A self-calibration optimization method for image denoising utilizing convexity

Abstract

This invention discloses a self-calibrated image denoising method utilizing convexity, belonging to the field of image processing technology. It is used for image denoising under conditions of unknown noise levels. The method includes generating a pair of noisy images and images with secondary noise, constructing a conditional denoising network; deriving the denoising estimator and calculating the variance of the estimation error based on the linear minimum mean square error estimation theory, and introducing a noise ratio parameter for normalization analysis; utilizing the convexity of the error variance function, determining the optimal synthetic noise variance through differentiation or a ternary search algorithm; calibrating the optimal synthetic noise variance using an empirical formula; training and executing the denoising network using the calibrated optimal variance, and outputting the final denoised image. This invention, by constructing a robust training process under conditions of unknown noise levels, can effectively learn the statistical characteristics of noise, overcoming the limitation of traditional methods that require prior knowledge of the noise level.

Description

Technical Field

[0001] This invention discloses an image denoising method that utilizes convexity for self-calibration optimization, belonging to the field of image processing technology. Background Technology

[0002] Image denoising is a fundamental and crucial task in image processing, aiming to recover a clean original image from a noisy observation. Denoising methods with unknown noise levels include supervised methods that do not require clean images. Noise2Noise, a pioneer in this field, proved that training can be performed using only two independent noisy images. However, its requirement for multiple independent observations of the same scene is difficult to meet in many practical scenarios. Blind spot methods, represented by Noise2Void, mask pixels in the input image and train the network to predict the masked center pixel using surrounding pixels, thus achieving training with only a single noisy image. Subsequent works such as Noise2Self, Self2Self, and Neighbor2Neighbor have extended and improved this paradigm. However, these methods inevitably lose some image information during training, making their performance difficult to match supervised methods. Deep image priors, such as the DIP method, have found that untrained convolutional neural network structures inherently possess inductive biases to the statistical properties of natural images, which can be used for denoising. However, this method is often not optimal, and determining the premature stopping point is difficult. General-purpose deep learning models, such as DnCNN-B, are trained on datasets with mixed noise levels to achieve a general model capable of handling different noise levels. While these models are highly flexible, their performance is generally inferior to models specifically trained for known noise levels due to a lack of specificity for particular noise levels. Noise level estimation methods: Another approach is to directly estimate the noise level from the noisy image to aid in denoising. For example, Foi et al. identified smooth regions through wavelet segmentation and used local statistics for parameter estimation; Pimpalkhute et al. combined the frequency separation capability of DWT with spatial edge localization to eliminate edge interference; and Rakhshanfar et al. used intensity-variance homogeneity classification for weighted noise peak estimation. These traditional estimation methods are effective to some extent, but their accuracy still needs improvement. Inaccurate noise level estimation directly limits the performance ceiling of subsequent denoising methods.

[0003] In summary, existing denoising methods suffer from the following main problems: methods with known noise levels are limited by the unpredictability of noise levels in practical applications; while methods with unknown noise levels suffer from information loss issues due to blind spot methods, insufficient performance of general models, or limited accuracy of traditional noise estimation methods, all of which prevent them from achieving denoising results comparable to methods with ideal known noise levels. Therefore, how to accurately estimate noise levels when they are completely unknown, and use this to drive the denoising model to achieve performance close to that of methods with known noise levels, is a pressing technical challenge in this field. Summary of the Invention

[0004] The purpose of this invention is to provide an image denoising method that utilizes convexity for self-calibration optimization, in order to solve the problem in the prior art of how to accurately estimate the noise level and drive the denoising model to achieve performance close to that of methods with known noise levels when the noise level is completely unknown.

[0005] A self-calibration optimization image denoising method utilizing convexity includes:

[0006] S1. Input a noisy image, generate a second-added noisy image, construct a conditional denoising network based on the variance of the second-added noisy image and the synthesized noise, and optimize the network parameters by minimizing the L2 norm distance between the network output and the noisy image.

[0007] S2. Based on the linear minimum mean square error estimation theory, the denoising estimator is derived and the estimation error variance is calculated. The noise ratio parameter is introduced for normalization analysis and a normalization expression is established. The optimal ratio between the optimal synthesized noise variance and the original noise variance is determined by differentiating the normalization expression.

[0008] S3. Using the normalized expression, the optimal ratio between the optimal synthesized noise variance and the original noise variance is obtained by differentiation. The optimal synthesized noise variance is then determined by the ternary search algorithm.

[0009] S4. Calibrate the optimal synthetic noise variance using an empirical calibration formula;

[0010] S5. Based on the DnCNN model architecture, a conditional mechanism is introduced to construct the DnCNN model as a conditional denoising network. The conditional denoising network is trained using a residual learning strategy to obtain the denoising network. Training pairs are constructed on a general dataset using the calibrated optimal synthetic noise variance and input into the denoising network. The denoising network is trained using a residual learning strategy. The noisy image is input into the trained denoising network to generate the final denoised image.

[0011] S1 includes, S1.1, the input noisy image. :

[0012] ;

[0013] In the formula, For a clean image; The original noise, Follow the mean The sum and variance are Gaussian distribution; This represents the original noise variance;

[0014] Synthetic noise is generated using a computer's pseudo-random number generator. , Follows a mean of 0 and a variance of Gaussian distribution, The variance of the synthesized noise;

[0015] Will injection Generate a second-order noisy image :

[0016] .

[0017] S1 includes S1.2, constructing a conditional denoising network. ,Will and As network input, As the training objective, minimize the network output and Optimize network parameters using the L2 norm distance between them. :

[0018] ;

[0019] In the formula, For loss function, As the expected value, The square of the L2 norm. for The variance.

[0020] S2 includes S2.1, based on observation. Constructing a pair The linear minimum mean square error estimator, derived The final denoising estimate :

[0021] ;

[0022] right Calculate the variance of the estimation error :

[0023] ;

[0024] In the formula, For variance, for The variance.

[0025] S2 includes S2.2, the introduced noise ratio. ,Will Represented as Multiples of:

[0026] ;

[0027] Will Substitution Get about normalized expression :

[0028] .

[0029] S2 includes, S2.3, and... Differentiation determines the critical point:

[0030] ;

[0031] In the formula, The differential symbol;

[0032] make Solving the equation yields the critical point. , hour It is the minimum value;

[0033] and The optimal ratio is:

[0034] .

[0035] S3 includes, S3.1, according to and The optimal ratio, using Convexity, optimized using a ternary search algorithm ;

[0036] Initialize search interval Includes the original optimal synthesized noise variance Set a convergence threshold. ,and ;

[0037] When the interval width During this process, steps S3.1.1, S3.1.2, and S3.1.3 are executed repeatedly.

[0038] S3.1.1, will Trisect the sample into three equal parts and calculate the two interior points. , :

[0039] ;

[0040] ;

[0041] S3.1.2, respectively in and Calculation The result is and ;

[0042] S3.1.3, according to and The comparison results narrow down the interval range: if ,renew ;like ,renew ;

[0043] S3 includes, S3.2, when When the loop terminates, the result is obtained. :

[0044] ;

[0045] set up = Used for image denoising.

[0046] S4 includes, for Adjusted using empirical calibration formulas:

[0047] ;

[0048] In the formula, The optimal variance of synthesized noise after calibration.

[0049] S5 includes S5.1, which introduces a conditional mechanism to construct the DnCNN model based on the DnCNN model architecture. Training using residual learning strategies After training, a denoising network is obtained. Determined by steps S3.1 and S3.2;

[0050] The DnCNN model consists of an input layer, an output layer, and multiple consecutive composite layers with the same structure. Each composite layer includes one convolutional layer, one batch normalization layer, and one ReLU function layer. The input layer takes in the data. The composite layer first performs convolution operations through the convolutional layer, outputting a set of feature maps. Then, it is input into the batch normalization layer to perform batch normalization on the feature maps. Finally, it is input into the ReLU function layer, which applies the ReLU activation function to each value output by the batch normalization layer. The output layer outputs the model's processing result.

[0051] S5 includes, S5.2, utilizing In constructing a training pair input denoising network on a general dataset, a residual learning strategy is used for training.

[0052] Will The trained denoising network is input into the denoising network, and the final denoised image is directly output through the forward propagation of the denoising network.

[0053] Compared with existing technologies, this invention has the following advantages: By constructing a robust training process under unknown noise levels, this invention can effectively learn the statistical characteristics of noise, overcoming the limitation of traditional methods that require prior knowledge of the noise level; through variance trend analysis and theoretical derivation, the theoretical conditions for the optimal synthetic noise variance are determined, providing theoretical guidance for noise level estimation; utilizing the convexity of the estimation error variance, a ternary search algorithm is used to quickly find the optimal noise level, improving estimation efficiency and accuracy; and an empirical calibration mechanism is introduced to effectively compensate for accuracy deviations during training, improving the practicality and robustness of the method. Attached Figure Description

[0054] Figure 1 The overall flowchart shows the image denoising method that utilizes convexity for self-calibration optimization.

[0055] Figure 2 The network architecture for DnCNN;

[0056] Figure 3 This is a variance trend diagram of the method of the present invention applied to the SET12 dataset;

[0057] Figure 4 This is a variance trend diagram of the method of the present invention applied to the BSD68 dataset. Detailed Implementation

[0058] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention are described clearly and completely below. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0059] A self-calibration optimization image denoising method utilizing convexity includes:

[0060] S1. Input a noisy image, generate a second-added noisy image, construct a conditional denoising network based on the variance of the second-added noisy image and the synthesized noise, and optimize the network parameters by minimizing the L2 norm distance between the network output and the noisy image.

[0061] S2. Based on the linear minimum mean square error estimation theory, the denoising estimator is derived and the estimation error variance is calculated. The noise ratio parameter is introduced for normalization analysis and a normalization expression is established. The optimal ratio between the optimal synthesized noise variance and the original noise variance is determined by differentiating the normalization expression.

[0062] S3. Using the normalized expression, the optimal ratio between the optimal synthesized noise variance and the original noise variance is obtained by differentiation. The optimal synthesized noise variance is then determined by the ternary search algorithm.

[0063] S4. Calibrate the optimal synthetic noise variance using an empirical calibration formula;

[0064] S5. Based on the DnCNN model architecture, a conditional mechanism is introduced to construct the DnCNN model as a conditional denoising network. The conditional denoising network is trained using a residual learning strategy to obtain the denoising network. Training pairs are constructed on a general dataset using the calibrated optimal synthetic noise variance and input into the denoising network. The denoising network is trained using a residual learning strategy. The noisy image is input into the trained denoising network to generate the final denoised image.

[0065] S1 includes, S1.1, the input noisy image. :

[0066] ;

[0067] In the formula, For a clean image; The original noise, Follow the mean The sum and variance are Gaussian distribution; This represents the original noise variance;

[0068] Synthetic noise is generated using a computer's pseudo-random number generator. , Follows a mean of 0 and a variance of Gaussian distribution, The variance of the synthesized noise;

[0069] Will injection Generate a second-order noisy image :

[0070] .

[0071] S1 includes S1.2, constructing a conditional denoising network. ,Will and As network input, As the training objective, minimize the network output and Optimize network parameters using the L2 norm distance between them. :

[0072] ;

[0073] In the formula, For loss function, As the expected value, The square of the L2 norm. for The variance.

[0074] S2 includes S2.1, based on observation. Constructing a pair The linear minimum mean square error estimator, derived The final denoising estimate :

[0075] ;

[0076] right Calculate the variance of the estimation error :

[0077] ;

[0078] In the formula, For variance, for The variance.

[0079] S2 includes S2.2, the introduced noise ratio. ,Will Represented as Multiples of:

[0080] ;

[0081] Will Substitution Get about normalized expression :

[0082] .

[0083] S2 includes, S2.3, and... Differentiation determines the critical point:

[0084] ;

[0085] In the formula, The differential symbol;

[0086] make Solving the equation yields the critical point. , hour It is the minimum value;

[0087] and The optimal ratio is:

[0088] .

[0089] S3 includes, S3.1, according to and The optimal ratio, using Convexity, optimized using a ternary search algorithm ;

[0090] Initialize search interval Includes the original optimal synthesized noise variance Set a convergence threshold. ,and ;

[0091] When the interval width During this process, steps S3.1.1, S3.1.2, and S3.1.3 are executed repeatedly.

[0092] S3.1.1, will Trisect the sample into three equal parts and calculate the two interior points. , :

[0093] ;

[0094] ;

[0095] S3.1.2, respectively in and Calculation The result is and ;

[0096] S3.1.3, according to and The comparison results narrow down the interval range: if ,renew ;like ,renew ;

[0097] S3 includes, S3.2, when When the loop terminates, the result is obtained. :

[0098] ;

[0099] set up = Used for image denoising.

[0100] S4 includes, for Adjusted using empirical calibration formulas:

[0101] ;

[0102] In the formula, The optimal variance of synthesized noise after calibration.

[0103] S5 includes S5.1, which introduces a conditional mechanism to construct the DnCNN model based on the DnCNN model architecture. Training using residual learning strategies After training, a denoising network is obtained. Determined by steps S3.1 and S3.2;

[0104] The DnCNN model consists of an input layer, an output layer, and multiple consecutive composite layers with the same structure. Each composite layer includes one convolutional layer, one batch normalization layer, and one ReLU function layer. The input layer takes in the data. The composite layer first performs convolution operations through the convolutional layer, outputting a set of feature maps. Then, it is input into the batch normalization layer to perform batch normalization on the feature maps. Finally, it is input into the ReLU function layer, which applies the ReLU activation function to each value output by the batch normalization layer. The output layer outputs the model's processing result.

[0105] S5 includes, S5.2, utilizing In constructing a training pair input denoising network on a general dataset, a residual learning strategy is used for training.

[0106] Will The trained denoising network is input into the denoising network, and the final denoised image is directly output through the forward propagation of the denoising network.

[0107] The core of this invention lies in accurately estimating the noise level when it is unknown, and achieving high-quality image denoising through theoretical analysis and algorithm optimization. Figure 1 As shown, noise is first input in stage one. Image, with added noise The image is used to obtain the estimation results for the noisy image. Then, the noise level is estimated using the variance convexity property; in stage two, the noise is first input. The image, combined with stage 1 paired noise, is input into the deep learning network to obtain a clean image. The specific implementation process is as follows:

[0108] Robust training process construction, during the training phase, for the input , Unknown. Random sampling synthesis and generate The denoising network adopts a conditional form. ,pass This is used to train the network, enabling it to implicitly learn the statistical properties of noise. During the training process, the network parameters... As weights and biases within the model, continuous optimization is achieved through four stages: forward propagation, loss calculation, back propagation, and parameter update. During the forward propagation stage... As fixed weights and bias parameters, and Mapped to predicted output Subsequently, the loss function is calculated... and The L2 norm distance between them is used to evaluate the current... Performance; during backpropagation, the loss function is calculated using the chain rule for each... The gradient explicitly indicates the direction and magnitude of adjustment for each parameter; finally, in the parameter update phase, the optimization algorithm adjusts the parameters based on the gradient information. Iterative updates are performed to continuously adjust the system along the direction of loss reduction. Through this repeated training cycle, The network gradually converges to the optimal configuration that can accurately model the relationship between input and output, thereby enabling the network to effectively estimate the target signal from noisy observations.

[0109] Analysis of variance trend, derived through theoretical derivation. The corresponding estimation error variance is .Will Represented as Analysis of variance term Follow The changing trend can be observed by differentiation when... hour Minimum, that is The time is the optimal condition.

[0110] Triple search optimization utilizes the convexity of the estimation error variance to find the optimal value using a ternary search algorithm. Algorithm initializes search interval In each iteration, two intermediate points are calculated. and Evaluate the variance at these two points, and update the search interval based on the variance until the interval length is less than a preset threshold. , ,return .

[0111] An empirical calibration mechanism is introduced to address the accuracy bias issue in actual training. This empirical calibration formula is derived from experimental results on a general dataset using the least squares method. The general dataset includes SET12, BSD68, and Train400. The general dataset is input into a denoising network for prediction. The prediction result is multiplied by 255, and then the relationship between the true and predicted values ​​is calculated using the least squares method, resulting in the empirical calibration formula. This empirical calibration formula effectively improves the accuracy of noise level estimation.

[0112] The final denoising implementation, based on optimized and calibrated noise levels, uses a denoising network trained on a general dataset, which is then used to perform final denoising on the input image. Specifically, the denoising network is trained using the general datasets Train400 and DFWB (DIV2K, Flickr2K, WED, and BSD). (Additionally...) The final optimized value determined through step S3 (ternary search optimization) and step S4 (empirical calibration) The training process was implemented using the DnCNN model within the PyTorch deep learning framework. DnCNN is a classic deep learning-based image denoising model whose core idea is to directly learn the end-to-end mapping from noisy images to clean images through a deep convolutional network. This model employs a residual learning strategy; instead of directly outputting the denoised image, the network predicts the residual (noise component) between the noisy and clean images, thus simplifying the learning objective and improving training efficiency and stability. The network structure is as follows: Figure 2As shown, the model consists of 15 stacked convolutional layers, Rectified Linear Unit (ReLU) activation functions, and batch normalization layers. This design effectively accelerates the model convergence process and enhances generalization ability. Specific training conditions and parameter settings are as follows: the Adam optimizer is used to minimize the mean squared error loss function, with a training learning rate of 0.001 on the dataset; the training batch size is 128; and the training epochs are set to 180. A fixed noise level was determined through previous optimization. Training under specific conditions allows the network to focus on learning the optimal denoising mapping for that particular noise level, thereby enabling the processing of noisy images. High-quality reconstruction.

[0113] Experimental results show that the present invention has been fully validated on multiple standard test datasets such as CBSD68, Kodak24, McMaster, and Urban100, as shown in Table 1:

[0114] Table 1. Quantitative Results of the Image Denoising Method Optimized by Self-Calibration Using Convexity

[0115] ;

[0116] The objective data in Table 1 clearly shows that this invention significantly outperforms other methods with unknown noise levels in both the PSNR (Peak Signal-to-Noise Ratio) and SSIM (Structural Similarity Index), demonstrating its superior overall performance. This indicates that while effectively suppressing noise, this invention better preserves the details and texture structure of the image, avoiding the problems of over-smoothing or loss of detail common in existing methods.

[0117] The core of this invention lies in the utilization of the convexity law of the estimation error variance. Figure 3 and Figure 4 The figure clearly demonstrates a high degree of agreement between the theoretically derived variance curve and the actual measured values, verifying the correctness of the theoretical analysis in this invention. This lays a solid theoretical foundation for subsequent optimization algorithms. Based on this theory, this invention has achieved a breakthrough in the accuracy of noise level estimation. To normalize the standard deviation of synthesized noise The values ​​are converted to the standard 8-bit image pixel value range (0-255) to represent their actual intensity. A comparison of the method of this invention with variance-finding methods is shown in Table 2:

[0118] Table 2. Quantitative experimental comparison between the method of this invention and the variance finding method.

[0119] ;

[0120] Table 2 shows that, regardless of whether it is on the SET12 or BSD68 dataset, the noise level estimation error of the present invention is significantly lower than that of the traditional estimation method when facing different levels of noise, demonstrating its higher estimation accuracy and stronger robustness. To estimate the noise standard deviation, %E represents the percentage error:

[0121] ;

[0122] CV-DnS identifies smooth regions through wavelet segmentation and uses local statistics for parameter estimation; DWT-NE combines the frequency separation capability of DWT with spatial edge localization to eliminate edge interference; and IVHC uses intensity-variance homogeneity classification for weighted noise peak estimation.

[0123] Further comparative experiments highlight the application value of this invention, as shown in Table 3:

[0124] Table 3. Quantitative experimental comparison between the method of the present invention and methods with unknown noise levels.

[0125] ;

[0126] On the SET12 and BSD68 datasets, the denoising performance of this invention comprehensively surpasses a series of mainstream methods with unknown noise levels, including Noise2Void, Noise2Self, Neighbor2Neighbor, and DnCNN-B, across all metrics. This demonstrates that this invention, while removing complex noise, has a stronger ability to preserve fine structures, inherently avoiding the information loss and performance bottlenecks caused by mask-based methods.

[0127] Most importantly, this invention successfully improves the denoising performance under unknown noise levels to a level close to that under known conditions, as shown in Table 4:

[0128] Table 4. Quantitative experimental comparison between the method of the present invention and methods with known noise levels.

[0129] ;

[0130] The comparative data in Table 4 show that when using the noise level estimated by this invention to drive well-known denoising methods such as BM3D, DnCNN-S, Noisier2Noise, and Restormer, the results are very similar to those obtained using the actual noise level (the maximum difference in PSNR is only 0.05 dB, and the maximum difference in SSIM is only 0.003). This fully demonstrates that the noise level estimation accuracy provided by this invention is sufficient to meet or even replace the actual noise level, enabling various denoising methods to achieve near-optimal performance under known conditions even in unknown scenarios.

[0131] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An image denoising method using convexity self-calibration optimization, characterized in that, The method comprises the following steps: S1, inputting a noise image, generating a secondary noise-added image, constructing a conditional denoising network based on the secondary noise-added image and a synthetic noise variance, and optimizing network parameters by minimizing the L2 norm distance between the network output and the noise image; S2, based on the linear least mean square error estimation theory, deducing a denoising estimator and calculating an estimation error variance, introducing a noise ratio parameter for normalization analysis and establishing a normalization expression, and determining the optimal ratio between the optimal synthetic noise variance and the original noise variance by derivation of the normalization expression; S3, using the normalization expression to obtain the optimal ratio between the optimal synthetic noise variance and the original noise variance by derivation, and determining the optimal synthetic noise variance by a ternary search algorithm; S4, calibrating the optimal synthetic noise variance by an empirical calibration formula; S5, based on the DnCNN model architecture, introducing a conditional mechanism to construct the DnCNN model into a conditional denoising network, training the conditional denoising network using a residual learning strategy, obtaining a denoising network, constructing training pairs on a general data set using the calibrated optimal synthetic noise variance and inputting the denoising network, training the denoising network using a residual learning strategy, inputting a noise image into the trained denoising network, and generating a final denoised image.

2. The method of image denoising with self-calibration optimization using convexity according to claim 1, wherein, S1 comprises, S1.1, inputting a noise image : ； wherein is a clean image; is a raw noise, obeys a Gaussian distribution with mean and variance ; and is a raw noise variance; Synthetic noise is generated using a computer's pseudo-random number generator. , Follows a mean of 0 and a variance of Gaussian distribution, The variance of the synthesized noise; Will injection Generate a second-order noisy image : 。 3. The method of image denoising with self-calibration optimization using convexity according to claim 2, wherein, S1 comprises, S1.2, constructing a conditional denoising network , wherein , and are taken as network inputs, and is taken as a training target, and the network parameters are optimized by minimizing the L2 norm distance between the network output and :​ ； wherein is a loss function, is an expected value, is a square of L2 norm, is a variance.

4. The method of image denoising with convexity-based self-calibration optimization of claim 3, wherein, S2 comprises, S2.1, based on the observations , constructing a linear minimum mean square error estimator for , deriving a final denoised estimator for :​ ； To Compute the estimated error variance : ； wherein is the variance, is the variance.

5. The method of image denoising with convexity-based self-calibration optimization according to claim 4, wherein, S2 comprises, S2.2, introducing a noise ratio , and is expressed as a multiple of : ； Substitute into to obtain the normalized expression for : 。 6. The method of image denoising with convexity-based self-calibration optimization according to claim 5, wherein, S2 comprises, S2.3, determining the critical point by derivative determination critical point: ； wherein is the differential operator; Let , solve the equation to get the critical point , when is a minimum. with the optimal ratio of 。 7. The method of image denoising with convexity-based self-calibration optimization according to claim 6, wherein, S3 comprises, S3.1, according to With the optimal ratio, using the convexity of ; initializing a search interval , including an original optimal synthesized noise variance , setting a convergence value threshold , and ; When the interval width steps S3.1.1, S3.1.2 and S3.1.3 are executed in a loop; S3.1.1, to trisect, compute two inpoints , : ； ； S3.1.2, at and and ;​​ S3.1.3、according to and the comparison result, reduce the interval range: if , update ; if , update ; S3 comprises, S3.2, when the cycle is aborted, resulting in : ； Setting = For image denoising.

8. The method of image denoising with convexity-based self-calibration optimization according to claim 7, wherein, S4 comprises, to Adjustment by empirical calibration formula: ； In the formula, Optimal synthesized noise variance after calibration.

9. The method of image denoising with convexity-based self-calibration optimization of claim 8, wherein, S5 comprises S5.1, based on the DnCNN model architecture, introducing a conditional mechanism to construct the DnCNN model into , using a residual learning strategy to train , obtaining a denoising network after training, determined by steps S3.1 and S3.2 processes; The DnCNN model comprises an input layer, an output layer and a plurality of continuous composite layers with the same structure, each composite layer comprising a convolution layer, a batch normalization layer and a ReLU function layer; the input layer inputs data; the composite layer first performs convolution operation through the convolution layer to output a group of feature maps, then inputs the batch normalization layer to perform batch normalization processing on the feature maps, and then inputs the ReLU function layer, which applies the ReLU activation function to each value output by the batch normalization layer; and the output layer outputs the model processing result.

10. The method of image denoising with convexity-based self-calibration optimization of claim 9, wherein, S5 comprises, S5.2, using In constructing the training pair input denoising network on the general data set, the residual learning strategy is adopted for training. Will The input training completed denoising network, through the forward propagation of the denoising network directly output final denoising image.

Images