A zero-reference retinal image enhancement method, system, device, and medium
By decomposing the light map and transmission map components of retinal images using an atmospheric scattering model, and combining deblurring and low-light enhancement techniques, the problem of image quality improvement for retinal images under complex distortions was solved, achieving effective enhancement of various distortions.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SHENZHEN UNIV
- Filing Date
- 2025-09-02
- Publication Date
- 2026-05-15
AI Technical Summary
Existing retinal image enhancement methods are ineffective in dealing with complex distortions, are difficult to apply to various coupled distortions in real-world scenarios, and rely on time-consuming and expensive paired image acquisition.
A zero-reference retinal image enhancement method is adopted, which decomposes atmospheric light map and transmission map components through atmospheric scattering model, and generates high-quality retinal images by combining deblurring and low-light enhancement curves, and uses reconstruction loss for constraint.
It effectively eliminates scattering interference, restores fine structures, improves brightness in low-light areas, enhances image detail recognition and contrast, and stably handles complex scenes with noise and uneven lighting.
Smart Images
Figure CN121120441B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of image processing technology, and specifically provides a zero-reference retinal image enhancement method, system, device, and medium. Background Technology
[0002] Retinoscopy is widely used for the early detection and monitoring of retinal diseases such as diabetic retinopathy, glaucoma, and macular degeneration. However, due to equipment limitations, patient factors, and environmental factors, acquired retinal images often contain various distortions. Low-quality retinal images can lead to misdiagnosis, missed diagnosis, and increased medical costs. Therefore, developing an efficient and practical retinal image enhancement method is of great significance for the diagnosis of ophthalmic diseases.
[0003] Traditional retinal image enhancement methods primarily utilize image transformation and manually annotated features to characterize and repair image distortions to improve image quality. However, these methods have limited ability to handle complex distortions, resulting in generally poor enhancement outcomes, which hinders clinical decision-making. In recent years, deep neural networks (DNNs) have attracted increasing attention in the field of image enhancement due to their powerful feature extraction and fusion capabilities. Existing methods can be broadly categorized into reference-based methods and zero-reference methods. The former learns the mapping relationship from low-quality images to high-quality images; the latter achieves enhancement by transferring knowledge through source-target image transformation from the low-quality domain to the high-quality domain, thus overcoming the dependence on strictly paired images, which leads to time-consuming and expensive clinical data acquisition. Despite their excellent performance in certain scenarios, existing DNN-based retinal image enhancement methods still have limitations. First, most methods rely on low-to-high-quality image pairs for model training. However, in clinical practice, acquiring paired retinal images is challenging due to the difficulty in controlling or reproducing complex image degradation. Second, most methods typically only enhance a single type of distortion, making them difficult to apply to real-world scenarios. Generally speaking, images captured from real-world scenes often exhibit multiple coupled distortions simultaneously, such as low light, blur, and noise.
[0004] Therefore, there is an urgent need for an image enhancement solution to address the above problems. Summary of the Invention
[0005] This invention provides a zero-reference retinal image enhancement method, comprising the following steps: obtaining the atmospheric light map component of the retinal image to be processed in an atmospheric scattering model; adding a random noise map to the grayscale image of the retinal image to be processed, and obtaining the transmission map component in the atmospheric scattering model; firstly deblurring the retinal image to be processed, and then generating a first image under the guidance of a transmission map generation module; improving the low-light performance of the V channel of the retinal image to be processed by learning an input low-light enhancement curve to generate a second image; fusing the first image and the second image to obtain an enhanced image; inputting the atmospheric light map component of the retinal image to be processed in the atmospheric scattering model, the transmission map component in the atmospheric scattering model, and the enhanced image into the atmospheric scattering model to obtain a quality-degraded image. The quality-degraded image and the retinal image to be processed are constrained using a reconstruction loss to achieve image enhancement.
[0006] In one technical solution of the aforementioned zero-reference retinal image enhancement method, the process of adding a random noise map to the grayscale image of the retinal image to be processed and obtaining the transmission map component in the atmospheric scattering model includes: adding the grayscale image of the retinal image to be processed with the random noise map to obtain an initial input map; using the initial input map and its 2x, 4x, and 8x downsampled images as inputs, respectively, and using the CONVB module to extract features from each downsampled image to generate a feature map; for the initial feature map in the feature map, considering the lack of high-level features, no fusion or CONVB processing is performed; all feature maps except the initial feature map are stitched together with the 2x downsampled adjacent high-level features along the channel direction, and the stitched features are input into the CONVB module to obtain guidance restoration features.
[0007] In one technical solution of the aforementioned zero-reference retinal image enhancement method, the process of adding a random noise map to the grayscale image of the retinal image to be processed and obtaining the transmission map component in the atmospheric scattering model further includes: After performing a 2x upsampling, the input is given to the CONVB module, and the resulting features are compared with... Element-wise multiplication is performed, where Sigmoid(·) represents the Sigmoid function. The product result is upsampled by a factor of 2 and passed through the CONVB module. The resulting feature is then multiplied by... Multiplication is performed step by step across feature decoding operations to obtain refined features. The refined features are then input into the CONVB module and a Sigmoid operation is performed again to generate a transmittance map.
[0008] In one technical solution of the aforementioned zero-reference retinal image enhancement method, the process of first deblurring the retinal image to be processed and then generating a first image under the guidance of the transillumination generation module includes: inputting the retinal image to be processed into the deblurring module for deblurring, the deblurring module comprising four encoding stages and four decoding stages; each encoding stage consists of a CONVB module, an attention-based feature integration module, and a downsampling operation, the attention-based feature integration module being used to guide the model to obtain detailed information from the guidance restoration features; each decoding stage contains four stages, each stage consisting of an upsampling operation, skip connection concatenation, and the CONVB module, in each decoder stage, the feature map of the previous stage is upsampled, concatenated with the features of the corresponding encoder stage, and then refined by the CONVB module, after four progressive decoder stages, generating a deblurred image E1.
[0009] In one technical solution of the aforementioned zero-reference retinal image enhancement method, the attention-based feature integration module includes: a spatial attention-based AFI module and a channel attention-based AFI module; the spatial attention-based AFI module integrates feature maps... A spatial attention map is generated by applying average and max pooling operations along the channel dimension. The resulting features are concatenated and then processed by a 7×7 convolution and a sigmoid function. The generated spatial attention map is then used for adaptive recalibration of the CONVB output results X1 / X2. The channel attention-based AFI module includes two parallel average and max pooling operations, followed by a multilayer perceptron to analyze the feature map. The multilayer perceptron consists of two 1×1 convolutions and a LeakyReLU activation function. The outputs of these two parallel branches are summed element-wise and then used by the Sigmoid function to generate a channel attention map. The multiplication operation is used to recalibrate the output results X3 / X4 of CONVB.
[0010] In one technical solution of the aforementioned zero-reference retinal image enhancement method, the V channel of the retinal image I(x) to be processed is used as the input of the low-light enhancement module. After passing through the first CONVB module, v1 is obtained. v1 then passes through the second CONVB module to obtain v2. Subsequently, v1 and v2 are concatenated along the channel direction and input into the third CONVB module for dimensionality reduction to obtain v3. At this point, v3 and the previous v2 are concatenated along the channel direction and then sent to the fourth CONVB module for dimensionality reduction to obtain v4. v4 is used as the pixel-by-pixel trainable parameter map of the higher-order curve. After n iterations, I is obtained. V (x) Enhanced version V e (x); V e The H and S channels of I(x) are concatenated with those of I(x), and then converted to RGB space to obtain the illumination-enhanced image E2.
[0011] In one technical solution of the aforementioned zero-reference retinal image enhancement method, the process of fusing the first image and the second image to obtain the enhanced image includes: stitching the deblurred image E1 and the illumination-enhanced image E2 along the channel direction, and processing the stitched features through the CONVB module; the obtained features are simultaneously processed by average pooling and max pooling operations in parallel; the acquired features are stitched along the channel direction and processed through a convolutional layer; the obtained features are normalized using the Sigmoid function to generate weights; and the weights are multiplied element-wise with E1 to generate the enhanced image.
[0012] Secondly, this invention discloses a zero-reference retinal image enhancement system, comprising: a first acquisition module for acquiring the atmospheric light map component of the retinal image to be processed in an atmospheric scattering model; a second acquisition module for adding a random noise map to the grayscale image of the retinal image to be processed and acquiring the transmission map component in the atmospheric scattering model; a first processing module for deblurring the retinal image to be processed and then generating a first image under the guidance of a transmission map generation module; a second processing module for improving the low-light performance of the V channel of the retinal image to be processed by learning an input low-light enhancement curve to generate a second image; a fusion module for fusing the first image and the second image to obtain an enhanced image; and a processing module for inputting the atmospheric light map component of the retinal image to be processed in the atmospheric scattering model, the transmission map component in the atmospheric scattering model, and the enhanced image into the atmospheric scattering model to obtain a degraded image. Reconstruction loss is used to constrain the degraded image and the retinal image to be processed to achieve image enhancement.
[0013] Thirdly, the present invention discloses an electronic device including a processor and a storage device, the storage device being adapted to store a plurality of program codes, characterized in that the program codes are adapted to be loaded and run by the processor to perform any of the above-described zero-reference retinal image enhancement methods.
[0014] Fourthly, the present invention discloses a computer-readable storage medium storing a plurality of program codes, characterized in that the program codes are adapted to be loaded and run by a processor to perform any of the above-described zero-reference retinal image enhancement methods.
[0015] The beneficial effects of the zero-reference retinal image enhancement method provided by this invention are as follows: On the one hand, based on the atmospheric scattering model, the atmospheric light map and transmission map components are accurately decomposed. Combined with deblurring processing and transmission map guidance to generate the first image, the scattering interference can be effectively eliminated and the fine structure can be restored. At the same time, the second image is generated through the learning mechanism of the V channel low-light enhancement curve to improve the brightness of the low-light area. After the two are fused, the image detail recognition and contrast can be comprehensively improved. On the other hand, the introduction of random noise during transmission map acquisition improves the robustness of the model to interference. The adaptive strategy of low-light enhancement avoids overexposure or under-enhancement caused by fixed parameters, enabling the algorithm to stably cope with complex scenes with noise and uneven illumination, thereby achieving the effect of simultaneously optimizing multiple distortions. Attached Figure Description
[0016] The disclosure of this invention will become more readily understood with reference to the accompanying drawings. It will be readily understood by those skilled in the art that these drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. Furthermore, similar numbers in the drawings are used to denote similar components, wherein:
[0017] Figure 1 This is a network framework diagram according to an embodiment of the present invention;
[0018] Figure 2 This is a structural diagram of a transmission image generation module according to an embodiment of the present invention;
[0019] Figure 3 This is a structural diagram of a deblurring module according to an embodiment of the present invention. Detailed Implementation
[0020] Some embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.
[0021] like Figure 1-3 As shown, this invention discloses a zero-reference retinal image enhancement method, comprising the following steps:
[0022] Step S1: Obtain the atmospheric light map components of the retinal image to be processed in the atmospheric scattering model;
[0023] Step S2: Add a random noise map to the grayscale image of the retinal image to be processed, and obtain the transmission map component in the atmospheric scattering model;
[0024] Step S3: First, the retinal image to be processed is deblurred, and then the first image is generated under the guidance of the transmission image generation module;
[0025] Step S4: Improve the low-light performance of the V channel of the retinal image to be processed by learning the low-light enhancement curve, and generate a second image;
[0026] Step S5: Fuse the first image and the second image to obtain the enhanced image;
[0027] Step S6: Input the atmospheric light map component of the retinal image to be processed in the atmospheric scattering model, the transmission map component of the atmospheric scattering model, and the enhanced image into the atmospheric scattering model to obtain the image with degraded quality. Constrain the image with degraded quality and the retinal image to be processed using reconstruction loss to achieve image enhancement.
[0028] In this embodiment, the following methods were specifically adopted to implement the above method: Figure 1 The network framework shown mainly consists of a Transmission Map Generation (TMG) module, a Deblurring (DB) module, a Low-Light Enhancement (LLE) module, and a fusion module. This network architecture primarily includes three parallel branches. Assuming the retinal image to be processed is used as the network's input image, in the upper branch, it is input into the Dark Channel Prior (DCP) module to obtain the atmospheric light map component as part of the atmospheric scattering model. In the middle branch, the grayscale image of the retinal image I(x) to be processed is processed using a method with parameters 1-I... gray The form of I(x) is added to the random noise map and used as the input to the TMG to obtain the transmission map component t(x) in the atmospheric scattering model. In the next branch, I(x) is first processed by the deblurring module DB, and then the deblurring result E1 is generated under the guidance of the transmission map generation module TMG. At the same time, in order to achieve low-light enhancement of the retinal image, the V channel of I(x) (denoted as I) is... V (x) is used as input to LLE. The LLE module improves the low-light performance of the image by learning the low-light enhancement curve of the input, generating the illumination enhancement result E2. Finally, in order to make full use of the brightness information in the low-light enhanced image E2, E1 and E2 are combined by the fusion module to output the enhanced image E.
[0029] The atmospheric light map component A(x) of the retinal image to be processed in the atmospheric scattering model, the transmission map component t(x) of the atmospheric scattering model, and the enhanced image E are input into the atmospheric scattering model to obtain the image after quality degradation.
[0030] In one embodiment, the transmittance map generation module is used to generate a transmittance map. Since grayscale images are adept at capturing key structural details, this module uses the grayscale value of I(x). Figure 1 -I gray (x) is used as input. Furthermore, to enhance the network's learning ability and robustness to noise, a Gaussian random noise map is added to the grayscale image to generate I. gn .
[0031] like Figure 2 As shown, the TMG module contains four parallel branches, each with a corresponding I... gn The input is then sampled at 2x, 4x, and 8x, and the CONVB module is used to extract features from the input to generate feature maps g. k , k∈{1,2,3,4}. To achieve feature fusion and complementarity of different sizes, the feature map g is... k 2x downsampled output of adjacent high-rise features The features are stitched together along the channel direction, and the stitched features are input into the CONVB block to obtain...
[0032] A CONVB block consists of two 3×3 convolutions, a LeadReLU, and a residual connection implemented by a 1×1 convolution.
[0033] In the first parallel branch, g1 is directly taken as... Without going through the CONVB block operation. Afterwards, After performing a 2x upsampling, the input to the CONVB block yields features that are... Perform element-wise multiplication, where Sigmoid(·) represents the Sigmoid function. Upsample the product by a factor of 2 and pass it through a CONVB block. The resulting feature is then multiplied by... Multiplication. By performing cross-feature decoding operations step by step, refined features are finally obtained. These features are then input into the CONVB module, and a Sigmoid operation is performed again to generate the transmittance map t(x).
[0034] In one embodiment, the DB module is used to reconstruct the enhanced image after blur restoration, and its structure is as follows: Figure 3 As shown, the DB module adopts a U-shaped architecture, comprising four encoding stages and four decoding stages. Each encoding stage consists of a CONVB module, an attention-based feature integration (AFI) module, and a downsampling operation. The TMG mentioned above generates multi-scale features. Containing a wealth of contextual information, the AFI module can be used to guide the model from... Get more details.
[0035] Considering that shallow layers have a large spatial dimension while deep layers have a large channel dimension, the first two encoding stages use spatial attention-based AFI (referred to as AFI / S), while the latter two encoding stages use channel attention-based AFI (referred to as AFI / C). AFI / S works by analyzing the feature maps... Spatial attention maps are generated by applying average and max pooling operations along the channel dimension. The resulting features are concatenated and processed using a 7×7 convolution and a sigmoid function. The generated spatial attention maps are then used for adaptive recalibration of the CONVB outputs X1 / X2. The AFI / C block focuses on inter-channel relationships, and two parallel average and max pooling operations are followed by a multilayer perceptron (MLP) for a more comprehensive analysis of the feature maps. This distributed MLP consists of two 1×1 convolutions and a LeakyReLU activation function. The outputs of these two parallel branches are element-wise summed and then used to generate a channel attention map via a sigmoid function. This attention map is then used to recalibrate the CONVB outputs X3 / X4 through multiplication operations.
[0036] Taking the k-th stage as an example, the output X of CONVB k Before downsampling, with G k Fusion is performed within the attention block. The decoder, in contrast to the encoder, also comprises four stages, each consisting of an upsampling operation, skip connection concatenation, and a CONVB block. In each decoder stage, the feature map from the previous stage is upsampled, concatenated with the features from the corresponding encoder stage, and then refined using the CONVB block. After four progressive decoder stages, the final deblurred image E1 is generated.
[0037] In one embodiment, a low-light enhancement module (LLE) addresses uneven illumination and insufficient brightness in retinal images. This module operates only on the V channel in the HSV color space, performing brightness correction to avoid artifacts such as color casts or oversaturation.
[0038] First, the V channel of the retinal image I(x) to be processed is used as the input to the low-light enhancement module. After passing through the first CONVB module, v1 is obtained. v1 then passes through the second CONVB module to obtain v2. Subsequently, v1 and v2 are concatenated along the channel direction and input into the third CONVB module for dimensionality reduction to obtain v3. Then, v3 and the previously obtained v2 are concatenated along the channel direction and fed into the fourth CONVB module for dimensionality reduction to obtain v4. v4 is used as the pixel-by-pixel trainable parameter map of the higher-order curve. After n iterations, I is obtained. V (x) Enhanced version V e (x). Finally, V e The H and S channels of I(x) are concatenated with those of I(x), and then converted to RGB space to obtain the illumination-enhanced image E2.
[0039] In one embodiment, after obtaining two intermediate results E1 and E2, a fusion module integrates them to generate the final enhanced image E. Specifically, first, the fusion module concatenates the deblurred image E1 and the illumination-enhanced image E2 along the channel direction, and processes the concatenated features through a CONVB block to halve the channel dimension. Subsequently, the resulting features are processed in parallel using both average pooling and max pooling operations. Next, the acquired features are concatenated along the channel direction and processed through a 7×7 convolutional layer. Afterward, the resulting features are normalized using the Sigmoid function to generate weights. Finally, these weights are multiplied element-wise with E1 to generate the enhanced image.
[0040] Multiple loss functions are used for joint optimization, mainly consisting of two parts: a loss function for low-light enhancement and a loss function for deblurring. The loss function for low-light enhancement includes illumination smoothing loss. and exposure control loss The loss used for deblurring includes smoothing loss. Spatial consistency loss Color constant loss and loss.
[0041] To verify the method, the following experiments were conducted: Retina images from the large-scale public dataset EyeQ were degraded to create the first dataset required for the application, named EyeDe1. Subsequently, 2000 additional images with a quality rating of "good" were randomly selected from EyeQ as high-quality images. Only two degradation models—illumination distortion and blur distortion—and their combinations were used to synthesize 6000 low-quality images. The degradation parameter values in the original degradation model were changed to make the synthesized low-quality images appear visually darker and more blurred, resulting in the second dataset, EyeDe2. For experimental convenience, all images were uniformly resized to 512×512 pixels during network training. The data from the EyeDe1 and EyeDe2 datasets were divided into training and testing sets in an 8:2 ratio for model training and testing.
[0042] The network model of this invention is implemented in the PyTorch deep learning framework and runs on the Ubuntu 18.04 operating system. The network model uses the Adam optimizer with a learning rate of 1×10⁻⁴ and an epoch count of 200.
[0043] Example 2
[0044] This invention also provides a zero-reference retinal image enhancement system, comprising: a first acquisition module for acquiring the atmospheric light map component of the retinal image to be processed in an atmospheric scattering model; a second acquisition module for adding a random noise map to the grayscale image of the retinal image to be processed and acquiring the transmission map component in the atmospheric scattering model; a first processing module for deblurring the retinal image to be processed and then generating a first image under the guidance of a transmission map generation module; a second processing module for improving the low-light performance of the V channel of the retinal image to be processed by learning an input low-light enhancement curve to generate a second image; a fusion module for fusing the first image and the second image to obtain an enhanced image; and a processing module for inputting the atmospheric light map component of the retinal image to be processed in the atmospheric scattering model, the transmission map component in the atmospheric scattering model, and the enhanced image into the atmospheric scattering model to obtain a degraded image. Reconstruction loss is used to constrain the degraded image and the retinal image to be processed to achieve image enhancement.
[0045] Example 3
[0046] The present invention also provides an electronic device. In one embodiment of the invention, the device includes a processor and a storage device. The storage device may be configured to store a program for performing the zero-reference retinal image enhancement method of the above-described method embodiments. The processor may be configured to execute the program in the storage device, which includes, but is not limited to, a program for performing the zero-reference retinal image enhancement method of the above-described method embodiments. For ease of explanation, only the parts related to the embodiments of the present invention are shown; for specific technical details not disclosed, please refer to the method section of the embodiments of the present invention. The control device may be a control device device comprising various electronic devices.
[0047] Example 4
[0048] The present invention also provides a computer-readable storage medium. In one embodiment of the computer-readable storage medium according to the present invention, the computer-readable storage medium can be configured to store a program for performing the zero-reference retinal image enhancement method of the above-described method embodiments. This program can be loaded and run by a processor to implement the above-described zero-reference retinal image enhancement method. For ease of explanation, only the parts related to the embodiments of the present invention are shown; for specific technical details not disclosed, please refer to the method section of the embodiments of the present invention. The computer-readable storage medium can be a storage device comprising various electronic devices. Optionally, in the embodiments of the present invention, the computer-readable storage medium is a non-transitory computer-readable storage medium.
[0049] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the original technical features, and the technical solutions resulting from these changes or substitutions will all fall within the scope of protection of the present invention.
Claims
1. A zero-reference retinal image enhancement method, characterized in that, Includes the following steps: Obtain the atmospheric light map components of the retinal image to be processed in the atmospheric scattering model; Adding a random noise map to the grayscale image of the retinal image to be processed and obtaining the transmission map component from the atmospheric scattering model specifically includes: adding the grayscale image of the retinal image to be processed with the random noise map to obtain an initial input map; using the initial input map and its 2x, 4x, and 8x downsampled images as inputs, extracting features from each downsampled image using the CONVB module to generate feature maps; for the initial feature map in the feature maps, considering the lack of high-level features, no fusion or CONVB processing is performed; all feature maps except the initial feature map are stitched together with the 2x downsampled adjacent high-level features along the channel direction, and the stitched features are input into the CONVB module to obtain guidance restoration features. , ;Will After performing a 2x upsampling, the input is given to the CONVB module, and the resulting features are compared with... Perform element-wise multiplication, where This indicates that the Sigmoid function upsamples the product by a factor of 2 and passes it through the CONVB module. The resulting feature is then multiplied by... Multiplication is performed step by step across feature decoding operations to obtain refined features. The refined features are then input into the CONVB module and a Sigmoid operation is performed again to generate a transmittance map. The retinal image to be processed is first deblurred, and then the first image is generated under the guidance of the transmission image generation module. The V channel of the retinal image to be processed is used to improve the low light of the image by learning the low light enhancement curve of the input, and a second image is generated. The first image and the second image are fused to obtain an enhanced image; the atmospheric light map component of the retinal image to be processed in the atmospheric scattering model, the transmission map component in the atmospheric scattering model, and the enhanced image are input into the atmospheric scattering model to obtain a degraded image. Image enhancement is achieved by constraining the degraded image and the retinal image to be processed using a reconstruction loss.
2. The method according to claim 1, characterized in that, The process of first deblurring the retinal image to be processed, and then generating the first image under the guidance of the transillumination generation module, includes: The retinal image to be processed is input into the deblurring module for deblurring. The deblurring module includes four encoding stages and four decoding stages. Each encoding stage consists of a CONVB module, an attention-based feature integration module, and a downsampling operation. The attention-based feature integration module is used to guide the model to obtain detailed information from the insulated features. Each decoding stage comprises four phases, each consisting of an upsampling operation, skip connection concatenation, and a CONVB module. In each decoder phase, the feature map from the previous phase is upsampled, concatenated with the features from the corresponding encoder phase, and then refined using the CONVB module. After four progressive decoder phases, a deblurred image is generated. E 1.
3. The method according to claim 2, characterized in that, The attention-based feature integration module includes: a first AFI module based on spatial attention and a second AFI module based on channel attention; The first AFI module based on spatial attention uses the feature map to be processed. Spatial attention maps are generated by applying average and max pooling operations along the channel dimension. The resulting features are concatenated and then processed by a 7×7 convolution and a sigmoid function. The generated spatial attention maps are then used for adaptive recalibration of the CONVB output. ; The second AFI module based on channel attention includes two parallel average and max pooling operations, followed by a multilayer perceptron to analyze the feature map. The multilayer perceptron consists of two 1×1 convolutions and a LeakyReLU activation function. The outputs of these two parallel branches are summed element-wise and then used to generate a channel attention map via a sigmoid function. The multiplication operation is then used to recalibrate the output of CONVB. .
4. The method according to claim 2, characterized in that, The process of improving the low-light performance of the retinal image by learning the low-light enhancement curve from the input image to generate a second image includes: The retinal image to be processed The V channel is used as the input to the low-light enhancement module, and after passing through the first CONVB module, it obtains... , Then, after passing through the second CONVB module, we obtain... Then and After splicing along the channel direction, the result is input into the third CONVB module for dimensionality reduction. At this time And the previous The data is then spliced along the channel direction and fed into the fourth CONVB module for dimensionality reduction. ; Will As a pixel-by-pixel trainable parameter map of higher-order curves After n iterations, we get Enhanced version ; Will and The H and S channels are concatenated and then converted to RGB space to obtain an enhanced illumination image. E 2.
5. The method according to claim 4, characterized in that, The process of fusing the first image and the second image to obtain the enhanced image includes: Stitching deblurred images along the channel direction E 1 and illumination enhancement image E 2. And process the spliced features through the CONVB module; The resulting features are processed in parallel by both average pooling and max pooling operations. The acquired features are concatenated along the channel direction and then processed through a convolutional layer. The obtained features are normalized using the Sigmoid function to generate weights; Connect the weights with E 1. Perform element-wise multiplication to generate an enhanced image.
6. A zero-reference retinal image enhancement system, characterized in that, include: The first acquisition module is used to acquire the atmospheric light map components of the retinal image to be processed in the atmospheric scattering model; The second acquisition module is used to add a random noise map to the grayscale image of the retinal image to be processed and to acquire the transmission map component in the atmospheric scattering model. Specifically, it includes: adding the grayscale image of the retinal image to be processed with the random noise map to obtain an initial input map; using the initial input map and its 2x, 4x, and 8x downsampled images as inputs, and using the CONVB module to extract features from each downsampled image to generate a feature map; for the initial feature map in the feature map, considering the lack of high-level features, no fusion or CONVB processing is performed; all feature maps except the initial feature map are stitched together with the 2x downsampled adjacent high-level features along the channel direction, and the stitched features are input into the CONVB module to obtain guidance restoration features. , ;Will After performing a 2x upsampling, the input is given to the CONVB module, and the resulting features are compared with... Perform element-wise multiplication, where This indicates that the Sigmoid function upsamples the product by a factor of 2 and passes it through the CONVB module. The resulting feature is then multiplied by... Multiplication is performed step by step across feature decoding operations to obtain refined features. The refined features are then input into the CONVB module and a Sigmoid operation is performed again to generate a transmittance map. The first processing module is used to first deblur the retinal image to be processed, and then generate the first image under the guidance of the transmission image generation module; The second processing module is used to improve the low-light performance of the V channel of the retinal image to be processed by learning the low-light enhancement curve input, and generate a second image. The fusion module is used to fuse the first image and the second image to obtain an enhanced image; The processing module is used to input the atmospheric light map component of the retinal image to be processed in the atmospheric scattering model, the transmission map component of the atmospheric scattering model, and the enhanced image into the atmospheric scattering model to obtain the image after quality degradation; and to use the reconstruction loss to constrain the image after quality degradation and the retinal image to be processed in order to achieve image enhancement.
7. An electronic device comprising a processor and a storage device, the storage device being adapted to store multiple lines of program code, characterized in that, The program code is adapted to be loaded and run by a processor to perform the zero-reference retinal image enhancement method of any one of claims 1 to 5.
8. A computer-readable storage medium storing a plurality of program codes, characterized in that, The program code is adapted to be loaded and run by a processor to perform the zero-reference retinal image enhancement method of any one of claims 1 to 5.