Image denoising method and system based on dynamic prior condition filtering
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- JIANGXI UNIVERSITY OF FINANCE AND ECONOMICS
- Filing Date
- 2022-01-24
- Publication Date
- 2026-07-21
AI Technical Summary
Existing image denoising methods are ineffective when dealing with real noise, especially since the distribution of real noise is complex and difficult to parameterize. Existing two-stage network structures cannot provide dynamic noise prior updates, resulting in suboptimal denoising performance.
A method based on dynamic prior condition filtering is adopted. By constructing a neural network model and combining feature encoding and decoding structures, it uses iterative learning affine transformation to extract real noise prior features, adaptively generates denoising convolution kernels, and dynamically updates noise prior features to achieve image feature denoising from coarse to fine.
It effectively removes Gaussian synthesized noise and has a robust removal effect on real noise, thus improving the quality of image denoising.
Smart Images
Figure CN114387188B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of computer image processing technology, and in particular to an image denoising method and system based on dynamic prior condition filtering. Background Technology
[0002] In recent years, with the continuous development of science and technology, digital imaging devices have emerged in large numbers. However, due to the limitations of electronic devices themselves and the influence of complex external environments, digital images are inevitably contaminated by various kinds of noise during acquisition and transmission, resulting in a significant reduction in image quality. Therefore, image denoising research has become an important part of the field of computer vision.
[0003] In early studies of image denoising, researchers often considered simple synthetic noise. Specifically, the synthetic noise distributions studied during this period mainly included Gaussian noise and salt-and-pepper noise. Related denoising methods can be roughly divided into: (1) methods based on traditional shallow models; (2) methods based on deep networks. Among them, methods based on traditional shallow models can be further divided into: (1.1) filtering-based methods; (1.2) methods based on predefined models. (1.1) Filtering-based methods often use manually designed low-pass filters to remove image noise. In addition, other representative works use nonlocal mean algorithms to achieve image denoising by finding similar sub-image patches in non-local regions of the image. (1.2) Methods based on predefined models usually explicitly design image prior information and model the denoising task as a maximum a posteriori probability estimation problem. Among them, sparsity and low rank are commonly used image priors. Although such methods have a solid theoretical background, shallow models based on predefined priors cannot describe the complex mapping relationship between noisy images and corresponding noise-free images. Especially when the noise intensity is large or the image texture is complex, such methods will also face challenges.
[0004] In recent years, thanks to the powerful representation capabilities of neural networks, significant progress has been made in image denoising targeting synthetic noise. Compared to methods based on traditional shallow models, deep network-based methods can utilize end-to-end training and significantly improve the removal capability of synthetic noise through distributed feature representation. However, since the distribution of real noise differs significantly from the simple synthetic noise distribution described above, the performance of these methods in removing real-world noise from images is less than satisfactory.
[0005] In recent years, with the release of more and more datasets of real-world noisy images, research on image denoising has gradually shifted towards real-world noise removal. Real-world noise is often caused by factors such as limitations of the shooting device itself, object motion, and color space mismatch. This type of noise is challenging due to its complex distribution, diverse types, and difficulty in parameterization. Currently, research on real-world noise removal in images mainly falls into two categories: one is implicitly learning noise priors, where methods directly train the model on real-world noisy images; the other is constructing a two-stage network structure, which first explicitly estimates the noise prior and then, based on this prior, reduces the blind real-world noise removal task to a non-blind image denoising task.
[0006] For methods that construct a two-stage network structure, to provide supervision information for the noise estimation task, such methods jointly train the model using real noisy images and synthetic noisy images. The synthetic noise is constructed using a heteroscedastic Gaussian distribution and the camera's internal processing function (ISP) to approximate the real noise distribution. However, this method has the following problems: (1) Existing methods often explicitly divide the network into two stages in time: a noise estimation stage and a non-blind denoising stage. The entire network only performs a one-time noise estimation on the input noisy image. This setting cannot provide dynamic noise prior updates, limiting the denoising effect; (2) Furthermore, existing methods always achieve non-blind denoising by merging estimated noise features and image texture features in the channel dimension and designing a way to share convolutional kernels in spatial locations. This method cannot adaptively adjust the denoising intensity at all feature locations according to specific texture features and noise priors, thus only achieving suboptimal results. Summary of the Invention
[0007] In view of the above, the main objective of this invention is to propose an image denoising method based on dynamic prior condition filtering to solve the aforementioned technical problems.
[0008] This invention provides an image denoising method based on dynamic prior condition filtering, wherein the method includes the following steps:
[0009] Step 1: Construct training data based on the obtained synthetic noise image dataset and real noise image dataset, and build a neural network model based on feature encoding and feature decoding structure, wherein the neural network model includes a non-blind denoising branch and a noise estimation branch.
[0010] Step 2: The noise estimation module in the noise estimation branch extracts the true noise prior features based on the affine transformation method of iterative learning.
[0011] Step 3: The non-blind denoising branch combines the image features of each local region with the prior features of the real noise to adaptively generate the denoising convolution kernel corresponding to the image features of the local region, wherein the denoising convolution kernel is used to improve the denoising effect of the image features of the local region.
[0012] Step four: Dynamically update the real noise prior features through multiple noise estimation modules in the neural network model to perform image feature denoising from coarse to fine.
[0013] Step 5: When the noise estimation module iterates to convergence, the denoising operation is completed and the denoised image is finally obtained.
[0014] This invention proposes an image denoising method based on dynamic prior condition filtering. It acquires supervised datasets of synthetically noisy images and datasets of real-world noisy images, and jointly trains a model using these two types of noisy images. A neural network model is constructed based on feature encoding and decoding structures to achieve multi-scale denoising from coarse to fine at the feature level. Based on the image features of each local region and the prior features of real-world noise, denoising convolutional kernels for each local region's image features are adaptively generated, specifically improving the denoising effect of the corresponding local region's image features. Along with the coarse-to-fine feature denoising process, multiple noise estimation modules are explicitly designed to dynamically update the prior features of real-world noise. The noise estimation modules utilize iteratively learned affine transformations, i.e., translation and scaling estimation structures, to extract the signal-correlated and signal-independent parts of the real-world noise distribution, respectively. When the noise estimation modules iterate to convergence, the denoising operation is completed, and the denoised image is finally obtained. This invention not only efficiently removes Gaussian synthetic noise but also has a robust removal effect on real-world noise.
[0015] The image denoising method based on dynamic prior condition filtering, wherein, in step one, the method for constructing the training data includes the following steps:
[0016] The training set and the test set are divided based on the synthetic noise image dataset and the real noise image dataset;
[0017] Obtain the synthetic noise image corresponding to the noise-free image from the synthetic noise image dataset, and obtain the real noise image corresponding to the noise-free image from the real noise image dataset;
[0018] The training data is constructed by cropping the synthetic noise image corresponding to the noise-free image and the real noise image corresponding to the noise-free image according to the preset image size, extracting sub-images and performing random enhancement.
[0019] The image denoising method based on dynamic prior condition filtering includes methods for randomly enhancing the sub-image, such as 90° rotation, 180° rotation, vertical flip, or horizontal flip.
[0020] The image denoising method based on dynamic prior condition filtering, wherein the neural network model corresponds to a loss function, the expression of which is:
[0021]
[0022] in, Denotes the loss function, This represents the fitted deep generative model. This represents the output noise-free image. This represents the noisy input image. This represents the set of learnable parameters of a neural network. Indicates the initial weights. This represents the estimation accuracy of the prior features of the true noise.
[0023] The image denoising method based on dynamic prior condition filtering, wherein the estimation accuracy of the true noise prior features... The expression is:
[0024]
[0025] in, The factor representing the degree of adjustment for overestimation. , Indicates an indicator function, Indicates pixel position The estimated value of noise variance. Indicates pixel position The true value of noise variance This indicates the serial number of the conditional denoising module.
[0026] The image denoising method based on dynamic prior condition filtering, wherein the expression for the noise estimation module used to extract the prior features of real noise is:
[0027]
[0028]
[0029] in, This represents the noise variance related to the irradiance of the original pixel. This represents the variance of heteroscedastic Gaussian distributed noise. This represents the irradiance of the original pixel. This represents a specific noise value that conforms to the distribution. Represents the signal-dependent noise component. Indicates signal-independent noise components. express A function used to simulate the in-camera processing steps to generate realistic noise.
[0030] The image denoising method based on dynamic prior conditional filtering, wherein the non-blind denoising branch includes a shallow feature extraction module, a conditional denoising module, and an output module, and each conditional denoising module includes a deep feature extraction module and a conditional filtering submodule, the expression of which is:
[0031]
[0032] in, This represents the output characteristics of the conditional filtering submodule. This represents the input characteristics of the conditional filtering submodule. All of these are spatial locations. For channel indexing, Indicated by feature position The local area centered on Represents the local feature domain of an image. The convolution kernel represents the local feature domain of noise.
[0033] The image denoising method based on dynamic prior condition filtering, wherein the expression for the denoising convolution kernel is:
[0034]
[0035] in, This refers to the denoising convolution kernel. This represents the Hadama product.
[0036] The present invention also proposes an image denoising system based on dynamic prior condition filtering, wherein the system includes a noise estimation branch and a non-blind denoising branch;
[0037] The noise estimation branch includes multiple noise estimation modules, and the non-blind denoising branch includes a shallow feature extraction module, a conditional denoising module, and an output module. The conditional denoising module includes multiple modules, including a deep feature extraction module and a conditional filtering submodule.
[0038] The non-blind denoising branch also includes an upsampling module and a downsampling module, both of which are connected to two adjacent conditional denoising modules.
[0039] The image denoising system based on dynamic prior condition filtering performs the image denoising method based on dynamic prior condition filtering as described above.
[0040] The image denoising system based on dynamic prior conditional filtering includes a conditional filter submodule, which in turn includes a kernel generation subnetwork, which includes multiple convolutional layers.
[0041] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by means of embodiments of the invention. Attached Figure Description
[0042] Figure 1 This is a flowchart of the image denoising method based on dynamic prior condition filtering proposed in this invention.
[0043] Figure 2 This is the network topology diagram of the image denoising system based on dynamic prior condition filtering proposed in this invention;
[0044] Figure 3 The topology diagram of the conditional filtering submodule in the proposed image denoising system based on dynamic prior conditional filtering is shown below.
[0045] Figure 4 This is a topology diagram of the noise estimation module in the image denoising system based on dynamic prior condition filtering proposed in this invention. Detailed Implementation
[0046] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.
[0047] These and other aspects of the embodiments of the invention will become clear from the following description and accompanying drawings. In these descriptions and drawings, some specific embodiments of the invention are specifically disclosed to illustrate some ways of implementing the principles of the embodiments of the invention; however, it should be understood that the scope of the embodiments of the invention is not limited thereto. Rather, the embodiments of the invention include all variations, modifications, and equivalents falling within the spirit and scope of the appended claims.
[0048] Please see Figure 1 This invention proposes an image denoising method based on dynamic prior condition filtering, wherein the method includes the following steps:
[0049] S101, training data is constructed based on the obtained synthetic noise image dataset and real noise image dataset, and a neural network model is constructed based on the feature encoding and feature decoding structure, wherein the neural network model includes a non-blind denoising branch and a noise estimation branch.
[0050] It should be noted that the neural network model in this invention is only used to predict the residual between the noiseless image and the corresponding noisy image.
[0051] In this step, the method for constructing the above training data includes the following steps:
[0052] S1011, the training set and test set are divided based on the synthetic noise image dataset and the real noise image dataset.
[0053] S1012, obtain the synthetic noise image corresponding to the noise-free image from the synthetic noise image dataset, and obtain the real noise image corresponding to the noise-free image from the real noise image dataset.
[0054] S1013, the synthetic noise image corresponding to the noiseless image and the real noise image corresponding to the noiseless image are cropped according to the preset image size, and sub-images are extracted and randomly enhanced to construct the training data.
[0055] In step S1013, the method for randomly enhancing the sub-image includes: 90° rotation, 180° rotation, vertical flip, or horizontal flip.
[0056] In this embodiment, the aforementioned neural network model corresponds to a loss function, the expression of which is:
[0057]
[0058] in, Denotes the loss function, This represents the fitted deep generative model. This represents the output noise-free image. This represents the noisy input image. This represents the set of learnable parameters of a neural network. Indicates the initial weights. This represents the estimation accuracy of the prior features of the true noise.
[0059] As a supplement, normative terms The weight is fixed at 1, and the initial weight is... It is 0.5. During the training process of the neural network model, A sliding assignment strategy is used for value assignment. As training progresses... The noise gradually decays to 0. In the early stages of training, the neural network model improves the accuracy of noise estimation and the effect of non-blind denoising in a balanced way. In the later stages of training, the neural network model focuses on reconstructing the residual between the noiseless image and the corresponding noisy image.
[0060] Furthermore, the estimation accuracy of the aforementioned true noise prior features... The expression is:
[0061]
[0062] in, The factor representing the degree of adjustment for overestimation. , Indicates an indicator function, Indicates pixel position The estimated value of noise variance. Indicates pixel position The true value of noise variance This indicates the serial number of the conditional denoising module.
[0063] S102, the noise estimation module in the noise estimation branch extracts the true noise prior features based on the affine transformation method of iterative learning.
[0064] In this step, the expression for the noise estimation module used to extract the prior features of true noise is:
[0065]
[0066]
[0067] in, This represents the noise variance related to the irradiance of the original pixel. This represents the variance of heteroscedastic Gaussian distributed noise. This represents the irradiance of the original pixel. This represents a specific noise value that conforms to the distribution. Represents the signal-dependent noise component. Indicates signal-independent noise components. express A function used to simulate the in-camera processing steps to generate realistic noise.
[0068] It should be added that, in the design of the neural network model, each noise estimation module is based on iterative learning of affine transformation, namely translation and scaling estimation structure, to extract the signal-independent part and the signal-dependent part of the real noise distribution, respectively.
[0069] S103, the non-blind denoising branch combines the image features of each local region with the prior features of the real noise to adaptively generate the denoising convolution kernel corresponding to the image features of the local region, wherein the denoising convolution kernel is used to improve the denoising effect of the image features of the local region.
[0070] In this embodiment, the non-blind denoising branch includes a shallow feature extraction module, a conditional denoising module, and an output module. Each conditional denoising module includes a deep feature extraction module and a conditional filtering submodule. Specifically, the expression for the conditional filtering submodule is:
[0071]
[0072] in, This represents the output characteristics of the conditional filtering submodule. This represents the input characteristics of the conditional filtering submodule. All of these are spatial locations. For channel indexing, Indicated by feature position The local area centered on Represents the local feature domain of an image. The convolution kernel represents the local feature domain of noise.
[0073] S104, the real noise prior features are dynamically updated through multiple noise estimation modules in the neural network model to perform image feature denoising from coarse to fine.
[0074] In this step, as noise in the feature domain is gradually removed, the noise prior features extracted by the noise estimation module change dynamically to assist the non-blind denoising task. Since only the first noise estimation module has explicit supervision information, its performance is optimized through training, and subsequent noise estimation modules share parameters with the first one.
[0075] S105, when the noise estimation module iterates to convergence, the denoising operation is completed and the denoised image is finally obtained.
[0076] In this step, after the noise estimation module completes its iteration to convergence, the residual output by the neural network model is added to the initial noisy image to obtain the predicted noiseless image.
[0077] Please see Figures 2 to 4 This invention proposes a topology diagram for an image denoising system based on dynamic prior condition filtering.
[0078] Specifically, using real noise images As input, this invention learns to Deep generative models with parameters To restore noise-free images and the input noisy image The residuals between them. In this embodiment of the invention, noise estimation and feature denoising tasks are performed respectively using a noise estimation branch and a non-blind denoising branch. As noise in the feature domain is gradually removed, the noise estimation branch extracts dynamic noise priors to assist the non-blind denoising task.
[0079] In this embodiment, the activation layer in the neural network model is PRelu, and the corresponding expression is:
[0080]
[0081] in, Input features The One portion, It is a learnable slope parameter with an initial value of 0.25.
[0082] The neural network model designed in this invention is jointly trained using real and synthetic noise images. The synthetic noise image is constructed using a heteroscedastic Gaussian distribution and an internal camera processing function (ISP) to approximate the real noise distribution. This synthetic noise image can provide supervision information for noise prior estimation tasks.
[0083] Specifically, the aforementioned non-blind denoising branch includes, in sequence, one shallow feature extraction module, six conditional denoising modules, and one output module. The shallow feature extraction module contains three convolutional layers. Both the upsampling and downsampling modules are connected to two adjacent conditional denoising modules.
[0084] Each downsampling module consists of one pooling layer and one convolutional layer. Correspondingly, one transposed convolutional layer is used to construct the upsampling module. Then, using the U-net neural network, symmetric structures with 64 to 256 channels are designed for the six conditional denoising modules. In each conditional denoising module, the number of feature channels remains constant, and it consists of one deep feature extraction module and two conditional filter sub-modules. The deep feature extraction module includes five convolutional layers. Furthermore, skip connections are introduced between identical deep feature extraction modules.
[0085] The noise estimation branch includes six noise estimation modules, the same number as the conditional denoising modules in the non-blind denoising branch. Since only the first noise estimation module has explicit supervision information, its performance is optimized through training, and subsequent noise estimation modules share parameters with it. In this embodiment, the conditional filtering submodule includes one conditional filter and two convolutional layers. A skip connection is introduced between the input and output features of the conditional filter; the specific principle expression of the conditional filtering submodule has been described above.
[0086] Specifically, the conditional filtering submodule of this invention includes a kernel generation subnetwork, which consists of two identical network structures that respectively model local region convolutional kernel generation functions for image features and noise features. In terms of implementation, the convolutional kernel generation function for each domain contains five convolutional layers. This convolutional kernel generation function characterizes the region convolutional kernel generation functions derived from features. Mapping to the corresponding convolutional kernel .in, B, C, H, W The batch size, number of channels, height, and width are the features. The final conditional filter kernel 210 is obtained by performing a Hadamard product on the convolution kernels corresponding to the image feature domain and the noise feature domain. That is, the expression for the denoising convolution kernel is:
[0087]
[0088] in, This refers to the denoising convolution kernel. This represents the Hadama product.
[0089] This invention proposes an image denoising method based on dynamic prior condition filtering. It acquires supervised datasets of synthetically noisy images and datasets of real-world noisy images, and jointly trains a model using these two types of noisy images. A neural network model is constructed based on feature encoding and decoding structures to achieve multi-scale denoising from coarse to fine at the feature level. Based on the image features of each local region and the prior features of real-world noise, denoising convolutional kernels for each local region's image features are adaptively generated, specifically improving the denoising effect of the corresponding local region's image features. Along with the coarse-to-fine feature denoising process, multiple noise estimation modules are explicitly designed to dynamically update the prior features of real-world noise. The noise estimation modules utilize iteratively learned affine transformations, i.e., translation and scaling estimation structures, to extract the signal-correlated and signal-independent parts of the real-world noise distribution, respectively. When the noise estimation modules iterate to convergence, the denoising operation is completed, and the denoised image is finally obtained. This invention not only efficiently removes Gaussian synthetic noise but also has a robust removal effect on real-world noise.
[0090] Please see Figure 2 The present invention also proposes an image denoising system based on dynamic prior condition filtering, wherein the image denoising system based on dynamic prior condition filtering is a neural network model, and the system includes a noise estimation branch and a non-blind denoising branch;
[0091] The noise estimation branch includes multiple noise estimation modules, and the non-blind denoising branch includes a shallow feature extraction module, a conditional denoising module, and an output module. The conditional denoising module includes multiple modules, including a deep feature extraction module and a conditional filtering submodule.
[0092] The non-blind denoising branch also includes an upsampling module and a downsampling module, both of which are connected to two adjacent conditional denoising modules.
[0093] The image denoising system based on dynamic prior condition filtering performs the image denoising method based on dynamic prior condition filtering as described above.
[0094] The conditional filtering submodule includes a conditional filter, which includes a kernel generation subnetwork, and the kernel generation subnetwork includes multiple convolutional layers.
[0095] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.
[0096] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.
[0097] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.
Claims
1. An image denoising method based on dynamic prior condition filtering, characterized in that, The method includes the following steps: Step 1: Construct training data based on the acquired synthetic noise image dataset and real noise image dataset, and build a neural network model based on feature encoding and feature decoding structures. The neural network model includes a non-blind denoising branch and a noise estimation branch. The method for constructing the training data includes the following steps: The training set and the test set are divided based on the synthetic noise image dataset and the real noise image dataset; Obtain the synthetic noise image corresponding to the noise-free image from the synthetic noise image dataset, and obtain the real noise image corresponding to the noise-free image from the real noise image dataset; The training data is constructed by cropping the synthetic noise image corresponding to the noiseless image and the real noise image corresponding to the noiseless image according to the preset image size, extracting sub-images and performing random enhancement. The method of performing random enhancement on the sub-images includes: 90° rotation, 180° rotation, vertical flip or horizontal flip. Step 2: The noise estimation module in the noise estimation branch extracts the true noise prior features based on the affine transformation method of iterative learning. Step 3: The non-blind denoising branch combines the image features of each local region with the prior features of the real noise to adaptively generate the denoising convolution kernel corresponding to the image features of the local region, wherein the denoising convolution kernel is used to improve the denoising effect of the image features of the local region. Step four: Dynamically update the real noise prior features through multiple noise estimation modules in the neural network model to perform image feature denoising from coarse to fine. Step 5: When the noise estimation module iterates to convergence, the denoising operation is completed and the denoised image is finally obtained.
2. The image denoising method based on dynamic prior condition filtering according to claim 1, characterized in that, The neural network model corresponds to a loss function, the expression of which is: ； in, Denotes the loss function, This represents the fitted deep generative model. This represents the output noise-free image. This represents the noisy input image. This represents the set of learnable parameters of a neural network. Indicates the initial weights. This represents the estimation accuracy of the prior features of the true noise.
3. The image denoising method based on dynamic prior condition filtering according to claim 2, characterized in that, The estimation accuracy of the true noise prior features The expression is: ； in, The factor representing the degree of adjustment for overestimation. , Indicates an indicator function, Indicates pixel position The estimated value of noise variance. Indicates pixel position The true value of noise variance This indicates the serial number of the conditional denoising module.
4. The image denoising method based on dynamic prior condition filtering according to claim 3, characterized in that, The expression for the noise estimation module used to extract prior features of real noise is: ； ； in, This represents the noise variance related to the irradiance of the original pixel. This represents the variance of heteroscedastic Gaussian distributed noise. This represents the irradiance of the original pixel. This represents a specific noise value that conforms to the distribution. Represents the signal-dependent noise component. Indicates signal-independent noise components. express A function used to simulate the in-camera processing steps to generate realistic noise.
5. The image denoising method based on dynamic prior condition filtering according to claim 4, characterized in that, The non-blind denoising branch includes a shallow feature extraction module, a conditional denoising module, and an output module. Each conditional denoising module includes a deep feature extraction module and a conditional filtering submodule. The expression for the conditional filtering submodule is as follows: ； in, This represents the output characteristics of the conditional filtering submodule. This represents the input characteristics of the conditional filtering submodule. All of these are spatial locations. For channel indexing, Indicated by feature position The local area centered on Represents the local feature domain of an image. The convolution kernel represents the local feature domain of noise.
6. The image denoising method based on dynamic prior condition filtering according to claim 5, characterized in that, The expression for the denoising convolution kernel is: ； in, This refers to the denoising convolution kernel. This represents the Hadama product.
7. An image denoising system based on dynamic prior condition filtering, characterized in that, The system includes a noise estimation branch and a non-blind denoising branch; The noise estimation branch includes multiple noise estimation modules, and the non-blind denoising branch includes a shallow feature extraction module, a conditional denoising module, and an output module. The conditional denoising module includes multiple modules, including a deep feature extraction module and a conditional filtering submodule. The non-blind denoising branch also includes an upsampling module and a downsampling module, both of which are connected to two adjacent conditional denoising modules. The image denoising system based on dynamic prior condition filtering performs the image denoising method based on dynamic prior condition filtering as described in any one of claims 1 to 6.
8. The image denoising system based on dynamic prior condition filtering according to claim 7, characterized in that, The conditional filtering submodule includes a conditional filter, which includes a kernel generation subnetwork, and the kernel generation subnetwork includes multiple convolutional layers.