Image feature extraction method based on multi-domain cooperation and full-link structure maintenance

Abstract

The application discloses an image feature extraction method based on multi-domain cooperation and full-link structure maintenance, and belongs to the technical field of image processing. KAN The network replaces the traditional convolution dimension reduction, effectively solves the structure distortion problem in the high-dimensional data mapping process, and is particularly suitable for image recovery with strong geometric features. GOIM 、 FDE and LKG work together in the model, realize all-around capture of image features, avoid the blind area of single model, and design LKG Branch replaces the traditional linear gate, and significantly improves the perception sensitivity of the network to local small structures and edges.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, particularly to the field of image artificial intelligence, specifically to an image feature extraction method based on multi-domain collaboration and end-to-end structure preservation. This invention is based on the Kolmogorov-Arnold network (…). KAN Image feature extraction using the improved state-space model (Mamba) can be widely applied to high-dimensional nonlinear tasks such as medical image reconstruction, remote sensing image super-resolution, geophysical exploration imaging, and image restoration. Background Technology

[0002] In deep learning-driven image inversion and reconstruction tasks, the goal is typically to recover / extract target images (such as subsurface velocity models, high-resolution images) with clear physical structure and high-frequency texture from observational data (such as low-resolution images, noisy images). However, existing technologies face numerous challenges in handling such tasks:

[0003] First, structural information is easily lost during downsampling. Mainstream convolutional neural networks (CNNs) typically use stride convolution or pooling operations for dimensionality reduction in the early stages of feature extraction. Because traditional convolutional kernels have fixed weights and are based on linear superposition, they are difficult to accurately fit the complex nonlinear topological structures in images, leading to distortion or blurring of key structural information during dimensionality reduction.

[0004] Second, balancing global dependency with computational efficiency is difficult. While the Transformer architecture possesses a global receptive field, the computational complexity of its self-attention mechanism increases quadratically with image resolution, making it difficult to handle large-scale data such as geophysical data. Although the recently proposed state-space model (Mamba) achieves global modeling with linear complexity, its original design is mainly for one-dimensional sequences and lacks explicit utilization of the frequency domain information of two-dimensional images, resulting in insufficient high-frequency detail recovery capabilities.

[0005] Third, local nonlinear representation is limited. Existing improved Mamba methods typically use ordinary convolution or linear projection to generate gated signals when enhancing local perception. This simple linear gating is difficult to handle nonlinear features with drastic changes in local pixels, limiting the model's ability to finely characterize small targets.

[0006] Furthermore, existing deep generative models generally exhibit a significant "spectral bias," meaning that neural networks tend to preferentially fit the low-frequency components of images, making it difficult to capture and recover high-frequency texture details. When processing high-dimensional data such as seismic wavefields or medical images, key features such as minute lesions, faults, or cracks often correspond to the high-frequency components of the signal. Existing methods are prone to edge smoothing effects during prediction, leading to the loss of small targets. Simultaneously, traditional neural networks rely on fixed node activation functions, lacking the ability to dynamically fit high-order nonlinear mappings in complex physical fields. This makes it difficult to maintain the integrity of the data's topological structure while reducing dimensionality, limiting the model's upper limit of accuracy and generalization ability in strongly nonlinear inversion tasks.

[0007] Therefore, there is an urgent need for an efficient image feature extraction method that can simultaneously achieve full-link structure fidelity, global long-distance interaction, frequency domain detail enhancement, and local strong nonlinear perception. Summary of the Invention

[0008] The purpose of this invention is to provide an image feature extraction method based on multi-domain collaboration and end-to-end structure preservation, which aims to solve the problems of structural ambiguity, loss of high-frequency details and insufficient expression of local nonlinear features in complex image processing tasks in existing technologies.

[0009] To achieve the above objectives, the present invention provides the following technical solution:

[0010] An image feature extraction method based on multi-domain collaboration and end-to-end structure preservation includes the following steps:

[0011] S1. Obtain the original image dataset and construct a paired dataset containing the input image and its corresponding target image;

[0012] S2. Construct a multi-domain collaborative image feature extraction model, wherein the image feature extraction model includes:

[0013] based on KAN The structure-fidelity preprocessing module is used to extract shallow structural features from the input image and output a first feature map. The structure-fidelity preprocessing module includes multiple cascaded... KAN The core of the downsampling layer lies in the following: KAN Interspersed in traditional convolutional layers with downsampling layers KAN Network layers perform feature dimensionality reduction; this module utilizes Kolmogorov - Arnold The learnable activation function property of the network maps the input image into a feature map, preserving the geometric topology during downsampling;

[0014] based on KAN Gated frequency domain Mamba The feature extraction group is used to perform deep feature extraction on the first feature map and output a second feature map; the frequency domainMamba Feature extraction groups consist of multiple cascaded KAN Composed of frequency domain gating modules, each KAN The frequency domain gating module includes multiple cascaded... KAN Frequency domain gating module, each KAN The frequency domain gating module includes the following parallel branches;

[0015] Branch 1, Global Omnidirectional Interaction Module: This module converts two-dimensional image features into one-dimensional sequences through a multi-path scanning mechanism and uses state-space equations for recursive processing to capture the global long-distance dependencies of the image with linear computational complexity.

[0016] Branch 2, Frequency Domain Detail Enhancement Module: This module uses Fast Fourier Transform to transform features to the frequency domain, filters and enhances high-frequency components through learnable frequency domain filters, and then fuses them through inverse transform to supplement the texture details lost in the spatial domain.

[0017] Branch 3, Local KAN The gating module utilizes KAN Convolutional layers extract local nonlinear features and generate spatial gating signals to perform nonlinear filtering and enhancement on the input feature map;

[0018] The feature maps of the three branches are then element-wise summed and linearly projected, and then compared with the... KAN The residual connection of the original input feature map of the frequency domain gated module serves as the basis for this... KAN Feature map output by the frequency domain gating module;

[0019] The feature fusion module uses an attention-guided dynamic fusion mechanism to fuse the first feature map and the second feature map, and outputs a fused feature map.

[0020] The image reconstruction module generates the corresponding target image based on the fused feature map;

[0021] S3. The image feature extraction model is trained using the original image dataset, and a joint loss function is used during the training process;

[0022] S4. The trained image feature extraction model is used to process the image to be processed to obtain the target image corresponding to the image to be processed.

[0023] The present invention KAN - PPM The core of this approach lies in interspersing elements within traditional convolutional layers. KAN The network layer performs feature dimensionality reduction. KAN Each weight parameter of the network layer is replaced with a parameterized univariate spline function. By adjusting the control points of the spline function, the nonlinear mapping relationship between the input and output is fitted, thereby preserving the edge and texture structure to the greatest extent while reducing the feature resolution.

[0024] Compared with the prior art, the present invention has the following beneficial effects:

[0025] (1) High structural fidelity: The front end introduces between convolutional layers KAN Networks replace traditional convolutional dimensionality reduction, effectively solving the structural distortion problem in the high-dimensional data mapping process, and are particularly suitable for image restoration with strong geometric features.

[0026] (2) Multi-domain synergy and complementarity: through GOIM (Global Spatial) FDE (Frequency domain details) and LKG The collaborative work of (local nonlinearity) enables comprehensive capture of image features, avoiding the blind spots of a single model.

[0027] (3) Nonlinear Gating Innovation: Design LKG Replacing traditional linear gating with branching significantly improves the network's sensitivity to local microstructures and edges. Attached Figure Description

[0028] Figure 1 This is an embodiment of the present invention. MDS - Net A schematic diagram of the model's structure;

[0029] Figure 2 This is an embodiment of the present invention. KFGB Module structural diagram;

[0030] Figure 3 This is an embodiment of the present invention. GOIM Module structural diagram;

[0031] Figure 4 This is an embodiment of the present invention. FDE Module structural diagram;

[0032] Figure 5 This is an embodiment of the present invention. LKG Module structural diagram;

[0033] Figure 6 This is a schematic diagram showing the results of a comparative experiment in an embodiment of the present invention. Detailed Implementation

[0034] To more clearly illustrate the present invention, specific embodiments are described below. Those skilled in the art should understand that the following description is illustrative rather than restrictive and should not be construed as limiting the scope of protection of the present invention.

[0035] An image feature extraction method based on multi-domain collaboration and end-to-end structure preservation includes the following steps:

[0036] S1. Obtain the original image dataset and perform standardized preprocessing to construct a paired dataset containing the input image and its corresponding target image;

[0037] S2. Construct a multi-domain collaborative image feature extraction model ( MDS - Net ), MDS - Net include:

[0038] based on KAN Structure fidelity preprocessing module ( KAN - PPM This is used to extract shallow structural features from the input image and output the first feature map;

[0039] based on KAN Gated frequency domain Mamba The feature extraction group is used to perform deep feature extraction on the first feature map and output the second feature map;

[0040] The feature fusion module uses an attention-guided dynamic fusion mechanism to fuse the first feature map and the second feature map, and outputs a fused feature map.

[0041] The image reconstruction module generates the corresponding target image based on the fused feature map.

[0042] S3. The image feature extraction model is trained using the original image dataset, and a joint loss function is used during the training process;

[0043] S4. The trained image feature extraction model is used to process the image to be processed to obtain the target image corresponding to the image to be processed.

[0044] In step S1 of this embodiment, corresponding data of the strata and rock mass are collected. The seismic images observed at the surface and their corresponding subsurface velocity model images are used as input and target images, respectively, to construct a paired dataset. The image data is normalized to ensure it is distributed between [0, 1]. Simultaneously, the input seismic images also need to be adjusted to a preset target size using bilinear interpolation to obtain the feature image. F As MDS - Net middle KAN - PPM The input feature maps and these preprocessing steps are routine operations and will not be described in detail here.

[0045] In this embodiment, MDS - Net This is a newly constructed model. To better understand this model, the following will be explained in conjunction with the accompanying diagrams. MDS - Net Each module will be described in detail.

[0046] Please refer to Figure 1 This demonstrates the present embodiment. ​ The network structure of this embodiment ​ - ​It contains a 3-layer downsampling structure, each of which does not use a traditional 2D convolutional layer. ​ Instead, use ​ downsampling layer ( ​ ​ ).

[0047] Each floor ​ - ​ The operation process is as follows:

[0048] For input feature image conduct Convolutional downsampling yields the first downsampled intermediate feature map. Then perform the first deformation operation ( ​ The second intermediate feature map obtained by downsampling is obtained. ,in, For batch size, The number of channels in the input feature map, For the input feature map height, The width of the input feature map;

[0049] Downsample the second intermediate feature map After hierarchical normalization ( ​ Enter after ) ​ Layer, to obtain the downsampled third intermediate feature map ; ​ The mathematical expression for a layer is: ,in Based on Learnable activation functions for splines It is a vector. , It represents A component, whose index is the component number;

[0050] Again Perform the second deformation operation ( ​ After layer normalization, the downsampled fourth intermediate feature map is obtained. As the ​ Output feature map of the downsampling layer;

[0051] After 3 floors ​ , ​ The final output is an initial feature map with the same size as the target image, i.e., the first feature map. As ​ Gated frequency domain ​ Input to the feature extraction group.

[0052] Please refer to ​ and ​ This demonstrates that, based on the present embodiment... ​Gated frequency domain ​ Network structure for feature extraction groups. Frequency domain of this embodiment. ​ Feature extraction groups consist of multiple cascaded ​ Frequency domain gated module ( ​ Composed of, each ​ Composed of multiple ​ Frequency domain gating module ( ​ It consists of a cascaded layer and a convolutional layer.

[0053] Each ​ The structure includes a parallel three-branch architecture, namely:

[0054] Branch 1, Global Omnidirectional Interaction Module ( ​ , ​ This module converts two-dimensional image features into one-dimensional sequences through a multi-path scanning mechanism and uses state-space equations for recursive processing to capture the global long-distance dependencies of the image with linear computational complexity.

[0055] Branch 2, Frequency Domain Detail Enhancement Module ( ​ , ​ This module utilizes the Fast Fourier Transform (FFT). ​ The features are transformed to the frequency domain, filtered and enhanced by a learnable frequency domain filter, and then fused by inverse transform to compensate for the texture details lost in the spatial domain.

[0056] Branch 3, Local ​ Gating module ( ​ , ​ This module utilizes ​ Convolutional layers extract local nonlinear features and generate spatial gating signals to perform nonlinear filtering and enhancement on the input features;

[0057] The feature maps output from the three branches are element-wise summed and linearly projected, then concatenated with the residuals of the original input feature maps to form the final feature map. ​ Feature map output by the frequency domain gating module.

[0058] Specifically, the first feature map After multiple ​ Deep feature extraction can be performed as follows:

[0059]

[0060] In the formula, m express ​ Quantity; Indicates a certain ​ The output feature map, with index _____ ​ The number; This is the second feature map;

[0061] For each ​ In this regard, the process of extracting features can be represented as follows:

[0062]

[0063] In the formula, Indicates the first indivual ​ Output feature map; Indicates the first -1 ​ Output feature map; For the first indivual ​ middle ​ Quantity; Indicates the first indivual ​ One of the middle ​ The output feature map, where the subscript indicates the feature map. ​ In the indivual ​ Chinese number; This indicates a convolution operation.

[0064] ​ Performing global and local modeling in the frequency-space domain, for a given... In other words, it contains multiple ​ , set the first indivual ​ Output feature map , No. indivual ​ Processed through three parallel branches get , can be represented as:

[0065]

[0066] In the formula, , , They represent ​ Feature maps output by the three modules; This indicates a linear projection.

[0067] Please refer to ​ It showcased ​ The network structure. ​ Responsible for capturing global features, it can maintain linear computational complexity. Simultaneously, it enables global information interaction between any two points in the image space, effectively overcoming the problem of limited receptive field in traditional convolutional neural networks. The specific operation is as follows:

[0068] Input feature map After normalization and linear projection, the first intermediate feature maps for the interaction module used for content interaction are generated. x and the second intermediate feature map of the interaction module used for gating adjustment z ;

[0069] To enhance the local context, the first intermediate feature map of the interaction module... x Through a depthwise convolution ( ​ )and ​ Activation function processing yields the third intermediate feature map of the interactive module. In the formula, This represents a depthwise convolution operation. express ​ Activation function;

[0070] The third intermediate feature map of the interaction module is a two-dimensional image. In order to convert it into a one-dimensional sequence suitable for processing by the state-space model, this module adopts a four-directional scanning strategy, starting from the top left, bottom left, top right, and bottom right vertices of the image and scanning along the diagonal direction. Expanded into four independent one-dimensional sequences This omnidirectional scanning eliminates the causal positional bias caused by unidirectional scanning.

[0071] Independent state-space modeling is performed for each of the above one-dimensional sequences. ​ This yields a unidirectional state feature sequence, the core operation of which is:

[0072]

[0073] In the formula, State-space modeling operations representing linear complexity; This is a one-dimensional sequence obtained by unfolding the feature map. Represents a unidirectional state feature sequence, with superscript indicating the direction of the sequence. k Indicates the sequence number;

[0074] ​ The model recursively updates the hidden state using discretized state equations, thereby achieving... The computational cost captures global correlations within the sequence;

[0075] Four passes ​ Processed sequence The reverse process is used to reconstruct the two-dimensional feature maps, which are then added and fused together. Finally, the processed data is used... ​ The second intermediate feature map of the interactive module processed by the activation function The final output of this branch is obtained by applying weighted activation to the feature map after addition and fusion. :

[0076]

[0077] In the formula, This indicates a reverse scan recovery operation, which will go through... ​ Processing one-way state feature sequences They are rearranged in reverse order according to their corresponding scan paths; This indicates point-by-point multiplication.

[0078] Please refer to ​ It showcased ​ The network structure. ​ It is responsible for directly capturing and recovering weak high-frequency signals lost due to downsampling in the frequency domain. It can explicitly compensate for the subtle high-frequency details that are difficult to capture by spatial domain convolution, significantly improving the clarity of the results at textures and edges. Specific operation steps:

[0079] Using two-dimensional fast Fourier transform ( ​ Input feature map Converting to frequency domain features; this step decomposes the image into components of different frequencies, separating high-frequency texture information from low-frequency background information. The transformation process is as follows:

[0080]

[0081] in, Let be a complex spectral tensor containing the real part. and the virtual part ; Represents a two-dimensional Fast Fourier Transform; Represents the set of complex numbers;

[0082] The spectral tensor is reconstructed by concatenating the real and imaginary parts along the channel dimension to obtain a frequency domain feature map.

[0083] ;

[0084] In the formula, Represents the frequency domain feature map; Indicates a splicing operation;

[0085] The frequency domain feature map is then input into a... Convolutional layers and ​ In a learnable filter constructed from activation functions, the filtered feature map is obtained. :

[0086] ,

[0087] In the formula, This represents the filtered feature map, i.e., the frequency domain feature map after filtering and enhancement.

[0088] The role of convolution is to interact between frequency channels, automatically learn and assign higher weights to key frequency components; the role of ReLU is to introduce nonlinear transformation and suppress noise.

[0089] The filtered feature map Recombined into a complex spectrum :

[0090]

[0091] In the formula, Represents a complex number recombination operation; Represents the real part after filtering; This represents the imaginary part after filtering. It is the number of channels;

[0092] The filtered and enhanced complex spectrum (Second complex spectrum), using two-dimensional inverse fast Fourier transform to restore it back to the spatial domain, yields a feature map containing rich high-frequency details. , as the output of FDE;

[0093]

[0094] Please refer to ​ It shows ​ The structure. ​ use ​ Its high-order fitting capability allows it to keenly capture minute velocity abrupt changes in the seismic wave field. The specific operation steps are as follows:

[0095] Input features Through a ​ Convolutional layers, unlike traditional convolutions which use fixed scalar weights, ​ Convolution applies a learnable non-linear activation function at each position of the convolution kernel. This allows for the extraction of local features rich in high-order geometric information. ;

[0096]

[0097] In the formula, Indicates that the convolution kernel is of ​ Convolution operation, ​ The core definition of a convolution kernel is a linear combination of basis functions and B-spline functions:

[0098]

[0099] in, These are trainable weight coefficients; It is a B-spline function;

[0100] Local features pass ​ The activation function compresses and maps numerical values ​​to the [0, 1] interval, generating a nonlinear spatial gated graph. :

[0101]

[0102] Using the generated gating graph In this part ​ Raw input characteristics of the gating module Element-wise multiplication is performed to adaptively activate the feature responses of key regions and suppress the smoothing background, thus obtaining the local... ​ The output of the gating module;

[0103]

[0104] The feature maps from the three branches are summed and then connected via residuals to obtain the output of the KFMB. ;

[0105]

[0106] The feature fusion module performs multi-scale dynamic fusion, which integrates shallow features rich in the original geometric structure (first feature map). ), and deep features rich in semantics and physical laws (second feature map) ) to perform fusion; introduce attention gates ​ Calculate the attention coefficient : , obtain fusion features ;

[0107] Finally, the image reconstruction module is used to... Reconstructed into the final velocity model image The image reconstruction module contains a sub-pixel convolutional layer (PixelShuffle), which can restore the image size while preserving the fine structure and high-frequency texture information extracted by the preceding module to the greatest extent.

[0108] Model training, setting the joint loss function: In the formula, Represents a true velocity model image. This represents the image of the predicted velocity model. Represents structural similarity calculation; and Represents the weighting coefficient; used ​ The optimizer is trained.

[0109] This embodiment uses the publicly available OpenFWI seismic dataset for model training and testing, and selects the mean square error (MSE) as the benchmark. ​ Mean absolute error ( ​ Learning to perceive image patch similarity ( ​ and structural similarity ( ​ As an evaluation metric, it was compared with current mainstream image inversion methods.

[0110] The dataset and experimental environment used were the Vel-Family subset of the OpenFWI dataset, which contains rich geological structures. The dataset was divided into a training set of 24,000 samples and a test set of 6,000 samples. The input seismic image size was [size missing]. (Time sampling) Detector), output velocity model image size is The experiment was conducted using the PyTorch deep learning framework. ​ To accelerate computation, the optimizer employs... ​ The initial learning rate is set to The learning rate is adjusted using a cosine annealing strategy, and the training rounds are 100.

[0111] The method of this invention is compared with three existing mainstream methods. The selected methods include... ​ (A classic inversion method based on pure convolutional neural networks) ​ (Based on pure) ​ The method has global modeling capabilities. ​ (Based on standards) ​ of U Type network, excluding frequency domain enhancement and ​ (Module).

[0112] (I) Comparison of Evaluation Indicators and Experimental Results

[0113] Table 1 shows a comparison of the quantitative metrics of each method on the test set.

[0114] Table 1. Performance comparison of different methods on the OpenFWI dataset.

[0115]

[0116] The results analysis, as shown in Table 1, demonstrates that the method of this invention achieves optimal performance across all indicators.

[0117] The present invention ​ compared to ​ It decreased by approximately 28%. ​ It was reduced by approximately 29%. This is thanks to the global omnidirectional interaction module ( ​The introduction of ) eliminates the directional bias of causal convolution by using an omnidirectional scanning mechanism, enabling the network to more accurately predict the velocity distribution throughout the entire work area.

[0118] The present invention ​ The index reached 0.941, significantly better than... ​ (0.865). This is mainly due to ​ Preprocessing module and local ​ Gated branches ( ​ ). ​ Convolution uses learnable spline functions to accurately fit the nonlinear geometry of geological faults and salt body boundaries, effectively avoiding the edge blurring phenomenon caused by traditional convolution.

[0119] exist ​ In terms of performance indicators, this invention achieves an extremely low value of 0.06, which is far superior to... ​ 0.112. This proves the frequency domain detail enhancement branch ( ​ The effectiveness of ). By explicitly utilizing ​ By capturing and enhancing high-frequency components, this invention successfully recovers minute texture details that were absorbed by deep strata or lost during downsampling, making the inverted image clearer and more realistic in visual perception.

[0120] (II) Comparison Results of Velocity Model Images

[0121] To intuitively evaluate the actual effectiveness of the method of this invention in the reconstruction of complex geological structures, three representative velocity model samples from typical geological types in the test set were selected. ​ In the middle, from left to right, the real models are shown ( ​ ​ ), comparison method ( ​ , ​ and ​ ) and the method of the present invention ( ​ The inversion results are visualized and compared.

[0122] 1. Strong nonlinear boundary recovery effect. In ​ In the real model samples in the first row (samples in the first column of the first row), there are typical irregular high-speed salt dome structures, which are relatively random in shape and have a violent velocity change interface with the surrounding rocks.

[0123] ​ When dealing with such strong nonlinear boundaries, the fixed weights of the convolution kernel result in a significant "oversmoothing" effect, leading to blurred edges of the salt dome and making it difficult to accurately segment the target contour (samples in the first row and second column).

[0124] Benefiting from local ​ Gated branches ( ​This invention excels in handling complex nonlinear function fitting, enabling it to accurately track step changes in the velocity field. ​ As can be seen in the image corresponding to the method of this invention in the first row (sample in the third column of the first row), the salt body boundary reconstructed by this invention is sharp and clear, and highly consistent with the real model, proving that... ​ Convolution has a remarkable ability to overcome edge blurring.

[0125] 2. Deep micro-texture restoration effect. ​ The real model samples in the second row (the samples in the first column of the second row) show the fine layered textures and micro-faults present in deep strata. This is a typical scenario of "high-frequency small-sized targets in a complex low-frequency background," which is extremely difficult to identify due to the attenuation of seismic wave energy. ​ Although it possesses global modeling capabilities, its accuracy in identifying minute faults in this sequence is poor (sample in the second row and second column). The reason for this is that... ​ The architecture suffers from a potential "spectral bias," tending to fit low-frequency components while ignoring high-frequency details. This results in the homogenization of deep textures, preventing the full realization of their potential. This invention addresses this by integrating a frequency domain detail enhancement branch (…). ​ The invention adaptively compensates for the attenuated high-frequency components in the frequency domain. The results show that the invention successfully recovers the sequence details of deep strata, making the fault displacement and dip of small faults clearly discernible in the inverted image (sample in the second row and third column).

[0126] 3. Global construction continuity. ​ The real model samples in the third row (the samples in the first column of the third row) are complex fold models with strong lateral heterogeneity from the OpenFWI dataset, requiring the model to have extremely strong context awareness to maintain the continuity of the formation. Although ​ Superior in overall structure ​ However, fractures or artifacts still appeared at large-angle bends in some strata (sample in the third row and second column), indicating that its ability to capture omnidirectional dependence in two-dimensional space is still limited.

[0127] This invention utilizes a global omnidirectional interaction module ( ​ ), through four directions ​ The scanning mechanism effectively captured long-distance geological dependencies throughout the entire work area. ​ In the image corresponding to the method of the present invention in the third row (sample in the third column of the third row), the inversion results generated by the present invention show the best performance in terms of lateral continuity and physical rationality of the strata, and no obvious structural fractures are found.

[0128] In summary, this invention addresses the two major bottlenecks in existing deep inversion networks: "structural information distortion" and "high-frequency texture loss." It innovatively constructs a feature extraction architecture that combines multi-domain collaboration with end-to-end structural preservation.

[0129] On the one hand, utilizing ​ With powerful high-order nonlinear function fitting capabilities, a structure-fidelity preprocessing module was constructed. ​ ) and local ​ Gated branches ( ​ This successfully solved the fuzziness and smoothing problem of strongly nonlinear geological interfaces (such as salt dome boundaries and faults) during the downsampling process;

[0130] On the other hand, by fusing frequency domain details to enhance the branch ( ​ The method, combined with the Global Omnidirectional Interaction Module (GOIM), overcomes the receptive field limitations and spectral biases of traditional convolutions, achieving cross-scale feature capture from deep micro-texture recovery to maintaining the continuity of the entire texture area. Experimental results show that the method of this invention... ​ , ​ and ​ Significantly superior to existing key indicators ​ and ​ The model demonstrates its high accuracy and robustness in handling geophysical image inversion tasks with high dimensions, nonlinearity, and low signal-to-noise ratio, providing an efficient and reliable new paradigm for fine imaging of complex geological structures.

[0131] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. An image feature extraction method based on multi-domain collaboration and end-to-end structure preservation, characterized in that, Includes the following steps: S1. Obtain the original image dataset and construct a paired dataset containing the input image and the target image; S2. Construct a multi-domain collaborative image feature extraction model, wherein the image feature extraction model includes: based on KAN The structure-fidelity preprocessing module is used to extract shallow structural features from the input image and output a first feature map. The structure-fidelity preprocessing module includes multiple cascaded... KAN downsampling layer, the KAN Interspersed in convolutional layers of downsampling layers KAN The network layer performs feature dimensionality reduction; based on KAN Gated frequency domain Mamba The feature extraction group is used to perform deep feature extraction on the first feature map and output a second feature map; the frequency domain Mamba Feature extraction groups consist of multiple cascaded... KAN Composed of frequency domain gating modules, each KAN The frequency domain gating module includes multiple cascaded... KAN Frequency domain gating module, each KAN The frequency domain gating module includes the following parallel branches; Branch 1, Global Omnidirectional Interaction Module: This module converts two-dimensional image features into one-dimensional sequences through a multi-path scanning mechanism and uses state-space equations for recursive processing to capture the global long-distance dependencies of the image with linear computational complexity. Branch 2, Frequency Domain Detail Enhancement Module: This module uses Fast Fourier Transform to transform the feature map to the frequency domain, filters and enhances high-frequency components through a learnable frequency domain filter, and then fuses them through inverse transform to supplement the texture details lost in the spatial domain. Branch 3, Local KAN The gating module utilizes KAN Convolutional layers extract local nonlinear features and generate spatial gating signals to perform nonlinear filtering and enhancement on the input feature map; The feature maps of the three branches are then element-wise summed and linearly projected, and then compared with the... KAN The residual connection of the original input feature map of the frequency domain gated module serves as the basis for this... KAN Feature map output by the frequency domain gating module; The feature map fusion module uses an attention-guided dynamic fusion mechanism to fuse the first feature map and the second feature map, and outputs a fused feature map. The image reconstruction module generates the corresponding target image based on the fused feature map; S3. The image feature extraction model is trained using the original image dataset, and a joint loss function is used during the training process; S4. The trained image feature extraction model is used to process the image to be processed to obtain the target image corresponding to the image to be processed.

2. The image feature extraction method based on multi-domain collaboration and end-to-end structure preservation according to claim 1, characterized in that, Each of the above KAN The downsampling layer includes the following operations: S211. Perform convolution downsampling on the input feature image to obtain the first downsampled intermediate feature map. ,in, For batch size, The number of channels in the input feature map, For the input feature map height, The width of the input feature map; S212, the downsampled first intermediate feature map The first deformation operation is performed to obtain the downsampled second intermediate feature map. ; S213, the downsampled second intermediate feature map Input after level normalization KAN Layer, to obtain the downsampled third intermediate feature map ; S214, Then, the downsampled third intermediate feature map is processed again. After performing the second deformation operation and layer normalization, the downsampled fourth intermediate feature map is obtained. As the KAN Output feature map of the downsampling layer.

3. The image feature extraction method based on multi-domain collaboration and end-to-end structure preservation according to claim 1, characterized in that, The global omnidirectional interaction module includes the following operations: S2211. The input feature map is normalized and linearly projected to generate the first intermediate feature map of the interaction module for content interaction and the second intermediate feature map of the interaction module for gating adjustment, respectively. S2212, The first intermediate feature map of the interaction module is processed by a depthwise convolution and SiLU The activation function is used to obtain the third intermediate feature map of the interactive module; S2213. The third intermediate feature map of the interaction module is transformed into multiple independent one-dimensional sequences by omnidirectional scanning. Each one-dimensional sequence is processed by the state space model to obtain the corresponding unidirectional state feature sequence. S2214. Reversely restore all the unidirectional state feature sequences to two-dimensional feature maps, and then add and fuse them; S2215, Utilization of the process SiLU The second intermediate feature map of the interaction module processed by the activation function performs weighted activation on the feature map after addition and fusion in step S2214 to obtain the output feature map of the global omnidirectional interaction module.

4. The image feature extraction method based on multi-domain collaboration and end-to-end structure preservation according to claim 1, characterized in that, The frequency domain detail enhancement module includes the following operations: S2221. The input spatial feature map is converted into frequency domain features using two-dimensional fast Fourier transform. The complex spectral tensor in the frequency domain features contains real and imaginary parts. S2222: Concatenate the real and imaginary parts of the complex spectrum tensor along the channel dimension to obtain a frequency domain feature map; S2223. Input the frequency domain feature map into a... Convolutional layers and ReLU The filtered feature map is obtained from the learnable filter constructed by the activation function; S2224. Recombine the filtered feature map into a second complex spectrum; S2225. After the second complex spectrum is filtered and enhanced, it is restored back to the spatial domain using a two-dimensional inverse fast Fourier transform to obtain a feature map containing rich high-frequency details, which is used as the output of the frequency domain detail enhancement module.

5. The image feature extraction method based on multi-domain collaboration and end-to-end structure preservation according to claim 1, characterized in that, The local KAN The gating module includes the following operations: S2231, Through KAN Convolutional layers extract local features rich in high-order geometric information from the input feature map; S2232, Transfer local features through Sigmoid Activation function to generate nonlinear spatial gated graph; S2233. Using the generated gating graph, perform local... KAN The original input feature map of the gating module is multiplied element-wise to obtain the local feature map. KAN Output of the gating module.

Images