Multi-scale complex feature extraction method and system based on lightweight two-step filter network

Abstract

The invention discloses a multi-scale complex feature extraction method and system based on a lightweight two-step filter network. The method comprises the following steps: S1, constructing a filter training data set based on a digital elevation model and Gaussian noise; s2, constructing a multi-scale complex feature extraction network model of lightweight two-step filtering; s3, training a multi-scale complex feature extraction network model by using the filtering training data set; and S4, performing phase filtering processing on the terrain interferogram by using the trained multi-scale complex feature extraction network model to complete feature extraction. According to the method, lightweight U-Net and depth separable convolution are introduced, so that the parameter quantity and calculation complexity of the model are greatly reduced, and meanwhile, a two-step filtering mechanism based on complex conjugate operation is combined, so that refined extraction and optimization of complex features are realized; and through the residual attention module and the multi-scale feature fusion module, the cross-scale feature expression capability of the network is enhanced, and the robustness and anti-noise performance of the model are improved.

Description

Technical Field

[0001] The present invention relates to the field of synthetic aperture radar interference data processing, and particularly to a multi-scale complex feature extraction method and system based on a lightweight two-step filtering network. Background Art

[0002] Currently, in the field of radar signal processing, filtering techniques are mainly used to remove noise, enhance image quality, and extract key features. However, the existing technologies have deficiencies such as high computational complexity, insufficient multi-scale feature extraction, and lack of feature enhancement mechanism.

[0003] High computational complexity: Traditional filtering methods, such as Fourier transform filtering and wavelet transform filtering, although they can operate in the complex domain, have high computational complexity. Especially when processing high-resolution data, it is difficult to meet the requirements of real-time processing.

[0004] Deep learning-based networks (such as the standard U-Net) although perform well in image restoration tasks, but they have a large number of parameters and high computational costs, and it is difficult to be applied to resource-constrained devices.

[0005] Insufficient multi-scale feature extraction: When existing convolutional neural networks extract multi-scale features, they mostly rely on downsampling and upsampling operations with fixed resolutions. This single-scale information may lead to insufficient robustness to small targets or detailed features.

[0006] The fusion of multi-resolution features is usually completed through simple splicing or weighted operations, and it is difficult to make full use of the context information at different resolutions.

[0007] Lack of feature enhancement mechanism: In complex domain signal processing, important features are often masked by noise. Traditional convolutional operations are difficult to effectively distinguish and enhance these important features. Although some methods attempt to introduce an attention mechanism, the implementation is complex and the parameter overhead is large, which is not suitable for lightweight application scenarios. Summary of the Invention

[0008] To solve the technical problems in the above background, the present invention provides a two-step filtering network based on lightweight parameters for efficient filtering and feature extraction of complex domain signals, aiming to solve the problems of high computational complexity, insufficient utilization of multi-scale features, and complex implementation of the attention mechanism existing in the existing filtering methods. While maintaining high efficiency and light weight, the present invention significantly improves the filtering accuracy and robustness.

[0009] To achieve the above object, the present invention provides a multi-scale complex feature extraction method based on a lightweight two-step filtering network, and the steps include:

[0010] S1. Construct a filtering training dataset based on a digital elevation model and Gaussian noise;

[0011] S2. Construct a multi-scale complex feature extraction network model with lightweight two-step filtering;

[0012] S3. Train the multi-scale complex feature extraction network model using the filtered training dataset;

[0013] S4. Use the trained multi-scale complex feature extraction network model to perform phase filtering on the terrain interferogram and complete feature extraction.

[0014] Preferably, step S1 includes:

[0015] Convert the elevation to the wrapped terrain interference phase by performing phase-height conversion and phase wrapping on the digital elevation model;

[0016] Add noise to the converted wrapped terrain interference phase;

[0017] Construct the filtered training dataset by mapping the real and imaginary parts of the wrapped terrain interference phase.

[0018] Preferably, the constructed multi-scale complex feature extraction network model includes: a U-Net module, a residual attention module, and an adaptive multi-scale feature fusion module;

[0019] The lightweight U-Net module is used to extract multi-scale prior features of complex signals;

[0020] The residual attention module is used to enhance the focusing ability on key features;

[0021] The adaptive multi-scale feature fusion module is used to integrate information of different resolutions.

[0022] Preferably, the lightweight U-Net module reduces model parameters and computational complexity by introducing depthwise separable convolutions, and includes: a depthwise separable convolution module, a multi-level residual block, and an upsampling module;

[0023] The depthwise separable convolution module decomposes the convolution operation into depthwise convolution and pointwise convolution, and processes the spatial features of each channel and the feature relationships between channels respectively, for extracting local features;

[0024] The number of parameters is:

[0025] D K *D K *M + M * N

[0026] The computational complexity is:

[0027] D K *D K *M * D F *DF +M*N*D F *D F

[0028] Among them, the convolution kernel size is D K *D K ; The number of input channels is M, the number of output channels is N, and the size of the output feature map is D F *D F ;

[0029] The multi-level residual block is used to combine the input features with the transformed residual features, enhancing the feature expression ability while retaining the original feature information;

[0030] The upsampling module performs progressive upsampling on the low-resolution features through deconvolution operations, and combines cross-layer feature splicing to achieve the integration of multi-layer features.

[0031] Preferably, the residual attention module is used to enhance the channel attention of the input feature map, dynamically adjust the feature weights based on statistical characteristics, and improve the robustness of feature representation; at the same time, avoid overfitting and numerical instability, and retain the original information to enhance the feature expression ability, including: a channel statistics calculation unit, an attention weight generation unit, and a residual enhancement unit;

[0032] The attention weight generation unit calculates and generates weights based on the channel mean μ and unbiased variance σ of the feature map 2 The calculation formula for generating weights is:

[0033]

[0034] Among them, x is the input feature map; H and W are the height and width of the feature map respectively; i and j represent the row and column indices of the feature map;

[0035] The adaptive attention weight is calculated by the following formula:

[0036]

[0037] Among them, δ is the smoothing term;

[0038] Finally, weighted enhancement is performed by combining the residual path and the attention features.

[0039] Preferably, the multi-scale feature fusion module integrates feature information of different resolutions, generates a fusion feature map containing global context information for optimizing the final output; at the same time, establishes the connection between different resolutions to improve the cross-scale integration ability, and realizes channel alignment through 1×1 to avoid information loss or redundancy, including: a multi-resolution pooling unit, a channel alignment convolution unit, and a fusion convolution unit;

[0040] The multi-resolution pooling unit downsamples the feature maps F at different levels l to extract multi-scale features:

[0041] F l,pool = MaxPool(F l )

[0042] where F l,pool represents the l-th feature map after max pooling; MaxPool represents the max pooling operation; F l represents the l-th feature map;

[0043] The channel alignment convolution unit performs channel alignment on the pooled feature maps through 1×1 convolution, and splices and fuses the features of different resolutions.

[0044] Preferably, the multi-scale complex feature extraction network model further includes a two-step filtering unit, which is used to gradually optimize the input data through complex conjugate filtering and multi-scale fusion to generate a high-precision complex feature output; the steps include: the first step is to perform rough optimization using the prior phase, and the second step is to refine by combining multi-scale information to improve the accuracy of the filtering result.

[0045] The present invention also provides a multi-scale complex feature extraction system based on a lightweight two-step filtering network, which is used to implement the above method, including: a dataset construction module, a model construction module, a training module, and an extraction module;

[0046] The dataset construction module is used to construct a filtering training dataset based on the digital elevation model and Gaussian noise;

[0047] The model construction module is used to construct a multi-scale complex feature extraction network model with lightweight two-step filtering;

[0048] The training module is used to train the multi-scale complex feature extraction network model using the filtering training dataset;

[0049] The extraction module is used to perform phase filtering on the terrain interferogram using the trained multi-scale complex feature extraction network model to complete feature extraction.

[0050] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0051] By introducing a lightweight U-Net and depthwise separable convolutions, the present invention significantly reduces the number of model parameters and computational complexity. Meanwhile, combined with a two-step filtering mechanism based on complex conjugate operations, it realizes the refined extraction and optimization of complex features. Through the residual attention module and multi-scale feature fusion module, the cross-scale feature expression ability of the network is enhanced, and the robustness and anti-noise performance of the model are improved. Experiments show that while ensuring high efficiency and lightweight, the present invention can significantly improve the extraction accuracy of complex features and is applicable to efficiently extracting the terrain phase in interferograms and filtering out phase redundant information and noise. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] To more clearly illustrate the technical solutions of the present invention, the drawings required for use in the embodiments are briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0053] Figure 1 It is a schematic flowchart of the method in the embodiment of the present invention.

[0054] Figure 2 It is a schematic diagram of the simulated data used in the embodiment of the present invention and its corresponding noisy phase;

[0055] Figure 3 It is a schematic diagram of the network structure of the multi-scale complex feature extraction method based on a lightweight two-step filtering network in the embodiment of the present invention;

[0056] Figure 4 It is a schematic diagram of the terrain wrapped phase, noisy data, and the filtering results of the noisy data after Boxfilter, Lee filter, Goldstein filter, ML filter, and the method of the present invention, as well as their corresponding filtering differences. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0057] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of them. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0058] To make the above objects, features, and advantages of the present invention more obvious and understandable, the present invention will be further described in detail below in conjunction with the drawings and specific embodiments.

[0059] Embodiment 1

[0060] As Figure 1As shown in the figure, it is a schematic flowchart of the method of this embodiment, and the steps include:

[0061] S1. Construct a filtering training data set based on the digital elevation model and Gaussian noise.

[0062] By performing phase height conversion and phase wrapping processing on the SRTM 90m DEM, the terrain-wrapped terrain interference phase is simulated. The phase height conversion and phase wrapping are shown in the following formula:

[0063]

[0064] Among them, H is the elevation of the SRTM 90m DEM; λ is the wavelength of the SAR satellite; B ⊥ is the vertical baseline length; R is the slant range; θ is the radar side view angle; ψ represents the absolute phase; exp(j·) is the Euler formula operation; angle(·) is the calculation of the radian value of the phase angle of the complex matrix; is the wrapped phase.

[0065] The noise addition process is to perform Gaussian noise addition on the terrain-wrapped terrain interference phase simulated by the above DEM. The two-dimensional Gaussian noise can be expressed as:

[0066]

[0067] In the formula, σ 2 is the variance of the two-dimensional Gaussian noise model, and μ is the mean of the two-dimensional Gaussian noise model.

[0068] The real part and the imaginary part of the simulated terrain-wrapped phase are:

[0069]

[0070] Among them, is the terrain-wrapped phase after being added with two-dimensional Gaussian noise; cos(ψ) is the real part of the complex terrain-wrapped phase; sin(ψ) is the imaginary part; exp(j·) is the Euler formula operation; is the real part of the noise-added complex terrain-wrapped phase; is the corresponding imaginary part; The network model training data set is constructed by mapping the real part and the imaginary part of the complex terrain-wrapped phase before and after noise addition.

[0071] S2. Construct a multi-scale complex feature extraction network model with lightweight two-step filtering.

[0072] The network model constructed in this embodiment uses the AdamW optimizer and the L1Loss as the model loss function. Its structure includes: a U-Net module, a residual attention module, and an adaptive multi-scale feature fusion module; the lightweight U-Net module is used to extract multi-scale prior features of complex signals; the residual attention module is used to enhance the focusing ability on key features; the adaptive multi-scale feature fusion module is used to integrate information of different resolutions.

[0073] Specifically, the lightweight U-Net module reduces the model parameters and computational amount by introducing depthwise separable convolutions, and includes: a depthwise separable convolution module, a multi-level residual block, and an upsampling module;

[0074] The depthwise separable convolution module decomposes the convolution operation into a depthwise convolution and a pointwise convolution, and processes the spatial features of each channel and the feature relationship between channels respectively, significantly reducing the model parameter amount and computational complexity, and is used to efficiently extract local features;

[0075] Its parameter amount is:

[0076] D K *D K *M + M * N(6)

[0077] The computational amount is:

[0078] D K *D K *M * D F *D F + M * N * D F *D F (7)

[0079] Among them, the convolution kernel size is D K *D K ; the input channels are M, the output channels are N, and the size of the output feature map is D F *D F ;

[0080] The multi-level residual block is used to combine the input features with the transformed residual features, enhancing the feature expression ability while retaining the original feature information; the upsampling module performs progressive upsampling on the low-resolution features through deconvolution operations, and realizes the integration of multi-layer features by combining cross-layer feature splicing.

[0081] The residual attention module is used to enhance the channel attention of the input feature map, and includes: a channel statistics calculation unit, an attention weight generation unit, and a residual enhancement unit.

[0082] Specifically, the channel statistical calculation unit dynamically adjusts the feature weights based on statistical characteristics to enhance the robustness of feature representation. At the same time, the residual enhancement unit is used to avoid overfitting and numerical instability, and retain the original information to enhance the feature expression ability;

[0083] (Based on mean and variance) The attention weight generation unit calculates and generates weights by calculating the channel mean μ and unbiased variance ρ of the feature map 2 The formula is:

[0084]

[0085] where x is the input feature map; H and W are the height and width of the feature map respectively; i and j represent the row and column indices of the feature map;

[0086] The adaptive attention weight is calculated by the following formula:

[0087]

[0088] where δ is the smoothing term, defaulting to e -5 ;

[0089] Finally, the residual path and the attention feature are combined for weighted enhancement.

[0090] The multi-scale feature fusion module integrates feature information for different resolutions, including: the multi-resolution pooling unit, the channel alignment convolutional unit, and the fusion convolutional unit.

[0091] Specifically, the fusion convolutional unit generates a fusion feature map containing global context information for optimizing the final output; at the same time, the channel alignment convolutional unit establishes the connection between different resolutions to improve the cross-scale integration ability, and realizes channel alignment through 1×1 to avoid information loss or redundancy.

[0092] The multi-resolution pooling unit downsamples the feature maps F at different levels l to extract multi-scale features:

[0093] F l,pool = MaxPool(F l ) (10)

[0094] In the formula, F l,pool represents the l-th feature map after max pooling; MaxPool represents the max pooling operation; F l represents the l-th feature map;

[0095] The channel alignment convolutional unit aligns the channels of the pooled feature map through 1×1 convolution, and splices and fuses the features of different resolutions.

[0096] In this embodiment, the multi-scale complex feature extraction network model further includes a two-step filtering unit, which is used to gradually optimize the input data through complex conjugate filtering and multi-scale fusion to generate a high-precision complex feature output. The steps include: the first step is to perform rough optimization using the prior phase, and the second step is to refine by combining multi-scale information to improve the accuracy of the filtering result. The network structure of this embodiment is as Figure 3 shown.

[0097] S3. Train the multi-scale complex feature extraction network model using the filtered training data set.

[0098] Use the real and imaginary parts of the complex terrain wrapped phase before and after adding noise simulated in S1 as the training data set, and train the multi-scale complex feature extraction model based on the lightweight two-step filtering network constructed in S2 with the network training parameters described in S2.

[0099] S4. Use the trained multi-scale complex feature extraction network model to perform phase filtering on the terrain interferogram to complete feature extraction.

[0100] Embodiment 2

[0101] To verify the technical effects of the present invention, in this embodiment, the Boxfilter, Goldstein filtering, Lee filtering, ML filtering, and the filtering method of the present invention are respectively used to perform phase filtering processing experiments on the same interferogram. The simulated data used in the experiment and its corresponding noisy interferogram are as Figure 2 shown. The results after filtering the simulated interferogram A by the Boxfilter, Goldstein filtering, Lee filtering, ML filtering, and the filtering method of the present invention, and the filtering results and filtering differences of each method are as Figure 4As shown in the figure; for the simulated data, the number of residual points of the Boxfilter filtering result is 143, the RMSE is 1.6772 rad, and the PSNR is 43.6390; the number of residual points of the Goldstein filtering result is 228, the RMSE is 1.5332 rad, and the PSNR is 44.4187; the number of residual points of the Lee filtering result is 136, the RMSE is 1.3886 rad, and the PSNR is 45.2791; the number of residual points of the ML filtering result is 34, the RMSE is 1.3634 rad, and the PSNR is 45.4386; the number of residual points of the filtering result of the method of the present invention is 19, the RMSE is 1.2683 rad, and the PSNR is 46.0664. The results are better than those of other methods. From the filtering difference diagrams of various methods, it can be found that the method of the present invention has better ability to retain phase details, and in the edge region of the phase fringes, the error of the method of the present invention is smaller, which is better than other methods. It can be clearly seen from the residual point distribution diagram that the number of residual point distributions of the method of the present invention is greatly reduced, especially in the areas where the interference fringes are dense and the terrain is steep. In summary, it shows that the phase filtering method of the present invention can effectively and efficiently extract the terrain phase in the real terrain interferogram and filter out the phase redundant information and noise.

[0102] Embodiment 3

[0103] This embodiment also provides a multi-scale complex feature extraction system based on a lightweight two-step filtering network, including: a data set construction module, a model construction module, a training module, and an extraction module; the data set construction module is used to construct a filtering training data set based on the digital elevation model and Gaussian noise; the model construction module is used to construct a multi-scale complex feature extraction network model with lightweight two-step filtering; the training module is used to train the multi-scale complex feature extraction network model by using the filtering training data set; the extraction module is used to perform phase filtering processing on the terrain interferogram by using the trained multi-scale complex feature extraction network model to complete feature extraction.

[0104] Next, in combination with this embodiment, it will be described in detail how the present invention solves the technical problems in actual work.

[0105] First, use the data set construction module to construct a filtering training data set based on the digital elevation model and Gaussian noise.

[0106] By performing phase-height conversion and phase wrapping processing on the DEM of SRTM 90m, the terrain-wrapped terrain interference phase is simulated. The phase-height conversion and phase wrapping are shown in the following formula:

[0107]

[0108] Among them, H is the elevation of the DEM of SRTM 90m; λ is the wavelength of the SAR satellite; B ⊥is the vertical baseline length; R is the slant range; θ is the radar side view angle; ψ represents the absolute phase; exp(j·) is the Euler's formula operation; angle(·) is the calculation of the radian value of the phase angle of the complex matrix; is the wrapped phase.

[0109] The noise addition process is to perform Gaussian noise addition on the terrain-wrapped terrain interference phase simulated by the above DEM. The two-dimensional Gaussian noise can be expressed as:

[0110]

[0111] In the formula, σ 2 is the variance of the two-dimensional Gaussian noise model, and μ is the mean of the two-dimensional Gaussian noise model. The real part and the imaginary part of the simulated terrain-wrapped phase are:

[0112]

[0113] where is the terrain-wrapped phase after adding two-dimensional Gaussian noise; cos(ψ) is the real part of the complex terrain-wrapped phase; sin(ψ) is the imaginary part; exp(j·) is the Euler's formula operation; is the real part of the noise-added complex terrain-wrapped phase; is the corresponding imaginary part; The real part and the imaginary part of the complex terrain-wrapped phase before and after adding noise are used to construct the network model training dataset by mapping.

[0114] The model construction module is used to construct a multi-scale complex feature extraction network model with lightweight two-step filtering.

[0115] The network model constructed in this embodiment uses the AdamW optimizer and takes the L1Loss as the model loss function. Its structure includes: U-Net module, residual attention module and adaptive multi-scale feature fusion module; The lightweight U-Net module is used to extract multi-scale prior features of complex signals; The residual attention module is used to enhance the focusing ability on key features; The adaptive multi-scale feature fusion module is used to integrate information of different resolutions.

[0116] Specifically, the lightweight U-Net module reduces the model parameters and computational amount by introducing depthwise separable convolution, and includes: depthwise separable convolution module, multi-level residual block and upsampling module;

[0117] The depthwise separable convolution module decomposes the convolution operation into depth convolution and point convolution, and processes the spatial features of each channel and the feature relationship between channels respectively, significantly reducing the model parameter amount and computational complexity, and is used to efficiently extract local features;

[0118] Its parameter amount is:

[0119] DK *D K *M + M * N(16)

[0120] The computational amount is:

[0121] D K *D K *M * D F *D F +M * N * D F *D F (17)

[0122] where the convolution kernel size is D K *D K ; the number of input channels is M, the number of output channels is N, and the size of the output feature map is D F *D F ;

[0123] The multi-level residual block is used to combine the input features with the transformed residual features, enhancing the feature expression ability while retaining the original feature information; the upsampling module performs progressive upsampling on the low-resolution features through deconvolution operations and combines cross-layer feature splicing to achieve the integration of multi-layer features.

[0124] The residual attention module is used to enhance the channel attention of the input feature map, including: a channel statistics calculation unit, an attention weight generation unit, and a residual enhancement unit.

[0125] Specifically, the channel statistics calculation unit dynamically adjusts the feature weights based on statistical characteristics to improve the robustness of feature representation. At the same time, the residual enhancement unit is used to avoid overfitting and numerical instability, and retain the original information to achieve the enhancement of feature expression ability.

[0126] (Based on mean and variance) The attention weight generation unit calculates and generates weights through the channel mean μ and unbiased variance ρ of the feature map 2 The formula is:

[0127]

[0128] where x is the input feature map; H and W are the height and width of the feature map respectively; i and j represent the row and column indices of the feature map;

[0129] The adaptive attention weight is through the following formula:

[0130]

[0131] where δ is the smoothing term, defaulting to e -5 ;

[0132] Finally, the residual path and the attention features are combined for weighted enhancement.

[0133] The multi-scale feature fusion module integrates feature information for different resolutions, including: a multi-resolution pooling unit, a channel alignment convolutional unit, and a fusion convolutional unit.

[0134] Specifically, the fusion convolutional unit generates a fused feature map containing global context information for optimizing the final output; meanwhile, the channel alignment convolutional unit establishes connections between different resolutions to improve the cross-scale integration ability, and achieves channel alignment through 1×1 convolution to avoid information loss or redundancy.

[0135] The multi-resolution pooling unit downsamples the feature maps F at different levels l to extract multi-scale features:

[0136] F l,pool = MaxPool(F l )(20)

[0137] In the formula, F l,pool represents the l-th feature map after max pooling; MaxPool represents the max pooling operation; F l represents the l-th feature map.

[0138] The channel alignment convolutional unit performs channel alignment on the pooled feature maps through 1×1 convolution, and splices and fuses the features of different resolutions.

[0139] In this embodiment, the multi-scale complex feature extraction network model further includes a two-step filtering unit, which is used to gradually optimize the input data through complex conjugate filtering and multi-scale fusion to generate a high-precision complex feature output; the steps include: the first step is to perform rough optimization using the prior phase, and the second step is to refine by combining multi-scale information to improve the accuracy of the filtering result.

[0140] After that, the training module uses the filtered training data set to train the multi-scale complex feature extraction network model.

[0141] Taking the real and imaginary parts of the complex terrain wrapped phase before and after adding noise simulated in S1 as the mapping for the training data set, the multi-scale complex feature extraction model based on the lightweight two-step filtering network constructed in S2 is trained with the network training parameters described in S2.

[0142] Finally, the extraction module uses the trained multi-scale complex feature extraction network model to perform phase filtering on the terrain interferogram to complete feature extraction.

[0143] The embodiments described above are only descriptions of the preferred embodiments of the present invention, and do not limit the scope of the present invention. Without departing from the design spirit of the present invention, various deformations and improvements made by those of ordinary skill in the art to the technical solutions of the present invention shall fall within the protection scope determined by the claims of the present invention.

Claims

1. A multi-scale complex feature extraction method based on a lightweight two-step filtering network, characterized in that the steps Including: S1. Construct a filtering training data set based on a digital elevation model and Gaussian noise; S2. Construct a multi-scale complex feature extraction network model for lightweight two-step filtering; S3. Train the multi-scale complex feature extraction network model using the filtering training data set; S4. Use the trained multi-scale complex feature extraction network model to perform phase filtering on the terrain interferogram to complete feature extraction.

2. The multi-scale complex feature extraction method based on a lightweight two-step filtering network according to claim 1, characterized in that, Step S1 includes: By performing phase-height conversion and phase wrapping on the digital elevation model, convert the elevation to the wrapped terrain interference phase; Add noise to the converted wrapped terrain interference phase; Construct the filtering training data set by mapping the real part and the imaginary part of the wrapped terrain interference phase.

3. The multi-scale complex feature extraction method based on a lightweight two-step filtering network according to claim 1, characterized in that The constructed multi-scale complex feature extraction network model includes: a U-Net module, a residual attention module, and an adaptive multi-scale feature fusion module; The lightweight U-Net module is used to extract multi-scale prior features of complex signals; The residual attention module is used to enhance the focusing ability on key features; The adaptive multi-scale feature fusion module is used to integrate information of different resolutions.

4. The multi-scale complex feature extraction method based on a lightweight two-step filtering network according to claim 3, wherein The lightweight U-Net module includes, by introducing depthwise separable convolution: a depthwise separable convolution module, a multi-level residual block, and an upsampling module; The depthwise separable convolution module decomposes the convolution operation into depth convolution and point convolution, and processes the spatial features of each channel and the feature relationship between channels respectively, for extracting local features; The number of parameters is: D K *D K *M + M * N The amount of computation is: D K *D K *M*D F *D F +M*N*D F *D F Among them, the convolution kernel size is D K *D K ; the number of input channels is M, the number of output channels is N, and the size of the output feature map is D F *D F ; The multi-level residual block is used to combine the input features with the transformed residual features, enhancing the feature expression ability while retaining the original feature information; The upsampling module performs progressive upsampling on the low-resolution features through deconvolution operations, and realizes the integration of multi-layer features by combining cross-layer feature splicing.

5. The multi-scale complex feature extraction method based on a lightweight two-step filtering network according to claim 3, characterized in that The residual attention module is used to enhance channel attention on the input feature map, dynamically adjust the feature weights based on statistical characteristics, including: a channel statistics calculation unit, an attention weight generation unit, and a residual enhancement unit; The attention weight generation unit calculates and generates weights based on the channel mean μ and unbiased variance ρ of the feature map. 2 The formula is as follows: Where x is the input feature map; H and W are the height and width of the feature map respectively; i and j represent the row and column indices of the feature map; The adaptive attention weight is calculated by the following formula: Where δ is a smoothing term; Finally, perform weighted enhancement by combining the residual path and the attention features.

6. The multi-scale complex feature extraction method based on the lightweight two-step filtering network according to claim 3, wherein, The multi-scale feature fusion module integrates feature information of different resolutions, generates a fusion feature map containing global context information for optimizing the final output; at the same time, establishes the connection between different resolutions to improve the cross-scale integration ability, and realizes channel alignment through 1×1, including: a multi-resolution pooling unit, a channel alignment convolution unit, and a fusion convolution unit; The multi-resolution pooling unit downsamples the feature maps F at different levels l to extract multi-scale features: F l,pool = MaxPool(F l ) where, F l,pool represents the l-th feature map after max pooling; MaxPool represents the max pooling operation; F l represents the l-th feature map; The channel alignment convolution unit performs channel alignment on the pooled feature map through 1×1 convolution, and fuses the features of different resolutions after splicing.

7. The multi-scale complex feature extraction method based on a lightweight two-step filtering network according to claim 1, wherein The multi-scale complex feature extraction network model further includes a two-step filtering unit, which is used to gradually optimize the input data through complex conjugate filtering and multi-scale fusion to generate a high-precision complex feature output. The steps include: the first step is to perform rough optimization using the prior phase, and the second step is to refine by combining multi-scale information.

8. A multi-scale complex feature extraction system based on a lightweight two-step filtering network, the system being used to implement the method according to any one of claims 1-7, characterized in that, Including: a dataset construction module, a model construction module, a training module, and an extraction module; The dataset construction module is used to construct a filtering training dataset based on the digital elevation model and Gaussian noise; The model construction module is used to construct a multi-scale complex feature extraction network model with lightweight two-step filtering; The training module is used to train the multi-scale complex feature extraction network model using the filtering training dataset; The extraction module is used to perform phase filtering on the terrain interferogram using the trained multi-scale complex feature extraction network model to complete feature extraction.

Images