InSAR phase unwrapping method based on multi-scale FiLM conditional generative adversarial network

Abstract

The present application relates to the technical field of interferometric synthetic aperture radar, and particularly relates to an InSAR phase unwrapping method based on a multi-scale FiLM conditional generative adversarial network, comprising the following steps: S1: obtaining a wrapped phase map to be processed; S2: calculating the fringe density, jump number and average coherence of the wrapped phase map to be processed, and generating a global statistical condition vector; S3: inputting the wrapped phase map to be processed and the corresponding global statistical condition vector into a generator of a trained multi-scale feature linear modulation conditional generative adversarial network, and outputting a corresponding generated unwrapped phase map. The generator can realize adaptive condition injection at the network level by using fringe density statistics, and can optimize the generation process by combining coherence weighting and a physical prior composite loss, so as to ensure the accuracy of unwrapping under complex fringe density, and to improve the point-by-point accuracy while maintaining the structure and perception quality through physical constraints.

Description

Technical Field

[0001] This invention relates to the field of interferometric synthetic aperture radar technology, specifically to an InSAR phase unwrapping method based on a multi-scale FiLM conditional generative adversarial network. Background Technology

[0002] Interferometric Synthetic Aperture Radar (InSAR) technology offers advantages such as all-weather, all-time coverage, high spatiotemporal resolution, and wide-area monitoring. It has been widely applied in digital elevation model (DEM) reconstruction and monitoring of surface deformation, including earthquakes, volcanoes, landslides, debris flows, mine subsidence, and urban and railway line deformation. The basic principle of InSAR is to separate the target signal from the phase difference (interferometric phase) of two or more mutually registered SAR images, thereby obtaining elevation or deformation information.

[0003] Due to limitations of the imaging hardware system, the actual obtained interferometric phase is wrapped around the principal value range. , [Within.] Obtaining the absolute phase from a wrapped phase is an ill-posed inverse problem: infinitely many absolute phases correspond to the same wrapped phase, therefore a unique solution cannot be obtained directly, and additional observation information or constraints must be introduced. Traditional methods typically rely on the phase continuity assumption (Itoh condition, i.e., the true phase difference between adjacent pixels does not exceed...) The methods can be categorized into three types, each with its own distinct drawbacks: 1) Path tracking methods (such as Goldstein's branching method, quality-guided method, and region growing method): These methods find suitable integration paths through local gradient integration, but their reliability is poor in regions with dense residual points and high noise, and errors are easily propagated along the path; 2) Minimum norm methods (such as least squares method and minimum cost flow MCF): These methods transform untangling into a global optimization problem, which is highly efficient, but they are prone to over-smoothing of real phase discontinuities (jumps), leading to error propagation and decreased robustness; 3) Bayesian / Kalman filtering methods: These methods rely on accurate statistical noise models, which are often difficult to obtain in practice.

[0004] All of the above traditional paradigms share a common failure mode: no method explicitly bases the untangling process on the estimation of "local fringe complexity (fringe density)," that is, it uses the exact same algorithm logic for dense and sparse fringe regions, lacking adaptability.

[0005] In recent years, deep learning has been introduced into phase unwrapping due to its ability to learn complex priors from data. Its greatest advantage lies in its ability to break free from the constraints of the phase continuity assumption and achieve end-to-end mapping in a data-driven manner. Representative works include PU-Net (U-Net directly regresses phase unwrapping) and PU-GAN (introducing PatchGAN adversarial training to improve output sharpness). However, existing deep learning methods still have the following fundamental shortcomings: 1) Unconditional generator: The generator only takes the entanglement phase as input and applies the same processing logic to entanglement maps with "5 transitions" and "100 transitions". It lacks a mechanism to adaptively adjust the behavior according to the stripe density, which leads to a significant decrease in unentanglement accuracy in high-density stripes, low coherence and steep terrain regions; 2) Single scale supervision: The discriminator often only evaluates one of the global structure or local texture, and it is difficult to take into account both global phase continuity and local texture fidelity at the same time; 3) Lack of physical constraints: The loss function is mostly pure pixel error, and the physical priors of phase unentanglement such as phase gradient consistency and boundary consistency are not encoded into the optimization objective; 4) Accuracy-perception imbalance: GAN-based methods often exhibit an imbalance phenomenon in which pixel accuracy is improved but perceptual quality deteriorates (or vice versa).

[0006] Therefore, this invention proposes an InSAR phase unwrapping method based on a multi-scale characteristic linear modulation (FiLM) conditional generative adversarial network. The aim is to solve the problem of how to explicitly inject global statistical information such as fringe density into the unwrapping network to achieve high-precision phase unwrapping with density adaptation, especially in high-density fringe, low coherence and steep terrain regions to ensure unwrapping accuracy. At the same time, physical constraints can improve point-by-point accuracy while maintaining structural and sensing quality. Summary of the Invention

[0007] To address the shortcomings of the existing technologies, the technical problem to be solved by this invention is: how to provide an InSAR phase unwrapping method based on a multi-scale FiLM conditional generative adversarial network, which can utilize fringe density statistics to achieve adaptive conditional injection across the entire network level, and combine coherence weighting and physical prior composite loss to optimize the generation process, thereby ensuring unwrapping accuracy under complex fringe densities while maintaining structural and sensing quality through physical constraints to improve point-by-point accuracy.

[0008] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0009] InSAR phase unwrapping methods based on multi-scale FiLM conditional generative adversarial networks include:

[0010] S1: Obtain the entanglement phase map to be processed;

[0011] S2: Calculate the fringe density, jump number, and average coherence of the entangled phase map to be processed, and generate a global statistical condition vector;

[0012] S3: Input the entangled phase map to be processed and the corresponding global statistical conditional vector into the generator of the trained multi-scale feature linear modulation conditional generative adversarial network, and output the corresponding generated unentangled phase map;

[0013] The generator includes an encoder with multiple downsampling modules, a bottleneck layer, and a decoder with multiple conditional generative adversarial network decoding blocks; the generator's processing steps include:

[0014] S301: Based on the global statistical conditional vector, the conditional encoder and multi-scale FiLM header generate characteristic linear modulation parameters of the corresponding scale for each conditional generative adversarial network decoding block of the bottleneck layer and decoder.

[0015] S302: The encoder uses multiple cascaded downsampling modules to downsample the wound phase map step by step to obtain the coded feature map; each downsampling module outputs a downsampled feature map of the corresponding scale, and the output of the last downsampling module is used as the coded feature map output by the encoder.

[0016] S303: By using the feature linear modulation parameters of the corresponding scale in the bottleneck layer to conditionally modulate the encoded feature map, a deep feature map is obtained;

[0017] S304: The decoder uses multiple conditional generative adversarial network (GAN) decoding blocks to perform progressive upsampling and conditional modulation on the deep feature map, and uses the output of the last GAN decoding block as the fused modulation feature map; where each GAN decoding block upsamples the input feature map, fuses it with the downsampled feature map of the corresponding scale in the encoder, and then performs conditional modulation using the feature linear modulation parameters of the corresponding scale.

[0018] S305: The fused modulation feature map is processed by the output head to obtain the generated unwrapped phase map.

[0019] Preferably, step S2, the processing step of generating the global statistical condition vector, includes:

[0020] S201: Calculate the horizontal and vertical phase gradients of adjacent pixels in the radian domain of the wound phase map, and statistically analyze the gradient magnitudes exceeding... The number of pixels as the jump number ;

[0021] S202: Based on jump numbers Calculate the stripe density that reflects the jump density per unit pixel. ;

[0022] The calculation formula is:

[0023] ;

[0024] In the formula: and These represent the height and width of the entanglement phase diagram, respectively.

[0025] S203: Wrap the phase diagram Complexity index Within a 5×5 sliding window, mean filtering is applied to both the real and imaginary parts, and the modulus is taken to obtain a pixel-by-pixel coherence map; the average value of the coherence map is calculated to obtain the average coherence. ;

[0026] S204: Regarding stripe density Jump variables and average coherence After normalization, the vectors are concatenated to generate a global statistical conditional vector. ;

[0027] The calculation formula is:

[0028] ;

[0029] In the formula: , , These represent the normalized fringe density, jump number, and average coherence, respectively.

[0030] Preferably, in step S301, the global statistical conditional vector is mapped to shared conditional features by using a two-layer fully connected layer with ReLU activation function through a conditional encoder;

[0031] By employing a multi-scale FiLM split-head approach, multiple independent linear projection heads are used to generate characteristic linear modulation parameters for each scale of the conditional generative adversarial network decoding block in the bottleneck layer and decoder, based on shared conditional features. Each scale outputs a 2×channel vector, which is then split into channel-wise scaling parameters. With bias parameters That is, characteristic linear modulation parameters .

[0032] Preferably, in step S302, the encoder includes a cascaded input layer and four downsampling modules;

[0033] The input layer consists of cascaded 7×7 convolutional layers, batch normalization layers, and ReLU activation functions;

[0034] The downsampling module consists of cascaded 4×4 convolutional layers, batch normalization layers, ReLU activation functions, and CBAM attention modules;

[0035] The encoder's processing steps include: inputting the wound phase map into the input layer to obtain an initial feature map; inputting the initial feature map into four cascaded downsampling modules for step-by-step downsampling; and using the output of the last downsampling module as the encoded feature map output by the encoder.

[0036] Preferably, in step S303, the bottleneck layer includes three cascaded FiLM residual blocks, and a CBAM attention module is connected between two adjacent FiLM residual blocks;

[0037] The FiLM residual block is used to conditionally modulate the input feature map using the feature linear modulation parameters of the corresponding scale to obtain the modulated feature map; the CBAM attention module is used to perform attention enhancement operations on the input feature map.

[0038] The output of the last FiLM residual block is used as the deep feature map of the bottleneck layer output.

[0039] Preferably, in step S304, the decoder includes four cascaded conditional generative adversarial network decoding blocks;

[0040] The conditional generative adversarial network (GAN) decoding block consists of cascaded transposed convolutional layers, batch normalization layers, ReLU activation functions, 3×3 convolutional layers, batch normalization layers, ReLU activation functions, FiLM residual blocks, and CBAM attention modules. The upsampled feature map output from the transposed convolutional layer is first concatenated with the corresponding downsampled feature map from the encoder along the channel dimension, and then fused through a 3×3 convolutional layer. The output of the last GAN decoding block is used as the fused modulation feature map output by the decoder.

[0041] The processing steps for each conditionally generated adversarial network decoding block include:

[0042] S3041: Upsample the input feature map by using a cascaded transposed convolutional layer, batch normalization layer, and ReLU activation function to obtain an upsampled feature map;

[0043] S3042: After concatenating the upsampled feature map with the downsampled feature map of the corresponding scale in the encoder along the channel dimension, the concatenated feature map is obtained by processing through a cascaded 3×3 convolutional layer, batch normalization layer, and ReLU activation function.

[0044] S3043: Linearly modulate the spliced ​​feature map using the feature linear modulation parameters based on the corresponding scale through the FiLM residual block to obtain the modulated feature map;

[0045] S3044: The modulated feature map is enhanced with attention through the CBAM attention module to obtain the output feature map of the conditional generative adversarial network decoding block.

[0046] Preferably, the calculation formula for conditional modulation of the FiLM residual block is as follows:

[0047] ;

[0048] In the formula: This represents the output feature map after FiLM residual block conditional modulation; This represents the input feature map of the FiLM residual block; Represents the characteristic linear modulation parameters. , These represent the channel-by-channel scaling parameter and the offset parameter, respectively.

[0049] Preferably, in step S3, the processing steps for training the multi-scale feature linear modulation conditional generative adversarial network include:

[0050] S311: Obtain a batch of training samples containing the entanglement phase map and the corresponding real unentanglement phase map;

[0051] S312: Calculate the fringe density, jump number, and average coherence of the entangled phase map, and generate a global statistical condition vector;

[0052] S313: Input the entanglement phase diagram and the corresponding global statistical condition vector into the generator, and output the unentanglement phase diagram;

[0053] S314: The entangled phase image is stitched together with the real unentangled phase image and the generated unentangled phase image to form a dual-channel image. The two dual-channel images are then input into the discriminator to obtain the real sample discrimination result and the generated sample discrimination result, respectively.

[0054] S315 calculates the discriminator loss based on the discrimination results of real samples and generated samples, and updates the network parameters of the discriminator through the backpropagation algorithm.

[0055] S316: Calculate the generator's composite loss based on the entangled phase map, the generated unentangled phase map, the true phase map, the coherence map, and the generated sample discrimination results, and update the generator's network parameters through the backpropagation algorithm;

[0056] S317: Repeat steps S311 to S316 to iteratively train the discriminator and generator until convergence or the maximum number of iterations is reached.

[0057] Preferably, in step S314, the processing steps of the discriminator include:

[0058] S3141: Combine the entangled phase map and the generated unentangled phase map to form a dual-channel image;

[0059] S3142: Stitch the entangled phase map and the real unentangled phase map into a real dual-channel image;

[0060] S3143: The generated dual-channel image and the real dual-channel image are sequentially passed through five 4×4 convolutional downsampling layers with a stride of 2 to extract multi-scale features, and then input into a 4×4 convolutional layer with a stride of 1 to obtain the corresponding two local discriminant confidence maps.

[0061] The first convolutional downsampling layer consists of cascaded 4×4 convolutional layers, LeakyReLU activation function, and CBAM attention module; the second to fifth convolutional downsampling layers consist of cascaded 4×4 convolutional layers, batch normalization layer, LeakyReLU activation function, and CBAM attention module.

[0062] S3144: Perform true / false judgment on each local block of the two local discriminant confidence maps and generate the judgment result.

[0063] Preferably, in step S316, the formula for calculating the composite loss is:

[0064] (Deleted)

[0065] In the formula: This represents the composite loss of the generator; This indicates the generation of adversarial loss; This represents the coherence-weighted pixel loss; This represents the phase gradient consistency loss; This represents the second-order smoothing loss; This represents the boundary consistency loss; This indicates the weight of the corresponding loss term;

[0066] Specifically:

[0067] (1) Generate adversarial loss

[0068] ;

[0069] In the formula: and These represent the height and width of the phase diagram, respectively. and These represent the indices in the height and width directions of the phase map, respectively; Indicates the generated sample discrimination result in ( , Local confidence level of location discrimination. Indicates the discriminator, This represents the entanglement phase diagram. This indicates the generation of an unwrapped phase diagram;

[0070] (2) Coherence-weighted pixel loss

[0071] ;

[0072] In the formula: express( , Average coherence at location; Indicates the generation of untangled phase diagrams and the true untangling phase diagram exist( , The absolute error of the position. , These represent the generation of untangled phase diagrams. and the true untangling phase diagram exist( , The pixel value at the location;

[0073] (3) Phase gradient consistency loss

[0074] ;

[0075] In the formula: This indicates that the difference will be wrapped back [ , ]; , These represent the first-order difference operators in the horizontal and vertical directions, respectively, used to calculate the gradient of the image; , These represent the generation of untangled phase diagrams. exist( , ) Horizontal and vertical gradients of the position; , These represent the entanglement phase diagrams respectively. exist( , ) Horizontal and vertical gradients of the position;

[0076] (4) Second-order smoothing loss

[0077] ;

[0078] In the formula: , , These represent the horizontal second-order, vertical second-order, and cross second-order difference operators, respectively. , , These represent the generation of untangled phase diagrams. exist( , The horizontal second gradient, vertical second gradient, and cross second gradient of the position;

[0079] (5) Boundary consistency loss

[0080] The four boundary regions are the upper boundary Lower boundary Left boundary Right boundary The loss for a single boundary region is:

[0081] ;

[0082] The boundary consistency loss is the average of the four regions:

[0083] ;

[0084] in: , These represent the generation of untangled phase diagrams. and entanglement phase diagram exist( , The pixel value at the location, and the boundary width of a single boundary region are dynamically determined based on the image size.

[0085] Compared with existing technologies, the InSAR phase unwrapping method based on multi-scale FiLM conditional generative adversarial networks in this invention has the following advantages:

[0086] This invention innovatively employs a multi-scale FiLM fringe density conditional modulation mechanism. By mapping fringe statistics to multi-level FiLM parameters through a shared conditional encoder, it enables scale-by-scale, non-attenuating conditional information injection throughout the encoding-decoding process. Compared to traditional methods that only inject the single conditional information of the entanglement phase, this invention allows the network to dynamically adjust the unwrapping strategy for different fringe density regions. Combining the conditional-skip-connection fusion structure of the FiLM residual block and the UNet decoding block, high-resolution skip-connection features are first fused before conditional modulation is applied, accurately matching density features while fully preserving spatial details. The step-by-step adaptive modulation mechanism effectively suppresses local periodic ambiguity, ensuring point-by-point unwrapping accuracy in complex phase modes.

[0087] This invention co-encodes a composite loss function by combining coherence-weighted pixel loss with physical priors such as phase gradient consistency, second-order smoothing, and boundary consistency. This ensures that the network optimization process synchronously follows physical constraints, avoiding overfitting to noise or disrupting the continuity of the phase structure. Simultaneously, this invention employs a single-scale PatchGAN discriminator combined with a CBAM matching design. With the generator already strongly guided by conditions, this effectively avoids noise supervision interference caused by local ambiguity in 2π areas from a small receptive field discriminator. Through the synergistic effect of these two approaches, the reliability of point-by-point estimation is ensured while maintaining the overall structural integrity and visual perceptual realism of the unwrapped results, achieving a balanced optimization of accuracy and quality. Attached Figure Description

[0088] To make the objectives, technical solutions, and advantages of the invention clearer, the invention will now be described in further detail with reference to the accompanying drawings, wherein:

[0089] Figure 1 Network structure diagram of a generative adversarial network for multi-scale feature linear modulation conditions. Detailed Implementation

[0090] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but only to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0091] The following detailed explanation illustrates the specific implementation methods:

[0092] Example:

[0093] This embodiment discloses an InSAR phase unwrapping method based on a multi-scale FiLM conditional generative adversarial network.

[0094] InSAR phase unwrapping methods based on multi-scale FiLM conditional generative adversarial networks include:

[0095] S1: Obtain the (single-channel) entangled phase map to be processed;

[0096] In this embodiment, a coherence map channel and / or a pixel-by-pixel local stripe density map channel can be added to the single channel of the wound phase to form a 2-channel or 3-channel input, providing explicit spatial priors for the network.

[0097] S2: Calculate the fringe density, jump number, and average coherence of the entangled phase map to be processed, and generate a global statistical condition vector;

[0098] S3: Input the entangled phase map to be processed and the corresponding global statistical conditional vector into the generator of the trained multi-scale feature linear modulation conditional generative adversarial network (Conditional GAN), and output the corresponding (single-channel) generated unentangled phase map.

[0099] The generator backbone adopts a U-Net encoder-decoder symmetrical structure and embeds FiLM conditional modulation and CBAM attention at multiple levels. The generator includes an encoder with multiple downsampling modules, a bottleneck layer, and a decoder with multiple conditional generative adversarial network decoding blocks;

[0100] like Figure 1 As shown, the generator's processing steps include:

[0101] S301: Based on the global statistical conditional vector, the conditional encoder and multi-scale FiLM header generate characteristic linear modulation parameters of the corresponding scale for each conditional generative adversarial network decoding block of the bottleneck layer and decoder.

[0102] S302: The encoder uses multiple cascaded downsampling modules to downsample the wound phase map step by step to obtain the coded feature map; each downsampling module outputs a downsampled feature map of the corresponding scale, and the output of the last downsampling module is used as the coded feature map output by the encoder.

[0103] S303: By using the feature linear modulation parameters of the corresponding scale in the bottleneck layer to conditionally modulate the encoded feature map, a deep feature map is obtained;

[0104] S304: The decoder uses multiple conditional generative adversarial network (GAN) decoding blocks to perform progressive upsampling and conditional modulation on the deep feature map, and uses the output of the last GAN decoding block as the fused modulation feature map; where each GAN decoding block upsamples the input feature map, fuses it with the downsampled feature map of the corresponding scale in the encoder, and then performs conditional modulation using the feature linear modulation parameters of the corresponding scale.

[0105] S305: The fused modulation feature map is processed by the output head to obtain the generated unwrapped phase map.

[0106] In this embodiment, the output header includes cascaded 7×7 convolutional layers with Tanh activation functions.

[0107] To better illustrate the technical solution of the present invention, this embodiment will be described in more detail through the following parts.

[0108] I. Conditional Vector Calculation Module

[0109] In the specific implementation process, the processing steps for the condition vector calculation module to generate the global statistical condition vector include:

[0110] S201: Calculate the horizontal and vertical phase gradients of adjacent pixels in the radian domain of the wound phase map, and statistically analyze the gradient magnitudes exceeding... The number of pixels as the jump number ;

[0111] S202: Based on jump numbers Calculate the stripe density that reflects the jump density per unit pixel. ;

[0112] The calculation formula is:

[0113] ;

[0114] In the formula: and These represent the height and width of the entanglement phase diagram, respectively.

[0115] S203: Wrap the phase diagram Complex index Within a 5×5 sliding window, mean filtering is applied to both the real and imaginary parts, and the modulus is taken to obtain a pixel-by-pixel coherence map; the average value of the coherence map is calculated to obtain the average coherence. ;

[0116] S204: Regarding stripe density Jump variables and average coherence After normalization, the vectors are concatenated to generate a global statistical conditional vector. ;

[0117] The calculation formula is:

[0118] ;

[0119] In the formula: , , These represent the normalized fringe density, jump number, and average coherence, respectively.

[0120] In this embodiment, stripe density Normalization truncated to 0.5, jump variables Normalized to a threshold of 50,000, average coherence It is already in the range [0,1].

[0121] This module uses completely consistent computational logic during the training and inference phases to ensure that the condition distribution is aligned.

[0122] II. Conditional Encoder and Multi-Scale FiLM Divider

[0123] In the specific implementation process, the global statistical conditional vector is mapped to a 128-dimensional shared conditional feature by using two fully connected layers (3→64→128) with ReLU activation function through the conditional encoder;

[0124] By employing a multi-scale FiLM split-head approach, multiple independent linear projection heads are used to generate characteristic linear modulation parameters for each scale of the conditional generative adversarial network decoding block in the bottleneck layer and decoder, based on shared conditional features. Each scale outputs a 2×channel vector, which is then split into channel-wise scaling parameters. With bias parameters That is, characteristic linear modulation parameters When there is no conditional input = =None, this expression automatically degenerates into an identity mapping, thus making it fully compatible with unconditional U-Net.

[0125] In this embodiment, the number of independent linear projection heads in the multi-scale FiLM split head is five, the scale corresponding to the bottleneck layer is 1024 channels, and the decoder includes four scale channels of 512, 256, 128 and 64.

[0126] III. Encoding Path

[0127] In practice, the encoder consists of a cascaded input layer and four downsampling modules;

[0128] The input layer consists of cascaded 7×7 convolutional layers, batch normalization layers, and ReLU activation functions;

[0129] The downsampling module consists of cascaded 4×4 convolutional layers, batch normalization layers, ReLU activation functions, and CBAM attention modules;

[0130] The encoder's processing steps include: inputting the wound phase map into the input layer to obtain an initial feature map; inputting the initial feature map into four cascaded downsampling modules for step-by-step downsampling; and using the output of the last downsampling module as the encoded feature map output by the encoder.

[0131] In this embodiment, the CBAM attention module uses an existing mature module, but it can also be replaced by channel / spatial attention modules such as SE and ECA.

[0132] IV. Bottleneck Layer

[0133] In the specific implementation process, the bottleneck layer includes three cascaded FiLM residual blocks, and a CBAM attention module is connected between two adjacent FiLM residual blocks;

[0134] The FiLM residual block is used to conditionally modulate the input feature map using the feature linear modulation parameters of the corresponding scale to obtain the modulated feature map; the CBAM attention module is used to perform attention enhancement operations on the input feature map.

[0135] The output of the last FiLM residual block is used as the deep feature map of the bottleneck layer output.

[0136] V. Decoding Path

[0137] In practice, the decoder consists of four cascaded conditional generative adversarial network decoding blocks;

[0138] The conditional generative adversarial network (GAN) decoding block consists of cascaded transposed convolutional layers, batch normalization layers, ReLU activation functions, 3×3 convolutional layers, batch normalization layers, ReLU activation functions, FiLM residual blocks, and CBAM attention modules. The upsampled feature map output from the transposed convolutional layer is first concatenated with the corresponding downsampled feature map from the encoder along the channel dimension, and then fused through a 3×3 convolutional layer. The output of the last GAN decoding block is used as the fused modulation feature map output by the decoder.

[0139] The processing steps for each conditionally generated adversarial network decoding block include:

[0140] S3041: Upsample the input feature map by using a cascaded transposed convolutional layer, batch normalization layer, and ReLU activation function to obtain an upsampled feature map;

[0141] S3042: After concatenating the upsampled feature map with the downsampled feature map of the corresponding scale in the encoder along the channel dimension, the concatenated feature map is obtained by processing through a cascaded 3×3 convolutional layer, batch normalization layer, and ReLU activation function.

[0142] S3043: Linearly modulate the spliced ​​feature map using the feature linear modulation parameters based on the corresponding scale through the FiLM residual block to obtain the modulated feature map;

[0143] S3044: The modulated feature map is enhanced with attention through the CBAM attention module to obtain the output feature map of the conditional generative adversarial network decoding block.

[0144] Specifically, the calculation formula for conditional modulation of the FiLM residual block is as follows:

[0145] ;

[0146] In the formula: This represents the output feature map after FiLM residual block conditional modulation; This represents the input feature map of the FiLM residual block; Represents the characteristic linear modulation parameters. , These represent the channel-by-channel scaling parameter and the offset parameter, respectively.

[0147] VI. Model Training

[0148] In practice, the steps for training a multi-scale feature linear modulation conditional generative adversarial network include:

[0149] S311: Obtain a batch of training samples containing the entanglement phase map and the corresponding real unentanglement phase map;

[0150] S312: Calculate the fringe density, jump number, and average coherence of the entangled phase map, and generate a global statistical condition vector;

[0151] S313: Input the entanglement phase diagram and the corresponding global statistical condition vector into the generator, and output the unentanglement phase diagram;

[0152] S314: The entangled phase image is stitched together with the real unentangled phase image and the generated unentangled phase image to form a dual-channel image. The two dual-channel images are then input into the discriminator to obtain the real sample discrimination result and the generated sample discrimination result, respectively.

[0153] S315 calculates the discriminator loss based on the discrimination results of real samples and generated samples, and updates the network parameters of the discriminator through the backpropagation algorithm.

[0154] In this embodiment, the discriminator loss can be achieved using the existing least-squares generative adversarial network loss.

[0155] S316: Calculate the generator's composite loss based on the entangled phase map, the generated unentangled phase map, the real phase map, and the generated sample discrimination results, and update the generator's network parameters through the backpropagation algorithm;

[0156] S317: Repeat steps S311 to S316 to iteratively train the discriminator and generator until convergence or the maximum number of iterations is reached.

[0157] Specifically, the discriminator's processing steps include:

[0158] S3141: Combine the entangled phase map and the generated unentangled phase map to form a dual-channel image;

[0159] S3142: Stitch the entangled phase map and the real unentangled phase map into a real dual-channel image;

[0160] S3143: The generated dual-channel image and the real dual-channel image are sequentially passed through five 4×4 convolutional downsampling layers with a stride of 2 to extract multi-scale features, and then input into a 4×4 convolutional layer with a stride of 1 to obtain the corresponding two local discriminant confidence maps.

[0161] The first convolutional downsampling layer consists of cascaded 4×4 convolutional layers, LeakyReLU activation function, and CBAM attention module; the second to fifth convolutional downsampling layers consist of cascaded 4×4 convolutional layers, batch normalization layer, LeakyReLU activation function, and CBAM attention module.

[0162] S3144: Perform true / false discrimination on each local block (effective receptive field approximately 70×70) of the two local discrimination confidence maps and generate discrimination results.

[0163] Compared to a dual-scale scheme that uses both a large-scale discriminator and a small-scale discriminator, the optimal solution of this invention retains only a single-scale discriminator: because a small receptive field discriminator cannot distinguish "phase difference of 2" within a local range of 35×35. The "reasonable value" and "true value" will introduce systematic noise supervision; when the generator has been strongly guided by multi-scale FiLM, removing the small-scale discriminator can obtain higher point-by-point accuracy.

[0164] Specifically, the formula for calculating the composite loss is as follows:

[0165] ;

[0166] In the formula: This represents the composite loss of the generator; This represents the adversarial loss generated, calculated based on the discriminator's discrimination results, where the target is to deceive the discriminator. Represents the coherence-weighted pixel loss, which is weighted by the coherence map based on the absolute error between the generated phase and the true value. This represents the phase gradient consistency loss; the difference between the forced unwrapping phase gradient and the wrapped phase gradient is 2. Integer multiples of; It represents a second-order smoothing loss, which penalizes the second-order curvature of the unwrapped phase and suppresses non-physical, drastic fluctuations. This represents the boundary consistency loss, which constrains the unwrapping phase and the wrapping phase to be consistent in the sense of wrapping around the boundaries of the image with the minimum unwrapping ambiguity. This indicates the weight of the corresponding loss term;

[0167] Specifically:

[0168] (1) Generating adversarial losses

[0169] ;

[0170] In the formula: and These represent the height and width of the phase diagram, respectively. and These represent the indices in the height and width directions of the phase map, respectively; Indicates the generated sample discrimination result in ( , Local confidence level of location discrimination. Indicates the discriminator, This represents the entanglement phase diagram. This indicates the generation of an unwrapped phase diagram;

[0171] (2) Coherence-weighted pixel loss

[0172] ;

[0173] In the formula: express( , Average coherence at location; Indicates the generation of untangled phase diagrams and the true untangling phase diagram exist( , The absolute error of the position. , These represent the generation of untangled phase diagrams. and the true untangling phase diagram exist( , The pixel value at the location;

[0174] (3) Phase gradient consistency loss

[0175] ;

[0176] In the formula: This indicates that the difference will be wrapped back [ , ]; , These represent the first-order difference operators in the horizontal and vertical directions, respectively, used to calculate the gradient of the image; , These represent the generation of untangled phase diagrams. exist( , ) Horizontal and vertical gradients of the position; , These represent the entanglement phase diagrams respectively. exist( , ) Horizontal and vertical gradients of the position;

[0177] (4) Second-order smoothing loss

[0178] ;

[0179] In the formula: , , These represent the horizontal second-order, vertical second-order, and cross second-order difference operators, respectively. , , These represent the generation of untangled phase diagrams. exist( , The horizontal second gradient, vertical second gradient, and cross second gradient of the position;

[0180] (5) Boundary consistency loss

[0181] The four boundary regions are the upper boundary Lower boundary Left boundary Right boundary The loss for a single boundary region is:

[0182] ;

[0183] The boundary consistency loss is the average of the four regions:

[0184] ;

[0185] in: , These represent the generation of untangled phase diagrams. and entanglement phase diagram exist( , The pixel value at the location, and the boundary width of a single boundary region are dynamically determined according to the image size, taking the minimum value between the preset width and one-quarter of the image height and width, to adapt to inputs of different resolutions.

[0186] Training: All training samples were split into training and test sets in an 8:2 ratio; a total of 150 epochs were trained, with a batch size of 16; both the generator and discriminator used the AdamW optimizer (generator learning rate 2×10⁻⁶). -4 Discriminator learning rate 5×10 -5 The weights decay by 0.01, β = (0.5, 0.999)), and the learning rate is based on cosine annealing (T). max =200, ηmin =1×10 -6 Adjustments are made; automatic blending precision (AMP) is used for acceleration.

[0187] Inference: Load generator weights, calculate the condition vector for the input entangled phase, and output the unentangled phase after a single forward propagation.

[0188] VII. Experimental Instructions

[0189] Unified experimental conditions: 1999 test samples, 150 epochs of training, batch size 16, AdamW + cosine annealing; consistent calculation methods for PSNR, SSIM, LPIPS, MSE, and RMSE.

[0190] The components for each configuration are added progressively as shown in the table below:

[0191]

[0192] in conclusion:

[0193] (1) Multi-scale FiLM brings the largest single improvement compared to bottleneck single-point FiLM (Comparison C → Comparison D, +1.54 dBPSNR, +0.0241 SSIM).

[0194] (2) Under the premise of multi-scale FiLM, the accuracy is further improved by changing the dual-scale discriminator to a single-scale discriminator (+0.30 dB PSNR, +0.0042 SSIM, MSE reduced to 0.0050).

[0195] The configuration of this invention is optimal in all four point-by-point / structural indices: PSNR, SSIM, MSE, and RMSE.

[0196] In summary, compared with existing technologies, this invention achieves the following positive effects (verified on a test set of 1999 256×256 simulated + real InSAR samples with fringe density ranging from 0 to 2400+ jumps / map):

[0197] High-precision density adaptive unwrapping: The optimal solution (multi-scale FiLM + single-scale discriminator) achieves PSNR of 31.76 dB, SSIM of 0.9103, MSE of 0.0050, and RMSE of 0.0441, which is the best in the series in terms of point-by-point accuracy indicators.

[0198] Significantly superior to unconditional and single-point conditional baselines: Compared to the unconditional single-scale baseline (PSNR 31.15) and the bottleneck-only FiLM scheme (PSNR 29.92), multi-scale FiLM delivers the largest single gain (approximately +1.84 dBPSNR and +0.028 SSIM improvement over bottleneck FiLM); compared to the scheme retaining the small-scale discriminator (PSNR 31.46), removing the small-scale discriminator further improves PSNR by +0.30 dB and SSIM by +0.004.

[0199] Robust in low-coherence and high-density fringe regions: Fringe density conditions and coherence weighting enable the network to maintain high accuracy in challenging regions;

[0200] End-to-end and highly efficient: After training, inference is performed in a single forward propagation, requiring no iterative optimization, and has the potential for real-time / near-real-time processing in practical InSAR systems.

[0201] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit the technical solutions. Those skilled in the art should understand that any modifications or equivalent substitutions to the technical solutions of the present invention without departing from the spirit and scope of the present invention should be covered within the scope of the claims of the present invention.

Claims

1. An InSAR phase unwrapping method based on multi-scale FiLM conditional generative adversarial networks, characterized in that, include: S1: Obtain the entanglement phase map to be processed; S2: Calculate the fringe density, jump number, and average coherence of the entangled phase map to be processed, and generate a global statistical condition vector; S3: Input the entangled phase map to be processed and the corresponding global statistical conditional vector into the generator of the trained multi-scale feature linear modulation conditional generative adversarial network, and output the corresponding generated unentangled phase map; The generator includes an encoder with multiple downsampling modules, a bottleneck layer, and a decoder with multiple conditional generative adversarial network decoding blocks; The generator's processing steps include: S301: Based on the global statistical conditional vector, the conditional encoder and multi-scale FiLM header generate characteristic linear modulation parameters of the corresponding scale for each conditional generative adversarial network decoding block of the bottleneck layer and decoder. S302: The encoder uses multiple cascaded downsampling modules to downsample the wound phase map step by step to obtain the coded feature map; each downsampling module outputs a downsampled feature map of the corresponding scale, and the output of the last downsampling module is used as the coded feature map output by the encoder. S303: By using the feature linear modulation parameters of the corresponding scale in the bottleneck layer to conditionally modulate the encoded feature map, a deep feature map is obtained; S304: The decoder uses multiple conditional generative adversarial network (GAN) decoding blocks to perform progressive upsampling and conditional modulation on the deep feature map, and uses the output of the last GAN decoding block as the fused modulation feature map; where each GAN decoding block upsamples the input feature map, fuses it with the downsampled feature map of the corresponding scale in the encoder, and then performs conditional modulation using the feature linear modulation parameters of the corresponding scale. S305: The fused modulation feature map is processed by the output head to obtain the generated unwrapped phase map.

2. The InSAR phase unwrapping method based on multi-scale FiLM conditional generative adversarial networks as described in claim 1, characterized in that: Step S2, the processing steps for generating the global statistical condition vector, include: S201: Calculate the horizontal and vertical phase gradients of adjacent pixels in the radian domain of the wound phase map, and statistically analyze the gradient magnitudes exceeding... The number of pixels as the jump number ; S202: Based on jump numbers Calculate the stripe density that reflects the jump density per unit pixel. ; The calculation formula is: ； In the formula: and These represent the height and width of the entanglement phase diagram, respectively. S203: Wrap the phase diagram Complexity index Within a 5×5 sliding window, mean filtering is applied to both the real and imaginary parts, and the modulus is taken to obtain a pixel-by-pixel coherence map; the average value of the coherence map is calculated to obtain the average coherence. ; S204: Regarding stripe density Jump variables and average coherence After normalization, the vectors are concatenated to generate a global statistical conditional vector. ; The calculation formula is: ； In the formula: , , These represent the normalized fringe density, jump number, and average coherence, respectively.

3. The InSAR phase unwrapping method based on multi-scale FiLM conditional generative adversarial networks as described in claim 1, characterized in that: In step S301, the global statistical conditional vector is mapped to shared conditional features by using two fully connected layers with ReLU activation function through a conditional encoder. By employing a multi-scale FiLM split-head approach, multiple independent linear projection heads are used to generate characteristic linear modulation parameters for each scale of the conditional generative adversarial network decoding block in the bottleneck layer and decoder, based on shared conditional features. Each scale outputs a 2×channel vector, which is then split into channel-wise scaling parameters. With bias parameters That is, characteristic linear modulation parameters .

4. The InSAR phase unwrapping method based on multi-scale FiLM conditional generative adversarial networks as described in claim 1, characterized in that: In step S302, the encoder includes a cascaded input layer and four downsampling modules; The input layer consists of cascaded 7×7 convolutional layers, batch normalization layers, and ReLU activation functions; The downsampling module consists of cascaded 4×4 convolutional layers, batch normalization layers, ReLU activation functions, and CBAM attention modules; The encoder's processing steps include: inputting the wound phase map into the input layer to obtain an initial feature map; The initial feature map is input into four cascaded downsampling modules for progressive downsampling, and the output of the last downsampling module is used as the encoded feature map output by the encoder.

5. The InSAR phase unwrapping method based on multi-scale FiLM conditional generative adversarial networks as described in claim 1, characterized in that: In step S303, the bottleneck layer includes three cascaded FiLM residual blocks, and a CBAM attention module is connected between two adjacent FiLM residual blocks; FiLM residual blocks are used to conditionally modulate the input feature map using the feature linear modulation parameters of the corresponding scale to obtain the modulated feature map. The CBAM attention module is used to perform attention enhancement operations on the input feature map; The output of the last FiLM residual block is used as the deep feature map of the bottleneck layer output.

6. The InSAR phase unwrapping method based on multi-scale FiLM conditional generative adversarial networks as described in claim 5, characterized in that: In step S304, the decoder includes four cascaded conditional generative adversarial network decoding blocks; The conditional generative adversarial network decoding block consists of cascaded transposed convolutional layers, batch normalization layers, ReLU activation functions, 3×3 convolutional layers, batch normalization layers, ReLU activation functions, FiLM residual blocks, and CBAM attention modules; The upsampled feature map output by the transposed convolutional layer is first concatenated with the downsampled feature map of the corresponding scale in the encoder along the channel dimension, and then fused through a 3×3 convolutional layer; the output of the last conditional generative adversarial network decoding block is used as the fused modulation feature map output by the decoder. The processing steps for each conditionally generated adversarial network decoding block include: S3041: Upsample the input feature map by using a cascaded transposed convolutional layer, batch normalization layer, and ReLU activation function to obtain an upsampled feature map; S3042: After concatenating the upsampled feature map with the downsampled feature map of the corresponding scale in the encoder along the channel dimension, the concatenated feature map is obtained by processing through a cascaded 3×3 convolutional layer, batch normalization layer, and ReLU activation function. S3043: Linearly modulate the spliced ​​feature map using the feature linear modulation parameters based on the corresponding scale through the FiLM residual block to obtain the modulated feature map; S3044: The modulated feature map is enhanced with attention through the CBAM attention module to obtain the output feature map of the conditional generative adversarial network decoding block.

7. The InSAR phase unwrapping method based on multi-scale FiLM conditional generative adversarial networks as described in claim 6, characterized in that: The formula for calculating conditional modulation of the FiLM residual block is as follows: ； In the formula: This represents the output feature map after FiLM residual block conditional modulation; This represents the input feature map of the FiLM residual block; Represents the characteristic linear modulation parameters. , These represent the channel-by-channel scaling parameter and the offset parameter, respectively.

8. The InSAR phase unwrapping method based on multi-scale FiLM conditional generative adversarial networks as described in claim 1, characterized in that: Step S3, the processing steps for training the multi-scale feature linear modulation conditional generative adversarial network include: S311: Obtain a batch of training samples containing the entanglement phase map and the corresponding real unentanglement phase map; S312: Calculate the fringe density, jump number, and average coherence of the entangled phase map, and generate a global statistical condition vector; S313: Input the entanglement phase diagram and the corresponding global statistical condition vector into the generator, and output the unentanglement phase diagram; S314: The entangled phase image is stitched together with the real unentangled phase image and the generated unentangled phase image to form a dual-channel image. The two dual-channel images are then input into the discriminator to obtain the real sample discrimination result and the generated sample discrimination result, respectively. S315 calculates the discriminator loss based on the discrimination results of real samples and generated samples, and updates the network parameters of the discriminator through the backpropagation algorithm. S316: Calculate the generator's composite loss based on the entangled phase map, the generated unentangled phase map, the true phase map, the coherence map, and the generated sample discrimination results, and update the generator's network parameters through the backpropagation algorithm; S317: Repeat steps S311 to S316 to iteratively train the discriminator and generator until convergence or the maximum number of iterations is reached.

9. The InSAR phase unwrapping method based on multi-scale FiLM conditional generative adversarial networks as described in claim 8, characterized in that: In step S314, the discriminator's processing steps include: S3141: Combine the entangled phase map and the generated unentangled phase map to form a dual-channel image; S3142: Stitch the entangled phase map and the real unentangled phase map into a real dual-channel image; S3143: The generated dual-channel image and the real dual-channel image are sequentially passed through five 4×4 convolutional downsampling layers with a stride of 2 to extract multi-scale features, and then input into a 4×4 convolutional layer with a stride of 1 to obtain the corresponding two local discriminant confidence maps. The first convolutional downsampling layer consists of cascaded 4×4 convolutional layers, LeakyReLU activation function, and CBAM attention module; the second to fifth convolutional downsampling layers consist of cascaded 4×4 convolutional layers, batch normalization layer, LeakyReLU activation function, and CBAM attention module. S3144: Perform true / false judgment on each local block of the two local discriminant confidence maps and generate the judgment result.

10. The InSAR phase unwrapping method based on multi-scale FiLM conditional generative adversarial networks as described in claim 8, characterized in that: In step S316, the formula for calculating the composite loss is: ； In the formula: This represents the composite loss of the generator; This indicates the generation of adversarial loss; This represents the coherence-weighted pixel loss; This represents the phase gradient consistency loss; This represents the second-order smoothing loss; This represents the boundary consistency loss; This indicates the weight of the corresponding loss term; Specifically: (1) Generate adversarial loss ； In the formula: and These represent the height and width of the phase diagram, respectively. and These represent the indices in the height and width directions of the phase map, respectively; Indicates the generated sample discrimination result in ( , Local confidence level of location discrimination. Indicates the discriminator, This represents the entanglement phase diagram. This indicates the generation of an unwrapped phase diagram; (2) Coherence-weighted pixel loss ； In the formula: express( , Average coherence at location; Indicates the generation of untangled phase diagrams and the true untangling phase diagram exist( , The absolute error of the position. , These represent the generation of untangled phase diagrams. and the true untangling phase diagram exist( , The pixel value at the location; (3) Phase gradient consistency loss ； In the formula: This indicates that the difference will be wrapped back [ , ]; , These represent the first-order difference operators in the horizontal and vertical directions, respectively, used to calculate the gradient of the image; , These represent the generation of untangled phase diagrams. exist( , ) Horizontal and vertical gradients of the position; , These represent the entanglement phase diagrams respectively. exist( , ) Horizontal and vertical gradients of the position; (4) Second-order smoothing loss ； In the formula: , , These represent the horizontal second-order, vertical second-order, and cross second-order difference operators, respectively. , , These represent the generation of untangled phase diagrams. exist( , The horizontal second gradient, vertical second gradient, and cross second gradient of the position; (5) Boundary consistency loss The four boundary regions are the upper boundary Lower boundary Left boundary Right boundary The loss for a single boundary region is: ； The boundary consistency loss is the average of the four regions: ； in: , These represent the generation of untangled phase diagrams. and entanglement phase diagram exist( , The pixel value at the location, and the boundary width of a single boundary region are dynamically determined based on the image size.

Images