An interference disentanglement method based on multi-scale pooling dilated pyramid network

Abstract

The present invention relates to an interference unwrapping method based on a multi-scale pooling dilated pyramid network. A sample data set is constructed by combining a real phase image with its corresponding wrapped phase image. A network to be trained is designed to include a multi-scale pooling module MSPM and a dilated pyramid module DCPM. Network training is performed to obtain a corresponding phase image unwrapping model, and interference phase unwrapping is achieved under low signal-to-noise ratio conditions. The method has the advantages of fast convergence speed, strong real-time performance, and strong robustness. In practical applications, the method can effectively improve the interferometric measurement accuracy of synthetic aperture radar.

Description

Technical Field

[0001] The invention relates to an interference unwrapping method based on a multi-scale pooling dilated pyramid network, and belongs to the technical field of synthetic aperture radar interferometry. Background Art

[0002] Interferometric Synthetic Aperture Radar (InSAR) is one of the most promising space-based earth observation technologies and an extension of SAR technology. InSAR technology has the advantages of all-weather, all-day, large-scale, low-cost, high-resolution, and the accuracy of detecting surface deformation can reach centimeters or even millimeters. It is widely used in the fields of surface deformation, earthquake research, urban and rural planning, volcanic activity, and military monitoring. In the process of InSAR data processing, phase unwrapping is a key step, but due to the influence of noise, phase discontinuity, and rapid phase changes, the phase unwrapping problem becomes extremely challenging, so it is very important to explore efficient and accurate unwrapping methods for InSAR technology.

[0003] Traditional phase unwrapping methods are mainly divided into two categories: path tracking method and minimum optimization method. Under the conditions of high noise and complex terrain, traditional interferometric phase unwrapping methods usually show the shortcomings of large errors and low efficiency. With the rise of artificial intelligence technology, deep learning has become an important branch of artificial intelligence. For the field of interferometric phase unwrapping, deep learning-based networks such as U-net (U-type convolutional neural network), QGPU (quality-guided phase unwrapping), PhaseNet (phase unwrapping neural network), and PUnet (phase unwrapping neural network) have been applied to interferometric phase unwrapping, and have achieved excellent performance in interferometric phase unwrapping scenarios. However, due to the huge data set required for deep learning and the influence of noise and complex terrain, the existing neural network models have the disadvantages of long training time, large training data set size, poor robustness, etc., and still have certain limitations for applications that require high precision and high timeliness. Summary of the invention

[0004] The technical problem to be solved by the present invention is to provide an interference unwrapping method based on a multi-scale pooling dilated pyramid network, construct a network to be trained including a multi-scale pooling module MSPM and a dilated pyramid module DCPM, and train a phase image unwrapping model using a sample data set to achieve interference phase unwrapping under low signal-to-noise ratio conditions, thereby improving the accuracy of synthetic aperture radar interferometry measurement.

[0005] In order to solve the above technical problems, the present invention adopts the following technical solutions: the present invention designs an interference unwrapping method based on a multi-scale pooling dilated pyramid network, and obtains a phase image unwrapping model corresponding to a target type synthetic aperture radar according to the following steps A to C, and then processes the wrapped phase image of the target type synthetic aperture radar using the phase image unwrapping model to obtain a corresponding unwrapped phase image;

[0006] Step A. Obtain a preset number of real phase images corresponding to the target type synthetic aperture radar, and the wrapped phase images corresponding to each real phase image, and use a single real phase image and its corresponding wrapped phase image to form a sample, thereby constructing a sample data set, and then proceeding to step B;

[0007] Step B. Construct a network to be trained including a multi-scale pooling module MSPM and a dilated pyramid module DCPM, and then proceed to step C;

[0008] Step C. Based on the sample data set, with the wrapped phase image as input and the corresponding real phase image as output, the network to be trained is trained to obtain a phase image unwrapping model corresponding to the target type synthetic aperture radar.

[0009] As a preferred technical solution of the present invention: in the step A, for each real phase image, firstly, a slope is randomly selected along the vertical direction and the horizontal direction to be added to the real phase image to obtain the corresponding interference phase image, then the interference phase image is updated by pixel phase wrapping, and finally, Gaussian noise is added to the interference phase image to obtain the corresponding wrapped phase image; thereby obtaining the wrapped phase images corresponding to each real phase image.

[0010] As a preferred technical solution of the present invention: in the step A, for each interference phase image and each pixel position in the interference phase image, the following formula is used:

[0011]

[0012] Get the winding phase corresponding to the pixel position Implement pixel phase wrapping update for pixel positions, where φ(x,y) represents the true phase of the (x,y) pixel position in the interferometric phase image, j represents the imaginary unit, exp() represents the exponential function with the natural constant e as the base, and ∠ represents the angle sign. Represents the wrap-around phase at the (x,y) pixel position in the interferometric phase image.

[0013] As a preferred technical solution of the present invention: the network to be trained includes an encoder network Encoder and a decoder network Decoder, wherein the encoder network Encoder includes a multi-scale pooling module MSPM, an expansion pyramid module DCPM, and a 3x3 convolution layer, a batch normalization and a linear activation combination layer and four encoding submodules sequentially connected in series from the input end to the output end, wherein the input end of the 3x3 convolution layer constitutes the input end of the encoder network Encoder, that is, the input end of the network to be trained, the output end of the fourth encoding submodule in the sequential series connection is connected to the input end of the expansion pyramid module DCPM, and the output end of the expansion pyramid module DCPM constitutes the main output end of the encoder network Encoder; the output end of the first encoding submodule in the sequential series connection in the encoder network Encoder is connected to the input end of the multi-scale pooling module MSPM; the structures of the encoding submodules are the same, and the encoding submodules are respectively connected in series with the maximum pooling layer and the convolution block in the direction from the input end to the output end, wherein the input end of the maximum pooling layer constitutes the input end of the encoding submodule, and the output end of the convolution block constitutes the output end of the encoding submodule;

[0014] The decoder network Decoder includes four decoding submodules connected in series from the input end to the output end, as well as a transposed convolution layer, a connection layer, a 3x3 convolution layer, a batch normalization and a linear activation combination layer, wherein the output end of the batch normalization and linear activation combination layer constitutes the output end of the decoder network Decoder, and the input end of the first decoding submodule in the series connection constitutes the input end of the decoder network Decoder; the structures of each decoding submodule are the same, and each decoding submodule is connected in series from the input end to the output end, wherein the input end of the transposed convolution layer constitutes the input end of the decoding submodule, and the output end of the convolution block constitutes the output end of the decoding submodule;

[0015] The output of the batch normalization and linear activation combination layer in the encoder network Encoder is connected to the input of the last connection layer in the decoder network Decoder in series. The four outputs of the multi-scale pooling module MSPM in the encoder network Encoder are connected to the input of the connection layer in the four decoding sub-modules in the decoder network Decoder in one-to-one correspondence. The multi-scale pooling module MSPM outputs four feature maps of different sizes, and the four feature maps are sequentially transmitted to the input of the connection layer in each decoding sub-module in the decoder network Decoder in series from small to large in size; the encoder network Encoder is sequentially connected in series. The output end of the second encoding submodule in the next sequence is connected to the input end of the connection layer of the third decoding submodule in the next sequence in series in the decoder network Decoder, the output end of the third encoding submodule in the next sequence in series in the encoder network Encoder is connected to the input end of the connection layer of the second decoding submodule in the next sequence in series in the decoder network Decoder, the output end of the fourth encoding submodule in the next sequence in series in the encoder network Encoder is connected to the input end of the connection layer of the first decoding submodule in the next sequence in series in the decoder network Decoder, and the main output end of the encoder network Encoder is connected to the input end of the decoder network Decoder.

[0016] As a preferred technical solution of the present invention: the multi-scale pooling module MSPM includes four branches corresponding to four feature maps of different sizes, the four branches are respectively connected in series from the input end to the output end, the average pooling layer, the 1x1 convolution layer, the batch normalization and the linear activation combination layer, the average pooling sizes of the average pooling layers on the four branches are different from each other, the input ends of the average pooling layers on the four branches are connected to each other, constituting the input end of the multi-scale pooling module MSPM, and the output ends of the batch normalization and the linear activation combination layers on the four branches respectively constitute the output ends of the multi-scale pooling module MSPM.

[0017] As a preferred technical solution of the present invention: the expansion pyramid module DCPM includes a multiplication fusion layer, a batch normalization and linear activation combination layer, two addition fusion layers, and four branches, wherein the input ends of the branches are connected to form the input end of the expansion pyramid module DCPM, the output end of the second branch and the output end of the third branch are connected to the input end of one of the addition fusion layers, the output end of the addition fusion layer is connected to the input end of the batch normalization and linear activation combination layer, the output end of the batch normalization and linear activation combination layer and the output end of the first branch are connected to the input end of the multiplication fusion layer, the output end of the multiplication fusion layer and the output end of the fourth branch are connected to the input end of another addition fusion layer, and the output end of the addition fusion layer constitutes the output end of the expansion pyramid module DCPM;

[0018] The first branch is connected in series from the input end to the output end in the direction of the 1x1 convolution layer with a step size of s=1, the 1x1 convolution layer with a step size of s=2, the batch normalization and the linear activation combination layer, wherein the input end of the 1x1 convolution layer with a step size of s=1 constitutes the input end of the first branch, and the output end of the batch normalization and the linear activation combination layer constitutes the output end of the first branch;

[0019] The second branch includes a 3x3 dilated convolution layer with a dilation coefficient of r=5, a 3x3 dilated convolution layer with a dilation coefficient of r=9, two 1x1 convolution layers with a step size of s=2, and two 3x3 convolution layers with a step size of s=1, wherein the input end of the 3x3 dilated convolution layer with a dilation coefficient of r=5 constitutes the input end of the second branch, and the two output ends of the 3x3 dilated convolution layer with a dilation coefficient of r=5 are respectively connected to the input end of the 3x3 dilated convolution layer with a dilation coefficient of r=9 and one of the 1x1 convolution layers with a step size of s=2. The input end of the 1x1 convolution layer, the output end of the 1x1 convolution layer with a step size of s=2 is connected to the input end of one of the 3x3 convolution layers with a step size of s=1, the output end of the 3x3 dilated convolution layer with a dilation coefficient of r=9 is connected to the input end of another 1x1 convolution layer with a step size of s=2, the output end of the 1x1 convolution layer with a step size of s=2 is connected to the input end of another 3x3 convolution layer with a step size of s=1, and the output ends of the two 3x3 convolution layers with a step size of s=1 constitute the output end of the second branch;

[0020] The third branch is connected in series from the input end to the output end in the direction of the 3x3 dilated convolution layer with a dilation coefficient of r=9, the 3x3 dilated convolution layer with a dilation coefficient of r=11, the 1x1 convolution layer with a step size of s=2, and the 3x3 convolution layer with a step size of s=1, wherein the input end of the 3x3 dilated convolution layer with a dilation coefficient of r=9 constitutes the input end of the third branch, and the output end of the 3x3 convolution layer with a step size of s=1 constitutes the output end of the third branch;

[0021] The fourth branch connects a 1x1 convolution layer with a step size of s=1, a global average pooling layer, a deformation layer, and a 1x1 convolution layer with a step size of s=1 in series from the input end to the output end; among them, the input end of the first 1x1 convolution layer with a step size of s=1 constitutes the input end of the fourth branch, and the output end of the last 1x1 convolution layer with a step size of s=1 constitutes the output end of the fourth branch.

[0022] As a preferred technical solution of the present invention: the structures of each convolution block are the same as each other, and each convolution block is respectively connected in series from the input end to the output end. A 1x1 convolution layer, two 3x3 convolution layers, and a batch normalization and linear activation combination layer are sequentially connected, wherein the input end of the 1x1 convolution layer constitutes the input end of the convolution block, and the output end of the batch normalization and linear activation combination layer constitutes the output end of the convolution block.

[0023] As a preferred technical solution of the present invention: in the process of training the network to be trained in step C, the Adam gradient optimization algorithm is used to optimize the network training, and the loss function used in the optimization process is as follows:

[0024]

[0025] Where (x, y) represents the pixel position in the phase image corresponding to the training sample, L(y′(x, y)) represents the loss result of the pixel position (x, y) in the wrapped phase image in the training sample, y′(x, y) represents the difference between the true phase and the predicted phase of the pixel position (x, y) in the wrapped phase image in the training sample, and b represents the preset positive coefficient.

[0026] As a preferred technical solution of the present invention: it also includes a phase image unwrapping model corresponding to the synthetic aperture radar of the target type, and performs evaluation according to the following formula;

[0027]

[0028] Where NRMSE is the normalized root mean square error, k = 1, 2, ..., N, N represents the number of training samples in the training process of the network to be trained, (x, y) represents the pixel position in the phase image corresponding to the training sample, and y′ k (x,y) represents the difference between the true phase and the predicted phase at the pixel position (x,y) in the wrapped phase image in the kth training sample.

[0029] The interference disentanglement method based on a multi-scale pooling dilated pyramid network described in the present invention adopts the above technical solution and has the following technical effects compared with the prior art:

[0030] (1) The present invention designs an interference unwrapping method based on a multi-scale pooling dilated pyramid network. A sample data set is constructed by combining a real phase image with its corresponding wrapped phase image. A network to be trained is designed to include a multi-scale pooling module MSPM and a dilated pyramid module DCPM. Network training is performed to obtain a corresponding phase image unwrapping model, thereby realizing interference phase unwrapping under low signal-to-noise ratio conditions. The method has the advantages of fast convergence speed, strong real-time performance, and strong robustness. In practical applications, the method can effectively improve the accuracy of synthetic aperture radar interferometry measurement. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 It is a schematic diagram of the structure of the network to be trained in the phase unwrapping method designed in the present invention;

[0032] Figure 2 It is a schematic diagram of the structure of the multi-scale pooling module MSPM in the network to be trained designed by the present invention;

[0033] Figure 3 It is a schematic diagram of the structure of the expansion pyramid module DCPM in the network to be trained designed by the present invention;

[0034] Figure 4 Schematic diagram of a wrapped phase image according to an embodiment of the present invention;

[0035] Figure 5 yes Figure 4 Schematic diagram of a real phase image corresponding to a wrapped phase image of an embodiment;

[0036] Figure 6 yes Figure 4 Schematic diagram of the corresponding unwrapped phase image obtained by implementing the wrapped phase image of the present invention. DETAILED DESCRIPTION

[0037] The specific implementation modes of the present invention will be further described in detail below in conjunction with the accompanying drawings.

[0038] The present invention designs an interference unwrapping method based on a multi-scale pooling dilated pyramid network. In practical applications, the following steps A to C are specifically performed to obtain a phase image unwrapping model corresponding to a target type synthetic aperture radar, and then the phase image unwrapping model is applied to the wrapped phase image of the target type synthetic aperture radar to obtain a corresponding unwrapped phase image.

[0039] Step A. Obtain a preset number of real phase images corresponding to the target type synthetic aperture radar, as well as the wrapped phase images corresponding to each real phase image, and use a single real phase image and its corresponding wrapped phase image to form a sample, thereby constructing a sample data set, and then proceed to step B.

[0040] In practical applications, for each real phase image, firstly, a slope is randomly selected in the vertical and horizontal directions and added to the real phase image to obtain the corresponding interference phase image; then, for each interference phase image, each pixel position in the interference phase image is further calculated according to the following formula:

[0041]

[0042] Get the winding phase corresponding to the pixel position The pixel phase winding update is implemented for the pixel position, and then the pixel phase winding update is performed on the interference phase image. Wherein, φ(x, y) represents the real phase of the (x, y) pixel position in the interference phase image, j represents the imaginary unit, exp() represents the exponential function with the natural constant e as the base, ∠ represents the angle sign, Represents the wrap-around phase at the (x,y) pixel position in the interferometric phase image.

[0043] Finally, Gaussian additive noises such as 0dB, 3d, 5dB, and 10dB are added to the interference phase image to obtain the corresponding wrapped phase image; and then the wrapped phase images corresponding to each real phase image are obtained.

[0044] In practical applications, for example, a sample data set of 1400 is created, where the pixel size of the phase image in each sample is 256×256 and the pixel value is -55 to 55.

[0045] Step B. Construct a network to be trained including a multi-scale pooling module MSPM and a dilated pyramid module DCPM, and then proceed to step C.

[0046] In practical applications, such as Figure 1 As shown, the network to be trained includes an encoder network Encoder and a decoder network Decoder, wherein the encoder network Encoder includes a multi-scale pooling module MSPM, an expansion pyramid module DCPM, and a 3x3 convolution layer, a batch normalization and a linear activation combination layer and four encoding submodules sequentially connected in series from the input end to the output end, wherein the input end of the 3x3 convolution layer constitutes the input end of the encoder network Encoder, that is, the input end of the network to be trained, the output end of the fourth encoding submodule in the sequential series connection is connected to the input end of the expansion pyramid module DCPM, and the output end of the expansion pyramid module DCPM constitutes the main output end of the encoder network Encoder; the output end of the first encoding submodule in the sequential series connection in the encoder network Encoder is connected to the input end of the multi-scale pooling module MSPM; the structures of the encoding submodules are the same, and each encoding submodule is sequentially connected in series with a maximum pooling layer and a convolution block from the input end to the output end, wherein the input end of the maximum pooling layer constitutes the input end of the encoding submodule, and the output end of the convolution block constitutes the output end of the encoding submodule.

[0047] like Figure 1 As shown, the decoder network Decoder includes four decoding submodules connected in series from the input end to the output end, as well as a transposed convolution layer, a connection layer, a 3x3 convolution layer, a batch normalization and a linear activation combination layer, wherein the output ends of the batch normalization and linear activation combination layers constitute the output end of the decoder network Decoder, and the input end of the first decoding submodule in the series connection constitutes the input end of the decoder network Decoder; the structures of each decoding submodule are the same, and each decoding submodule is connected in series from the input end to the output end, respectively, with a transposed convolution layer, a connection layer, and a convolution block, wherein the input end of the transposed convolution layer constitutes the input end of the decoding submodule, and the output end of the convolution block constitutes the output end of the decoding submodule.

[0048] like Figure 1As shown in the figure, the output of the batch normalization and linear activation combination layer in the encoder network Encoder is connected to the input of the last connection layer in the decoder network Decoder in series, and the four outputs of the multi-scale pooling module MSPM in the encoder network Encoder are connected to the input of the connection layer in the four decoding sub-modules in the decoder network Decoder in one-to-one correspondence. The multi-scale pooling module MSPM outputs four feature maps of different sizes, and the four feature maps are sequentially transmitted to the input of the connection layer in each decoding sub-module in the decoder network Decoder in series from small to large in size; the encoder network Encoder is sequentially connected to the decoder network Decoder in series. The output end of the second encoding submodule in series is connected to the input end of the connection layer in the third decoding submodule in series in the decoder network Decoder, the output end of the third encoding submodule in series in the encoder network Encoder is connected to the input end of the connection layer in the second decoding submodule in series in the decoder network Decoder, the output end of the fourth encoding submodule in series in the encoder network Encoder is connected to the input end of the connection layer in the first decoding submodule in series in the decoder network Decoder, and the main output end of the encoder network Encoder is connected to the input end of the decoder network Decoder.

[0049] In the above design, if Figure 2 As shown, the multi-scale pooling module MSPM is further designed to include four branches corresponding to four feature maps of different sizes. The four branches are respectively connected in series from the input end to the output end, the average pooling layer, the 1x1 convolution layer, the batch normalization layer and the linear activation combination layer. The pooling sizes of the average pooling layers on the four branches are different from each other. The input ends of the average pooling layers on the four branches are connected to each other to constitute the input end of the multi-scale pooling module MSPM. The output ends of the batch normalization and the linear activation combination layers on the four branches respectively constitute the output ends of the multi-scale pooling module MSPM.

[0050] In practical applications, such as Figure 3As shown, the designed expansion pyramid module DCPM includes a multiplication and fusion layer (Multiply), a batch normalization and linear activation combination layer, two addition and fusion layers, and four branches, wherein the input ends of the branches are connected to form the input end of the expansion pyramid module DCPM, the output end of the second branch and the output end of the third branch are connected to the input end of one of the addition and fusion layers, the output end of the addition and fusion layer is connected to the input end of the batch normalization and linear activation combination layer, the output end of the batch normalization and linear activation combination layer and the output end of the first branch are connected to the input end of the multiplication and fusion layer (Multiply), the output end of the multiplication and fusion layer (Multiply) and the output end of the fourth branch are connected to the input end of another addition and fusion layer, and the output end of the addition and fusion layer constitutes the output end of the expansion pyramid module DCPM.

[0051] In the expansion pyramid module DCPM, Figure 3 As shown, the first branch is connected in series from the input end to the output end in sequence: a 1x1 convolution layer with a step size of s=1, a 1x1 convolution layer with a step size of s=2, a batch normalization and a linear activation combination layer, wherein the input end of the 1x1 convolution layer with a step size of s=1 constitutes the input end of the first branch, and the output end of the batch normalization and linear activation combination layer constitutes the output end of the first branch.

[0052] like Figure 3 As shown, the second branch includes a 3x3 dilated convolution layer with a dilation coefficient of r=5, a 3x3 dilated convolution layer with a dilation coefficient of r=9, two 1x1 convolution layers with a step size of s=2, and two 3x3 convolution layers with a step size of s=1, wherein the input end of the 3x3 dilated convolution layer with a dilation coefficient of r=5 constitutes the input end of the second branch, and the two output ends of the 3x3 dilated convolution layer with a dilation coefficient of r=5 are respectively connected to the input end of the 3x3 dilated convolution layer with a dilation coefficient of r=9, and one of the 1x1 convolution layers with a step size of s=2. The input end of the 1x1 convolution layer with step size s=2 is connected to the input end of one of the 3x3 convolution layers with step size s=1, the output end of the 3x3 dilated convolution layer with expansion coefficient r=9 is connected to the input end of another 1x1 convolution layer with step size s=2, the output end of the 1x1 convolution layer with step size s=2 is connected to the input end of another 3x3 convolution layer with step size s=1, and the output ends of the two 3x3 convolution layers with step size s=1 constitute the output end of the second branch.

[0053] like Figure 3As shown, the third branch is connected in series from the input end to the output end in sequence: a 3x3 dilated convolution layer with a dilation coefficient of r=9, a 3x3 dilated convolution layer with a dilation coefficient of r=11, a 1x1 convolution layer with a step size of s=2, and a 3x3 convolution layer with a step size of s=1, wherein the input end of the 3x3 dilated convolution layer with a dilation coefficient of r=9 constitutes the input end of the third branch, and the output end of the 3x3 convolution layer with a step size of s=1 constitutes the output end of the third branch.

[0054] like Figure 3 As shown, the fourth branch sequentially connects a 1x1 convolution layer with a step size of s=1, a global average pooling layer, a deformation layer (shape), and a 1x1 convolution layer with a step size of s=1 from the input end to the output end; wherein, the input end of the first 1x1 convolution layer with a step size of s=1 constitutes the input end of the fourth branch, and the output end of the last 1x1 convolution layer with a step size of s=1 constitutes the output end of the fourth branch.

[0055] Regarding the convolution blocks involved in the above-mentioned network design to be trained, in actual applications, the structures of the convolution blocks are designed to be the same as each other. Each convolution block is connected in series from the input end to the output end in sequence: a 1x1 convolution layer, two 3x3 convolution layers, and a batch normalization and linear activation combination layer, wherein the input end of the 1x1 convolution layer constitutes the input end of the convolution block, and the output end of the batch normalization and linear activation combination layer constitutes the output end of the convolution block.

[0056] Based on the above step A about establishing the sample data set and step B about creating the network to be trained, the following step C is finally performed.

[0057] Step C. Based on the sample data set, with the wrapped phase image as input and the corresponding real phase image as output, the network to be trained is trained to obtain a phase image unwrapping model corresponding to the target type synthetic aperture radar.

[0058] In the actual application of training the network to be trained, the starting learning rate of training is set to 0.0001, the maximum learning rate is set to 0.01, the number of training batches is set to 4, and the number of training rounds is set to 100; and the Adam gradient optimization algorithm is used to optimize the network training, and the loss function used in the optimization process is as follows:

[0059]

[0060] Where (x, y) represents the pixel position in the phase image corresponding to the training sample, L(y′(x, y)) represents the loss result of the pixel position (x, y) in the wrapped phase image in the training sample, y′(x, y) represents the difference between the true phase and the predicted phase of the pixel position (x, y) in the wrapped phase image in the training sample, and b represents the preset positive coefficient.

[0061] The above is to obtain the phase image unwrapping model corresponding to the target type synthetic aperture radar, and then apply the phase image unwrapping model to process the wrapped phase image of the target type synthetic aperture radar to obtain the corresponding unwrapped phase image. In practical applications, the following formula can be further designed and executed to evaluate the phase image unwrapping model.

[0062]

[0063] NRMSE is the normalized root mean square error. The smaller the NRMSE value, the higher the phase unwrapping accuracy, indicating that the network performance of the phase image unwrapping model is good. k = 1, 2, ..., N, N represents the number of training samples in the training process of the network to be trained, (x, y) represents the pixel position in the phase image corresponding to the training sample, y′ k (x,y) represents the difference between the true phase and the predicted phase at the pixel position (x,y) in the wrapped phase image in the kth training sample.

[0064] Apply the above design scheme to practice, such as Figure 4 The wrapped phase image schematic diagram of the embodiment shown in FIG. Figure 5 The corresponding real phase image shown in Figure 6 As shown in the figure, the corresponding unwrapped phase image obtained by implementing the design scheme of the present invention, the experimental results prove that the present invention can achieve accurate unwrapping of the interference entangled phase under low signal-to-noise ratio, and the NRMSE result reaches 1.5%. Compared with other networks such as U-net, QGPU, PhaseNet, and PUnet methods under the same conditions, the time used by the present invention is the least compared with the time used by other networks. At the same time, the calculation results of NRMSE of other networks are only 3.91%, 4.89%, 10.61% and 2.97%, respectively.

[0065] The interference unwrapping method based on the multi-scale pooling dilated pyramid network designed in the present invention constructs a sample data set with a real phase image in combination with its corresponding wrapped phase image, performs network training on a network to be trained that includes a multi-scale pooling module MSPM and a dilated pyramid module DCPM, obtains a corresponding phase image unwrapping model, and realizes interference phase unwrapping under low signal-to-noise ratio conditions. The method has the advantages of fast convergence speed, strong real-time performance, and strong robustness. In practical applications, the accuracy of synthetic aperture radar interferometry can be effectively improved.

[0066] The embodiments of the present invention are described in detail above with reference to the accompanying drawings, but the present invention is not limited to the above embodiments, and various changes can be made within the knowledge scope of ordinary technicians in this field without departing from the purpose of the present invention.

Claims

1. An interference disentanglement method based on a multi-scale pooling dilated pyramid network, characterized in that: According to the following steps A to C, a phase image unwrapping model corresponding to the target type synthetic aperture radar is obtained, and then the phase image unwrapping model is applied to the wrapped phase image of the target type synthetic aperture radar to obtain the corresponding unwrapped phase image; Step A. A preset number of real phase images corresponding to the target type synthetic aperture radar and wrapped phase images corresponding to each real phase image are obtained, and a single real phase image and its corresponding wrapped phase image are used as a sample to construct a sample data set, and then step B is entered; Step B. Construct a network to be trained including a multi-scale pooling module MSPM and a dilated pyramid module DCPM, and then proceed to step C; Step C. Based on the sample data set, with the wrapped phase image as input and the corresponding real phase image as output, the network to be trained is trained to obtain a phase image unwrapping model corresponding to the target type synthetic aperture radar; The network to be trained includes an encoder network Encoder and a decoder network Decoder, wherein the encoder network Encoder includes a multi-scale pooling module MSPM, an expansion pyramid module DCPM, and a 3x3 convolution layer, a batch normalization and a linear activation combination layer and four encoding submodules sequentially connected in series from the input end to the output end, wherein the input end of the 3x3 convolution layer constitutes the input end of the encoder network Encoder, that is, the input end of the network to be trained, the output end of the fourth encoding submodule in the sequential series is connected to the input end of the expansion pyramid module DCPM, and the output end of the expansion pyramid module DCPM constitutes the main output end of the encoder network Encoder; the output end of the first encoding submodule in the sequential series in the encoder network Encoder is connected to the input end of the multi-scale pooling module MSPM; the structures of the encoding submodules are the same, and the encoding submodules are respectively connected in series with a maximum pooling layer and a convolution block in the direction from the input end to the output end, wherein the input end of the maximum pooling layer constitutes the input end of the encoding submodule, and the output end of the convolution block constitutes the output end of the encoding submodule; The decoder network Decoder includes four decoding submodules connected in series from the input end to the output end, as well as a transposed convolution layer, a connection layer, a 3x3 convolution layer, a batch normalization and a linear activation combination layer, wherein the output end of the batch normalization and linear activation combination layer constitutes the output end of the decoder network Decoder, and the input end of the first decoding submodule in the series connection constitutes the input end of the decoder network Decoder; the structures of each decoding submodule are the same, and each decoding submodule is connected in series from the input end to the output end, wherein the input end of the transposed convolution layer constitutes the input end of the decoding submodule, and the output end of the convolution block constitutes the output end of the decoding submodule; The output of the batch normalization and linear activation combination layer in the encoder network Encoder is connected to the input of the last connection layer in the decoder network Decoder in series. The four outputs of the multi-scale pooling module MSPM in the encoder network Encoder are connected to the input of the connection layer in the four decoding sub-modules in the decoder network Decoder in one-to-one correspondence. The multi-scale pooling module MSPM outputs four feature maps of different sizes, and the four feature maps are sequentially transmitted to the input of the connection layer in each decoding sub-module in the decoder network Decoder in series from small to large in size; the encoder network Encoder is sequentially connected in series. The output end of the second encoding submodule in the next sequence is connected to the input end of the connection layer of the third decoding submodule in the next sequence in series in the decoder network Decoder, the output end of the third encoding submodule in the next sequence in series in the encoder network Encoder is connected to the input end of the connection layer of the second decoding submodule in the next sequence in series in the decoder network Decoder, the output end of the fourth encoding submodule in the next sequence in series in the encoder network Encoder is connected to the input end of the connection layer of the first decoding submodule in the next sequence in series in the decoder network Decoder, and the main output end of the encoder network Encoder is connected to the input end of the decoder network Decoder.

2. The interference disentanglement method based on a multi-scale pooling dilated pyramid network according to claim 1, characterized in that: In step A, for each real phase image, firstly, a slope is randomly selected along the vertical direction and the horizontal direction to be added to the real phase image to obtain the corresponding interference phase image, then the interference phase image is updated by pixel phase wrapping, and finally, Gaussian noise is added to the interference phase image to obtain the corresponding wrapped phase image; thereby, the wrapped phase images corresponding to each real phase image are obtained.

3. The interference disentanglement method based on multi-scale pooling dilated pyramid network according to claim 2, characterized in that: In step A, for each interference phase image and each pixel position in the interference phase image, the following formula is used: Get the winding phase corresponding to the pixel position Implement pixel phase wrapping update for pixel positions, where φ(x,y) represents the true phase of the (x,y) pixel position in the interferometric phase image, j represents the imaginary unit, exp() represents the exponential function with the natural constant e as the base, and ∠ represents the angle sign. Represents the wrap-around phase at the (x,y) pixel position in the interferometric phase image.

4. The interference disentanglement method based on a multi-scale pooling dilated pyramid network according to claim 1, characterized in that: The multi-scale pooling module MSPM includes four branches corresponding to four feature maps of different sizes. The four branches are respectively connected in series from the input end to the output end, and the average pooling layer, the 1x1 convolution layer, the batch normalization and the linear activation combination layer are respectively connected in series. The pooling sizes of the average pooling layers on the four branches are different from each other. The input ends of the average pooling layers on the four branches are connected to each other to constitute the input end of the multi-scale pooling module MSPM. The output ends of the batch normalization and the linear activation combination layers on the four branches respectively constitute the output ends of the multi-scale pooling module MSPM.

5. The interference disentanglement method based on multi-scale pooling dilated pyramid network according to claim 1, characterized in that: The expansion pyramid module DCPM includes a multiplication fusion layer, a batch normalization and linear activation combination layer, two addition fusion layers, and four branches, wherein the input ends of the branches are connected to form the input end of the expansion pyramid module DCPM, the output end of the second branch and the output end of the third branch are connected to the input end of one of the addition fusion layers, the output end of the addition fusion layer is connected to the input end of the batch normalization and linear activation combination layer, the output end of the batch normalization and linear activation combination layer and the output end of the first branch are connected to the input end of the multiplication fusion layer, the output end of the multiplication fusion layer and the output end of the fourth branch are connected to the input end of another addition fusion layer, and the output end of the addition fusion layer constitutes the output end of the expansion pyramid module DCPM; The first branch is connected in series from the input end to the output end in the direction of the 1x1 convolution layer with a step size of s=1, the 1x1 convolution layer with a step size of s=2, the batch normalization and the linear activation combination layer, wherein the input end of the 1x1 convolution layer with a step size of s=1 constitutes the input end of the first branch, and the output end of the batch normalization and the linear activation combination layer constitutes the output end of the first branch; The second branch includes a 3x3 dilated convolution layer with a dilation coefficient of r=5, a 3x3 dilated convolution layer with a dilation coefficient of r=9, two 1x1 convolution layers with a step size of s=2, and two 3x3 convolution layers with a step size of s=1, wherein the input end of the 3x3 dilated convolution layer with a dilation coefficient of r=5 constitutes the input end of the second branch, and the two output ends of the 3x3 dilated convolution layer with a dilation coefficient of r=5 are respectively connected to the input end of the 3x3 dilated convolution layer with a dilation coefficient of r=9 and one of the 1x1 convolution layers with a step size of s=2. The input end of the 1x1 convolution layer, the output end of the 1x1 convolution layer with a step size of s=2 is connected to the input end of one of the 3x3 convolution layers with a step size of s=1, the output end of the 3x3 dilated convolution layer with a dilation coefficient of r=9 is connected to the input end of another 1x1 convolution layer with a step size of s=2, the output end of the 1x1 convolution layer with a step size of s=2 is connected to the input end of another 3x3 convolution layer with a step size of s=1, and the output ends of the two 3x3 convolution layers with a step size of s=1 constitute the output end of the second branch; The third branch is connected in series from the input end to the output end in the direction of the 3x3 dilated convolution layer with a dilation coefficient of r=9, the 3x3 dilated convolution layer with a dilation coefficient of r=11, the 1x1 convolution layer with a step size of s=2, and the 3x3 convolution layer with a step size of s=1, wherein the input end of the 3x3 dilated convolution layer with a dilation coefficient of r=9 constitutes the input end of the third branch, and the output end of the 3x3 convolution layer with a step size of s=1 constitutes the output end of the third branch; The fourth branch connects a 1x1 convolution layer with a step size of s=1, a global average pooling layer, a deformation layer, and a 1x1 convolution layer with a step size of s=1 in series from the input end to the output end; among them, the input end of the first 1x1 convolution layer with a step size of s=1 constitutes the input end of the fourth branch, and the output end of the last 1x1 convolution layer with a step size of s=1 constitutes the output end of the fourth branch.

6. The interference disentanglement method based on a multi-scale pooling dilated pyramid network according to any one of claims 1 to 5, characterized in that: The structures of each convolution block are the same. Each convolution block connects a 1x1 convolution layer, two 3x3 convolution layers, and a batch normalization and linear activation combination layer in series from the input end to the output end, where the input end of the 1x1 convolution layer constitutes the input end of the convolution block, and the output end of the batch normalization and linear activation combination layer constitutes the output end of the convolution block.

7. The interference disentanglement method based on multi-scale pooling dilated pyramid network according to claim 1, characterized in that: In the process of training the network to be trained in step C, the Adam gradient optimization algorithm is used to optimize the network training, and the loss function used in the optimization process is as follows: Where (x, y) represents the pixel position in the phase image corresponding to the training sample, L(y′(x, y)) represents the loss result of the pixel position (x, y) in the wrapped phase image in the training sample, y′(x, y) represents the difference between the true phase and the predicted phase of the pixel position (x, y) in the wrapped phase image in the training sample, and b represents the preset positive coefficient.

8. The interference disentanglement method based on multi-scale pooling dilated pyramid network according to claim 1, characterized in that: It also includes a phase image unwrapping model corresponding to the synthetic aperture radar of the target type, and performs evaluation according to the following formula; Where NRMSE is the normalized root mean square error, k = 1, 2, ..., N, N represents the number of training samples in the training process of the network to be trained, (x, y) represents the pixel position in the phase image corresponding to the training sample, and y′ k (x,y) represents the difference between the true phase and the predicted phase at the pixel position (x,y) in the wrapped phase image in the kth training sample.

Images