A variational inference based blind poisson denoising deep learning method

Abstract

The application discloses a blind Poisson denoising deep learning method based on variational inference, first, the variational inference based on Poisson distribution and the deep convolution network are used to infer the Poisson noise statistical information, and the final Poisson information is sampled through the resampling technology based on the Gamma distribution.Combining the Poisson information, the proposed denoising sub-network can effectively complete the blind Poisson denoising task, and the denoising performance is relatively excellent.Because the variational inference is combined with the deep learning, the proposed method has strong interpretability and efficient denoising capacity.Therefore, the method can realize the Poisson denoising task under different conditions.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of image denoising, and relates to a model-driven and data-driven combined method for blind denoising of Poisson noise images, which is suitable for various photon counting imaging application scenarios under extremely low light conditions. BACKGROUND

[0002] In CCD solid-state photoelectric detector arrays, computed tomography (CT), fluorescence microscopic imaging, astronomical imaging and other imaging applications based on photon counting, images are obtained by counting the photons hitting the sensor. When capturing images under low-intensity illumination or short exposure time, the main noise source can be well modeled as Poisson noise, and the strength of the Poisson noise depends on the pixel intensity. In recent years, Poisson data has great prospects in the medical image and commercial fields, and Poisson denoising has become a hot topic in the field of image denoising in recent years.

[0003] At present, the existing Poisson denoising methods are divided into VST denoising methods, model-driven denoising and data-driven denoising methods. The main idea of the VST denoising method is: first, the variance-stabilizing transformation (VST) is performed on the Poisson noise image, and the common transformations are Anscombe and Haar-Fisz transformation. The purpose of VST is to replace the Poisson noise with additive Gaussian white noise with constant variance, so as to reduce the dependence of the Poisson noise on the signal. Secondly, the classic Gaussian denoising method is used to estimate the latent clean image. Finally, the inverse VST transformation is used to obtain the denoised image. This method can achieve fast Poisson denoising performance when the light is strong, and has good Poisson denoising ability. The second method is the model denoising method. The model-driven method is based on the Bayesian perspective, and the objective function is constructed according to the statistical properties of the Poisson noise. The objective function usually includes two items. One is the data fidelity term, which plays a role in restoring the image; the other is the prior term, which constrains the latent distribution. The model-driven method has strong interpretability and natural blind denoising ability. The third method is the data-driven method. Benefiting from the development of deep learning, the Poisson denoising method represented by convolutional neural network is studied and proposed. By learning clean-noisy image pairs, the deep learning method has strong Poisson denoising performance.

[0004] Although the Poisson denoising has made considerable progress, there are still the following challenging problems to be solved for blind Poisson denoising. Since the distribution of the VST post-noise image is still far from the assumption of Gaussian distribution, the performance of the indirect method is poor under extremely low light conditions (i.e. low signal-to-noise ratio). For the model-driven method, the existing priori still cannot accurately describe the complex characteristics of natural scenes. Therefore, the constructed denoising model is inaccurate. At the same time, the difficulty of selecting the optimal model parameter is another urgent problem to be solved. Although the data-driven method can successfully avoid the above problems by learning from data, most of the existing methods are always trained as a black box, lack of good explanation, and lead to overfitting problem. Therefore, it is a key problem to design a Poisson denoising method to have strong generalization ability while maintaining efficient denoising ability. SUMMARY

[0005] In order to solve the problems existing in the prior art, the present application provides a blind Poisson denoising deep learning method based on variational inference.

[0006] The present application provides a blind Poisson denoising deep learning method based on variational inference, comprising the following steps:

[0007] Step 1, under the variational inference framework, a target function of variational inference is constructed based on Poisson distribution;

[0008] Step 2, the target function of variational inference is solved to obtain the specific form of the lower bound of evidence;

[0009] Step 3, the inference subnetwork is constructed and used to parameterize the variational distribution, and the variational inference is completed to obtain the optimal solution;

[0010] Step 4, the inference result of the variational inference is sampled by a Gamma-based reparameterization technique to obtain Poisson noise statistical information, a denoising subnetwork is constructed, and the Poisson noise image is denoised in combination with the variational Poisson noise statistical information, and finally the denoised result is obtained;

[0011] Step 5, the inference subnetwork and the denoising subnetwork are trained by using clean image-Poisson noise image pairs, and after the loss function is stable, the trained model is obtained, and the Poisson denoising and blind denoising results are obtained by using the trained model network parameters and test data.

[0012] Further, the specific implementation mode of step 1 is as follows:

[0013] Step 1.1, let y, x represent the image containing Poisson noise and the clean noise-free image respectively, and the imaging process of Poisson noise is represented as follows:

[0014]

[0015] where P(y|x) is the degenerate model of Poisson noise, N is the total number of image pixels;

[0016] Step 1.2, using the variable λ to represent the latent variable, according to the statistical properties of the Poisson distribution, the Gamma conjugate prior is imposed on the latent variable:

[0017]

[0018] where, represents the Gamma conjugate prior, s i ,γ i represents the parameters of the Gamma conjugate prior;

[0019] Step 1.3, based on formula (2), the variational posterior form of the latent variable λ is represented as follows:

[0020]

[0021] where q(λ|) represents the variational posterior of the latent variable λ, α i (y i ;Θ),β i (y i ;) represents two inference functions of the inference subnetwork Θ;

[0022] Step 1.4, under the variational framework, the target optimization function of the Poisson distribution is constructed:

[0023] q * (λ)=argmin q(λ)∈Q D KL (q(λ|y;Θ)||p(λ|y)) (4)

[0024] where D KL represents the Kullback-Leibler divergence, q(λ|y;Θ) represents the variational posterior distribution, p(λ|y) is the true posterior distribution, Q represents the variational distribution family, q(γ) is regarded as a variable, and q*(γ) is used to represent when q(γ) satisfies the KL divergence with the true posterior distribution.

[0025] Further, the specific implementation manner of step 2 is as follows:

[0026] Step 2.1, by equivalent replacement of the target optimization function (4), the following form is obtained:

[0027]

[0028] where p(y) is the marginal likelihood;

[0029] Step 2.2, the lower bound of the evidence in equation (5) Formal decomposition, based on equations (1), (2) and (4), can obtain the specific form of the evidence lower bound;

[0030]

[0031] where, Γ(·), Ψ(·) represent gamma and digamma functions respectively.

[0032] Further, the specific implementation mode of step 3 is as follows:

[0033] Step 3.1, construct an inference subnetwork to complete the variational inference task, and the inference subnetwork is a U-Net network with a 4-layer structure, wherein all convolution kernel sizes are 3x3, the step is 1x1, and the padding is 1x1;

[0034] Step 3.2, the input of the inference subnetwork is the Poisson noise image, and the output is two parameters of the variational posterior;

[0035] Step 3.3, according to the evidence lower bound obtained by inference according to the variational framework, the loss function of the inference subnetwork is constructed:

[0036]

[0037] Further, the U-Net network is a symmetric network with an encoder-decoder structure, in the encoder part, three convolution operations are performed at each layer, and LeakyReLU is used as the activation function after each convolution, and when reaching the next layer, mean pooling is used to complete the down sampling operation, wherein the pooling kernel size is 2x2 and the step is 2x2; in the decoder part, each layer first performs a deconvolution operation, the deconvolution kernel size is 2x2 and the step is 2x2, and then two convolution operations are performed after deconvolution, the convolution operation is set to be consistent with the decoder part, and the encoder and the decoder are connected through the jump connection.

[0038] Further, the specific implementation mode of step 4 is as follows:

[0039] Step 4.1, the purpose of variational inference is to infer the parameters of the variational posterior, in order to obtain the Poisson statistical information, sampling needs to be performed from the variational posterior, however, ordinary random sampling will result in that the sampling result has no gradient, so that optimization cannot be performed; therefore, resampling technology needs to be used for sampling, based on the statistical characteristics of the Gamma distribution, the resampling mode is as follows:

[0040]

[0041] Step 4.2, a denoiser network using the statistical information of Poisson noise is constructed, the denoiser network is a deep network structure integrating residual learning and skip connection, a residual network is constructed using three three-layer sub-decoders, in each sub-decoder, convolution, skip connection and deconvolution are used to form a symmetric network, and each layer is connected through the channel attention based on contrast perception.

[0042] Further, the convolution operations in the first layer, the second layer and the third layer of the sub-decoder use 64, 128 and 256 convolution kernels respectively.

[0043] Further, the convolution kernel sizes of the three sub-decoders are 7*7, 5*5 and 3*3 respectively, and the sizes and steps of the pooling kernel and the deconvolution kernel are both 2*2.

[0044] Further, in step 5, first, the Poisson noise image is input into the inference sub-network based on the U-Net network, then the output of the inference sub-network is sampled using the resampling method, finally the sampled information is spliced with the noise image in the usual dimension to form a tensor with two channels, and then input into the denoiser network, and finally the denoised image is obtained, the loss value between the denoised image and the clean image is calculated, and the Adam optimizer is used for optimization.

[0045] Further, step 5 also includes using indicators PSNR and SSIM to evaluate the denoising.

[0046] Compared with the prior art, the advantages and beneficial effects of the present application are:

[0047] The present application is constructed from the Bayesian point of view, in order to improve the blind Poisson denoising ability, the variational inference method is used to infer the distribution parameters of the latent variables in the Poisson noise image, and a resampling technology based on Gamma distribution is particularly used. Combined with the inferred information, a Poisson denoising residual network with multiple sub-decoders is designed, which successfully avoids the overfitting problem. The present application combines the advantages of model-driven and data-driven methods, has strong interpretability and denoising performance, strong generalization ability and extremely competitive blind denoising ability. The method can realize blind Poisson denoising tasks in different situations, and is suitable for denoising tasks in single photon imaging, astronomical imaging, computer tomography and other imaging applications based on photon counting. BRIEF DESCRIPTION OF DRAWINGS

[0048] Figure 1 It is the overall network structure diagram for blind Poisson denoising, wherein the upper half is the inference sub-network participating in variational inference, and the lower half is the denoising sub-network for denoising.

[0049] Figure 2is the specific structure diagram for the denoising subnetwork, wherein Figure 2 (a) is the specific structure of the sub-decoder, Figure 2 (b) is the contrast-based channel attention used.

[0050] Figure 3 is the test result diagram of the embodiment in a weak light environment (peak value of 20), wherein (a) is the original noise-free image, (b) is the Poisson noise image, (c) is the result of FOTV, (d) is the result of Poisson2Sparse, (e) is the result of I+VST+BM3D, (f) is the result of DnCNN, (g) is the result of CADNet, (h) is the result of MC 2 RNet, (i) is the result of SNLDP, and (j) is the result of the method proposed in the application (Proposed).

[0051] Figure 4 is the test result diagram of blind denoising of the embodiment, wherein (a) is the original noise-free image, (b) is the Poisson noise image, (c) is the result of FOTV, (d) is the result of Poisson2Sparse, (e) is the result of I+VST+BM3D, (f) is the result of DnCNN, (g) is the result of CADNet, (h) is the result of MC 2 RNet, (i) is the result of SNLDP, and (j) is the result of the method proposed in the application (Proposed). DETAILED DESCRIPTION

[0052] In order to facilitate those skilled in the art to understand and implement the application, the application will be further described in detail below in combination with the drawings and embodiments. It should be understood that the embodiments described herein are only used to explain the application and do not limit the application.

[0053] The application mainly aims at blind Poisson denoising and improves the denoising ability in a weak light environment. In the Bayesian variational inference framework, we derive the variational evidence lower bound of the Poisson distribution, parameterize the variational posterior with a neural network, effectively combine variational inference with deep learning, and thus can effectively infer the statistical information of Poisson noise. Adding the Poisson noise statistical information to the denoising network proposed by us effectively improves the Poisson denoising performance and generalization ability, and the blind Poisson denoising effect is particularly significant.

[0054] APPENDIX Figure 1 is the blind Poisson denoising deep network structure based on variational inference of the embodiment, and APPENDIX Figure 2 is the specific network structure. The embodiment provides a blind Poisson denoising deep convolutional network method based on variational inference to realize Poisson denoising under different signal-to-noise ratios, specifically including the following steps:

[0055] Step 1: In the variational inference framework, the objective function of variational inference is constructed based on the Poisson distribution. The specific implementation includes the following sub-steps:

[0056] Step 1.1: y, x represent the image containing Poisson noise and the clean noise-free image respectively, and the imaging process of Poisson noise is represented as follows:

[0057]

[0058] where P(y|x) is the degradation model of Poisson noise, and N is the total number of image pixels;

[0059] Step 1.2: Use the variable λ to represent the hidden variable, and according to the statistical properties of the Poisson distribution, impose a Gamma conjugate prior on the hidden variable:

[0060]

[0061] where represents the Gamma conjugate prior, and s i ,γ i represent the parameters of the Gamma conjugate prior.

[0062] Step 1.3: Based on formula (2), the variational posterior form of the hidden variable λ is represented as follows.

[0063]

[0064] where q(λ|y) represents the variational posterior of the hidden variable λ, and α i (y i ;Θ),β i (y i ;) represent two inference functions of the inference subnetwork Θ.

[0065] Step 1.4: In the variational framework, the objective optimization function of the Poisson distribution is constructed:

[0066] q * (λ)=argmin q(λ)∈Q D KL (q(λ|y;Θ)||p(λ|y)) (4)

[0067] where D KL represents the Kullback-Leibler divergence, q(λ|y; Θ) represents the variational posterior distribution, p(λ|y) is the true posterior distribution, Q represents the variational distribution family, and q(γ) is considered as a variable. When q(γ) satisfies the KL divergence with the true posterior distribution, use q*(γ) to represent it;

[0068] Step 2: Solve the objective function of the variational inference to obtain the specific form of the evidence lower bound. The specific implementation includes the following sub-steps:

[0069] Step 2.1: Replace the objective optimization function (4) by an equivalent replacement to obtain the following form:

[0070]

[0071] where p(y) is the marginal likelihood;

[0072] Step 2.2: Formally decompose the evidence lower bound in equation (5) Based on equations (1), (2), and (4), the specific form of the evidence lower bound can be obtained.

[0073]

[0074] where Γ(·), Ψ(·) represent the gamma and digamma functions, respectively.

[0075] Step 3: Construct and use the reasoning subnetwork to parameterize the variational distribution, and complete the variational inference to obtain the optimal solution. The specific implementation includes the following sub-steps:

[0076] Step 3.1: Construct the reasoning subnetwork to complete the variational inference task, as shown in the upper half of Figure 1 The reasoning subnetwork is a U-Net network with a 4-layer structure, where all convolution kernel sizes are 3x3, the step size is 1x1, and the padding is 1x1. The U-Net network is a symmetric network with an encoder-decoder structure. In the encoder part, three convolution operations are performed at each layer, and LeakyReLU is used as the activation function after each convolution. When reaching the next layer, mean pooling is used to complete the downsampling operation, where the pooling kernel size is 2x2 and the step size is 2x2. In the decoder part, each layer first performs an inverse convolution operation, with an inverse convolution kernel size of 2x2 and a step size of 2x2. After inverse convolution, two convolution operations are performed, with the same settings as the decoder part. The encoder and decoder are connected through a skip connection.

[0077] Step 3.2: The input of the reasoning subnetwork is the Poisson noise image, and the output is the two parameters of the variational posterior, α, β.

[0078] Step 3.3: Based on the evidence lower bound obtained by the variational inference, construct the loss function of the reasoning subnetwork:

[0079]

[0080] Step 4: Sample the inference results of the variational inference by the Gamma-based reparameterization technique to obtain the Poisson noise statistics, and construct a denoiser network to realize noise removal. The specific implementation includes the following sub-steps:

[0081] Step 4.1: The purpose of variational inference is to infer the parameters of the variational posterior. In order to obtain the Poisson statistics, sampling needs to be performed from the variational posterior. However, ordinary random sampling will result in sampling results without gradients, so that optimization cannot be performed. Therefore, a resampling technique needs to be used for sampling. Based on the statistical characteristics of the Gamma distribution, the resampling method is represented as follows:

[0082]

[0083] Step 4.2: A denoiser network using Poisson noise statistics is constructed, as shown in the lower half of Figure 1 . The specific results of the Sub-Decoder are shown in Figure 2 (a) and 2(b). The denoiser network is a deep network structure integrating residual learning and skip connection. Three three-layer sub-decoders are used to form a residual network, as shown in Figure 2 (a). In each sub-decoder, convolution, skip connection and deconvolution are used to form a symmetric network, and each layer of the network is connected through a channel attention based on contrast perception. The convolution operations in the first, second and third layers of the sub-decoder use 64, 128 and 256 convolution kernels, respectively. The mean pooling layer is used to connect between each layer in the left half, and the size of the pooling layer is 2x2. The deconvolution layer is used to connect between each layer in the right half, and the size of the deconvolution layer is 2x2. The specific structure of the channel attention is shown in Figure 2 (b), which uses two convolution layers, each with a convolution kernel size of 1x1. The first convolution layer uses ReLU as the activation function, and the second convolution layer uses Sigmoid as the activation function. The convolution kernel sizes of the three Sub-Decoders in the denoiser network are 7x7, 5x5 and 3x3, respectively, and the sizes and steps of the pooling kernel and the deconvolution kernel are both 2x2. In the sub-decoder, after each convolution operation, batch normalization is performed, and then LeakyReLU is used as the activation function.

[0084] Step 5: Train the above network (the network composed of the inference sub-network and the denoiser network) using clean image-Poisson noise image pairs. After the loss function (formula 7) is stable and does not decrease, the trained model is obtained. The trained model network parameters and test data are used to obtain the experimental results of Poisson denoising and blind denoising, and compared with other algorithms. The specific implementation includes the following sub-steps:

[0085] Step 5.1: Training preparation, the test results obtained are compared with each of the comparative methods in terms of visual and objective evaluation indicators. In this embodiment, a total of 4000 high-definition images are used as training data, the noise peak values are {1, 2, 4, 8, 20}, the size of the training simulation image block is 128x128, the size of the reference image block is 128x128, the learning rate is set to 0.0001, and the batch size is set to 64.

[0086] Step 5.2: Training network, in the training process, each time the data is loaded, a peak value is randomly selected to synthesize Poisson noise data. In the training process, whether to use GPU acceleration is set according to the device used. First, the noise image is input into the inference subnetwork based on the U-Net network, then the output of the inference subnetwork is sampled using the resampling method described above, finally the sampled information is spliced with the noise image in the normal dimension to form a tensor with two channels, and then input into the denoising subnetwork, and finally the denoised image is obtained. The loss value between the denoised image and the real noise-free image is calculated, and the Adam optimizer is used for optimization, and the parameters of the Adam optimizer are set to betas=[0.99, 0.999], eps=1e-8.

[0087] Step 5.3: Test network, load the trained model and restore the model parameters. The test image is input into the network model, the denoised image is saved, and the PSNR and SSIM indicators are calculated.

[0088] Step 5.4: The comparative methods are FOTV, Poisson2Sparse, I+VST+BM3D, DnCNN, CADNet, and MC 2 RNet. The objective evaluation indicator is the peak signal-to-noise ratio (PSNR), and the BSD68 dataset is used for comparison, and the average results are shown in Tables 1 and 2. The visual graphs are shown in Figures 3, 4, and 5. Figure 3

[0089] Table 1 Comparison of average PSNR (dB) of different methods under BSD68 dataset (ideal value: +∞)

[0090]

[0091] Table 2 Comparison results of blind Poisson denoising experiment, PSNR (ideal value: +∞)

[0092] FOTV PURE-LET SNLDP IVST+BM3D DnCNN CADNet MCRNet Proposed 16.79 23.67 24.63 24.75 25.24 23.83 25.06 25.57

[0093] ​It can be seen that the method proposed in this paper can efficiently blind Poisson denoise based on variational inference and deep neural network. The combination of variational inference and neural network can quickly and accurately infer the Poisson noise statistical information. Using the inferred Poisson noise information, the generalization ability of the denoising network is effectively enhanced, thereby achieving excellent blind Poisson denoising performance.

[0094] It should be understood that parts not elaborated in the specification are all prior art.

[0095] It should be understood that the above description of the embodiments is more detailed and should not be considered as a limitation on the scope of patent protection of the present application. Those skilled in the art can make substitutions or modifications without departing from the scope of protection claimed by the present application, which falls within the scope of protection of the present application. The scope of protection of the present application shall be subject to the appended claims.

Claims

1. A blind Poisson denoising deep learning method based on variational inference, characterized in that, Comprising the following steps: Step 1, in the variational inference framework, a target function of variational inference is constructed based on Poisson distribution; The specific implementation mode of step 1 is as follows: Step 1.1, Note Let I and I0denote the image with Poisson noise and the clean noise-free image, respectively. The imaging process with Poisson noise is represented as follows: (1) Wherein, P(y|x) is the degradation model of Poisson noise, N is the total number of image pixels; Step 1.2, use of variables represents a latent variable, a Gamma conjugate prior is imposed on the latent variable according to the statistical properties of the Poisson distribution: (2) wherein, denotes a Gamma conjugate prior, denotes a parameter of the Gamma conjugate prior; Step 1.3, On the basis of equation (2), the variational posterior form of the hidden variable is given by (3) wherein, represents a latent variable the variational posterior, represents two inference functions of an inference subnetwork ; Step 1.4, in the variational framework, the target optimization function of Poisson distribution is constructed: (4) in, Indicates the Kullback-Leibler divergence. Describes the variational posterior distribution. It is the true posterior distribution, and Q represents the family of variational distributions. Consider it as a variable, when When the KL divergence with the true posterior distribution is minimized, use To indicate; Step 2, the target function of variational inference is solved to obtain the specific form of the evidence lower bound; Step 3, construct and use the reasoning subnetwork to parameterize the variational distribution, and complete the variational inference to obtain the optimal solution; Step 4, the reasoning result of variational inference is sampled by Gamma-based reparameterization technology to obtain the Poisson noise statistical information, construct the denoising subnetwork, and combine the variational Poisson noise statistical information to denoise the Poisson noise image, and finally obtain the denoised result; Step 5, use clean image-Poisson noise image pair to train the reasoning subnetwork and denoising subnetwork, and obtain the trained model after the loss function is stable, and use the trained model network parameters and test data to obtain the Poisson denoising and blind denoising results.

2. The variational inference based blind Poisson denoising deep learning method of claim 1, wherein: The specific implementation mode of step 2 is as follows: Step 2.1, by equivalent replacement of the target optimization function (4), the following form is obtained: (5) wherein is the marginal likelihood; Step 2.

2. Derivation of the lower bound of evidence in equation (5) Performing formal decomposition, based on equations (1), (2) and (4), the specific form of the lower bound of evidence can be obtained; (6) wherein denote the gamma and digamma functions, respectively.

3. The variational inference based blind Poisson denoising deep learning method of claim 2, wherein: The specific implementation mode of step 3 is as follows: Step 3.1, constructing the inference subnetwork to complete the variational inference task, the inference subnetwork is a U-Net network with a 4-layer structure, wherein all the convolution kernel sizes are , the step is , and the padding is ; Step 3.2, the input of the reasoning subnetwork is the Poisson noise image, and the output is the two parameters of the variational posterior; Step 3.3, according to the evidence lower bound obtained by reasoning in the variational framework, the loss function of the reasoning subnetwork is constructed: (7)。 4. The variational inference based blind Poisson denoising deep learning method of claim 3, wherein: The U-Net network is a symmetrical network with an encoder-decoder structure. In the encoder part, each layer performs three convolution operations, and after each convolution, LeakyReLU is used as the activation function. When reaching the next layer, mean pooling is used to complete the down-sampling operation, wherein the size of the pooling kernel is , and the step is . In the decoder part, each layer first performs an inverse convolution operation, the size of the inverse convolution kernel is , and the step is . After the inverse convolution, two convolution operations are performed, and the convolution operation settings are consistent with the decoder part. The encoder and the decoder are connected through a jump connection.

5. The variational inference based blind Poisson denoising deep learning method of claim 3, wherein: The specific implementation mode of step 4 is as follows: Step 4.1, the purpose of variational inference is to infer the parameters of variational posterior, in order to obtain the Poisson statistical information, sampling is needed from the variational posterior, however, ordinary random sampling will result in no gradient of the sampling result, so that optimization cannot be performed; Therefore, resampling technology needs to be used for sampling, based on the statistical characteristics of Gamma distribution, the resampling method is represented as follows: (8) Step 4.2, a denoising subnetwork using Poisson noise statistical information is constructed, which is a deep network structure integrating residual learning and skip connection, using three three-layer subdecoders to form a residual network, in each subdecoder, convolution, skip connection and deconvolution are used to form a symmetric network, and each layer is connected through channel attention based on contrast perception.

6. The variational inference based blind Poisson denoising deep learning method of claim 5, wherein: The convolution operations in the first, second and third layers of the subdecoder use 64, 128 and 256 convolution kernels respectively.

7. The variational inference based blind Poisson denoising deep learning method of claim 5, wherein: The convolution kernel sizes of the three subdecoders are 7x7, 5x5 and 3x3 respectively, and the pool kernel and deconvolution kernel sizes and steps are both 2x2.

8. The variational inference based blind Poisson denoising deep learning method of claim 1, wherein: In step 5, first, input the Poisson noise image into the reasoning subnetwork based on the U-Net network, then sample the output of the reasoning subnetwork using the resampling method, finally, the sampled information is spliced with the noise image in the normal dimension to form a tensor with two channels, and then input into the denoising subnetwork, finally get the denoised image, calculate the loss value between the denoised image and the clean image, and use the Adam optimizer for optimization.

9. The variational inference based blind Poisson denoising deep learning method of claim 1, wherein: Step 5 also includes evaluating the denoising using the metrics PSNR and SSIM.

Images