Image denoising method and device

Abstract

The application discloses an image denoising method and device, and the method comprises the following steps: constructing a hybrid partial differential image denoising model based on an adaptive edge detection algorithm, wherein the model combines ROF and LLT two denoising models in a manner of weighting by a fractional order edge detection function; determining an optimal order of the image fractional order edge function by a pollination algorithm according to feature information of the image; and solving the denoising model to output a denoised image. The application integrates the advantages of total variation and fourth-order partial differential equations, and compared with other advanced denoising algorithms, the application can maintain image edges and texture details while avoiding the "step effect" of smooth areas.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, specifically, it relates to an image denoising method and apparatus. Background Technology

[0002] Image denoising is a crucial research step in the field of digital image processing. It not only suppresses noise, providing more accurate information for subsequent processing such as image segmentation and compression, but also helps facilitate the solution of other image processing problems, such as image enhancement and image detection. Among these methods, image denoising based on partial differential equations has developed rapidly since the 1980s and has become a very important research direction.

[0003] Image denoising methods based on partial differential equations (PDEs) have their own advantages: PDEs provide a continuous model for image analysis, mathematically representing discrete images as continuous differential operators, facilitating local nonlinear filtering and preserving image edge information as much as possible. In the past few decades, fractional-order differential operators have been increasingly used in image processing. Fractional-order differential operators outperform traditional first- and second-order differential operators; experiments show that compared to integer-order edge detection operators, they not only effectively extract image edges but also suppress noise well. However, current techniques all use the same fractional order to process images, so the detection results still do not meet ideal requirements. Summary of the Invention

[0004] To address the above problems, this application discloses an image denoising method, comprising the following steps:

[0005] Step S1: Construct a hybrid partial differential image denoising model based on an adaptive edge detection algorithm. This model combines two denoising models, ROF and LLT, by weighting fractional edge detection functions.

[0006] Step S2: Determine the optimal order of the fractional-order edge function of the image based on the image's feature information using a flower pollination algorithm. ;

[0007] Step S3: Solve the denoising model and output the denoised image. .

[0008] Optionally, the energy function of the hybrid partial differential image denoising model is as follows:

[0009] (19)

[0010] in, It is a fractional-order edge detection function:

[0011] (14)

[0012] In the formula, For the denoised image Domain;

[0013] Noisy image;

[0014] The image after denoising;

[0015] for Weight parameters;

[0016] The gradient of the image;

[0017] , , , , All images are after denoising. The second derivative;

[0018] In order to ensure And add a positive number;

[0019] It is the contrast ratio;

[0020] It is dimensional order matrix;

[0021] , It is an image The fractional gradient vector magnitude.

[0022] Optional, .

[0023] Optionally, the step of determining the fractional order based on the image's feature information using a flower pollination algorithm is described. This includes the following steps:

[0024] Step S21, set the number of pollen gametes Randomly initialize the position of pollen gametes Number of iterations and conversion probability ;

[0025] Step S22: Calculate the local optimum for each gamete according to Formula 15, calculate the fitness value for each pollen gamete, and determine a local optimum as the global optimum based on the fitness value. The corresponding position is the position of the global optimum for the gamete.

[0026] (15)

[0027] in, , and These represent the peak signal-to-noise ratio, structural similarity, and mean absolute error of the image, respectively.

[0028] Step S23: Determine whether the number of iterations has been reached. If the number of iterations has been reached, proceed to step S24. If the number of iterations has not been reached, increment the number of iterations by 1, update the position of the pollen gametes, and return to step S22.

[0029] Step S24: Output the global optimal solution, that is, find the optimal order for the fractional-order edge function of the image. .

[0030] Optionally, determining a local optimum as the global optimum based on the fitness value means taking the largest local optimum as the global optimum.

[0031] Optional, peak signal-to-noise ratio for:

[0032] (16)

[0033] Structural similarity for:

[0034] (17)

[0035] Mean Absolute Error for:

[0036] (18)

[0037] In the formula, Indicates the size of the image. and These represent the length and width of the image, respectively. The image's feature information is... and They represent The mean and variance, and They represent The mean and variance, express and covariance, and The constants in the stability equation are represented.

[0038] Optionally, updating the location of the pollen gametes includes:

[0039] when When a global search is used, the position of the gametes is updated according to Formula 11. When using a local search, the position of the gametes is updated according to Formula 13.

[0040] The formula for updating the location of pollen in the global search is:

[0041] (11)

[0042] in, For the first The pollen in the first Position vector in the next iteration For the first The pollen in the first Position vector in the next iteration This is the optimal solution found among all current iterations. The parameters for pollination intensity of flowers, and Its expression is as follows:

[0043] (12)

[0044] In the formula, It is the standard gamma function. It indicates the large stride of the pollen gamete during movement.

[0045] The formula for updating the location of pollen in a local search is:

[0046] (13)

[0047] in, and These represent pollen gametes from different flowers of the same plant at the [number]th [time]. The position vector at the next iteration Indicates in Independent and uniform distribution on.

[0048] Optionally, the denoising model is solved to output the denoised image. ,include:

[0049] The Euler-Lagrange equation is obtained by using the finite difference method for the energy function of the denoising model;

[0050] Nonlinear partial differential equations are obtained by solving the Euler-Lagrange equations using the gradient descent method.

[0051] By dividing the image into a grid and discretizing it in time and space, information about the denoised image is obtained. Discrete representation of the iteration:

[0052] The denoised image is calculated using the discrete representation of the iteration. This process continues until the set number of iterations is reached, at which point the final denoised image is output. .

[0053] This application also discloses an image denoising apparatus, comprising:

[0054] The model construction module is used to construct a hybrid partial differential image denoising model based on an adaptive edge detection algorithm. This model combines two denoising models, ROF and LLT, by weighting fractional-order edge detection functions.

[0055] The optimal order determination module is used to determine the optimal order of the fractional-order edge function of an image based on its feature information using a flower pollination algorithm. ;

[0056] The model solving module is used to solve the denoising model and output the denoised image. .

[0057] This application also discloses an electronic device, the electronic device comprising:

[0058] At least one processor; and,

[0059] A memory communicatively connected to the at least one processor; wherein,

[0060] The memory stores instructions that can be executed by the at least one processor to cause the at least one processor to perform the image denoising method as described above.

[0061] This application designs a novel fractional-order edge detection function using the flower pollination algorithm and proposes a denoising algorithm that weights and naturally fuses total variational and fourth-order partial differential equations using the fractional-order edge detection function. The proposed model is applied to synthetic images and real ophthalmic ultrasound images, demonstrating the effectiveness of the proposed denoising method. Experimental results show that the model combines the advantages of total variational and fourth-order partial differential equations. Compared with other advanced denoising algorithms, the proposed model can preserve image edge and texture details while avoiding the "staircase effect" in smooth regions. Attached Figure Description

[0062] Figure 1 This is a flowchart of an image denoising method according to an embodiment of the present invention;

[0063] Figure 2 This invention provides an embodiment for determining the optimal order of an image fractional edge function using a flower pollination algorithm. Flowchart;

[0064] Figure 3The test images used in the comparative experiments of this invention;

[0065] Figure 4 The comparative experiment results of this invention show the results of different models processing the magnified portion of the "Peppers" image;

[0066] Figure 5 The comparative experiment results of the present invention are shown below, showing the results of different models processing the magnified portion of the "Lena" image.

[0067] Figure 6 The comparative experiment of this invention presents the experimental results of different models after processing the "Airplane" image;

[0068] Figure 7 The comparative experiment of this invention shows the experimental results of different models after processing the "House" image;

[0069] Figure 8 This refers to the grayscale value of the feature line in the 50th row of the comparative experimental image of this invention;

[0070] Figure 9 These are comparative experiments of the present invention, showing the effect of different models on the denoising of retinal detachment images.

[0071] Figure 10 These are comparative experiments of the present invention, showing the effect of different models on the denoising of choroidal detachment images.

[0072] Figure 11 The images show the denoising effects of different models on images of astrocytic degeneration in a comparative experiment of this invention. Detailed Implementation

[0073] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0074] The image denoising method of this embodiment includes the following steps:

[0075] Step S1: Construct a hybrid partial differential image denoising model based on an adaptive edge detection algorithm. This model effectively combines ROF (Rudin, Osher, Fatemi, also known as the TV model, e.g., "LI Rudin, S. Osher, and E. Fatemi, “Nonlinear total variation based noise removal algorithms,” Phys. D, Nonlinear Phenomena, vol. 60, nos. 1-4, pp. 259-268, 1992.) and LLT (Lysaker, Lundervold, and Tai, a fourth-order partial differential denoising model, e.g., "M. Lysaker, A. Lundervold, and XC Tai, “Noise removal using fourth-order partial differential equation with applications to medical magnetic resonance images in space and time,” IEEE Trans. Image Process, vol. 12, no. 12, pp.1579–1590, Dec.) by weighting fractional-order edge detection functions. (2003) The advantages of the two denoising models are as follows: The energy function of the denoising model is as follows:

[0076] (19)

[0077] in, It is a fractional-order edge detection function:

[0078] (14)

[0079] In the formula, For the denoised image Domain;

[0080] Noisy image;

[0081] The image after denoising;

[0082] for Weight parameters;

[0083] The gradient of the image;

[0084] , , , , All are images The second derivative;

[0085] In order to ensure And add a positive number;

[0086] It is the contrast ratio;

[0087] It is dimensional order matrix;

[0088] , It is an image To ensure the linearity of the fractional gradient vector, the absolute value is used instead of the square root operation, i.e. .

[0089] The ROF denoising model is characterized by its ability to preserve image edges, while the LLT denoising model is characterized by its ability to overcome the "staircase effect" in smooth regions. To remove image noise while preserving more texture details, a hybrid partial differential image denoising model is employed, combining the advantages of both models. This model utilizes an edge detection function... g This achieves a natural fusion of the ROF and LLT models. For regions with complex image textures, i.e., where edges are most likely to occur, When the value is large, the ROF model plays a major role; conversely, for regions with simple textures, i.e., locations where the signal is smooth, When the value is small, the LLT model plays a major role. Therefore, the new model retains the denoising advantages of the ROF model and preserves the edges of the image, while also inheriting the advantages of the LLT model, reducing the "staircase effect" in smooth areas.

[0090] Step S2: Determine the optimal order of the fractional-order edge function of the image based on the image's feature information using a flower pollination algorithm. The flower pollination algorithm is a heuristic optimization algorithm that simulates the biological pollination process of flowering plants. The main purpose of flower pollination is survival of the fittest and plant optimization. Its pollination process can be simulated as a global search and a local search in the optimization process, which includes the following steps:

[0091] Step S21: Initialize parameters and randomly initialize the positions of pollen gametes. Set the number of pollen gametes Number of iterations and conversion probability .

[0092] Step S22: Calculate the local optimum value of each gamete according to formula (15), calculate the fitness value of each pollen gamete, and determine a certain local optimum value as the global optimum value based on the fitness value. The corresponding position is the position of the global optimum value of the gamete. Preferably, the largest local optimum value is used as the global optimum value.

[0093] (15)

[0094] in, , and Peak signal-to-noise ratio (PSNR), structural similarity (SSIM), and mean absolute error (MAE) of an image are defined as follows:

[0095] Peak signal-to-noise ratio for:

[0096] (16)

[0097] Structural similarity for:

[0098] (17)

[0099] Mean Absolute Error for:

[0100] (18)

[0101] In the formula, This represents the size of the image; the image's feature information is... and They represent The mean and variance, and They represent The mean and variance, express and covariance, and This represents the constants in the stability equation. Among them, the larger the PSNR and SSIM values, and the smaller the MAE value, the better the quality of the denoised image.

[0102] Step S23: Determine if the iteration count has been reached. If it has, proceed to step S24. If not, increment the iteration count and update the position of the pollen gametes. When this happens, a global search is used, that is, the position of the gametes is updated according to formula (11). When using local search, the position of the gametes is updated according to formula (13), and then the process returns to step S22.

[0103] The formula for updating the location of pollen in the global search is:

[0104] (11)

[0105] in, For the first The pollen in the first Position vector in the next iteration This is the optimal solution found among all current iterations. The parameters for pollination intensity of flowers, and Its expression is as follows:

[0106] (12)

[0107] In the formula, It is the standard gamma function, and in general, it is selected as... This distribution only applies to large strides. efficient.

[0108] The formula for updating the location of pollen in a local search is:

[0109] (13)

[0110] in, and These represent pollen gametes from different flowers of the same plant at the [number]th [time]. The position vector at the next iteration Indicates in Independent and uniform distribution on.

[0111] Step S24: Output the global optimal solution, that is, find the optimal order for the fractional-order edge function of the image. .

[0112] Step S3: Solve the denoising model and output the denoised image. .

[0113] In step S3, the denoising model can be solved using the finite difference method, as detailed below:

[0114] Step S31: Obtain the Euler-Lagrange equation from the energy function of the denoising model using the finite difference method:

[0115] (28)

[0116] and Images exist direction and The derivative of the direction.

[0117] Step S32, by solving equation (28) using the gradient descent method, the nonlinear partial differential equation can be obtained:

[0118] (29)

[0119] Step S33, assuming the size of the image involved can be expressed as The matrix, Indicates the time step. If the spatial grid size is represented, then the time and space discretization are:

[0120] (30)

[0121] In the formula

[0122] Indicates the image in Directional discrete unit;

[0123] Indicates the image in Directional discrete unit;

[0124] ,

[0125] ,

[0126] Thus, we obtain information about the denoised image. Discrete representation of the iteration:

[0127] (31)

[0128] in, Indicates the first The next iteration;

[0129] , , Representing an image exist First-order partial derivative in the direction;

[0130] , , Representing an image exist First-order partial derivative in the direction;

[0131] , , , Representing images respectively exist Second-order partial derivatives of the direction, with respect to After guidance Guide, to After guidance Guide and in Second-order partial derivative in direction;

[0132] Indicates the image The second derivative of .

[0133] The denoised image is calculated using the discrete representation of the iteration. This process continues until the set number of iterations is reached, at which point the final denoised image is output. .

[0134] The performance of the image denoising method in this application will be verified below.

[0135] In this comparative experiment, the image denoising method of this application was compared with the classic ROF model, LLT model, and references 36 (D. Wang, JJ Nieto, X. Li, and Y. Li, “A spatially adaptive edge-preserving denoising method based on fractional-order variational PDEs,” IEEE Access, vol. 8, pp. 163115-163128, Sep. 2020), 19 (X. Zhang, R. Liu, J. Ren, and Q. Gui, “Adaptive fractional image enhancement algorithm based on rough set and particle swarm optimization,” Fractal Fract., vol. 6, no. 100, Jun. 2022), and 12 (F. Li, C. Shen, J. Fan, and C. Shen, “Image restoration combining a total variational filter and a fourth-order filter,” J. Vis. Commun. Image R., vol. 18, no. 4). The proposed model is compared with advanced denoising algorithms (pp. 322-330, Aug. 2007) to verify its effectiveness. A magnified region within the bounding box of the "Peppers" image is selected, such as... Figure 3 To verify the effectiveness of the model in processing smooth regions of images, we use 'a' as an example; Figure 3 The enlarged area within the bounding box of the "Lena" image (b) is used to verify the ability of the proposed model to preserve the edges of noisy images during denoising; [the model is then used in conjunction with other methods]. Figure 3 c, Figure 3 Images of "Airplane" and "House" were used to verify the ability of the proposed model to preserve image information during denoising. Furthermore, images of retinal detachment, choroidal detachment, and vitreous hemorrhage—images that are difficult for clinicians to distinguish—were selected for denoising. Figure 3 e- Figure 3 The g of the method further demonstrates that the proposed method is also suitable for processing real-world noisy images.

[0136] To further verify the effectiveness of the proposed model, this paper selects formulas (16)-(18) as the evaluation index of the image to objectively evaluate the quality of the denoised image from a numerical perspective.

[0137] The experiment was first conducted using images from an image library, as follows:

[0138] In this comparative experiment, for Figure 3 a- Figure 3 Adding different noise standard deviations to d , The effectiveness of the proposed model for image denoising is verified by comparing the denoised Gaussian noise images. Specifically, Figures 4 to 7 Add standard deviation to different model pairs Experimental images after Gaussian noise image processing. Figure 8 The 50th row of the image is the feature line of the ROF model, LLT model and the model in this paper. Tables 1 to 3 show the PSNR, SSIM and MAE of the images after denoising images with different degrees of Gaussian noise by different models.

[0139] from Figure 4 The magnified results of the Peppers image show that the ROF model excessively smooths the surface of the peppers, causing the image to become blurry; after processing by the LLT model, many isolated black and white dots are generated on the surface of the peppers; after processing by Reference 36, noise still exists on the surface of the peppers; after processing by Reference 19, the noise in the image can be removed, but the smoothness of the image surface is also destroyed; after processing by Reference 12, some black and white spots are still left. However, after processing the image, the model in this paper retains the gloss of the surface of the peppers in the image, and there is no problem of excessive noise removal or destruction of smooth areas, which shows the effectiveness of the model in this paper in processing the smooth areas of the image.

[0140] Figure 5 These are experimental results of local magnification of the Lena image. The ROF model and reference 19 enhanced edge and texture details after image processing, but destroyed smooth areas of the image; the LLT model preserved the edges of the hat, but left isolated black and white dots in the processed image, affecting the visual effect; compared with other denoising models, the model in this paper has a significant enhancement effect on the edges of the entire hat and the outline, while better preserving smooth areas such as the background wall in the image, indicating that the model in this paper can better preserve the edges of noisy images during denoising.

[0141] Figure 6 and Figure 7 The figures show the experimental results for denoising airplane and house images, respectively. As can be seen from the figures, the ROF model destroys the image quality. Figure 6 The smooth region of the cloud in b, processed by the LLT model Figure 6The letter part of the 'c' becomes blurred; Reference 36 pairs Figure 7 After processing by the model in this paper, the images still contain noise. In reference 19, the images become blurred after processing, and in reference 12, a small number of black and white dots remain in both images. However, after processing by the model in this paper... Figure 6 The letter 'a' and the outline of the clouds are clearer. Figure 7 The edges of houses in model a are more distinct, and the overall outlines of the paths and trees in front of the buildings are clearer, which further illustrates that the model in this paper can effectively balance the enhancement of textured areas and the preservation of smooth areas.

[0142] From a data perspective, this paper selects an image feature line. Appropriate fluctuations in the grayscale values ​​of the feature line indicate rich image details, while excessive and frequent fluctuations indicate that the image is affected by noise. The feature line selected in this experiment is... Figure 3 a- Figure 3 The 50th row of the image in d shows the feature line data fluctuation curve as shown in the figure. Figure 8 The curves obtained by the model in this paper retain the key feature peaks of the original image while smoothing out noise. The ROF model curve removes too many peaks, and the curve is too smooth, destroying the edge and detail information of the original image. The curves obtained by the LLT model not only fail to effectively reduce peaks, but also cause larger and more frequent fluctuations compared to the original image. This once again proves the effectiveness of the model in this paper.

[0143] As can be seen from the data in Tables 1-3, the PSNR value of the image processed by the model in this paper is relatively large, the SSIM is closer to 1, and the MAE value is relatively small. This further demonstrates that the model in this paper enhances the edge and texture details of the image while keeping the smooth area as consistent as possible with the original image.

[0144] Table 1. PSNR values ​​of four images with different noise intensities processed by different models.

[0145]

[0146] Table 2. SSIM values ​​of four images with different noise intensities processed by different models.

[0147]

[0148] Table 3. MAE values ​​of four images with different noise intensities processed by different models.

[0149]

[0150] The following experiment uses real ophthalmic ultrasound images:

[0151] This experiment investigated the addition of speckle noise of varying intensities. Figure 3 e- Figure 3 Denoising was performed on g-band ophthalmic ultrasound images to verify the effectiveness of the proposed model for denoising real ultrasound images. Experimental results are as follows: Figures 9 to 11 As shown, where Figure 9 The images show the results of different models after processing retinal detachment images with speckle noise added at a variance of 0.02. Figure 10 The results of denoising images of choroid detachment with speckle noise of 0.02 variance are presented for different models. Figure 11 Tables 4 to 6 show the performance metrics of various models on images of vitreous volume blood with speckle noise added at variances of 0.02, 0.04, and 0.06.

[0152] As can be seen from the figures, although the ROF model can also remove speckle noise, its denoising effect is not ideal. The overall brightness becomes darker, the overall image becomes blurry, and it is difficult to distinguish whether the lesion is caused by retinal detachment, choroidal detachment, or vitreous organization caused by vitreous hemorrhage. The LLT model produces some distortion after processing. The images processed by references 36 and 12 still have noise. Reference 19 has a better denoising effect than the other models, but it blurs the image edges while denoising. Visually, the model in this paper has the best denoising effect. The internal structure of the denoised image is well preserved, the brightness information is also improved compared with other models, and the overall shape of the image is well preserved.

[0153] Tables 4 to 6 show the PSNR, SSIM, and MAE of various models for different ophthalmic ultrasound images under different noise intensities. It can be seen that as the noise intensity increases, the performance values ​​of each model decrease rapidly, while the performance of the proposed model decreases slowly. Furthermore, for these three images, the proposed model outperforms the other models in all performance indicators under different noise intensities, objectively reflecting the effectiveness of the proposed model. Experimental comparisons demonstrate that the proposed model is applicable not only to synthetic images but also to real ultrasound images.

[0154] Table 4. PSNR values ​​of different models after processing ophthalmic images with added noise of different intensities.

[0155]

[0156] Table 5. SSIM values ​​of different models after processing ophthalmic images with added noise of different intensities.

[0157]

[0158] Table 6. MAE values ​​of different models after processing ophthalmic images with added noise of different intensities.

[0159]

[0160] Of course, the present invention may have other various embodiments. Without departing from the spirit and essence of the present invention, those skilled in the art can make various corresponding changes and modifications according to the present invention, but these corresponding changes and modifications are all within the protection scope of the claims of the present invention.

Claims

1. An image denoising method, characterized in that, Includes the following steps: Step S1: Construct a hybrid partial differential image denoising model based on an adaptive edge detection algorithm. This model combines two denoising models, ROF and LLT, by weighting fractional edge detection functions. Step S2: Determine the optimal order of the fractional-order edge function of the image based on the image's feature information using a flower pollination algorithm. ; Step S3: Solve the denoising model and output the denoised image. , The energy function of the hybrid partial differential image denoising model is as follows: (19) in, It is a fractional-order edge detection function: ,(14) In the formula, For the denoised image Domain; Noisy image; The image after denoising; for Weight parameters; The gradient of the image; , , , , All images are after denoising. The second derivative; In order to ensure And add a positive number; It is the contrast ratio; It is dimensional order matrix; , It is an image The fractional gradient vector magnitude, The fractional order is determined based on the image's feature information using a flower pollination algorithm. This includes the following steps: Step S21, set the number of pollen gametes Randomly initialize the position of pollen gametes Number of iterations and conversion probability ; Step S22: Calculate the local optimum for each gamete according to Formula 15, calculate the fitness value for each pollen gamete, and determine a local optimum as the global optimum based on the fitness value. The corresponding position is the position of the global optimum for the gamete. (15) in, , and These represent the peak signal-to-noise ratio, structural similarity, and mean absolute error of the image, respectively. Step S23: Determine whether the number of iterations has been reached. If the number of iterations has been reached, proceed to step S24. If the number of iterations has not been reached, increment the number of iterations by 1, update the position of the pollen gametes, and return to step S22. Step S24: Output the global optimal solution, that is, find the optimal order for the fractional-order edge function of the image. .

2. The image denoising method according to claim 1, characterized in that, 。 3. The image denoising method according to claim 1, characterized in that, The statement that a local optimum is determined as the global optimum based on the fitness value means that the largest local optimum is taken as the global optimum.

4. The image denoising method according to claim 1, characterized in that, Peak signal-to-noise ratio for: ,(16) Structural similarity for: ,(17) Mean Absolute Error for: ,(18) In the formula, and These represent the length and width of the image, respectively. The image's feature information is... and They represent The mean and variance, and They represent The mean and variance, express and covariance, and The constants in the stability equation are represented.

5. The image denoising method according to claim 1, characterized in that, The location of the updated pollen gametes includes: when When a global search is used, the position of the gametes is updated according to Formula 11. When using a local search, the position of the gametes is updated according to Formula 13. The formula for updating the location of pollen in the global search is: ,(11) in, For the first The pollen in the first Position vector in the next iteration For the first The pollen in the first Position vector in the next iteration This is the optimal solution found among all current iterations. The parameters for pollination intensity of flowers, and Its expression is as follows: ,(12) In the formula, It is the standard gamma function. This indicates the large stride length of pollen gametes during movement. The formula for updating the location of pollen in a local search is: ,(13) in, and These represent pollen gametes from different flowers of the same plant at the [number]th [time]. The position vector at the next iteration Indicates in Independent and uniform distribution on.

6. The image denoising method according to claim 1, characterized in that, The denoising model is solved to output the denoised image. ,include: The Euler-Lagrange equation is obtained by using the finite difference method for the energy function of the denoising model; Nonlinear partial differential equations are obtained by solving the Euler-Lagrange equations using the gradient descent method. By dividing the image into a grid and discretizing it in time and space, information about the denoised image is obtained. Discrete representation of the iteration: The denoised image is calculated using the discrete representation of the iteration. This process continues until the set number of iterations is reached, at which point the final denoised image is output. .

7. An image denoising device, characterized in that, include: The model construction module is used to construct a hybrid partial differential image denoising model based on an adaptive edge detection algorithm. This model combines two denoising models, ROF and LLT, by weighting fractional-order edge detection functions. The optimal order determination module is used to determine the optimal order of the fractional-order edge function of an image based on its feature information using a flower pollination algorithm. ; The model solving module is used to solve the denoising model and output the denoised image. , The energy function of the hybrid partial differential image denoising model is as follows: (19) in, It is a fractional-order edge detection function: ,(14) In the formula, For the denoised image Domain; Noisy image; The image after denoising; for Weight parameters; The gradient of the image; , , , , All images are after denoising. The second derivative; In order to ensure And add a positive number; It is the contrast ratio; It is dimensional order matrix; , It is an image The fractional gradient vector magnitude, The fractional order is determined based on the image's feature information using a flower pollination algorithm. This includes the following steps: Step S21, set the number of pollen gametes Randomly initialize the position of pollen gametes Number of iterations and conversion probability ; Step S22: Calculate the local optimum for each gamete according to Formula 15, calculate the fitness value for each pollen gamete, and determine a local optimum as the global optimum based on the fitness value. The corresponding position is the position of the global optimum for the gamete. (15) in, , and These represent the peak signal-to-noise ratio, structural similarity, and mean absolute error of the image, respectively. Step S23: Determine whether the number of iterations has been reached. If the number of iterations has been reached, proceed to step S24. If the number of iterations has not been reached, increment the number of iterations by 1, update the position of the pollen gametes, and return to step S22. Step S24: Output the global optimal solution, that is, find the optimal order for the fractional-order edge function of the image. .

8. An electronic device, characterized in that, The electronic device includes: At least one processor; and, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to cause the at least one processor to perform the image denoising method as described in any one of claims 1 to 6.

Images