Image denoising model establishment, processing method, system, device and medium

Abstract

The application discloses a kind of establishment, processing method, system, equipment and medium of image denoising model, the method comprises: constructing asynchronous network;Synchronization main network with the network parameters of subnetwork;Get the new debugging parameter of preset image denoising algorithm under current training round;Determine the image texture loss parameter after original noisy image is denoised after current training round;Get the cumulative gradient parameter corresponding to subnetwork under current training round;New main network parameter is obtained using the cumulative gradient parameter corresponding to subnetwork, and the next training round is executed until training round reaches preset iteration round, to train and obtain image denoising model.In the application, by constructing asynchronous policy gradient model, an image noise automatic removal technology is proposed, which effectively improves the precision and efficiency of image noise removal processing;And it can adapt to different debugging environment, thereby effectively saving time and manpower cost in the image noise removal link.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image processing, in particular to an image denoising model establishing method, system, device and medium. BACKGROUND

[0002] Images are often disturbed and affected by various noises in the process of generation and transmission, thereby causing image degradation, which has an adverse effect on subsequent image processing and image visual effect. With the continuous development of cameras, major camera device manufacturers have increasingly high requirements for the quality of photographed images, and image noise removal is an important part of the final imaging quality.

[0003] At present, the effect of image noise removal is mostly debugged by major camera manufacturers relying on manual debugging, which often requires manual modification of parameters, shooting, comparison and other operations, and the process consumes a lot of time and labor cost, and the accuracy is difficult to control. At the same time, due to the differences in hardware, image denoising algorithms and other aspects of various camera device manufacturers, the environment for image noise removal debugging is variable, and different hardware and different denoising algorithms cannot adopt unified debugging measures, thereby further reducing the image denoising debugging efficiency. SUMMARY

[0004] The technical problem to be solved by the present application is to overcome the defects of low image denoising efficiency, insufficient accuracy and restriction by hardware environment factors in the prior art, and to provide an image denoising model establishing method, system, device and medium.

[0005] The present application solves the above technical problems by the following technical solutions:

[0006] The present application provides an image denoising model establishing method, which comprises:

[0007] An asynchronous network is constructed;

[0008] The asynchronous network comprises a main network and a plurality of sub-networks, and the same preset policy gradient model is deployed in the main network and the sub-networks;

[0009] For the first training round, the main network parameters of the main network are used to synchronize the sub-network parameters of the sub-networks;

[0010] The adjustable debugging parameters in the preset image denoising algorithm are input into the preset policy gradient model to obtain new debugging parameters corresponding to the current training round;

[0011] Based on the new debugging parameters, the image texture loss parameters of the original noise image after denoising by the current training round are determined.

[0012] Based on the image texture loss parameter, an accumulated gradient parameter corresponding to the sub-network in the current training round is obtained;

[0013] The accumulated gradient parameter is obtained through the main network, and a new main network parameter is updated using the accumulated gradient parameter corresponding to the sub-network, and the next training round is performed, and the step of inputting the modifiable debugging parameter in the preset image denoising algorithm into the preset policy gradient model is re-executed until the training round reaches a preset iteration round, so as to train the image denoising model.

[0014] Preferably, the preset network model is constructed based on a preset neural network.

[0015] Preferably, the number of neurons of the input layer, the number of neurons of the output layer in the preset network model, and the number of modifiable debugging parameters in the preset image denoising algorithm are equal.

[0016] Preferably, each training round corresponds to a plurality of training moments.

[0017] For each training round, the step of inputting the modifiable debugging parameter in the preset image denoising algorithm into the preset policy gradient model to obtain a new debugging parameter corresponding to the current training round includes:

[0018] The modifiable debugging parameter in the preset image denoising algorithm at the last training moment is obtained and input into the preset policy gradient model to output a new debugging parameter corresponding to the current training moment, which is used as the modifiable debugging parameter corresponding to the next training moment; wherein the first training moment corresponds to the initialized debugging parameter.

[0019] When the current training moment reaches the preset iteration moment corresponding to the training round, the modifiable debugging parameter in the preset image denoising algorithm at the current training moment is used as the new debugging parameter corresponding to the current training round.

[0020] Preferably, the step of determining the image texture loss parameter of the original noise image after denoising by the current training round based on the new debugging parameter includes:

[0021] When the current training moment reaches the preset iteration moment corresponding to the training round, the new debugging parameter corresponding to the current training moment is brought into the preset image denoising algorithm, and the original noise image is denoised using the updated preset image denoising algorithm to obtain a denoised image.

[0022] Based on the denoised image and the original noise image, an image texture loss value corresponding to the original noise image is calculated.

[0023] Preferably, the step of calculating the image texture loss value based on the denoised image and the original noise image comprises:

[0024] calculating a reward value of the image denoising quality at the current training moment based on the image texture loss value;

[0025] wherein the image texture loss value and the reward value are negatively correlated.

[0026] Preferably, when the current training moment reaches the preset iteration moment corresponding to the training round, the step of obtaining the cumulative gradient parameter corresponding to the sub-network at the current training round based on the image texture loss parameter comprises:

[0027] calculating the cumulative gradient parameter corresponding to the sub-network at the current training round based on the first gradient parameter corresponding to the sub-network at the current training moment, the new debugging parameter output by the preset policy gradient model and the reward value, and the cumulative gradient parameter corresponding to the sub-network at the last training round.

[0028] Preferably, when the training round reaches the preset iteration round, the step of training the image denoising model comprises:

[0029] when the cumulative gradient parameter meets the preset condition, it is determined that the model converges, and the trained image denoising model is obtained.

[0030] wherein the preset condition indicates that the cumulative gradient parameters of a continuous set number of training rounds are all tending to the same stable value.

[0031] The present application also provides an image denoising processing method, which comprises:

[0032] obtaining a to-be-denoised image;

[0033] inputting the to-be-denoised image into the image denoising model trained by the image denoising model establishment method to output a target denoised image after denoising processing.

[0034] wherein the input of the target network corresponding to the image denoising model is a target denoising parameter corresponding to a preset image denoising algorithm.

[0035] The present application also provides an image denoising model establishment system, which comprises:

[0036] an asynchronous network construction module for constructing an asynchronous network;

[0037] The asynchronous network comprises a main network and a plurality of sub-networks, and the same preset policy gradient model is deployed in the main network and the sub-networks.

[0038] The parameter synchronization module is configured to synchronize, for a first training round, main network parameters of the main network to sub-network parameters of the sub-networks.

[0039] The debugging parameter updating module is configured to input modifiable debugging parameters in a preset image denoising algorithm to the preset policy gradient model to obtain new debugging parameters corresponding to the current training round.

[0040] The texture loss parameter determination module is configured to determine, based on the new debugging parameters, an image texture loss parameter of the original noisy image after denoising in the current training round.

[0041] The cumulative gradient parameter acquisition module is configured to acquire, based on the image texture loss parameter, a cumulative gradient parameter corresponding to the sub-network in the current training round.

[0042] The main network parameter updating module is configured to acquire the cumulative gradient parameter through the main network, update the new main network parameters by using the cumulative gradient parameter corresponding to the sub-network, and perform a next training round, re-call the debugging parameter updating module, until the training round reaches a preset iteration round, to call the model acquisition module.

[0043] The model acquisition module is configured to acquire the image denoising model obtained through training.

[0044] The application further provides an image denoising processing system, which comprises:

[0045] The image denoising processing system further comprises a to-be-denoised image acquisition module configured to acquire a to-be-denoised image.

[0046] The image denoising processing system further comprises an image denoising processing module configured to input the to-be-denoised image into the image denoising model obtained through training of the image denoising model establishment system to output a target denoised image after denoising processing.

[0047] The target network corresponding to the image denoising model comprises a preset image denoising algorithm corresponding to a target denoising parameter.

[0048] The application further provides an electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the image denoising model establishment method or the image denoising processing method.

[0049] The present invention also provides a computer-readable storage medium having a computer program stored thereon, characterized in that when the computer program is executed by a processor, it implements the above-mentioned method for establishing an image denoising model; or, implements the above-mentioned image denoising processing method.

[0050] On the basis of conforming to the common sense in this field, the above-mentioned preferred conditions can be arbitrarily combined to obtain the preferred embodiments of the present invention.

[0051] The positive progress effect of the present invention is:

[0052] In the present invention, by constructing an asynchronous policy gradient model, an automatic image noise removal technology is proposed. Compared with the existing manual debugging method of removing image noise, the technology is more efficient and more accurate. At the same time, the technology can adapt to different debugging environments, such as different hardware and different image denoising algorithms, thereby saving more time and manpower costs in the image noise removal link, effectively improving the accuracy and efficiency of image noise removal processing, and making debugging unconstrained by the environment, eliminating the impact of the environment, and improving the universality of the image noise removal technology, so that it can meet the needs of higher-demand image denoising processing scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 This is a flowchart of a method for establishing an image denoising model according to Example 1 of the present invention.

[0054] Figure 2 This is a flowchart of a method for establishing an image denoising model according to embodiment 2 of the present invention.

[0055] Figure 3 Schematic diagram of image denoising model training according to Example 2 of the present invention.

[0056] Figure 4 Schematic diagram of an asynchronous network in the image denoising model of Example 2 of the present invention.

[0057] Figure 5 Schematic diagram of the BP neural network in the image denoising model of Example 2 of the present invention.

[0058] Figure 6 This is a flowchart of an image denoising method according to embodiment 3 of the present invention.

[0059] Figure 7 This is a flowchart of a system for establishing an image denoising model according to Example 4 of the present invention.

[0060] Figure 8 This is a flowchart of the image denoising processing system according to embodiment 6 of the present invention.

[0061] Figure 9 This is a structural diagram of an electronic device according to embodiment 7 of the present invention. DETAILED DESCRIPTION

[0062] The present application is further illustrated by the following examples without limiting the present application to the examples.

[0063] Example 1

[0064] As shown in the figure, the method for establishing the image denoising model of the present embodiment comprises: Figure 1

[0065] S101, constructing an asynchronous network;

[0066] The asynchronous network comprises a main network and a plurality of sub-networks, and the same preset policy gradient model is deployed in the main network and the sub-networks.

[0067] Specifically, the asynchronous network comprises one main network and N sub-networks, where N is an integer greater than or equal to 1.

[0068] In an implementable scheme, the preset policy gradient model is constructed based on a preset neural network.

[0069] The preset neural network includes but is not limited to a BP (a kind of multilayer feedforward network trained according to error back propagation algorithm) neural network.

[0070] The number of neurons in the input layer, the number of neurons in the output layer of the preset network model, and the number of adjustable debugging parameters in the preset image denoising algorithm are equal.

[0071] Of course, the preset policy gradient model or the preset network model can also be constructed by using other network models as long as the actual training requirements of the image denoising model can be met.

[0072] S102, for the first training round, synchronizing the sub-network parameters of the sub-networks with the main network parameters of the main network;

[0073] Specifically, the asynchronous network belongs to an asynchronous multi-thread network, and at the beginning of training, the main network parameters of the main network are initialized and synchronized to the sub-network parameters of the sub-networks, so that the network parameters of the main network and the sub-networks remain consistent.

[0074] Before the synchronization operation, the network parameters of the sub-network parameters of each sub-network are generated in a random manner, which can effectively reduce the correlation between samples, thereby improving the anti-interference ability of the model, effectively improving the accuracy of model training, and further ensuring the precision of image denoising processing.

[0075] S103, inputting the adjustable debugging parameters in the preset image denoising algorithm to the preset policy gradient model to obtain new debugging parameters corresponding to the current training round.​

[0076] S104, determine the image texture loss parameter of the original noise image after denoising by the current training round based on the new debugging parameter;

[0077] S105, obtain the cumulative gradient parameter corresponding to the sub-network under the current training round based on the image texture loss parameter;

[0078] S106, obtain the cumulative gradient parameter through the main network, update the new main network parameter by using the cumulative gradient parameter corresponding to the sub-network, and execute the next training round to re-execute step S103;

[0079] S107, when the training round reaches the preset iteration round, obtain the image denoising model by training.

[0080] In the embodiment, by constructing a preset policy gradient model, the debugging parameters of the image denoising algorithm are adjusted by the computer according to the change of the debugging environment, and the manual debugging is no longer needed, thereby effectively improving the efficiency and precision of the image noise removal debugging process.

[0081] In addition, only the main network parameter of the main network needs to be updated, and the main network parameter of the main network is updated according to the cumulative value of the gradient in the sub-network, without other redundant calculations, thereby effectively reducing the calculation amount, and ensuring the efficiency and accuracy of the model training.

[0082] It should be noted that in the asynchronous multi-thread network structure corresponding to the asynchronous network of the embodiment, each sub-network corresponds to a sub-thread for control, which can effectively utilize computer resources and achieve the effect of accelerating the convergence of the model, thereby improving the efficiency of the model training.

[0083] In the embodiment, by constructing an asynchronous policy gradient model, an image noise automatic removal technology is proposed, which is more efficient and more accurate than the existing manual debugging image noise removal method. At the same time, the technology can adapt to different debugging environments, such as different hardware and different image denoising algorithms, thereby saving more time and labor cost in the image noise removal link, effectively improving the precision and efficiency of the image noise removal processing, and making the debugging not restricted by the environment, eliminating the influence caused by the environment, and improving the universality of the image noise removal technology, thereby meeting the higher demand of image denoising processing scene.

[0084] Embodiment 2

[0085] As shown in Figure 2 the establishment method of the image denoising model of the embodiment is a further improvement of embodiment 1, specifically:

[0086] In an implementable scheme, each training round corresponds to a plurality of training time points;

[0087] For the first training round, step S103 comprises:

[0088] obtaining the modifiable debugging parameter in the preset image denoising algorithm at the last training moment and inputting it into the preset policy gradient model to output the corresponding new debugging parameter at the current training moment as the modifiable debugging parameter corresponding to the next training moment; wherein the first training moment corresponds to the initialized debugging parameter;

[0089] When the current training moment reaches the preset iteration moment corresponding to the training round, the modifiable debugging parameter in the preset image denoising algorithm at the current training moment is taken as the new debugging parameter corresponding to the current training round.

[0090] In an implementable scheme, step S104 comprises:

[0091] When the current training moment reaches the preset iteration moment corresponding to the training round, the new debugging parameter corresponding to the current training moment is brought into the preset image denoising algorithm, and the original noisy image is denoised using the updated preset image denoising algorithm to obtain a denoised image;

[0092] Based on the denoised image and the original noisy image, an image texture loss value corresponding to the original noisy image is calculated.

[0093] Based on the image texture loss value, a reward value of the image denoising quality at the current training moment is calculated;

[0094] Wherein, the image texture loss value and the reward value are negatively correlated.

[0095] In an implementable scheme, when the current training moment reaches the preset iteration moment corresponding to the training round, step S105 comprises:

[0096] S1051, based on the first gradient parameter corresponding to the subnetwork at the current training moment, the new debugging parameter and the reward value output by the preset policy gradient model, and based on the cumulative gradient parameter corresponding to the subnetwork at the last training round, the cumulative gradient parameter corresponding to the subnetwork at the current training round is calculated.

[0097] In an implementable scheme, step S107 comprises:

[0098] S1071, when the cumulative gradient parameter meets the preset condition, it is determined that the model converges, and the trained image denoising model is obtained;

[0099] Wherein, the preset condition represents that the cumulative gradient parameters of a continuous number of training rounds are all tending to the same stable value.

[0100] The implementation principle of the method for establishing the image denoising model of the embodiment will be specifically explained below with examples:

[0101] Step 1, preparing an initial environment;

[0102] (11) obtaining an original noise image with noise (as shown in Figure 3 );

[0103] (12) providing an interface for modifying the algorithm parameters (or debugging parameters) of the image denoising algorithm;

[0104] (13) taking the debugging parameters that can be modified in the image denoising algorithm as environment variables S;

[0105] Step 2, constructing an asynchronous network (or an asynchronous policy gradient model);

[0106] (21) constructing a policy gradient model;

[0107] (211) constructing a BP neural network to fit the policy function;

[0108] (212) calculating the image texture loss and setting it as the reward value R every time a new parameter is generated;

[0109] As shown in Figure 4 , the BP neural network structure in step (211) is divided into four layers: an input layer, a hidden layer 1, a hidden layer 2, and an output layer. The number of input layer neurons is equal to the number N of parameters that can be modified in the denoising algorithm, the hidden layer 1 contains 64 neurons, the hidden layer 2 contains 32 neurons, and the number of output layer neurons is equal to the number of input layer neurons.

[0110] (22) constructing a 1-N asynchronous network to obtain an asynchronous policy gradient model

[0111] (221) deploying a policy gradient model created above in the main thread as the main network;

[0112] Starting N sub-threads, each of which deploys the same preset policy gradient model as the main thread as a sub-network, for details, see Figure 5 , each sub-network corresponds to the debugging parameters of a different debugging environment;

[0113] Step 3, inputting the original denoising image into the asynchronous policy gradient model for training

[0114] (31) initializing the main network parameters θ of the policy gradient model in the main network, the network parameters θ' of the policy gradient model in the sub-network, the global shared iteration number T, the global maximum iteration number T max , the maximum length t of a single iteration sequence within a thread local , and the initial time tstart , step size a, decay factor g, and the like;

[0115] (32) Synchronize the subnetwork parameters of the preset policy gradient model in the subnetwork with the main network parameters of the main network, i.e., θ' = θ;

[0116] (33) Reset the cumulative gradient update amount dθ of the subnetwork to 0, and the initial cumulative gradient of the subnetwork is 0;

[0117] (34) Input the debugging parameters S of the image denoising algorithm into the BP neural network for forward propagation to obtain new debugging parameters S';

[0118] (35) Bring the new debugging parameters S' into the denoising algorithm and perform denoising on the original image to obtain the denoised image img t at time t, and obtain the reward value R t at time t by calculating the image texture loss of img t ;

[0119] (36) Enter the next training time t <- t + 1; if t - t start = t local , enter step (37), otherwise return to step (34);

[0120] Alternatively, before step (36), if t - t start ≠ t local , execute step (36); otherwise, execute step (37);

[0121] (37) for i∈(t-1,t-2,…,t start )

[0122] Execute

[0123] wherein t represents the training time t, dθ corresponds to the cumulative gradient of the subnetwork, R t corresponds to the reward value at time t, and S t corresponds to the debugging parameters at time t.

[0124] (38) Update the main network parameters of the main network: θ = θ - a dθ;

[0125] (39) Enter the next training round T <- T + 1;

[0126] (310) If T = T max , the training is completed, and the public network parameters θ are output, i.e., the best denoising parameters, which are applied to the image denoising algorithm to perform denoising processing on the input image, otherwise, step (32) is executed again;

[0127] Step 4, the best denoising image is obtained after the model converges or reaches the iteration threshold value;

[0128] (41) the best denoising image is obtained after the model converges or reaches the maximum iteration threshold value, specifically by determining that the cumulative gradient dθ tends to a value stably, then the model converges, T=T max Then it represents that the iteration threshold value is reached, all training rounds are completed, and the image denoising model is trained.

[0129] In this embodiment, by constructing an asynchronous policy gradient model, an image noise automatic removal technology is proposed, which is more efficient and more accurate than the existing manual debugging image noise removal method. At the same time, this technology can adapt to different debugging environments, such as different hardware and different image denoising algorithms, thereby saving more time and labor costs in the image noise removal link, effectively improving the precision and efficiency of image noise removal processing, and making the debugging not restricted by the environment, eliminating the influence caused by the environment, improving the universality of the image noise removal technology, so as to meet the higher image denoising processing scene demand.

[0130] Embodiment 3

[0131] The image denoising processing method of this embodiment is realized based on the image denoising model establishment method in Embodiment 1 or Embodiment 2.

[0132] As shown in Figure 6 , the image denoising processing method of this embodiment includes:

[0133] S201, obtaining a to-be-denoised image;

[0134] S202, inputting the to-be-denoised image into the image denoising model to output a target denoised image after denoising processing.

[0135] The input of the target network corresponding to the image denoising model is the target denoising parameter corresponding to the preset image denoising algorithm.

[0136] In this embodiment, by constructing an asynchronous policy gradient model, an image noise automatic removal technology is proposed, which is more efficient and more accurate than the existing manual debugging image noise removal method. At the same time, this technology can adapt to different debugging environments, such as different hardware and different image denoising algorithms, thereby saving more time and labor costs in the image noise removal link, effectively improving the precision and efficiency of image noise removal processing, and making the debugging not restricted by the environment, eliminating the influence caused by the environment, improving the universality of the image noise removal technology, so as to meet the higher image denoising processing scene demand.

[0137] Embodiment 4

[0138] AsFigure 7 The image denoising model establishment system of the embodiment includes:

[0139] An asynchronous network construction module 1 is configured to construct an asynchronous network.

[0140] The asynchronous network includes a main network and a plurality of sub-networks, and the same preset policy gradient model is deployed in the main network and the sub-networks.

[0141] A parameter synchronization module 2 is configured to synchronize the main network parameters of the main network to the sub-network parameters of the sub-networks for a first training round.

[0142] A debugging parameter updating module 3 is configured to input the modifiable debugging parameters in the preset image denoising algorithm to the preset policy gradient model to obtain new debugging parameters corresponding to the current training round.

[0143] A texture loss parameter determination module 4 is configured to determine the image texture loss parameters of the original noise image after denoising by the current training round based on the new debugging parameters.

[0144] An accumulated gradient parameter acquisition module 5 is configured to acquire the accumulated gradient parameters of the sub-networks in the current training round based on the image texture loss parameters.

[0145] A main network parameter updating module 6 is configured to acquire the accumulated gradient parameters by the main network, update the new main network parameters by using the accumulated gradient parameters of the sub-networks, and perform the next training round, re-call the debugging parameter updating module 3, until the training round reaches the preset iteration round, to call the model acquisition module 7.

[0146] The model acquisition module 7 is configured to acquire the image denoising model obtained by training.

[0147] In the embodiment, the image noise automatic removal technology is proposed by constructing the asynchronous policy gradient model, which is more efficient and more accurate than the existing manual debugging image noise removal method. Meanwhile, the technology can adapt to different debugging environments, such as different hardware and different image denoising algorithms, thereby saving more time and labor cost in the image noise removal link, effectively improving the accuracy and efficiency of the image noise removal processing, and making the debugging not restricted by the environment, eliminating the influence caused by the environment, improving the universality of the image noise removal technology, thereby meeting the higher image denoising processing scene demand.

[0148] Embodiment 5

[0149] The image denoising model establishment system of the embodiment is a further improvement of the embodiment 4, specifically:

[0150] In an implementable solution, the preset network model is constructed based on a preset neural network.

[0151] The preset neural network includes but is not limited to a BP neural network.

[0152] The number of neurons in the input layer, the number of neurons in the output layer in the preset network model, and the number of adjustable debugging parameters in the preset image denoising algorithm are equal.

[0153] In an implementable solution, each training round corresponds to a plurality of training time points.

[0154] For each training round, the debugging parameter updating module 3 is configured to obtain the adjustable debugging parameters in the preset image denoising algorithm at the last training time point and input the same into the preset policy gradient model to output the corresponding new debugging parameters at the current training time point, which are used as the adjustable debugging parameters corresponding to the next training time point. The debugging parameters at the first training time point are initialized.

[0155] The debugging parameter updating module 3 is further configured to, when the current training time point reaches the preset iteration time point corresponding to the training round, take the adjustable debugging parameters in the preset image denoising algorithm at the current training time point as the new debugging parameters corresponding to the current training round.

[0156] In an implementable solution, the texture loss parameter determination module 4 includes:

[0157] The denoising processing image obtaining unit is configured to, when the current training time point reaches the preset iteration time point corresponding to the training round, bring the new debugging parameters corresponding to the current training time point into the preset image denoising algorithm, and perform denoising processing on the original noise image using the updated preset image denoising algorithm to obtain a denoising processing image.

[0158] The texture loss value calculating unit is configured to calculate an image texture loss value corresponding to the original noise image based on the denoising processing image and the original noise image.

[0159] The reward value obtaining unit is configured to calculate a reward value of the image denoising quality at the current training time point based on the image texture loss value.

[0160] The image texture loss value and the reward value are negatively correlated.

[0161] In an implementable solution, when the current training time point reaches the preset iteration time point corresponding to the training round, the cumulative gradient parameter obtaining module 5 is configured to calculate the cumulative gradient parameter corresponding to the sub-network at the current training round based on the first gradient parameter corresponding to the sub-network at the current training time point, the new debugging parameters and the reward value output by the preset policy gradient model, and the cumulative gradient parameter corresponding to the sub-network at the last training round.

[0162] In an implementable solution, when the training round reaches the preset iteration round, the model obtaining module 7 is configured to determine that the model converges and obtain the trained image denoising model when the accumulated gradient parameters of the preset number of continuous training rounds all tend to the same stable value.

[0163] The preset condition indicates that the accumulated gradient parameters of the preset number of continuous training rounds all tend to the same stable value.

[0164] In this embodiment, by constructing an asynchronous policy gradient model, an image noise automatic removal technology is proposed. Compared with the existing manual debugging image noise removal method, the image noise automatic removal technology is more efficient and more accurate. Meanwhile, the image noise automatic removal technology can adapt to different debugging environments, such as different hardware and different image denoising algorithms. Thus, more time and labor costs are saved in the image noise removal link, the accuracy and efficiency of the image noise removal processing are effectively improved, the debugging is not restricted by the environment, the influence caused by the environment is eliminated, the universality of the image noise removal technology is improved, and thus the image noise removal processing scene demand of higher requirement can be met.

[0165] Embodiment 6

[0166] The image denoising processing system in this embodiment is implemented based on the image denoising model establishment system in the embodiment 4 or the embodiment 5.

[0167] As shown in Figure 8 , the image denoising processing system in this embodiment includes:

[0168] The to-be-denoised image obtaining module 8 is configured to obtain a to-be-denoised image.

[0169] The image denoising processing module 9 is configured to input the to-be-denoised image into the image denoising model to output a target denoised image after denoising processing.

[0170] The input of the target network corresponding to the image denoising model is a target denoising parameter corresponding to a preset image denoising algorithm.

[0171] In this embodiment, by constructing an asynchronous policy gradient model, an image noise automatic removal technology is proposed. Compared with the existing manual debugging image noise removal method, the image noise automatic removal technology is more efficient and more accurate. Meanwhile, the image noise automatic removal technology can adapt to different debugging environments, such as different hardware and different image denoising algorithms. Thus, more time and labor costs are saved in the image noise removal link, the accuracy and efficiency of the image noise removal processing are effectively improved, the debugging is not restricted by the environment, the influence caused by the environment is eliminated, the universality of the image noise removal technology is improved, and thus the image noise removal processing scene demand of higher requirement can be met.

[0172] Embodiment 7

[0173] Figure 9 Figure 7 shows a schematic block diagram of an electronic device according to an embodiment of the application. The electronic device comprises a memory, a processor and a computer program stored on the memory and loadable into the processor, the processor implementing the method of the above embodiments when executing the program. Figure 9 The electronic device 30 shown is merely an example and should not be taken as limiting the functionality or the applicability of embodiments of the application.

[0174] As shown in Figure 9 Figure 7, the electronic device 30 can take the form of a general purpose computing device, for example it can be a server device. Components of the electronic device 30 can include, but are not limited to, the at least one processor 31 described above, the at least one memory 32 described above, a bus 33 that couples various system components including the memory 32 and the processor 31.

[0175] The bus 33 can include an address bus, a data bus, and a control bus, although not all buses of these types are required.

[0176] The memory 32 can include volatile memory, such as random access memory (RAM) 321 and / or cache memory 322, and non-volatile memory, such as read only memory (ROM) 323.

[0177] The memory 32 can include program / utility 325 having one or more sets of programs / modules 324, which can include an operating system, one or more applications, other program modules, and program data, and each of the above examples, or some combination thereof, can include implementations of a network environment.

[0178] The processor 31 can execute the various embodiments of the methods of the application by running the computer program from the memory 32 as described above.

[0179] The electronic device 30 can also communicate with one or more external devices 34 such as a keyboard or a pointing device, for example, through an input / output (I / O) interface 35. Further, the electronic device 30 can communicate with one or more networks, such as a local area network (LAN), a wide area network (WAN), and / or a public network, such as the Internet, through a network adapter 36. As Figure 9 illustrated, the network adapter 36 can communicate with the other components of the electronic device 30 through the bus 33. It should be understood that although not shown, other hardware and / or software components could be used in conjunction with the electronic device 30. These include, but are not limited to, microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data archival storage systems, etc.

[0180] It should be noted that although several units / modules or sub-units / modules of the electronic device are mentioned in the above detailed description, such division is merely exemplary and not mandatory. Indeed, according to an embodiment of the application, the features and functionalities of two or more units / modules described above can be embodied in one unit / module. Conversely, the features and functionalities of one unit / module described above can be further divided into units / modules embodied by several units / modules.

[0181] Embodiment 8

[0182] The embodiment provides a computer readable storage medium, and a computer program is stored on the computer readable storage medium. The computer program is executed by a processor to implement the steps in the method of the above embodiment.

[0183] More specifically, the readable storage medium can include, but is not limited to, a portable disc, a hard disk, a random access memory, a read-only memory, an erasable programmable read-only memory, an optical storage device, a magnetic storage device, or any suitable combination of the above.

[0184] In a possible implementation, the application can also be implemented in the form of a program product, which includes program codes for causing a terminal device to execute the steps in the method of the above embodiment when the program product is run on the terminal device.

[0185] The program codes for executing the application can be written in any combination of one or more programming languages, and the program codes can be executed completely on the user device, partially on the user device, as a separate software package, partially on the user device and partially on a remote device, or completely on a remote device.

[0186] Although the specific embodiments of the application are described above, those skilled in the art should understand that this is only an illustration, and the protection scope of the application is defined by the appended claims. Those skilled in the art can make various changes or modifications to these embodiments without departing from the principles and essence of the application, and such changes and modifications fall within the protection scope of the application.

Claims

1. A method for establishing an image denoising model, characterized in that, The establishing method comprises: constructing an asynchronous network; wherein the asynchronous network comprises a main network and a plurality of sub-networks, and the same preset policy gradient model is deployed in the main network and the sub-networks; for a first training round, synchronizing sub-network parameters of the sub-networks with main network parameters of the main network; inputting modifiable debugging parameters in a preset image denoising algorithm into the preset policy gradient model to obtain new debugging parameters corresponding to the current training round; each training round corresponds to a plurality of training instants; for each training round, the step of inputting modifiable debugging parameters in a preset image denoising algorithm into the preset policy gradient model to obtain new debugging parameters corresponding to the current training round comprises: obtaining the modifiable debugging parameters in the preset image denoising algorithm at the last training instant and inputting them into the preset policy gradient model to output the new debugging parameters corresponding to the current training instant, which are used as the modifiable debugging parameters corresponding to the next training instant; wherein the first training instant corresponds to the initialized debugging parameters; when the current training instant reaches the preset iteration instant corresponding to the training round, the modifiable debugging parameters in the preset image denoising algorithm at the current training instant are used as the new debugging parameters corresponding to the current training round; based on the new debugging parameters, determining an image texture loss parameter of the original noise image after denoising by the current training round; based on the image texture loss parameter, obtaining the cumulative gradient parameter corresponding to the sub-network at the current training round; obtaining the cumulative gradient parameter through the main network, updating the new main network parameters by using the cumulative gradient parameter corresponding to the sub-network, and performing the next training round to re-perform the step of inputting the modifiable debugging parameters in the preset image denoising algorithm into the preset policy gradient model, until the training round reaches the preset iteration round, so as to train the image denoising model.

2. The method of claim 1, wherein the image denoising model is established by: The preset network model is constructed based on a preset neural network; the preset network model is the preset policy gradient model; wherein the number of neurons of the input layer, the number of neurons of the output layer in the preset network model, and the number of modifiable debugging parameters in the preset image denoising algorithm are equal.

3. The method of claim 1, wherein the image denoising model is established by using a neural network. The step of determining the image texture loss parameter of the original noise image after denoising by the current training round based on the new debugging parameters comprises: when the current training instant reaches the preset iteration instant corresponding to the training round, the new debugging parameters corresponding to the current training instant are brought into the preset image denoising algorithm, and the original noise image is denoised by using the updated preset image denoising algorithm to obtain a denoising processing image; based on the denoising processing image and the original noise image, the image texture loss value corresponding to the original noise image is calculated.

4. The method of claim 3, wherein the image denoising model is established by using a neural network. The step of calculating the image texture loss value based on the denoising processing image and the original noise image comprises: calculating a reward value of the image denoising quality at the current training instant based on the image texture loss value; The image texture loss value is negatively correlated with the reward value.

5. The method for establishing an image denoising model according to claim 4, wherein: When the current training time reaches a preset iteration time corresponding to the training round, the step of obtaining the accumulated gradient parameter corresponding to the sub-network in the current training round based on the image texture loss parameter comprises: Based on the first gradient parameter corresponding to the sub-network in the current training time, the new debugging parameter output by the preset policy gradient model and the reward value, and based on the accumulated gradient parameter corresponding to the sub-network in the last training round, the accumulated gradient parameter corresponding to the sub-network in the current training round is calculated.

6. The method of claim 1, wherein the image denoising model is established by: When the training round reaches a preset iteration round, the step of training the image denoising model comprises: When the accumulated gradient parameter meets a preset condition, it is determined that the model converges, and the image denoising model obtained by training is obtained. The preset condition indicates that the accumulated gradient parameters of a continuous set number of training rounds are all in the same stable value.

7. An image denoising processing method, characterized by, The image denoising processing method comprises: Obtaining a to-be-denoised image; Inputting the to-be-denoised image into the image denoising model obtained by training the image denoising model establishment method of any one of claims 1-6 to output a target denoised image after denoising processing. The input of the target network corresponding to the image denoising model is a target denoising parameter corresponding to a preset image denoising algorithm.

8. A system for establishing an image denoising model, characterized by, The establishment system comprises: An asynchronous network construction module for constructing an asynchronous network; The asynchronous network comprises a main network and a plurality of sub-networks, and the same preset policy gradient model is deployed in the main network and the sub-networks; A parameter synchronization module for synchronizing the sub-network parameters of the sub-networks with the main network parameters of the main network in the first training round; A debugging parameter updating module for inputting the modifiable debugging parameters in the preset image denoising algorithm into the preset policy gradient model to obtain new debugging parameters corresponding to the current training round; Each training round corresponds to a plurality of training times; For each training round, the debugging parameter updating module is further configured to obtain the modifiable debugging parameters in the preset image denoising algorithm in the last training time and input them into the preset policy gradient model to output the new debugging parameters corresponding to the current training time as the modifiable debugging parameters corresponding to the next training time; wherein the first training time corresponds to the initialized debugging parameters; The debugging parameter updating module is further configured to, when the current training time reaches the preset iteration time corresponding to the training round, take the modifiable debugging parameters in the preset image denoising algorithm in the current training time as the new debugging parameters corresponding to the current training round; A texture loss parameter determination module for determining the image texture loss parameter after the original noise image is denoised in the current training round based on the new debugging parameters; An accumulated gradient parameter obtaining module for obtaining the accumulated gradient parameter corresponding to the sub-network in the current training round based on the image texture loss parameter. The main network parameter updating module is configured to obtain the accumulated gradient parameter from the main network, update the main network parameter using the accumulated gradient parameter corresponding to the sub-network, perform a next training round, and re-call the debugging parameter updating module until the training round reaches a preset iteration round, so as to call the model obtaining module. The model obtaining module is configured to obtain the image denoising model obtained through training.

9. An image denoising processing system, characterized by comprising: The image denoising processing system comprises: An image to be denoised obtaining module configured to obtain an image to be denoised; An image denoising processing module configured to input the image to be denoised into the image denoising model established by the image denoising model establishing system of claim 8 to output a target denoised image after denoising processing. The input of the target network corresponding to the image denoising model is a target denoising parameter corresponding to a preset image denoising algorithm.

10. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the computer program to implement the image denoising model establishing method of any one of claims 1-6; or, implement the image denoising processing method of claim 7.

11. A computer readable storage medium having stored thereon a computer program, characterized in that The computer program is executed by the processor to implement the image denoising model establishing method of any one of claims 1-6; or, implement the image denoising processing method of claim 7.

Images