Interferometric phase unwrapping method based on double-branch decoding residual network

Abstract

The invention discloses an interferometric phase unwrapping method based on a double-branch decoding residual network, and the method comprises the following steps: S1, constructing an interferometric phase data set which comprises an interferometric phase diagram of pixel phase wrapping and a corresponding real interferometric phase diagram; s2, constructing a double-branch decoding residual network, wherein the double-branch decoding residual network comprises an input layer, an encoder network, a double-branch decoder network comprising a first branch decoder module and a second branch decoder module, and an output layer which are connected in sequence; training the double-branch decoding residual network by using the interferometric phase data set to obtain a double-branch decoding residual network interferometric phase unwrapping model and the weight of the model; and S3, inputting the interferometric phase diagram subjected to pixel phase winding into a double-branch decoding residual network interferometric phase unwrapping model to obtain a predicted interferometric unwrapping phase diagram. According to the method, the interferometric phase unwrapping under the condition of low signal-to-noise ratio is realized, and the interferometric measurement precision of the synthetic aperture radar is improved.

Description

Technical Field

[0001] The invention belongs to the field of synthetic aperture radar interferometry, and in particular relates to an interferometric phase unwrapping method based on a double-branch decoding residual network. Background Art

[0002] InSAR (Interferometric Synthetic Aperture Radar) can be widely used in the production of DEM (Digital Elevation Model) under all-weather conditions. Phase unwrapping is one of the core steps of interferometric synthetic aperture radar technology, and the accuracy of phase unwrapping will directly affect the accuracy of DEM. Phase unwrapping aims to recover the absolute phase from interferometric observations, and accurate phase unwrapping is essential for applications such as three-dimensional terrain reconstruction and surface deformation monitoring. However, practical challenges such as decorrelation noise, terrain variability, and algorithm limitations can significantly reduce the unwrapping accuracy. These challenges are particularly prominent in areas with low coherence and steep gradients, where traditional methods often fail, limiting the reliability of InSAR-derived measurements.

[0003] Traditional phase unwrapping methods can be divided into three categories: path tracking method, minimum norm method and Bayesian method. The path tracking method uses prior information to formulate a suitable integral path. It has the advantage of fast speed, but the method is sensitive to residual points. The minimum norm method seeks the global optimal solution, so it will lead to low local accuracy and also has the disadvantage of large amount of calculation. The Bayesian method is based on the maximum a posteriori estimation algorithm and uses the Markov random field statistical distribution model to describe the prior distribution of the target height. It can make more accurate estimates, but the calculation complexity is high. Since timeliness and high precision are required in the application of interferometric phase unwrapping, it is of great significance to explore and develop a new phase unwrapping method.

[0004] In recent years, with the rise of artificial intelligence technology, deep learning, as an important branch of artificial intelligence, has been widely used in various fields. In the field of interferometric phase unwrapping, there are existing deep learning-based networks such as U-net (U-type convolutional neural network), PhaseNet (phase unwrapping neural network), PUnet (phase unwrapping neural network) and PSPNet (phase unwrapping neural network) DCNN models that improve the efficiency and accuracy of phase unwrapping, but they still face problems such as incomplete multi-scale feature learning, high feature redundancy, and strong data set dependence. These problems undermine the effectiveness of these networks in areas with low coherence and high gradient deformation. Therefore, for interferometric phase unwrapping using neural networks, how to solve these problems and achieve an interferometric phase unwrapping method that reduces dependence on data sets, has high timeliness, good robustness and high unwrapping accuracy is exactly the problem that needs to be solved at present. Summary of the invention

[0005] An interference phase unwrapping method based on a double-branch decoding residual network comprises the following steps: S1. Construct an interference phase data set, including an interference phase map of pixel phase entanglement and its corresponding real interference phase map; S2, constructing a dual-branch decoding residual network, including a sequentially connected input layer, an encoder network, a dual-branch decoder network including a first branch decoder module and a second branch decoder module, and an output layer; The interference phase data set is used to train a two-branch decoding residual network to obtain a two-branch decoding residual network interference phase unwrapping model and the weight of the model; S3. Input the interferometric phase image with pixel phase entanglement into the dual-branch decoding residual network interferometric phase unwrapping model to obtain the predicted interferometric unwrapping phase image.

[0006] Furthermore, in the aforementioned step S1, constructing the interferometric phase data set includes the following sub-steps: S1.1, creating random-shaped interference phase images by adding and subtracting several Gaussian functions of different shapes and positions; S1.2, randomly select ramps and add them to the interference phase diagram in vertical and horizontal directions to form an interference phase diagram with ramp phase; S1.3, the phase of the interference phase pattern with the ramp phase is pixel-wound, and the phase is wound The calculation formula is as follows: , Among them, exp is a natural constant e The exponential function with base , is the original phase of the interference phase image pixel, is the spatial coordinate of the pixel in the phase image, To find the angle sign.

[0007] Furthermore, the encoder network in the aforementioned step S2 includes a convolutional layer with a step size of 2, a maximum pooling layer with a pooling size of 3 and a step size of 2, and sixteen residual modules connected in sequence; the sixteenth residual module is connected to the dual-branch decoder network as the output end of the encoder network, and the structure of the residual module is set according to the input step size, as follows: When the stride is not 1, the input layer of the residual module is connected to a 1x1 convolution layer with the same stride and channel as the input and a 3x3 convolution layer with the same stride and channel as the input, the branch connecting the 3x3 convolution layer is sequentially connected to a batch normalization and linear activation function combination layer, a 3x3 convolution layer and a batch normalization layer as the first input end of the additive fusion, the branch connecting the 1x1 convolution layer is connected to a batch normalization layer as the second input end of the additive fusion, and the output of the additive fusion is output as the residual module result after a linear activation. When the stride is 1, the input layer of the residual module is respectively connected to an additive fusion layer and a 3x3 convolution layer with the same stride and channel as the input, the branch connecting the 3x3 convolution layer is sequentially connected to a batch normalization and linear activation function combination layer, a 3x3 convolution layer and a batch normalization layer as the other input end of the additive fusion, and the output of the additive fusion is output as the residual module result after a linear activation.

[0008] Further, the aforementioned first branch decoder module includes: an input layer, three convolutional layers with a stride of 1x1, five dilated convolutional layers, nine batch normalization and linear activation combination layers, three 1x1 convolutional layers with a stride of 2, a global average pooling layer, two deformation layers, two fully connected layers, a sigmoid function layer, two multiplication fusion layers, four upsampling layers, an addition fusion layer, a connection layer, a 3x3 convolutional layer with a stride of 1, a linear activation layer, and two output layers; The first branch of the input layer of the first branch decoder module sequentially connects the first 1x1 convolution layer with a stride of 1, the first batch normalization and linear activation combination layer, the output of the first batch normalization and linear activation combination layer is used as the first input of the connection layer, and the first 1x1 convolution layer with a stride of 2, the output of the fifth batch normalization and linear activation combination layer are sequentially connected as the first input of the second multiplication fusion layer; The second branch of the input layer of the first branch decoder module sequentially connects the first dilated convolution layer, the second batch normalization and linear activation combination layer. The second batch normalization and linear activation combination layer is used as the second input of the connection layer, and the second 1x1 convolution layer with a stride of 2, the fourth dilated convolution layer, the sixth batch normalization and linear activation combination layer are sequentially connected. The sixth batch normalization and linear activation combination layer is used as the first input of the output of the first additive fusion layer. The third branch of the input layer of the first branch decoder module is sequentially connected to the second dilated convolution layer, the third batch normalization and linear activation combination layer, the third batch normalization and linear activation combination layer is used as the third input end of the connection layer, and is sequentially connected to the third 1x1 convolution layer with a stride of 2 and the fifth dilated convolution layer, and the fifth dilated convolution layer is used as the second input end of the output end of the first additive fusion layer; The fourth branch of the input layer of the first branch decoder module is sequentially connected to the third dilated convolution layer, the fourth batch normalization and linear activation combination layer, and the fourth batch normalization and linear activation combination layer serves as the fourth input terminal of the connection layer; The fifth branch of the input layer of the first branch decoder module is sequentially connected to the second 1x1 convolution layer with a step size of 1, the first global average pooling layer, the first deformation layer, and the first upsampling layer, and the output end of the first deformation layer is connected to the input end of the first fully connected layer, and the branch of the first upsampling layer serves as the fifth input end of the connection layer; after the first fully connected layer, a linear activation function, a second fully connected layer, a sigmoid function layer, and a second deformation layer are sequentially connected; the output end of the second upsampling layer serves as the second input end of the first multiplication fusion layer; The connection layer is followed by the third 1x1 convolutional layer with a stride of 1, the eighth batch normalization and linear activation combination layer, the fourth upsampling layer, and the first output of the first branch decoder module. The output of the first additive fusion layer is sequentially connected to a 3x3 convolution layer with a stride of 1, the seventh batch normalization and linear activation combination layer, and the output of the seventh batch normalization and linear activation combination layer is used as the first input of the first multiplicative fusion layer; the output of the first multiplicative fusion layer is used as the second input of the second multiplicative fusion layer, and after the second multiplicative fusion layer, it is sequentially connected to the third upsampling layer, the ninth batch normalization and linear activation function combination layer, and the second output of the first branch decoder module.

[0009] Furthermore, the aforementioned second branch decoder module includes: an input layer, four average pooling layers, five 1x1 convolutional layers, six batch normalization and linear activation combination layers, eight upsampling layers, three fully connected layers, a global average pooling layer, two deformation layers, a linear activation layer, two sigmoid activation function layers, two connection layers, a multiplication fusion layer and an output layer; The input layer of the second branch decoder module is connected to a global average pooling layer and the first upsampling layer respectively. The branch connected to the global average pooling layer is sequentially connected to the first fully connected layer, a linear activation layer, the second fully connected layer, the first sigmoid activation function layer, the first deformation layer, the second upsampling layer, and then serves as the first input end of the multiplication fusion layer. The branches connected to the first upsampling layer are respectively connected to the second connection layer, the first average pooling layer, the second average pooling layer, the third average pooling layer, and the fourth average pooling layer; The branch connected to the first average pooling layer is connected to the first 1x1 convolution layer, the first batch normalization and linear activation combination layer, and the third upsampling layer in sequence. The output of the third upsampling layer is used as the first input of the first connection layer. The branch connected to the second average pooling layer is connected to the second 1x1 convolution layer, the second batch normalization and linear activation combination layer, and the fourth upsampling layer in sequence. The output of the fourth upsampling layer serves as the second input of the first connection layer. The branch connected to the third average pooling layer is connected to the third 1x1 convolution layer, the third batch normalization and linear activation combination layer, and the fifth upsampling layer in sequence. The output of the fifth upsampling layer serves as the third input of the first connection layer. The branch connected to the fourth average pooling layer is connected to the fourth 1x1 convolution layer, the fourth batch normalization and linear activation combination layer, and the sixth upsampling layer in sequence. The output of the sixth upsampling layer serves as the fourth input of the first connection layer. The output of the first connection layer is connected to the input of the second connection layer and the input of the third fully connected layer; The third fully connected layer is sequentially connected to the second sigmoid activation function layer, the second deformation layer, and the seventh upsampling layer. The output of the seventh upsampling layer is used as the second input of the multiplication fusion. The output of the multiplication fusion is sequentially connected to the fifth batch normalization linear activation combination layer and the eighth upsampling layer. The output of the eighth upsampling layer is used as the third input of the second connection layer. The output of the second connection layer is sequentially connected to the fifth 1x1 convolutional layer, the sixth batch normalization and linear activation combination layer, and then to the output layer of the second branch decoder module.

[0010] Furthermore, in the aforementioned step S2, training the dual-branch decoding residual network includes the following sub-steps: S2.1, set the training starting learning rate, maximum learning rate, training batch size and number of training rounds; S2.2, during the training process, L2 norm regularization is used to prevent the network from overfitting; S2.3. Use the Adam gradient optimization algorithm to optimize the network training. The following function is used as the loss function during the optimization process. The calculation formula is as follows: , In the formula, is the difference between the true value of the training data set sample and the estimated value of the network training, , represents the phase pixel coordinate, b is a positive coefficient; S2.4. Return to execute steps S2.2 to S2.3 until the network converges, obtain the dual-branch decoding residual network interference phase unwrapping model and weights, and save the network training weights.

[0011] Furthermore, the aforementioned interference phase unwrapping method based on a two-branch decoding residual network also includes verifying the two-branch decoding residual network interference phase unwrapping model, specifically: sending the verification data set to the trained model, continuing the verification, and obtaining the interference phase unwrapping phase map.

[0012] Furthermore, the aforementioned interference phase unwrapping method based on a dual-branch decoding residual network also includes evaluating the phase unwrapping accuracy using the normalized root mean square error NRMES. The smaller the NRMES value, the higher the phase unwrapping accuracy and the better the network performance. The calculation formula is as follows: , Among them, NRMES is the normalized root mean square error, N is the training data set sample, is the difference between the true value of sample k in the training data set and its network training estimate, Represents the phase pixel coordinates.

[0013] Compared with the prior art, the beneficial technical effects of the above technical solutions adopted in the present invention are as follows: The present invention proposes an interference phase unwrapping method based on a dual-branch decoding residual network, which includes the creation of an interference phase data set, the use of residual encoder network design, the design of a dual-branch decoding module, the construction, training and verification of a dual-branch decoding network, etc., to achieve interference phase unwrapping under low signal-to-noise ratio conditions, improve the accuracy of synthetic aperture radar interferometry measurement, and is an interference phase unwrapping method with fast convergence speed, strong real-time performance and strong robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 It is a process flow chart of the present invention.

[0015] Figure 2 It is based on the structure diagram of the dual-branch decoding residual network.

[0016] Figure 3 This is the structural diagram of the first branch decoding module.

[0017] Figure 4This is a structural diagram of the second branch decoding module.

[0018] Figure 5 This is the interference winding phase diagram of the present invention under noise.

[0019] Figure 6 It is the real interference phase diagram corresponding to the interference winding phase diagram under noise of the present invention.

[0020] Figure 7 It is the interference unwrapping phase diagram predicted by the present invention through the network. DETAILED DESCRIPTION

[0021] In order to better understand the technical content of the present invention, specific embodiments are given and described as follows in conjunction with the accompanying drawings.

[0022] Various aspects of the invention are described herein with reference to the accompanying drawings, in which many illustrative embodiments are shown. The embodiments of the invention are not limited to those described in the accompanying drawings. It should be understood that the invention is implemented by any of the various concepts and embodiments described above, as well as the concepts and embodiments described in detail below, because the concepts and embodiments disclosed in the invention are not limited to any implementation. In addition, some aspects disclosed in the invention may be used alone or in any appropriate combination with other aspects disclosed in the invention.

[0023] like Figure 1 As shown, the present invention provides an interference phase unwrapping method based on a double-branch decoding residual network, comprising the following steps: S1. Construct an interference phase data set, including an interference phase map of pixel phase entanglement and its corresponding real interference phase map; S2, constructing a two-branch decoding residual network, including a sequentially connected input layer (Input), an encoder network (Encoder), a two-branch decoder network (Decoder) including a first branch decoder module and a second branch decoder module, and an output layer (Output); The interference phase data set is used to train a two-branch decoding residual network to obtain a two-branch decoding residual network interference phase unwrapping model and the weight of the model; S3. Input the interferometric phase image with pixel phase entanglement into the dual-branch decoding residual network interferometric phase unwrapping model to obtain the predicted interferometric unwrapping phase image.

[0024] As a preferred embodiment of the present invention, step S1 creates an interferometric phase image of irregular and arbitrary shapes by adding and subtracting several Gaussian functions of different shapes and positions, adds slopes along the vertical and horizontal directions and Gaussian additive noise to the interferometric phase image, and constructs an interferometric phase data set, as follows: S1.1. Create an interferometric phase image of random shape by adding and subtracting several Gaussian functions of different shapes and positions.

[0025] S1.2. Randomly select ramps in the vertical and horizontal directions and add them to the interference phase diagram to form an interference phase diagram with ramp phase.

[0026] S1.3, the phase of the interference phase image with the ramp phase is pixel-wound, and the phase is wound The calculation formula is as follows: , Among them, exp is a natural constant e The exponential function with base , is the original phase of the interference phase image pixel, is the spatial coordinate of the pixel in the phase image, To find the angle sign.

[0027] S1.4. According to the above method, create an interference phase image dataset with a data scale of 1400, and randomly divide the interference phase dataset into a training dataset, a validation dataset and a test dataset according to a ratio of 7:2:1. The pixel size of each interference phase image is 256´256, and the pixel value ranges from -55 to 55.

[0028] S1.5. Randomly assign 0dB, 3dB, 5dB and 10dB Gaussian additive noise to the interference phase image data in the data set to construct an interference phase data set.

[0029] refer to Figure 2 As a preferred embodiment of the present invention, the encoder network in step S2 includes a sequentially connected convolution layer with a step size of 2, a maximum pooling layer with a pooling size of 3 and a step size of 2, and sixteen residual modules; the sixteenth residual module is connected to the dual-branch decoder network as the output end of the encoder network, and the structure of the residual module is set according to the input step size, specifically as follows: When the stride is not 1, the input layer of the residual module is connected to a 1x1 convolution layer with the same stride and channel as the input and a 3x3 convolution layer with the same stride and channel as the input, the branch connecting the 3x3 convolution layer is connected to a batch normalization and linear activation function combination layer, a 3x3 convolution layer and a batch normalization layer in turn as the first input of the additive fusion, the branch connecting the 1x1 convolution layer is connected to a batch normalization layer and then as the second input of the additive fusion, and the output of the additive fusion is output as the residual module result after a linear activation. When the stride is 1, the input layer of the residual module is connected to an additive fusion layer and a 3x3 convolution layer with the same stride and channel as the input, the branch connecting the 3x3 convolution layer is connected to a batch normalization and linear activation function combination layer, a 3x3 convolution layer and a batch normalization layer in turn as the other input of the additive fusion, and the output of the additive fusion is output as the residual module result after a linear activation.

[0030] like Figure 3 As shown, as a preferred embodiment of the present invention, the first branch decoder module includes: an input layer, three convolutional layers with a step size of 1x1, five dilated convolutional layers, nine batch normalization and linear activation combination layers, three 1x1 convolutional layers with a step size of 2, a global average pooling layer, two deformation layers, two fully connected layers, a sigmoid function layer, two multiplication fusion layers, four upsampling layers, an addition fusion layer, a connection layer, a 3x3 convolutional layer with a step size of 1, a linear activation layer and two output layers; The first branch of the input layer of the first branch decoder module sequentially connects the first 1x1 convolution layer with a stride of 1, the first batch normalization and linear activation combination layer, the output of the first batch normalization and linear activation combination layer is used as the first input of the connection layer, and the first 1x1 convolution layer with a stride of 2, the output of the fifth batch normalization and linear activation combination layer are sequentially connected as the first input of the second multiplication fusion layer; The second branch of the input layer of the first branch decoder module sequentially connects the first dilated convolution layer, the second batch normalization and linear activation combination layer. The second batch normalization and linear activation combination layer is used as the second input of the connection layer, and the second 1x1 convolution layer with a stride of 2, the fourth dilated convolution layer, the sixth batch normalization and linear activation combination layer are sequentially connected. The sixth batch normalization and linear activation combination layer is used as the first input of the output of the first additive fusion layer. The third branch of the input layer of the first branch decoder module is sequentially connected to the second dilated convolution layer, the third batch normalization and linear activation combination layer, the third batch normalization and linear activation combination layer is used as the third input end of the connection layer, and is sequentially connected to the third 1x1 convolution layer with a stride of 2 and the fifth dilated convolution layer, and the fifth dilated convolution layer is used as the second input end of the output end of the first additive fusion layer; The fourth branch of the input layer of the first branch decoder module is sequentially connected to the third dilated convolution layer, the fourth batch normalization and linear activation combination layer, and the fourth batch normalization and linear activation combination layer serves as the fourth input terminal of the connection layer; The fifth branch of the input layer of the first branch decoder module is sequentially connected to the second 1x1 convolution layer with a step size of 1, the first global average pooling layer, the first deformation layer, and the first upsampling layer, and the output end of the first deformation layer is connected to the input end of the first fully connected layer, and the branch of the first upsampling layer serves as the fifth input end of the connection layer; after the first fully connected layer, a linear activation function, a second fully connected layer, a sigmoid function layer, and a second deformation layer are sequentially connected; the output end of the second upsampling layer serves as the second input end of the first multiplication fusion layer; The connection layer is followed by the third 1x1 convolutional layer with a stride of 1, the eighth batch normalization and linear activation combination layer, the fourth upsampling layer, and the first output of the first branch decoder module. The output of the first additive fusion layer is sequentially connected to a 3x3 convolution layer with a stride of 1, the seventh batch normalization and linear activation combination layer, and the output of the seventh batch normalization and linear activation combination layer is used as the first input of the first multiplicative fusion layer; the output of the first multiplicative fusion layer is used as the second input of the second multiplicative fusion layer, and after the second multiplicative fusion layer, it is sequentially connected to the third upsampling layer, the ninth batch normalization and linear activation function combination layer, and the second output of the first branch decoder module.

[0031] like Figure 4 As shown, as a preferred embodiment of the present invention, the second branch decoder module includes: an input layer, four average pooling layers, five 1x1 convolution layers, six batch normalization and linear activation combination layers, eight upsampling layers, three fully connected layers, a global average pooling layer, two deformation layers, a linear activation layer, two sigmoid activation function layers, two connection layers, a multiplication fusion layer and an output layer; The input layer of the second branch decoder module is connected to a global average pooling layer and the first upsampling layer respectively. The branch connected to the global average pooling layer is sequentially connected to the first fully connected layer, a linear activation layer, the second fully connected layer, the first sigmoid activation function layer, the first deformation layer, the second upsampling layer, and then serves as the first input end of the multiplication fusion layer. The branches connected to the first upsampling layer are respectively connected to the second connection layer, the first average pooling layer, the second average pooling layer, the third average pooling layer, and the fourth average pooling layer; The branch connected to the first average pooling layer is connected to the first 1x1 convolution layer, the first batch normalization and linear activation combination layer, and the third upsampling layer in sequence. The output of the third upsampling layer is used as the first input of the first connection layer. The branch connected to the second average pooling layer is connected to the second 1x1 convolution layer, the second batch normalization and linear activation combination layer, and the fourth upsampling layer in sequence. The output of the fourth upsampling layer serves as the second input of the first connection layer. The branch connected to the third average pooling layer is connected to the third 1x1 convolution layer, the third batch normalization and linear activation combination layer, and the fifth upsampling layer in sequence. The output of the fifth upsampling layer serves as the third input of the first connection layer. The branch connected to the fourth average pooling layer is connected to the fourth 1x1 convolution layer, the fourth batch normalization and linear activation combination layer, and the sixth upsampling layer in sequence. The output of the sixth upsampling layer serves as the fourth input of the first connection layer. The output of the first connection layer is connected to the input of the second connection layer and the input of the third fully connected layer; The third fully connected layer is sequentially connected to the second sigmoid activation function layer, the second deformation layer, and the seventh upsampling layer. The output of the seventh upsampling layer is used as the second input of the multiplication fusion. The output of the multiplication fusion is sequentially connected to the fifth batch normalization linear activation combination layer and the eighth upsampling layer. The output of the eighth upsampling layer is used as the third input of the second connection layer. The output of the second connection layer is sequentially connected to the fifth 1x1 convolutional layer, the sixth batch normalization and linear activation combination layer, and then to the output layer of the second branch decoder module.

[0032] As a preferred embodiment of the present invention, in step S2, training the dual-branch decoding residual network includes the following sub-steps: S2.1, set the training starting learning rate, maximum learning rate, training batch size and number of training rounds; S2.2, during the training process, L2 norm regularization is used to prevent the network from overfitting; S2.3. Use the Adam gradient optimization algorithm to optimize the network training. The following function is used as the loss function during the optimization process. The calculation formula is as follows: , In the formula, is the difference between the true value of the training data set sample and the estimated value of the network training, , represents the phase pixel coordinate, b is a positive coefficient; S2.4. Return to execute steps S2.2 to S2.3 until the network converges, obtain the dual-branch decoding residual network interference phase unwrapping model and weights, and save the network training weights.

[0033] Finally, the model is verified and evaluated to verify the dual-branch decoding residual network interference phase unwrapping model. Specifically, the verification data set is sent to the trained model, and the verification is continued to obtain the interference phase unwrapping phase map. The normalized root mean square error NRMES is used to evaluate the accuracy of star phase unwrapping. The smaller the NRMES value, the higher the phase unwrapping accuracy and the better the network performance. The calculation formula is as follows: , Among them, NRMES is the normalized root mean square error, N is the training data set sample, is the difference between the true value of sample k in the training data set and its network training estimate, Represents the phase pixel coordinates.

[0034] The present invention is based on a double-branch decoding residual network to interfere with the entanglement phase diagram under noise. Figure 5 The real interference phase diagram is shown as Figure 6 As shown, the interference unwrapping phase diagram under noise is as follows Figure 7 As shown, by observing Figure 6 The real interferometric phase pattern and Figure 5 The winding phase diagram under noise can intuitively show the impact of noise on the interference phase diagram. Figure 6 True phase diagram and Figure 7 The predicted interference unwrapping phase diagram shows that the two have extremely high similarities, which illustrates the accuracy of the prediction. At the same time, the network is calculated through the root mean square error NRMSE. The present invention can achieve high-precision unwrapping of interference entangled phase under low signal-to-noise ratio, and the NRMSE result reaches 1.39%. Under the same other conditions, the root mean square error NRMSE obtained by phase unwrapping of other networks such as U-net, PhaseNet, PUnet and PSPNet methods are 3.91%, 10.61%, 2.97% and 1.95% respectively. The root mean square error describes the gap between the predicted phase and the true phase. The smaller the value, the smaller the gap between the two, and the higher the accuracy of unwrapping. The experimental results show that the present invention is an interference phase unwrapping method with strong robustness and high accuracy under low signal-to-noise ratio conditions, which can effectively improve the accuracy of synthetic aperture radar interferometry.

[0035] Although the present invention has been described above with preferred embodiments, it is not intended to limit the present invention. A person skilled in the art of the present invention may make various modifications and improvements without departing from the spirit and scope of the present invention. Therefore, the scope of protection of the present invention shall be determined by the definition of the claims.

Claims

1. An interferometric phase unwrapping method based on a double-branch decoding residual network, characterized in that: The following steps are involved: S1. Construct an interference phase data set, including an interference phase map of pixel phase entanglement and its corresponding real interference phase map; S2, constructing a dual-branch decoding residual network, including a sequentially connected input layer, an encoder network, a dual-branch decoder network including a first branch decoder module and a second branch decoder module, and an output layer; The interference phase data set is used to train a two-branch decoding residual network to obtain a two-branch decoding residual network interference phase unwrapping model and the weight of the model; S3. Input the interferometric phase image with pixel phase entanglement into the dual-branch decoding residual network interferometric phase unwrapping model to obtain the predicted interferometric unwrapping phase image.

2. The interference phase unwrapping method based on a double-branch decoding residual network according to claim 1, characterized in that: In step S1, constructing the interferometric phase data set includes the following sub-steps: S1.1, creating random-shaped interference phase images by adding and subtracting several Gaussian functions of different shapes and positions; S1.2, randomly select ramps and add them to the interference phase diagram in vertical and horizontal directions to form an interference phase diagram with ramp phase; S1.3, the phase of the interference phase image with the ramp phase is pixel-wound, and the phase is wound The calculation formula is as follows: ， Among them, exp is a natural constant e The exponential function with base , is the original phase of the interference phase image pixel, is the spatial coordinate of the pixel in the phase image, To find the angle sign.

3. The interferometric phase unwrapping method based on a dual-branch decoding residual network according to claim 1, characterized in that: The encoder network in step S2 includes a convolutional layer with a step size of 2, a maximum pooling layer with a pooling size of 3 and a step size of 2, and sixteen residual modules connected in sequence; the sixteenth residual module is connected to the dual-branch decoder network as the output end of the encoder network, and the structure of the residual module is set according to the input step size, as follows: When the stride is not 1, the input layer of the residual module is connected to a 1x1 convolution layer with the same stride and channel as the input and a 3x3 convolution layer with the same stride and channel as the input, the branch connecting the 3x3 convolution layer is sequentially connected to a batch normalization and linear activation function combination layer, a 3x3 convolution layer and a batch normalization layer as the first input end of the additive fusion, the branch connecting the 1x1 convolution layer is connected to a batch normalization layer as the second input end of the additive fusion, and the output of the additive fusion is output as the residual module result after a linear activation. When the stride is 1, the input layer of the residual module is respectively connected to an additive fusion layer and a 3x3 convolution layer with the same stride and channel as the input, the branch connecting the 3x3 convolution layer is sequentially connected to a batch normalization and linear activation function combination layer, a 3x3 convolution layer and a batch normalization layer as the other input end of the additive fusion, and the output of the additive fusion is output as the residual module result after a linear activation.

4. The interference phase unwrapping method based on a double-branch decoding residual network according to claim 1, characterized in that: The first branch decoder module includes: an input layer, three 1x1 convolutional layers with a stride of 1, five dilated convolutional layers, nine batch normalization and linear activation combination layers, three 1x1 convolutional layers with a stride of 2, a global average pooling layer, two deformation layers, two fully connected layers, a sigmoid function layer, two multiplication fusion layers, four upsampling layers, an addition fusion layer, a connection layer, a 3x3 convolutional layer with a stride of 1, a linear activation layer, and two output layers; The first branch of the input layer of the first branch decoder module sequentially connects the first 1x1 convolution layer with a stride of 1, the first batch normalization and linear activation combination layer, the output of the first batch normalization and linear activation combination layer is used as the first input of the connection layer, and the first 1x1 convolution layer with a stride of 2, the output of the fifth batch normalization and linear activation combination layer are sequentially connected as the first input of the second multiplication fusion layer; The second branch of the input layer of the first branch decoder module sequentially connects the first dilated convolution layer, the second batch normalization and linear activation combination layer. The second batch normalization and linear activation combination layer is used as the second input of the connection layer, and the second 1x1 convolution layer with a stride of 2, the fourth dilated convolution layer, the sixth batch normalization and linear activation combination layer are sequentially connected. The sixth batch normalization and linear activation combination layer is used as the first input of the output of the first additive fusion layer. The third branch of the input layer of the first branch decoder module is sequentially connected to the second dilated convolution layer, the third batch normalization and linear activation combination layer, the third batch normalization and linear activation combination layer is used as the third input end of the connection layer, and is sequentially connected to the third 1x1 convolution layer with a stride of 2 and the fifth dilated convolution layer, and the fifth dilated convolution layer is used as the second input end of the output end of the first additive fusion layer; The fourth branch of the input layer of the first branch decoder module is sequentially connected to the third dilated convolution layer, the fourth batch normalization and linear activation combination layer, and the fourth batch normalization and linear activation combination layer serves as the fourth input terminal of the connection layer; The fifth branch of the input layer of the first branch decoder module is sequentially connected to the second 1x1 convolution layer with a step size of 1, the first global average pooling layer, the first deformation layer, and the first upsampling layer, and the output end of the first deformation layer is connected to the input end of the first fully connected layer, and the branch of the first upsampling layer serves as the fifth input end of the connection layer; after the first fully connected layer, a linear activation function, a second fully connected layer, a sigmoid function layer, and a second deformation layer are sequentially connected; the output end of the second upsampling layer serves as the second input end of the first multiplication fusion layer; The connection layer is followed by the third 1x1 convolutional layer with a stride of 1, the eighth batch normalization and linear activation combination layer, the fourth upsampling layer, and the first output of the first branch decoder module. The output of the first additive fusion layer is sequentially connected to a 3x3 convolution layer with a stride of 1, the seventh batch normalization and linear activation combination layer, and the output of the seventh batch normalization and linear activation combination layer is used as the first input of the first multiplicative fusion layer; the output of the first multiplicative fusion layer is used as the second input of the second multiplicative fusion layer, and after the second multiplicative fusion layer, it is sequentially connected to the third upsampling layer, the ninth batch normalization and linear activation function combination layer, and the second output of the first branch decoder module.

5. The interference phase unwrapping method based on a double-branch decoding residual network according to claim 1, characterized in that: The second branch decoder module includes: an input layer, four average pooling layers, five 1x1 convolution layers, six batch normalization and linear activation combination layers, eight upsampling layers, three fully connected layers, a global average pooling layer, two deformation layers, a linear activation layer, two sigmoid activation function layers, two connection layers, a multiplication fusion layer and an output layer; The input layer of the second branch decoder module is connected to a global average pooling layer and the first upsampling layer respectively. The branch connected to the global average pooling layer is sequentially connected to the first fully connected layer, a linear activation layer, the second fully connected layer, the first sigmoid activation function layer, the first deformation layer, the second upsampling layer, and then serves as the first input end of the multiplication fusion layer. The branches connected to the first upsampling layer are respectively connected to the second connection layer, the first average pooling layer, the second average pooling layer, the third average pooling layer, and the fourth average pooling layer; The branch connected to the first average pooling layer is connected to the first 1x1 convolution layer, the first batch normalization and linear activation combination layer, and the third upsampling layer in sequence. The output of the third upsampling layer is used as the first input of the first connection layer. The branch connected to the second average pooling layer is connected to the second 1x1 convolution layer, the second batch normalization and linear activation combination layer, and the fourth upsampling layer in sequence. The output of the fourth upsampling layer serves as the second input of the first connection layer. The branch connected to the third average pooling layer is connected to the third 1x1 convolution layer, the third batch normalization and linear activation combination layer, and the fifth upsampling layer in sequence. The output of the fifth upsampling layer serves as the third input of the first connection layer. The branch connected to the fourth average pooling layer is connected to the fourth 1x1 convolution layer, the fourth batch normalization and linear activation combination layer, and the sixth upsampling layer in sequence. The output of the sixth upsampling layer serves as the fourth input of the first connection layer. The output of the first connection layer is connected to the input of the second connection layer and the input of the third fully connected layer; The third fully connected layer is sequentially connected to the second sigmoid activation function layer, the second deformation layer, and the seventh upsampling layer. The output of the seventh upsampling layer is used as the second input of the multiplication fusion. The output of the multiplication fusion is sequentially connected to the fifth batch normalization linear activation combination layer and the eighth upsampling layer. The output of the eighth upsampling layer is used as the third input of the second connection layer. The output of the second connection layer is sequentially connected to the fifth 1x1 convolutional layer, the sixth batch normalization and linear activation combination layer, and then to the output layer of the second branch decoder module.

6. The interference phase unwrapping method based on a double-branch decoding residual network according to claim 1, characterized in that: In step S2, training the dual-branch decoding residual network includes the following sub-steps: S2.1, set the training starting learning rate, maximum learning rate, training batch size and number of training rounds; S2.2, during the training process, L2 norm regularization is used to prevent the network from overfitting; S2.

3. Use the Adam gradient optimization algorithm to optimize the network training. The following function is used as the loss function during the optimization process. The calculation formula is as follows: ， In the formula, is the difference between the true value of the training data set sample and the estimated value of the network training, , represents the phase pixel coordinate, b is a positive coefficient; S2.

4. Return to execute steps S2.2 to S2.3 until the network converges, obtain the dual-branch decoding residual network interference phase unwrapping model and weights, and save the network training weights.

7. The interference phase unwrapping method based on a double-branch decoding residual network according to claim 1, characterized in that: The method also includes verifying the dual-branch decoding residual network interference phase unwrapping model, specifically: sending the verification data set to the trained model, continuing the verification, and obtaining the interference phase unwrapping phase map.

8. The interference phase unwrapping method based on a double-branch decoding residual network according to claim 1, characterized in that: The method further includes evaluating the phase unwrapping accuracy by using the normalized root mean square error NRMES. The smaller the NRMES value, the higher the phase unwrapping accuracy and the better the network performance. The calculation formula is as follows: , Among them, NRMES is the normalized root mean square error, N is the training data set sample, is the difference between the true value of sample k in the training data set and its estimated value during network training. Represents the phase pixel coordinates.

Images