A solar image feature enhancement method and system fusing space-frequency domain information

Abstract

The application discloses a kind of solar image feature enhancement method and system based on space-frequency domain feature learning.The method comprises: obtaining solar image training set;The image to be processed is input into the solar feature enhancement model based on space-frequency domain feature collaborative learning, and the feature enhancement result is obtained, the space-frequency domain collaborative enhancement loss function is constructed based on feature enhancement result and high-quality reference image, and the solar feature enhancement model is trained based on space-frequency domain collaborative enhancement loss function;Obtain the solar image to be tested, input into the solar feature enhancement model trained, and obtain the solar image after feature enhancement.The application can effectively improve the feature expression capability of solar image under complex imaging conditions, and has significant advantages in overall radiation consistency maintenance, small structure detail enhancement in solar surface area, noise suppression and quality improvement of structural degradation image.

Description

Technical Field

[0001] This invention relates to the field of astronomical solar observation technology, and in particular to a method and system for enhancing solar image features by fusing spatial and frequency domain information. Background Technology

[0002] Solar activity is the core driving force influencing changes in the Sun-Earth space environment. Events such as solar flares, coronal mass ejections, and the release of high-energy particle streams trigger intense electromagnetic radiation and charged particle streams, causing severe disturbances in the Earth's magnetosphere and ionosphere. This, in turn, affects satellite attitude control, navigation signal accuracy, communication link stability, and the safe operation of power systems. Therefore, continuous and accurate monitoring of solar activity processes has significant scientific and applied value for space weather forecasting and disaster prevention.

[0003] As a key payload for observing the structure of the solar corona and disk, the space solar telescope can acquire multi-band radiation information with high spatiotemporal resolution, providing important data support for studying the mechanisms of solar atmospheric motion, coronal dynamics, and magnetic reconnection processes. However, due to factors such as the complex space environment and hardware limitations, images acquired by space solar telescopes generally suffer from varying degrees of solar radiation and structural feature attenuation, including low contrast of solar surface structures and mixed noise interference. This directly affects the accuracy of downstream tasks based on feature recognition, such as solar active region identification, flare prediction, and coronal mass ejection detection.

[0004] Most existing solar image feature enhancement methods, such as radiometric calibration, flat-field correction, or spatial enhancement filters, primarily focus on improving visual quality or suppressing local noise. They struggle to simultaneously preserve and enhance multi-scale features crucial for downstream tasks in both the spatial and frequency domains. While emerging deep learning-based feature enhancement methods have improved processing accuracy, most models are directly transferred from natural image processing tasks. Coupled with their limited receptive fields, these methods often result in over-smoothing of high-frequency features (such as fine flare structures) or distortion in compensation for low-frequency background features (such as coronal background radiation). Consequently, they fall short in feature discriminability and other aspects, hindering the accuracy of downstream analysis and models. Summary of the Invention

[0005] This invention provides a solar image feature enhancement method that integrates spatial and frequency domain information. It aims to improve the quality of discriminative features in solar images for downstream analysis tasks, effectively overcome the interference of complex hybrid degradation on feature extraction, and has significant advantages in preserving and enhancing the features of fine solar structures such as flares, coronal rings, and filaments. This invention provides a method for enhancing solar image features by fusing spatial and frequency domain information, including: Obtain a solar image training set; wherein, the solar image training set includes an image to be enhanced and a high-quality reference image corresponding to the imaging time of the image to be enhanced, for guiding feature learning; The image to be enhanced is input into the solar image feature enhancement model to obtain the feature enhancement result. Based on the feature enhancement result and the high-quality reference image, a spatial-frequency domain collaborative enhancement loss function is constructed. The solar image feature enhancement model is trained based on the spatial-frequency domain collaborative enhancement loss function. Acquire a solar image to be tested, and input the solar image to be tested into a trained solar image feature enhancement model to obtain a solar image with enhanced features.

[0006] Furthermore, the solar image feature enhancement model includes a generator and a discriminator; the generator includes an encoding module, several spatial frequency domain residual interaction guidance modules, a decoding module, a global residual connection layer, and a normalized activation layer.

[0007] Furthermore, the image to be enhanced is input into the solar image feature enhancement model to obtain the feature enhancement result, including: The image to be processed is input into the encoding module to obtain a shallow encoded feature map; The shallow coding feature map is input into the spatial-frequency domain residual interaction guidance module to extract and fuse spatial and frequency domain information to obtain spatial-frequency domain fused features. The spatial-frequency domain fusion feature output from the last spatial-frequency domain residual interactive guidance module is input into the decoding module to obtain the decoding feature map; The image to be processed and the decoded feature map are input into the global residual connection layer for fusion, and then the fusion result is input into the normalized activation layer to obtain the feature enhancement result.

[0008] Furthermore, the spatial frequency domain residual interaction guidance module is composed of two modulated fast Fourier convolution units and a residual connection layer connected in series.

[0009] Furthermore, the modulated fast Fourier convolutional unit splits the input features into local features and global features along the channel dimension; The local features are input into two standard convolutional layers respectively to obtain local spatial features. Figure 1 and local spatial features Figure 2 ; The global features are processed by a standard convolutional layer and a frequency domain processing unit to obtain a global spatial feature map and a frequency domain enhanced feature map. The local spatial features Figure 1The local features are fused with the global spatial feature map and then input into batch normalization and nonlinear activation units to obtain optimized local features. The local spatial features Figure 2 The feature is fused with the frequency domain enhanced feature map and then input into the batch normalization and nonlinear activation unit to obtain the optimized global feature; The optimized local features and optimized global features are reintegrated into context-enhanced features through channel concatenation operations; The global residual connection layer further adds and fuses the input features of the spatial-frequency domain residual interaction guidance module with the context enhancement features output by the last modulated fast Fourier convolution unit to obtain spatial-frequency domain fused features.

[0010] Furthermore, the frequency domain processing unit includes a fast Fourier transform layer, a complex convolution layer, a batch normalization layer, a modulation-type complex activation layer, and an inverse fast Fourier transform layer. In the frequency domain processing unit, the input feature map is first subjected to a two-dimensional fast Fourier transform to map the spatial domain features to a complex representation in the frequency domain. Then, the real and imaginary parts of the spectrum are respectively input into independent 1×1 convolutional layers for complex convolution operations. The complex weight mapping is completed through the linear combination of the real and imaginary parts, thereby realizing the joint modeling of the spectrum amplitude and phase. After complex convolution processing, the output results of the real and imaginary parts are batch normalized to stabilize the network training process and improve the consistency of feature distribution. Then, the normalized real and imaginary parts are recombined into a complex spectrum and nonlinearly activated through a modulated complex activation layer. Finally, the activated frequency domain features are restored to the spatial domain through inverse fast Fourier transform, resulting in the frequency domain enhanced feature map containing global frequency information and phase directionality.

[0011] Furthermore, the global residual connection layer performs element-wise addition of the output of the encoding module with the original input to form a global residual fusion result, which enables the network to focus on learning enhancement features while maintaining the overall brightness and structural information of the original input image; The decoding module uses n transposed convolutional modules in series and a convolutional optimization layer to perform feature decoding and reconstruction. Each transposed convolutional module includes a transposed convolutional layer, a batch normalization layer, and a ReLU nonlinear activation layer to progressively restore the spatial resolution of the feature map and reconstruct detailed texture information. Finally, a convolutional optimization layer is used to further optimize the decoded features.

[0012] Furthermore, the spatial-frequency domain collaborative enhancement loss function is: ; in , and These are the set weight parameters. For spatial frequency domain collaborative enhancement loss function; Spatial feature enhancement loss is used to enhance solar contrast and radiance information; This is a frequency domain structure enhancement loss, used to enhance frequency-related features such as solar details and periodic structures; To generate adversarial loss, which is used to improve the realism and physical consistency of feature augmentation results.

[0013] Furthermore, spatial feature enhancement loss Aiming to minimize the feature enhancement results of the model output and reference images Pixel-level differences in the spatial domain, spatial feature enhancement loss Represented as: ; in, Expressing expectations, Indicates the L1 distance; Frequency domain structure enhancement loss minimizes the feature enhancement result of the model output. and reference images The spectral distribution gap in the frequency domain constrains frequency domain feature learning to enhance the model's ability to augment global structure and periodic features, as expressed as: ; in, Represents the two-dimensional Fast Fourier Transform operator; Generate adversarial loss The aim is to generate adversarial learning patterns and train the discriminator alternately. D and generator G This is used to assist the generator in producing higher quality feature enhancement results, represented as: ; During the training of the solar image feature enhancement model, the network model parameters are continuously optimized through the stochastic gradient descent algorithm and backpropagation mechanism until the spatial-frequency domain collaborative enhancement loss function is converged by comparing the model output results with the reference image, and then the training ends.

[0014] The present invention also provides a solar image feature enhancement system that integrates spatial and frequency domain information, including a processor and a memory. The memory is used to store program instructions, and the processor is used to call the program instructions in the memory to execute a solar image feature enhancement method that integrates spatial and frequency domain information as described in the above technical solution.

[0015] This invention provides a solar image feature enhancement method based on the fusion of spatial and frequency domain information. It can achieve high-precision structural restoration and detail enhancement for complex degradation phenomena in solar extreme ultraviolet and other multi-band imaging data. Compared with existing methods based solely on spatial filtering or traditional enhancement algorithms, this invention has the following advantages: ① The spatial-frequency domain joint modeling method of this invention introduces a spatial-frequency domain residual interaction guidance module into the generator network. By combining spatial convolution and frequency convolution, it realizes the collaborative modeling of local detail features and global context features, thereby capturing the local texture information and long-range dependencies of solar images, effectively achieving feature enhancement under complex degradation conditions, and improving the contrast and discriminability of solar surface feature structures.

[0016] ② The frequency domain processing unit proposed in this invention can perform learnable convolution operations in the complex domain and introduce a modulated complex activation function to maintain phase stability, effectively separating and weakening erroneous features introduced by spatially periodic interference artifacts, geometric blurs, etc., improving feature purity while significantly improving the structural integrity of solar images.

[0017] ③ The spatial-frequency domain collaborative enhancement loss function proposed in this invention enables the network model to achieve collaborative enhancement of features such as structure, brightness, and texture during the optimization process. This strategy effectively improves the training stability and generalization ability of the model, avoids overfitting and excessive smoothing, and accelerates the convergence speed of the model.

[0018] ④ The method of the present invention does not depend on specific imaging equipment or bands and can be applied to image feature enhancement tasks of various solar observation payloads. The method has strong adaptability and automation features, and can achieve effective purification and enhancement of solar features without human intervention. It has good versatility and engineering application prospects. Attached Figure Description

[0019] The technical solution and other beneficial effects of this application will become apparent from the following detailed description of specific embodiments in conjunction with the accompanying drawings.

[0020] Figure 1 The flowchart illustrates a solar image feature enhancement method based on the fusion of spatial and frequency domain information, as provided in an embodiment of the present invention.

[0021] Figure 2 This is a schematic diagram of the structure of the solar image feature enhancement model provided in an embodiment of the present invention.

[0022] Figure 3 This is a schematic diagram of the structure of the modulation-type fast Fourier convolution unit and the frequency domain processing unit provided in the embodiments of the present invention.

[0023] Figure 4The image shows a comparison of the feature enhancement before and after, as provided in an embodiment of the present invention.

[0024] Figure 5 The diagram shows a comparison between the method provided in this embodiment of the invention and existing image feature enhancement methods. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] Please see Figure 1 , Figure 1 A flowchart of a solar image feature enhancement method based on fusion of spatial and frequency domain information provided in this embodiment of the invention includes the following steps: S1, Obtain the solar image training set; wherein, the solar image training set includes a large number of images to be processed and a large number of high-quality reference images corresponding to the imaging time of the images to be processed.

[0027] Specifically, the solar image training set should include multiple pairs of spatiotemporally consistent low-quality solar remote sensing images to be processed and their corresponding high-quality solar reference remote sensing images to ensure a sufficient data foundation for training the solar feature enhancement network model, thereby ensuring the model's feature enhancement accuracy and generalization.

[0028] S2, input the image to be processed into the solar image feature enhancement model to obtain the feature enhancement result; construct the spatial-frequency domain collaborative enhancement loss function based on the feature enhancement result and the reference image; train the solar image feature enhancement model based on the spatial-frequency domain collaborative enhancement loss function.

[0029] Figure 2 This is a schematic diagram of the structure of the solar image feature enhancement model provided in an embodiment of the present invention, as shown below. Figure 2 As shown, the solar image feature enhancement model innovatively adopts a spatial-frequency domain feature collaborative learning mode. The overall solar image feature enhancement model includes a generator G and a discriminator D. Step S2 specifically includes the following steps: S21, input the image to be processed into the generator to obtain the feature enhancement result.

[0030] Specifically, the generator G includes an encoding module, n spatial frequency domain residual interaction guidance modules, a decoding module, a global residual connection layer, and a normalized activation layer. Step S21 includes the following steps: S211, Input the image to be processed into the encoding module to obtain a shallow encoded feature map.

[0031] In one embodiment, the encoding module includes n concatenated convolutional modules. Each convolutional module includes a convolutional layer, a batch normalization layer, and a LeakyReLU activation function to extract multi-scale shallow encoding features. The first convolutional module has a 7×7 kernel to capture a large range of features. The subsequent n-1 convolutional modules have 4×4 kernels to downsample stepwise, reducing the width and height of the feature map and decreasing the computational memory burden of subsequent modules. In this embodiment, n is 4.

[0032] S212, input the shallow coding feature map into the spatial-frequency domain residual interaction guidance module to extract and fuse spatial and frequency domain information to obtain spatial-frequency domain fused features.

[0033] like Figure 2 As shown, the spatial frequency domain residual interaction guidance module is located in the network bottleneck layer, and there are n modules in total. Each spatial frequency domain residual interaction guidance module consists of two modulated fast Fourier convolutional units and a residual connection layer connected in series; in this embodiment, n is 6.

[0034] Figure 3 This is a schematic diagram of the structure of a modulation-type fast Fourier convolution unit and a frequency domain processing unit, as shown below. Figure 3 As shown, the modulated fast Fourier convolutional unit splits the input features into local and global features along the channel dimension; the local features are then fed into two 3×3 standard convolutional layers to obtain local spatial features. Figure 1 and local spatial features Figure 2 The global features are then processed by a standard convolutional layer and a frequency domain processing unit, respectively, to obtain a global spatial feature map and a frequency domain enhanced feature map; subsequently, the local spatial features... Figure 1 The local spatial features are fused with the global spatial feature map and then input into a batch normalization layer and a nonlinear activation layer to obtain optimized local features; local spatial features Figure 2 The feature is fused with the frequency domain enhanced feature map, and then input into the batch normalization layer and nonlinear activation layer to obtain the optimized global feature; finally, the optimized local feature and the optimized global feature are re-integrated into the context-enhanced feature through channel concatenation operation. like Figure 3 As shown, the frequency domain processing unit includes a fast Fourier transform layer, a complex convolution layer, a batch normalization layer, a modulation complex activation layer, and an inverse fast Fourier transform layer. Specifically, the frequency domain processing unit first processes the input feature map. Performing a two-dimensional Fast Fourier Transform (FFT) maps the spatial domain features to a complex frequency domain representation, yielding the spectral features. : (1) in, This represents the two-dimensional orthogonal fast Fourier transform operator. These represent the real part and the imaginary part, respectively.

[0035] Then, in the frequency domain, two independent 1×1 learnable convolution kernels are used to perform convolution operations on the real and imaginary parts respectively to achieve complex convolution mapping. Let the real and imaginary parts of the convolution kernel be respectively... and Then the real and imaginary parts of the complex convolution result can be expressed as follows: (2) in, This represents a two-dimensional convolution operation. After complex convolution processing, the convolution result is... and Batch normalization is performed to stabilize the network training process and improve the consistency of feature distribution, resulting in normalized results. and And recombine into a complex spectrum. .

[0036] Subsequently, nonlinear activation is performed through a modulated complex activation layer using the ModReLU activation function. This allows for adaptive modulation of the amplitude while preserving the phase information of the complex features, thereby achieving nonlinear enhancement of the frequency domain features. For complex features... Where r is the amplitude, Let b be the phase and b be the learnable channel bias parameter. Then the ModReLU activation function can be expressed as: (3) Finally, after the above-described modulation-type complex activation, the output complex spectrum is denoted as... Then through inverse fast Fourier transform ,Will Return to airspace: (4) Obtain a frequency domain enhanced feature map containing global frequency information and phase directionality. This is used to perform subsequent spatial-frequency domain feature fusion processing in the modulated fast Fourier convolution unit, ultimately obtaining context-enhanced features.

[0037] like Figure 2As shown, the residual connection layer further adds and fuses the input features of the spatial-frequency domain residual interaction guidance module with the context enhancement features output by the last modulated fast Fourier convolution unit to obtain the spatial-frequency domain fused features.

[0038] S213, input the spatial frequency domain fusion feature output by the last spatial frequency domain residual interactive guidance module into the decoding module to obtain the decoding feature map.

[0039] Specifically, the decoding module uses n cascaded transposed convolutional modules and a convolutional optimization layer for feature decoding and reconstruction. Each transposed convolutional module sequentially includes a transposed convolutional layer, a batch normalization layer, and a ReLU nonlinear activation layer to progressively restore the spatial resolution of the feature map and reconstruct detailed texture information. Finally, a 7×7 convolutional layer further optimizes the decoded features.

[0040] S214, the image to be processed and the decoded feature map are input into the global residual connection layer for fusion, and then the fusion result is input into the normalized activation layer to obtain the feature enhancement result.

[0041] Specifically, the global residual connection layer first performs element-wise addition of the output of the encoding module with the original input to form a global residual fusion result. This allows the network to focus on learning enhanced features while maintaining the overall brightness and structural information of the original input image. Finally, the normalized activation layer uses the Tanh activation function to compress and map the global residual fusion result to the range of [-1,1]. This improves the stability of the network while effectively suppressing individual enhanced aberration pixels, resulting in the final feature enhancement result.

[0042] S22, the feature enhancement result and the reference image are input into the discriminator to distinguish between real and fake, and the generator is assisted to output high-fidelity feature enhancement result based on the discrimination result.

[0043] Specifically, the discriminator D consists of multiple feature extraction convolutional modules and one output convolutional layer connected in series. Each feature extraction convolutional module consists of a 4×4 convolution, a normalization layer, and a LeakyReLU activation function. The first few convolutional blocks use a stride of s=2 to achieve progressive downsampling and expand the receptive field. In subsequent convolutional blocks, a stride of s=1 is used to maintain denser local discrimination. Finally, a 4×4 input convolutional layer maps the features to a single-channel discrimination result, resulting in a block-level discrimination map used to distinguish the generator output from the reference image. This allows the generator to generate high-fidelity feature enhancement results by calculating adversarial loss.

[0044] S23. Based on the feature enhancement results and reference images, a spatial-frequency domain collaborative enhancement loss function is constructed, and the solar image feature enhancement model is trained based on the spatial-frequency domain collaborative enhancement loss function.

[0045] Specifically, the spatial-frequency domain collaborative enhancement loss function is a weighted combination of spatial domain feature enhancement loss, frequency domain structure enhancement loss, and generative adversarial loss: (5) in , and These are the set weight parameters. For the spatial frequency domain collaborative enhancement loss function, For spatial feature enhancement loss, For frequency domain structure enhancement loss, To generate adversarial losses.

[0046] Among them, spatial feature enhancement loss Aiming to minimize the feature enhancement results of the model output and reference images The pixel-level difference in spatial domain between them. Spatial feature enhancement loss. It can be represented as: (6) in, Expressing expectations, This represents the L1 distance.

[0047] Frequency domain structure enhancement loss minimizes the feature enhancement result of the model output. and reference images The spectral distribution gap in the frequency domain, constraining frequency domain feature learning to enhance the model's ability to augment global structure and periodic features, can be expressed as: (7) in, This represents the two-dimensional Fast Fourier Transform (FFT) operator.

[0048] Generate adversarial loss The aim is to generate adversarial learning patterns and train the discriminator alternately. D and generator G To assist the generator in producing higher quality feature enhancement results, it can be represented as: (8) During the training of the solar image feature enhancement model, the network model parameters are continuously optimized through the stochastic gradient descent algorithm and backpropagation mechanism until the spatial-frequency domain collaborative enhancement loss function is converged by comparing the model output results with the reference image, and then the training ends.

[0049] S3. Obtain the solar image to be tested, input the solar image to be tested into the trained solar image feature enhancement model, and obtain the feature-enhanced solar image.

[0050] Specifically, by inputting the solar image to be tested, which is of the same origin as the low-quality images in the training set, into the trained solar image feature enhancement model, a solar image with high signal-to-noise ratio and high contrast can be obtained after feature enhancement. The overall structure of the image is highly interpretable, and the details of the solar surface and corona are clear. Figure 4 This is a comparison diagram showing the effect of feature enhancement before and after in an embodiment of this application. Figure 5 The comparison diagrams provided in this application embodiment show the effect of this method compared with existing feature enhancement methods. Figure 4 and Figure 5 It is evident that the method proposed in this invention has significant advantages in improving the overall image radiation quality, enhancing the small feature structure of the solar region, and suppressing noise and structural degradation (such as grids, bright spots, bright bands, etc.).

[0051] On the other hand, embodiments of the present invention also provide a solar image feature enhancement system that integrates spatial and frequency domain information, including a processor and a memory. The memory is used to store program instructions, and the processor is used to call the program instructions in the memory to execute a solar image feature enhancement method that integrates spatial and frequency domain information as described in the above technical solution.

[0052] The above provides a detailed description of a solar image feature enhancement method based on the fusion of spatial and frequency domain information provided in the embodiments of this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The description of the above embodiments is only for the purpose of helping to understand the method and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for enhancing solar image features by fusing spatial and frequency domain information, characterized in that, The method includes: Obtain a solar image training set; wherein, the solar image training set includes an image to be enhanced and a high-quality reference image corresponding to the imaging time of the image to be enhanced, for guiding feature learning; The image to be enhanced is input into the solar image feature enhancement model to obtain the feature enhancement result. Based on the feature enhancement result and the high-quality reference image, a spatial-frequency domain collaborative enhancement loss function is constructed. The solar image feature enhancement model is trained based on the spatial-frequency domain collaborative enhancement loss function. Acquire a solar image to be tested, and input the solar image to be tested into a trained solar image feature enhancement model to obtain a solar image with enhanced features.

2. The solar image feature enhancement method according to claim 1, characterized in that: The solar image feature enhancement model includes a generator and a discriminator; the generator includes an encoding module, several spatial frequency domain residual interaction guidance modules, a decoding module, a global residual connection layer and a normalized activation layer.

3. The solar image feature enhancement method according to claim 2, characterized in that: The image to be enhanced is input into the solar image feature enhancement model to obtain the feature enhancement result, including: The image to be processed is input into the encoding module to obtain a shallow encoded feature map; The shallow coding feature map is input into the spatial-frequency domain residual interaction guidance module to extract and fuse spatial and frequency domain information to obtain spatial-frequency domain fused features. The spatial-frequency domain fusion feature output from the last spatial-frequency domain residual interactive guidance module is input into the decoding module to obtain the decoding feature map; The image to be processed and the decoded feature map are input into the global residual connection layer for fusion, and then the fusion result is input into the normalized activation layer to obtain the feature enhancement result.

4. The solar image feature enhancement method according to claim 2, characterized in that: The spatial frequency domain residual interaction guidance module consists of two modulated fast Fourier convolutional units and a residual connection layer connected in series.

5. The solar image feature enhancement method according to claim 4, characterized in that: The modulated fast Fourier convolutional unit splits the input features into local features and global features along the channel dimension; The local features are respectively input into two standard convolutional layers to obtain local spatial feature map 1 and local spatial feature map 2; The global features are processed by a standard convolutional layer and a frequency domain processing unit to obtain a global spatial feature map and a frequency domain enhanced feature map. The local spatial feature map 1 is fused with the global spatial feature map, and then input into the batch normalization and nonlinear activation unit to obtain the optimized local features; The local spatial feature map 2 is fused with the frequency domain enhanced feature map, and then input into the batch normalization and nonlinear activation unit to obtain the optimized global features; The optimized local features and optimized global features are reintegrated into context-enhanced features through channel concatenation operations; The global residual connection layer further adds and fuses the input features of the spatial-frequency domain residual interaction guidance module with the context enhancement features output by the last modulated fast Fourier convolution unit to obtain spatial-frequency domain fused features.

6. The solar image feature enhancement method according to claim 5, characterized in that: The frequency domain processing unit includes a fast Fourier transform layer, a complex convolution layer, a batch normalization layer, a modulation-type complex activation layer, and an inverse fast Fourier transform layer. In the frequency domain processing unit, the input feature map is first subjected to a two-dimensional fast Fourier transform to map the spatial domain features to a complex representation in the frequency domain. Then, the real and imaginary parts of the spectrum are respectively input into independent 1×1 convolutional layers for complex convolution operations. The complex weight mapping is completed through the linear combination of the real and imaginary parts, thereby realizing the joint modeling of the spectrum amplitude and phase. After complex convolution processing, the output results of the real and imaginary parts are batch normalized to stabilize the network training process and improve the consistency of feature distribution. Subsequently, the normalized real and imaginary parts are recombine into a complex spectrum, and nonlinear activation is performed through a modulated complex activation layer; Finally, the activated frequency domain features are restored to the spatial domain through inverse fast Fourier transform, resulting in the frequency domain enhanced feature map containing global frequency information and phase directionality.

7. The solar image feature enhancement method according to claim 2, characterized in that: The global residual connection layer performs element-wise addition of the output of the encoding module with the original input to form a global residual fusion result, which enables the network to focus on learning enhancement features while maintaining the overall brightness and structural information of the original input image. The decoding module uses n transposed convolutional modules in series and a convolutional optimization layer to perform feature decoding and reconstruction. Each transposed convolutional module includes a transposed convolutional layer, a batch normalization layer, and a ReLU nonlinear activation layer to progressively restore the spatial resolution of the feature map and reconstruct detailed texture information. Finally, a convolutional optimization layer is used to further optimize the decoded features.

8. The solar image feature enhancement method according to claim 1, characterized in that: The spatial frequency domain collaborative enhancement loss function is: ； in , and These are the set weight parameters. For spatial frequency domain collaborative enhancement loss function; Spatial feature enhancement loss is used to enhance solar contrast and radiance information; This is a frequency domain structure enhancement loss, used to enhance frequency-related features such as solar details and periodic structures; To generate adversarial loss, which is used to improve the realism and physical consistency of feature augmentation results.

9. A method for enhancing solar image features by fusing spatial and frequency domain information as described in claim 8. Its features are: Spatial feature enhancement loss Aiming to minimize the feature enhancement results of the model output and reference images Pixel-level differences in the spatial domain, spatial feature enhancement loss Represented as: ； in, Expressing expectations, Indicates the L1 distance; Frequency domain structure enhancement loss minimizes the feature enhancement result of the model output. and reference images The spectral distribution gap in the frequency domain constrains frequency domain feature learning to enhance the model's ability to augment global structure and periodic features, as expressed as: ； in, Represents the two-dimensional Fast Fourier Transform operator; Generate adversarial loss The aim is to generate adversarial learning patterns and train the discriminator alternately. D and generator G This is used to assist the generator in producing higher quality feature enhancement results, represented as: ； During the training of the solar image feature enhancement model, the network model parameters are continuously optimized through the stochastic gradient descent algorithm and backpropagation mechanism until the spatial-frequency domain collaborative enhancement loss function is converged by comparing the model output results with the reference image, and then the training ends.

10. A solar image feature enhancement system that integrates spatial and frequency domain information, characterized in that: It includes a processor and a memory, the memory being used to store program instructions, and the processor being used to call the program instructions in the memory to execute the solar image feature enhancement method that fuses spatial-frequency domain information as described in any one of claims 1-9.

Images