Image adaptive denoising method based on deep learning
The deep learning-based image adaptive denoising method addresses noise variation and feature preservation in diverse images by employing block splitting and multi-dimensional evaluation, enhancing precision and adaptability in endoscopy, satellite cloud, and medical imaging.
Patent Information
- Application Number
- CN202510288602.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-03-12
- Publication Date
- 2025-06-27
AI Technical Summary
Traditional image denoising methods treat all regions of an image uniformly, failing to account for noise variations, while adaptive methods often overlook color and distribution features, which are crucial in images like endoscopy, satellite cloud images, and medical imaging.
A deep learning-based image adaptive denoising method that employs block splitting algorithms, multi-dimensional noise evaluation functions, and intelligent optimization algorithms to capture complex noise distributions and structural relationships, preserving shape, size, color, and distribution features.
Enhances denoising precision and adaptability by effectively capturing noise and structural complexities, improving diagnostic accuracy and retaining critical image features in endoscopy, satellite cloud, and medical imaging.
Smart Images
Figure CN120219220A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of image processing, and particularly to an image adaptive denoising method based on deep learning. Background Art
[0002] Traditional image denoising methods do not pay attention to the noise differences existing in different regions of an image, and adopt a one-size-fits-all approach to denoise the entire image; on the other hand, adaptive denoising pays attention to the noise differences in different regions, but some images not only focus on features such as shape and size, but also need to consider features such as color and distribution, for example, endoscope images, satellite cloud images, medical images (such as CT, MRI, etc.) and remote sensing images, etc.
[0003] Among them, an endoscope is an important medical tool for directly observing the tissues and organs inside the human body, and is crucial for diagnosing and treating diseases. However, due to the influence of lighting conditions, lens quality, and the biological tissue itself on endoscope imaging, endoscope images are often accompanied by various noises and distortions, and these factors may interfere with doctors' diagnostic and treatment decisions.
[0004] For this reason, many scholars have focused on the field of endoscope image denoising. Denoising can significantly improve the visual quality and clarity of images. By reducing the noise in the image, it is easier for doctors to observe and analyze the minute details of tissues, improving the accuracy and reliability of diagnosis. This is crucial for the detection of early lesions and the confirmation of diseases, especially in tumor screening and endoscope surgery. Secondly, denoising technology helps to retain important information in the image. In endoscope images, doctors need to observe fine features such as cell structure and blood vessel distribution, and these features are often blurred or masked by noise. Through precise denoising processing, these key information can be effectively retained and enhanced, thereby improving the medical experts' understanding and analysis ability of the condition. Existing denoising methods may ignore these features. For this reason, proposing an efficient and highly usable adaptive image algorithm is one of the problems urgently needed to be solved in the existing image processing field. Summary of the Invention
[0005] For this reason, the present invention provides an image adaptive denoising method based on deep learning to solve the problems of the one-size-fits-all processing method existing in traditional denoising and the incomplete denoising evaluation existing in adaptive denoising as proposed in the background art.
[0006] To achieve the above object, the present invention provides the following technical solution: An image adaptive denoising method based on deep learning, comprising:
[0007] Design an image block splitting algorithm to block the input image to improve the denoising accuracy and applicability;
[0008] Introduce a deep learning model MDL Achieve adaptive denoising of split images, which can capture the relationship between complex noise distribution and image structure in the image;
[0009] Design a multidimensional noise evaluation function as a deep learning model M DL The learning function in is used to evaluate the image denoising effect from a multi-dimensional level, retaining the shape, size, color and distribution characteristics of the image to achieve adaptive denoising of different characteristic noises;
[0010] Introducing intelligent optimization algorithm as deep learning model M DL The loss function is used to accelerate the convergence speed and efficiency of the entire adaptive denoising.
[0011] Preferably, the images include endoscopic images, satellite cloud images, medical images (such as CT, MRI, etc.) and remote sensing images; wherein the endoscopic images are processed by the deep learning model M DL It can preserve the subtle features of cell structure and vascular distribution tissue.
[0012] Preferably, the image block splitting algorithm is specifically as follows:
[0013] Use edge detection operators to obtain edge intensity images of an image:
[0014]
[0015] Among them, G x (x,y) and G y (x, y) are the gradient values in the horizontal and vertical directions respectively;
[0016] From the seed point (x s ,y s ) and perform region growing based on the similarity of gray values between pixels; define a merging criterion T, let R i is the i-th region, initially R i =(x s ,y s ), for each seed point (x s ,y s ), perform the following steps until convergence:
[0017] R i =(x,y)|(x,y)∈R i or satisfies T(x,y,x s ,y s ) Guidelines
[0018] (x,y)|(x,y)∈R i or R i Satisfies T(x,y,x s ,ys )Criterion
[0019] where T(x, y, x s , y s ) is the merging criterion function used to determine whether to merge the pixel (x, y) into the region R i ;
[0020] According to the results of region growing, the distribution of noise in different regions is obtained, and the image is segmented into multiple blocks B i , each block B i is defined by the corresponding region R i ;
[0021] B i = {(x, y)|(x, y) ∈ R i}.
[0022] Preferably, the specific design of the deep learning architecture M DL to achieve adaptive denoising of the split image is as follows:
[0023] Design the deep learning model M DL , and perform adaptive denoising through the learning function f(I orig ; θ):
[0024] I denoised = f(I orig ; θ);
[0025] In the above formula, I orig represents the input image block, θ includes the reduction degree S of the noise level, the retention degree M of details, the improvement P of the overall image quality, and the visual evaluation V, and I denoised represents the denoised image.
[0026] Preferably, design a multi-dimensional noise evaluation function as the learning function in the deep learning model M DL , and the design is as follows:
[0027] The multi-dimensional evaluation function includes:
[0028] (1) The reduction degree S of the noise level, evaluating the overall denoising effect of the image:
[0029]
[0030] where represents the signal variance, represents the noise variance, and the calculation is as follows:
[0031]
[0032] In the formula, N represents the number of pixels in the image, I origrepresents the original image, represents the mean value of the original image;
[0033] (2) The degree of detail retention M, which evaluates the retention of the shape, size, color, and distribution characteristics of the image:
[0034]
[0035] In the formula, represents the mean value of the image before and after denoising, represents the variance of the image before and after denoising, represents the covariance of the image before and after denoising, c1, c2 are small constants to increase numerical stability, take c1 = (k1L) 2 , c2 = (k2L) 2 , L represents the dynamic range of pixel values, k1, k2 are weight factors for controlling brightness and contrast;
[0036] (3) The improvement P of the overall image quality, which evaluates the pixel difference of the fine features of the cell structure and blood vessel distribution tissue before and after image denoising:
[0037]
[0038] represents the maximum value of the image pixels, MSE represents the mean square error, which measures the difference between the original image and the processed image;
[0039] (4) Visual evaluation V, which evaluates the visual difference before and after image denoising:
[0040]
[0041] The higher the reduction degree S of the noise level, the degree of detail retention M, the improvement P of the overall image quality, and the visual evaluation V, the better. Then the multi-dimensional noise evaluation function f(I orig ; θ) is expressed as:
[0042] f(I orig ; θ) = aS + bM + cP + dV;
[0043] θ = {a, b, c, d} is a set of parameters.
[0044] Preferably, an intelligent optimization algorithm is introduced as the loss function of the deep learning model M DL as follows:
[0045] Taking the multi-dimensional noise evaluation function f(I orig ; θ) as the optimization function; first, perform Gaussian process prior:
[0046] f(I orig; θ) ~ gp(m(I orig ; θ), k(θ, θ'));
[0047] where gp(.) represents the Gaussian process prior, m(I orig ; θ) represents the mean function, and k(θ, θ') represents the covariance function;
[0048] Given some parameter - function value pairs Given the observed data D, the posterior distribution of the objective function at the new point is as follows:
[0049]
[0050] where
[0051]
[0052] In the above formula, Θ n = {θ1,..., θ n}}, is the variance of the observation noise, I is an n×n identity matrix, and k[(X n , Θ n ), (X n , Θ n )] are the parameters substituting those in k(θ, θ'); represents the nth image before denoising, θ n represents the nth parameter set, represents the nth image after denoising.
[0053] Select the parameter that maximizes the value of the optimization function f(I orig ; θ) as the optimal parameter θ = {a, b, c, d} for this round, and iterate the steps of the intelligent optimization algorithm as the loss function of the deep - learning model until the predetermined stopping condition is reached, and select the optimal parameter as the denoising parameter.
[0054] The present invention has the following advantages:
[0055] (1) The adaptive denoising designed in the present invention pays attention to the noise differences in different regions, introduces a deep - learning model to implement image denoising, and can effectively capture the complex relationship between the complex noise distribution and the image structure in the image. Compared with traditional linear filtering methods, the deep - learning model designed in the present invention can provide a higher denoising effect and is applicable to images that not only need to pay attention to features such as shape and size, but also need to consider features such as color and distribution, such as endoscopic images, satellite cloud images, medical images (such as CT, MRI, etc.) and remote - sensing images;
[0056] (2) Design an image block splitting algorithm to accurately divide the input image into blocks, improving the denoising accuracy and applicability;
[0057] (3) Design a multi-dimensional noise evaluation function as the learning function in the deep learning model, replacing the traditional learning objective function, evaluating the image denoising effect from multiple dimensions, further improving the accuracy, and simultaneously achieving adaptive denoising for different noises;
[0058] (4) Introduce an intelligent optimization algorithm as the loss function of the deep learning model to accelerate the convergence speed and efficiency of the entire adaptive denoising, which is applicable to denoising large-scale image datasets; Brief Description of the Drawings
[0059] Figure 1 It is the overall framework diagram of the present invention;
[0060] Figure 2 It is the schematic diagram of the image block splitting process of the present invention;
[0061] Figure 3 It is the schematic diagram of the multi-dimensional noise evaluation process of the present invention;
[0062] Figure 4 It is the schematic diagram of the intelligent optimization process of the present invention. Detailed Embodiments
[0063] The following specific embodiments illustrate the implementation manners of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of them. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.
[0064] Example 1: In endoscopic images, doctors need to observe fine features such as cell structure and blood vessel distribution, which are often blurred or masked by noise. Through precise denoising, these key information can be effectively retained and enhanced, thereby improving the medical experts' ability to understand and analyze the condition. In this example, a deep learning model is introduced to achieve endoscopic image denoising, which can effectively capture the complex relationship between the complex noise distribution in endoscopic images and the endoscopic image structure, and retain the fine features of tissues such as cell structure and blood vessel distribution. Compared with traditional linear filtering methods, the deep learning model can provide higher denoising effect; design an image block splitting algorithm to precisely block the input endoscopic image to improve the denoising accuracy and applicability; design a multi-dimensional noise evaluation function as the learning function in the deep learning model to replace the traditional learning objective function, evaluate the image denoising effect from multiple dimensions, consider features such as the shape, size, color and distribution of tissues, further improve the accuracy, and at the same time achieve adaptive denoising for different feature noises; introduce an intelligent optimization algorithm as the loss function of the deep learning model to accelerate the convergence speed and efficiency of the entire adaptive denoising, and be applicable to denoising large-scale endoscopic image datasets;
[0065] Such as Figure 1 is the overall framework schematic diagram of the present invention. In this example, the present invention is applied to the field of endoscopic image denoising, and a method for adaptive denoising of endoscopic images based on deep learning is designed, specifically including:
[0066] Such as Figure 2 As shown, design an image block splitting algorithm to precisely block the input endoscopic image to improve the denoising accuracy and applicability;
[0067] The specific image block splitting algorithm is as follows:
[0068] Use an edge detection operator to obtain the edge intensity image of the endoscopic image:
[0069]
[0070] Among them, G x (x,y) and G y (x,y) are the gradient values in the horizontal and vertical directions respectively;
[0071] Starting from the seed point (x s ,y s ), perform region growing according to the gray value similarity between pixels; define a merging criterion T, let R i be the i-th region, initially R i =(x s ,y s ), for each seed point (x s ,y s), perform the following steps until convergence:
[0072] R i = {(x, y)|(x, y) ∈ R i or satisfies the T(x, y, x s , y s ) criterion
[0073] {(x, y)|(x, y) ∈ R i or R i satisfies the T(x, y, x s , y s ) criterion
[0074] where T(x, y, x s , y s ) is the merging criterion function for determining whether to merge the pixel (x, y) into the region R i ;
[0075] According to the result of region growing, obtain the distribution of noises in different regions, and segment the endoscopic image into multiple blocks B i , each block B i is defined by the corresponding region R i ;
[0076] B i = {(x, y)|(x, y) ∈ R i .
[0077] Introduce the deep learning model M DL to achieve adaptive denoising of the segmented endoscopic image, which can effectively capture the complex relationship between the complex noise distribution in the endoscopic image and the endoscopic image structure;
[0078] The specific design of the deep learning architecture M DL for achieving adaptive denoising of the segmented image is as follows:
[0079] Design the deep learning model M DL , and perform adaptive denoising by learning the function f(I orig ; θ) to retain the subtle features of tissues such as cell structure and blood vessel distribution:
[0080] I denoised = f(I orig ; θ);
[0081] In the above formula, I orig represents the input endoscopic image block, θ includes the reduction degree S of the noise level, the retention degree M of details, the improvement P of the overall endoscopic image quality, and the visual evaluation V, and I denoised represents the denoised endoscopic image.
[0082] As Figure 3 shown, a multi-dimensional noise evaluation function is designed as the learning function in the deep learning model M DL to evaluate the denoising effect of endoscopic images from a multi-dimensional perspective and simultaneously achieve adaptive denoising of different noises;
[0083] The multi-dimensional noise evaluation function is designed as the learning function in the deep learning model M DL and is designed as follows:
[0084] The multi-dimensional evaluation function includes:
[0085] (1) The degree of reduction S of the noise level, which evaluates the overall denoising effect of the endoscopic image:
[0086]
[0087] Among them, represents the signal variance, represents the noise variance, and the calculation is as follows:
[0088]
[0089] In the formula, N represents the number of pixels of the endoscopic image, and I orig represents the original endoscopic image, represents the mean value of the original endoscopic image;
[0090] (2) The degree of preservation M of details, which evaluates the preservation of the shape, size, color, and distribution characteristics of tissues in the endoscopic image:
[0091]
[0092] In the formula, represents the mean value of the endoscopic image before and after denoising, represents the variance of the endoscopic image before and after denoising, represents the covariance of the endoscopic image before and after denoising, and c1 and c2 are small constants for increasing numerical stability, taking c1 = (k1L) 2 , c2 = (k2L) 2 , L represents the dynamic range of pixel values, and k1 and k2 are weight factors for controlling brightness and contrast;
[0093] (3) The improvement P in the overall quality of the endoscopic image, which evaluates the pixel differences in the subtle features of tissues such as cell structure and blood vessel distribution before and after denoising of the endoscopic image:
[0094]
[0095] Let \(I_{max}\) represent the maximum value of the endoscopic image pixels, and MSE represent the mean square error, which measures the difference between the original endoscopic image and the processed endoscopic image;
[0096] (4) Visual evaluation \(V\), which evaluates the visual difference before and after endoscopic image denoising:
[0097]
[0098] The higher the reduction degree \(S\) of the noise level, the retention degree \(M\) of details, the improvement \(P\) of the overall endoscopic image quality, and the visual evaluation \(V\), the better. Then the multi-dimensional noise evaluation function \(f(I\) orig ;\(\theta)\) is expressed as:
[0099] \(f(I\) orig ;\(\theta)=aS + bM + cP + dV\);
[0100] \(\theta=\{a,b,c,d\}\) is a set of parameters.
[0101] As Figure 4 shown, an intelligent optimization algorithm is introduced as the loss function of the deep learning model \(M\) DL to accelerate the convergence speed and efficiency of the entire adaptive denoising.
[0102] Introduce an intelligent optimization algorithm as the loss function of the deep learning model \(M\) DL , specifically as follows:
[0103] Take the multi-dimensional noise evaluation function \(f(I\) orig ;\(\theta)\) as the optimization function; first perform Gaussian process prior:
[0104] \(f(I\) orig ;\(\theta)\sim gp(m(I\) orig ;\(\theta),\theta(\theta,\theta'))\);
[0105] In the formula, \(gp(.)\) represents the Gaussian process prior, \(m(I\) orig ;\(\theta)\) represents the mean function, and \(k(\theta,\theta')\) represents the covariance function;
[0106] Given some parameter - function value pairs Given the observed data \(D\), the posterior distribution of the objective function at the new point is:
[0107]
[0108] In the formula,
[0109]
[0110] In the above formula, \(\Theta\) n={θ1,...,θ n}, is the variance of the observation noise, I is an n×n identity matrix, k[(X n ,Θ n ),(X n ,Θ n )] is a parameter that replaces k(θ,θ'); represents the nth endoscopic image before denoising, θ n Represents the nth parameter set, represents the nth denoised endoscopic image.
[0111] Choose the optimization function f(I orig ; θ) The parameter with the largest value is taken as the optimal parameter θ = {a, b, c, d} of this round, and the iterative intelligent optimization algorithm is used as the step of the loss function of the deep learning model until the predetermined stopping condition is reached, and the optimal parameter is selected as the denoising parameter.
[0112] Embodiment 2: This embodiment also discloses the application of the present invention in satellite cloud images, medical images (such as CT, MRI, etc.) and remote sensing images;
[0113] (1) Satellite cloud image application:
[0114] Shape and size characteristics: Different cloud systems in satellite cloud images have different shapes, such as dense blocks of cumulonimbus clouds, thin filaments of cirrus clouds, and large sheets of stratus clouds. These shape characteristics are crucial for weather forecasting and climate research. Cloud systems also have different sizes, from small-scale convective cloud clusters to large-scale cyclone and anticyclone systems. Their size and coverage are important bases for analyzing and judging meteorological changes.
[0115] Color and distribution characteristics: The color of satellite cloud images usually reflects information such as the temperature and height of the cloud tops. For example, blue indicates cold cloud areas, and white indicates areas with thicker clouds or lower temperatures. The distribution of clouds is closely related to factors such as atmospheric circulation and topography. By observing the distribution of clouds, we can understand the movement and development of weather systems.
[0116] (2) Application of medical imaging (such as CT, MRI, etc.):
[0117] Shape and size characteristics: In CT images, different organs and diseased tissues have specific shapes and sizes. For example, the shape of a normal liver is relatively regular and wedge-shaped; while tumor tissue may present an irregular round, lobed or infiltrative growth shape, and its size can be used to assess the severity of the disease and formulate a treatment plan. MRI images can also clearly show the shape and structure of human tissues and organs, which is of great significance for detecting brain lesions, spinal cord diseases, etc.
[0118] Color and distribution characteristics: Although most medical images are grayscale images, different grayscale values can reflect the density differences of tissues, indirectly reflecting certain "color" characteristics. For example, in CT images, bones appear as bright white, soft tissues are gray, and air shows as black. The grayscale distribution of diseased tissues is different from that of normal tissues, and analyzing the grayscale distribution can assist in disease diagnosis.
[0119] (3) Applications in remote sensing images:
[0120] Shape and size characteristics: The ground object targets in remote sensing images have various shapes and sizes, such as the rectangular, circular or irregular shapes of urban buildings, the linear form of rivers, and the planar form of lakes. These shape and size information are of great significance for the extraction of geographical information, land use classification, and environmental monitoring. For example, by analyzing the shape and area changes of forest cover, the situation of deforestation and ecological evolution can be studied.
[0121] Color and distribution characteristics: The color information of remote sensing images is rich, and different ground objects present different color characteristics on images of different bands, which helps to identify and distinguish various ground object types. For example, vegetation has a high reflectance in the near-infrared band and appears red, while it reflects green and blue light in the visible light band. By methods such as the normalized difference vegetation index (NDVI), the contrast between vegetation and other ground objects can be enhanced, thus more accurately monitoring and analyzing the distribution of vegetation.
[0122] According to the image adaptive denoising method based on deep learning designed by the present invention above, it can denoise satellite cloud images, medical images (such as CT, MRI, etc.) and remote sensing images while preserving their shape and size, color and distribution characteristics.
[0123] Although the present invention has been described in detail with general descriptions and specific embodiments above, based on the present invention, some modifications or improvements can be made, which are obvious to those skilled in the art. Therefore, these modifications or improvements made without departing from the spirit of the present invention all fall within the scope of protection required by the present invention.
Claims
1. An image adaptive denoising method based on deep learning, characterized in that: include: Design an image block splitting algorithm to divide the input image into blocks to improve the denoising accuracy and applicability; Introducing the deep learning model M DL Achieve adaptive denoising of split images, which can capture the relationship between complex noise distribution and image structure in the image; Design a multidimensional noise evaluation function as a deep learning model M DL The learning function in is used to evaluate the image denoising effect from a multi-dimensional level, retaining the shape, size, color and distribution characteristics of the image to achieve adaptive denoising of different characteristic noises; Introducing intelligent optimization algorithm as deep learning model M DL The loss function is used to accelerate the convergence speed and efficiency of the entire adaptive denoising.
2. The image adaptive denoising method based on deep learning according to claim 1, characterized in that: The images include endoscopic images, satellite cloud images, medical images and remote sensing images; the endoscopic images are processed by the deep learning model M DL It can preserve the subtle features of cell structure and vascular distribution tissue.
3. The image adaptive denoising method based on deep learning according to claim 1, characterized in that: The image block splitting algorithm is specifically as follows: Use edge detection operators to obtain edge intensity images of an image: Among them, G x (x,y) and G y (x, y) are the gradient values in the horizontal and vertical directions respectively; From the seed point (x s ,y s ) and perform region growing based on the similarity of gray values between pixels; define a merging criterion T, let R i is the i-th region, initially R i =(x s ,y s ), for each seed point (x s ,y s ), perform the following steps until convergence: R i =(x,y)|(x,y)∈R i or satisfies T(x,y,x s ,y s ) Guidelines (x,y)|(x,y)∈R i or R i Satisfies T(x,y,x s ,y s ) Guidelines Among them, T(x,y,x s ,y s ) is the merging criterion function used to determine whether to merge the pixel (x, y) into the region R i middle; According to the results of region growing, the distribution of noise in different regions is obtained, and the image is divided into multiple blocks B i , each block B i By the corresponding area R i definition; B i =(x,y)|(x,y)∈R i 。 4. The method for adaptive image denoising based on deep learning according to claim 1, characterized in that: Deep Learning Architecture M DL The specific design for implementing adaptive denoising of split images is as follows: Design deep learning model M DL , by learning the function f(I orig ; θ) to perform adaptive denoising: I denoised =f(I orig ;θ); In the above formula, I orig Represents the input image block, θ includes the degree of noise level reduction S, the degree of detail preservation M, the improvement of overall image quality P, and the visual evaluation V, I denoised Represents the denoised image.
5. The image adaptive denoising method based on deep learning according to claim 1, characterized in that: Design a multidimensional noise evaluation function as a deep learning model M DL The learning function in is designed as follows: The multidimensional evaluation functions include: (1) The degree of reduction of noise level S, which evaluates the overall denoising effect of the image: in, represents the signal variance, represents the noise variance, which is calculated as follows: Where N is the number of pixels in the image, I orig represents the original image, Represents the mean of the original image; (2) The degree of detail preservation M, which evaluates the preservation of the shape, size, color, and distribution characteristics of the image: In the formula, represents the mean value of the image before and after denoising, represents the variance of the image before and after denoising, Represents the covariance of the image before and after denoising. c1 and c2 are small constants to increase numerical stability. c1 = (k1L) 2 , c2=(k2L) 2 , L represents the dynamic range of pixel values, k1 and k2 are weight factors for controlling brightness and contrast; (3) Improvement of overall image quality P, evaluating the pixel differences in the subtle features of cell structure and vascular distribution tissue before and after image denoising: It represents the maximum value of the image pixels, and MSE represents the mean square error, which measures the difference between the original image and the processed image; (4) Visual evaluation V, which evaluates the visual difference before and after image denoising: The higher the noise level reduction S, the detail retention M, the overall image quality improvement P, and the visual evaluation V, the better. The multidimensional noise evaluation function f(I orig ; θ) is expressed as: f(I orig ;θ)=aS+bM+cP+dV; θ = a set of {a, b, c, d} parameters.
6. The method for adaptive image denoising based on deep learning according to claim 1, characterized in that: Introducing intelligent optimization algorithm as deep learning model M DL The loss function is as follows: The multidimensional noise evaluation function f(I orig ; θ) as the optimization function; first perform Gaussian process prior: f(I orig ;θ)~gp(m(I orig ;θ),k(θ,θ')); In the formula, gp(.) represents the Gaussian process prior, m(I orig ; θ) represents the mean function, k(θ,θ') represents the covariance function; Given some parameter-function value pairs Given the observed data D, the objective function is at the new point The posterior distribution of is: In the formula, In the above formula, Θ n ={θ1,...,θ n }, is the variance of the observation noise, I is an n×n identity matrix, k[(X n ,Θ n ),(X n ,Θ n )] is a parameter that replaces k(θ,θ'); represents the nth image before denoising, θ n Represents the nth parameter set, Represents the nth denoised image. Choose the optimization function f(I orig ; θ) The parameter with the largest value is taken as the optimal parameter θ = {a, b, c, d} of this round, and the iterative intelligent optimization algorithm is used as the step of the loss function of the deep learning model until the predetermined stopping condition is reached, and the optimal parameter is selected as the denoising parameter.