Image denoising method, device and terminal based on partial differential equation learning network

Abstract

The present invention belongs to the field of image processing technology, and discloses an image denoising method, device and terminal based on a partial differential equation learning network, including: image denoising data set acquisition and input, output of predicted denoised image time series through a partial differential equation learning network, comparison with the real denoised image time series to obtain loss, minimizing loss to train parameters, and then introducing random integral theory to construct an improved numerical solver for calculating the solution of the partial differential equation, and finally the network outputs the partial differential equation corresponding to the image denoising data set and the solution of the equation, and then the trained network is combined with a denoising post-processing module for image denoising to obtain a denoising result. The present invention can solve the problem of difficulty in adjusting parameters of the image denoising algorithm based on partial differential equations, and improve the error problem of the existing numerical solver, improve the accuracy of the solution, and finally use the image deconstruction operation of the denoising post-processing module to obtain a more competitive denoising result.

Description

Technical Field

[0001] The present invention belongs to the technical field of image processing, and in particular relates to an image denoising method, device and terminal based on a partial differential equation learning network. Background Art

[0002] At present, images are one of the main media for humans to obtain and transmit information in social activities, and have become an important way for humans to perceive the world and then transform the world. However, noise pollution will inevitably occur in the process of image acquisition and transmission, which will seriously reduce the visual quality of the acquired image. In order to improve the image quality and increase the efficiency of reading image data, it is necessary to process the acquired noisy image before analyzing and using the image to restore important information from the real world. With the development and popularization of computer technology, image denoising algorithms have continued to develop in recent decades and have been widely used in important national defense and science and technology fields such as biomedicine, remote sensing aerospace, and industrial design.

[0003] Traditional image denoising algorithms mainly include image smoothing represented by mean filtering and median filtering, which processes the image with a sliding window and performs simple numerical calculations on the pixels in the window; however, it processes the image as a whole and does not retain the image features. Wavelet transform, represented by spatial correlation and wavelet threshold filtering, constructs corresponding rules based on the mechanism that the wavelet coefficients of signals and noise have different properties at different scales, and uses other mathematical methods in the wavelet domain to process the wavelet coefficients of noisy signals; however, if the threshold is not properly selected, the negative effect of blurring the image edge will occur while removing the noise. Linear and nonlinear filtering, represented by Kalman filtering and differential filtering, processes noisy input and observation signals based on linear or nonlinear state space representation to obtain the system state or real signal; however, the algorithm requires many prior conditions, especially the moment mean and covariance of the noise, which limits its flexibility in practical applications. Partial differential equation-based algorithms, represented by isotropic diffusion, anisotropic diffusion, and variational methods, have two methods: process-oriented and object-oriented. However, the algorithm needs to define the algorithm equation form based on people's experience, and spend a lot of time exploring the most appropriate parameter configuration. In addition, model-based image denoising algorithms and neural network-based image denoising algorithms have also appeared in recent years. Model-based denoising algorithms include the BM3D algorithm of K.Dabov et al., which divides the image into 3D groups according to features, then filters each group, and finally fuses the 3D groups to obtain the final filtered image. However, the BM3D algorithm has several disadvantages: the optimization algorithm is usually time-consuming and cannot be directly used to remove spatial variation noise; manually made image priors (such as sparsity and non-local self-similarity) are required, and these prior information may not be sufficient to describe complex image structures. Denoising algorithms based on neural networks include HCBurger et al., who used multi-layer perceptron MLP for image denoising, with the aim of learning the underlying image prior from the training set of noise-result image pairs. Later, in order to extract and utilize potential features, convolutional neural networks (CNNs) were more widely used. The Denoising Convolutional Neural Network (DnCNN) proposed by K. Zhang combines batch normalization, rectified linear units, and residual learning into the CNN model to complete image denoising. The Generative Convolutional Blind Denoising (GCBD) proposed by J. Chen uses GAN to solve the problem of unpaired noisy images.D. Valsesia proposed a graph convolutional denoising network (GCDN) to capture self-similar information. However, the existing neural network-based denoising algorithms have limited flexibility, and the learning models are usually customized for specific noise levels. The most important data sets for network training have no unified standards and are difficult to obtain. In addition, neither model-based algorithms nor neural network-based algorithms can summarize their corresponding physical laws, which is an urgent problem to be solved by image denoising algorithms. With the increase in visual sensors, the composition of image noise has become more complex, and parameter adjustment has become increasingly difficult; real image data is difficult to obtain in terms of time sample data, and often only a small amount of data is available, which cannot meet the needs of network training; denoising is difficult on images with complex texture edges.

[0004] Through the above analysis, the problems and defects of the existing technology are as follows: the traditional image denoising algorithm needs to define the algorithm equation form based on people's experience, and spends a lot of time to explore the most suitable parameter configuration, which has certain limitations in image denoising applications; the image denoising algorithm based on models and neural networks is difficult to obtain the large amount of high-quality data sets required for its training, and the model has a large amount of calculation and weak generalization ability. Summary of the invention

[0005] In view of the problems existing in the prior art, the present invention provides an image denoising method, device and terminal based on a partial differential equation learning network.

[0006] The present invention is implemented as follows: an image denoising method based on a partial differential equation learning network includes: acquiring and inputting an image denoising data set, outputting a predicted denoised image time series through a partial differential equation learning network, comparing the predicted denoised image time series with the real denoised image time series to obtain a loss, minimizing the loss to train parameters, and then introducing random integral theory to construct an improved numerical solver for calculating the solution of the partial differential equation. Finally, the network outputs the partial differential equation corresponding to the image denoising data set and the solution of the equation, and then the trained network is combined with a denoising post-processing module for image denoising to obtain a denoising result.

[0007] Furthermore, the image denoising method based on the partial differential equation learning network specifically includes:

[0008] (1) Obtaining and inputting an image denoising dataset: Obtaining an image denoising dataset and inputting the time series data in the image denoising dataset into the partial differential equation learning network one by one;

[0009] (2) Construct and train a partial differential equation learning network: The partial differential equation learning network consists of multiple blocks. The image denoising time series data passes through multiple blocks, and each block outputs a predicted denoising image time series. Compared with the real denoised image time series Compare and calculate the loss function Minimize the loss function by gradient descent, complete parameter training once, and perform layered training on multiple blocks;

[0010] (3) Obtaining output by improving the numerical solver: After training is completed, the parameters are output to obtain the partial differential equation form corresponding to the image denoising data set, as well as the partial differential equation solution of the improved numerical solver by introducing stochastic integral theory;

[0011] (4) Image denoising using the partial differential equation learning network and the denoising post-processing module: The noisy image to be denoised is input into the trained partial differential equation learning network, then passes through the denoising post-processing module, and finally outputs the denoised clean image.

[0012] Furthermore, in step (1) of acquiring and inputting the image denoising dataset, in order to obtain the image denoising dataset, simulated white noise is added to the original clean image dataset to obtain a simulated noise dataset, and then denoising is performed using an existing image denoising method based on partial differential equations to obtain a series of intermediate denoising results and a final denoising result, forming an image denoising dataset with a spatial size of N×M and a time length of T.

[0013] Furthermore, in step (2) of constructing and training the partial differential equation learning network, the more blocks there are, the deeper the network is. When the network is at a certain depth, the training effect is best.

[0014] Furthermore, the Block structure of the partial differential equation learning network includes:

[0015] The first step is to calculate the partial differential terms of the partial differential equation candidate library:

[0016] It is known that a partial differential equation is an equation involving partial differentials of an unknown function, and its general form is:

[0017]

[0018] Where t represents time, x and y represent space, ...is a partial differential term. The Chebyshev polynomial interpolation method is used to find the partial differential term of the Block input data as a candidate library for partial differential equations part of;

[0019] The second step is to build a candidate library for partial differential equations:

[0020] The edge features of the image are also added to the candidate library. After comparing multiple edge operators, the edge stop function is used It means, where k is a constant, is the gradient of the image; the partial differential items and edge feature items in the candidate library are arranged and combined, and added to the candidate library. The build is complete;

[0021] The third step is to obtain the output of Block by improving the numerical solver:

[0022] Candidate Library and network parameters Multiply and calculate to get the partial differential in time dimension Then, random integral theory is introduced to construct a new numerical solution formula, and finally the output of Block is obtained. in The calculation formula is:

[0023]

[0024] in N(0,1) is a random number from a standard normal distribution, and max and min are the maximum and minimum values ​​of the data respectively.

[0025] Further, the hot start, layered training and loss function of the partial differential equation learning network include:

[0026] (1) Network hot start: The parameters of the partial differential equation learning network are initialized by hot start, and the initial values ​​of the parameters are obtained by STRidge regression;

[0027] (2) Hierarchical training: Input a batch of data into the first block and get the output After that, the loss is calculated and the parameters are trained. After the training is completed, another batch of data is used to restart the training on the first two blocks. This step is repeated until all blocks are completed in the present invention. The parameters are shared between blocks during the whole training process.

[0028] (3) Loss function: The loss function of the partial differential equation learning network consists of a data loss term and a regularization term. The data loss term uses mean square error, and the regularization term uses L1 regularization:

[0029]

[0030] in It represents the mean square error between two sets of data.

[0031] Furthermore, the processing method of the denoising post-processing module includes:

[0032] The first step is to group similar blocks: divide the image into multiple reference blocks. For each reference block, use the Euclidean distance in the surrounding area to measure the similarity of nearby blocks, select similar blocks with insignificant differences, and finally integrate them to obtain the 3D similar block group corresponding to the reference block.

[0033] The second step is DCT transformation: the obtained three-dimensional similar block group is transformed into DCT in the third dimension, which can remove the information redundancy brought by the similar block image, improve the efficiency, and distinguish the similar blocks with different noise pollution degrees in the same similar block matrix, so as to distinguish and process them in the subsequent denoising process and improve the denoising effect;

[0034] The third step is PDE denoising: the PDE learned by the partial differential deep network is used to solve the input similar block group, and the output result is the similar block matrix after denoising;

[0035] The fourth step is inverse transformation and aggregation operation: perform inverse DCT transformation on the similar block matrix, then put the inverse transformed similar blocks back to their original positions, use the number of non-zero components to statistically superimpose the weights, and finally divide the superimposed image by the weights of the corresponding points to obtain the aggregated denoised image.

[0036] Another object of the present invention is to provide a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the image denoising method based on the partial differential equation learning network.

[0037] Another object of the present invention is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, enables the processor to execute the image denoising method based on a partial differential equation learning network.

[0038] Another object of the present invention is to provide an information data processing terminal, which is used to implement the image denoising method based on the partial differential equation learning network.

[0039] In combination with the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solutions to be protected by the present invention are analyzed from the following aspects:

[0040] First, in view of the technical problems existing in the above-mentioned prior art and the difficulty of solving the problems, the technical solutions to be protected by the present invention and the results and data during the research and development process are closely combined to analyze in detail and deeply how the technical solutions of the present invention solve the technical problems, and some creative technical effects brought about after solving the problems. The specific description is as follows:

[0041] Since the present invention proposes to combine the partial differential equation learning theory with the field of image denoising, it solves the problem that traditional image algorithms need to define the algorithm equation form based on people's experience and spend a lot of time exploring the most suitable parameter configuration, so that the present invention has the ability to learn partial differential equations.

[0042] Since the present invention designs a partial differential equation learning network with a multi-block structure, its hierarchical training system solves the problem that the image denoising data has a small amount of time samples, which affects the accuracy of the partial differential equation learning results. This makes the present invention have the advantage of higher accuracy under training with less time sample data.

[0043] Since the present invention introduces the theory of random integration into the numerical solver, the error problem existing in the traditional numerical solver for the solution of partial differential equations is solved, so that the present invention has the advantages of higher solution accuracy and better denoising effect.

[0044] Since the present invention introduces image deconstruction operation in the denoising post-processing module, the problem that the best denoising effect cannot be achieved due to complex image information during the image denoising process is solved, so that the present invention has the denoising ability to process complex texture edge noise images.

[0045] Second, considering the technical solution as a whole or from the perspective of the product, the technical effects and advantages of the technical solution to be protected by the present invention are described in detail as follows:

[0046] The present invention provides a practical solution, solves the parameter adjustment problem of image denoising, improves the accuracy and robustness of partial differential equation learning, effectively solves the network training problem under a small amount of time sample data, and has a better denoising effect on images with complex textures. The present invention has higher solution accuracy and better denoising effect.

[0047] Third, as auxiliary evidence of the inventiveness of the claims of the present invention, it is also reflected in the following important aspects:

[0048] (1) The technical solution of the present invention fills the technical gap in the industry at home and abroad: the existing algorithms have not yet combined the field of image denoising with the field of partial differential learning based on neural networks. The present invention combines the two fields and introduces deep learning, which can not only better learn partial differential equations, but also achieve better image denoising results.

[0049] (2) Whether the technical solution of the present invention solves the technical problems that people have been eager to solve but have never been able to solve: In the existing field of image denoising, there are problems with the traditional method of manually defining equations and complex parameter adjustments, and the network method has difficulty in obtaining data. At the same time, the field of partial differential equation learning has the limitation of being difficult to learn with a small amount of time sample data. The present invention solves the problems in both fields, saving labor costs and the difficulty of data acquisition. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 Flowchart of the image denoising method based on partial differential equation learning network provided by the embodiment of the present invention

[0051] Figure 2 It is a network structure diagram of a partial differential equation learning network provided by an embodiment of the present invention.

[0052] Figure 3 It is a single Block structure diagram provided by the implementation of the present invention.

[0053] Figure 4 It is a flow chart of the denoising post-processing module provided by the implementation of the present invention.

[0054] Figure 5 This is a graph of image denoising results in an image with complex edges provided by the implementation of the present invention.

[0055] Figure 6 This is a graph of image denoising results in an image with simple edges provided by the implementation of the present invention.

[0056] Figure 7 This is an image denoising result diagram in an image with complex textures provided by the implementation of the present invention.

[0057] Figure 8 This is an image denoising result diagram in an image with simple texture provided by the implementation of the present invention. DETAILED DESCRIPTION

[0058] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0059] 1. Explanatory Examples In order to enable those skilled in the art to fully understand how to implement the present invention, this section provides an illustrative example that expands and describes the technical solution of the claims.

[0060] like Figure 1 As shown, the image denoising method based on partial differential equation learning network provided by the present invention comprises the following steps:

[0061] S101: Acquisition and input of image denoising dataset;

[0062] S102: Build and train a partial differential equation learning network;

[0063] S103: obtaining output by improving the numerical solver;

[0064] S104: Perform image denoising using a partial differential equation learning network and a denoising post-processing module.

[0065] The technical solution of the present invention is further described below in conjunction with the accompanying drawings.

[0066] like Figure 2 As shown, the network structure of the partial differential equation learning network provided by the present invention is further described:

[0067] The partial differential equation learning network consists of multiple blocks. The more blocks there are, the deeper the network is. When the network is at a certain depth, its training effect will be best. First, the network parameters are initialized using a hot start method, and the initial values ​​of the parameters are obtained using STRidge regression. Then, the image denoising time series data passes through multiple blocks, and each block outputs a predicted denoised image time series. Compared with the real denoised image time series Compare and calculate the loss function The loss function consists of a data loss term and a regularization term. The data loss term uses mean square error, and the regularization term uses L1 regularization. Regularization:

[0068]

[0069] in It represents the mean square error between two sets of data.

[0070] Minimize the loss function by gradient descent to complete parameter training. Perform layered training on multiple blocks, input a batch of data into the first block, and get the output After that, the loss is calculated and the parameters are trained. After the training is completed, another batch of data is used to restart the training on the first two blocks. This step is repeated until all blocks are completed in the present invention. The parameters are shared among the blocks during the whole training process.

[0071] like Figure 3 As shown, a single Block structure diagram constructed by the present invention is further described:

[0072] The first step is to calculate the partial differential terms of the partial differential equation candidate library:

[0073] It is known that the partial differential equation contains the equation of the partial differential of the unknown function, and its general form is:

[0074]

[0075] Where t represents time, x and y represent space, ...is a partial differential term. Apply Chebyshev polynomial interpolation to find the partial differential term of the Block input data as a candidate library for partial differential equations part of.

[0076] The second step is to build a candidate library for partial differential equations:

[0077] In order to better learn the partial differential equations of the image denoising process, the edge features of the image are also added to the candidate library. After comparing multiple edge operators, the edge stop function is used. To represent, where k is a constant, is the gradient of the image. The partial differential items and edge feature items in the candidate library are arranged and combined and added to the candidate library. Build completed.

[0078] The third step is to obtain the output of Block by improving the numerical solver:

[0079] Candidate Library and network parameters Multiply and calculate to get the partial differential in time dimension Then, random integral theory is introduced to construct a new numerical solution formula, and finally the output of Block is obtained. in The calculation formula is:

[0080]

[0081] in N(0,1) is a random number from a standard normal distribution, and max and min are the maximum and minimum values ​​of the data respectively.

[0082] like Figure 4 As shown, the denoising post-processing module flow chart constructed by the present invention includes the following steps:

[0083] The first step is grouping similar blocks: divide the image into multiple reference image blocks. For each reference image block, use the Euclidean distance in the surrounding area to measure the similarity of nearby blocks, select similar blocks with insignificant differences, and finally integrate them to obtain the three-dimensional similar block group corresponding to the reference image block.

[0084] The second step is DCT transformation: the obtained three-dimensional similar block group is transformed into DCT in the third dimension. Its function is to remove the information redundancy brought by the similar block image, improve efficiency, and distinguish similar blocks with different noise pollution degrees in the same similar block matrix, and perform differentiated processing in the subsequent denoising process to improve the denoising effect.

[0085] The third step is PDE denoising: the PDE learned by the partial differential deep network is used to solve the input similar block group, and the output result is the similar block matrix after denoising.

[0086] The fourth step is inverse transformation and aggregation operation: perform inverse DCT transformation on the similar block matrix, then put the inverse transformed similar blocks back to their original positions, use the number of non-zero components to statistically superimpose the weights, and finally divide the superimposed image by the weights of the corresponding points to obtain the aggregated denoised image.

[0087] 2. Application Examples The embodiments of the present invention have achieved some positive effects in the process of research and development or use, and indeed have great advantages over the prior art. The following content is described in combination with data, charts, etc. of the test process.

[0088] 1. Simulation conditions:

[0089] The hardware platform of the simulation experiment of the present invention is: the processor is Intel(R) Core(TM) i7-10700K CPU, the main frequency is 2.9GHz, the memory is 64GB, and the graphics card is NVIDIA GeForce RTX 3060.

[0090] The software platform of the simulation experiment of the present invention is: Ubuntu 20.04 operating system, Pycharm2021 software, python3.7 and Pytorch deep learning framework.

[0091] 2. Simulation content and result analysis:

[0092] The simulation experiment of the present invention is verified by experiments on SIDD and BSDS datasets. 100 pictures in SIDD dataset are selected to obtain image denoising dataset as the training set of partial differential equation learning network, and 500 BSDS datasets are selected as test datasets. The number of training iterations of partial differential equation learning network is 20 epochs, the batch size of the network is set to 1, the number of blocks is 30, the reference image block size blocksize in denoising post-processing module is 10, and the number of similar blocks is set to 16.

[0093] In the simulation experiment, the existing technologies used are:

[0094] The total variation image denoising algorithm proposed by Chan TF, Shen J et al. in "Image Processing and Analysis: Variational, PDE, Wavelet, and Stochastic Methods[M]. Siam, 2005", referred to as the TV model.

[0095] In order to evaluate the partial differential equation learning ability of the present invention, five classic partial differential equations in the field of physics and image denoising are selected, namely the Burgers equation, the Navier-Stokes equation, the heat diffusion equation, the PM model and the TV model. As shown in Table 1, the partial differential equations output by the present invention, the errors and average errors of the coefficients of the partial differential equations, and the errors of the solutions of the equations are listed, and calculated using MSE.

[0096] Table 1 The partial differential equation learning results of the present invention in the simulation experiment

[0097]

[0098] As can be seen from Table 1, the formulas of the five equations are complex, but the present invention still accurately learns each term of the equation. In addition, it also has a high degree of accuracy in learning the coefficients of each term. At the same time, in solving partial differential equations, the error of the solution is relatively low.

[0099] In order to qualitatively evaluate the simulation effect of the present invention, four images with different texture and edge complexity are used to subjectively compare the denoising results. The denoising results obtained are as follows: Figures 5 to 8 shown. Figures 5 and 6 It is the denoising effect of different complexity of image edge information. Figure 7 and Figure 8 denoising effects of different image texture information complexities, where (a) is the original image, (b) is the noisy image, (c) is the TV model denoising result, and (d) is the denoising result of the present invention.

[0100] Depend on Figures 5 to 8It can be concluded that the present invention can be applied to image denoising of natural images. It can be seen that on images with complex edges, the method of the present invention has a good ability to identify and protect edges compared to the prior art, so that edge information can be retained while denoising; and on images with simple edges, when edge information exists, the method of the present invention can better achieve the denoising effect. On images with complex textures, the present invention can perform denoising without changing the texture structure; on images with simple textures, the present invention has a better denoising effect on smooth areas and can clearly retain the texture. When the image edge and texture information are complex, the subjective evaluation effect of image denoising is better than that of the TV model. It shows that under the structure of the partial differential equation learning network and the improved numerical solver, the solution of the partial differential equation is more accurate, that is, the image denoising result is better, and there is a denoising post-processing module to perform image deconstruction operations, which further improves the image denoising effect.

[0101] In order to quantitatively evaluate the simulation effect of the present invention, the present invention uses peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) as performance evaluation indicators to compare with the prior art. The comparison results are shown in Table 2:

[0102] Table 2 Comparison of evaluation indicators of the present invention and the prior art in simulation experiments

[0103]

[0104]

[0105] From Table 2, we can see that Figures 5 to 8 and BSDS datasets, the present invention is superior to the prior art in all evaluation indicators, which proves that the present invention can achieve better and more robust denoising results on images with complex edge and texture information, and is a very practical image denoising method.

[0106] It should be noted that the embodiments of the present invention can be implemented by hardware, software, or a combination of software and hardware. The hardware part can be implemented using dedicated logic; the software part can be stored in a memory and executed by an appropriate instruction execution system, such as a microprocessor or dedicated design hardware. It can be understood by a person of ordinary skill in the art that the above-mentioned devices and methods can be implemented using computer executable instructions and / or contained in a processor control code, such as a carrier medium such as a disk, CD or DVD-ROM, a programmable memory such as a read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. Such code is provided on the carrier medium. The device and its modules of the present invention can be implemented by hardware circuits such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, etc., or programmable hardware devices such as field programmable gate arrays, programmable logic devices, etc., can also be implemented by software executed by various types of processors, and can also be implemented by a combination of the above-mentioned hardware circuits and software, such as firmware.

[0107] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by any technician familiar with the technical field within the technical scope disclosed by the present invention and within the spirit and principle of the present invention should be covered by the protection scope of the present invention.

Claims

1. An image denoising method based on a partial differential equation learning network, characterized in that: The image denoising method based on the partial differential equation learning network includes: acquiring and inputting an image denoising data set, outputting a predicted denoising image time series through the partial differential equation learning network, comparing the predicted denoising image time series with the real denoising image time series to obtain a loss, minimizing the loss to train parameters, and then introducing the random integral theory to construct an improved numerical solver for calculating the solution of the partial differential equation, and finally the network outputs the partial differential equation corresponding to the image denoising data set and the solution of the equation, and then combining the trained network with the denoising post-processing module for image denoising to obtain a denoising result; The image denoising method based on the partial differential equation learning network specifically includes: (1) Obtaining and inputting an image denoising dataset: Obtaining an image denoising dataset and inputting the time series data in the image denoising dataset into the partial differential equation learning network one by one; (2) Construct and train a partial differential equation learning network: The partial differential equation learning network consists of multiple blocks. The image denoising time series data passes through multiple blocks, and each block outputs a predicted denoised image time series. Compared with the real denoised image time series Compare and calculate the loss function Minimize the loss function by gradient descent, complete parameter training once, and perform layered training on multiple blocks; (3) Obtaining output by improving the numerical solver: After training is completed, the parameters are output to obtain the partial differential equation form corresponding to the image denoising data set, as well as the partial differential equation solution of the improved numerical solver by introducing stochastic integral theory; (4) Image denoising using a partial differential equation learning network and a denoising post-processing module: The noisy image to be denoised is input into the trained partial differential equation learning network, then passes through the denoising post-processing module, and finally outputs a denoised clean image; The Block structure of the partial differential equation learning network includes: The first step is to calculate the partial differential terms of the partial differential equation candidate library: It is known that a partial differential equation is an equation involving partial differentials of an unknown function, and its general form is: Where t represents time, x and y represent space, ...is a partial differential term. The Chebyshev polynomial interpolation method is used to find the partial differential term of the Block input data as a candidate library for partial differential equations part of; The second step is to build a candidate library for partial differential equations: The edge features of the image are also added to the candidate library. After comparing multiple edge operators, the edge stop function is used Indicates, where k is a constant; the partial differential items and edge feature items in the candidate library are arranged and combined, and added to the candidate library. The build is complete; The third step is to obtain the output of Block by improving the numerical solver: Candidate Library and network parameters Multiply and calculate Then, random integral theory is introduced to construct a new numerical solution formula, and finally the output of Block is obtained. in The calculation formula is: in N(0,1) is a random number from a standard normal distribution.

2. The image denoising method based on partial differential equation learning network as claimed in claim 1, characterized in that: In step (1), in acquiring and inputting the image denoising dataset, in order to acquire the image denoising dataset, simulated white noise is added to the original clean image dataset to obtain a simulated noise dataset, and then denoising is performed using an existing image denoising method based on partial differential equations to obtain a series of intermediate denoising results and a final denoising result, forming an image denoising dataset with a spatial size of N×M and a time length of T.

3. The image denoising method based on partial differential equation learning network as claimed in claim 1, characterized in that: Warm start, layer-wise training, and loss functions for PDE learning networks include: (1) Network hot start: The parameters of the partial differential equation learning network are initialized by hot start, and the initial values ​​of the parameters are obtained by STRidge regression; (2) Hierarchical training: Input a batch of data into the first block and get the output After that, the loss is calculated and the parameters are trained. After the training is completed, another batch of data is used to restart the training on the first two blocks. This step is repeated until all blocks are completed in the present invention. The parameters are shared between blocks during the whole training process. (3) Loss function: The loss function of the partial differential equation learning network consists of a data loss term and a regularization term. The data loss term uses mean square error, and the regularization term uses L1 regularization:

4. The image denoising method based on partial differential equation learning network as claimed in claim 1, characterized in that: The processing method of the denoising post-processing module includes: The first step is to group similar blocks: divide the image into multiple reference blocks. For each reference block, use the Euclidean distance in the surrounding area to measure the similarity of nearby blocks, select similar blocks with insignificant differences, and finally integrate them to obtain the 3D similar block group corresponding to the reference block. The second step is DCT transformation: the obtained three-dimensional similar block group is transformed into DCT in the third dimension, which can remove the information redundancy brought by the similar block image, improve the efficiency, and distinguish the similar blocks with different noise pollution degrees in the same similar block matrix, so as to distinguish and process them in the subsequent denoising process and improve the denoising effect; The third step is PDE denoising: the PDE learned by the partial differential deep network is used to solve the input similar block group, and the output result is the similar block matrix after denoising; The fourth step is inverse transformation and aggregation operation: perform inverse DCT transformation on the similar block matrix, then put the inverse transformed similar blocks back to their original positions, use the number of non-zero components to statistically superimpose the weights, and finally divide the superimposed image by the weights of the corresponding points to obtain the aggregated denoised image.

5. A computer device, characterized in that: The computer device includes a memory and a processor, the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the image denoising method based on a partial differential equation learning network as described in any one of claims 1 to 4.

6. A computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the processor executes the image denoising method based on a partial differential equation learning network as claimed in any one of claims 1 to 4.

7. An information data processing terminal, characterized in that: The information data processing terminal is used to implement the image denoising method based on partial differential equation learning network as described in any one of claims 1 to 4.

Images