Color enhancement method and system for perception image of color barrier patient

Abstract

The invention provides a color enhancement method and system for a color barrier patient (CVD) perception image. The method comprises the following steps: step S1, based on visual characteristics of color vision impaired people, establishing a color vision impaired friendly image database through subjective experiments; s2, a CVD Perceptual-Guided Enhanced Model (CPEM) is constructed, and the CVD Perceptual-Guided Enhanced Model (CPEM) comprises a CVD-guided feature extraction module and a diffusion transformation module for people suffering from color vision disorder; s3, completing the training of a color vision impairment perception guide enhancement model by using the color vision impairment friendly image database; and S4, taking the original image and the pure noise image as model input, performing color reconstruction on the original image through the trained color vision disorder perception guide enhancement model, and generating an enhanced image conforming to visual perception of a CVD user. The model provided by the invention has a remarkable advantage in the aspect of improving the color perception of an individual with color vision disorder.

Description

Technical Field

[0001] The present invention belongs to the field of image quality enhancement, and in particular relates to a color enhancement method and system for images perceived by color-blind patients. Background Art

[0002] Color perception is a key function of the human visual system and is essential for information exchange and decision-making. However, approximately 350 million people worldwide suffer from color vision disorders, of which red-green color vision disorders are the most common, causing them to face color discrimination difficulties in transportation, education, etc. Although some physical intervention methods such as colored filters have been developed for color correction, the effects are limited and difficult to adapt to different scenarios. With the development of visual processing technology and artificial intelligence technology, AI-driven color reconstruction technology has become a research hotspot for improving the visual experience of CVD users. However, current research has the limitations of lack of data sets for real user feedback and mostly unsupervised learning algorithms, which affects its actual application effect. Summary of the invention

[0003] In order to fill the gaps and shortcomings of the prior art, further improve the performance of color reconstruction technology, and optimize the visual experience of users with color vision impairment, the present invention proposes a color enhancement method and system for color-blind patients to perceive images, so as to improve the contrast and naturalness of the image, while enhancing the color discrimination ability, thereby better meeting the needs of color-blind patients.

[0004] To achieve the above-mentioned purpose, the technical solution of the present invention is: a color enhancement method and system for color-blind patients to perceive images.

[0005] In a first aspect, the present invention provides a method for color enhancement of images perceived by patients with color impairment, which specifically comprises the following steps:

[0006] Step S1, establishing a color vision-impaired friendly image database through subjective experiments based on the visual characteristics of color vision-impaired people;

[0007] Step S2, constructing a CVD Perception-guided Enhancement Model (CPEM), wherein the CVD Perception-guided Enhancement Model includes a feature extraction module and a diffusion transformation module guided by CVD for people with color vision impairment;

[0008] Step S3, using a color vision impairment friendly image database to complete the training of a color vision impairment perception guidance enhancement model;

[0009] Step S4: using the original image and the pure noise image as model inputs, reconstructing the color of the original image through the trained color vision impairment perception guidance enhancement model, and generating an enhanced image that meets the visual perception of CVD users.

[0010] Preferably, the establishment of a color vision-impaired friendly image database through subjective experiments specifically comprises the following steps:

[0011] Step S11, collecting colorful images from online resource libraries and related literature; the images cover a variety of visual signals and color perception content, and contain color pairs that are easily confused by CVD individuals;

[0012] Step S12, randomly dividing the obtained images into N groups, applying N different image recoloring methods to generate corresponding recolored images, pairing the generated recolored images with the initial images, and performing subjective quality assessment based on the ITU-R BT.500-11 standard by a paired comparison method;

[0013] Step S13, designing and implementing a graphical user interface system for automatically displaying the test image pairs and collecting subjective feedback from participants, wherein the participants are CVD individuals;

[0014] Step S14, after obtaining the subjective evaluation results of all image pairs, apply the half-voting strategy to screen the image pairs; if more than half of the participants are dissatisfied with the expression of a recolored image, the image pair is discarded, thereby screening out all image pairs for which more than half of the participants are satisfied with the expression of the recolored images, forming the color vision-impaired friendly image database.

[0015] Preferably, the CVD-guided feature extraction module specifically performs the following steps:

[0016] Step S21-1, by simulating the color perception difference of CVD users, a CVD simulated image is generated based on the original image x; and the features of the original image and the CVD simulated image are respectively extracted through the FEM feature extraction submodule, which is denoted as F r and F c ,Right now:

[0017] F r =FEM(x)

[0018] F c =FEM(CVD(x))

[0019] Wherein, x is the original image, CVD is the color blindness simulation matrix, and FEM is the feature extraction submodule;

[0020] Step S21-2: Through element-level difference operation, the extracted features are fused to generate the final difference feature map F d :

[0021] F d =F r -F c

[0022] Step S21-3: The difference feature map F d Downsampling obtains different feature maps of multiple scales, including F d1 、F d2 、F d3 , the difference feature map F d 、F d1 、F d2 、F d3 Input to the diffusion transformation module for further processing.

[0023] Preferably, the FEM feature extraction submodule specifically performs the following steps: the FEM feature extraction submodule first batch normalizes the input feature map, reduces the dimension through 1×1 convolution and introduces Swish activation function to increase nonlinearity, then uses 3×3 convolution to capture the spatial hierarchy, and then selects important features through Simple Gate to input into the SCA module. The SCA module divides the features into two parts and applies global average pooling and global maximum pooling respectively, uses 1×1 convolution to establish channel correlation and then splices the features, and finally processes the features output by the SCA module through 3×3 convolution and Swish activation function to generate a robust feature output that integrates global and local information.

[0024] The FEM feature extraction submodule extracts and enhances the global and local features of the image by using the combined operation of global average pooling and global maximum pooling through the channel attention SCA module.

[0025] Preferably, the diffusion transformation module specifically performs the following steps:

[0026] Step S22-1, feature extraction layer: original image x and noise image y t After concatenation in the channel dimension, a convolution operation is performed, and the feature map obtained by the convolution operation is combined with the difference feature map F d The time step t reflecting the current image noise level is input into the first CFormer submodule for feature fusion, and the first output feature map is obtained and downsampled together with the difference feature map F d1 The time step t is input into the second CFormer submodule for feature fusion, and the second output feature map is obtained and downsampled together with the difference feature map F d2 The time step t is input into the third CFormer submodule for feature fusion, and the third output feature map is obtained and downsampled together with the difference feature map F d3 and time step t are input into the fourth CFormer submodule for feature fusion to obtain the fourth output feature map as the final output feature map of the feature extraction layer;

[0027] Step S22-2, feature reconstruction layer: upsample the fourth output feature map and input it together with the third output feature map into the fifth CFormer submodule for feature fusion, obtain the fifth output feature map and upsample it and input it together with the second output feature map into the sixth CFormer submodule for feature fusion, obtain the sixth output feature map and upsample it and input it together with the first output feature map Figure 1 The seventh CFormer submodule is input to perform feature fusion, and after obtaining the seventh output feature map, a convolution operation is performed to obtain the predicted noise map, and then y t The image is subtracted from the predicted noise image to obtain the noise image y to be re-input into the feature extraction layer in the next round t-1 .

[0028] Preferably, in step S3, the noise image y t is the recolored image y in the database in step S1 0 Add noise step by step and generate the noise image y according to the forward diffusion process t :

[0029]

[0030] Among them, ∈ represents Gaussian noise that follows the standard normal distribution, β t is the hyperparameter of the Gaussian distribution variance; α t is the degree of noise added, Indicates that from y 0 to y t The cumulative scaling factor of

[0031] In step S4, the noise image y t is a pure noise image.

[0032] Preferably, any CFormer submodule from the first CFormer submodule to the seventh CFormer submodule performs the following steps:

[0033] S23-1. Input feature y in After batch normalization (BN), the time step t is embedded into the normalized features through a multi-layer perceptron (MLP) to form a noise feature map yˉ in ; The input feature y in Represents the output of the previous module adjacent to the current CFormer module in the diffusion transformation module; including the output feature map of the original image x and the noise image yt after convolution in the channel dimension, the output feature map of the previous CFormer submodule after downsampling or upsampling, etc. (see Figure 3 );

[0034] S23-2. Cross-attention mechanism for noise feature map yˉin and feature difference map F di Perform bidirectional fusion, and use 1×1 convolution and 3×3 depth convolution to respectively obtain the in and F di Extract query Q, key K and value V, and cross-calculate to generate attention graph F in two directions A and Y A , respectively learning the relationship between the noise feature map and the feature difference map; the feature difference map F di Represents the difference feature map F in the feature extraction module d 、F d1 、F d2 or F d3 ;

[0035] S23-3, the attention map F A and Y A The fused feature map is batch normalized, and the time step t is embedded into the normalized fused feature map through a multi-layer perceptron and input into a gated feedforward network. The network combines 3×3, 5×5, and 7×7 deep convolutions to capture multi-scale information, and further optimizes the feature representation through a channel attention module (SCA) and a simple gate.

[0036] S23-4. The multi-scale features output by the gated feedforward network are fused through Hadamard product and 1×1 convolution, and then fused with the normalized fused feature map to generate the final output.

[0037] On the fused feature map, multi-scale convolution, SCA module and gating mechanism are applied to further refine the features and generate the final feature representation to guide image recoloring.

[0038] Preferably, the step S3 trains the model by minimizing the loss function to ensure that the generated enhanced image that meets the visual perception of the CVD user is consistent with the target visual characteristics:

[0039]

[0040] Among them, q(y0) represents the real data distribution, y0~q(y0) represents the recolored image y0 sampled from the real data distribution; c is the original image x and the feature difference map F di Conditional information related to CVD patient perception;∈ θ (y t ,c,t) represents the noise predicted according to the CVD related condition information c and time step t. The goal is to make the predicted noise consistent with the actual noise ∈ t Minimize the difference between them.

[0041] Preferably, the CVD-related condition information c is used to guide the reverse denoising process by calculating the conditional probability distribution p θ (y t-1 |y t , c) to estimate the noise removal at each step:

[0042]

[0043] Among them, p θ represents the Gaussian distribution predicted by the model, θ is the model parameter; μ θ (y t ,t,c) and∑ θ (y t ,t,c) represent the mean and variance of the model prediction distribution respectively.

[0044] In a second aspect, the present invention provides a color enhancement system for images perceived by patients with color impairment, comprising a processor, a memory, and a computer program stored on the memory. When the processor executes the computer program, it specifically performs the steps of the above-mentioned color enhancement method for images perceived by patients with color impairment.

[0045] Compared with the prior art, the present invention has the following beneficial effects:

[0046] The present invention establishes a color vision disorder friendly image database based on subjective experiments, adopts a deep learning method, proposes a CVD-guided feature extraction module and a diffusion transformation module, which are used to convert CVD-related signals into CVD-specific features; and constructs a novel diffusion model, CPEM, based on the CVD-guided feature extraction module and diffusion transformation module, which drives the latest generative diffusion technology to achieve color reconstruction of CVD-friendly images. The present invention can effectively improve the color perception ability of CVD users, and shows superior performance and objectivity in terms of color reconstruction accuracy and user experience improvement. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 A schematic diagram of a data set construction process according to an embodiment of the present invention;

[0048] Figure 2 It is a schematic diagram of the overall process of an embodiment of the present invention;

[0049] Figure 3 It is a schematic diagram of the overall structure of the model of an embodiment of the present invention. DETAILED DESCRIPTION

[0050] The technical solution of the present invention is described in detail below in conjunction with the accompanying drawings.

[0051] refer to Figure 1-3The present invention proposes a color enhancement method for images perceived by color-blind patients, which specifically includes the following steps:

[0052] Step S1, establishing a color vision-impaired friendly image database through subjective experiments based on the visual characteristics of color vision-impaired people;

[0053] Step S2, constructing a CVD Perception-guided Enhancement Model (CPEM), wherein the CVD Perception-guided Enhancement Model includes a feature extraction module and a diffusion transformation module guided by CVD for people with color vision impairment;

[0054] Step S3, using a color vision impairment friendly image database to complete the training of a color vision impairment perception guidance enhancement model;

[0055] Step S4: using the original image and the pure noise image as model inputs, reconstructing the color of the original image through the trained color vision impairment perception guidance enhancement model to generate an enhanced image that meets the visual perception of CVD users, wherein the pure noise image can be a Gaussian noise image.

[0056] In this embodiment, the establishment of a color vision-impaired friendly image database through subjective experiments specifically includes the following steps:

[0057] Step S11, collecting colorful images from online resource libraries and relevant literature; the images cover a variety of visual signals and color perception content, and contain color pairs that are easily confused by CVD individuals;

[0058] Step S12, randomly dividing the obtained images into N groups, applying N different image recoloring methods to generate corresponding recolored images, pairing the generated recolored images with the initial images, and performing subjective quality assessment based on the ITU-R BT.500-11 standard by a paired comparison method;

[0059] Step S13, designing and implementing a graphical user interface system for automatically displaying the test image pairs and collecting subjective feedback from participants, wherein the participants are CVD individuals;

[0060] Step S14, after obtaining the subjective evaluation results of all image pairs, apply the half-voting strategy to screen the image pairs; if more than half of the participants are dissatisfied with the expression of a recolored image, the image pair is discarded, thereby screening out all image pairs for which more than half of the participants are satisfied with the expression of the recolored images, forming the color vision-impaired friendly image database.

[0061] In this embodiment, the CVD-guided feature extraction module specifically performs the following steps:

[0062] Step S21-1, by simulating the color perception difference of CVD users, a CVD simulated image is generated based on the original image x; and the features of the original image and the CVD simulated image are respectively extracted through the FEM feature extraction submodule, which is denoted as F r and F c ,Right now:

[0063] F r =FEM(x)

[0064] F c =FEM(CVD(x))

[0065] Wherein, x is the original image, CVD is the color blindness simulation matrix, and FEM is the feature extraction submodule;

[0066] Step S21-2: Through element-level difference operation, the extracted features are fused to generate the final difference feature map F d :

[0067] F d =F r -F c

[0068] Step S21-3: The difference feature map F d Downsampling obtains different feature maps of multiple scales, including F d1 、F d2 、F d3 , the difference feature map F d 、F d1 、F d2 、F d3 Input to the diffusion transformation module for further processing.

[0069] In this embodiment, the FEM feature extraction submodule specifically performs the following steps: the FEM feature extraction submodule first batch normalizes the input feature map, reduces the dimension through 1×1 convolution and introduces Swish activation function to increase nonlinearity, then uses 3×3 convolution to capture the spatial hierarchy, and then selects important features through Simple Gate to input into the SCA module. The SCA module divides the features into two parts and applies global average pooling and global maximum pooling respectively, uses 1×1 convolution to establish channel correlation and then splices the features, and finally processes the features output by the SCA module through 3×3 convolution and Swish activation function to generate a robust feature output that integrates global and local information.

[0070] The FEM feature extraction submodule extracts and enhances the global and local features of the image by using the combined operation of global average pooling and global maximum pooling through the channel attention SCA module.

[0071] In this embodiment, the diffusion transformation module specifically performs the following steps:

[0072] Step S22-1, feature extraction layer: original image x and noise image y t After concatenation in the channel dimension, a convolution operation is performed, and the feature map obtained by the convolution operation is combined with the difference feature map F d The time step t reflecting the current image noise level is input into the first CFormer submodule for feature fusion, and the first output feature map is obtained and downsampled together with the difference feature map F d1 The time step t is input into the second CFormer submodule for feature fusion, and the second output feature map is obtained and downsampled together with the difference feature map F d2 The time step t is input into the third CFormer submodule for feature fusion, and the third output feature map is obtained and downsampled together with the difference feature map F d3 and time step t are input into the fourth CFormer submodule for feature fusion to obtain the fourth output feature map as the final output feature map of the feature extraction layer;

[0073] Step S22-2, feature reconstruction layer: upsample the fourth output feature map and input it together with the third output feature map into the fifth CFormer submodule for feature fusion, obtain the fifth output feature map and upsample it and input it together with the second output feature map into the sixth CFormer submodule for feature fusion, obtain the sixth output feature map and upsample it and input it together with the first output feature map Figure 1 The seventh CFormer submodule is input to perform feature fusion, and after obtaining the seventh output feature map, a convolution operation is performed to obtain the predicted noise map, and then y t The image is subtracted from the predicted noise image to obtain the noise image y to be re-input into the feature extraction layer in the next round (the total number of rounds is the preset value). t-1 .

[0074] In this embodiment, in step S3, the noise image y t is the recolored image y in the database in step S1 0 Add noise step by step and generate the noise image y according to the forward diffusion process t :

[0075]

[0076] Among them, ∈ represents Gaussian noise that follows the standard normal distribution, β t is the hyperparameter of the Gaussian distribution variance; α t is the degree of noise added, Indicates that from y 0 to y tThe cumulative scaling factor of , which is independent and additivity of Gaussian distribution, is obtained in the second line of the derivation process with a standard deviation of A mixture of Gaussian distributions;

[0077] In step S4, the noise image y t is a pure noise image.

[0078] In this embodiment, any CFormer submodule from the first CFormer submodule to the seventh CFormer submodule performs the following steps:

[0079] S23-1. Input feature y in After batch normalization (BN), the time step t is embedded into the normalized features through a multi-layer perceptron (MLP) to form a noise feature map yˉ in ; The input feature y in Represents the output of the previous module adjacent to the current CFromer module in the diffusion transformation module;

[0080] S23-2. Cross-attention mechanism for noise feature map yˉ in and feature difference map F di Perform bidirectional fusion, and use 1×1 convolution and 3×3 depth convolution to respectively obtain the in and F di Extract query Q, key K and value V, and cross-calculate to generate attention graph F in two directions A (Q y ,K F ,V F ) and Y A (Q F ,K y ,V y ), respectively learning the relationship between the noise feature map and the feature difference map; the feature difference map F di Represents the difference feature map F in the feature extraction module d 、F d1 、F d2 or F d3 ;

[0081] S23-3, the attention map F A and Y A The fused feature map is batch normalized, and the time step t is embedded into the normalized fused feature map through a multi-layer perceptron and input into a gated feedforward network. The network combines 3×3, 5×5, and 7×7 deep convolutions to capture multi-scale information, and further optimizes the feature representation through a channel attention module (SCA) and a simple gate.

[0082] S23-4. The multi-scale features output by the gated feedforward network are fused through Hadamard product and 1×1 convolution, and then fused with the normalized fused feature map to generate the final output.

[0083] On the fused feature map, multi-scale convolution, SCA module and gating mechanism are applied to further refine the features and generate the final feature representation to guide image recoloring.

[0084] In this embodiment, step S3 trains the model by minimizing the loss function to ensure that the generated enhanced image that meets the visual perception of the CVD user is consistent with the target visual characteristics:

[0085]

[0086] Among them, q(y0) represents the real data distribution, y0~q(y0) represents the recolored image y0 sampled from the real data distribution; c is the original image x and the feature difference map F di Conditional information related to CVD patient perception;∈ θ (y t ,c,t) represents the noise predicted according to the CVD related condition information c and time step t. The goal is to make the predicted noise consistent with the actual noise ∈ t Minimize the difference between, where the actual noise ∈ t refers to the recolored image y 0 Noise added gradually.

[0087] In this embodiment, the CVD-related condition information c is used to guide the reverse denoising process by calculating the conditional probability distribution p θ (y t-1 |y t , c) to estimate the noise removal at each step:

[0088]

[0089] Among them, p θ represents the Gaussian distribution predicted by the model, θ is the model parameter; μ θ (y t ,t,c) and∑ θ (y t ,t,c) represent the mean and variance of the model prediction distribution respectively.

[0090] The present invention also provides a color enhancement system for images perceived by patients with color blindness, comprising a processor, a memory, and a computer program stored on the memory. When the processor executes the computer program, it specifically performs the steps in the above-mentioned color enhancement method for images perceived by patients with color blindness.

[0091] The present invention conducts experimental simulation for improving the visual perception of users with red-green color vision impairment: 5,800 colorful images are collected from online resource libraries and related literature, and the original images are randomly divided into four groups. Four different image recoloring methods are used to generate corresponding recolored images, and a total of 5,800 pairs of images are obtained. The participants in the experiment are individuals with red-green color vision impairment, preferably aged 20 to 26 years old. To avoid fatigue and improve the accuracy of the evaluation, the experiment is divided into 10 stages, and the participants get sufficient rest time between each stage. When selecting images through the graphical user interface system, the participants mainly consider information richness and aesthetic appeal; finally, 3,386 pairs of images are selected using a half-voting strategy to form the color vision impairment-friendly image database. The experimental results show that the proposed model has significant advantages in improving the color perception of individuals with color vision impairment.

[0092] The above are preferred embodiments of the present invention. Any changes made according to the technical solution of the present invention, as long as the resulting functions do not exceed the scope of the technical solution of the present invention, belong to the protection scope of the present invention.

Claims

1. A color enhancement method for images perceived by patients with color impairment, characterized in that: The specific steps include: Step S1, establishing a color vision-impaired friendly image database through subjective experiments based on the visual characteristics of color vision-impaired people; Step S2, constructing a color vision impairment perception guidance enhancement model CPEM, wherein the color vision impairment perception guidance enhancement model includes a feature extraction module and a diffusion transformation module guided by CVD for color vision impaired people; Step S3, using a color vision impairment friendly image database to complete the training of a color vision impairment perception guidance enhancement model; Step S4: using the original image and the pure noise image as model inputs, reconstructing the color of the original image through the trained color vision impairment perception guidance enhancement model, and generating an enhanced image that meets the visual perception of CVD users.

2. The color enhancement method for images perceived by patients with color impairment according to claim 1, characterized in that: The establishment of a color vision-impaired friendly image database through subjective experiments specifically includes the following steps: Step S11, collecting colorful images from online resource libraries and relevant literature; the images cover a variety of visual signals and color perception content, and contain color pairs that are easily confused by CVD individuals; Step S12, randomly dividing the obtained images into N groups, applying N different image recoloring methods to generate corresponding recolored images, pairing the generated recolored images with the initial images, and performing subjective quality assessment based on the ITU-R BT.500-11 standard by a paired comparison method; Step S13, designing and implementing a graphical user interface system for automatically displaying the test image pairs and collecting subjective feedback from participants, wherein the participants are CVD individuals; Step S14, after obtaining the subjective evaluation results of all image pairs, apply the half-voting strategy to screen the image pairs; if more than half of the participants are dissatisfied with the expression of a recolored image, the image pair is discarded, thereby screening out all image pairs for which more than half of the participants are satisfied with the expression of the recolored images, forming the color vision-impaired friendly image database.

3. The color enhancement method for images perceived by color-blind patients according to claim 1, characterized in that: The CVD-guided feature extraction module specifically performs the following steps: Step S21-1, by simulating the color perception difference of CVD users, a CVD simulated image is generated based on the original image x; and the features of the original image and the CVD simulated image are respectively extracted through the FEM feature extraction submodule, which is denoted as F r and F c ,Right now: F r =FEM(x) F c =FEM(CVD(x)) Wherein, x is the original image, CVD is the color blindness simulation matrix, and FEM is the feature extraction submodule; Step S21-2: Through element-level difference operation, the extracted features are fused to generate the final difference feature map F d : F d =F r -F c Step S21-3: The difference feature map F d Downsampling obtains different feature maps of multiple scales, including F d1 、F d2 、F d3 , the difference feature map F d 、F d1 、F d2 、F d3 Input to the diffusion transformation module for further processing.

4. The method for color enhancement of images for patients with color impairment according to claim 3, characterized in that: The FEM feature extraction submodule specifically performs the following steps: the FEM feature extraction submodule first batch normalizes the input feature map, reduces the dimension through 1×1 convolution and introduces Swish activation function to increase nonlinearity, then uses 3×3 convolution to capture the spatial hierarchy, and then selects important features through SimpleGate to input the SCA module. The SCA module divides the features into two parts and applies global average pooling and global maximum pooling respectively, uses 1×1 convolution to establish channel correlation and then splices the features, and finally processes the features output by the SCA module through 3×3 convolution and Swish activation function to generate a robust feature output that integrates global and local information.

5. The method for color enhancement of images for patients with color impairment according to claim 3, characterized in that: The diffusion transformation module specifically performs the following steps: Step S22-1, feature extraction layer: original image x and noise image y t After concatenation in the channel dimension, a convolution operation is performed, and the feature map obtained by the convolution operation is combined with the difference feature map F d The time step t reflecting the current image noise level is input into the first CFormer submodule for feature fusion, and the first output feature map is obtained and downsampled together with the difference feature map F d1 The time step t is input into the second CFormer submodule for feature fusion, and the second output feature map is obtained and downsampled together with the difference feature map F d2 The time step t is input into the third CFormer submodule for feature fusion, and the third output feature map is obtained and downsampled together with the difference feature map F d3 and time step t are input into the fourth CFormer submodule for feature fusion to obtain the fourth output feature map as the final output feature map of the feature extraction layer; Step S22-2, feature reconstruction layer: after upsampling the fourth output feature map, input it together with the third output feature map into the fifth CFormer submodule for feature fusion, obtain the fifth output feature map, upsample it, and input it together with the second output feature map into the sixth CFormer submodule for feature fusion, obtain the sixth output feature map, upsample it, and input it together with the first output feature map into the seventh CFormer submodule for feature fusion, obtain the seventh output feature map, perform a convolution operation to obtain the predicted noise map, and then convert y t The image is subtracted from the predicted noise image to obtain the noise image y to be re-input into the feature extraction layer in the next round t-1 .

6. The method for color enhancement of images for patients with color impairment according to claim 5, characterized in that: In step S3, the noise image y t The noise is gradually added to the recolored image y0 in the database of step S1, and the noise image y is generated according to the forward diffusion process. t : α t ＝1-β t ， Among them, ∈ represents Gaussian noise that follows the standard normal distribution, β t is the hyperparameter of the Gaussian distribution variance; α t is the degree of noise added, From y0 to y t The cumulative scaling factor of In step S4, the noise image y t is a pure noise image.

7. The method for color enhancement of images for patients with color impairment according to claim 5, characterized in that: Any CFormer submodule from the first CFormer submodule to the seventh CFormer submodule performs the following steps: S23-1. Input feature y in After batch normalization BN, the time step t is embedded into the normalized features through the multi-layer perceptron MLP to form a noise feature map y - in ; The input feature y in Represents the output of the previous module adjacent to the current CFromer module in the diffusion transformation module; S23-2. Cross-attention mechanism for noise feature map y - in and feature difference map F di Perform bidirectional fusion, using 1×1 convolution and 3×3 depth convolution to extract the convolutional layers from y - in and F di Extract query Q, key K and value V, and generate attention maps FA and YA in two directions by cross calculation, respectively learning the relationship between the noise feature map and the feature difference map; the feature difference map F di Represents the difference feature map F in the feature extraction module d 、F d1 、F d2 or F d3 ; S23-3, the attention map F A and Y A The fused feature map is batch normalized, and the time step t is embedded into the normalized fused feature map through a multi-layer perceptron and input into a gated feedforward network. The network combines 3×3, 5×5, and 7×7 deep convolutions to capture multi-scale information, and further optimizes the feature representation through the channel attention module SCA and simple gating Simple Gate; S23-4. The multi-scale features output by the gated feedforward network are fused through Hadamard product and 1×1 convolution, and then fused with the normalized fused feature map to generate the final output for guiding image recoloring.

8. The method for color enhancement of images for patients with color impairment according to claim 1, characterized in that: The step S3 trains the model by minimizing the loss function to ensure that the generated enhanced image that meets the visual perception of the CVD user is consistent with the target visual characteristics: Among them, q(y0) represents the real data distribution, y0~q(y0) represents the recolored image y0 sampled from the real data distribution; c is the original image x and the feature difference map F di Conditional information related to CVD patient perception;∈ θ (y t , c, t) represents the noise predicted according to the CVD related condition information c and the time step t. The goal is to make the predicted noise consistent with the actual noise ∈ t Minimize the difference between them.

9. The method for color enhancement of images for patients with color impairment according to claim 8, characterized in that: The CVD-related condition information c is used to guide the reverse denoising process by calculating the conditional probability distribution p θ (y t-1 |y t , c) to estimate the noise removal at each step: Among them, p θ represents the Gaussian distribution predicted by the model, θ is the model parameter; μ θ (y t ,t,c) and∑ θ (y t ,t,c) represent the mean and variance of the model prediction distribution respectively.

10. A color enhancement system for color-blind patients to perceive images, characterized by: The method comprises a processor, a memory and a computer program stored in the memory. When the processor executes the computer program, the method specifically performs the steps in the color enhancement method for color-blind patients to perceive images as described in any one of claims 1 to 7.

Images