Phase unwrapping method based on channel-space-frequency three-branch U-Net

Abstract

The invention provides a phase unwrapping method based on a channel-space-frequency three-branch U-Net. The method comprises the following steps: acquiring a phase unwrapping data set and dividing the phase unwrapping data set into a training data set, a verification data set and a test data set; secondly, constructing a channel-space-frequency three-branch U-Net network; inputting the training data set into the channel-space-frequency three-branch U-Net network for iterative training, and storing an optimal model weight after training is completed; and finally, inputting a test data set into the trained channel-space-frequency three-branch U-Net network, and outputting a predicted unwrapping phase diagram. According to the method, the unification of noise suppression and detail reservation can be realized, the phase unwrapping precision and robustness in a low signal-to-noise ratio scene are effectively improved, the unwrapping result of a data missing area is enabled to better conform to a physical law, and a reliable phase unwrapping scheme is provided for ground surface deformation monitoring of synthetic aperture radar interference measurement.

Description

Technical Field

[0001] This invention relates to the field of synthetic aperture radar interferometry, and more particularly to a phase unwrapping method based on a channel-space-frequency three-branch U-Net. Background Technology

[0002] Synthetic Aperture Radar Interferometry (SAR) has become a core method for monitoring surface deformation due to its advantages of all-weather, all-time capability and high spatial resolution. Phase unwrapping, a crucial step in SAR data processing, directly determines the reliability of deformation inversion results. The core objective of phase unwrapping is to recover continuous phase information reflecting the true surface elevation or deformation from the wrapped phase within the range of [-π, π]. However, in practical applications, issues such as noise interference, dense phase jumps, and feature redundancy consistently hinder the improvement of unwrapping accuracy.

[0003] Current phase unwrapping techniques are mainly divided into two categories: traditional methods and deep learning methods. Traditional phase unwrapping methods have inherent drawbacks: methods such as branch cutting and least squares rely on manually designed rules and lack adaptability. In low signal-to-noise ratio scenarios, they are susceptible to speckle noise and measurement errors, leading to misjudgment of residual points. When dealing with discontinuous regions, it is difficult to balance noise suppression and detail preservation, often resulting in unwrapped phase breaks or over-smoothing.

[0004] Existing deep learning methods suffer from optimization bottlenecks. First, the attention mechanism design is flawed. Most mainstream attention modules employ a sequential channel and spatial structure, failing to consider the frequency domain characteristics where phase signal noise is concentrated at high frequencies and effective information at low frequencies, thus hindering multi-dimensional feature collaborative optimization. Second, frequency domain processing methods are limited. Traditional frequency domain processing is only used as an independent pre-unwrapping or post-processing step, employing fixed-rule filtering, which cannot be linked with channel and spatial features and is difficult to adapt to scenarios with dynamically changing noise distribution. Third, the loss function constraints are too simplistic. Existing models mostly rely on pixel-level loss, only optimizing the numerical difference between the predicted phase and the true value, ignoring the physical characteristics of the phase. This often results in numerically accurate but physically distorted unwrapping results, such as blurred edges or contradictions with the original wrapped phase. Summary of the Invention

[0005] This invention provides a phase unwrapping method based on a channel-space-frequency three-branch U-Net, mainly including:

[0006] S1: Obtain the phase unwrapping dataset and divide the phase unwrapping dataset into a training dataset, a validation dataset, and a test dataset;

[0007] S2: Construct a channel-space-frequency three-branch U-Net network; wherein the channel-space-frequency three-branch U-Net network includes an input layer, an encoder, a bottleneck layer, a decoder, a post-processing layer, and an output layer;

[0008] S3: Input the training dataset into the channel-space-frequency three-branch U-Net network for iterative training to generate a trained channel-space-frequency three-branch U-Net network. The validation dataset is used to monitor the training effect in real time.

[0009] Among them, the loss value is calculated through a multi-constraint loss function, the model parameters are iteratively optimized through backpropagation, and the optimal model weights are saved after training is completed;

[0010] S4: Input the test dataset into the trained channel-space-frequency three-branch U-Net network and output the predicted unwrapped phase map.

[0011] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects:

[0012] This invention proposes a phase unwrapping method based on a channel-space-frequency three-branch U-Net. The method uses a U-Net architecture as its foundation, embedding a parallel attention module with three branches (channel, space, and frequency) after the encoder convolutional block: the channel branch filters effective feature channels, the spatial branch focuses on key phase regions, and the frequency branch extracts frequency domain information through two-dimensional fast Fourier transform and dynamically generates noise weights, achieving multi-dimensional feature collaborative purification. Simultaneously, frequency domain processing is integrated as an embedded module within the network, replacing traditional independent filtering steps, allowing it to deeply integrate with channel and spatial features and adapt to scenarios with dynamic noise distribution. Furthermore, a multi-constraint loss function is designed, fusing pixel-level loss, gradient consistency loss, and phase consistency loss, balancing numerical accuracy and physical properties. This method effectively addresses the problems of single attention dimension, disconnected frequency domain processing, and insufficient loss constraints through multi-dimensional attention collaboration and physical constraint optimization, achieving a balance between noise suppression and detail preservation, and improving the accuracy and robustness of phase unwrapping.

[0013] (1) In existing phase unwrapping techniques, attention modules mostly process channels first and then space, with channel and spatial attention being sequential and lacking frequency domain branches. In this invention, the channel-space-frequency three branches are processed in parallel and synergistically through residual fusion. The effective channels enhanced by the channel branch are given priority weights for high-gradient regions by the spatial branch, while the low-frequency components are retained by the frequency domain branch, achieving multi-dimensional dynamic purification of features. This solves the problem that existing attention modules only process channels and space sequentially, only weighting in two dimensions: feature type and spatial location, ignoring the significant high-frequency concentration of phase noise in the frequency domain, making it difficult to specifically distinguish between structural stripes and random noise from a frequency perspective, and making it difficult to simultaneously consider the joint optimization of feature type, spatial location, and frequency components. By linking the channel-space-frequency three-branch attention modules, the model accurately captures the effective feature types, key location information, and stable low-frequency components required for phase unwrapping, ultimately achieving a more robust unwrapping effect. The resulting encoded features show a significant reduction in noise and a more complete structural outline on the entire image. Since the spatial branch focuses on high gradient locations such as phase discontinuities and dense fringes, while the frequency branch adaptively weakens random noise components based on high-frequency energy, the multi-domain consistent high-response regions that are ultimately retained are often the real fringe structures and phase transition locations required for phase unwrapping; thus, the feature representation at fringe breaks is more continuous; and the responses of pseudo-fringe and isolated noise points in low-coherence regions are suppressed.

[0014] (2) In traditional frequency domain unwrapping methods, the solution is only used to solve the pre-unwrapping result and does not involve noise perception weighting at the feature level. In existing deep learning, the frequency domain module is mostly replaced by attention and no weight generation method is designed for the physical characteristics of high-frequency concentration of phase noise. In this invention, the channel weight is calculated in real time by the high-frequency energy of the current feature in the frequency domain and obtained by mapping through a learnable multilayer perceptron. Different inputs will generate different weighting coefficients instead of using a fixed filter. This solves the problems of traditional frequency domain filtering, which can only set a fixed low-frequency threshold, which cannot adapt to the dynamic changes in InSAR phase noise distribution, and the problem of processing high-frequency components of all interferograms in a "one-size-fits-all" manner, which cannot adapt to the changes in InSAR phase noise in different scenarios to adjust the suppression strength, thus making it difficult to distinguish and treat useful high-frequency structures and random high-frequency noise separately.

[0015] (3) Traditional phase unwrapping methods mostly focus on pixel-level error loss and gradient consistency loss, ignoring the physical characteristics of the phase. This invention incorporates phase consistency loss, enabling the network to simultaneously satisfy numerical accuracy and physical rules. The multi-constraint loss directly incorporates physical laws such as phase periodicity, phase gradient continuity, and consistency with the observed interferogram into the loss. During training, the network is forced to simultaneously satisfy the fitted data and adhere to physics, forming a dual-driven unwrapping mechanism of physics and data. In areas of low coherence, occlusion, or severe fringe breakage, the physical constraint loss binds these areas to constraints such as interferometric geometry and phase gradient consistency, ensuring that the network provides solutions that satisfy the overall physical unwrapping equation even in areas lacking data, thus achieving the effect of physical laws filling data gaps. This addresses the problem that existing loss functions easily lead to numerical accuracy but physical distortion. (See attached figures for details.)

[0016] Figure 1 This is a flowchart illustrating an embodiment of the present invention.

[0017] Figure 2 This is a partial training set winding phase diagram of an embodiment of the present invention.

[0018] Figure 3 This is a partial training set unwrapped phase diagram of an embodiment of the present invention.

[0019] Figure 4 This is a schematic diagram of the channel-space-frequency three-branch attention U-Net network according to an embodiment of the present invention.

[0020] Figure 5 This is a schematic diagram of the channel-space-frequency three-branch attention module network according to an embodiment of the present invention.

[0021] Figure 6 This is a schematic diagram of a multi-constraint loss function network according to an embodiment of the present invention.

[0022] Figure 7 This is a test winding phase diagram of an embodiment of the present invention.

[0023] Figure 8 This is a test unwrapping phase diagram according to an embodiment of the present invention. Detailed implementation method.

[0024] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0025] like Figure 1 This embodiment of a phase unwrapping method based on a channel-space-frequency three-branch U-Net may include:

[0026] S1: Obtain the phase unwrapping dataset and divide the phase unwrapping dataset into a training dataset, a validation dataset, and a test dataset.

[0027] Specifically, the phase unwrapping dataset used was proposed by Zhou Lifan's team and published in *IEEE Geoscience and Remote Sensing Magazine*. The simulation data is based on the NASA Space Shuttle Radar Terrain Mission Digital Elevation Model (DEM) database, selecting real terrain data covering the entire territory of China, including various landform types such as plains, mountains, plateaus, and basins. A total of 1000 original DEM images were extracted, each with a size of 3000×3000 pixels, corresponding to a real geographic resolution of 30 meters per pixel. The wrapped phase map was generated by simulating the principle of synthetic aperture radar interferometry (SAR), while Gaussian noise and Perlin noise were introduced to simulate atmospheric turbulence phase, thus restoring the noise interference characteristics of real SAR interferometry data. The real data used was the interferogram measured by the 100TanDEM-X satellite. This satellite data has the advantages of high resolution and high precision imaging. After radiometric and geometric correction, the interferogram was combined with terrain data for phase calibration, forming wrapped-unwrapped phase annotation pairs in a real-world scenario, used to verify the model's generalization ability. The training set in this dataset contains 30,000 pairs of simulated entanglement phases and their corresponding unentanglement phases. A partial training set entanglement phase diagram from this embodiment of the invention is shown below. Figure 2 The training set unwrapped phase diagrams shown in the figure and the corresponding diagrams are as follows: Figure 3 As shown, the validation set contains 1000 pairs of simulated wrapped phases and their corresponding unwrapped phases, while the test set contains 100 pairs of real-world wrapped phases and their corresponding unwrapped phases. The training set is used for the core iterative training of the model parameters; the validation set is responsible for verifying the model's basic unwrapping performance in standard simulated scenarios; and the test set is specifically used to evaluate the model's generalization ability in real InSAR data.

[0028] For example.

[0029] S2: Construct a channel-space-frequency three-branch U-Net network; wherein the channel-space-frequency three-branch U-Net network includes an input layer, an encoder, a bottleneck layer, a decoder, a post-processing layer, and an output layer.

[0030] Specifically, a schematic diagram of the channel-space-frequency three-branch attention U-Net network of this invention is shown below. Figure 4 As shown. The specific structure and data processing flow of each part are as follows:

[0031] The input layer receives a single-channel wrapped phase image and transforms it into a multi-channel feature representation through a residual block structure, laying the foundation for deep feature extraction and phase unwrapping processing in the subsequent encoder. The residual block contains two 3×3 convolutional layers, two batch normalization layers, a ReLU activation function, and a 1×1 convolutional residual connection for channel dimension matching, converting the single-channel wrapped phase map into 8-channel features.

[0032] The encoder is responsible for extracting multi-scale phase features from the 8-channel features obtained from the input layer and dynamically refining them through a three-branch attention module. It consists of five encoding layers, each composed of a convolutional block and a channel-space-frequency three-branch attention module connected in series. The number of feature channels increases from 8 to 128 layer by layer, while the spatial resolution halves from 256×256 to 8×8 layer by layer. The convolutional block uses a double-layer convolutional structure to address the gradient vanishing problem in deep networks, specifically including two 3×3 convolutional layers. The first layer handles channel transformation and initial feature extraction, while the second layer refines features within a unified channel space. The number of convolutional kernels increases with the encoder layer, with a stride of 1. Copy padding is used to avoid losing boundary information. Batch normalization layers accelerate model training convergence and reduce overfitting. The ReLU activation function introduces non-linear feature representation. Residual connections directly add the input features to the convolutional layer output, ensuring smooth information flow. The output of each convolutional block undergoes feature enhancement through the corresponding three-branch attention module, achieving multi-dimensional adaptive feature refining. The calculation formula is as follows:

[0033] ;

[0034] in, and These are the first and second layers of 3×3 convolutions, respectively. For batch normalization operations, The input is the residual block. The three-branch attention module adopts a parallel structure of "channel-space-frequency domain" to perform multi-dimensional purification of the output features of the convolutional block. The network diagram of the channel-space-frequency three-branch attention module in this embodiment is shown below. Figure 5 As shown. The structure includes: the first branch, channel attention to filter effective feature channels and suppress redundant channel interference. Global average pooling and global max pooling are performed on the output after feature extraction from the convolutional block to extract global statistical information in the channel dimension, resulting in two 1x1 channel feature maps; the two pooling results are input into a shared fully connected layer to complete dimensionality transformation and feature fusion, and then a sigmoid activation layer is used to generate adaptive weight coefficients for each channel; finally, these weight coefficients are weighted and modulated with the original input feature map in the channel dimension to complete the feature filtering in the channel dimension. The calculation formula is:

[0035] ;

[0036] in, It is the Sigmoid activation function. and These are global average pooling and global max pooling, respectively. The first branch is the channel weight map; the second branch focuses on spatial attention in key regions of phase change. For each spatial location of the output of the convolutional block, average pooling and max pooling operations are performed in the channel dimension to compress multi-channel features into a single-channel spatial feature map. The two pooling results are concatenated and input into a 7×7 convolutional layer to complete spatial feature dimensionality reduction and fusion. Adaptive weight coefficients for each spatial location are generated through a Sigmoid activation layer. Finally, these weight coefficients are weighted and modulated with the original input feature map in the spatial dimension to complete spatial feature focusing. The calculation formula is as follows:

[0037] ;

[0038] in, This indicates concatenation along the channel dimension. This is a 7×7 convolution operation. The spatial weight map is used; the third branch, frequency domain attention, adaptively suppresses high-frequency noise. After feature extraction from the convolutional block, the output uses a two-dimensional fast Fourier transform (FFT) in the frequency attention module to convert the spatial domain features to the frequency domain, extracting the high-frequency residuals of the amplitude spectrum. Low-frequency components are extracted using 3×3 average pooling, and the high-frequency components are calculated by subtracting the low-frequency components from the amplitude. The high-frequency components are averaged across the spatial dimension to obtain a high-frequency descriptor. This high-frequency descriptor dynamically generates channel-level weights using a multilayer perceptron. The input feature map is multiplied channel-by-channel by the adaptive weights, dynamically adjusting the channel response based on the frequency domain features of different inputs to achieve adaptive suppression of high-frequency noise. The calculation formula is as follows:

[0039] ;

[0040] ;

[0041] ;

[0042] in, For two-dimensional fast Fourier transform, For the frequency domain amplitude spectrum, For high-frequency residuals, For global mean pooling, The first part is a frequency domain weight map. After the three branches perform parallel operations to optimize features in each dimension, the weighted modulation feature maps output by each branch are concatenated together through a connection layer. Finally, a 1×1 convolutional layer is used to adjust the number of channels, and a residual connection is made with the original input feature map of the module to serve as the final output of the three-branch attention module. The calculation formula is as follows:

[0043] ;

[0044] ;

[0045] in, For channel-dimensional splicing, A 1×1 convolution merges the 3 channels into a single channel. This is the final output feature.

[0046] The bottleneck layer, located at the deepest part of the encoder, consists of one convolutional block and one channel-space-frequency domain three-branch attention module, with 128 input channels, 256 output channels, and a spatial resolution of 8×8. This module is responsible for capturing the most abstract phase semantic features, integrating multi-scale information, and providing a high-quality feature foundation for decoder reconstruction.

[0047] The decoder is responsible for restoring spatial resolution and fusing encoder features. It consists of five layers. Each decoder layer comprises one upsampled residual block and one transposed convolutional layer, enhanced with a skip connection by combining the attention mechanism module of the convolutional block input from each encoder layer. This refines the encoder features before fusing them, with the number of feature channels decreasing layer by layer from 256 to 1, and the spatial resolution gradually restored from 8×8 to 256×256. The transposed convolutional layer uses transposed convolution to achieve upsampling with a stride of 2 and a kernel size of 2×2, ensuring that the spatial resolution is doubled. Its calculation formula is as follows:

[0048] ;

[0049] in, For the decoder input features, Features after upsampling The kernel size is [size]. The step size is [value]. The structure of the upsampled residual block is similar to that of the encoder convolutional block. The number of channels is adjusted to adapt to the feature fusion requirements of the decoder. The calculation formula is:

[0050] ;

[0051] in, It is a 3×3 convolution, and the input channels are the number of channels obtained by concatenating the decoder features and encoder features. The convolution is 3×3, and the output channels are the target number of channels. A convolutional block attention mechanism module is introduced between the encoder and decoder to refine the encoder features before fusion. The steps are as follows:

[0052] (1) The output features of each layer of the encoder are input into the convolutional block attention mechanism module, and after sequential processing from channel attention to spatial attention, weighted features are generated. The calculation formula is as follows:

[0053] ;

[0054] in, The original features of the encoder, Weighted features of the convolutional block attention mechanism module;

[0055] (2) The weighted encoder features and the transposed convolutional output features of the decoder are concatenated along the channel dimension;

[0056] (3) The spliced ​​features are input into the upsampled residual block to further refine the features and improve the reconstruction accuracy.

[0057] The post-processing layer consists of two convolutional layers, which refine the decoder output features. The first convolutional layer is a 3×3 convolution with batch normalization and ReLU activation, 1 input channel and 16 output channels, enhancing feature representation. The second convolutional layer is a 1×1 convolution with 16 input channels and 1 output channel, mapping the features to a single-channel unwrapped phase map. The output phase map is slightly smoothed to ensure phase continuity, and the final output is an unwrapped phase tensor with a size of 1×256×256.

[0058] For example.

[0059] S3: Input the training dataset into the channel-space-frequency three-branch U-Net network for iterative training to generate a trained channel-space-frequency three-branch U-Net network. The validation dataset is used to monitor the training effect in real time.

[0060] The loss value is calculated using a multi-constraint loss function, the model parameters are iteratively optimized through backpropagation, and the optimal model weights are saved after training.

[0061] Specifically, the steps are as follows:

[0062] (1) Set the initial learning rate, maximum learning rate, number of training batches, and number of training rounds; the initial learning rate is 0.0001, the maximum learning rate is 0.01, the number of training batches is 4, and the number of training rounds is 100.

[0063] (2) During the training process, batch normalization and early stopping strategies are used to prevent network overfitting;

[0064] (3) The Adam gradient optimization algorithm is used to optimize the training of the channel-space-frequency three-branch U-Net network structure. During the optimization process, a composite function of pixel-level precision loss, gradient consistency loss, and phase consistency loss is used as the loss function. The network diagram of the multi-constraint loss function in this embodiment is shown below. Figure 6 As shown.

[0065] The pixel-level accuracy loss is calculated by taking the absolute value of the difference between the predicted phase and the true phase pixel by pixel, averaging it over all pixels. The formula is as follows:

[0066] ;

[0067] in, To predict the untangling phase, For the true untangling phase, Image size;

[0068] Gradient consistency loss is used to constrain the consistency of the predicted unwrapped phase and the true unwrapped phase in terms of spatial gradient distribution, ensuring that the edge structure, texture details, and local rate of change of the phase are not lost, and avoiding edge blurring or structural distortion caused by overfitting of the model to pixel-level errors. A first-order central difference operator is used to describe the local rate of change of the phase, and boundary pixels are supplemented with front or back differences to balance the gradient magnitudes of the edges and interior. The horizontal and vertical gradients of the true and predicted phases are calculated separately, and the gradient differences in the horizontal and vertical directions of the predicted and true unwrapped phases are calculated separately. The absolute values ​​of the gradient differences in the horizontal and vertical directions are summed, and then divided by twice the total number of image pixels to normalize, yielding the final loss value. The calculation formula is as follows:

[0069] ;

[0070] in, and These are the gradient operators in the x and y directions, respectively. To predict the untangling phase, For the true untangling phase, Image size;

[0071] Phase consistency loss involves converting the predicted unwrapped phase into a complex number, taking an angle from this complex number to obtain the rewrapped phase, and calculating the difference between the rewrapped phase and the original wrapped phase. Since the phase has a 2π periodicity, this difference needs to be adjusted to the range of -π to π. The absolute average of the adjusted difference is taken as the phase consistency loss; its calculation formula is as follows:

[0072] ;

[0073] in, The winding phase function remaps the unwrapped phase back to the winding phase, and its calculation formula is as follows:

[0074] ;

[0075] in, To predict the untangling phase, This is the original entanglement phase.

[0076] Finally, a multi-constraint loss function that integrates pixel-level accuracy, gradient structure consistency, and phase physical compatibility guides comprehensive model optimization. The total loss formula is as follows:

[0077] ;

[0078] in, For pixel-level precision loss, For gradient consistency loss, This represents the loss of phase consistency.

[0079] (4) Repeat steps (2) and (3) for iterative training until the channel-space-frequency three-branch U-Net network structure converges, and obtain the final channel-space-frequency three-branch U-Net network model and weights for interferometric phase unwrapping, and save the optimal model weights.

[0080] For example.

[0081] S4: Input the test dataset into the trained channel-space-frequency three-branch U-Net network and output the predicted unwrapped phase map.

[0082] Specifically, the optimal model weights saved during the training phase are loaded into the channel-space-frequency three-branch U-Net network to complete initialization; then, the entangled phase data of the test set is input into the network for inference, wherein the test entangled phase map of this embodiment is as follows. Figure 7 As shown; after forward computation via the network, the corresponding interference phase unwrapping prediction result is directly obtained. The test unwrapped phase diagram of this embodiment of the invention is shown below. Figure 8 As shown.

[0083] It is worth noting that this application employs a parallel structure for the channel-space-frequency three-branch attention module. This allows the model to accurately capture the effective feature types, key location information, and stable low-frequency components required for phase unwrapping, achieving a more robust unwrapping effect. The resulting encoded features exhibit significantly reduced noise and a more complete structural outline across the entire image. Since the spatial branch focuses on high-gradient locations such as phase discontinuities and dense stripes, while the frequency branch adaptively weakens random noise components based on high-frequency energy, the ultimately retained multi-domain consistent high-response regions are often precisely the true stripe structures and phase transition locations required for phase unwrapping. This results in more continuous feature representation at stripe breaks and reduced responses to pseudo-stripes and isolated noise points in low-coherence regions. Furthermore, the dynamic weighting design of the frequency domain attention precisely adapts to the characteristics of phase noise. Instead of a one-size-fits-all low-pass filter, this application achieves adaptive noise suppression by finely distinguishing between useful high-frequency structures and useless high-frequency noise. Traditional frequency domain filtering often simply cuts off based on a frequency threshold, uniformly removing frequencies above a certain threshold. This easily leads to the loss of useful high-frequency structures such as stripe edges and phase transitions. In this application, high frequencies in different channels are preserved or suppressed to varying degrees, rather than being uniformly filtered out.

[0084] For the loss function module, this application employs a multi-constraint loss function design that integrates physical constraints to ensure the physical rationality of the unwrapped phase. A composite function of pixel-level accuracy loss, gradient consistency loss, and phase consistency loss is used as the loss function. In regions of low coherence, occlusion, or severe fringe breakage, the physical constraint loss binds these regions to constraints such as interference geometry and phase gradient consistency. This ensures that the network must provide solutions that satisfy the overall physical unwrapping equation even in areas lacking data, effectively filling data gaps with physical laws. The physical constraint loss, by additionally penalizing non-closed, inconsistent, and discontinuous solutions, eliminates spurious solutions, making the network more inclined to select physical solutions.

[0085] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. A phase unwrapping method based on a channel-space-frequency three-branch U-Net, characterized in that, The method includes: S1: Obtain the phase unwrapping dataset and divide the phase unwrapping dataset into a training dataset, a validation dataset, and a test dataset; S2: Construct a channel-space-frequency three-branch U-Net network; wherein the channel-space-frequency three-branch U-Net network includes an input layer, an encoder, a bottleneck layer, a decoder, a post-processing layer, and an output layer; S3: Input the training dataset into the channel-space-frequency three-branch U-Net network for iterative training to generate a trained channel-space-frequency three-branch U-Net network. The validation dataset is used to monitor the training effect in real time. Among them, the loss value is calculated through a multi-constraint loss function, the model parameters are iteratively optimized through backpropagation, and the optimal model weights are saved after training is completed; S4: Input the test dataset into the trained channel-space-frequency three-branch U-Net network and output the predicted unwrapped phase map.

2. The method according to claim 1, characterized in that, The input layer is used to receive a single-channel wound phase image and convert it into a multi-channel feature representation through a residual block structure; The bottleneck layer is located at the deepest part of the encoder and consists of one convolutional block and one channel-space-frequency domain three-branch attention module, with 128 input channels, 256 output channels, and a spatial resolution of 8×8. The post-processing layer consists of two convolutional layers and is used to refine the output features of the decoder.

3. The method according to claim 1, characterized in that, The encoder contains five coding layers, each consisting of a convolutional block and a channel-space-frequency three-branch attention module connected in series to form the main branch for feature extraction.

4. The method according to claim 3, characterized in that, The channel-space-frequency three-branch attention module adopts a channel-space-frequency three-branch parallel structure to perform multi-dimensional purification of the convolutional block output features, specifically: The first branch is the channel attention branch, which performs global average pooling and global max pooling on the output of the convolutional block, with a window size of 1×1 for each. It extracts global statistical information in the channel dimension, resulting in two 1×1 channel feature maps. The two pooling results are then input into a shared fully connected layer to complete the dimensionality transformation and feature fusion. A sigmoid activation layer is then used to generate adaptive weight coefficients for each channel. Finally, these weight coefficients are weighted and modulated with the original input feature map in the channel dimension to complete the feature selection in the channel dimension. The second branch is the spatial attention branch. For each spatial location output by the convolutional block, average pooling and max pooling operations are performed in the channel dimension to compress multi-channel features into a single-channel spatial feature map. The two pooling results are concatenated and then input into a 7×7 convolutional layer to complete spatial feature dimensionality reduction and fusion. Adaptive weight coefficients for each spatial location are generated through a Sigmoid activation layer. Finally, the weight coefficients are weighted and modulated with the original input feature map in the spatial dimension to complete spatial feature focusing. The third branch is the frequency attention branch. After the convolutional block extracts features, the output uses a two-dimensional fast Fourier transform to convert the spatial domain features to the frequency domain in the frequency attention module. The high-frequency residual of the amplitude spectrum is extracted, and the low-frequency component is extracted using 3x3 average pooling. The high-frequency component is the amplitude minus the low-frequency component. The high-frequency component is averaged in the spatial dimension to obtain the high-frequency descriptor. The high-frequency descriptor generates channel-level weights through a multilayer perceptron. The input feature map is multiplied with the weights channel by channel to suppress the channel features corresponding to high-frequency noise. Finally, the output features of the three branches are weighted and concatenated, then multi-dimensional information is fused through 1×1 convolution, and then connected with the original input residual to generate an enhanced feature map; the enhanced feature map enters the max pooling layer for downsampling and continues to participate in the deep feature extraction of the encoder. After processing through five coding layers, the network reaches the bottleneck layer, which has the largest number of channels (256) and the smallest spatial resolution. It is also equipped with the channel-space-frequency three-branch attention module to extract the highest level of semantic features.

5. The method according to claim 1, characterized in that, The encoder contains five decoding layers. Each decoding layer is upsampled through transposed convolution and skipped connections are made with attention-enhanced features from each layer of the encoder. The features of the skip connections are further processed by convolutional blocks and then fused with upsampled features to provide the network with multi-scale contextual information. Finally, the decoder output passes through a post-processing layer consisting of 3×3 convolutions and 1×1 convolutions to output the final unwrapped phase map.

6. The method according to claim 1, characterized in that, The multi-constraint loss function includes pixel-level accuracy loss, gradient consistency loss, and phase consistency loss.

7. The method according to claim 6, characterized in that, The multi-pixel level accuracy loss is obtained by calculating the pixel-by-pixel difference between the predicted phase and the true phase, taking the absolute value, and averaging it over all pixels.

8. The method according to claim 6, characterized in that, The gradient consistency loss is used to constrain the consistency of the predicted unwrapped phase and the true unwrapped phase in the spatial gradient distribution, so as to ensure that the edge structure, texture details and local rate of change of the phase are not lost, and to avoid edge blurring or structural distortion caused by overfitting the model to pixel-level errors. The gradient consistency loss uses a first-order central difference operator to describe the local phase change rate, and the boundary pixels are supplemented with front or back differences to balance the gradient magnitudes of the edges and the interior. The horizontal and vertical gradients of the true phase and the predicted phase are calculated respectively, and the gradient differences in the horizontal and vertical directions of the predicted unwrapped phase and the true unwrapped phase are calculated respectively. The absolute values ​​of the gradient differences in the horizontal and vertical directions are summed and then divided by twice the total number of image pixels to normalize, thus obtaining the final loss value.

9. The method according to claim 6, characterized in that, The phase consistency loss is obtained by converting the predicted unwrapped phase into a complex number, taking an angle on the complex number to obtain the rewrapped phase, calculating the difference between the rewrapped phase and the original wrapped phase. Since the phase has a 2π periodicity, this difference needs to be adjusted to the range of -π to π. The absolute average of the adjusted difference is taken as the phase consistency loss.

10. The method according to any one of claims 7 to 9, characterized in that, The multi-constraint loss function, which integrates the pixel-level precision loss, gradient consistency loss, and phase consistency loss, guides the comprehensive optimization of the channel-space-frequency three-branch U-Net network.

Images