Long baseline interference array image reconstruction method based on multi-scale hybrid architecture

Abstract

The invention discloses a long baseline interference array image reconstruction method based on a multi-scale hybrid architecture. The method comprises the following steps: constructing a training set and a test set by using a spatial frequency undersampling image; the method comprises the following steps: constructing an image reconstruction model of which a main body frame is a multi-resolution coding-decoding structure, wherein the model comprises a mixed sensing module and a shallow feature extraction module, wherein the mixed sensing module is composed of a spatial domain convolution branch, a frequency domain convolution branch and a multi-scale feedforward network; inputting the training set into an image reconstruction model, introducing a multi-scale spatial domain and frequency domain information reconstruction loss function, and training the model by constructing a multi-dimensional constraint condition and a progressive training strategy; and performing image reconstruction by using the trained image reconstruction model. According to the method, the ringing effect and the artifact phenomenon caused by undersampling can be effectively reduced in a complex target imaging scene with extremely low frequency sampling coverage rate, high-quality image reconstruction is realized, and the method is more suitable for long-baseline interference array high-resolution imaging detection.

Description

Technical Field

[0001] This invention belongs to the field of optical synthetic aperture imaging technology, specifically relating to a long baseline interferometric array image reconstruction method based on a multi-scale hybrid architecture. Background Technology

[0002] With the rapid development of Earth observation and space reconnaissance technologies, the demand for ultra-large aperture and ultra-high resolution remote sensing images is growing exponentially. Traditional single-aperture telescopes, limited by their physical size, can no longer meet the technical requirements for detailed detection of complex scenes. Sparse optical interferometer arrays, through spatiotemporal synchronous signal synthesis from distributed telescopes, can construct large-scale observation systems with equivalent apertures of hundreds or even kilometers, overcoming physical aperture limitations and achieving ultra-high resolution observations using ultra-long baselines. However, limited by the number of array telescopes and baseline geometry, the spatial frequency sampling of the interferometer array cannot meet the Nyquist completeness condition, leading to severe ill-conditioned ill-posedness in image reconstruction. Taking the Very Large Telescope Interferometer (VLTI) as an example, its 30-200 meter baseline distribution only covers a limited spatial frequency range. The sparse sampling of high-frequency information directly leads to degradation phenomena such as spectral aliasing and blurred edge features in the reconstructed image. Therefore, how to accurately reconstruct the spatial frequency of the original signal based on limited baseline sampling has become a key breakthrough point for current high-resolution remote sensing imaging technology.

[0003] Traditional undersampled image reconstruction algorithms mainly include the CLEAN algorithm, the Maximum Entropy Method (MEM), and compressed sensing algorithms. Among them, the CLEAN algorithm is insufficient in representing the complex structure of extended targets and suffers from problems such as low computational efficiency and unstable solutions in low signal-to-noise ratio scenarios. The maximum entropy method relies on high-density observation data and its non-convex optimization process is prone to getting trapped in local extrema, which limits its application performance under long baseline conditions. Compressed sensing algorithms impose strict incoherence requirements on the geometric configuration design of the observation array and need to ensure low correlation between the measurement matrix and the sparse basis during the hardware layout stage, which increases the complexity of engineering implementation.

[0004] Recent research has explored deep learning-based methods for reconstructing undersampled images from interferometric arrays. These methods mine deep semantic information from images through multi-level representation learning, constructing a high-dimensional nonlinear mapping model from low-quality observations to high-fidelity outputs. However, these methods only model local correlations between spatial pixels, and their effectiveness depends on the assumption of high spatial frequency sampling density and low complexity of the target structure. In highly undersampled Earth observation tasks, artifacts and blurring are greatly increased, and single spatial feature modeling is insufficient to support the recovery of detailed features. Therefore, a method for image reconstruction of highly undersampled images is urgently needed. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a long baseline interferometric array image reconstruction method based on a multi-scale hybrid architecture. By constructing a multi-scale hybrid framework to capture and fuse spatial-frequency domain information, and analyzing the spectral distribution pattern based on the spatial frequency information of finite baseline sampling, unknown frequency components are recovered, thereby achieving high-quality image reconstruction of a highly undersampled interferometric array imaging system. This method has a positive promoting effect on the development of major applications such as high-resolution scientific observation.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] This invention provides a method for long baseline interferometric array image reconstruction based on a multi-scale hybrid architecture, comprising:

[0008] Step 1: Construct a dataset and divide it into a training set and a test set; the dataset includes spatial frequency undersampled images and their paired sharp target images;

[0009] Step 2: Input the training set into the image reconstruction model, introduce multi-scale spatial domain information and frequency domain information to construct the loss function of the image reconstruction model, and train the image reconstruction model by constructing multi-dimensional constraints and progressive training strategies.

[0010] Step 3: Quantitatively evaluate the reconstruction effect of the image reconstruction model using the test set. If the test evaluation index is met, a well-trained image reconstruction model is obtained; otherwise, adjust the model parameters and continue training.

[0011] Step 4: Reconstruct the undersampled image to be reconstructed using the trained image reconstruction model;

[0012] The main framework of the image reconstruction model adopts a multi-resolution encoder-decoder structure. The image reconstruction model includes a hybrid sensing module and a shallow feature extraction module. The hybrid sensing module consists of a spatial convolution branch, a frequency convolution branch, and a multi-scale feedforward network. It generates high-level semantic features through multi-level spatial and frequency domain feature extraction and aggregation. The shallow feature extraction module provides the encoder with low-level spatial information from low-level features. It directly transmits multi-resolution inputs to the corresponding scale outputs through skip connections to provide pixel-level image structure support.

[0013] The beneficial effects of this invention are as follows:

[0014] (1) This invention proposes an improved hybrid sensing network model for undersampled image reconstruction of long-baseline optical interferometer arrays. This model replaces the self-attention mechanism in the traditional Transformer structure with a spatial-frequency domain dual convolutional branch structure, reducing computational complexity while integrating the ability to capture local feature details and model global semantic relationships. Furthermore, this invention ensures the consistency between spatial detail features and frequency domain phase information in the feature space through multi-resolution scale feature fusion and spatial-frequency domain loss construction, ensuring that the model generates a structural distribution closer to that of natural images.

[0015] (2) This invention solves the problem of aliasing distortion and edge blurring caused by spatial frequency undersampling in long baseline optical interferometric array imaging scenarios. It can achieve high-quality reconstruction of complex targets under extremely low frequency coverage, and breaks through the resolution bottleneck caused by insufficient baseline in traditional interferometric imaging, providing technical support for high-resolution remote sensing imaging. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the multi-scale hybrid architecture of the image reconstruction model of the present invention.

[0017] Figure 2 This is a schematic diagram of the spatial frequency domain stitching and fusion operation of the hybrid sensing module in the image reconstruction model of the present invention.

[0018] Figure 3 The diagram shows the spatial convolution branch and the frequency convolution branch of the image reconstruction model of the present invention; wherein, (a) is a schematic diagram of the spatial convolution branch; and (b) is a schematic diagram of the frequency convolution branch.

[0019] Figure 4 This is a schematic diagram of an interferometric array imaging model according to an embodiment of the present invention; wherein, (a) is a schematic diagram of the arrangement of 20 sub-stars, (b) is a schematic diagram of the UV sampling distribution, (c) is a clear target image, and (d) is a dirty image.

[0020] Figure 5 The above are test data reconstruction results according to an embodiment of the present invention; wherein, (a) is an undersampled dirty image, (b) is a reconstructed image, and (c) is a clear target image. Detailed Implementation

[0021] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0022] The term “exemplary” as used herein means “serving as an example, embodiment, or illustration.” Any embodiment illustrated herein as “exemplary” is not necessarily to be construed as superior to or better than other embodiments.

[0023] The terms “first”, “second”, etc., are used to distinguish similar objects, not to describe or indicate a specific order or sequence.

[0024] The term "comprising" or any other similar term is intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus / device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent in such process, method, article, or apparatus / device.

[0025] Furthermore, to better illustrate the present invention, numerous specific details are provided in the following detailed embodiments. Those skilled in the art should understand that the present invention can be practiced without certain specific details. In some instances, methods, means, elements, and circuits well known to those skilled in the art have not been described in detail in order to highlight the spirit of the invention.

[0026] Existing deep learning-based methods for reconstructing undersampled images of interferometric arrays only model the local correlations between spatial pixels. In highly undersampled Earth observation tasks, modeling a single spatial feature is insufficient to support the recovery of detailed features. It is necessary to combine frequency domain prior constraints and specially designed learning strategies to guide the network to accurately describe the complex target structural features.

[0027] In view of this, the present invention proposes a long baseline interferometric array image reconstruction method based on a multi-scale hybrid architecture, which reconstructs undersampled images of long baseline optical interferometric arrays through an improved image reconstruction model. Figure 1 A schematic diagram of the multi-scale hybrid architecture of the image reconstruction model of the present invention is shown, such as... Figure 1 As shown, the main framework of this image reconstruction model adopts a multi-resolution encoder-decoder structure. The core module of this model is a hybrid perception module composed of spatial domain convolutional branches, frequency domain convolutional branches, and a multi-scale feedforward network. It generates high-quality high-level semantic features through multi-level spatial and frequency domain feature extraction and aggregation. The image reconstruction model also includes a shallow feature extraction module, which provides the encoder with low-level spatial information from low-level features. Through skip connections, it directly transmits the multi-resolution input to the corresponding scale output to provide pixel-level image structure support.

[0028] For example, the hybrid perception module is constructed by introducing a parallel dual convolutional branch consisting of a spatial convolutional branch and a frequency convolutional branch in the Transformer architecture to replace the self-attention mechanism; wherein, the spatial convolutional branch is used to mine the local spatial correlation at the pixel level; the frequency convolutional branch is used to model the global frequency features of the image as a whole; and the multi-scale feedforward network is used to further refine the contextual information of the spatial-frequency dual-branch aggregated features to enhance the expressive power of complex features.

[0029] For example, the spatial convolution branch selects the first 1 / 4 of the input feature channels for convolution operation, leaving the other channels unchanged, thereby reducing redundant calculations and improving computational efficiency; then the obtained spliced ​​features are rearranged by channels, and the expression of local feature diversity is enhanced by information exchange between channels, so as to capture the subtle structural and texture features in the image.

[0030] For example, the frequency domain convolution branch transforms the input spatial domain features into the frequency domain via discrete Fourier transform for modulation and fusion, and then transforms them back into the spatial domain via inverse Fourier transform. The update of frequency information affects the distribution of global features in the spatial domain, and can establish long-distance dependencies with low computational complexity to enhance the expression of key frequency components.

[0031] For example, the shallow feature extraction module extracts the low-level spatial information through multiple cascaded convolutional layers, and then the feature fusion module in the encoder combines the low-level spatial information with the input features of the encoder to enhance information transmission; the decoder directly learns the residual between the input dirty image (i.e., the input spatial frequency undersampled image) and the clear target image, and adds it pixel by pixel with the input image of the corresponding resolution to obtain the reconstructed image.

[0032] The steps of the long baseline interferometric array image reconstruction method based on a multi-scale hybrid architecture of the present invention include:

[0033] Step 1: Construct a dataset and divide the dataset into a training set and a test set; the dataset includes spatial frequency undersampled images and their paired sharp target images.

[0034] For example, step 1 specifically includes:

[0035] (1) Perform autocorrelation calculation on the array distribution of the interferometric array to obtain the frequency sampling distribution;

[0036] (2) Multiply the frequency sampling distribution with the Fourier transform image of the clear target to obtain the original complex coherence measurement data;

[0037] (3) Perform inverse Fourier transform on the original complex coherence measurement data to obtain the spatial frequency undersampled image;

[0038] (4) Construct the dataset using the spatial frequency undersampled image and its paired sharp target image.

[0039] Step 2: Input the training set into the image reconstruction model, introduce multi-scale spatial domain information and frequency domain information to construct the loss function of the image reconstruction model, and train the image reconstruction model by constructing multi-dimensional constraints and progressive training strategies to improve network convergence efficiency and training stability.

[0040] For example, step 2 specifically includes:

[0041] In the spatial domain, Charbonnier loss function and edge loss function are used, and in the frequency domain, frequency reconstruction L1 loss function is used to optimize low-frequency structure and high-frequency details from low-resolution scale to high-resolution scale. At the same time, a learning rate warm-up strategy and periodic cosine annealing mechanism are introduced during training to balance training stability and accuracy.

[0042] Step 3: Quantitatively evaluate the reconstruction effect of the image reconstruction model using the test set. If the test evaluation index is met, a well-trained image reconstruction model is obtained; otherwise, adjust the model parameters and continue training.

[0043] For example, the test evaluation metric is constructed using Peak Signal-to-Noise Ratio (PSNR), which measures the pixel-level difference between the reconstructed image output by the image reconstruction model and the sharp target image. Step 3 specifically includes:

[0044] Adjust the network learning rate, batch size, spatial frequency domain loss function, and other model parameters based on the PSNR results until the PSNR reaches the expected value, indicating that a high-quality reconstruction output has been obtained. At this point, stop training and obtain the trained image reconstruction model. If the PSNR does not reach the expected value, adjust the model parameters and continue training.

[0045] Step 4: Reconstruct the undersampled image to be reconstructed using the trained image reconstruction model.

[0046] This invention presents a long baseline interferometric array image reconstruction method based on a multi-scale hybrid architecture. First, the original complex coherence measurement data is processed into an undersampled dirty image (i.e., a spatial frequency undersampled image) through inverse Fourier transform, which serves as the model input to construct the dataset. The core model employs a multi-resolution encoder-decoder architecture, embedding a hybrid sensing module to model pixel-level local correlation features and overall global frequency features of the image. In the feature processing flow, spatial semantic features of the image are captured and aggregated at different depths through a multi-level encoder-decoder structure, and the skip connection mechanism of the multi-resolution input is combined to strengthen the support of underlying details. During the training phase, spatial and frequency reconstruction losses are used to iteratively optimize the model parameters, improving image reconstruction performance in complex scenes based on limited sampling data. This invention's method, through hierarchical modeling and fusion of local spatial correlation and global frequency features, can effectively mine the implicit information in the original complex coherence measurement data. In complex target imaging scenarios with extremely low frequency sampling coverage, it effectively reduces ringing effects and artifacts introduced by undersampling, achieving high-quality image reconstruction and is more suitable for high-resolution imaging detection using long baseline interferometric arrays.

[0047] The following detailed description of the long baseline interferometric array image reconstruction method based on a multi-scale hybrid architecture of the present invention, with reference to the accompanying drawings and embodiments, is provided in detail.

[0048] This invention employs a 20-sub-satellite long baseline interferometric array. A schematic diagram of the interferometric array imaging model of this invention is shown below. Figure 4 As shown. The specific implementation steps of the image reconstruction method of the present invention are as follows:

[0049] (1) A numerical simulation research platform for a 20-sub-satellite long baseline interferometric array was established. The optimized array arrangement and UV sampling coverage are as follows: Figure 4 As shown in (a) and (b), 380 effective sampling points were obtained in a 200×200 UV plane matrix through a sparse sampling strategy, with a sampling rate of 0.95%. The dataset used was the publicly available remote sensing image military aircraft target recognition dataset MAR20 (Military Aircraft Recongnition), covering 3842 images of 20 military aircraft models. 3500 high-quality images were selected as the original target data to construct the model training and testing sample library. The imaging process was based on Fourier optics principles. First, a two-dimensional Fourier transform was performed on the original target image, mapping it to the frequency domain. Then, according to the preset UV sampling point distribution, the frequency domain data was selectively acquired, with the frequency values ​​of uncovered areas filled with zeros, forming a sparse frequency domain representation. Finally, a two-dimensional inverse Fourier transform was used to invert the sampled frequency data to the spatial domain, resulting in clear target images and dirty images (i.e., undersampled images). Figure 4 As shown in (c) and (d) in the figure, its imaging process can be simplified to the following formula:

[0050] (1)

[0051] in, Indicates a dirty image. Represents a clear target image. and These represent the Fourier transform and the inverse Fourier transform, respectively. This represents the UV coverage of the interferometer array (i.e., the frequency sampling distribution of the interferometer array).

[0052] (2) Build a hybrid sensing module based on the Transformer structure. Figure 2 This is a schematic diagram of the spatial-frequency domain stitching and fusion operation of the hybrid sensing module, as shown below. Figure 2 As shown, the attention mechanism is replaced by a spatial-frequency domain dual convolution branch, and a multi-scale feedforward network is introduced to replace the traditional single-scale feedforward network, thereby reducing computational complexity and achieving multi-level analysis of multi-dimensional features.

[0053] Specifically, the input features of the hybrid sensing module double in number after convolution operation and are divided into two paths for input to the spatial domain convolution branch and the frequency domain convolution branch. Figure 3 This diagram illustrates the spatial and frequency domain convolution branches. The spatial convolution branch employs partial convolution to further reduce redundant computation and leverages channel rearrangement to reorganize and enhance local features, effectively improving the diversity of feature representation. Figure 3 As shown in (a); the frequency domain convolution branch transforms the input features to the frequency domain via Fourier transform, modulates the frequency domain features through convolution, and finally transforms them back to the spatial domain via inverse Fourier transform, as shown in (a). Figure 3 As shown in (b) above, the output features of the two branches are concatenated and then processed through channel attention and convolutional layers to enhance the expression of key features, forming spatial-frequency domain aggregated features. Finally, the multi-scale feedforward network extracts local feature associations from different receptive fields through a multi-kernel convolution strategy, further enhancing the expressive power of complex features.

[0054] (3) Construct a 6-layer encoding-decoding framework. For example... Figure 1 As shown, the encoding stage consists of three encoder levels. Each encoder first extracts low-level visual information from the input features through a 3×3 convolutional layer, then introduces a hybrid perception module for deep feature mining, and performs multiple downsampling operations to gradually convert the input data into feature maps with decreasing spatial resolution and increasing channel dimensions. To enhance the efficiency of low-level spatial information transmission, a feature fusion module is embedded in the intermediate layer between encoder 2 and encoder 3 to interactively aggregate the shallow features of the input image at the corresponding resolution with the features output by the 3×3 convolutional layer. In the decoding stage, the features output by the encoder are upsampled through deconvolution operations to gradually restore the spatial size and channel dimensions of the feature maps. During this process, the decoder outputs residual information between the dirty image and the clear target image, and uses a skip connection mechanism to fuse the input image at the corresponding resolution with the residual pixel by pixel, ultimately generating a multi-scale reconstructed image.

[0055] (4) Multi-scale loss functions are introduced to optimize the stability and accuracy of model training. Among them, spatial Charbonnier loss reduces the model's sensitivity to data noise and outliers due to its smoothing properties, enhancing the robustness of the training process; spatial edge loss indirectly improves the rationality of the overall image structure by constraining the accuracy of edges; frequency reconstruction L1 loss strengthens the recovery of high-frequency components and specifically compensates for the excessive smoothing caused by spatial loss, ensuring the richness of details in the reconstructed image. In terms of the loss function construction strategy, a hierarchical supervision mechanism is adopted, which effectively avoids blurring or artifacts caused by insufficient single-scale supervision information, thereby obtaining a more natural reconstructed image. Total loss function The expression is:

[0056] (2)

[0057] in, This represents the Charbonnier loss function. This represents the frequency reconstruction L1 loss function. Represents the edge loss function; , , The balance coefficients of the total loss function are set to 0.1, 0.2, and 0.05, respectively. The expressions for the three loss functions are as follows:

[0058] (3)

[0059] (4)

[0060] (5)

[0061] in, This indicates the number of multi-scale outputs, which is set to 3 here; and They represent the network's first and second generations respectively. Reconstructed images and sharp target images output at each scale; Indicates Fourier transform; Represents the Laplace operator; It is a non-zero positive number.

[0062] (5) The PSNR index is used to quantitatively evaluate the reconstruction performance of the image reconstruction model on the undersampled image. This index directly calculates the gray values ​​of the reconstructed image and the clear target image pixel by pixel. By quantifying the pixel-level error, it reflects the degree of image distortion. The higher the PSNR value, the higher the similarity between the reconstructed image and the original target image in terms of gray distribution and detail preservation. The formula for calculating PSNR is as follows:

[0063] (6)

[0064] in, and These represent the grayscale matrix of the clear target image and the grayscale matrix of the reconstructed image, respectively. and These represent the length and width of the image grayscale matrix, respectively.

[0065] (6) Model training was conducted on a deep learning server running Ubuntu 20.04 LTS, equipped with an NVIDIA GeForce RTX 4090 graphics card for computing power. Simulation studies showed that effective aircraft type discrimination could only be achieved when the PSNR of the reconstructed image exceeded 17dB. Based on this, the model parameters were further optimized to improve reconstruction performance until the optimal index was reached. The training process used 3000 sets of data as the training set and 500 sets of data as the test set. The network achieved stable convergence after 180 training cycles. In addition, the predicted frequency components of the sampling locations were directly replaced with the measured sampling frequency data to reduce reconstruction errors. The image reconstruction test results are as follows: Figure 5 As shown, Figure 5 In the figures (a), (b), and (c), the images are the undersampled dirty image, the reconstructed image, and the clear target image, respectively. Quantitatively, the average PSNR of the undersampled dirty image test set was 14.02 dB, while the average PSNR of the reconstructed image increased to 19.19 dB, an increase of 5.17 dB. Visually, the information loss caused by the sparse baseline resulted in severe aliasing artifacts and geometric distortion in the undersampled dirty image, making the target structure difficult to identify. After model reconstruction, the key contour features and local details of the aircraft were effectively restored, and the image quality was significantly improved.

[0066] The contents not described in detail in this specification are existing technologies known to those skilled in the art.

[0067] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the present invention. The above descriptions are exemplary and not exhaustive. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A long-baseline interferometric array image reconstruction method based on a multi-scale hybrid architecture, characterized in that, The method comprises the following steps: Step 1, constructing a data set and dividing the data set into a training set and a test set; the data set comprises a spatial frequency undersampling image and a paired clear target image thereof; Step 2, inputting the training set into an image reconstruction model, introducing multi-scale spatial domain information and frequency domain information to construct a loss function of the image reconstruction model, and training the image reconstruction model by constructing a multi-dimensional constraint condition and a progressive training strategy; Step 3, quantitatively evaluating the reconstruction effect of the image reconstruction model by using the test set, obtaining a trained image reconstruction model when the test evaluation index is met, or adjusting the model parameters and continuing the training; Step 4, reconstructing the undersampling image to be reconstructed by using the trained image reconstruction model; The main framework of the image reconstruction model adopts a multi-resolution encoder-decoder structure; the image reconstruction model comprises a hybrid perception module and a shallow feature extraction module; the hybrid perception module is composed of a spatial domain convolution branch, a frequency domain convolution branch and a multi-scale feedforward network, and high-level semantic features are generated by multi-level spatial and frequency domain feature extraction and aggregation; the shallow feature extraction module provides bottom space information in low-level features for the encoder, and directly transmits multi-resolution input to the output of the corresponding scale through a skip connection to provide pixel-level image structure support.

2. The method of claim 1, wherein, The step 1 specifically comprises: performing autocorrelation calculation on the array distribution of the interference array to obtain a frequency sampling distribution; point-multiplying the frequency sampling distribution and the Fourier transform image of the clear target to obtain original complex coherence measurement data; performing inverse Fourier transform on the original complex coherence measurement data to obtain the spatial frequency undersampling image; constructing the data set by using the spatial frequency undersampling image and the paired clear target image thereof.

3. The method of claim 1, wherein, The hybrid perception module is constructed by introducing a parallel double convolution branch composed of a spatial domain convolution branch and a frequency domain convolution branch into a Transformer architecture to replace the self-attention mechanism; The spatial domain convolution branch is used to mine the local spatial correlation at the pixel level; the frequency domain convolution branch is used to model the global frequency features of the whole image; and the multi-scale feedforward network is used to further extract the context information of the spatial-frequency dual-branch aggregated features to enhance the expression ability of complex features.

4. The method of claim 3, wherein, The spatial domain convolution branch selects the first 1 / 4 channels from the input feature channels for convolution operation, and the remaining channels remain unchanged, so as to reduce redundant calculation and improve calculation efficiency; then the obtained spliced features are rearranged in channels, the local feature diversity expression is enhanced through information exchange between channels, and subtle structure and texture features in the image are captured.

5. The method of claim 3, wherein, The frequency domain convolution branch converts the input spatial domain features into the frequency domain for modulation and fusion through discrete Fourier transform, and then converts them back to the spatial domain through inverse Fourier transform; wherein the distribution of global features in the spatial domain is affected by the update of frequency information, which can establish long-distance dependence with low computational complexity to enhance the expression of key frequency components.

6. The method of claim 1, wherein, The shallow feature extraction module extracts the bottom layer spatial information through multiple cascaded convolution layers, and then combines the bottom layer spatial information with input features of the encoder through a feature fusion module in the encoder to enhance information transmission; The decoder directly learns the residual between the input dirty image and the clear target image, and the reconstructed image is obtained by pixel-by-pixel addition with the input image of the corresponding resolution.

7. The method of claim 1, wherein, The step 2 specifically comprises: In the spatial domain, a Charbonnier loss function and an edge loss function are adopted, and in the frequency domain, a frequency reconstruction L1 loss function is adopted, so as to optimize low-frequency structures and high-frequency details from a low-resolution scale to a high-resolution scale; meanwhile, a learning rate warm-up strategy and a periodic cosine annealing mechanism are introduced in the training process to balance the training stability and precision.

8. The method of claim 1, wherein, The test evaluation index is constructed by using a peak signal-to-noise ratio (PSNR), which is used to measure the pixel-level difference between the reconstructed image output by the image reconstruction model and the clear target image; the step 3 specifically comprises: According to the PSNR result, the model parameters are adjusted until the PSNR reaches an expected value, so as to obtain a trained image reconstruction model; the model parameters include a learning rate, a batch size and spatial-frequency domain loss functions.

Images