose of this invention is to provide a deep learning-based stripe pattern phase recovery method to solve the problems in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] A deep learning-based method for stripe pattern phase retrieval includes the following steps:

[0009] Step S1: Obtain the stripe image sequence to be processed, wherein the stripe image sequence contains multiple stripe images of the same object being measured, collected from different viewpoints and at different times;

[0010] Step S2: Input the stripe image sequence into a pre-trained phase retrieval neural network model. The phase retrieval neural network model is a lightweight convolutional neural network based on an encoder-decoder structure, which is configured to receive a single stripe image and directly output the corresponding reconstructed phase map.

[0011] Step S3: Obtain and output the restored phase map generated by the phase retrieval neural network model, and evaluate the phase retrieval accuracy by calculating the root mean square error between the restored phase map and the real phase map.

[0012] Furthermore, the training steps of the phase retrieval neural network model specifically include:

[0013] A training dataset is constructed, which includes multiple sample pairs covering different noise intensities, stripe frequencies and geometric features of object surfaces. Each sample pair consists of a stripe image and its corresponding real phase map label. The training dataset is divided into a training set, a validation set and a test set according to a preset ratio.

[0014] Construct an initial neural network model, which includes an encoder part for feature extraction and a decoder part for phase map reconstruction;

[0015] A hybrid precision training strategy and an exponential decay learning rate strategy are employed to perform supervised training on the initial neural network model using the training dataset. The model parameters are updated by optimizing the loss function to obtain the phase retrieval neural network model that has been trained and converged.

[0016] Furthermore, the encoder section adopts a SlowFast dual-branch structure, including:

[0017] The first branch is used to downsample and sparsify the input features, process the stripe image sequence with a low view sampling rate, and extract the phase-related macroscopic geometric global features in the stripe image.

[0018] The second branch, running in parallel with the first branch, extracts phase-related micro-detail features from the striped image at a high viewpoint sampling rate.

[0019] The fusion module is used to concatenate and fuse the macroscopic geometric global features output from the first branch with the microscopic detail features output from the second branch to form a fused feature, which is then output to the decoder. The fusion process satisfies the following: ,in, The output features after fusion at the Lth layer are... For feature splicing operations, The output features of the Lth layer of the first branch, This is the output feature of the Lth layer of the second branch.

[0020] Furthermore, the decoder portion has a lightweight structure, including:

[0021] The upsampling unit uses nearest-neighbor upsampling to upsample the fused features in order to restore spatial resolution;

[0022] The feature refinement unit uses depthwise separable convolution to perform convolution operations on the upsampled feature map. Depthwise separable convolution separates spatial convolution from channel convolution, adapting to the frequency and orientation feature extraction requirements of stripe images to obtain comprehensive features.

[0023] The output unit adjusts the number of feature map channels through 1×1 convolution to generate a phase map corresponding to the stripe image.

[0024] Furthermore, the steps for constructing the training dataset specifically include:

[0025] Generate a true absolute phase map based on a preset phase distribution function;

[0026] Based on the principle of fringe projection, the real absolute phase image is converted into a simulated ideal sinusoidal fringe image;

[0027] Add simulated noise or simulated non-ideal optical effects in projection imaging to the ideal sinusoidal fringe image to generate the fringe image for training;

[0028] The stripe image is paired with the corresponding true absolute phase image, and abnormal samples with phase errors greater than a preset threshold are removed to construct the training dataset.

[0029] Furthermore, the loss function is a weighted sum of the mean square error loss function and the edge composite loss function. The mean square error loss function is used to measure pixel-level error, and the edge composite loss function is used to enhance the phase prediction accuracy of the stripe edge region.

[0030] Furthermore, before inputting the stripe image sequence into the phase retrieval neural network model, a preprocessing step is included: grayscale normalization is performed on the stripe images in the stripe image sequence, and bilinear interpolation is used to adjust the size of the normalized stripe images to the input size specified by the model.

[0031] Furthermore, after obtaining the recovered phase map, a post-processing step is included: performing phase unwrapping processing on the recovered phase map to obtain a continuous absolute phase distribution.

[0032] Furthermore, the hybrid precision training strategy uses FP16 half-precision floating-point numbers for forward propagation and backward gradient calculation, and FP32 full-precision floating-point numbers for updating and storing sovereign re-parameters.

[0033] Furthermore, it also includes a model optimization step: regularly collecting new stripe image samples and corresponding real phase map labels, incrementally training the trained phase retrieval neural network model, updating the model parameters, and improving the model's adaptability to phase retrieval of stripe images in new scenes.

[0034] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0035] 1. This invention adopts a combination architecture of SlowFast dual-branch encoder and lightweight decoder. The number of channels in the first branch and the second branch are configured proportionally. The decoder uses depthwise separable convolution and nearest neighbor upsampling to replace traditional convolution and transposed convolution. Combined with group normalization to reduce batch dependency, the computational load and parameter load of the model are greatly reduced, so that the model can be directly deployed on mobile terminals or embedded devices to meet real-time measurement requirements.

[0036] 2. This invention decouples macroscopic and microscopic feature extraction through a SlowFast dual-branch structure, focusing on the macroscopic geometric global features of the overall periodic structure of the stripe pattern and the microscopic details such as edge regions, local deformations and frequency changes. After channel splicing and fusion, multi-scale information complementarity is achieved, and the prediction accuracy of key areas is enhanced by combining a loss function to reduce phase recovery error.

[0037] 3. The training dataset of this invention covers different fringe frequencies, directions, noise intensities, and object surface features. Combined with an incremental training strategy, it can continuously optimize the model's adaptability to new scenes. It is suitable for fringe pattern phase recovery tasks in multiple fields such as industrial inspection and 3D scanning, without the need to redesign or adjust the model structure for different scenes. Attached Figure Description

[0038] Figure 1 This is a schematic diagram illustrating the steps of the fringe pattern phase recovery method of the present invention.

[0039] Figure 2This is a schematic diagram of the phase retrieval neural network model of the present invention.

[0040] Figure 3 This is a schematic flowchart of the fringe pattern phase recovery method of the present invention. Detailed Implementation

[0041] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0042] Example 1

[0043] In this embodiment, as Figures 1-3 As shown, a deep learning-based stripe pattern phase retrieval method specifically includes the following steps:

[0044] Step S1: Obtain the stripe image sequence to be processed. The stripe image sequence contains multiple stripe images of the same object being measured, collected from different perspectives and at different times. A stripe pattern is an image with a periodic or directional structure. Its core feature is that it presents a stripe pattern of alternating light and dark or colors. The main types include: stripe patterns in image processing (grayscale images used to analyze image mixing patterns, which intuitively show the pixel change rules through the interaction between the primary color and the mixed color), physical interference stripe patterns (describing the alternating light and dark patterns formed by light wave interference phenomena, whose shape is directly related to the optical path difference), and general stripe features (referring to images with periodic structures such as interference fringes and grating images). The phase of a stripe pattern refers to the phase information contained in the stripe pattern. Stripe patterns usually refer to alternating light and dark patterns formed by interference, projection, or natural textures. Its phase distribution encodes key data of spatial changes.

[0045] Step S2: Input the stripe image sequence into the pre-trained phase retrieval neural network model. The phase retrieval neural network model is a lightweight convolutional neural network based on an encoder-decoder structure. It is configured to receive a single stripe image and directly output the corresponding reconstructed phase map.

[0046] Step S3: Obtain and output the restored phase map generated by the phase retrieval neural network model, and evaluate the phase retrieval accuracy by calculating the root mean square error between the restored phase map and the true phase map;

[0047] Before inputting the stripe image sequence into the phase retrieval neural network model, a preprocessing step is also included: grayscale normalization is performed on the stripe images in the stripe image sequence to eliminate the interference of illumination changes on the model input. If the model input layer is fixed at 512×512 pixels (an integer multiple of 32 to adapt to the encoder downsampling factor), the size of the normalized stripe image is adjusted to the input size specified by the model using bilinear interpolation. For images with a resolution higher than 512×512, they are scaled down to the target size proportionally to ensure that the periodic structure of the stripes is not destroyed. For images with a resolution lower than 512×512, they are scaled up proportionally to avoid feature distortion caused by stretching and deformation. Then, the processed image is converted into a tensor format that the model can accept to adapt to subsequent network calculations.

[0048] After obtaining the restored phase map, a post-processing step is also included: performing phase unwrapping on the restored phase map to obtain a continuous absolute phase distribution. The phase unwrapping adopts a quality-guided phase unwrapping algorithm, which calculates the pixel quality value through the phase gradient. The higher the quality value, the stronger the phase reliability. Then, unwrapping is performed in order of quality value from high to low to avoid error propagation caused by crossing unreliable areas and eliminate 2π periodic faults.

[0049] Specifically, the above-mentioned deep learning-based stripe pattern phase retrieval method is applied to a stripe pattern phase retrieval system, which includes a preprocessing module, a model processing module, and a postprocessing module. The preprocessing module is used to implement the preprocessing steps, and the postprocessing module is used to implement the postprocessing steps.

[0050] Among them, such as Figure 2 As shown, the training steps for the phase retrieval neural network model specifically include:

[0051] Construct a training dataset, which includes multiple sample pairs covering different noise intensities, stripe frequencies and geometric features of object surfaces. Each sample pair consists of a stripe image and its corresponding real phase map label. The training dataset is divided into a training set, a validation set and a test set according to a preset ratio.

[0052] Construct an initial neural network model, which includes an encoder part for feature extraction and a decoder part for phase map reconstruction;

[0053] A hybrid precision training strategy and an exponential decay learning rate strategy are adopted. The initial neural network model is supervised and trained using the training dataset. The model parameters are updated by optimizing the loss function, and a phase retrieval neural network model with training convergence is obtained.

[0054] Specifically, the phase retrieval neural network model construction process includes:

[0055] The preprocessing module performs grayscale normalization on the input striped image according to the formula. The grayscale values ​​are mapped to the [0,1] range. In this embodiment, it is assumed that the image size is adjusted to 512×512 pixels (an integer multiple of 32) using bilinear interpolation, and the output tensor has a dimension of 512×512×1.

[0056] The encoder (SlowFast dual-branch structure): The first branch input tensor is downsampled by layers.AveragePooling3D (kernel size 3×3×3, stride 2×2×2), and then passed through three convolutional blocks (each containing two 3×3 convolutional layers, group normalization, and ReLU activation function), with the number of channels set to 32, 64, and 128 (base number of channels = 32), ultimately outputting a 128×128×128 macroscopic global feature map. The second branch inputs in parallel with the first branch, with the number of channels set to 1 / 8 of the first branch (4, 8, 16), and is processed by three convolutional blocks (structure same as the first branch, no downsampling), outputting a 512×512×16 microscopic detail feature map. The fusion module concatenates the output features of the two branches according to the channel dimension using layers.Concatenate, obtaining a 128×128×144 fused feature map, satisfying... The 1:8 configuration ensures that the second branch can capture enough detailed features while effectively controlling the total number of model parameters.

[0057] Decoder (lightweight structure): Divide the fused feature map into 8 groups by channel, calculate the mean and variance of each group and normalize them. The upsampling unit uses nearest neighbor upsampling to restore the feature map size to 256×256, while keeping the number of channels unchanged at 144. The feature refining unit performs 3×3 depthwise convolution and 1×1 pointwise convolution, adjusting the number of channels to 64. Then, it restores the size to 512×512 again through nearest neighbor upsampling, and adjusts the number of channels to 1 through 1×1 convolution. The output is a 512×512×1 wrapped phase map.

[0058] The weighted sum of the mean squared error loss function (MSE) and the edge composite loss function (EdgeLoss) is expressed as follows: ,in, This is the weighting coefficient, with a value ranging from 0.3 to 0.7. The loss function;

[0059] Training strategy configuration: The optimizer used is AdamW, the initial learning rate is set to 1e-3, the weight decay factor is 1e-4, FP16 mixed precision training is used, FP16 is used for forward propagation and backpropagation gradient calculation, FP32 is used for weight storage and update, and an exponential decay strategy is adopted for the learning rate. The update formula is as follows: ,in, For the updated learning rate, The initial learning rate, The attenuation coefficient is... For the current training round, Assuming the decay step size, in this embodiment, we assume... It is 0.92. The value is 10, meaning the learning rate decays to 92% of its original value every 10 rounds of training, as shown by the formula. The error in calculating and validating the true absolute phase map of the set. The number of sample pixels, For the first The true absolute phase value of each pixel. For the first The absolute phase value of each pixel is recovered. If it does not decrease for 10 consecutive rounds, training is stopped and the optimal model parameters are saved. The model training parameters can be adjusted according to actual needs.

[0060] Among them, such as Figure 2 and Figure 3 As shown, considering that the SlowFast structure was originally used to process spatiotemporal features in videos, its core idea is to use a dual-path parallel architecture to capture motion and appearance information at different time scales. The stripe pattern itself has multi-scale features, containing both macroscopic periodic structures reflecting the overall phase distribution and microscopic high-frequency information reflecting local deformation and edge details. The Slow branch is suitable for extracting macroscopic features, while the Fast branch is suitable for extracting microscopic details. By splicing or adding the two feature paths, the model can simultaneously utilize macroscopic phase trends and microscopic detail changes, thus preserving the accuracy of edges and local deformations while achieving overall phase smoothing. Therefore, the encoder part of this invention adopts a SlowFast dual-branch structure, a parallel feature extraction architecture composed of a Slow branch and a Fast branch. The Slow branch extracts macroscopic geometric global features at a low viewing angle sampling rate, while the Fast branch captures microscopic detail features at a high viewing angle sampling rate. Multi-scale information complementarity is achieved through feature fusion, including:

[0061] The first branch (Slow branch) is used to downsample and sparsify the input features, process the stripe image sequence with a low view sampling rate, and extract the phase-related macroscopic geometric global features in the stripe image.

[0062] The second branch (Fast branch) runs in parallel with the first branch and extracts phase-related micro-detail features in the striped image at a high viewpoint sampling rate.

[0063] The fusion module is used to concatenate and fuse the macroscopic geometric global features output from the first branch with the microscopic detail features output from the second branch to form a fused feature that is output to the decoder. The fusion process satisfies the following: ,in, The output features after fusion at the Lth layer are... For feature splicing operations, The output features of the Lth layer of the first branch, The output features of the second branch at layer L are combined by the fusion module to achieve multi-scale information complementarity, which not only ensures the overall smoothness of phase recovery, but also improves the prediction accuracy of edge and detail regions.

[0064] Among them, such as Figure 2 and Figure 3 As shown, the decoder section has a lightweight structure, including:

[0065] The upsampling unit uses nearest-neighbor upsampling to upsample the fused features in order to restore spatial resolution;

[0066] The feature refinement unit employs depthwise separable convolution to perform convolution operations on the upsampled feature map. Depthwise separable convolution separates spatial convolution from channel convolution, adapting to the frequency and orientation feature extraction requirements of stripe images to obtain comprehensive features. Depthwise separable convolution decomposes standard convolution into two independent steps: depthwise convolution and pointwise convolution. Depthwise convolution involves independent spatial convolution for each input channel, used to extract local spatial features of the stripes such as edges and frequencies. Pointwise convolution uses 1×1 convolutions to fuse information across channels, integrating features from different channels. Compared to ordinary convolution, depthwise separable convolution significantly reduces the number of parameters and computational cost. For example, for an input channel with... The number of output channels is A convolutional layer with a K×K kernel, and a normal convolutional parameter count of . The number of parameters for depthwise separable convolution is The parameter count was reduced by approximately ,like , Then the number of parameters can be reduced to 1 / 8 to 1 / 9 of that of ordinary convolution, which is suitable for the lightweight design of this invention;

[0067] The output unit adjusts the number of feature map channels through 1×1 convolution to generate a phase map corresponding to the stripe image;

[0068] In this embodiment, the steps for constructing the training dataset specifically include:

[0069] Generate a true absolute phase map based on a preset phase distribution function;

[0070] Based on the principle of fringe projection, the real absolute phase image is converted into a simulated ideal sinusoidal fringe image;

[0071] Add simulated noise or simulated non-ideal optical effects in projection imaging to an ideal sinusoidal fringe image to generate a fringe image for training.

[0072] The stripe image is paired with the corresponding true absolute phase image, and abnormal samples with phase errors greater than a preset threshold are removed to construct the training dataset.

[0073] Specifically, according to the phase distribution function Multiple sets of true absolute phase maps are generated, among which, For pixel coordinates, , These are the fringe frequencies in the x and y directions, respectively, with values ​​ranging from 5 to 30 lp / mm. They can be adjusted according to actual needs. For the initial phase offset, four values ​​are used: 0, π / 2, π, and 3π / 2. The phase map size is set to 512×512 pixels.

[0074] Based on the principle of stripe generation The real absolute phase image is converted into a simulated ideal sinusoidal fringe image, where, The background light intensity ranges from 50 to 200. To modulate the light intensity, the value range is 30~80, generating the corresponding ideal sinusoidal fringe image;

[0075] Gaussian noise, salt-and-pepper noise, and other simulated noises are added to an ideal sinusoidal stripe image, along with non-ideal factors such as uneven lighting, shadows, and system interference, to generate training stripe images with different signal-to-noise ratios. For example, Gaussian noise with an intensity of 5 to 20 and salt-and-pepper noise with a density of 0.01 to 0.05 are added, while simulating non-ideal imaging effects such as uneven lighting (brightness gradient range of 0.1 to 0.3), shadow occlusion (occlusion area ratio of 5% to 15%), and lens distortion (distortion coefficient k1 of -0.3 to 0.3) to generate several sets of training stripe images.

[0076] Pair the stripe image with the corresponding true absolute phase image, remove abnormal samples with a phase error greater than a preset threshold of 5°, and divide the final effective sample pairs into training set, validation set and test set in a ratio of 7:2:1.

[0077] Through the above operations, the training dataset can be widely covered with different imaging environments, different equipment conditions and different object surface characteristics, thereby improving the phase recovery robustness and generalization ability of the model in complex real-world scenarios.

[0078] In this embodiment, the loss function is a weighted sum of the mean square error loss function and the edge composite loss function. The mean square error loss function is used to measure pixel-level error, and the edge composite loss function is used to enhance the phase prediction accuracy of the stripe edge region.

[0079] In this embodiment, the mixed precision training strategy uses FP16 half-precision floating-point numbers for forward propagation and backward gradient calculation to reduce memory usage and improve calculation speed, and uses FP32 full-precision floating-point numbers for updating and storing sovereign re-parameters to ensure training stability.

[0080] In this embodiment, a model optimization step is also included: periodically collecting new stripe image samples and corresponding real phase map labels, incrementally training the trained phase retrieval neural network model, updating the model parameters, and improving the model's adaptability to phase retrieval of new scene stripe images. Specifically, new scene samples such as surface stripe patterns of parts in industrial inspection and tissue interference stripe patterns in biomedical imaging are collected periodically, and the corresponding real absolute phase maps are obtained through measurement devices. An incremental training dataset is constructed, the trained model parameters are loaded, the initial learning rate is adjusted to 1e-4 (1 / 10 of the original initial learning rate), fine-tuning training is performed using the incremental training dataset, the performance is verified every 5 rounds, the optimal model parameters are retained, and the optimized model parameters are deployed to the inference device to replace the original model, thereby achieving continuous improvement in the phase retrieval accuracy of new scene stripe images.

[0081] Specifically, this invention normalizes and resizes the stripe image through a preprocessing module, and inputs it into a phase retrieval neural network model based on a SlowFast dual-branch architecture. The encoder extracts macroscopic geometric global features through the first branch and captures microscopic detail features through the second branch. After splicing and fusion, the images are input into a lightweight decoder. Feature refinement and phase map wrapping are achieved through group normalization, nearest neighbor upsampling, and depthwise separable convolution. Finally, a continuous absolute phase distribution is obtained after phase unwrapping and denoising postprocessing.

[0082] Example 2

[0083] Unlike Embodiment 1, the present invention also provides an electronic device, including a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it implements the above-described deep learning-based stripe pattern phase recovery method and steps. The present invention also provides a computer-readable storage medium storing a computer program, and when the computer program is executed by the processor, it implements the above-described deep learning-based stripe pattern phase recovery method.

[0084] In summary, this method combines the SlowFast dual-branch structure with a lightweight design, solving the problems of traditional methods such as reliance on multiple frames of images and low efficiency. It overcomes the shortcomings of existing deep learning models, such as high computational cost and insufficient detail capture. It can be directly deployed on resource-constrained platforms, adapts to the application needs of complex scenarios, and has good application prospects.

[0085] The above description is merely a specific embodiment 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 technical scope 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 deep learning-based stripe pattern phase retrieval method, characterized in that, Specifically, the following steps are included: Step S1: Obtain the stripe image sequence to be processed, wherein the stripe image sequence contains multiple stripe images of the same object being measured, collected from different viewpoints and at different times; Step S2: Input the stripe image sequence into a pre-trained phase retrieval neural network model. The phase retrieval neural network model is a lightweight convolutional neural network based on an encoder-decoder structure, which is configured to receive a single stripe image and directly output the corresponding reconstructed phase map. Step S3: Obtain and output the restored phase map generated by the phase retrieval neural network model, and evaluate the phase retrieval accuracy by calculating the root mean square error between the restored phase map and the real phase map.

2. The deep learning-based stripe pattern phase retrieval method according to claim 1, characterized in that, The training steps of the phase retrieval neural network model specifically include: A training dataset is constructed, which includes multiple sample pairs covering different noise intensities, stripe frequencies and geometric features of object surfaces. Each sample pair consists of a stripe image and its corresponding real phase map label. The training dataset is divided into a training set, a validation set and a test set according to a preset ratio. Construct an initial neural network model, which includes an encoder part for feature extraction and a decoder part for phase map reconstruction; A hybrid precision training strategy and an exponential decay learning rate strategy are employed to perform supervised training on the initial neural network model using the training dataset. The model parameters are updated by optimizing the loss function to obtain the phase retrieval neural network model that has been trained and converged.

3. The deep learning-based stripe pattern phase retrieval method according to claim 2, characterized in that, The encoder section adopts a SlowFast dual-branch structure, including: The first branch is used to downsample and sparsify the input features, process the stripe image sequence with a low view sampling rate, and extract the phase-related macroscopic geometric global features in the stripe image. The second branch, running in parallel with the first branch, extracts phase-related micro-detail features from the striped image at a high viewpoint sampling rate. The fusion module is used to concatenate and fuse the macroscopic geometric global features output from the first branch with the microscopic detail features output from the second branch to form a fused feature, which is then output to the decoder. The fusion process satisfies the following: ,in, The output features after fusion at the Lth layer are... For feature splicing operations, The output features of the Lth layer of the first branch, This is the output feature of the Lth layer of the second branch.

4. The deep learning-based stripe pattern phase retrieval method according to claim 3, characterized in that, The decoder section has a lightweight structure, including: The upsampling unit uses nearest-neighbor upsampling to upsample the fused features in order to restore spatial resolution; The feature refinement unit uses depthwise separable convolution to perform convolution operations on the upsampled feature map. Depthwise separable convolution separates spatial convolution from channel convolution, adapting to the frequency and orientation feature extraction requirements of stripe images to obtain comprehensive features. The output unit adjusts the number of feature map channels through 1×1 convolution to generate a phase map corresponding to the stripe image.

5. The deep learning-based stripe pattern phase retrieval method according to claim 2, characterized in that, The specific steps for constructing the training dataset include: Generate a true absolute phase map based on a preset phase distribution function; Based on the principle of fringe projection, the real absolute phase image is converted into a simulated ideal sinusoidal fringe image; Add simulated noise or simulated non-ideal optical effects in projection imaging to the ideal sinusoidal fringe image to generate the fringe image for training; The stripe image is paired with the corresponding true absolute phase image, and abnormal samples with phase errors greater than a preset threshold are removed to construct the training dataset.

6. The deep learning-based stripe pattern phase retrieval method according to claim 2, characterized in that, The loss function is a weighted sum of the mean square error loss function and the edge composite loss function. The mean square error loss function is used to measure pixel-level error, and the edge composite loss function is used to enhance the phase prediction accuracy of the stripe edge region.

7. The deep learning-based stripe pattern phase retrieval method according to claim 1, characterized in that, Before inputting the stripe image sequence into the phase retrieval neural network model, a preprocessing step is also included: performing grayscale normalization on the stripe images in the stripe image sequence, and using bilinear interpolation to adjust the size of the normalized stripe images to the input size specified by the model.

8. The deep learning-based stripe pattern phase retrieval method according to claim 1, characterized in that, After obtaining the recovered phase map, a post-processing step is also included: performing phase unwrapping processing on the recovered phase map to obtain a continuous absolute phase distribution.

9. The deep learning-based stripe pattern phase retrieval method according to claim 2, characterized in that, The hybrid precision training strategy uses FP16 half-precision floating-point numbers for forward propagation and backward gradient calculation, and FP32 full-precision floating-point numbers for updating and storing sovereign re-parameters.

10. The deep learning-based stripe pattern phase retrieval method according to claim 1, characterized in that, It also includes model optimization steps: regularly collecting new stripe image samples and corresponding real phase map labels, incrementally training the trained phase retrieval neural network model, updating model parameters, and improving the model's adaptability to phase retrieval of stripe images in new scenes.

Images