Low-light face super-resolution method and system based on retinex and diffusion model

Abstract

The application discloses a low-light face super-resolution method and system based on a Retinex and diffusion model, and comprises the following steps: constructing a low-light face super-resolution network model based on the Retinex and diffusion model, the model comprising a noise image estimation module and a denoising diffusion module; performing bicubic interpolation on an input face image to obtain an up-sampling image; decomposing the up-sampling image by the noise image estimation module to obtain a face image with noise; fusing the face image with noise with pure Gaussian noise, and obtaining a result after noise elimination by the denoising diffusion module; minimizing diffusion loss to complete training of the low-light face super-resolution model; and testing a collected face image test sample by using the trained low-light face super-resolution network model to obtain a super-resolution result and perform evaluation. The application can better restore a low-light low-resolution face image, and extensive experiments show that the method is effective and can be applied to a low-light face super-resolution task.

Description

Technical Field

[0001] The present invention relates to a low-light face super-resolution technology, and in particular to a low-light face super-resolution method and system based on Retinex and diffusion models. Background Art

[0002] Face super-resolution, a long-standing and highly sought-after topic in computer vision, has achieved remarkable success in recent years, leveraging advances in deep learning technology and the booming availability of large-scale datasets. It has also flourished in numerous practical applications. However, within the specific field of face super-resolution, the problem of face super-resolution in low-light environments, as an emerging topic, remains understudied. Currently, solutions for low-light face super-resolution fall into two main categories: algorithms based on task splitting and algorithms specifically designed for this task. The former decomposes low-light face super-resolution into two independent steps: low-light enhancement and face super-resolution, which are performed sequentially to restore the image. The latter focuses on building a new model directly targeting low-light face super-resolution.

[0003] Unfortunately, both existing methods face challenges when it comes to low-light face super-resolution. Task-splitting algorithms, by breaking the original problem into two subtasks, often lead to mutual constraints between the two, affecting overall performance. Algorithms designed specifically for low-light face super-resolution generally face the challenges of high model complexity and difficulty in training. Therefore, there is an urgent need to develop a new low-light face super-resolution method that is both tailored and easy to train to more effectively address this challenge. Summary of the Invention

[0004] Purpose of the invention: One purpose of the present invention is to provide a low-light face super-resolution method based on Retinex and diffusion model, which can better restore low-light and low-resolution face images.

[0005] Another object of the present invention is to provide a low-light face super-resolution system based on Retinex and diffusion model.

[0006] Technical solution: The low-light face super-resolution method based on Retinex and diffusion model described in the present invention includes the following steps:

[0007] A low-light face super-resolution network model based on Retinex and diffusion models was constructed and trained using a face image dataset. The trained low-light face super-resolution network model was used to test the collected face image test samples, obtaining super-resolution results that were evaluated using evaluation metrics.

[0008] The low-light face super-resolution network model includes: a noise image estimation module and a denoising diffusion module. The denoising diffusion module is a U-Net architecture, including an encoding layer, a bottleneck layer, and a decoding layer. The noise image estimation module performs a Retinex transform on the model input image to obtain a noisy face image R noise ; R obtained by the noise image estimation module noise After being spliced ​​with pure Gaussian noise, it is processed through three consecutive coding layers to obtain the coding layer output image I E ; Coding layer output image I E Further information integration is performed in the bottleneck layer, and the integrated information is sequentially passed through three decoding layers for information integration and image restoration; then the convolution layer outputs a noise-free high-resolution face output image I HH ; The entire network model is trained using diffusion loss.

[0009] Furthermore, before using the face image dataset to train the low-light face super-resolution network model, it is necessary to normalize the face image data in the face image dataset and obtain the upsampled image I by cubic interpolation. LH .

[0010] Furthermore, the input image is subjected to Retinex transformation through the noise image estimation module to obtain the noisy face image R noise , including: first, performing three-channel averaging on the input image to obtain the color mean image I mean ; Then the input image and the color mean image I mean Splicing is performed in the channel dimension, and the brightness inverse component is obtained by sequentially passing through the convolution layer, DW convolution layer and convolution layer. Finally, according to the Retinex theory, with I LH Multiply to get the noisy face image R noise ; expressed as:

[0011] I mean =F(I LH )

[0012]

[0013]

[0014] Among them, the function F represents taking the mean on three channels, the function Con represents splicing in the channel dimension, Conv represents ordinary convolution, and DWConv represents DW convolution.

[0015] Furthermore, the encoding layer includes a face feature extraction block and a downsampling module, which is expressed as:

[0016]

[0017] in, Represents the input of the encoding layer. For the first encoding layer, Represents a noisy face image R noise , for the second and third coding layers, represents the output of the previous encoding layer, represents the output of the i-th encoding layer, FFER represents the face feature extraction block, and downSample represents the downsampling block.

[0018] The facial feature extraction block consists of two normalization layers, a global facial structure information extraction unit, and a gated convolutional layer that replaces the fully connected layer. The global facial structure information extraction unit uses linear layers, convolutional layers, and hourglass networks to implement the attention algorithm to extract global facial structure information, which can be expressed as:

[0019]

[0020]

[0021]

[0022] Among them, Linear represents the linear layer, HourGlass represents the hourglass network, Sigmod represents the activation function, I att represents the calculated attention feature, I filter Represents the calculated attention weight, I G It represents the output result of the global face structure information unit;

[0023] GFSR is used to represent the global face structure information extraction unit, and the calculation process of the entire face feature extraction block is simplified to:

[0024]

[0025]

[0026] Among them, GateConv represents gated convolution, I noise represents a pure Gaussian noise, and LN represents a normalization layer.

[0027] Furthermore, the bottleneck layer includes a global face structure information extraction unit, which is expressed as:

[0028] I B =GFSR(I E )

[0029] Among them, I B represents the integration result of the bottleneck layer, I ERepresents the output of the last encoding layer.

[0030] Furthermore, the decoding layer corresponds to the encoding layer. Each decoding layer includes an upsampling block and a face feature extraction block. The decoding layer not only receives the output from the previous decoding layer, but also receives the output of the encoding layer corresponding to its scale. The calculation process is:

[0031]

[0032] Among them, upSample represents the upsampling block, Represents the input of the decoding layer. For the first decoding layer, Represents the integration result of the bottleneck layer I B , for the second and third decoding layers, Represents the output of the previous decoding layer, represents the output of the i-th decoding layer, represents the output of the i-th encoding layer.

[0033] Finally, a convolutional layer is used to transform the output of the last decoding layer into the final result, which is represented as follows:

[0034] I HH =Conv(I D )

[0035] Among them, I HH represents the high-resolution face output image without noise under normal lighting, I D The resulting graph representing the output of the last decoding layer.

[0036] Furthermore, when calculating the diffusion loss, the Gaussian noise added to the output is estimated using the inverse diffusion transform, and then the L1 Loss is calculated with the true value of the Gaussian noise to constrain the model training, which is expressed as:

[0037] I ε =INV(I HH )

[0038] Loss diff =||I ε -I noise ||1

[0039] Where INV represents the inverse diffusion transformation, I ε Represents the estimated value of Gaussian noise, Loss diff represents diffusion loss.

[0040] Furthermore, the evaluation indicators include peak signal-to-noise ratio PSNR and structural similarity SSIM;

[0041] The peak signal-to-noise ratio (PSNR) is calculated as follows:

[0042]

[0043]

[0044] Where MSE represents mean square error, n represents the total number of test samples, and x i and y i They respectively represent the face images I HH and real image I GT Pixels, MAX I Indicates the maximum value of the image point color;

[0045] The calculation method of structural similarity SSIM is:

[0046] c1=(k1L) 2

[0047] c2=(k2L) 2

[0048]

[0049] in, isI HH The average value of isI GT The average value of isI HH The variance of isI GT variance, L is the dynamic range of pixel values, k1 and k2 are parameters, and c1 and c2 represent intermediate variables.

[0050] The system corresponding to the method includes:

[0051] The model building and training unit is used to build a low-light face super-resolution network model based on Retinex and diffusion models, and train it using a face image dataset;

[0052] and a test evaluation unit for testing the collected face image test samples using the trained low-light face super-resolution network model, obtaining super-resolution results, and evaluating them using evaluation indicators;

[0053] The low-light face super-resolution network model includes: a noise image estimation module and a denoising diffusion module. The denoising diffusion module is a U-Net architecture, including an encoding layer, a bottleneck layer, and a decoding layer. The noise image estimation module performs a Retinex transform on the model input image to obtain a noisy face image R noise ; R obtained by the noise image estimation module noiseAfter being concatenated with pure Gaussian noise, the image is processed through three consecutive coding layers to obtain image I E ; Coding layer output image I E Further information integration is performed in the bottleneck layer, and the integrated information is sequentially passed through three decoding layers for information integration and image restoration; then the convolution layer outputs a high-resolution output image I without noise. HH ; The entire network model is trained using diffusion loss.

[0054] A device for storing and executing the method includes a memory and a processor, wherein:

[0055] a memory for storing computer programs capable of running on the processor;

[0056] The processor is configured to execute the steps of the low-light face super-resolution method based on Retinex and diffusion model when running the computer program.

[0057] Beneficial effects: Compared with the existing technology, the significant technical effects of the present invention are: by introducing Retinex theory and diffusion model, the performance of the face super-resolution algorithm in low-light environment is significantly improved. Specifically, the noise image estimation module is first used to perform Retinex transformation on the input image, and then the denoising diffusion module based on the U-Net architecture is used to denoise the above-mentioned noisy face image; the present invention cleverly integrates Retinex theory to decompose the face image, and combines the diffusion model to repair the face image. The effectiveness and practicality of this method have been verified through a series of extensive experiments; in particular, the test results on the real face super-resolution dataset show that the method proposed in the present invention exhibits excellent performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 It is a flow chart of the method of the present invention;

[0059] Figure 2 Flowchart for training low-light face super-resolution network model;

[0060] Figure 3 This is the low-light face super-resolution network framework based on Retinex and diffusion model;

[0061] Figure 4 It is a structural diagram of the noise image estimation module of the present invention;

[0062] Figure 5 This is a diagram of the face feature extraction block structure of the present invention;

[0063] Figure 6 This is a structural diagram of the global face structure information extraction unit of the present invention. DETAILED DESCRIPTION

[0064] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0065] The present invention provides a low-light face super-resolution method based on Retinex and diffusion model, such as Figure 1 As shown in the figure, a low-light face super-resolution network model based on Retinex and diffusion model is first constructed, and then trained using a face image dataset; finally, the trained low-light face super-resolution network model is used to test the collected face image test samples to obtain the super-resolution results, which are then evaluated using evaluation indicators. The low-light face super-resolution network model includes two main parts: a noise image estimation module and a denoising diffusion module. First, the noise image estimation module performs a Retinex transform on the input face image to obtain a noisy face image R noise ; Secondly, the denoising diffusion module of the U-Net architecture is used to eliminate the noise and obtain a high-resolution face output image I without noise HH The entire low-light face super-resolution network model is trained using diffusion loss.

[0066] like Figures 1 to 3 As shown, the specific steps include:

[0067] S1, use data normalization method to input face image I LL Perform range constraints and obtain the upsampled image I through bicubic interpolation LH ;

[0068] The details are as follows: First, the maximum and minimum normalization method is used to control the data range of the face image dataset to between 0 and 1. Then, bicubic interpolation is used to convert the input face image I LL Enlarge to the target size and get the upsampled image I LH .

[0069] S2, using the noise image estimation module and Retinex theory to upsample the image I LH Decompose and get the noisy face image R noise ;

[0070] The details are as follows: In the noise image estimation module, the upsampled image I LH You need to first take the average value to get the color mean map I mean , and then concatenate the two and estimate the brightness inverse component through three layers of convolution Finally, the noisy face image R is obtained by dot multiplication. noise , the process is as follows:

[0071] I mean =F(I LH )

[0072]

[0073]

[0074] Among them, the function F represents taking the mean on three channels, the function Con represents splicing in the channel dimension, Conv represents ordinary convolution, and DWConv represents DW convolution.

[0075] like Figure 3 and Figure 4 As shown, in this embodiment, the upsampled image I LH It will first be sent to the noise image estimation module, and then in the noise image estimation module, a color mean image I will be obtained by averaging the three channels. mean , then I LH and I mean Splicing is performed on the channel dimension, and the convolution layer, DW convolution layer and convolution layer are obtained in sequence. Finally, according to the Retinex theory, with I LH Multiply to get the noisy face image R noise .

[0076] S3, denoising diffusion module for the noisy face image R in the previous step noise Perform denoising and obtain the final super-resolution result, that is, a high-resolution output image without noise.

[0077] The details are as follows: The denoising diffusion module mainly consists of three parts: the encoding layer, the bottleneck layer and the decoding layer. noise It will first be concatenated with a pure Gaussian noise, and then processed by three consecutive coding layers to obtain the coding layer result I E Each encoding layer consists of a face feature extraction block and a downsampling module, which can be expressed as follows.

[0078]

[0079] in, Represents the input of the encoding layer. For the first encoding layer, Represents a noisy face image R noise , for the second and third coding layers, represents the output of the previous encoding layer, represents the output of the i-th encoding layer, FFER denotes the facial feature extraction block, and downSample denotes the downsampling block. The facial feature extraction block consists of two normalization layers, a global facial structure information extraction unit, and a gated convolutional layer that replaces the fully connected layer. The global facial structure information extraction unit primarily utilizes linear layers, convolutional layers, and an hourglass network to implement a simple attention algorithm to extract global facial structure information. The calculation process is shown below.

[0080]

[0081]

[0082]

[0083] Among them, Linear represents the linear layer, HourGlass represents the hourglass network, Sigmod represents the activation function, I att represents the calculated attention feature, I filter Represents the calculated attention weight, I G It represents the output result of the global face structure information unit.

[0084] GFSR is used to represent the global face structure information extraction unit, and the calculation process of the entire face feature extraction block is simplified to:

[0085]

[0086]

[0087] Among them, GateConv represents gated convolution, I noise represents a pure Gaussian noise, and LN represents a normalization layer.

[0088] After being processed by the encoding layer, the feature information obtained is further integrated in the bottleneck layer. The bottleneck layer is relatively simple in design, consisting of only a global facial structure information extraction unit, which can be expressed as follows.

[0089] I B =GFSR(I E )

[0090] Among them, I B represents the integration result of the bottleneck layer, I E Represents the output of the last encoding layer.

[0091] After information integration, the results of the bottleneck layer are fed into the decoding layer for information integration and image restoration. Similar to the encoding layer, there are also three decoding layers, each consisting of an upsampling block and a facial feature extraction block. Furthermore, to obtain richer information, the decoding layer receives not only the output from the previous decoding layer but also the output of the encoding layer at its corresponding scale. The calculation process is shown below.

[0092]

[0093] Among them, upSample represents the upsampling block, Represents the input of the decoding layer. For the first decoding layer, Represents the output of the bottleneck layer, that is, the integration result of the bottleneck layer I B , for the second and third decoding layers, Represents the output of the previous decoding layer, represents the output of the i-th decoding layer.

[0094] Finally, a convolutional layer is used to transform the output of the last decoding layer into the final result, which can be expressed as follows.

[0095] I HH =Conv(I D )

[0096] Among them, I HH Represents the high-resolution face output image without noise under normal lighting, that is, the output image, I D Represents the result graph of the last decoding layer.

[0097] like Figure 3 、 Figure 5 and Figure 6 As shown, in this embodiment, the noisy image R obtained in the previous step S2 is noise It is sent to the denoising diffusion module of the U-shaped architecture. In the denoising diffusion module, R noise It will first be followed by a pure Gaussian noise I noise The image is then stitched together and processed sequentially through the encoding layer, bottleneck layer, and decoding layer to fully extract information and use it to repair the image. Finally, the output image I is obtained through a convolutional layer. HH .

[0098] S4. Complete the training of the low-light face super-resolution network model through minimum diffusion loss;

[0099] Specifically, the low-light face super-resolution network model is trained using diffusion loss. When calculating the diffusion loss, the Gaussian noise added to the output result is estimated using the inverse diffusion transform, and then the L1 loss is calculated with the true value of the Gaussian noise to constrain the model training. It can be expressed as follows:

[0100] I ε =INV(I HH )

[0101] Loss diff =||I ε -I noise ||1

[0102] Where INV represents the inverse diffusion transformation, I ε Represents the estimated value of Gaussian noise, Loss diff represents diffusion loss.

[0103] S5. Use the trained low-light face super-resolution network model to predict the collected face image test samples, obtain the super-resolution results, and evaluate them using evaluation indicators;

[0104] Specifically, the test samples are predicted and the super-resolution results are compared with the real results. In order to evaluate the performance of the low-light face super-resolution results, the evaluation indicators selected are Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity (SSIM).

[0105] The peak signal-to-noise ratio (PSNR) is calculated as follows:

[0106]

[0107]

[0108] Where MSE represents mean square error, n represents the total number of test samples, and x i and y i They respectively represent the face images I HH and real image I GT Pixels, MAX I Indicates the maximum value of the image point color, usually set to 255.

[0109] The calculation method of structural similarity SSIM is:

[0110] c1=(k1L) 2

[0111] c2=(k2L) 2

[0112]

[0113] in, isI HH The average value of isI GT The average value of isI HH The variance of isI GT The variance of L is the dynamic range of pixel values, k1 and k2 are parameters with values ​​of k1=0.01 and k2=0.03, and c1 and c2 represent intermediate variables.

[0114] The test data was tested to obtain the super-resolution results, which were evaluated using the above indicators. The effectiveness of this method was demonstrated through extensive experiments and applied to the low-light face super-resolution task.

[0115] This method uses PSNR and SSIM to evaluate the super-resolution results. Two representative algorithms in the field of face super-resolution are selected: SR3 and IDM. The indicator results are shown in Table 1. The method of the present invention (Ours) achieves the best results in both PSNR and SSIM, indicating that the super-resolution results of the present invention are closer to the real results.

[0116] Table 1 Low-light face super-resolution index results

[0117]

[0118] Based on the same concept, the present invention proposes a low-light face super-resolution system based on Retinex and diffusion models, including a model construction and training unit that constructs a low-light face super-resolution model based on Retinex and diffusion models and trains it using a facial image dataset. A testing and evaluation module is used to test collected facial image test samples using the trained low-light face super-resolution network model, obtain super-resolution results, and evaluate them using evaluation metrics.

[0119] The model construction and training unit includes a data processing module, a noise image estimation module, a denoising diffusion module and a training module. The data processing module converts the input face image I LL Normalize and use bicubic interpolation to get the upsampled image I LH , the noise image estimation module uses the Retinex theory to transform I LH Decomposed into a noisy face image R noise ; The denoising diffusion module will first convert R noise Followed by a pure Gaussian noise I noise The image is then stitched together and processed sequentially through the encoding layer, bottleneck layer, and decoding layer to fully extract information and use it to repair the image. Finally, the output image I is obtained through a convolutional layer. HHThe training module completes the training of the low-light face super-resolution network by minimizing the diffusion loss, and a test evaluation unit calculates an index according to a super-resolution result of a test sample and evaluates the low-light face super-resolution result.

[0120] The low-light face super-resolution network model comprises two main parts, namely a noise image estimation module and a denoising diffusion module. noise In the first step, the noise image estimation module is used to perform Retinex transformation on the input image to obtain a face image R with noise. HH In the second step, the denoising diffusion module with the U-Net architecture is used to eliminate the noise to obtain a high-resolution face output image I without noise.

[0121] Based on the same inventive concept, the present application also provides a low-light face super-resolution device based on a Retinex and diffusion model, which comprises a memory and a processor.

[0122] The memory is used to store a computer program capable of running on the processor.

[0123] The processor is used to execute the steps of the low-light face super-resolution device method based on the Retinex and diffusion model when the computer program is running.

[0124] Based on the same inventive concept, the present application also provides a storage medium, wherein the storage medium stores a computer program, and the computer program is executed by at least one processor to implement the steps of the low-light face super-resolution method based on the Retinex and diffusion model.

Claims

1. A low-light face super-resolution method based on Retinex and diffusion model, characterized by: The following steps are involved: A low-light face super-resolution network model based on Retinex and diffusion model is constructed and trained using face image datasets. The trained low-light face super-resolution network model is used to test the collected face image test samples to obtain super-resolution results, which are evaluated using evaluation indicators. Before using the face image dataset to train the low-light face super-resolution network model, the face image data in the face image dataset needs to be normalized and the upsampled image I is obtained by cubic interpolation. LH ; The low-light face super-resolution network model includes: a noise image estimation module and a denoising diffusion module. The denoising diffusion module is a U-Net architecture, including an encoding layer, a bottleneck layer, and a decoding layer. The encoding layer includes a face feature extraction block and a downsampling module, which is expressed as: in, Represents the input of the encoding layer. For the first encoding layer, Represents a noisy face image R noise , for the second and third coding layers, represents the output of the previous encoding layer, represents the output of the i-th encoding layer, FFER represents the face feature extraction block, and downSample represents the downsampling block; The facial feature extraction block consists of two normalization layers, a global facial structure information extraction unit, and a gated convolutional layer that replaces the fully connected layer. The global facial structure information extraction unit uses linear layers, convolutional layers, and hourglass networks to implement the attention algorithm to extract global facial structure information, which can be expressed as: Among them, Linear represents the linear layer, HourGlass represents the hourglass network, Sigmod represents the activation function, I att represents the calculated attention feature, I filter Represents the calculated attention weight, I G It represents the output result of the global face structure information unit; GFSR is used to represent the global face structure information extraction unit, and the calculation process of the entire face feature extraction block is simplified to: Among them, GateConv represents gated convolution, I noise represents a pure Gaussian noise, LN represents the normalization layer; The bottleneck layer includes a global face structure information extraction unit, which is expressed as: I B =GFSR(I E ) Among them, I B represents the integration result of the bottleneck layer, I E Represents the output of the last encoding layer; The decoding layer corresponds to the encoding layer. Each decoding layer includes an upsampling block and a face feature extraction block. The decoding layer not only receives the output from the previous decoding layer, but also receives the output of the encoding layer corresponding to its scale. The calculation process is: Among them, upSample represents the upsampling block, Represents the input of the decoding layer. For the first decoding layer, Represents the integration result of the bottleneck layer I B , for the second and third decoding layers, Represents the output of the previous decoding layer, represents the output of the i-th decoding layer, represents the output of the i-th encoding layer; Finally, a convolutional layer is used to transform the output of the last decoding layer into the final result, which is represented as follows: I HH =Conv(I D ) Among them, I HH represents the high-resolution face output image without noise under normal lighting, I D The result graph representing the output of the last decoding layer; The model input image is subjected to Retinex transformation through the noise image estimation module to obtain the noisy face image R noise ; including: first, the input image is averaged over three channels to obtain a color mean image I mean ; Then the input image and the color mean image I mean Splicing is performed in the channel dimension, and the brightness inverse component is obtained by sequentially passing through the convolution layer, DW convolution layer and convolution layer. Finally, according to the Retinex theory, with I LH Multiply to get the noisy face image R noise ; expressed as: I mean =F(I LH ) Among them, the function F represents taking the mean on three channels, the function Con represents splicing on the channel dimension, Conv represents ordinary convolution, and DWConv represents DW convolution; R obtained by the noise image estimation module noise After being spliced ​​with pure Gaussian noise, it is processed through three consecutive coding layers to obtain the coding layer output image I E ; Coding layer output image I E Further information integration is performed in the bottleneck layer, and the integrated information is sequentially passed through three decoding layers for information integration and image restoration; then the convolution layer outputs a noise-free high-resolution face output image I HH ; The entire network model is trained using diffusion loss.

2. The low-light face super-resolution method based on Retinex and diffusion model according to claim 1, characterized in that: When calculating the diffusion loss, the Gaussian noise added to the output is estimated using the inverse diffusion transform, and then the L1 Loss is calculated with the true value of the Gaussian noise to constrain the model training, which is expressed as: I ε =INV(I HH ) Loss diff =||I ε -I noise ||1 Where INV represents the inverse diffusion transformation, I ε Represents the estimated value of Gaussian noise, Loss diff represents diffusion loss.

3. The low-light face super-resolution method based on Retinex and diffusion model according to claim 1, characterized in that: Evaluation indicators include peak signal-to-noise ratio (PSNR) and structural similarity (SSIM); The peak signal-to-noise ratio (PSNR) is calculated as follows: Where MSE represents mean square error, n represents the total number of test samples, and x i and y i They respectively represent the face images I HH and the real image I GT Pixels, MAX I Indicates the maximum value of the image point color; The calculation method of structural similarity SSIM is: c1=(k1L) 2 c2=(k2L) 2 in, isI HH The average value of isI GT The average value of isI HH The variance of isI GT The variance of L is the dynamic range of pixel values, k1 and k2 are parameters, and c1 and c2 represent intermediate variables.

4. A system for the low-light face super-resolution method based on Retinex and diffusion model according to any one of claims 1 to 3, characterized in that: include: The model building and training unit is used to build a low-light face super-resolution network model based on Retinex and diffusion models, and train it using a face image dataset; and a test evaluation unit for testing the collected face image test samples using the trained low-light face super-resolution network model, obtaining super-resolution results, and evaluating them using evaluation indicators; The low-light face super-resolution network model includes: a noise image estimation module and a denoising diffusion module. The denoising diffusion module is a U-Net architecture, including an encoding layer, a bottleneck layer, and a decoding layer. The model input image is subjected to Retinex transformation through the noise image estimation module to obtain the noisy face image R noise ; R obtained by the noise image estimation module noise After being concatenated with pure Gaussian noise, the image is processed through three consecutive coding layers to obtain image I E ; Coding layer output image I E Further information integration is performed in the bottleneck layer, and the integrated information is sequentially passed through three decoding layers for information integration and image restoration; then the convolution layer outputs a noise-free high-resolution face output image I HH ; The entire network model is trained using diffusion loss.

5. A low-light face super-resolution device based on Retinex and diffusion model, characterized in that: comprising a memory and a processor, wherein: a memory for storing computer programs capable of running on the processor; The processor is configured to execute the steps of the low-light face super-resolution method based on Retinex and diffusion model as described in any one of claims 1 to 3 when running the computer program.

Images