A phase unwrapping method based on a global attention upsampling network

Abstract

This invention aims to provide a phase unwrapping method based on a global attention upsampling network, comprising the following steps: A. Constructing a deep neural network structure, specifically an encoding network and a decoding network; wherein the encoding network includes five sequentially connected 2×2 max pooling layers and six PU-M-Net modules; the decoding network includes five GAU modules, five PU-M-Net modules, and a stacked module I; B. Constructing multiple training sets with different noise levels to train the deep neural network and obtain unwrapping networks suitable for unwrapping interferograms with different noise levels; C. Determining the noise level of the interferogram to be unwrapped, and then matching it to the unwrapping network of the corresponding noise level for unwrapping. The method of this invention has high unwrapping efficiency and robustness.

Description

Technical Field

[0001] This invention relates to the field of image processing, and more specifically to a phase unwrapping method based on a global attention upsampling network. Background Technology

[0002] Phase unwrapping (PU), a core step in many interferometric techniques, is widely used in fields such as interferometric synthetic aperture radar, optical interferometry, and nuclear magnetic resonance imaging. The principal phase value obtained from interferometric techniques is confined to the interval (-π, π]. The process of recovering the original phase by adding an integer multiple of the phase period to the principal phase value is called phase unwrapping. Currently, existing unwrapping algorithms have various shortcomings. For example, traditional algorithms such as QGPU and ILS can achieve good unwrapping results when the signal-to-noise ratio of the interferogram is high; however, at low signal-to-noise ratios, QGPU suffers from global noise error propagation during unwrapping, leading to inconsistent unwrapped phases or even unwrapping failure; while ILS unwrapped phases are roughly continuous, but its noise error range is relatively large. Summary of the Invention

[0003] This invention aims to provide a phase unwrapping method based on a global attention upsampling network. The method proposes a phase unwrapping algorithm based on global attention upsampling, which effectively integrates the PU-M-Net network and the attention mechanism. It achieves one-step phase unwrapping by establishing a nonlinear mapping relationship between the wrapped phase and the unwrapped phase, exhibiting high unwrapping efficiency and robustness.

[0004] The technical solution of the present invention is as follows:

[0005] The phase unwrapping method based on a global attention upsampling network includes the following steps:

[0006] A. Construct a deep neural network structure, the specific structure of which is as follows:

[0007] Encoding networks, decoding networks;

[0008] The encoding network consists of five sequentially connected 2×2 max pooling layers and six PU-M-Net modules;

[0009] The decoding network comprises five GAU modules, five PU-M-Net modules, and stacked module I;

[0010] B. Construct multiple training sets with different noise levels, and train the deep neural network to obtain an unwrapped network suitable for unwrapping interferograms with different noise levels;

[0011] C. Determine the noise level of the interferogram to be unwrapped, and then match it to the unwrapping network with the corresponding noise level for unwrapping. The specific process is as follows:

[0012] a. The interferogram to be unwrapped is input into the encoding network and processed by five 2×2 max pooling layers in sequence to obtain five max pooling images.

[0013] The interferogram to be unwrapped is processed by the first PU-M-Net module, and the output of the first PU-M-Net is then input into the decoding network.

[0014] The first max-pooled image is stacked with the first output PU-M-Net, then processed by the second PU-M-Net module to obtain the second PU-M-Net output, which is then input into the decoding network.

[0015] The second max-pooled image, the output of the second PU-M-Net, and the PU-M-Net stacked together are then input into the third PU-M-Net module for processing, resulting in the output of the third PU-M-Net, which is then input into the decoding network.

[0016] The third max-pooled image and the output of the third PU-M-Net are stacked with the PU-M-Net, then processed by the fourth PU-M-Net module to obtain the output of the fourth PU-M-Net, which is then input into the decoding network.

[0017] The fourth max-pooling image and the output of the fourth PU-M-Net are stacked with the PU-M-Net, then processed by the fifth PU-M-Net module to obtain the output of the fifth PU-M-Net, which is then input into the decoding network.

[0018] The fifth max-pooling image and the output of the fifth PU-M-Net are stacked with the PU-M-Net, then processed by the sixth PU-M-Net module to obtain the output of the sixth PU-M-Net, which is then input into the decoding network.

[0019] b. In the decoding network, the output of the sixth PU-M-Net is used as a high-level feature input to the fifth PU-M-Net module, and the output of the sixth PU-M-Net is input into the stacking module I after being processed above.

[0020] The output of the fifth PU-M-Net is used as a low-level feature input to the fifth GAU module. After the fifth GAU module processes the high-level and low-level features, it is input into the fifth PU-M-Net module for further processing, resulting in output a. Output a is used as a high-level feature input to the fourth PU-M-Net module. At the same time, output a is upsampled and input into stacked module I.

[0021] The output of the fourth PU-M-Net is used as a low-level feature input to the fourth GAU module. After the fourth GAU module processes the high-level and low-level features, it is input into the fourth PU-M-Net module for further processing, resulting in output b. Output b is used as a high-level feature input to the third PU-M-Net module. At the same time, output b is upsampled and input into stacked module I.

[0022] The output of the third PU-M-Net is used as a low-level feature input to the third GAU module. After the third GAU module processes the high-level and low-level features, it is input into the third PU-M-Net module for further processing, resulting in output c. Output c is used as a high-level feature input to the second PU-M-Net module. At the same time, output c is upsampled and input into stacked module I.

[0023] The output of the second PU-M-Net is used as a low-level feature input to the second GAU module. After the second GAU module processes the high-level and low-level features, it is input into the second PU-M-Net module for further processing, resulting in the output d. The output d is then used as a high-level feature input to the first PU-M-Net module. Simultaneously, the output d is upsampled and input into the stacked module I.

[0024] The output of the first PU-M-Net is used as a low-level feature input to the first GAU module. After the first GAU module processes the high-level and low-level features, it is input into the first PU-M-Net module for further processing to obtain the output result e. The output result e is then input into the stacked module I.

[0025] The output results a, b, c, and d, after upsampling, have the same resolution as output result e.

[0026] In stacking module I, output results a, b, c, d, and e are stacked to obtain the final phase unwrapping result.

[0027] In step B, the specific process of constructing multiple training sets with different noise levels is as follows;

[0028] Based on the phase noise of the interferogram, it is divided into 4 levels according to the range of noise estimates:

[0029]

[0030] Based on the noise level, four training sets with different noise levels are constructed to train the deep neural network.

[0031] The method for calculating the noise estimate Ne is as follows:

[0032]

[0033] Where M and N are the number of rows and columns of the interferogram, respectively, Res(x,y) is the residual map after the modulo operation, l is the noise weighting coefficient, and W(x,y) is the interferogram quality map obtained using the pseudo-coherence coefficient method.

[0034] In the encoding and decoding networks, each PU-M-Net module has the same structure, including C module I, R module, stacked module II, and C module II connected in sequence. The input result is processed by C module I and R module in sequence. The result processed by R module and the input result are stacked and then processed by C module II to obtain the output result of the PU-M-Net module.

[0035] The C module I and C module II have the same structure, each including a 3×3 convolutional layer, batch normalization (BN), and ReLU activation function. The input result is processed sequentially through the 3×3 convolutional layer, BN, and ReLU activation function to obtain the output result.

[0036] The R module includes a 3×3 convolutional layer, batch normalization (BN), and ReLU activation function. The input result is processed sequentially through a 3×3 convolutional layer, batch normalization (BN), ReLU activation function, and a 3×3 convolutional layer, and then batch normalization (BN). The processed result is added to the input result and fused, and then processed by the ReLU activation function to obtain the output result.

[0037] The processing procedure in the GAU module is as follows:

[0038] The high-level feature input is first processed by global average pooling to obtain the average pooling result. The average pooling result is then passed through a 1×1 convolutional layer to reduce the number of channels in the feature map to 64 / r. Then, a 1×1 convolutional layer is used to increase the dimensionality of the obtained feature map so that the number of output channels is consistent with the number of low-level feature channels. The output channel is then multiplied by the low-level feature to obtain a weighted low-level feature. The weighted low-level feature is then added to the average pooling result and fused to obtain the output result.

[0039] The method of this invention employs upsampling between the encoding and decoding paths, connecting the low-level feature layer with its corresponding high-level feature layer. The GAU module can effectively utilize multi-scale feature mapping and use high-level features to provide weighted guidance for low-level features, thereby further enhancing the interferogram fringe feature information.

[0040] The PU-M-Net of this invention adopts simulated skip connection operation, which promotes the effective fusion of phase detail information and semantic fringe information and improves the utilization rate of initial features in phase unwrapping.

[0041] The method of this invention divides the noisy interferogram into four noise levels according to the interferogram noise level assessment system, constructs datasets of four noise levels to train the network model, and the trained network can directly obtain the unwrapped phase from the entangled phase at different noise levels, thereby achieving high unwrapping efficiency and robustness, and has good application prospects. Attached Figure Description

[0042] Figure 1 This is a schematic diagram of the untangling process in Embodiment 1 of the present invention;

[0043] Figure 2 This is a schematic diagram of the overall structure of the deep neural network in Embodiment 1 of the present invention;

[0044] Figure 3 This is a schematic diagram of the structure of the GAU module in Embodiment 1 of the present invention;

[0045] Figure 4 This is a simulated interferogram of Embodiment 2 of the present invention;

[0046] Figure 5 For different phase unwrapping methods Figure 4 (e) Unwrapping results; the rows from top to bottom show the unwrapping results of the "QGPU method", "ILS method", and "GAUPU method"; the columns from left to right show the interferogram unwrapping phase, unwrapping phase error map, and unwrapping phase error histogram.

[0047] Figure 6 For different phase unwrapping methods Figure 4 (f) Unwrapping results; the rows from top to bottom show the unwrapping results of the "QGPU method", "ILS method", and "GAUPU method"; the columns from left to right show the interferogram unwrapping phase, unwrapping phase error map, and unwrapping phase error histogram.

[0048] Figure 7 For different phase unwrapping methods Figure 4 (g) Unwrapping results; the row directions from top to bottom are the unwrapping results of "QGPU method", "ILS method" and "GAUPU method"; the column directions from left to right are the interferogram unwrapping phase, unwrapping phase error map and unwrapping phase error histogram;

[0049] Figure 8 For different phase unwrapping methods Figure 4 (h) Unwrapping results; the rows from top to bottom are the unwrapping results of the "QGPU method", "ILS method" and "GAUPU method"; the columns from left to right are the interferogram unwrapping phase, unwrapping phase error map and unwrapping phase error histogram;

[0050] Figure 9 The measured interferogram is shown in Example 3.

[0051] Figure 10 For different untangling methods Figure 9 (a) Unwrapping results; the rows from top to bottom are the interferogram unwrapped phase and the unwrapped phase rewrapped phase; the columns from left to right are the unwrapping results of the "QGPU method", "ILS method" and "GAUPU method";

[0052] Figure 11 For different untangling methods Figure 9 (b) Unwrapping results; the rows from top to bottom are the interferogram unwrapped phase and the unwrapped phase rewrapped phase; the columns from left to right are the unwrapping results of the "QGPU method", "ILS method" and "GAUPU method";

[0053] Figure 12 For different untangling methods Figure 9 (c) Unwrapping results; the rows from top to bottom are the interferogram unwrapped phase and the unwrapped phase rewrapped phase; the columns from left to right are the unwrapping results of the "QGPU method", "ILS method" and "GAUPU method";

[0054] Figure 13 For different untangling methods Figure 9 (d) Unwrapping results; the row direction from top to bottom shows the unwrapped phase of the interferogram and the rewrapped phase of the unwrapped phase; the column direction from left to right shows the unwrapping results of the "QGPU method", "ILS method" and "GAUPU method". Detailed Implementation

[0055] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0056] Example 1

[0057] The phase unwrapping method based on a global attention upsampling network includes the following steps:

[0058] The phase unwrapping method based on a global attention upsampling network includes the following steps:

[0059] A. Constructing as follows Figure 2 and 3 The deep neural network structure shown is as follows:

[0060] Encoding networks, decoding networks;

[0061] The encoding network consists of five sequentially connected 2×2 max pooling layers and six PU-M-Net modules;

[0062] The decoding network comprises five GAU modules, five PU-M-Net modules, and stacked module I;

[0063] In the encoding and decoding networks, each PU-M-Net module has the same structure, including C module I, R module, stacked module II, and C module II connected in sequence. The input result is processed by C module I and R module in sequence. The result processed by R module and the input result are stacked and then processed by C module II to obtain the output result of the PU-M-Net module.

[0064] The C module I and C module II have the same structure, each including a 3×3 convolutional layer, batch normalization (BN), and ReLU activation function. The input result is processed sequentially through the 3×3 convolutional layer, BN, and ReLU activation function to obtain the output result.

[0065] The R module includes a 3×3 convolutional layer, batch normalization (BN), ReLU activation function, and so on. The input result is processed sequentially through a 3×3 convolutional layer, batch normalization (BN), ReLU activation function, and so on. The processed result is then added to the input result and fused, and finally processed by the ReLU activation function to obtain the output result.

[0066] B. Construct multiple training sets with different noise levels, and train the deep neural network to obtain an unwrapped network suitable for unwrapping interferograms with different noise levels;

[0067] like Figure 1 (a) As shown in the diagram of GAUPU network training, the specific process of constructing multiple training sets with different noise levels is as follows;

[0068] Based on the phase noise of the interferogram, it is divided into 4 levels according to the range of noise estimates:

[0069]

[0070] The method for calculating the noise estimate Ne is as follows:

[0071]

[0072] Where M and N are the number of rows and columns of the interferogram, respectively; Res(x,y) is the residual image after the modulo operation; l is the noise weighting coefficient, which is set to 100 here; W(x,y) is the interferogram quality image obtained using the pseudo-coherence coefficient method;

[0073] Four training sets with different noise levels were constructed, and the same deep learning network architecture was trained to obtain unwrapping networks suitable for interferogram unwrapping at different noise levels. Each dataset contains 35,000 data points and was generated in the following four ways:

[0074] (I) 3000 sets of data were generated according to the dataset construction method described in the literature. The image size was 256×256 and the phase range of the label image was 0-60 radians.

[0075] (II) 4000 sets of data were generated according to the data set construction method described in the literature. The image size was 256×256 and the phase range of the label image was 0-60 radians.

[0076] (III) 24,000 sets of InSAR interferometric phase map data were converted from DEM data from Datong City, Shanxi Province, Huangshan City, Anhui Province, and Jinhua City, Zhejiang Province. That is, the DEM data was first converted into the true interferometric phase according to InSAR theory, and then the wrapped phase map with different noise was generated, resulting in a total of 24,000 sets of InSAR interferometric data. The image size is 256×256, and the phase range of the labeled image is 0-60 radians.

[0077] (IV) 3000 sets of InSAR interferometric phase map data obtained by the traditional phase unwrapping method, that is, the unwrapped phase of the measured InSAR interferogram is obtained by unwrapping the measured InSAR interferogram using the traditional phase unwrapping method, and then wrapped phase maps with different noise are generated, resulting in a total of 3000 sets of quasi-measured InSAR interferogram data. The image size is 256×256, and the phase range of the tag image is 0 to 60 radians.

[0078] For details on the construction of the above dataset, please refer to the references.

[0079] [1]ZHANG T,JIANG S,ZHAO Z,DIXIT K,ZHOU

[0080] [2]WANG K, Li Y, KEMAO Q, DI J, ZHAO J. One-step robust deep learning phase unwrapping[J]. Opt Express, 2019, 27(10): 15100-15115.

[0081] [3] Liang Feng, Xie Xianming, Xu Youmiao, Song Minghui, Zeng Qingning. An improved U-Net phase unwrapping method [J]. Remote Sensing Information, 2021, 36(05): 134-141.

[0082] [4]XIE XM,TIAN

[0083] C. For example Figure 1 (b) As shown in the diagram of GAUPU network unwrapping, the noise level of the interferogram to be unwrapped is determined, and then it is matched to the unwrapping network with the corresponding noise level for unwrapping. The specific process is as follows:

[0084] a. The interferogram to be unwrapped is input into the encoding network and processed by five 2×2 max pooling layers in sequence to obtain five max pooling images.

[0085] The interferogram to be unwrapped is processed by the first PU-M-Net module, and the output of the first PU-M-Net is then input into the decoding network.

[0086] The first max-pooled image is stacked with the first output PU-M-Net, then processed by the second PU-M-Net module to obtain the second PU-M-Net output, which is then input into the decoding network.

[0087] The second max-pooled image, the output of the second PU-M-Net, and the PU-M-Net stacked together are then input into the third PU-M-Net module for processing, resulting in the output of the third PU-M-Net, which is then input into the decoding network.

[0088] The third max-pooled image and the output of the third PU-M-Net are stacked with the PU-M-Net, then processed by the fourth PU-M-Net module to obtain the output of the fourth PU-M-Net, which is then input into the decoding network.

[0089] The fourth max-pooling image and the output of the fourth PU-M-Net are stacked with the PU-M-Net, then processed by the fifth PU-M-Net module to obtain the output of the fifth PU-M-Net, which is then input into the decoding network.

[0090] The fifth max-pooling image and the output of the fifth PU-M-Net are stacked with the PU-M-Net, then processed by the sixth PU-M-Net module to obtain the output of the sixth PU-M-Net, which is then input into the decoding network.

[0091] b. In the decoding network, the output of the sixth PU-M-Net is used as a high-level feature input to the fifth PU-M-Net module, and the output of the sixth PU-M-Net is input into the stacking module I after upsampling;

[0092] The output of the fifth PU-M-Net is used as a low-level feature input to the fifth GAU module. After the fifth GAU module processes the high-level and low-level features, it is input into the fifth PU-M-Net module for further processing, resulting in output a. Output a is used as a high-level feature input to the fourth PU-M-Net module. At the same time, output a is upsampled and input into stacked module I.

[0093] The output of the fourth PU-M-Net is used as a low-level feature input to the fourth GAU module. After the fourth GAU module processes the high-level and low-level features, it is input into the fourth PU-M-Net module for further processing, resulting in output b. Output b is used as a high-level feature input to the third PU-M-Net module. At the same time, output b is upsampled and input into stacked module I.

[0094] The output of the third PU-M-Net is used as a low-level feature input to the third GAU module. After the third GAU module processes the high-level and low-level features, it is input into the third PU-M-Net module for further processing, resulting in output c. Output c is used as a high-level feature input to the second PU-M-Net module. At the same time, output c is upsampled and input into stacked module I.

[0095] The output of the second PU-M-Net is used as a low-level feature input to the second GAU module. After the second GAU module processes the high-level and low-level features, it is input into the second PU-M-Net module for further processing, resulting in the output d. The output d is then used as a high-level feature input to the first PU-M-Net module. Simultaneously, the output d is upsampled and input into the stacked module I.

[0096] The output of the first PU-M-Net is used as a low-level feature input to the first GAU module. After the first GAU module processes the high-level and low-level features, it is input into the first PU-M-Net module for further processing to obtain the output result e. The output result e is then input into the stacked module I.

[0097] The output results a, b, c, and d, after upsampling, have the same resolution as output result e.

[0098] In stacking module I, output results a, b, c, d, and e are stacked to obtain the final phase unwrapping result.

[0099] The processing procedure in the GAU module is as follows:

[0100] The high-level feature input is first processed by global average pooling to obtain the average pooling result. The average pooling result is then passed through a 1×1 convolutional layer to reduce the number of channels in the feature map to 64 / r. Then, a 1×1 convolutional layer is used to increase the dimensionality of the obtained feature map so that the number of output channels is consistent with the number of low-level feature channels. The output channel is then multiplied by the low-level feature to obtain a weighted low-level feature. The weighted low-level feature is then added to the average pooling result and fused to obtain the output result.

[0101] Example 2: Simulated Interferogram Phase Unwrapping Experiment

[0102] To verify the performance of the network architecture of this invention, the network algorithm was used to unwrap different types of simulated interferograms, and compared with traditional phase unwrapping algorithms such as the Quality Guided GPU (QGPU) and Iterative Least Squares (ILS). Simulated interferograms are shown below. Figure 4 As shown, Figure 4 (ad) represents the true phase of multi-peak terrain, a circle, a pyramid, and a simulated mountain, respectively, all with a resolution of 256×256. Figure 4 (eh) is Figure 4 (ad) The noise wrapping diagrams corresponding to the true phase have signal-to-noise ratios of 4.94dB, 1.42dB, -1.07dB, and -1.07dB, respectively, and the corresponding noise levels are 1, 2, 3, and 3, respectively.

[0103] Using QGPU, ILS, and the method of Embodiment 1 of this invention respectively, Figure 4 (eh) is used for untangling, and the untangling results are as follows: Figures 5-8 , Figures 5-8 The left, middle, and right columns correspond to the algorithms described above. Figure 4 The unwrapped phase diagram, unwrapped phase error diagram, and unwrapped phase error histogram of (eh). Figures 5-8It can be seen that both traditional algorithms, QGPU and ILS, achieve good unwrapping results when the signal-to-noise ratio of the interferogram is high. However, under low signal-to-noise ratio conditions, QGPU suffers from global noise propagation during unwrapping, leading to inconsistent unwrapping phases and even unwrapping failure. ILS achieves roughly continuous unwrapping phases, but its noise error range is relatively large. Meanwhile, the method of this invention achieves more ideal unwrapping results with a smaller noise error range. Especially when the noise content of the interferogram is high, its error range is much smaller than the two methods mentioned above, indicating that this method has good noise robustness while being able to reconstruct the original interferogram phase information relatively completely. The QGPU and ILS methods are run in MATLAB R2016b, while the method of this invention is run in Python 3.8.

[0104] Example 3: Experimental Interferogram Phase Unwrapping

[0105] Measured interferogram as follows Figure 9 As shown, where Figure 9 (a) is a local interferogram of the Three Gorges Dam. Figure 9 (b) A partial interferogram of Yunnan Province acquired by the Sentinel 1A satellite at 21:48 on May 21, 2021, in Yangbi County, Dali Prefecture, Yunnan Province (25.67°N, 99.87°E). Figure 9 (c) is a MEMS interferogram. Figure 9 (d) is the Rahul interferogram.

[0106] Table 1. Measured noise levels of interferograms

[0107]

[0108] Table 1 above is Figure 9 The noise level of the interferogram shown in (ad) Figures 10-13 For QGPU and ILS, the method of this invention respectively targets... Figure 9 (ad) The result of untangling, where Figures 10-13 Each row represents QGPU, ILS, the unwrapping phase of the method of this invention, and the corresponding rewrapping diagram. From Figures 10-13 In the left column, it can be observed that there are obvious phase discontinuities in the unwrapped phase diagram of the QGPU method. The unwrapped phase rewrapping diagram is roughly consistent with the measured interferogram fringes, but it still has a lot of noise, which can easily cause noise error to spread globally, leading to a decrease in phase unwrapping accuracy or unreliable unwrapping results, such as the unwrapped phase diagram of MEMS interferogram and the unwrapped phase diagram of Rahul interferogram. Figures 10-13The unwrapped phase obtained by ILS in the middle column is relatively smooth and continuous, but the unwrapped phase rewound pattern has obvious inconsistent fringes compared with the measured interferogram, resulting in the loss of fringe detail information and reducing the reliability of its unwrapped phase. Examples include the unwrapped phase rewound patterns of the local Yunnan interferogram, MEMS interferogram, and Rahul interferogram. Compared with the above two types of algorithms, the unwrapped phase obtained by the method of this invention is continuous and smooth, and the fringes of its unwrapped phase rewound pattern are basically consistent with the fringes of the measured interferogram. Moreover, there is almost no noise in the rewound pattern, indicating that the algorithm effectively suppresses the influence of noise while achieving better unwrapping results.

[0109] In summary, the method of this invention combines attention upsampling GAU and PU-M-Net, and achieves interferogram unwrapping by constructing a mapping relationship between the wrapped phase and the true phase. The trained network can unwrap interferograms with different noise levels. The simulated interferogram in Example 2 and the measured interferogram phase unwrapping experiments in Example 3 demonstrate the effectiveness of the algorithm of this invention. Compared with traditional algorithms such as QGPU and ILS, it achieves more robust phase unwrapping results while consuming significantly less time.

Claims

1. A phase unwrapping method based on a global attention upsampling network, characterized in that, Includes the following steps: A. Construct a deep neural network structure, the specific structure of which is as follows: Encoding networks, decoding networks; The encoding network consists of five sequentially connected 2×2 max pooling layers and six PU-M-Net modules; The decoding network comprises five GAU modules, five PU-M-Net modules, and stacked module I; B. Construct multiple training sets with different noise levels, and train the deep neural network to obtain an unwrapped network suitable for unwrapping interferograms with different noise levels; C. Determine the noise level of the interferogram to be unwrapped, and then match it to the unwrapping network with the corresponding noise level for unwrapping. The specific process is as follows: a. The interferogram to be unwrapped is input into the encoding network and processed by five 2×2 max pooling layers in sequence to obtain five max pooling images. The interferogram to be unwrapped is processed by the first PU-M-Net module, and the output of the first PU-M-Net is then input into the decoding network. The first max-pooled image is stacked with the first output PU-M-Net, then processed by the second PU-M-Net module to obtain the second PU-M-Net output, which is then input into the decoding network. The second max-pooled image, the output of the second PU-M-Net, and the PU-M-Net stacked together are then input into the third PU-M-Net module for processing, resulting in the output of the third PU-M-Net, which is then input into the decoding network. The third max-pooled image and the output of the third PU-M-Net are stacked with the PU-M-Net, then processed by the fourth PU-M-Net module to obtain the output of the fourth PU-M-Net, which is then input into the decoding network. The fourth max-pooling image and the output of the fourth PU-M-Net are stacked with the PU-M-Net, then processed by the fifth PU-M-Net module to obtain the output of the fifth PU-M-Net, which is then input into the decoding network. The fifth max-pooling image and the output of the fifth PU-M-Net are stacked with the PU-M-Net, then processed by the sixth PU-M-Net module to obtain the output of the sixth PU-M-Net, which is then input into the decoding network. b. In the decoding network, the output of the sixth PU-M-Net is used as a high-level feature input to the fifth PU-M-Net module, and the output of the sixth PU-M-Net is input into the stacking module I after being processed above. The output of the fifth PU-M-Net is used as a low-level feature input to the fifth GAU module. After the fifth GAU module processes the high-level and low-level features, it is input into the fifth PU-M-Net module for further processing, resulting in output a. Output a is used as a high-level feature input to the fourth PU-M-Net module. At the same time, output a is upsampled and input into stacked module I. The output of the fourth PU-M-Net is used as a low-level feature input to the fourth GAU module. After the fourth GAU module processes the high-level and low-level features, it is input into the fourth PU-M-Net module for further processing, resulting in output b. Output b is used as a high-level feature input to the third PU-M-Net module. At the same time, output b is upsampled and input into stacked module I. The output of the third PU-M-Net is used as a low-level feature input to the third GAU module. After the third GAU module processes the high-level and low-level features, it is input into the third PU-M-Net module for further processing, resulting in output c. Output c is used as a high-level feature input to the second PU-M-Net module. At the same time, output c is upsampled and input into stacked module I. The output of the second PU-M-Net is used as a low-level feature input to the second GAU module. After the second GAU module processes the high-level and low-level features, it is input into the second PU-M-Net module for further processing, resulting in the output d. The output d is then used as a high-level feature input to the first PU-M-Net module. Simultaneously, the output d is upsampled and input into the stacked module I. The output of the first PU-M-Net is used as a low-level feature input to the first GAU module. After the first GAU module processes the high-level and low-level features, it is input into the first PU-M-Net module for further processing to obtain the output result e. The output result e is then input into the stacked module I. The output results a, b, c, and d, after upsampling, have the same resolution as output result e. In stacking module I, output results a, b, c, d, and e are stacked to obtain the final phase unwrapping result.

2. The phase unwrapping method based on a global attention upsampling network as described in claim 1, characterized in that: In step B, the specific process of constructing multiple training sets with different noise levels is as follows; Based on the phase noise of the interferogram, it is divided into 4 levels according to the range of noise estimates: Based on the noise level, four training sets with different noise levels are constructed to train the deep neural network.

3. The phase unwrapping method based on a global attention upsampling network as described in claim 2, characterized in that: The noise estimate Ne The calculation method is as follows: (1) in M and N Here, represents the number of rows and columns of the interferogram, Res(x,y) is the residual plot after modulo operation, and l is the noise weighting coefficient. W(x,y) This is the quality map of the interferogram obtained using the pseudo-coherence coefficient method.

4. The phase unwrapping method based on a global attention upsampling network as described in claim 1, characterized in that: In the encoding and decoding networks, each PU-M-Net module has the same structure, including C module I, R module, stacked module II, and C module II connected in sequence. The input result is processed by C module I and R module in sequence. The result processed by R module and the input result are stacked and then processed by C module II to obtain the output result of the PU-M-Net module.

5. The phase unwrapping method based on a global attention upsampling network as described in claim 4, characterized in that: The C module I and C module II have the same structure, each including a 3×3 convolutional layer, batch normalization (BN), and ReLU activation function. The input result is processed sequentially through the 3×3 convolutional layer, BN, and ReLU activation function to obtain the output result.

6. The phase unwrapping method based on a global attention upsampling network as described in claim 4, characterized in that: The R module includes a 3×3 convolutional layer, batch normalization (BN), ReLU activation function, and a 3×3 convolutional layer, batch normalization (BN), and ReLU activation function. The input result is processed sequentially through a 3×3 convolutional layer, batch normalization (BN), ReLU activation function, 3×3 convolutional layer, and batch normalization (BN). The processed result is added to the input result and fused. Then, it is processed by the ReLU activation function to obtain the output result.

7. The phase unwrapping method based on a global attention upsampling network as described in claim 1, characterized in that: The processing procedure in the GAU module is as follows: The high-level feature input is first processed by global average pooling to obtain the average pooling result. The average pooling result is then passed through a 1×1 convolutional layer to reduce the number of channels in the feature map to 64 / r. Then, a 1×1 convolutional layer is used to increase the dimensionality of the obtained feature map so that the number of output channels is consistent with the number of low-level feature channels. The output channel is then multiplied by the low-level feature to obtain a weighted low-level feature. The weighted low-level feature is then added to the average pooling result and fused to obtain the output result.

Images