Self-supervised image denoising and enhancement method and device based on noise iterative estimation

Abstract

The present invention proposes a self-supervised image denoising and enhancement method based on iterative noise estimation, including obtaining a noisy image and constructing a blind spot preprocessing denoising network; constructing and training a noise estimation network based on the blind spot preprocessing denoising network, and generating paired noisy images through the noise estimation network; using paired noisy images to train the blind spot preprocessing denoising network in depth to obtain a deep denoising network; iteratively optimizing the deep denoising network and the noise estimation network; and using the trained deep denoising network and the noise estimation network to denoise and image a real noisy image. The method proposed by the present invention can complete the noise parameter estimation and denoising network training by using only a single noisy data set, freeing the deep denoising network from the strong dependence on paired data.

Description

Technical Field

[0001] The invention belongs to the field of computational imaging science and technology. Background Art

[0002] Image denoising is an important research problem in the field of image processing, and its goal is to remove noise while preserving the details and structure of the original image. Image denoising not only contributes to the visual quality perceived by the human eye, but also helps downstream imaging problems, including remote sensing, medical imaging, and microscopy imaging. In recent years, with the rapid development of computing resources, deep learning-based methods, especially convolutional neural network (CNN)-based methods, have attracted more and more attention due to their powerful representation and spatial feature extraction capabilities, and have achieved performance far higher than traditional denoising methods. However, the training of such methods relies on high-quality noisy / clean training pairs, so their application in practical scenarios is limited.

[0003] Currently, there are usually two ways to solve the above problems.

[0004] One is to use synthetic noisy / clean image pairs for training, thus avoiding the collection of real image pairs. For this type of method, it is necessary to ensure that the noisy image is as close to the real noisy image as possible, so accurate modeling of the real noise distribution is critical. Early methods used simple Gaussian models to model the noise distribution. In recent years, Wei et al. proposed more accurate generative models and complex statistical models that can better describe the real sensor noise distribution. Although accurate noise modeling can serve as a good substitute for paired real data, the modeling and calibration of noise are also limited by manpower, material resources, and the environment, and cannot cover more usage scenarios.

[0005] The second is to design unsupervised and self-supervised training strategies so that the network no longer relies on noise / clean image pairs for training. On the contrary, such methods can be trained with more easily accessible paired images or even single noisy images. For example, Noise2Noise requires two different noisy images of the same scene for network learning. In order to further improve practicality, more methods have tried to recover clear images from a single noisy image in recent years. Among these methods, some methods are further optimizations of Noise2Noise. They generate noise / noise image pairs from a single noisy image, so that the learning of Noise2Noise no longer requires paired noisy images. However, most of these methods require the distribution and size of the noise model to be known as a prerequisite. Another type of method proposes a blind spot network to learn the denoising process from a single noisy image. The design of the blind spot network is to avoid identity mapping, but the design process results in the discarding of important pixel information, which also affects the denoising effect. Summary of the invention

[0006] The present invention aims to solve one of the technical problems in the related art at least to a certain extent.

[0007] To this end, the purpose of the present invention is to propose a self-supervised image denoising and enhancement method based on iterative noise estimation, which is used to complete noise estimation and removal when there is only a single noisy image.

[0008] To achieve the above-mentioned purpose, the first embodiment of the present invention proposes a self-supervised image denoising and enhancement method based on noise iterative estimation, comprising:

[0009] Get noisy images and build a blind spot preprocessing denoising network;

[0010] Constructing and training a noise estimation network based on the blind spot preprocessing denoising network, and generating paired noisy images through the noise estimation network;

[0011] Using paired noisy images to train the deep blind spot preprocessing denoising network, a deep denoising network is obtained;

[0012] Iteratively optimizing the deep denoising network and the noise estimation network;

[0013] Use the trained deep denoising network and noise estimation network to denoise and image the real noisy image.

[0014] In addition, the self-supervised image denoising and enhancement method based on iterative noise estimation according to the above embodiment of the present invention may also have the following additional technical features:

[0015] Furthermore, in one embodiment of the present invention, the step of constructing a blind spot preprocessing denoising network includes:

[0016] Given a noisy image y, randomly select 10% of the pixels in y as blind spots through mask M;

[0017] The blind spot mask is applied to the noisy image y and sent as input to the denoising network f(·,θ) to obtain the blind spot network output

[0018]

[0019] Among them, θ is the parameter of the denoising network.

[0020] The L2 loss function is used to train the blind spot network. The loss function of the blind spot network is expressed as:

[0021]

[0022] Further, in one embodiment of the present invention, the step of constructing and training a noise estimation network based on the blind spot preprocessing denoising network includes:

[0023] Use the blind spot preprocessing denoising network to denoise the noisy image y to obtain the clean image predicted by the network at the current stage:

[0024]

[0025] The clean image Input the noise estimation network simultaneously with the noisy image y Get the noise parameter map:

[0026]

[0027] For the clean signal x, the signal n generated during the sensor acquisition process is represented by a heteroscedastic Gaussian distribution:

[0028]

[0029] in, represents a Gaussian distribution, and Represent the variance coefficient related to the signal and the noise variance independent of the signal respectively; then the noise estimation network Output M nl For each pixel in the image and value;

[0030] Using M nl And use the reparameterization technique to regenerate two synthetic noisy images:

[0031]

[0032] Among them, z i represents the noise sampled from a standard normal distribution, that is, i∈{1,2}, I represents the identity matrix.

[0033] Furthermore, in one embodiment of the present invention, it is characterized in that it also includes:

[0034] The loss is calculated by comparing the block variance of the two synthetic noisy images with the block variance of the real noisy images. The block variance function is expressed as:

[0035]

[0036] in, Indicates that the image is converted into multiple p×p blocks.

[0037] Furthermore, in one embodiment of the present invention, it is characterized in that the deep blind spot preprocessing denoising network is trained using paired noisy images, and the loss function of the training is expressed as:

[0038]

[0039] Among them, y n1 and n2 represents a pair of noisy images.

[0040] Further, in one embodiment of the present invention, the iterative optimization of the deep denoising network and the noise estimation network includes:

[0041] Iteratively optimize the denoising-noising-denoising training process, where the iterative total loss function is:

[0042]

[0043] Among them, μ, γ, and λ represent the weights corresponding to the block variance loss, the deep denoising network loss, and the blind spot preprocessing denoising network, respectively.

[0044] To achieve the above-mentioned purpose, the second embodiment of the present invention proposes a self-supervised image denoising and enhancement device based on noise iterative estimation, comprising the following modules:

[0045] The first training module is used to obtain noisy images and build a blind spot preprocessing denoising network;

[0046] A second training module is used to construct and train a noise estimation network based on the blind spot preprocessing denoising network, and generate paired noisy images through the noise estimation network;

[0047] A third training module is used to train the deep blind spot preprocessing denoising network using paired noisy images to obtain a deep denoising network;

[0048] An iterative module, used for iteratively optimizing the deep denoising network and the noise estimation network;

[0049] The denoising module is used to denoise and image real noisy images using the trained deep denoising network and noise estimation network.

[0050] To achieve the above-mentioned purpose, the third aspect of the present invention proposes a computer device, characterized in that it includes a memory, a processor, and a computer program stored in the memory and executable on the processor, and when the processor executes the computer program, it implements the self-supervised image denoising and enhancement method based on noise iterative estimation as described above.

[0051] To achieve the above-mentioned objectives, the fourth aspect of the present invention proposes a computer-readable storage medium on which a computer program is stored, characterized in that when the computer program is executed by a processor, it implements the self-supervised image denoising and enhancement method based on noise iterative estimation as described above.

[0052] The self-supervised image denoising and enhancement method based on iterative noise estimation proposed in the embodiment of the present invention does not rely on expensive hardware costs and is a generally applicable denoising method. The method can be used to remove noise from images of different sensors, and can be performed without paired data sets or calibration data taken by a specific camera. The method of the present invention can complete the noise parameter estimation and denoising network training by using only a single noisy data set, freeing the deep denoising network from the strong dependence on paired data. The present invention is of great significance to the development of scientific research, industrial detection and other fields. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] The above and / or additional aspects and advantages of the present invention will become apparent and easily understood from the following description of the embodiments in conjunction with the accompanying drawings, in which:

[0054] Figure 1 A flowchart of a self-supervised image denoising and enhancement method based on iterative noise estimation provided by an embodiment of the present invention.

[0055] Figure 2 The present invention provides a general flow chart of denoising-noising-denoising unsupervised denoising.

[0056] Figure 3 A flowchart of noise estimation and paired noisy image generation provided by an embodiment of the present invention.

[0057] Figure 4 A schematic diagram of the flow of a self-supervised image denoising and enhancement device based on iterative noise estimation provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0058] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present invention, and should not be construed as limiting the present invention.

[0059] The following describes a self-supervised image denoising and enhancement method based on iterative noise estimation according to an embodiment of the present invention with reference to the accompanying drawings.

[0060] Figure 1 A flowchart of a self-supervised image denoising and enhancement method based on iterative noise estimation provided by an embodiment of the present invention.

[0061] like Figure 1 As shown, the self-supervised image denoising and enhancement method based on noise iterative estimation includes the following steps:

[0062] S101: Obtain a noisy image and build a blind spot preprocessing denoising network;

[0063] In order to ensure stable training of the network, it is necessary to construct a blind spot preprocessing denoising network to provide preliminary denoising processing for the self-supervised image denoising and enhancement method based on noise iterative estimation.

[0064] Specifically, given a noisy image y, 10% of the pixels in y are randomly selected as blind spots through the mask M. Then, the blind spot mask is applied to the noisy image y and sent as input to the denoising network f(·,θ), and the blind spot network output is obtained.

[0065]

[0066] In Formula 1, θ is the parameter of the denoising network.

[0067] The blind spot network is trained using the L2 loss function, and in order to avoid identity mapping, only the reconstruction results at the blind spot are used to constrain the network training. The loss function of the blind spot network is expressed as:

[0068]

[0069] At this point, the construction of the blind spot preprocessing denoising network has been completed.

[0070] Furthermore, in one embodiment of the present invention, a blind spot preprocessing denoising network is constructed, including:

[0071] Given a noisy image y, randomly select 10% of the pixels in y as blind spots through mask M;

[0072] Apply the blind spot mask to the noisy image y and send it as input to the denoising network f(·,θ) to get the blind spot network output

[0073]

[0074] Among them, θ is the parameter of the denoising network.

[0075] The L2 loss function is used to train the blind spot network. The loss function of the blind spot network is expressed as:

[0076]

[0077] S102: construct and train a noise estimation network based on the blind spot preprocessing denoising network, and generate paired noisy images through the noise estimation network;

[0078] like Figure 2 As shown, the present invention proposes a denoising-noising-denoising self-supervised denoising process to iteratively optimize the deep denoising network and the noise estimation network.

[0079] First, if Figure 2 As shown, the present invention proposes a self-supervised denoising process of denoising-noising-denoising, which iteratively optimizes the deep denoising network and the noise estimation network. To train the noise estimation network, the denoising network is first used to denoise the input noisy image to obtain the clean image predicted by the network at the current stage:

[0080]

[0081] Then, if Figure 3 As shown on the left, the clean image predicted by the network Input the noise estimation network simultaneously with the noisy input image y Thus, the noise parameter diagram is obtained

[0082]

[0083] In particular, this embodiment takes the Poisson-Gaussian noise generated during the acquisition process of a real image sensor as an example. This type of noise is usually represented by a heteroscedastic Gaussian distribution, that is, for a clean signal x, the signal n generated during the sensor acquisition process conforms to the following distribution:

[0084]

[0085] In formula 5, represents a Gaussian distribution, and Represent the variance coefficient related to the signal and the noise variance independent of the signal respectively. Based on this, the noise estimation network Output M nl For each pixel in the image and Then, if Figure 3 As shown on the right, use M nl And use the reparameterization technique to regenerate two synthetic noisy images

[0086]

[0087] Among them, z i represents the noise sampled from a standard normal distribution, i.e. i∈{1,2}, I represents the identity matrix.

[0088] Further, in one embodiment of the present invention, constructing and training a noise estimation network based on the blind spot preprocessing denoising network includes:

[0089] Use the blind spot preprocessing denoising network to denoise the noisy image y and get the clean image predicted by the network at the current stage:

[0090]

[0091] Clean Image Input the noise estimation network simultaneously with the noisy image y Get the noise parameter map:

[0092]

[0093] For the clean signal x, the signal n generated during the sensor acquisition process is represented by a heteroscedastic Gaussian distribution:

[0094]

[0095] in, represents a Gaussian distribution, and Represent the variance coefficient related to the signal and the noise variance independent of the signal respectively; then the noise estimation network Output M nl For each pixel in the image and value;

[0096] Using M nl And use the reparameterization technique to regenerate two synthetic noisy images:

[0097]

[0098] Among them, z i represents the noise sampled from a standard normal distribution, that is, i∈{1,2}, I represents the identity matrix.

[0099] In order to constrain the noise estimation network, the present invention proposes a block variance loss so that the generated y ni The noise statistics are close to the original input noisy image y. Specifically, the input image is stacked into multiple p×p blocks, and the variance of each block is calculated as the target statistic. Finally, the block variance of the generated noisy image is compared with the block variance of the real noisy image to calculate the loss. The block variance function can be expressed as

[0100]

[0101] In formula 7 Indicates that the image is converted into multiple p×p blocks.

[0102] Furthermore, in one embodiment of the present invention, it is characterized in that it also includes:

[0103] The loss is calculated by comparing the block variance of the two synthetic noisy images with the block variance of the real noisy images. The block variance function is expressed as:

[0104]

[0105] in, Indicates that the image is converted into multiple p×p blocks.

[0106] S103: Using the paired noisy images to train a deep blind spot preprocessing denoising network to obtain a deep denoising network;

[0107] In step 102, a pair of noisy images y with the same underlying clean image has been generated using the noise estimation network. n1 and n2 Obviously, y n1 and n2 The assumption that noise is independent and signal is not independent is satisfied, so y is used directly. n1 and n2 As the input and output of the denoising network, the L2 loss function is used to complete the deep denoising network training based on the Noise2Noise strategy. The loss function of using the generated paired noisy images to train the denoising network can be expressed as

[0108]

[0109] At this point, the training of the deep denoising network is completed.

[0110] Further, in one embodiment of the present invention, a deep blind spot preprocessing denoising network is trained using paired noisy images, and the loss function of the training is expressed as:

[0111]

[0112] Among them, y n1 and n2 represents a pair of noisy images.

[0113] S104: Iteratively optimize the deep denoising network and the noise estimation network;

[0114] Iteratively optimize the above denoising-noising-denoising training process. After each processing of steps 101 to 103, the noise estimation network will be closer to outputting more accurate noise parameters, and the deep denoising network will be closer to denoising parameters with better performance. The present invention continues to repeat steps 101 to 103 to complete the iterative optimization of the noise estimation network and the deep denoising network.

[0115] Specifically, each iteration requires the noise estimation network to be The noise estimation network and the denoising network f(·,θ) are optimized separately. First, the noise estimation network is optimized according to Formula 7. In order to avoid learning the identity mapping, the gradient is immediately backpropagated and the gradient is cleared after the loss function described in Formula 7 is calculated to avoid affecting the subsequent deep denoising network. Subsequently, the gradient of the blind spot denoising network in step 101 and the deep denoising network loss in step 103 are backpropagated to optimize the denoising network.

[0116] In general, the total loss function for each iteration from step 101 to step 103 is

[0117]

[0118] Among them, μ, γ, and λ represent the weights corresponding to the block variance loss, the deep denoising network loss, and the blind spot preprocessing denoising network, respectively.

[0119] In order to maintain the stability of training, the present invention specifies the weights of three loss functions. Specifically, during the entire training process, the coefficient λ of the pre-processing denoising loss function gradually changes from 1 to 0, ensuring that it serves to provide an initial denoising model; since the noise estimation network is easier to train than the denoising network, the coefficient γ of its loss function gradually changes from 1 to 0; in order to avoid gradient explosion, the loss weight of the deep denoising network in step 103 gradually changes from 0 to 1.

[0120] At this point, the iterative training of the noise estimation network and the deep denoising network is completed by minimizing the loss function to obtain a trained deep denoising network.

[0121] Furthermore, in one embodiment of the present invention, the deep denoising network and the noise estimation network are iteratively optimized, including:

[0122] Iteratively optimize the denoising-noising-denoising training process, where the iterative total loss function is:

[0123]

[0124] Among them, μ, γ, and λ represent the weights corresponding to the block variance loss, the deep denoising network loss, and the blind spot preprocessing denoising network, respectively.

[0125] S105: Use the trained deep denoising network and noise estimation network to denoise and image the real noisy image.

[0126] Use a deep denoising network to denoise the noisy image and obtain a denoised predicted clean image.

[0127] The self-supervised image denoising and enhancement method based on iterative noise estimation proposed in the embodiment of the present invention does not rely on expensive hardware costs and is a generally applicable denoising method. The method can be used to remove noise from images of different sensors, and can be performed without paired data sets or calibration data taken by a specific camera. The method of the present invention can complete the noise parameter estimation and denoising network training by using only a single noisy data set, freeing the deep denoising network from the strong dependence on paired data. The present invention is of great significance to the development of scientific research, industrial detection and other fields.

[0128] Compared with the prior art, the advantages of the present invention are:

[0129] 1) The present invention optimizes the deep denoising network and the noise estimation network in an iterative manner so that the two are optimized simultaneously, and the noise estimation network gradually and accurately predicts the noise distribution of the image. This method can gradually approach the performance of the Noise2Noise method trained with paired data, overcome the disadvantage of single noisy data relative to paired training data, and achieve the improvement and enhancement of the quality of the real image. The efficient training strategy proposed in the present invention provides important support for this method.

[0130] 2) The blind spot network preprocessing scheme proposed in the present invention enhances the stability of network training, making it possible to iteratively optimize the deep denoising network and the noise estimation network.

[0131] 3) The unsupervised denoising scheme proposed in the present invention does not rely on expensive hardware costs and is a generally applicable denoising method. The method can be used to remove noise from images of different sensors and can be performed without paired data sets or calibration data taken by a specific camera. The method of the present invention can complete the noise parameter estimation and denoising network training by using only a single noisy data set, freeing the deep denoising network from the strong dependence on paired data. Therefore, the present invention is of great significance to the development of scientific research and industrial detection and other fields.

[0132] In order to implement the above embodiment, the present invention also proposes a self-supervised image denoising and enhancement device based on noise iterative estimation.

[0133] Figure 4 A schematic diagram of the structure of a self-supervised image denoising and enhancement device based on iterative noise estimation provided by an embodiment of the present invention.

[0134] like Figure 4 As shown, the self-supervised image denoising and enhancement device based on noise iterative estimation includes: a first training module 100, a second training module 200, a third training module 300, an iteration module 400, and a denoising module 500, wherein:

[0135] The first training module is used to obtain noisy images and build a blind spot preprocessing denoising network;

[0136] The second training module is used to construct and train a noise estimation network based on the blind spot preprocessing denoising network, and generate paired noisy images through the noise estimation network;

[0137] The third training module is used to train a deep blind spot preprocessing denoising network using paired noisy images to obtain a deep denoising network;

[0138] Iterative module, used to iteratively optimize the deep denoising network and the noise estimation network;

[0139] The denoising module is used to denoise and image real noisy images using the trained deep denoising network and noise estimation network.

[0140] To achieve the above-mentioned purpose, the third aspect of the present invention proposes a computer device, characterized in that it includes a memory, a processor, and a computer program stored in the memory and executable on the processor, and when the processor executes the computer program, it implements the self-supervised image denoising and enhancement method based on noise iterative estimation as described above.

[0141] To achieve the above-mentioned objectives, the fourth aspect of the present invention proposes a computer-readable storage medium on which a computer program is stored, characterized in that when the computer program is executed by a processor, it implements the self-supervised image denoising and enhancement method based on noise iterative estimation as described above.

[0142] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, without contradiction.

[0143] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one of the features. In the description of the present invention, the meaning of "plurality" is at least two, such as two, three, etc., unless otherwise clearly and specifically defined.

[0144] Although the embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and are not to be construed as limitations of the present invention. A person skilled in the art may change, modify, replace and modify the above embodiments within the scope of the present invention.

Claims

1. A self-supervised image denoising and enhancement method based on noise iterative estimation, characterized in that: The following steps are involved: Get noisy images and build a blind spot preprocessing denoising network; Constructing and training a noise estimation network based on the blind spot preprocessing denoising network, and generating paired noisy images through the noise estimation network; Using paired noisy images to train the deep blind spot preprocessing denoising network, a deep denoising network is obtained; Iteratively optimizing the deep denoising network and the noise estimation network; Use the trained deep denoising network and noise estimation network to denoise and image the real noisy image; The blind spot preprocessing denoising network is constructed, comprising: Given a noisy image y, randomly select 10% of the pixels in y as blind spots through mask M; The blind spot mask is applied to the noisy image y and sent as input to the denoising network f(·,θ) to obtain the blind spot network output Among them, θ is the parameter of the denoising network; The L2 loss function is used to train the blind spot network. The loss function of the blind spot network is expressed as: The constructing and training a noise estimation network based on the blind spot preprocessing denoising network includes: Use the blind spot preprocessing denoising network to denoise the noisy image y to obtain the clean image predicted by the network at the current stage: The clean image Input the noise estimation network simultaneously with the noisy image y Get the noise parameter map: For the clean signal x, the signal n generated during the sensor acquisition process is represented by a heteroscedastic Gaussian distribution: in, represents a Gaussian distribution, and Represent the variance coefficient related to the signal and the noise variance independent of the signal respectively; then the noise estimation network Output M nl For each pixel in the image and value; Using M nl And use the reparameterization technique to regenerate two synthetic noisy images: Among them, z i represents the noise sampled from a standard normal distribution, that is, i∈{1,2}, I represents the identity matrix.

2. The method according to claim 1, characterized in that Also includes: The loss is calculated by comparing the block variance of the two synthetic noisy images with the block variance of the real noisy images. The block variance function is expressed as: in, Indicates that the image is converted into multiple p×p blocks.

3. The method according to claim 2, characterized in that The deep blind spot preprocessing denoising network is trained using paired noisy images, and the training loss function is expressed as: Among them, y n1 and n2 represents a pair of noisy images.

4. The method according to claim 3, characterized in that The iterative optimization of the deep denoising network and the noise estimation network comprises: Iteratively optimize the denoising-noising-denoising training process, where the iterative total loss function is: Among them, μ, γ, and λ represent the weights corresponding to the block variance loss, the deep denoising network loss, and the blind spot preprocessing denoising network loss, respectively.

5. A self-supervised image denoising and enhancement device based on noise iterative estimation, characterized in that: Includes the following modules: The first training module is used to obtain noisy images and build a blind spot preprocessing denoising network; A second training module is used to construct and train a noise estimation network based on the blind spot preprocessing denoising network, and generate paired noisy images through the noise estimation network; A third training module is used to train the deep blind spot preprocessing denoising network using paired noisy images to obtain a deep denoising network; An iterative module, used for iteratively optimizing the deep denoising network and the noise estimation network; The denoising module is used to denoise and image real noisy images using the trained deep denoising network and noise estimation network; The blind spot preprocessing denoising network is constructed, comprising: Given a noisy image y, randomly select 10% of the pixels in y as blind spots through mask M; The blind spot mask is applied to the noisy image y and sent as input to the denoising network f(·,θ) to obtain the blind spot network output Among them, θ is the parameter of the denoising network; The L2 loss function is used to train the blind spot network. The loss function of the blind spot network is expressed as: The constructing and training a noise estimation network based on the blind spot preprocessing denoising network includes: Use the blind spot preprocessing denoising network to denoise the noisy image y to obtain the clean image predicted by the network at the current stage: The clean image Input the noise estimation network simultaneously with the noisy image y Get the noise parameter map: For the clean signal x, the signal n generated during the sensor acquisition process is represented by a heteroscedastic Gaussian distribution: in, represents a Gaussian distribution, and Represent the variance coefficient related to the signal and the noise variance independent of the signal respectively; then the noise estimation network Output M nl For each pixel in the image and value; Using M nl And use the reparameterization technique to regenerate two synthetic noisy images: Among them, z i represents the noise sampled from a standard normal distribution, that is, i∈{1,2}, I represents the identity matrix.

6. A computer device, characterized in that: The method comprises a memory, a processor and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the method for self-supervised image denoising and enhancement based on iterative noise estimation as described in any one of claims 1 to 4 is implemented.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the self-supervised image denoising and enhancement method based on noise iterative estimation as described in any one of claims 1 to 4 is implemented.

Images