A training method and system of an image denoising model

Abstract

The application discloses a kind of training method and system of image denoising model, it is related to digital image processing and computer vision field, method includes: first, noisy image data and real world image data are input to student network, obtain noisy image denoising data and first real world image denoising data;Real world image data is input to teacher network, and second real world image denoising data is obtained;According to the data obtained, determine the noise reduction loss of student network;Then with the minimum denoising loss as the goal, the weight parameter of student network is updated;According to the weight parameter of student network after updating, the weight parameter of teacher network is updated.The training method based on the application, when training teacher-student network, only needs to use noisy image data and real world image data, can train image denoising model, and does not need to obtain high-quality clean image pair.

Description

Technical Field

[0001] This invention relates to the fields of digital image processing and computer vision, and in particular to a training method and system for an image denoising model. Background Technology

[0002] Image denoising is currently a crucial task in image processing and computer vision. Deep learning-based image denoising methods are emerging, but these methods may struggle to generalize well to noisy real-world scenes. To overcome these issues, researchers have attempted to design supervised models trained on real-world noisy-clean image pairs for denoising.

[0003] However, obtaining high-quality noise-clean image pairs requires shooting under stringent conditions. Furthermore, the clean images obtained are not always usable, and even clean images in real-world datasets may be unreliable. Summary of the Invention

[0004] The purpose of this invention is to provide a training method and system for an image denoising model that does not rely on clean real-world images.

[0005] This invention provides a method for training an image denoising model, comprising:

[0006] Step 1: Input the noisy image data and real-world image data into the student network to obtain the noisy image denoising data and the first real-world image denoising data; the noisy image data is the image data after adding pseudo-noise to the sample image; the student network is a denoising network based on the UDNet network framework;

[0007] Step 2: Input the real-world image data into the teacher network to obtain the second real-world image denoising data; the teacher network is a denoising network based on the UDNet network framework;

[0008] Step 3: Determine the denoising loss of the student network based on the noisy image denoising data, the first real-world image denoising data, and the second real-world image denoising data; the denoising loss includes denoising loss and consistency loss; the denoising loss is determined based on the sample image data and the noisy image denoising data; the consistency loss is determined based on the first real-world image denoising data and the second real-world image denoising data.

[0009] Step 4: Update the weight parameters of the student network with the goal of minimizing the noise reduction loss;

[0010] Step 5: Update the weight parameters of the teacher network according to the updated weight parameters of the student network.

[0011] Optionally, prior to step 1, the following steps are also included:

[0012] Obtain an image dataset; the image dataset includes the real-world image data and the sample image data;

[0013] The pseudo-noise data is added to the sample image data to obtain the noisy image data.

[0014] Optionally, the student network includes: a first encoder module, a first decoder module, and a first context connection module;

[0015] The first encoder module is used to progressively reduce the size of the image input to the first encoder module in layers to obtain a first feature map;

[0016] The first decoder module is used to obtain a denoised image based on the first feature map;

[0017] The first context connection module is connected to the first encoder module and the first decoder module, and is used for the fusion of information in the first encoder module and the first decoder module.

[0018] Optionally, the teacher network includes: a second encoder module, a second decoder module, and a second context connection module;

[0019] The second encoder module is used to progressively reduce the size of the image input to the second encoder module in layers to obtain a second feature map;

[0020] The second decoder module is used to obtain a denoised image based on the second feature map;

[0021] The second context connection module is connected to the second encoder module and the second decoder module, and is used for the fusion of information in the second encoder module and the second decoder module.

[0022] Optionally, the loss function for the noise reduction loss is as follows:

[0023]

[0024] Among them, L total For noise reduction loss, A and R represent the sets of noisy image data and real-world image data, respectively, N and M represent the number of noisy image data and real-world image data in each iteration, respectively, and L s (x i Let L be the denoising loss during the i-th iteration. c (y j Let x be the consistency loss number at the i-th iteration. iThe sample image data at the i-th iteration, y j For the i-th iteration, λ is a coefficient that varies with the number of iterations.

[0025] Optionally, the loss function expression for the denoising loss is as follows:

[0026] L s (x)=||xP x ||1

[0027] The loss function expression for the consistency loss is as follows:

[0028] L c (y)=||S y -T y ||1

[0029] Among them, L s (x) represents the denoising loss, x represents the sample image data, and P x For denoising data of noisy images; ||·||1 is the L1 loss function, L c (y) represents the consistency loss, S y T is the first real-world image denoising data generated for student network predictions. y Denoising data for the second real-world image generated by the teacher network prediction.

[0030] Optionally, the calculation formula for the coefficient λ that varies with the number of iterations is as follows:

[0031]

[0032] Where t is the current iteration number, t max λ is the maximum number of iterations. max This represents the maximum value of the coefficient λ, which varies with the number of iterations.

[0033] Optionally, step 3 is as follows:

[0034] θ' t =ηθ' t-1 +(1-η)θ t

[0035] Where, θ t Let θ' be the parameter of the student network during the t-th iteration of training. t Let θ' be the parameter of the teacher network during the t-th iteration of training. t-1 Let η be the parameters of the teacher network during the (t-1)th iteration of training, and η be the smoothing exponent.

[0036] Optionally, the UDNet network framework includes a multi-scale convergent attention module (MSAA) and a multi-scale linear cross attention module (MLCA).

[0037] The present invention also provides a training system for an image denoising model, comprising:

[0038] The student network denoising module is used to input noisy image data and real-world image data into the student network to obtain noisy image denoised data and first real-world image denoised data; the noisy image data is the image data after adding pseudo-noise to the sample image; the student network is a denoising network based on the UDNet network framework.

[0039] The teacher network denoising module is used to input the real-world image data into the teacher network to obtain the second real-world image denoised data; the teacher network is a denoising network based on the UDNet network framework;

[0040] The loss calculation module is used to determine the denoising loss of the student network based on the noisy image denoising data, the first real-world image denoising data, and the second real-world image denoising data; the denoising loss includes denoising loss and consistency loss; the denoising loss is determined based on the sample image data and the noisy image denoising data; the consistency loss is determined based on the first real-world image denoising data and the second real-world image denoising data.

[0041] The student network parameter update module is used to update the weight parameters of the student network with the goal of minimizing the noise reduction loss.

[0042] The teacher network parameter update module is used to update the weight parameters of the teacher network based on the updated weight parameters of the student network.

[0043] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0044] This invention provides a training method and system for an image denoising model, comprising: inputting noisy image data and real-world image data into a student network to obtain noisy image denoising data and first real-world image denoising data; the noisy image data is image data after adding pseudo-noise to sample images; the student network is a denoising network based on the UDNet network framework; inputting real-world image data into a teacher network to obtain second real-world image denoising data; the teacher network is a denoising network based on the UDNet network framework; determining the denoising loss of the student network based on the noisy image denoising data, the first real-world image denoising data, and the second real-world image denoising data; the denoising loss includes denoising loss and consistency loss; the denoising loss is determined based on the sample image data and the noisy image denoising data; the consistency loss is determined based on the first real-world image denoising data and the second real-world image denoising data; updating the weight parameters of the student network with the goal of minimizing the denoising loss; and updating the weight parameters of the teacher network based on the updated weight parameters of the student network. This invention combines a teacher-student training network with a denoising UDNet to train a student network with denoising capabilities. Furthermore, during model training, the proposed training method does not require high-quality clean image pairs. It only needs to use noisy image data with pseudo-noise added to sample images and real-world image data to train the image denoising model, reducing the training conditions for the image denoising model and improving the reliability of the training data. Attached Figure Description

[0045] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0046] Figure 1 This is a flowchart illustrating the training process of the image denoising model provided in an embodiment of the present invention.

[0047] Figure 2 This is a framework diagram of the MTUDet model provided in an embodiment of the present invention;

[0048] Figure 3 This is a diagram of the UDNet network framework for a noise reducer provided in an embodiment of the present invention.

[0049] Figure 4 A schematic diagram of the Convolutional Block in the UDNet noise denoiser provided in this embodiment of the invention.

[0050] Figure 5This is a schematic diagram of the structure of the Multi-Scale Aggregation Attention Module (MSAA) provided in an embodiment of the present invention;

[0051] Figure 6 This is a schematic diagram of the structure of the Multi-Scale Linear Cross-Attention Module (MLCA) provided in an embodiment of the present invention. Detailed Implementation

[0052] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0053] The purpose of this invention is to provide a training method and system for an image denoising model that does not rely on clean real-world images.

[0054] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0055] like Figure 1 As shown, the present invention provides an unsupervised training method for an image denoising model, comprising:

[0056] Step 1: Input the noisy image data and real-world image data into the student network to obtain the noisy image denoising data and the first real-world image denoising data; the noisy image data is the image data after adding pseudo-noise to the sample image; the student network is a denoising network based on the UDNet network framework.

[0057] Step 2: Input the real-world image data into the teacher network to obtain the second real-world image denoising data; the teacher network is a denoising network based on the UDNet network framework.

[0058] Step 3: Determine the denoising loss of the student network based on the denoised noisy image data, the first real-world image denoising data, and the second real-world image denoising data; the denoising loss includes denoising loss and consistency loss; the denoising loss is determined based on the denoised noisy image data and the denoised noisy image data; the consistency loss is determined based on the first real-world image denoising data and the second real-world image denoising data.

[0059] Step 4: Update the weight parameters of the student network with the goal of minimizing the noise reduction loss.

[0060] Step 5: Update the weight parameters of the teacher network according to the updated weight parameters of the student network.

[0061] In some embodiments, prior to step 1, the method further includes:

[0062] Obtain an image dataset; the image dataset includes real-world image data and sample image data; the sample image data is arbitrary image data without spurious noise.

[0063] The pseudo-noise data is added to the sample image data to obtain the noisy image data.

[0064] Specifically, during model training, 78 images from the BSD500 image dataset and 1400 images from the MIT-Adobe FiveK image dataset were used as sample images, while images from the SIDD Medium dataset were used as real-world image data. The validation dataset from SIDD was used for model validation, and then this data was imported into the student network.

[0065] Specifically, adding pseudo-noise data to the sample image data to obtain the noisy image data can be done as follows:

[0066] A pseudo-noise generation unit is used to simulate pseudo-noise for the sample data, resulting in a pseudo-noise dataset. The noise image data in this paper refers to the noise image data input to the model in step 1. This data is used to drive the learning of real-world noise and alleviate the problem of noise distribution differences. After adding pseudo-noise to the sample data, the resulting noise image data is obtained; that is, the pseudo-noise image is generated from a given sample image, which is used as the ground truth, also known as the clean image. Therefore, during noise reduction training, it is not necessary to obtain the noise image; instead, it can be obtained by adding noise to the sample module through the pseudo-noise generation unit.

[0067] In addition, when constructing the noisy image data, a data augmentation unit is used to expand the amount of training data. This involves cropping the original image into 128×128 image blocks required for training and expanding the training data using random cropping, horizontal flipping, vertical flipping, and transposition.

[0068] Specifically, to improve the performance of deep learning models, more training data is usually required. However, obtaining large amounts of high-quality labeled data is costly. Therefore, data augmentation techniques are an important means to generate more dummy data by performing a series of transformations and expansions on the original data, thereby improving the model's generalization performance on both the training and test sets.

[0069] Data augmentation methods are widely used in tasks such as image classification, object detection, and image segmentation. Among these, the data augmentation unit is a core component of data augmentation techniques. This unit can crop the original image according to the required size, scale, lighting, and other factors to generate the desired image patches. Furthermore, the data augmentation unit can employ different data augmentation strategies for different problems. For example, it can use random cropping, horizontal flipping, vertical flipping, and transposition to augment the original image.

[0070] Random cropping is a common data augmentation strategy that generates image patches of different positions, sizes, and proportions by randomly cropping an image. Horizontal and vertical flipping expands the training data by flipping an image left-right or top-bottom. Transpose is a common matrix operation that rotates an image to increase its variation in the rotation direction. These data augmentation strategies effectively increase the amount of training data, improve the model's generalization performance, and do not increase annotation costs.

[0071] In some embodiments, the student network further includes a first encoder module, a first decoder module, and a first context connection module.

[0072] The first encoder module is used to progressively reduce the size of the image input to the first encoder module in layers to obtain a first feature map.

[0073] The first decoder module is configured to construct a denoised image of the image input to the first encoder module based on the first feature map.

[0074] The first context connection module is connected to the first encoder module and the first decoder module, and is used for the fusion of information in the first encoder module and the first decoder module.

[0075] The student network uses the denoising network UDNet, whose overall structure is based on the U-Net architecture. The UDNet in the student network has four encoder stages and four corresponding decoder stages, with context connection modules between the encoders and decoders.

[0076] Specifically, such as Figure 3As shown, the basic convolutional building blocks in the encoder, decoder, and context connection modules follow the same residual convolutional structure. This invention uses LeakyReLU as the activation function for each convolutional layer. The feature maps are downsampled by 1 / 2×scale through 2×2 max pooling at the end of each encoder stage and upsampled to 2×scale at the decoder stage through 2×2 deconvolution before each encoder stage. Large-scale low-level feature maps are passed from each encoder stage to its corresponding decoder stage via skip connections. The number of input channels is initially 3, transforming to 32, 64, 128, and 256 in the encoder stages, and then to 512 at the context connection block. In the decoder stage, the number of input channels is progressively restored to 32. The output of the final decoder is processed by a 3×3 convolutional layer as a global residual to handle noisy input and output a denoised result. Furthermore, an attention module is added at the end of each encoder stage. In the encoder structure, features are split into two branches after passing through the attention module: one is downsampled to the next layer through max pooling, and the other is connected to the corresponding decoder feature map via skip-connections. For features at different levels, this invention designs two attention modules: the Multi-Scale Aggregated Attention Module (MSAA) for shallow features and the Multi-Scale Linear Cross Attention Module (MLCA) for deep features.

[0077] Furthermore, to better aggregate and fuse features, this invention proposes two attention modules, one for shallow features and one for deep features. The Multi-Scale Aggregation Attention Module (MSAA) for shallow features aims to improve the encoder's ability to aggregate multi-scale features, thereby enhancing local information and capturing multi-scale context. The Multi-Scale Linear Cross Attention Module (MLCA) for deep features aims to improve the feature compression capability of the bottleneck layer. The cross attention module improves upon traditional self-attention modules. MSAA enables multi-scale fusion of feature maps; inserting it into the encoder improves encoder efficiency. MLCA aims to extract long-range structural information from high-dimensional feature maps, using linear cross attention blocks to fuse multi-scale information. Inserting MLCA blocks into the encoder-decoder connection enhances semantic features, highlights effective information, and aids in image decoding and reconstruction. These attention module designs effectively extract image features and improve the performance of the neural network.

[0078] Specifically, such as Figure 5The diagram shows the Multi-Scale Aggregated Attention (MSAA) module structure in the UDNet denoiser. MSAA consists of three parts: a depthwise convolution for aggregating local information, a multi-branch depthwise convolution of different sizes for capturing multi-scale context, and a 1×1 convolution for adjusting the number of output channels. The adjusted channel outputs are directly used as attention weights to reweight the MSAA input. In MSAA, k×1 and 1×k asymmetric convolutions are used instead of standard k×k 2D convolutions, reducing computational cost. The depthwise convolution kernel sizes for each branch are set to 3 and 5, respectively. Multi-scale fusion in MSAA is achieved through simple addition.

[0079] Specifically, MLCA is the multi-scale linear cross-attention module in the denoising UDNet, and its structure diagram is shown below. Figure 6 This module consists of a combination of multi-scale and multi-head linear cross-attention modules. Its main goal is to interact channel information on a multi-scale basis, filtering out irrelevant regions and highlighting regions that are significant for denoising. MLCA has three characteristics: First, it uses cross-attention to fuse multi-scale features, which provides more information than pure additive fusion methods; second, the interaction of cross-attention occurs at the channel level, unlike traditional self-attention mechanisms, making it easier to generalize to input noisy images with higher resolution during inference; finally, MLCA uses computationally efficient filtering gates to obtain attention weights, excluding pairwise interactions of query keys. These characteristics make MLCA perform well in denoising and improve the performance of neural networks. Finally, the MLCA in this invention excludes pairwise interactions of query keys and uses computationally efficient filtering gates to obtain attention weights. Therefore, it has linear computation and memory complexity. After strengthening the input feature map through the two branches of MSAA, this invention obtains k and v of the linear cross-attention module. Input k,v∈R H×W×C After being horizontally normalized, the number of channels of k and v is adjusted to D through a 1×1 convolution. module In MLCA, this invention uses eight parallel attention layers. The attention weight A in MLCA is... mlc Obtained through filtering. Specifically, K for each attention head. i ∈R N×D By key linear weight W kl ∈R D×1 The projection filters the information, which is then normalized using softmax and concatenated to obtain attention map A. mlc ∈R h×N MLCA has linear complexity. This invention fuses multi-scale features and adjusts the output channel size through 1×1 convolution.

[0080] In some embodiments, the teacher network includes: a second encoder module, a second decoder module, and a second context connection module.

[0081] The second encoder module is used to progressively reduce the size of the image input to the second encoder module in layers to obtain a second feature map.

[0082] The second decoder module is used to construct a denoised image of the image input to the second encoder module based on the second feature map.

[0083] The second context connection module is connected to the second encoder module and the second decoder module, and is used for the fusion of information in the first encoder module and the first decoder module.

[0084] The teacher network uses the denoising network UDNet, whose overall structure is based on the U-Net architecture. The student network's UDNet has four encoder stages and four corresponding decoder stages, with context connection modules between the encoders and decoders. In other words, the teacher and student networks have the same structure, which will not be elaborated further here.

[0085] In some embodiments, when calculating the denoising loss of the model, the denoising loss and consistency loss in the denoising loss can be calculated separately, and then the denoising loss can be calculated, as follows:

[0086] The loss function for denoising loss can be specifically defined as follows:

[0087] L s (x)=||xP x ||1

[0088] In the formula, L s (x) represents the sample image data, and P is the P sample image data. x Denoising data for noisy images.

[0089] The loss function for consistency loss is as follows:

[0090] L c (y)=||S y -T y ||1

[0091] In the formula, L c (y) is the L1 loss function, ||·||1 is the y-value, and S is the loss function. y T is the first real-world image denoising data generated for student network predictions. y Denoising data for the second real-world image generated by the teacher network prediction.

[0092] The loss function for noise reduction is as follows:

[0093]

[0094] In the formula, A and R represent the sets of noisy image data and real-world image data, respectively; N and M are the number of noisy image data and real-world image data in each iteration; and λ is a coefficient that varies with the number of iterations. t is the current iteration number, t max λ is the maximum number of iterations. max =1.2.

[0095] In some embodiments, parameter updates are performed from the student network's parameters θ' using an exponential moving average (EMA) strategy, based on the teacher network's parameters θ'. The parameter update formula for the teacher network during training iterations is θ'. t =ηθ' t-1 +(1-η)θ t θ t Let θ' be the parameter of the student network during training iterations. t-1 Let η be the parameters of the teacher network during the (t-1)th iteration of training, and let η be the smoothing parameter, η = 0.99.

[0096] The present invention also provides an unsupervised training system for an image denoising model, comprising:

[0097] The student network denoising module is used to input noisy image data and real-world image data into the student network to obtain noisy image denoised data and first real-world image denoised data; the noisy image data is the image data after adding pseudo-noise to the sample image; the student network is a denoising network based on the UDNet network framework.

[0098] The teacher network denoising module is used to input the same real-world image data into the teacher network to obtain second real-world image denoised data; the teacher network is a denoising network based on the UDNet network framework.

[0099] The loss calculation module is used to determine the denoising loss of the student network based on the denoised noisy image data, the first real-world image denoising data, and the second real-world image denoising data; the denoising loss includes denoising loss and consistency loss; the denoising loss is determined based on the sample image data and the denoised noisy image data; the consistency loss is determined based on the first real-world image denoising data and the second real-world image denoising data.

[0100] The parameter update module is used to update the weight parameters of the student network with the goal of minimizing the noise reduction loss.

[0101] The detection module is used to detect whether the current iteration number is the set iteration number; if yes, it outputs the student network at the current iteration number; if no, it performs the next iteration training based on the student network with updated weight parameters.

[0102] Furthermore, this invention also provides an unsupervised learning framework workflow for real-world image denoising, as detailed below:

[0103] The first step is data collection. This invention uses 78 images from the BSD500 dataset and 1400 images from the MIT-Adobe FiveK dataset as sample data, and images from the SIDDMedium dataset as real-world noise data. The validation dataset from SIDD is used to validate the model. This data is imported into the data import module.

[0104] The second step is data preprocessing, which uses 128×128 image patches and augments the training data using random cropping, horizontal flipping, vertical flipping, and transposition.

[0105] The third step is the noise reduction module, which uses the Adam optimizer to update the network and sets the momentum terms to β1 = 0.9 and β2 = 0.999. The initial learning rate is set to 2 × 10⁻⁶. -4 The learning rate decreasing scheduling is set to cosine annealing. Simultaneously, the weight decay is set to 10. -8 The minimum learning rate is 5×10 -5 The training batch size is 32, and the entire training process requires 100,000 small-step iterations. During training, the optimizer only updates the student network weights.

[0106] The fourth step is to use the SIDD validation dataset to validate the effectiveness of the model.

[0107] In summary, the present invention has the following advantages:

[0108] The MTUDNet used in this invention is a novel unsupervised image denoising framework designed to reduce the domain difference between synthetic and real noise. This model combines a Mean Teacher and a denoiser UDNet, and does not require any clean real-world images to be used.

[0109] The UDNet denoiser used in the student and teacher networks of this invention includes attention modules MSAA and MLCA, which can further improve the denoising capability of the model. MTUDNet outperforms existing unsupervised denoising methods in SIDD and DND benchmarks and competes with some supervised denoising methods in the DND benchmark.

[0110] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.

[0111] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for training an image denoising model, characterized in that, The method comprises the following steps: Step 1: inputting noise image data and real world image data into a student network to obtain noise image denoising data and first real world image denoising data; the noise image data is image data obtained by adding pseudo noise to sample image data; the student network is a denoising network based on a UDNet network framework; the overall structure of the denoising network UDNet is based on a U-Net architecture, and the UDNet has four encoder stages and four corresponding decoder stages, and there is a context connection module between the encoder and the decoder; Step 2: inputting the real world image data into a teacher network to obtain second real world image denoising data; the teacher network is a denoising network based on a UDNet network framework; Step 3: determining a noise reduction loss of the student network according to the noise image denoising data, the first real world image denoising data and the second real world image denoising data; the noise reduction loss comprises a denoising loss and a consistency loss; the denoising loss is determined according to sample image data and the noise image denoising data; the consistency loss is determined according to the first real world image denoising data and the second real world image denoising data; the loss function expression of the denoising loss is specifically as follows: wherein, is a denoising loss, A and R represent a set of noise image data and a set of real-world image data, respectively, and N and M represent the number of noise image data and real-world image data at each iteration, respectively, is a denoising loss at the i-th iteration, is a consistency loss at the j-th iteration, is sample image data at the i-th iteration, is real-world image denoising data at the j-th iteration, is a coefficient that varies with the number of iterations. the loss function expression of the denoising loss is specifically as follows: the loss function expression of the consistency loss is specifically as follows: wherein, is a denoising loss, x is sample image data, P x is noise image denoising data; is an L1 loss function, is a consistency loss, S y is first real-world image denoising data generated by student network prediction, T y is second real-world image denoising data generated by teacher network prediction; Step 4: updating the weight parameters of the student network with the minimum noise reduction loss as the target; Step 5: updating the weight parameters of the teacher network according to the updated weight parameters of the student network.

2. The training method of claim 1, wherein, Before step 1, it also comprises the following steps: obtaining an image data set; the image data set comprises the real world image data and the sample image data; adding pseudo noise data to the sample image data to obtain the noise image data.

3. The training method of claim 1, wherein, The student network comprises a first encoder module, a first decoder module and a first context connection module; the first encoder module is used for reducing the size of an image input into the first encoder module layer by layer and step by step to obtain a first feature map; the first decoder module is used for obtaining a denoised image according to the first feature map; the first context connection module is connected with the first encoder module and the first decoder module, and is used for fusing information in the first encoder module and the first decoder module.

4. The training method of claim 1, wherein, The teacher network comprises a second encoder module, a second decoder module and a second context connection module; the second encoder module is used for reducing the size of an image input into the second encoder module layer by layer and step by step to obtain a second feature map; the second decoder module is used for obtaining a denoised image according to the second feature map; the second context connection module is connected with the second encoder module and the second decoder module, and is used for fusing information in the second encoder module and the second decoder module.

5. The training method of claim 1, wherein, the coefficient that varies with the iteration number the calculation formula of the coefficient that varies with the iteration number is as follows: where t is the current iteration number, t max is the maximum iteration number, is the coefficient varying with the iteration number is the maximum value of the function 6. The training method of claim 1, wherein, Step 3 is specifically as follows: wherein, is the parameter of the student network at the tth iteration of training, is the parameter of the teacher network at the tth iteration of training, is the parameter of the teacher network at the (t-1)th iteration of training, is a smoothing exponent. 7.A training system of an image denoising model, characterized in that, comprises the following steps: The student network denoising module is configured to input noise image data and real-world image data into a student network to obtain noise image denoising data and first real-world image denoising data; the noise image data is image data obtained by adding pseudo-noise to sample image data; the student network is a denoising network based on a UDNet network framework; the overall structure of the denoising network UDNet is based on a U-Net architecture, and the UDNet has four encoder stages and four corresponding decoder stages, and there is a context connection module between the encoder and the decoder; The teacher network denoising module is configured to input the real-world image data into a teacher network to obtain second real-world image denoising data; the teacher network is a denoising network based on a UDNet network framework; The loss calculation module is configured to determine a denoising loss of the student network according to the noise image denoising data, the first real-world image denoising data, and the second real-world image denoising data; the denoising loss includes a denoising loss and a consistency loss; The denoising loss is determined according to sample image data and the noise image denoising data; the consistency loss is determined according to the first real-world image denoising data and the second real-world image denoising data; The loss function expression of the denoising loss is specifically as follows: wherein, is a denoising loss, A and R represent a set of noisy image data and a set of real-world image data, respectively, and N and M represent the number of noisy image data and real-world image data at each iteration, respectively, is a denoising loss at the i-th iteration, is a consistency loss at the j-th iteration, is a sample image data at the i-th iteration, is a real-world image denoising data at the j-th iteration, is a coefficient that varies with the number of iterations. The loss function expression of the denoising loss is specifically as follows: The loss function expression of the consistency loss is specifically as follows: wherein, is a denoising loss, x is sample image data, P x is noise image denoising data; is an L1 loss function, is a consistency loss, S y is first real-world image denoising data generated by student network prediction, T y is second real-world image denoising data generated by teacher network prediction; The student network parameter updating module is configured to update the weight parameters of the student network with the minimum denoising loss as the target; The teacher network parameter updating module is configured to update the weight parameters of the teacher network according to the updated weight parameters of the student network.

Images