Color eye fundus image detail enhancement method of mixed double-domain structure

Abstract

The invention relates to the technical field of medical image processing and deep learning, in particular to a color eye fundus image detail enhancement method of a mixed double-domain structure, and the method comprises the steps: an image data obtaining step, which is configured to obtain to-be-processed target image data; a shallow feature extraction step configured to input the target image data into a preset generative network and extract a shallow feature map; a double-flow parallel processing step, which is configured to simultaneously input the shallow feature map into a spatial domain enhancement branch and a frequency domain enhancement branch which are arranged in parallel; a feature fusion step configured to calculate a weight vector based on a channel attention mechanism and generate an enhanced feature map; an image reconstruction step configured to reconstruct and output a target image after detail enhancement and illumination correction based on the enhanced feature map; according to the method, the problem that in the prior art, when high-frequency details of the microvessels are recovered, low-frequency signal interference such as uneven illumination is likely to happen is solved, and the definition of medical images is remarkably improved.

Description

Technical Field

[0001] This invention relates to the fields of medical image processing and deep learning technology, specifically to a method for enhancing details in color fundus images with a hybrid dual-domain structure. Background Technology

[0002] With the continuous development of medical imaging technology, color fundus images have become an important basis for the clinical auxiliary diagnosis of ophthalmic diseases such as diabetic retinopathy and glaucoma. In order to obtain higher quality medical images, medical imaging workstations equipped with high-performance graphics processors are widely used for digital processing and detail enhancement of fundus images.

[0003] Currently, spatial domain enhancement techniques based on convolutional neural networks are commonly used for processing fundus images. These methods primarily extract spatial features from images using pre-defined convolutional kernels, aiming to restore the visual characteristics of retinal vessels, optic discs, microaneurysms, exudates, and other lesions, while also suppressing sensor noise during the imaging process. However, due to the nonlinear degradation of the ocular optical system and scattering by the imaging medium, raw fundus images often suffer from severe illumination unevenness, low contrast, and complex background interference. Traditional single-network architectures struggle to handle such pathological inverse problems while also considering the high frequencies of microvessels. The restoration of details and the suppression of low-frequency interference from uneven illumination are key challenges. Specifically, during denoising, the structures at the ends of microvessels and the edges of lesions are easily blurred. When enhancing local textures, noise in low-light areas is often amplified, and the halo effect of a bright center and dark edges cannot be effectively corrected. This coupling in feature processing limits the structural fidelity of images in clinical diagnosis. Therefore, how to achieve effective correction of global illumination distribution and high-fidelity enhancement of microscopic tissue structures, thereby improving the safety and reliability of fundus images in clinical applications, has become a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a method for enhancing details in color fundus images with a hybrid dual-domain structure. Specifically, the technical solution of this invention includes:

[0005] The image data acquisition step is configured to acquire the target image data to be processed.

[0006] The shallow feature extraction step is configured to input the target image data into a preset generative network to extract shallow feature maps;

[0007] The dual-stream parallel processing step is configured to simultaneously input the shallow feature map into a spatial domain enhancement branch and a frequency domain enhancement branch set in parallel; wherein, the spatial domain enhancement branch is configured to extract local texture features of the shallow feature map based on multi-scale processing, and the frequency domain enhancement branch is configured to extract global illumination and structural features of the shallow feature map based on frequency domain transformation.

[0008] The feature fusion step is configured to use a spatial pixel-level attention mechanism to calculate a pixel-level weight map, and then dynamically weight and fuse the local texture features output from the spatial domain enhancement branch and the global illumination and structural features output from the frequency domain enhancement branch based on the pixel-level weight map to generate an enhanced feature map.

[0009] The image reconstruction step is configured to reconstruct and output a target image with enhanced details and illumination correction based on the enhanced feature map.

[0010] Preferably, the process of extracting global illumination and structural features by the frequency domain enhancement branch includes:

[0011] Perform a two-dimensional fast Fourier transform on the input shallow feature map to convert the spatial domain features into frequency domain complex features;

[0012] Extract the real and imaginary parts of the complex features in the frequency domain;

[0013] The real and imaginary data are concatenated into independent feature channels to construct multi-channel frequency domain features.

[0014] Perform convolution operations on the spliced ​​multi-channel frequency domain features to perform cross-channel information interaction and feature modulation in the frequency domain, and generate modulated frequency domain features.

[0015] A two-dimensional inverse fast Fourier transform is performed on the modulated frequency domain features to restore them to the spatial domain representation, thereby obtaining global illumination and structural features.

[0016] Preferably, a convolution operation is performed on the spliced ​​multi-channel frequency domain features, including:

[0017] A 1×1 convolution kernel is used to perform convolution processing on the spliced ​​multi-channel frequency domain features;

[0018] By establishing a linear mapping relationship between frequency components through convolution processing, adaptive adjustment of the global illumination distribution of the target image data can be achieved.

[0019] Preferably, the process of extracting local texture features by the spatial domain enhancement branch includes:

[0020] Multi-scale dilated convolution is used to process the input shallow feature map;

[0021] By setting different preset dilation rates, the receptive field is expanded while keeping the feature map resolution unchanged, capturing local structural information at different scales as local texture features.

[0022] Preferably, the enhanced feature map is generated in the feature fusion step, including:

[0023] The output features of the spatial domain enhancement branch and the output features of the frequency domain enhancement branch are concatenated by channels;

[0024] The concatenated features are subjected to continuous convolution processing to extract spatial context weights while maintaining spatial resolution, and a pixel-level weight map with spatial resolution is calculated.

[0025] The output features of the two branches are weighted and summed using a pixel-level weight map, thereby adaptively adjusting the weight ratio of the frequency domain enhancement branch in flat regions and the weight ratio of the spatial domain enhancement branch in texture regions to generate an enhanced feature map.

[0026] Preferably, the generative network is trained using a joint loss function during the training phase. The joint loss function includes:

[0027] Charbonnier loss used to constrain pixel-level intensity;

[0028] Edge loss used to constrain high-frequency details;

[0029] And frequency domain loss used to constrain frequency domain consistency.

[0030] Preferably, the target image data is a color image of the fundus;

[0031] Local texture features include retinal vascular features and lesion edge features;

[0032] Global illumination and structural features include the illumination distribution characteristics of the fundus background.

[0033] Compared with the prior art, the present invention has the following beneficial effects:

[0034] 1. This invention achieves effective decoupling of features by constructing a dual-stream architecture that runs in both the spatial and frequency domains. The spatial domain branch focuses on protecting the topology of microvessels, while the frequency domain branch utilizes the global receptive field to efficiently correct illumination distribution. This design breaks the deadlock of conflict between denoising and sharpening in traditional methods from an architectural perspective, and solves the problem that existing technologies are easily interfered with by low-frequency signals due to uneven illumination when restoring high-frequency details of microvessels, thus significantly improving the clarity of medical images.

[0035] 2. This invention changes the limitation of traditional frequency domain enhancement, which is limited to mask filtering, by convolving the real and imaginary parts of the complex features in the frequency domain. This mechanism can capture the illumination pattern of the entire image with extremely low computational cost, effectively eliminating the halo effect that is bright in the center and dark around the edges that is common in fundus images. At the same time, this learnable frequency domain modulation method avoids the ringing effect produced by traditional hard threshold filtering, making the background of the enhanced image smoother and more natural, and greatly improving the contrast balance of the image.

[0036] 3. In the spatial domain enhancement branch, this invention expands the receptive field without reducing image resolution by introducing convolutional combinations with different dilation rates. This design can simultaneously and accurately capture different scale features from the main blood vessel trunk to the terminal microvessels, avoiding the permanent loss of information caused by conventional downsampling operations. This ensures that the enhanced image has extremely high fidelity at the terminal of the blood vessel and the edge of small lesions, providing clinicians with more accurate visual evidence for observing early lesions.

[0037] 4. This invention utilizes spatial pixel-level weight calculation, enabling the network to intelligently arbitrate image region attributes: emphasizing illumination smoothing in flat background areas and detail sharpening in textured areas; combined with a joint loss function that includes edge and frequency domain consistency constraints, it mathematically limits the network's tendency to arbitrarily create false textures, effectively preventing the generation of phantom blood vessels in medical images; this not only improves the visual quality of the image but also ensures the reliability and topological security of the enhancement results in medical clinical auxiliary diagnosis. Attached Figure Description

[0038] The present invention will be further explained below with reference to the accompanying drawings and embodiments:

[0039] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0040] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0041] Example 1:

[0042] Please see Figure 1 A method for enhancing details in color fundus images with a hybrid dual-domain structure, comprising:

[0043] The image data acquisition step is configured to acquire the target image data to be processed.

[0044] The shallow feature extraction step is configured to input the target image data into a preset generative network to extract shallow feature maps;

[0045] The dual-stream parallel processing step is configured to simultaneously input the shallow feature map into a spatial domain enhancement branch and a frequency domain enhancement branch set in parallel; wherein, the spatial domain enhancement branch is configured to extract local texture features of the shallow feature map based on multi-scale processing, and the frequency domain enhancement branch is configured to extract global illumination and structural features of the shallow feature map based on frequency domain transformation.

[0046] The feature fusion step is configured to calculate a pixel-level weight map with spatial resolution based on a spatial attention mechanism, and dynamically weight and fuse the local texture features output from the spatial domain enhancement branch and the global illumination and structural features output from the frequency domain enhancement branch based on the weight map to generate an enhanced feature map.

[0047] The image reconstruction step is configured to reconstruct and output a target image with enhanced details and illumination correction based on the enhanced feature map.

[0048] This embodiment proposes a hybrid dual-domain structure for enhancing the details of color fundus images, aiming to solve the pathological inverse problem of existing technologies in processing fundus images, which makes it difficult to simultaneously restore high-frequency details of microvessels and suppress low-frequency interference from uneven illumination. The method of this embodiment runs on a computing device with image processing capabilities, such as a medical imaging workstation equipped with a high-performance graphics processing unit (GPU).

[0049] Image data acquisition step: This step is configured to acquire the target image data to be processed;

[0050] In this embodiment, the target image data refers to the raw or color two-dimensional fundus image captured by a fundus camera or after preliminary format conversion; the image typically includes retinal vessels, optic discs, macula, and possible lesions such as microaneurysms and exudates; due to the nonlinear degradation of the ocular optical system and the scattering of the imaging medium, the image is often accompanied by uneven illumination, low contrast, and sensor thermal noise.

[0051] Shallow feature extraction step: This step is configured to extract features from shallow layers. Input is fed into a pre-defined generative network to extract shallow feature maps;

[0052] In this embodiment, the generator network adopts an improved encoder-decoder structure based on the U-Net architecture; the specific improvement lies in the introduction of a residual connection module before each downsampling layer of the encoder, which combines the input features with those of the two layers... The features after the convolutional layer are summed at the pixel level to alleviate the gradient vanishing problem in deep networks and ensure that shallow spatial information can be effectively transferred to deeper layers; the shallow feature map is denoted as... This refers to the feature tensor output after processing by the first convolutional layer of the generative network;

[0053] Shallow features contain the most original pixel-level spatial information and have not yet undergone geometric information loss due to multiple downsampling, making them suitable as reference data for subsequent dual-stream processing;

[0054] Dual-stream parallel processing step: This step is the core architecture design of this invention and is configured to... Simultaneously input to the spatial domain enhancement branch and frequency domain enhancement branch set in parallel;

[0055] Spatial domain enhancement branch: focuses on local high-frequency signals; this branch is configured to extract based on multi-scale processing. The local texture features are denoted as The focus is on restoring the sharpness of the microvascular ends and the edges of the lesions;

[0056] Frequency domain enhancement branch: Focuses on global low-frequency signals; this branch is configured to extract shallow feature maps based on frequency domain transform. The global illumination and structural features are denoted as The focus is on correcting the halo effect and overall contrast imbalance commonly seen in fundus images, which is bright in the center and dark around the edges.

[0057] Feature fusion step: This step is configured to calculate a weight vector based on a channel attention mechanism, and then... right and Dynamic weighted fusion is performed to generate an enhanced feature map;

[0058] Blending logic: The system automatically identifies image region attributes; for flat background regions, it assigns... Higher weighting for smoother lighting; for areas with dense blood vessels, assign... Greater weighting to emphasize details;

[0059] Image reconstruction step: This step is configured to be based on the enhanced feature map. The target image, after detail enhancement and illumination correction, is output through the decoder and reconstruction layer of the generative network.

[0060] By constructing a parallel dual-stream architecture in the spatial domain and frequency domain, this invention successfully achieves feature decoupling. Unlike traditional single CNN networks that attempt to process noise and details simultaneously using the same set of convolutional kernels, this scheme utilizes spatial branches to protect the easily lost microvascular topology, while simultaneously using the global receptive field characteristics of the frequency domain branches to efficiently correct the illumination distribution. This design mathematically breaks the zero-sum game of denoising leading to blurring, significantly improving the usability of images in clinical diagnosis.

[0061] Example 2:

[0062] The process of extracting global illumination and structural features by the frequency domain enhancement branch includes:

[0063] Perform a two-dimensional fast Fourier transform on the input shallow feature map to convert the spatial domain features into frequency domain complex features;

[0064] Extract the real and imaginary parts of the complex features in the frequency domain;

[0065] The real and imaginary data are concatenated into independent feature channels to construct multi-channel frequency domain features.

[0066] Perform convolution operations on the spliced ​​multi-channel frequency domain features to perform cross-channel information interaction and feature modulation in the frequency domain, and generate modulated frequency domain features.

[0067] A two-dimensional inverse fast Fourier transform is performed on the modulated frequency domain features to restore them to the spatial domain representation in order to obtain global illumination and structural features.

[0068] Perform convolution operations on the concatenated multi-channel frequency domain features, including:

[0069] A 1×1 convolution kernel is used to perform convolution processing on the spliced ​​multi-channel frequency domain features;

[0070] By establishing a linear mapping relationship between frequency components through convolution processing, adaptive adjustment of the global illumination distribution of the target image data can be achieved.

[0071] This embodiment further illustrates the specific implementation process of the frequency domain enhancement branch; this is the key innovation of the present invention that overcomes the technical bias that frequency domain processing is limited to mask filtering;

[0072] For input The calculation formula for performing a two-dimensional Fast Fourier Transform (2D-FFT) is as follows:

[0073] in, The height of the feature map, The width of the feature map. Number of feature channels; It is a complex feature in the frequency domain, containing information about the amplitude spectrum and the phase spectrum; This represents the two-dimensional Fast Fourier Transform operation. Represents the set of complex numbers;

[0074] extract Real data and imaginary part data ;

[0075] in, This indicates the operation of taking the real part. This indicates the operation of taking the imaginary part;

[0076] This embodiment does not employ traditional frequency domain masking operations, but creatively uses... and The channels are spliced ​​together to form independent feature channels, and the spliced ​​multi-channel frequency domain features are constructed. ;

[0077] in, Represents the set of real numbers;

[0078] right Perform a channel-dimensional fully hybrid convolution; specifically, use a kernel size of [missing value]. The convolutional layers are processed;

[0079] Specifically, the number of convolutional kernels in this convolutional layer is configured as follows: This ensures that the channel dimension of the output feature map is consistent with that of the input, thereby meeting the dimensional requirements for subsequent separation of the real and imaginary parts;

[0080] Therefore, the frequency domain modulation function is defined as follows:

[0081] in, It is the ReLU activation function; This refers to the number of output channels. of Convolution operation;

[0082] In the frequency domain, each pixel corresponds to a frequency component of the entire image; by... After convolution, a nonlinear activation function is introduced. While keeping the total number of frequency components unchanged, the network learns the complex nonlinear interaction between the real and imaginary parts, thereby achieving adaptive adjustment of the global illumination distribution of the target image data.

[0083] Modulated features Separate the data back into real and imaginary parts, perform a 2D inverse fast Fourier transform (2D-Inverse FFT), and restore the representation to the spatial domain to obtain the desired result. ;

[0084] Low-cost implementation of global receptive field: Traditional methods require stacking dozens of convolutional layers to obtain the full receptive field to handle uneven illumination, while this embodiment utilizes the global instantaneous nature of FFT and Convolution, with extremely low computational cost, enables the capture and correction of the lighting pattern of the entire image;

[0085] Avoiding ringing effects: Traditional hard threshold filtering can cause Gibbs ringing, while this embodiment uses learnable convolutional modulation to achieve soft adjustment of frequency components, effectively avoiding the generation of artifacts;

[0086] Example 3:

[0087] The process of extracting local texture features through spatial domain enhancement branches includes:

[0088] Multi-scale dilated convolution is used to process the input shallow feature map;

[0089] By setting different preset dilation rates, the receptive field is expanded while keeping the feature map resolution unchanged, and local structural information at different scales is captured as local texture features.

[0090] This embodiment details how the spatial domain enhancement branch extracts local texture features;

[0091] This branch uses multi-scale dilated convolution to process the input shallow feature map;

[0092] This embodiment sets up a set of parallel convolutional layers, whose dilation rates are respectively set to... ;

[0093] Standard convolution is used to capture features of fine blood vessels with a diameter of <3 pixels.

[0094] Dilated convolution expands the receptive field to capture features of intermediate blood vessels and main blood vessels while maintaining the feature map resolution (i.e., without pooling downsampling).

[0095] The final output local texture features It is a fusion of the above-mentioned features at different scales; the specific fusion logic is: to combine the expansion rate... The feature maps output by the three parallel branches are concatenated along the channel dimension to obtain a dimension of The temporary feature tensor is input into a convolution kernel of size . The number of output channels is The convolutional layers perform information integration and dimensionality reduction, ultimately generating a dimension of... Local texture features ;

[0096] The microvessels in the fundus have extremely high spatial density and variable diameters; conventional downsampling operations can directly lead to the permanent loss of microvascular information; this embodiment achieves precise capture of blood vessels of different thicknesses while maintaining full resolution through multi-scale dilated convolution, ensuring that the enhanced image has extremely high structural fidelity at the edge of the lesion and the end of the blood vessel;

[0097] Example 4:

[0098] The feature fusion step generates enhanced feature maps, including:

[0099] The output features of the spatial domain enhancement branch and the output features of the frequency domain enhancement branch are concatenated by channels;

[0100] The concatenated features are subjected to continuous convolution processing to extract spatial context weights while maintaining spatial resolution, and a pixel-level weight map is calculated.

[0101] The output features of the two branches are weighted and summed using a pixel-level weight map, thereby adaptively adjusting the weight ratio of the frequency domain enhancement branch in flat regions and the weight ratio of the spatial domain enhancement branch in texture regions to generate an enhanced feature map.

[0102] This embodiment details the adaptive weighting mechanism in the feature fusion step;

[0103] To address the difference in importance between spatial features and frequency domain features in different regions, this embodiment introduces a spatial pixel-level attention mechanism;

[0104] Feature splicing: and By concatenating the features along the channel dimension, a joint feature map is obtained. ;

[0105] Spatial pixel-level weight calculation: To achieve accurate differentiation between different regions of the image, such as blood vessel areas and background areas, this step abandons the global pooling operation that loses spatial information, and instead uses a convolutional structure that preserves spatial resolution to generate a spatial weight map. ;

[0106] in, This is the spliced ​​joint feature map. Represents the ReLU activation function. This represents the Sigmoid activation function. For the generated spatial weight map; First layer The weights of the convolution are used to change the number of channels from... Dimensional reduction , In this embodiment, the scaling factor is used. The range of values ​​is The preferred value is 4, which is used to balance the computational cost and feature compression ratio in order to fuse cross-domain features;

[0107] The ReLU activation function introduces non-linear characteristics.

[0108] Second layer The weights of the convolution are used to restore the number of channels to [the desired value]. ;

[0109] The sigmoid activation function maps the output value to... interval;

[0110] The generated spatial weight graph has the following dimensions: Unlike traditional channel attention, this weight map is applied at each pixel location. Each has an independent weight vector;

[0111] Dynamic fusion is calculated using the following formula:

[0112] Using this spatial weight map The network can achieve pixel-level intelligent arbitration: in flat background areas of fundus images, The network primarily handles low-frequency information, and automatically reduces its frequency. The value increases the frequency domain branch. The weighting is adjusted based on smooth lighting; in the vascular texture area, The main information is high-frequency, and the network automatically upgrades. The value increases the spatial domain branch. The weighting percentage is used to sharpen the edges;

[0113] This fusion mechanism acts as an intelligent arbitrator, overcoming the limitations of single-domain processing. It eliminates the problem of excessive texture sharpening across the entire image that may result from traditional frequency domain enhancement, and also solves the problem that simple spatial domain enhancement not only fails to brighten but also amplifies noise in low-light areas, thus achieving a pixel-level dynamic optimal solution for image enhancement.

[0114] Example 5:

[0115] The generative network is trained using a joint loss function, which includes:

[0116] Charbonnier loss used to constrain pixel-level intensity;

[0117] Edge loss used to constrain high-frequency details;

[0118] And frequency domain loss used to constrain frequency domain consistency.

[0119] This embodiment details the training constraint mechanism of the generative network; to ensure the reliability of the enhanced images in medical applications, this invention employs a joint loss function. ;

[0120] in, , and These are preset loss weight coefficients; to balance the dimensional differences of loss functions in different domains, this embodiment calculates... Logarithmically scaling was applied to the spectral magnitude to ensure that it remained consistent with the spatial domain loss in terms of numerical magnitude.

[0121] In order to ensure that loss functions of different magnitudes have similar gradient contributions in the early stages of training, this embodiment sets... , , ;

[0122] loss( ): Used to constrain pixel-level intensity; it is Robust approximation of norm;

[0123] in, This represents a preset color fundus standard reference image. This represents the target image reconstructed by the generator network. For example, a preset minimum constant. Compared to The function does not vanish when the error is close to zero, and it is more robust to outliers, thus better restoring the color fidelity of the image.

[0124] Edge loss Used to constrain high-frequency details;

[0125] The difference is calculated after extracting edges by applying the Laplacian operator to the image. The specific mathematical expression is as follows:

[0126] in, The Laplacian convolution operator is defined with its discrete convolution kernel as:

[0127] By calculating real images With generated image After convolution by this operator Norm distance forces the network to focus on the geometric contours of blood vessel walls and tiny lesions, preventing structural ambiguity;

[0128] Frequency domain loss ( ): Used to constrain frequency domain consistency;

[0129] This loss function directly constrains the spectral distribution of the generated image in the frequency domain, ensuring that the enhanced image does not produce spurious frequency components, thus preventing hallucinations and guaranteeing the topological security of medical images. To achieve this constraint, in the calculation... When performing this operation, the image needs to be first centered, which involves calculating the mean of all pixels in the image and subtracting the mean from each pixel value to achieve zero mean of the data. Then, a fast Fourier transform is performed to eliminate the interference of the DC component on the spectral energy distribution.

[0130] Introducing frequency domain loss is a key risk control method in this invention; in medical AI, generating non-existent blood vessels is a fatal flaw; by applying hard constraints in the frequency domain, this embodiment mathematically limits the network's tendency to arbitrarily create textures, significantly improving the safety and reliability of the algorithm in clinical applications.

[0131] Example 6:

[0132] The target image data is a color image of the fundus;

[0133] Local texture features include retinal vascular features and lesion edge features;

[0134] Global illumination and structural features include the illumination distribution characteristics of the fundus background.

[0135] This embodiment defines the target image data as a fundus color image;

[0136] In this application scenario:

[0137] Local texture features include, but are not limited to: the diameter and course of retinal arteries and veins, the edges of microaneurysms, the outlines of hemorrhages, and the texture of soft / hard exudates;

[0138] The specific global illumination and structural features include: the overall orange-red background distribution of the retina, the peripheral damascene caused by the curvature of the eyeball, and the distribution of reflected light spots caused by the imaging light source;

[0139] By explicitly limiting the application to fundus images, the technical parameters of this invention, such as the selection of dilation rate and the configuration of frequency domain convolution, can be specifically optimized for fundus anatomy, thereby providing high-quality images with greater diagnostic value than general image enhancement algorithms in the auxiliary diagnosis of ophthalmic diseases such as diabetic retinopathy (DR) and glaucoma.

[0140] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for enhancing details in color fundus images with a hybrid dual-domain structure, characterized in that, include: The image data acquisition step is configured to acquire the target image data to be processed. The shallow feature extraction step is configured to input the target image data into a preset generative network to extract shallow feature maps; The dual-stream parallel processing step is configured to simultaneously input the shallow feature map into a spatial domain enhancement branch and a frequency domain enhancement branch set in parallel; wherein, the spatial domain enhancement branch is configured to extract local texture features of the shallow feature map based on multi-scale processing, and the frequency domain enhancement branch is configured to extract global illumination and structural features of the shallow feature map based on frequency domain transformation. The feature fusion step is configured to calculate a pixel-level weight map based on a spatial pixel-level attention mechanism, and then dynamically weight and fuse the local texture features output from the spatial domain enhancement branch and the global illumination and structural features output from the frequency domain enhancement branch based on the pixel-level weight map to generate an enhanced feature map. The image reconstruction step is configured to reconstruct and output a target image with enhanced details and illumination correction based on the enhanced feature map.

2. The method for enhancing details in a color fundus image with a hybrid dual-domain structure according to claim 1, characterized in that, The process of extracting global illumination and structural features by the frequency domain enhancement branch includes: Perform a two-dimensional fast Fourier transform on the input shallow feature map to convert the spatial domain features into frequency domain complex features; Extract the real and imaginary parts of the complex features in the frequency domain; The real and imaginary data are concatenated into independent feature channels to construct multi-channel frequency domain features. Perform convolution operations on the spliced ​​multi-channel frequency domain features to perform cross-channel information interaction and feature modulation in the frequency domain, and generate modulated frequency domain features. A two-dimensional inverse fast Fourier transform is performed on the modulated frequency domain features to restore them to the spatial domain representation, thereby obtaining global illumination and structural features.

3. The method for enhancing details in a color fundus image with a hybrid dual-domain structure according to claim 2, characterized in that, Perform convolution operations on the concatenated multi-channel frequency domain features, including: A 1×1 convolution kernel is used to perform convolution processing on the spliced ​​multi-channel frequency domain features; By establishing a linear mapping relationship between frequency components through convolution processing, adaptive adjustment of the global illumination distribution of the target image data can be achieved.

4. The method for enhancing details in a color fundus image with a hybrid dual-domain structure according to claim 1, characterized in that, The process of extracting local texture features through spatial domain enhancement branches includes: Multi-scale dilated convolution is used to process the input shallow feature map; By setting different preset dilation rates, the receptive field is expanded while keeping the feature map resolution unchanged, capturing local structural information at different scales as local texture features.

5. The method for enhancing details in a color fundus image with a hybrid dual-domain structure according to claim 1, characterized in that, The feature fusion step generates enhanced feature maps, including: The output features of the spatial domain enhancement branch and the output features of the frequency domain enhancement branch are concatenated by channels; The concatenated features are subjected to continuous convolution processing to extract spatial context weights while maintaining spatial resolution, and a pixel-level weight map with spatial resolution is calculated. The output features of the two branches are weighted and summed using a pixel-level weight map, thereby adaptively adjusting the weight ratio of the frequency domain enhancement branch in flat regions and the weight ratio of the spatial domain enhancement branch in texture regions to generate an enhanced feature map.

6. The method for enhancing details in a color fundus image with a hybrid dual-domain structure according to claim 1, characterized in that, The generative network is trained using a joint loss function, which includes: Charbonnier loss used to constrain pixel-level intensity; Edge loss used to constrain high-frequency details; And frequency domain loss used to constrain frequency domain consistency.

7. The method for enhancing details in a color fundus image with a hybrid dual-domain structure according to claim 1, characterized in that, The target image data is a color image of the fundus; Local texture features include retinal vascular features and lesion edge features; Global illumination and structural features include the illumination distribution characteristics of the fundus background.

Images