A cascaded image reconstruction method and device based on a physical residual feedback mechanism and a medium

Abstract

The present application relates to the technical field of computer vision and data processing, and particularly relates to a cascade image reconstruction method based on a physical residual feedback mechanism, equipment and medium, comprising inputting observation data into a pre-trained first-level reconstruction network to output an initial reconstruction image; subsequently, a physical forward projection calculation data domain residual is used, and inverse projection is performed to generate a spatial attention mask; finally, the initial reconstruction image is combined with the spatial attention mask to obtain a reconstruction image embedded with the spatial attention mask, which is input into a second-level refining network; and thus a final reconstruction image is output. The method effectively suppresses artifacts that violate physical facts, guarantees the physical authenticity of the reconstruction result, overcomes the edge oversmoothing problem caused by the traditional global unified updating strategy, and significantly improves the edge sharpness and detail integrity of key structures in the reconstruction image.

Description

Technical Field

[0001] This invention relates to the field of computer vision and data processing technology, and in particular to a cascaded image reconstruction method, device and medium based on a physical residual feedback mechanism. Background Technology

[0002] In recent years, data-driven methods based on deep learning have attracted much attention due to their fast inference speed in image reconstruction tasks. The mainstream approach typically employs end-to-end convolutional neural networks to directly establish a nonlinear mapping relationship from multidimensional observation data to the target image. However, these methods still face significant challenges in practical applications.

[0003] To overcome the lack of physical constraints in end-to-end networks, deep learning solutions incorporating physical models have emerged in existing technologies. These solutions employ an encoder-decoder architecture to directly map data to images. However, this is essentially a purely data-driven black-box model that only fits the statistical regularities of data distribution. It lacks an effective backtracking mechanism to verify whether the image conforms to the physical observation data defined by the wave equation. Therefore, the network is prone to generating visually plausible artifacts that severely violate physical laws.

[0004] Another type of existing technology embeds differentiable physical forward projection operators between network layers to introduce physical gradients. However, when processing physical gradients using convolutional networks, a globally uniform update strategy is employed. This approach ignores the significant spatial non-uniformity of reconstruction errors, causing the network to apply equal computational weights to simple background regions and complex construction boundaries. Key high-frequency details are difficult to effectively enhance, ultimately resulting in overly smoothed edges in the reconstructed image, failing to accurately restore the sharp topological structure of the target in the reconstructed image.

[0005] Therefore, current deep learning-based image reconstruction techniques share a common technical problem: when performing image reconstruction with physical consistency constraints, there is a lack of a mechanism that can adaptively and selectively target and refine high-error regions in the reconstructed image based on the non-uniform spatial distribution of physical residuals. This results in a difficulty in achieving both physical consistency and high-frequency detail recovery accuracy in the reconstruction results. Summary of the Invention

[0006] The purpose of this invention is to provide a cascaded image reconstruction method based on a physical residual feedback mechanism. This method can adaptively and selectively target and refine high-error regions in the reconstructed image according to the non-uniform spatial distribution of physical residuals. This solves the problem that existing physical constraint reconstruction methods, which use a global uniform update strategy, cannot adaptively target and refine high-error regions according to the spatial distribution differences of physical residuals, resulting in insufficient physical consistency, overly smooth edges, and loss of high-frequency details in the reconstructed image.

[0007] To achieve the above objectives,

[0008] Firstly, a cascaded image reconstruction method based on a physical residual feedback mechanism is provided, including:

[0009] Data acquisition and preprocessing, the preprocessing including adding Gaussian noise and random missing channel processing;

[0010] The observation data is input into the pre-trained first-level reconstruction network, and the output is the initial reconstruction map;

[0011] The initial reconstructed image is input into the physical simulation layer for forward projection to obtain the corresponding simulated observation data. The difference between the observed data and the simulated observation data is calculated in the physical domain to obtain the data domain residual. By introducing a physical forward projection and residual feedback mechanism into the cascaded network, this invention can calculate the physical difference between the observed data and the predicted data in real time during the network inference process, thereby establishing a closed-loop constraint between the data domain and the image domain, effectively suppressing artifacts and structural distortions that violate physical observation facts, and ensuring the physical authenticity of the reconstruction results.

[0012] Using the inverse projection unit, the data domain residual is mapped back to the image domain to generate a spatial error feature map;

[0013] The spatial error feature map is subjected to LeakyReLU nonlinear activation processing to generate a spatial attention mask;

[0014] Using an attention residual gating mechanism, the initial reconstruction map is combined with the spatial attention mask to obtain a reconstruction map with the embedded spatial attention mask, which is then input into the second-level refinement network. The spatial attention mask is used as a structure selector to extract features. At the same time, a non-linear convolutional layer is used for local reconstruction, and finally the features are superimposed back to the original features in the form of residuals.

[0015] Output the final reconstructed image.

[0016] Furthermore, it also includes designing a joint loss function, which is used to effectively constrain both the image domain and the physical data domain simultaneously.

[0017] Furthermore, the first-level reconstruction network adopts a U-Net architecture based on depthwise separable convolutions. Downsampling consists of six convolution operations, each of which sequentially executes a max pooling layer with a stride of 2, a depthwise separable convolutional layer with a stride of 1 and a kernel of 3, batch normalization, and LeakyReLU nonlinear activation. Upsampling consists of six deconvolution operations, each of which sequentially executes a skip connection, a depthwise separable convolutional layer with a stride of 1 and a kernel of 3, batch normalization, LeakyReLU nonlinear activation, and a deconvolutional layer with a stride of 2.

[0018] Furthermore, the second-level refinement network adopts the U-Net architecture based on depthwise separable convolution, where downsampling consists of five convolution operations and upsampling consists of five deconvolution operations.

[0019] Furthermore, the joint loss function is:

[0020] ,in,

[0021] Represents the joint loss function;

[0022] This represents the image domain fidelity loss;

[0023] Indicates the physical domain fidelity loss;

[0024] This indicates a localized loss of phase consistency in the sliding window.

[0025] , respectively, represent the learnable parameters in the three networks;

[0026] , respectively represent the weight coefficients of the three constraints;

[0027] Represents the natural constant.

[0028] Furthermore, the inverse projection unit is used to map the data domain residual back to the image domain to generate a spatial error feature map;

[0029] Specifically, this includes data domain residuals. Each channel Apply spatial filtering independently. Given a set of real numbers, extract single-channel wavefield features;

[0030] ,in,

[0031] The number of channels representing the residual in the data field;

[0032] Indicates the height of the data domain residual;

[0033] Indicates the width of the data field residual;

[0034] Represents data domain residuals The first in Individual channel characteristics;

[0035] Represents a two-dimensional convolution operation;

[0036] For the first Each channel corresponds to Spatial convolution kernel, The preset kernel size;

[0037] Indicates the extracted first... Each single-channel wavefield feature constitutes an intermediate feature map. .

[0038] This step does not involve inter-channel information exchange; utilizing Convolution kernel pair Perform a linear combination across channels:

[0039] ,in,

[0040] The total number of channels in the output spatial error feature map;

[0041] For output channel index;

[0042] For the first The corresponding output channel of the first One input channel Convolution weights;

[0043] For bias terms, For the first The bias term corresponding to each output channel;

[0044] Based on the characteristics of each channel The cross-channel combined feature map;

[0045] Use bilinear interpolation upsampling operator The feature maps are aligned from the data domain scale to the image domain scale to obtain the final spatial error features:

[0046] ,in, Let LeakyReLU be the nonlinear activation function; apply LeakyReLU nonlinear activation to the above spatial error feature map to generate a spatial attention mask, denoted as: ,in, This is the spatial error feature map; For spatial attention masks, their dimensions are... Keep it consistent.

[0047] Furthermore, the attention residual gating mechanism combines the initial reconstruction map with the spatial attention mask to obtain a reconstruction map embedded with the spatial attention mask, which is then input into the second-level refinement network; the specific formula is as follows:

[0048] ,in, Indicates by Local refinement operators consisting of convolution and ReLU activation functions;

[0049] This is a learnable residual scaling factor;

[0050] Spatial attention mask;

[0051] Represents the initial reconstruction image;

[0052] This represents the reconstructed image after embedding a spatial attention mask.

[0053] This represents element-wise multiplication. By constructing an attention residual gating mechanism, this invention can map physical errors into spatial location weights, achieving targeted gating refinement of high-error regions. This changes the traditional method of uniform convolution processing of the entire image, thereby significantly improving the edge sharpness and integrity of complex geological target topology while suppressing background noise.

[0054] Furthermore, the local sliding window phase consistency loss is calculated through the following steps:

[0055] Along the time dimension, the simulated observation data and the real observation data sequence are divided into multiple overlapping local time windows to obtain the corresponding simulated subsequence and observation subsequence;

[0056] The simulated subsequence and observed subsequence within each window are respectively subjected to mean removal processing;

[0057] For the mean-removed simulated subsequence and the observed subsequence within each window, calculate their normalized cross-correlation coefficient;

[0058] The local sliding window phase consistency loss is calculated based on the normalized cross-correlation coefficients of all windows.

[0059] Secondly, a cascaded image reconstruction device based on a physical residual feedback mechanism is provided, the device comprising:

[0060] processor;

[0061] And a memory storing executable code, which, when executed, causes the processor to perform a cascaded image reconstruction method based on a physical residual feedback mechanism.

[0062] Thirdly, a non-volatile computer storage medium is provided, characterized in that it stores computer instructions thereon, which, when executed, implement a cascaded image reconstruction method based on a physical residual feedback mechanism.

[0063] The beneficial effects of this invention are:

[0064] 1. By introducing a physical residual feedback mechanism, false artifacts are eliminated and physical authenticity is guaranteed. A "data-image" closed-loop constraint is established in network inference, which solves the problem of lack of physical verification in traditional end-to-end networks and effectively suppresses false textures and structural distortions that violate physical facts.

[0065] 2. By utilizing an attention residual gating mechanism, the problem of overly smooth edges is solved, improving the sharpness of complex structures. This enables "targeted refinement" of high-error areas such as faults and boundaries. This changes the traditional global unified update model, significantly improving the edge clarity and integrity of complex geological targets. Attached Figure Description

[0066] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0067] Figure 1 This is a schematic diagram of the structure of the first-level reconstruction network of the present invention.

[0068] Figure 2 This is a schematic diagram of the structure of the second-level refinement network of the present invention.

[0069] Figure 3 This is a flowchart illustrating the physical residual verification and attention mask generation process of the present invention.

[0070] Figure 4 This is the initial reconstruction map output by the first-level reconstruction network of this invention.

[0071] Figure 5 This is the reconstruction image after attention-targeting refinement output by the second-level refinement network of this invention.

[0072] Figure 6 This is a schematic diagram of the actual speed model used in this invention.

[0073] Figure 7 This is a schematic diagram of the internal structure of the device provided in the embodiments of this application. Detailed Implementation

[0074] The embodiments of the present invention are described in detail below. Examples of the embodiments are shown in the accompanying drawings. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, but should not be construed as limiting the present invention.

[0075] The first embodiment of this application is as follows: Since velocity models exhibit significant spatial correlation and structural characteristics in geological formations, which is highly consistent with the statistical regularities of natural images, this invention treats velocity models as a type of geological structural image for processing.

[0076] Please see Figures 1 to 7 This invention provides a cascaded image reconstruction method based on a physical residual feedback mechanism, comprising the following steps:

[0077] S101: Acquire multidimensional observation data and preprocess the observation data;

[0078] Specifically, taking seismic exploration as an example, the observation data is multi-shot seismic data, possessing three dimensions: shot gather, time domain, and spatial domain. The data used can be sourced from publicly available datasets such as OpenFWI, SEGsalt, and SEGsimulation. The reconstructed target image is a two-dimensional subsurface velocity model, which exhibits significant spatial correlation and structural features in geological structures and can be treated as a geological structural image for processing.

[0079] The preprocessing of the observation data includes compound noise injection and random missing trace processing. First, to simulate interference and data incompleteness in a real exploration environment, the following operations are performed on the raw seismic data: Gaussian random noise with a standard deviation of 0.02 is superimposed; then, 20% of the shot sets are randomly selected, and 20% of the seismic traces within these selected shot sets are further randomly removed to simulate irregular missing traces caused by detector failure or data transmission loss.

[0080] This preprocessing strategy aims to improve the network's robustness to reconstruction, generalization performance, and training stability on noisy and incomplete observation data. Simultaneously, the velocity model data used as training labels is normalized to unify the data scale, accelerate training convergence, and ensure gradient stability.

[0081] S102: Input the preprocessed observation data into the first-level reconstruction network, and the first-level reconstruction network outputs the initial predicted velocity model;

[0082] Specifically, the observation data The input is fed into the pre-trained first-level reconstruction network, which primarily uses the U-Net architecture based on depthwise separable convolutions. Downsampling consists of six convolutional operations, each of which sequentially performs a max-pooling layer with a stride of 2, a depthwise separable convolutional layer with a stride of 1 and a kernel of 3, batch normalization, and LeakyReLU non-linear activation. Upsampling consists of six deconvolutional operations, which sequentially perform skip connections, a depthwise separable convolutional layer with a stride of 1 and a kernel of 3, batch normalization, LeakyReLU non-linear activation, and a deconvolutional layer with a stride of 2. The first-level reconstruction network is as follows: Figure 1 As shown. The output yields the initial predicted velocity model. At this point, the model only recovered the low-frequency background information of the geological structure, but there were a lot of blurs or artifacts at complex boundaries.

[0083] S103: Input the initial predicted velocity model into the physical simulation layer for forward projection operation, calculate the corresponding simulated observation data, calculate the difference between the real observation data and the simulated observation data to obtain the data domain residual, use the inverse projection unit to map the data domain residual back to the image domain to generate a spatial error feature map, perform activation and normalization processing on the spatial error feature map to generate a spatial attention mask.

[0084] Specifically, the initial predicted velocity model The physical simulation layer is input for forward projection to simulate the seismic wave propagation process, and the corresponding simulated observation data is calculated. ; , This represents the forward projection operator.

[0085] Calculate observation data in the physical domain With simulated observation data The difference is used to obtain the data domain residual. The residual reflects the degree of inconsistency between the current prediction model and the physical data.

[0086] Using the inverse projection unit, the data domain residuals are... Mapping back to the image domain generates a spatial error feature map. Due to the large number of parameters and computational redundancy in this process, this embodiment does not use the traditional fully connected transposed convolution, but instead constructs a depth-separable projection unit. This operator decouples the mapping process into two stages: channel-by-channel spatial filtering and cross-channel linear projection. The specific mathematical expression is as follows:

[0087] Channel-by-channel spatial filtering:

[0088] First, the data domain residuals Each channel Apply spatial filtering independently. Represented as a set of real numbers, single-channel wavefield features are extracted:

[0089] ,in,

[0090] The number of channels representing the residual in the data field;

[0091] Indicates the height of the data domain residual;

[0092] Indicates the width of the data field residual;

[0093] Represents data domain residuals The first in Individual channel characteristics;

[0094] Represents a two-dimensional convolution operation;

[0095] For the first Each channel corresponds to Spatial convolution kernel, The preset kernel size;

[0096] Indicates the extracted first... Each single-channel wavefield feature constitutes an intermediate feature map. This step does not involve inter-channel information exchange;

[0097] Cross-channel linear projection: utilizing Convolution kernel pair Perform a linear combination across channels:

[0098] ,in,

[0099] The total number of channels in the output spatial error feature map;

[0100] For output channel index;

[0101] For the first The corresponding output channel of the first One input channel Convolution weights;

[0102] For bias terms, For the first The bias term corresponding to each output channel;

[0103] Based on the characteristics of each channel The cross-channel combined feature map;

[0104] This step is responsible for fusing the physical information of different wave field components into the target feature space.

[0105] Domain scale mapping:

[0106] Use bilinear interpolation upsampling operator The feature maps are aligned from the data domain scale to the velocity model image domain scale to obtain the final spatial error features:

[0107] ,in, Using the LeakyReLU nonlinear activation function, the spatial error feature map above is subjected to LeakyReLU nonlinear activation processing to generate a spatial attention mask, denoted as: ,in,

[0108] This is the spatial error feature map;

[0109] For spatial attention masks, their dimensions are... Keep it consistent.

[0110] Generate spatial attention mask:

[0111] Spatial error feature map LeakyReLU nonlinear activation is applied, and the values ​​are normalized to the [0,1] interval to generate a spatial attention mask. .

[0112] S104: Using an attention residual gating mechanism, the initial predicted velocity model is combined with the spatial attention mask to obtain a reconstructed image after embedding the spatial attention mask, which is then input into the second-level refinement network; the second-level refinement network uses the attention residual gating mechanism to locally reconstruct the high-error regions identified by the mask and outputs the final reconstructed velocity model.

[0113] Specifically, the initial predicted velocity model Spatial attention mask The initial reconstruction image and the spatial attention mask are input together into the second-level refinement network. An attention residual gating mechanism is used to combine the initial reconstruction image with the spatial attention mask to obtain a reconstruction image with the embedded spatial attention mask, which is then used as the input to the second-level refinement network.

[0114] The second-level refinement network also uses the U-Net architecture based on depthwise separable convolutions. However, due to the different input data, downsampling here uses five convolution operations, while upsampling uses five deconvolution operations.

[0115] like Figure 2 As shown, spatial attention masking is used. As a structure selector, it extracts high-frequency detail region features, performs local reconstruction through a non-linear convolutional layer, and finally superimposes the residuals back into the original features. The mathematical expression is as follows:

[0116] ,in, Indicates by Local refinement operators consisting of convolution and ReLU activation functions.

[0117] This is a learnable residual scaling factor;

[0118] Represents the initial reconstruction image;

[0119] This represents the reconstructed image after embedding a spatial attention mask.

[0120] and The approach focuses on processing complex structural regions that are highlighted by masking. The designed attention residual gating mechanism allows the network to target and improve the imaging accuracy of key geological targets without compromising the consistency of the overall velocity model.

[0121] S105: The first-level reconstruction network and the second-level refinement network are trained and optimized using the physical consistency joint loss function.

[0122] Specifically, a joint loss function is designed to effectively constrain both the image domain and the physical data domain: the physical consistency joint loss.

[0123] ,in,

[0124] Represents the joint loss function;

[0125] This represents the image domain fidelity loss;

[0126] Indicates the physical domain fidelity loss;

[0127] This indicates a localized loss of phase consistency in the sliding window.

[0128] , representing the learnable parameters in the three networks respectively;

[0129] , representing the weight coefficients of the three constraints respectively;

[0130] Represents the natural constant.

[0131] Considering the significant differences in numerical magnitude and convergence difficulty among different constraints, simple linear weighting often leads to gradient dominance. This invention employs an adaptive weighting strategy based on homoscedastic uncertainty. It automatically balances the loss contributions of each task by maximizing the Gaussian likelihood function. In the original theory, the minimization objective function for each regression task is typically defined as:

[0132] ,in, This represents the objective function of the original theory. Indicates the current task. The variance of the observation noise for this task. The larger the value, the higher the corresponding loss weight. The smaller, This represents a constant term.

[0133] To improve the stability of numerical computation and avoid the risk of gradient explosion due to zero denominator, a new learnable parameter is defined. To satisfy Based on this variable substitution, the weight term can be transformed into... The regular term can be transformed into However, to simplify the calculation, the intermediate form above is multiplied by a constant 2, and the constant term is omitted. Thus, the final loss function form for a single task was obtained. ,

[0134] ,in,

[0135] Represents the natural constant;

[0136] The single-task optimization loss is after adaptive weighting.

[0137] The original task loss item is to be optimized.

[0138] To simplify the loss function and avoid introducing unnecessary constraints, the most commonly used and effective mean squared error (MSE) is used in the image domain reconstruction loss part. Specifically, the image domain fidelity loss... Represented as:

[0139] ,in, Denote the pixel value of the predicted image output by the network at the coordinate . Denote the pixel value of the true label image at the corresponding coordinate; Denote the height and width of the image.

[0140] Physical fidelity loss . This part aims to verify the correctness of the reconstruction result using the physical forward model. By mapping the reconstructed image back to the physical data space, it is required that the physical properties of the predicted data be consistent with the true observed data. Constructed using norm.

[0141] , where Denote the three dimensions of the input physical observation data; Denote the simulated observation data value obtained by forward projection at this coordinate; Denote the actual observation data value at this coordinate.

[0142] Local sliding window phase consistency loss . Seismic waves are typical non-stationary signals, and their energy decays exponentially with time. The traditional global loss function will be dominated by the high-amplitude direct waves in the shallow layer, causing the network to ignore the weak reflection signals in the deep layer. By introducing an overlapping sliding window mechanism, the long seismic record is segmented into several local short sequences. Independent mean removal and normalization are performed within each local window, so that the weak signals in the deep layer have the same weight as the strong signals in the shallow layer when calculating the loss.

[0143] Intercept the sliding window according to the time dimension. Define a time window with length L and a sliding step of , where < L, forming an overlap to ensure continuity. Then for the th window, the time interval it covers is , . We obtain the simulated subsequence and the observed subsequence according to and . Denote the total number of sliding windows, is the set of real numbers.

[0144] Remove the mean value to lock the phase of the waveform. To eliminate the DC component and low-frequency background drift in the signal and lock the undulating characteristics of the waveform, perform mean removal on the local vector to pull it back to the same zero reference line:

[0145] ;

[0146] ;

[0147] and This represents the simulated subsequence after removing the mean and the observed subsequence.

[0148] Local normalized cross-correlation. The local waveform similarity after removing the mean is calculated using the ratio of the vector inner product to the norm. This step, through L2 norm normalization, forces deep, weak signals and shallow, strong signals to have equal weight.

[0149] ,in, Indicates the first Normalized similarity within a local window;

[0150] Represents the dot product of vectors; express Norm, This represents a tiny constant used to prevent division by zero errors.

[0151] when The closer the correlation is to 1, the stronger the correlation between the two.

[0152] Global consistency aggregation: Aggregate the local similarities of all local windows into a local sliding window phase consistency loss. : To verify the effectiveness of this invention, the trained network was tested on a public dataset, and the initial prediction speed model output by the first-level reconstruction network was quantitatively compared with the final reconstruction speed model processed by the second-level refinement network. Evaluation metrics included mean squared error (MSE), mean absolute error (MAE), universal image quality index (UIQ), and learned perceptual patch similarity (LPIPS). The results are shown in the table below:

[0153] Model output <![CDATA[MSE (×10 -3 )]]> MAE UIQ LPIPS Level 1 Reconstruction Network Output 3.325 0.027 0.978 0.072 Second-level refined network output 3.201 0.025 0.986 0.060

[0154] As shown in the table above, the output of the second-stage refinement network after processing by the cascaded process proposed in this invention outperforms the direct output of the first-stage reconstruction network in all four evaluation metrics: MSE, MAE, UIQ, and LPIPS. In particular, the LPIPS metric, reflecting perceptual quality, is significantly reduced, while the UIQ metric is improved. This quantitatively demonstrates that the "physical residual feedback" and "attention-targeted refinement" mechanisms proposed in this invention can effectively improve the overall fidelity and visual perceptual quality of the reconstructed image, especially in restoring complex structural details.

[0155] This invention establishes a "data-image" closed-loop constraint in network inference by introducing a physical residual feedback mechanism, which solves the problem of lack of physical verification in traditional end-to-end networks and effectively suppresses false textures and structural distortions that violate physical facts.

[0156] By utilizing an attention residual gating mechanism, physical errors are mapped to spatial weights, enabling "targeted refinement" of high-error areas such as faults and boundaries. This changes the traditional global unified update model and significantly improves the edge clarity and integrity of complex geological targets.

[0157] For non-stationary signals such as seismic waves, a local sliding window phase consistency loss method is proposed. By performing local mean removal and normalization processing through windowing, the energy difference between deep and shallow signals is eliminated, ensuring the weight balance of weak deep signals in the inversion, thereby achieving accurate reconstruction of deep geological structures.

[0158] A cascaded image reconstruction device based on a physical residual feedback mechanism is provided, the device comprising:

[0159] processor;

[0160] And a memory storing executable code, which, when executed, causes the processor to perform a cascaded image reconstruction method based on a physical residual feedback mechanism.

[0161] A non-volatile computer storage medium is provided, on which computer instructions are stored, which, when executed, implement a cascaded image reconstruction method based on a physical residual feedback mechanism.

[0162] The above-disclosed embodiments are merely one or more preferred embodiments of this application and should not be construed as limiting the scope of this application. Those skilled in the art can understand that all or part of the processes for implementing the above embodiments and equivalent changes made in accordance with the claims of this application still fall within the scope of this application.

Claims

1. A cascaded image reconstruction method based on a physical residual feedback mechanism, characterized in that, include: Data acquisition and preprocessing, the preprocessing including adding Gaussian noise and random missing channel processing; The observation data is input into the pre-trained first-level reconstruction network, and the output is the initial reconstruction map. The first-level reconstruction network adopts the U-Net architecture based on depthwise separable convolution. The initial reconstructed image is input into the physical simulation layer for forward projection to obtain the corresponding simulated observation data; the difference between the observed data and the simulated observation data is calculated in the physical domain to obtain the data domain residual. Using an inverse projection unit, the data domain residuals are mapped back to the image domain to generate a spatial error feature map, specifically including: Data domain residuals Each channel Apply spatial filtering independently. Given a set of real numbers, extract single-channel wavefield features; ,in, The number of channels representing the residual in the data field; Indicates the height of the data domain residual; Indicates the width of the data field residual; Represents data domain residuals The first in Individual channel characteristics; Represents a two-dimensional convolution operation; For the first Each channel corresponds to Spatial convolution kernel, The preset kernel size; Indicates the extracted first... Each single-channel wavefield feature constitutes an intermediate feature map. This step does not involve information exchange between channels. use Convolution kernel pair Perform a linear combination across channels: ,in, The total number of channels in the output spatial error feature map; For output channel index; For the first The corresponding output channel of the first One input channel Convolution weights; For bias terms, For the first The bias term corresponding to each output channel; Based on the characteristics of each channel The cross-channel combined feature map; Use bilinear interpolation upsampling operator The feature map is aligned from the data domain scale to the image domain scale to obtain the final spatial error features. ,in, LeakyReLU is a non-linear activation function. The spatial error feature map is subjected to LeakyReLU nonlinear activation processing to generate a spatial attention mask, which is represented as follows: ,in, This is the spatial error feature map; For spatial attention masks, their dimensions are... Keep it consistent; The spatial error feature map is subjected to LeakyReLU nonlinear activation processing to generate a spatial attention mask; Using an attention residual gating mechanism, the initial reconstructed image is combined with the spatial attention mask to obtain a reconstructed image with the embedded spatial attention mask, which is then input into the second-level refinement network. The second-level refinement network adopts a U-Net architecture based on depthwise separable convolution. The spatial attention mask is used as a structure selector to extract features. At the same time, a non-linear convolutional layer is used for local reconstruction, and finally the features are superimposed back to the original features in the form of residuals. Output the final reconstructed image.

2. The cascaded image reconstruction method based on physical residual feedback mechanism according to claim 1, characterized in that, It also includes designing a joint loss function that effectively constrains both the image domain and the physical data domain simultaneously.

3. The cascaded image reconstruction method based on physical residual feedback mechanism as described in claim 1 or 2, characterized in that, The first-level reconstruction network adopts a U-Net architecture based on depthwise separable convolutions. Downsampling consists of six convolution operations, each of which sequentially executes a max pooling layer with a stride of 2, a depthwise separable convolutional layer with a stride of 1 and a kernel of 3, batch normalization, and LeakyReLU nonlinear activation. Upsampling consists of six deconvolution operations, which sequentially execute skip connections, a depthwise separable convolutional layer with a stride of 1 and a kernel of 3, batch normalization, LeakyReLU nonlinear activation, and a deconvolutional layer with a stride of 2.

4. The cascaded image reconstruction method based on physical residual feedback mechanism as described in claim 1 or 2, characterized in that, The second-level refinement network adopts the U-Net architecture based on depthwise separable convolution, where downsampling consists of five convolution operations and upsampling consists of five deconvolution operations.

5. The cascaded image reconstruction method based on physical residual feedback mechanism according to claim 2, characterized in that, The joint loss function is: ,in, Represents the joint loss function; This represents the image domain fidelity loss; Indicates the physical domain fidelity loss; This indicates a localized loss of phase consistency in the sliding window. , representing the learnable parameters in the three networks respectively; , representing the weight coefficients of the three constraints respectively; Represents the natural constant.

6. The cascaded image reconstruction method based on physical residual feedback mechanism according to claim 1, characterized in that, The attention residual gating mechanism combines the initial reconstruction map with the spatial attention mask to obtain a reconstruction map embedded with the spatial attention mask, which is then input into the second-level refinement network; the specific formula is as follows: ,in, This represents a local refinement operator consisting of a 3x3 convolution and a ReLU activation function; This is a learnable residual scaling factor; Spatial attention mask; Represents the initial reconstruction image; This represents the reconstructed image after embedding a spatial attention mask. This indicates element-wise multiplication.

7. The cascaded image reconstruction method based on physical residual feedback mechanism as described in claim 5, characterized in that, The local sliding window phase consistency loss is calculated through the following steps: Along the time dimension, the simulated observation data and the real observation data sequence are divided into multiple overlapping local time windows to obtain the corresponding simulated subsequence and observation subsequence; The simulated subsequence and observed subsequence within each window are respectively subjected to mean removal processing; For the mean-removed simulated subsequence and the observed subsequence within each window, calculate their normalized cross-correlation coefficient; The local sliding window phase consistency loss is calculated based on the normalized cross-correlation coefficients of all windows.

8. A cascaded image reconstruction device based on a physical residual feedback mechanism, characterized in that, The device includes: processor; And a memory having executable code stored thereon, which, when executed, causes the processor to perform a cascaded image reconstruction method based on a physical residual feedback mechanism as described in any one of claims 1-7.

9. A non-volatile computer storage medium, characterized in that, It stores computer instructions, which, when executed, implement a cascaded image reconstruction method based on a physical residual feedback mechanism as described in any one of claims 1-7.

Images