Phase recovery method of single fringe image based on GD-UNet network

Abstract

The invention discloses a phase recovery method of a single fringe image based on a GD-UNet network, which belongs to the field of three-dimensional reconstruction, and comprises the following steps: building a monocular structured light three-dimensional reconstruction experiment platform, collecting a fringe image and generating a data set; then a GD-UNet model is constructed, the input of the GD-UNet model is a three-channel fringe image with three-frequency fusion, and simultaneous prediction of wrapped phase molecules, denominator terms and fringe levels is achieved through an information integration and distribution module. And after training is completed, inputting a to-be-detected image into the network, outputting a prediction result, and correcting the fringe order by using an arc tangent function and a connected domain segmentation strategy to finally obtain a high-precision absolute phase diagram. According to the method, multi-task prediction is completed in a single network, the reasoning efficiency is improved, error accumulation is effectively inhibited through joint modeling, and the adaptability to complex scenes is high.

Description

Technical Field

[0001] The present invention belongs to the field of three-dimensional reconstruction, and in particular relates to a phase recovery method for a single fringe image based on a GD-UNet network. Background Art

[0002] In recent years, three-dimensional reconstruction technology has been widely used in many fields, including medical imaging, robot navigation, virtual reality, 3D animation modeling, and online product inspection. Among the many three-dimensional reconstruction technologies, Fringe Projection Profilometry (FPP) projects a coded pattern onto the surface of an object to accurately locate the same-name point pairs in triangulation, and then achieves three-dimensional reconstruction. It has the advantages of strong anti-interference, non-contact, and high precision. At present, FPP is widely used in industrial inspection, reverse engineering, cultural relics scanning, and medical diagnosis. Mainstream FPP technology can generally be divided into two types: Fourier transform profilometry (FTP) and phase shift profilometry (PSP). Because PSP technology has stronger robustness to phase noise caused by ambient light and surface reflectivity, and can achieve pixel-by-pixel phase measurement results with higher resolution and accuracy, it has been widely used in practical applications.

[0003] However, traditional phase-shift profilometry typically requires projecting multiple grating patterns, resulting in long reconstruction times and difficulty meeting real-time requirements. Furthermore, while Fourier transform profilometry can extract phase information from a single image frame, reconstruction accuracy is difficult to guarantee on complex surfaces or in high-noise environments. In recent years, researchers have integrated deep learning with phase extraction and phase unwrapping processes, significantly reducing the number of required projection patterns while maintaining high reconstruction accuracy, thereby improving reconstruction efficiency. Currently, deep learning-based absolute phase recovery methods fall into two main categories: one predicts the numerator and denominator of the three frequency-wrapped phases separately and combines them with multi-frequency heterodyning or number theory methods to calculate the absolute phase; the other employs two networks or a dual-decoding network structure to predict the numerator and denominator of the high-frequency wrapped phase and the fringe order to obtain the absolute phase. The former approach, due to the accumulation of wrapped phase errors during the multi-frequency heterodyning calculation process, results in large absolute phase errors and poor stability. The latter, due to the introduction of multi-network collaboration or a dual-decoding network structure, suffers from high computational complexity and low inference efficiency. Summary of the Invention

[0004] To solve the above technical problems, the present invention provides a phase recovery method for a single fringe image based on a GD-UNet network, comprising:

[0005] Build a monocular structured light 3D reconstruction experimental platform and obtain fringe projection images based on the experimental platform;

[0006] Constructing a GD-UNet model based on a deep convolutional neural network, and training the GD-UNet model using a training set;

[0007] The trained GD-UNet network model is used to predict the fringe projection image, and the fringe order is corrected according to the prediction result to achieve high-precision phase recovery of the object to be measured.

[0008] Preferably, the monocular structured light 3D reconstruction experimental platform is composed of a DLP projector, a camera, and a computer equipped with automatic acquisition software;

[0009] The computer controls the DLP projector to project an image and the camera to capture an image, thereby obtaining the fringe projection image.

[0010] Preferably, the process of obtaining a fringe projection image based on the experimental platform includes:

[0011] Generate a phase-shifted fringe pattern and burn it into the projector;

[0012] The fringe image is collected using the built experimental platform, and the target area in the fringe image is cropped using the target detection method to obtain a cropped image.

[0013] The cropped image is processed using an N-step phase shift method and a triple-frequency heterodyne method to generate a fringe projection image. The data set includes an input image, fringe order, and a numerator and denominator of a wrapped phase. The input image is used as input to a GD-UNet model, and the fringe order, numerator, and denominator of the wrapped phase are used as supervisory labels for model training.

[0014] Preferably, the process of cropping around the target area in the fringe image using the target detection method includes: using the target detection model to identify the object area in the fringe image and generate a corresponding rectangular frame, and cropping all fringe images corresponding to the object according to the pixel position of the rectangular frame to generate an image of a fixed size.

[0015] Preferably, the process of processing the cropped image using the N-step phase shift method and the triple-frequency heterodyne method includes:

[0016] The first frame is extracted from each set of three-frequency phase-shift images and mapped to RGB channels from high to low frequency. The three-channel image is fused and used as the input of GD-UNet. The numerator and denominator of the wrapped phase are calculated using the N-step phase-shift algorithm, and the corresponding fringe order is calculated using the three-frequency heterodyne algorithm to generate the fringe projection image.

[0017] Preferably, the GD-UNet model based on deep convolutional neural network includes a downsampling module, an upsampling module, a residual module and an information integration and distribution module.

[0018] Preferably, the residual module adopts a dual-branch structure design, wherein the first branch is composed of a single-layer 3×3 convolution, which is used to extract local detail features;

[0019] The second branch is composed of multiple convolution kernels of different scales and depths to achieve the extraction and fusion of multi-scale features.

[0020] After the two branches complete feature extraction respectively, the outputs are fused by element-by-element addition. The fused features are used as the input of subsequent modules to enhance the model's perception of features at different scales.

[0021] Preferably, the information integration and distribution module fuses and collects multi-scale feature information through a horizontal connection with the encoder and an upsampling operation, and further refines and enhances the fused features through a 3×3 convolutional layer. After feature enhancement, the feature map is divided by channel and sent to two subtask branches respectively. One branch outputs the numerator and denominator of the wrapped phase through a convolutional layer with a channel number of 2, and the other branch is used for classification prediction of the fringe order. The prediction results of the numerator and denominator of the wrapped phase and the fringe order are used for subsequent recovery of the phase of the object to be measured.

[0022] Preferably, the process of achieving high-precision phase recovery of the object to be measured includes:

[0023] According to the numerator and denominator of the wrapped phase output by the trained GD-UNet network model, the wrapped phase map is calculated by the inverse tangent function;

[0024] The wrapped phase image is binarized to extract the connected regions. The distribution of fringe orders in each connected region is counted. The fringe order with the highest frequency is used as the final prediction result of the region, thereby correcting the initially predicted fringe order.

[0025] According to the corrected fringe order and wrapped phase, the absolute phase is calculated by the formula, and then according to the original cropping strategy, the seamless splicing of the absolute phase is achieved, and finally the complete absolute phase map is restored.

[0026] Preferably, for the prediction task of the wrapped phase numerator and denominator, a weighted combination of the SmoothL1 loss function and the structural similarity index (SSIM) loss is used as the joint loss function;

[0027] For stripe-level prediction, multi-class cross entropy loss is used as the loss function of the network.

[0028] Compared with the prior art, the present invention has the following advantages and technical effects:

[0029] 1. By designing an information integration and distribution module, the prediction of the numerator, denominator, and fringe order of the wrapped phase is simultaneously completed in a single network, avoiding the computational redundancy brought by the multi-network structure in traditional methods and significantly improving the inference efficiency. 2. The deep neural network is used to jointly model the numerator, denominator, and fringe order of the wrapped phase, effectively suppressing the cumulative effect of errors in the traditional multi-frequency heterodyne algorithm, and maintaining a high absolute phase recovery accuracy and stability in complex texture scenes. 3. By introducing three-frequency fusion images as input, the network's perception of information at different frequencies is enhanced, making the model more generalizable and adaptable to surfaces of objects with a variety of materials and complex morphologies. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] The accompanying drawings, which constitute part of this application, are intended to provide a further understanding of this application. The exemplary embodiments and descriptions of this application are intended to explain this application and do not constitute an improper limitation on this application. In the accompanying drawings:

[0031] Figure 1 This is a flowchart of the GD-UNet fringe projection profilometry method according to an embodiment of the present invention;

[0032] Figure 2 Schematic diagram of the network structure of GD-UNet according to an embodiment of the present invention;

[0033] Figure 3 These are absolute phase diagrams of different test scenarios according to an embodiment of the present invention. DETAILED DESCRIPTION

[0034] It should be noted that, in the absence of conflict, the embodiments and features of the embodiments in this application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0035] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0036] Example 1

[0037] This embodiment provides a phase recovery method for a single fringe image based on a GD-UNet network, including:

[0038] In this embodiment, a monocular structured light 3D reconstruction experimental platform is first built. A projector is used to project three sets of twelve-step phase-shifted fringe images of different frequencies onto the object under test, and the deformed fringe images are synchronously collected by a camera. The phase shift algorithm is combined with a multi-frequency heterodyne algorithm to calculate the numerator, denominator and fringe order of the wrapped phase, and a data set for supervised training of a deep neural network is constructed. On this basis, a deep learning model based on a convolutional neural network is constructed, named GD-UNet (U-Net with Information Gather-and-Distribute Mechanism). Based on the typical U-Net structure, the model introduces a residual module and an information integration and distribution module to enhance the network's ability to extract multi-scale features. The input of the network is a three-channel fringe image after three-frequency fusion, where each channel corresponds to the first frame of the phase-shifted fringe image at three frequencies. The information integration and distribution module is used to effectively integrate the intermediate features and distribute them to multiple decoding branches according to the task type, ultimately achieving simultaneous prediction and output of the wrapped phase numerator, denominator and fringe order. Subsequently, the constructed dataset was divided into training, test, and validation sets in proportion to train the GD-UNet network. The model parameters were then optimized based on the training and test results to improve prediction accuracy and model generalization. After network training was completed, the actual three-frequency fused three-channel image was input into the trained GD-UNet model to obtain the prediction results of the numerator, denominator, and fringe order of the wrapped phase. Substituting the predicted numerator and denominator into the inverse tangent function, a preliminary wrapped phase map was calculated. Furthermore, a fringe order correction method based on connected domain segmentation was adopted to improve the accuracy of absolute phase calculation. Combining the corrected fringe order with the wrapped phase can accurately restore the absolute phase map.

[0039] As an embodiment, a single-frame fringe image phase recovery method based on the GD-UNet network can obtain high-precision absolute phase information from a single-frame fringe image. This embodiment includes the following four steps:

[0040] Step 1: Build a monocular structured light 3D reconstruction experimental platform. The platform consists of a DLP projector, a monocular camera, and a computer equipped with automatic acquisition software.

[0041] Step 2: Construct a fringe projection image dataset and generate input images and corresponding labels suitable for GD-UNet model training. The details are as follows:

[0042] Step 1: During a complete projection cycle, the projector sequentially projects 36 fringe patterns onto the object surface, covering three sets of 12-step phase-shifted fringe patterns with different frequencies (70, 64, and 59). In addition to the above frequency combinations, other frequency configurations that meet the phase decoding constraints are also applicable to this method.

[0043] Step 2: Use a monocular color camera to collect 36 deformed fringe images modulated by the object surface. The specific image intensity is expressed as follows:

[0044]

[0045] For the collected fringe images, the three-channel fringe images obtained by fusing the first frame of the phase-shifted fringe images at three frequencies are used as the input data of the GD-UNet network. For the 36 collected images, the numerator and denominator of the wrapped phase are obtained using the phase shift method:

[0046]

[0047] Where M(x, y) and D(x, y) are the numerator and denominator of the wrapping phase. Fringe order can be obtained using multifrequency heterodyning based on the wrapping phase obtained at three frequencies (see "Temporal Phase Unwrapping Algorithms for Fringe Projection Profilometry: A Comparative Review," by C Zuo et al.).

[0048] In order to enhance the learning ability of the network, an appropriate modulation threshold (THr) is set through the modulation function and the mask function to mask the invalid points in the training data graph.

[0049]

[0050] Step 3: In order to improve the efficiency of feature extraction and reduce the computational burden, the dataset constructed in this paper does not directly use the complete image (1600×1200) originally captured by the camera, but instead uses a sub-image of 800×640 size obtained by cropping the target area using the target detection model.

[0051] Step 4: Repeat the above steps for different application scenarios to generate a total of 500 training samples. The dataset is divided into training, validation, and test sets in an 8:1:1 ratio, containing 400 training samples, 50 validation samples, and 50 test samples, respectively. To improve the generalization ability of the trained model, all samples in the test set are from new scenarios not covered in the training and validation sets, ensuring the model's robustness and adaptability in unfamiliar environments.

[0052] Step 3: Build a GD-UNet model based on a deep convolutional neural network and input the dataset for training, as follows:

[0053] Step 1: Build the GD-UNet model;

[0054] The constructed GD-UNet network structure is as follows Figure 2 As shown in Figure 1, its input is a three-channel three-dimensional tensor of size (H, W, C), where H represents the image height (pixels), W represents the image width, and C represents the number of channels. The output of the model is the numerator and denominator of the wrapped phase and the fringe order.

[0055] GD-UNet consists of four core modules: the Downsampling Module (DownM), the Upsampling Module (UPB), the Residual Module (ResB), and the Information Gather-and-Distribute Mechanism (GDM). The Residual Module employs a dual-branch design: one branch consists of a single 3×3 convolution layer for rapid feature extraction; the other branch is composed of multiple convolution kernels of varying scales and depths to effectively extract multi-scale features. The outputs of the two branches are fused element-wise and then coupled to a LeakyReLU activation function to enhance the model's nonlinear representation capabilities.

[0056] During the encoding phase, the downsampling module (DownM) consists of a residual module and a maximum pooling layer. The residual module is responsible for extracting local and global feature information, while the pooling operation is used to compress the spatial dimension. The output of each residual module layer is retained and spliced ​​with the corresponding upsampling module through a skip connection in the subsequent decoding phase to enhance the flow and fusion of contextual information. The entire downsampling process is repeated four times, and each downsampling operation doubles the number of channels in the feature map while halving the spatial resolution, thereby gradually extracting deeper features with more semantic levels.

[0057] In the decoding stage, the upsampling module (UPM) first uses transposed convolution to restore the spatial resolution of the feature map (upsampling), and then splices it with the feature map of the corresponding layer in the encoder. Then, the residual module further extracts and fuses semantic information to improve the decoding quality. The final feature map is input to the information aggregation and distribution module (GDM), whose structure is as follows: Figure 3As shown in the figure, the GDM module first fuses multi-scale feature information through lateral connections and upsampling operations, and then uses a 3×3 convolution to further refine the feature expression capability. After convolution, the feature map is divided according to the channel dimension and fed into two subtask branches: the first branch outputs the numerator and denominator of the wrapped phase through a convolutional layer with 2 channels; the second branch is used for classification prediction of the stripe level, using a fully convolutional structure instead of the traditional fully connected layer design to reduce network complexity while maintaining prediction accuracy and improving model inference efficiency.

[0058] Step 2: Before training the GD-UNet network, the constructed dataset must be uniformly preprocessed to ensure that it meets the neural network's input format requirements. Specifically, the network's input image can remain in its original data type, while the corresponding label data (including the numerator and denominator of the wrapped phase and the fringe order map) must be uniformly converted to 32-bit floating point (float32) format. Subsequently, the image and label data are converted to a tensor format compatible with the PyTorch framework, providing a data foundation for subsequent model training and optimization.

[0059] Step 3: The preprocessed input data and standard data are fed into the GD-UNet network model for training. During training, the SmoothL1 loss function and the Structural Similarity Index (SSIM) are used as a joint loss function for the prediction of the numerator and denominator of the wrapped phase. This ensures regression accuracy while enhancing the perception of structural details. The specific expression of the loss function is as follows:

[0060]

[0061] Among them, λ1, λ2 are constants, indicating that SmoothL1 and SSIM losses are in Loss ND The importance of μ x and μ y are the mean values ​​of image x and y, respectively, indicating brightness; and are images x and y, respectively, indicating contrast; σ xy is the covariance of images x and y, indicating structural similarity; C1 and C2 are constants used to avoid the denominator being zero; is the difference between the predicted value and the true value.

[0062] For stripe-level prediction, multi-class cross entropy loss is used:

[0063]

[0064] Among them, k i is the one-hot encoding of the true label; is the probability of the i-th class predicted by the model.

[0065] During model training, the network weight parameters are continuously adjusted through the back-propagation algorithm, and the Adam optimizer is used to optimize the network to accelerate convergence and improve stability. The entire training process is executed in a cyclic iterative manner, continuously updating the parameters until the loss function converges to the set threshold or reaches the expected accuracy. In each training round, because the mapping relationship learned by the GD-UNet network is relatively stable and the training data itself does not show a significant overfitting trend, in this embodiment, the input order of the training data is not randomly shuffled, thereby simplifying the process configuration while ensuring training efficiency.

[0066] Step 4: Use the trained GD-UNet network model to predict the input image and correct the fringe order based on the prediction results to achieve high-precision phase recovery of the object to be measured, as follows:

[0067] Step 1: Input the three-channel fringe image, resulting from the three-frequency fusion, into the trained GD-UNet model. The network outputs the numerator M, denominator D, and corresponding fringe order k of the parcel phase. Substituting the numerator and denominator into the inverse tangent function, the high-frequency parcel phase φ is calculated. Simultaneously, the fringe order output by the network can be used to further infer the highly accurate absolute phase Φ.

[0068]

[0069] Φ=φ+2πk;

[0070] Step 2: In practical applications, due to many interference factors, the network may have errors in predicting the fringe order of some points. To improve the stability and accuracy of the fringe order prediction results, a connected domain segmentation strategy based on the wrapped phase map is used to correct the initial fringe order.

[0071] First, perform a binarization operation on the wrapped phase to obtain a binary image mask:

[0072]

[0073] The bwlabel function is then used to segment the white areas in the mask into connected domains, obtaining different connected domain labels, Label1 and Label2. Based on these labels, a complete connected domain segmentation map, Label, is constructed, thereby extracting multiple independent connected domains. All pixels in each connected domain theoretically have the same fringe level. To achieve this, the fringe level can be effectively corrected by counting the most frequently occurring fringe level in each connected domain and using it as the final prediction for that connected domain.

[0074] Step 3: Based on the original cropping strategy, padding is used to fill and align the edges of each sub-region to achieve seamless splicing of the absolute phase map and finally restore the complete absolute phase.

[0075] Example 2

[0076] This embodiment provides a phase recovery method for a single fringe image based on a GD-UNet network, including:

[0077] Build a monocular structured light 3D reconstruction experimental platform;

[0078] Construct fringe projection images and generate input images and corresponding labels suitable for GD-UNet model training;

[0079] Build a GD-UNet model based on a deep convolutional neural network and input the dataset for training;

[0080] The trained GD-UNet network model is used to predict the input image, and the fringe order is corrected according to the prediction results, thereby achieving high-precision phase recovery of the object to be measured.

[0081] Furthermore, the construction of a monocular structured light 3D reconstruction experimental platform includes:

[0082] The monocular structured light 3D reconstruction experimental platform consists of a DLP projector, a camera, and a computer equipped with automatic acquisition software.

[0083] Furthermore, the construction of the fringe projection image to generate input images and corresponding labels suitable for GD-UNet model training includes:

[0084] Generate a phase-shifted fringe pattern and burn it into the projector;

[0085] The computer controls the projector to project images and the camera to capture images;

[0086] Use target detection method to crop around the target area in the fringe image;

[0087] The dataset generated by the N-step phase shift method and triple-frequency heterodyne method includes the input image, fringe order, and the numerator and denominator of the wrapped phase.

[0088] Furthermore, a fringe projection image is constructed to generate an input image and corresponding labels suitable for GD-UNet model training, wherein the target detection method is used to crop around the target area in the fringe image, including:

[0089] The target detection model is used to identify the object area in the fringe image and generate the corresponding rectangular box;

[0090] According to the pixel position of the rectangular frame, all stripe images corresponding to the object are cropped to generate an image of fixed size.

[0091] Furthermore, a fringe projection image is constructed to generate an input image and corresponding labels suitable for GD-UNet model training. The dataset generated by the N-step phase shift method and the triple-frequency heterodyne method includes the input image, the fringe order, the numerator and denominator of the wrapped phase, including:

[0092] The first frame is extracted from each set of three-frequency phase shift images and mapped to RGB channels from high to low frequency, and the three-channel image is fused to form the input of GD-UNet.

[0093] The numerator and denominator of the wrapped phase are calculated using the N-step phase shift algorithm;

[0094] The corresponding fringe order is calculated using a triple-frequency heterodyne algorithm.

[0095] Furthermore, the construction of the GD-UNet model based on a deep convolutional neural network includes:

[0096] The GD-UNet network model is mainly composed of four core modules: downsampling module, upsampling module, residual module, and information integration and distribution module.

[0097] Furthermore, the residual module includes:

[0098] The residual module adopts a dual-branch structure. The first branch consists of a single 3×3 convolution layer to extract local detail features. The second branch is composed of multiple convolution kernels of different scales and depths to extract and fuse multi-scale features. After the two branches complete feature extraction, the outputs are fused through element-by-element addition.

[0099] Furthermore, the information integration and distribution module includes:

[0100] Through lateral connections with the encoder and upsampling operations, multi-scale feature information is fused and collected;

[0101] The fused features are further refined and enhanced through a 3×3 convolutional layer;

[0102] After feature enhancement, the feature map is divided by channel and sent to two subtask branches: one outputs the numerator and denominator of the wrapped phase through a convolutional layer with a channel number of 2, and the other is used for classification prediction of the stripe order.

[0103] Furthermore, after feature enhancement in the information integration and distribution module, the feature map is divided by channel and fed into two subtask branches: one branch outputs the numerator and denominator of the wrapped phase through a convolutional layer with a channel number of 2, and the other branch is used for classification prediction of the fringe order, including:

[0104] Different from the traditional classification method that uses fully connected layers, this method uses a fully convolutional structure to complete the classification task.

[0105] Furthermore, the construction of the GD-UNet model based on a deep convolutional neural network and inputting a data set for training includes:

[0106] For the prediction task of the wrapped phase numerator and denominator, a weighted combination of the SmoothL1 loss function and the structural similarity index (SSIM) loss is used as the joint loss function;

[0107] For stripe-level prediction, multi-class cross entropy loss is used as the loss function of the network.

[0108] Furthermore, the trained GD-UNet network model is used to predict the input image, and the fringe order is corrected according to the prediction result, thereby realizing phase recovery of the object to be measured, including:

[0109] According to the wrapped phase numerator and denominator output by the GD-UNet network, the formula Get the wrapped phase;

[0110] By binarizing the wrapped phase image, extracting the connected areas, and counting the distribution of fringe orders in each connected area, the fringe order value with the highest frequency is used as the final prediction result of the area, thereby correcting the abnormal prediction.

[0111] The absolute phase is obtained according to the wrapped phase and the corrected fringe order using the formula Φ=φ+2πk. Then, according to the original cropping strategy, the seamless splicing of the absolute phase is achieved, and finally the complete absolute phase map is restored.

[0112] The above are merely preferred embodiments of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

Claims

1. A phase recovery method for a single fringe image based on a GD-UNet network, characterized in that: include: Build a monocular structured light 3D reconstruction experimental platform and obtain fringe projection images based on the experimental platform; Constructing a GD-UNet model based on a deep convolutional neural network, and training the GD-UNet model using a training set; The trained GD-UNet network model is used to predict the fringe projection image, and the fringe order is corrected according to the prediction result to achieve high-precision phase recovery of the object to be measured.

2. The method according to claim 1, characterized in that The monocular structured light 3D reconstruction experimental platform consists of a DLP projector, a camera, and a computer equipped with automatic acquisition software; The computer controls the DLP projector to project an image and the camera to capture an image, thereby obtaining the fringe projection image.

3. The method according to claim 1, characterized in that The process of obtaining a fringe projection image based on the experimental platform includes: Generate a phase-shifted fringe pattern and burn it into the projector; The fringe image is collected using the constructed experimental platform, and the target detection method is used to crop the fringe image around the target area in the fringe image to obtain a cropped image; The cropped image is processed using an N-step phase shift method and a triple-frequency heterodyne method to generate a fringe projection image. The fringe projection image includes an input image, a fringe order, and a numerator and denominator of a wrapped phase. The input image is used as input to a GD-UNet model, and the fringe order and numerator and denominator of the wrapped phase are used as supervisory labels for model training.

4. The method according to claim 3, characterized in that The process of cropping around the target area in the fringe image using the target detection method includes: using the target detection model to identify the object area in the fringe image and generate a corresponding rectangular frame, and cropping all fringe images corresponding to the object according to the pixel position of the rectangular frame to generate an image of a fixed size.

5. The method according to claim 3, characterized in that The process of processing the cropped image using the N-step phase shift method and the triple-frequency heterodyne method includes: The first frame is extracted from each set of three-frequency phase-shift images and mapped to RGB channels from high to low frequency. The three-channel image is fused and used as the input of GD-UNet. The numerator and denominator of the wrapped phase are calculated using the N-step phase-shift algorithm, and the corresponding fringe order is calculated using the three-frequency heterodyne algorithm to generate the fringe projection image.

6. The method according to claim 1, wherein The GD-UNet model based on deep convolutional neural network includes a downsampling module, an upsampling module, a residual module and an information integration and distribution module.

7. The method according to claim 6, characterized in that The residual module adopts a dual-branch structure design, where the first branch consists of a single-layer 3×3 convolution to extract local detail features; The second branch is composed of multiple convolution kernels of different scales and depths to achieve the extraction and fusion of multi-scale features. After the two branches complete feature extraction respectively, the outputs are fused by element-by-element addition. The fused features are used as the input of subsequent modules to enhance the model's perception of features at different scales.

8. The method according to claim 6, characterized in that The information integration and distribution module fuses and collects multi-scale feature information through a horizontal connection with the encoder and an upsampling operation, and further refines and enhances the fused features through a 3×3 convolutional layer. After feature enhancement, the feature map is divided by channel and sent to two subtask branches. One branch outputs the numerator and denominator of the wrapped phase through a convolutional layer with a channel number of 2, and the other branch is used for classification prediction of the stripe order. The prediction results of the numerator and denominator of the wrapped phase and the stripe order are used for subsequent phase recovery of the object to be measured.

9. The method according to claim 1, characterized in that The process of realizing high-precision phase recovery of the object to be measured includes: According to the numerator and denominator of the wrapped phase output by the trained GD-UNet network model, the wrapped phase map is calculated by the inverse tangent function; The wrapped phase image is binarized to extract the connected regions. The distribution of fringe orders in each connected region is counted. The fringe order with the highest frequency is used as the final prediction result of the region, thereby correcting the initially predicted fringe order. According to the corrected fringe order and wrapped phase, the absolute phase is calculated by the formula, and then according to the original cropping strategy, the seamless splicing of the absolute phase is achieved, and finally the complete absolute phase map is restored.

10. The method according to claim 9, characterized in that For the prediction task of the wrapped phase numerator and denominator, a weighted combination of the SmoothL1 loss function and the structural similarity index (SSIM) loss is used as the joint loss function; For stripe-level prediction, multi-class cross entropy loss is used as the loss function of the network.

Images