Image noise reduction enhancement method based on additive Gaussian diffusion model

Abstract

The invention provides an image noise reduction enhancement method based on an additive Gaussian diffusion model, which comprises the following steps: selecting a plurality of clear images, gradually adding additive Gaussian white noise to the clear images until the images are completely polluted by the noise, and generating a series of noise images with different pollution measures; forming a training data set by the noise images with different pollution measures, the corresponding time steps and the global noise samples obeying standard normal distribution, and training a U-Net network; establishing an additive Gaussian diffusion model for reversely deriving and estimating a noise image at the previous moment; collecting a noise image to be denoised, obtaining an initial value of the noise level of the noise image, and calculating a time step number corresponding to the initial value of the noise level through a time coding function; and inputting a polluted image to be subjected to noise reduction and the corresponding time step number into the trained U-Net network, connecting a vertical additive Gaussian diffusion model in parallel, and removing the noise of the noise image by adopting a loop iteration optimization strategy through back diffusion calculation.

Description

Technical Field

[0001] The present invention belongs to the technical field of image processing, and relates to an image denoising and enhancement method based on an additive Gaussian diffusion model. Background Art

[0002] With the continuous development of image processing technology, image denoising has become an important research direction in the field of image processing. Image noise not only affects the visual quality of images but may also affect subsequent image analysis and computer vision tasks, such as object detection, image segmentation, medical image analysis, etc. Images are affected by various factors during acquisition and transmission, including environmental light changes, sensor performance, signal interference, etc., resulting in a decline in image quality and inevitable noise pollution. Existing image denoising techniques can generally be divided into traditional methods, deep learning-based methods, and composite methods combined with noise estimation.

[0003] Traditional image denoising methods mainly include filtering methods, regularization methods, etc. Filtering methods such as mean filtering, median filtering, and Gaussian filtering remove noise by weighted averaging of neighboring pixels, but these methods usually cannot effectively remove high-frequency noise and are prone to loss of image details. For example, although mean filtering can remove noise, it blurs the edges of the image and reduces the retention of detail information. Another type of method, such as the regularization method based on total variation (TV), can retain the edges of the image while denoising, but when dealing with complex noise and image details, it is prone to detail blurring or image distortion. In addition, these traditional methods usually rely on prior assumptions about noise and are difficult to adapt to different types and intensities of noise changes, resulting in unstable effects in complex scenarios.

[0004] In recent years, deep learning methods have become an important research direction in image denoising. In particular, technologies such as convolutional neural networks (CNNs) and generative adversarial networks (GANs) have been widely applied to denoising tasks. Deep learning models automatically learn the characteristics of noise by training large-scale datasets and generate clearer images. Such methods have strong denoising capabilities and show good effects especially when dealing with complex noise. However, the disadvantages of deep learning methods lie in their dependence on large-scale labeled data. In some specific application fields (such as medical images, remote sensing images, etc.), the lack of labeled data limits the effectiveness and promotion of these methods. In addition, deep learning models often require a large amount of computing resources, and the training and tuning processes of the models are complex, increasing the application difficulty.

[0005] Existing image denoising techniques usually separate the noise level estimation from the image denoising process. Although noise estimation and denoising are two important steps in image denoising, independent processing often leads to unstable denoising effects. For example, traditional noise estimation methods such as those based on principal component analysis (PCA) or statistical models, although able to provide certain noise estimation values, often produce inaccurate estimation results, especially in complex noise environments, where they cannot fully consider the spatio-temporal characteristics of the noise. In addition, the results of noise estimation cannot be fed back to the denoising process in a timely manner, resulting in poor denoising effects, such as over-smoothed images or incomplete noise removal, which is particularly obvious when the image contains rich details or complex textures.

[0006] Although existing image denoising methods have achieved certain results in different scenarios, there are still some significant deficiencies. First, traditional methods relying on prior noise estimation often cannot adapt to dynamically changing noise environments. The error in noise level estimation will directly affect the denoising effect, resulting in the loss of image details or the appearance of artifacts. Second, although deep learning methods can achieve good denoising effects, their dependence on labeled data and consumption of computing resources limit their practical applications in specific fields. Finally, the separate processing of noise estimation and the denoising process makes it impossible to optimize the denoising effect, especially in complex noise environments, resulting in unstable image processing results. Summary of the Invention

[0007] To solve the above technical problems, the present invention proposes an image denoising and enhancement method based on an additive Gaussian diffusion model. By coupling the noise estimation and the denoising process with each other, the collaborative optimization of the two is achieved, thus effectively solving the problem of the disconnection between noise estimation and the denoising process in the prior art.

[0008] The present invention provides an image denoising and enhancement method based on an additive Gaussian diffusion model, including:

[0009] Step 1: Select multiple clear images, and gradually add additive Gaussian white noise to the clear images until the images are completely contaminated by noise, generating a series of noise images with different contamination metrics.

[0010] Step 2: Use the noise images with different contamination metrics generated in Step 1, the corresponding time steps, and the global noise samples obeying the standard normal distribution to form a training dataset, and train the U-Net network.

[0011] Step 3: Establish an additive Gaussian diffusion model for backward derivation to estimate the noise image at the previous moment.

[0012] Step 4: Collect the noise image to be denoised, obtain the initial value of the noise level of the noise image, and calculate the corresponding time step through the time encoding function.

[0013] Step 5: Input the polluted image to be denoised and the corresponding time steps into the trained U-Net network, and use the additive Gaussian diffusion model established in step 3 to remove the noise of the noisy image through reverse diffusion calculation and a cyclic iterative optimization strategy.

[0014] The image denoising and enhancement method based on the additive Gaussian diffusion model of the present invention has at least the following beneficial effects:

[0015] 1. In existing image denoising technologies, noise estimation and denoising processes are usually performed independently, and effective collaborative optimization cannot be achieved. This separate processing method leads to problems such as over-smoothing or residual noise in the image during the denoising process. The present invention combines noise level estimation with the denoising process and adopts a bidirectional adaptive optimization strategy to ensure that the noise estimation and image restoration processes interact with each other to form a benign feedback. The bidirectional adaptive optimization strategy continuously adjusts the denoising process and the noise level estimation value in multiple iterations, making the image denoising process more accurate and stable, especially when dealing with complex noise environments.

[0016] 2. The present invention models the gradual addition of noise through a diffusion process, thereby accurately simulating the diffusion characteristics of additive white Gaussian noise in the image. By constructing a gradually degraded image sequence and a Gaussian distribution of accumulated noise, the present invention can accurately capture the noise accumulation effect at each time step and perform reverse denoising based on the accumulated noise level. Compared with traditional direct denoising and denoising methods, the present invention can describe the diffusion and accumulation process of noise in more detail, ensuring a smoother noise removal process and avoiding over-denoising or loss of details.

[0017] 3. The reverse diffusion process of the present invention utilizes the Markov chain and Bayesian framework to gradually approach the true distribution of the image from the Gaussian distribution. In the parameter optimization process, the noise parameter sequence is optimized by deep learning technology, and the U-Net architecture is used for parameter estimation to ensure the efficiency and accuracy of the reverse diffusion process. Compared with the traditional denoising algorithm, the reverse diffusion process of the present invention combines neural network optimization to improve the denoising accuracy. The optimized noise parameters effectively reduce the calculation error in the denoising process and ensure the high fidelity of image restoration.

[0018] 4. In order to further improve the accuracy of noise level estimation and denoising, the present invention introduces a time coding mechanism to dynamically adjust the number of time steps in the back diffusion process according to the noise level estimation value in each iteration. This mechanism ensures that the noise level estimation of each time step can effectively guide the denoising process and avoid denoising errors caused by inaccurate noise estimation that may occur in complex images. Compared with traditional static denoising methods, the adaptive adjustment mechanism of the present invention can effectively improve the image detail recovery and noise suppression capabilities.

[0019] 5. The present invention shows excellent generalization ability when processing various types of noise. By comprehensively applying models and deep neural networks, it can adapt to a variety of different noise conditions and effectively remove various types of noise. This generalization ability makes the present invention not only applicable to additive white Gaussian noise, but also capable of processing other complex noise types, such as salt and pepper noise and Poisson noise. The technical solution of the present invention is not only applicable to the denoising of natural images, but also can be widely used in multiple fields such as medical image processing and real-time monitoring video processing. In medical images (such as X-ray images, MRI images, etc.), the present invention can significantly remove noise in the imaging process, improve the diagnostic accuracy of the image, and maintain important pathological details. In real-time video monitoring, the present invention can efficiently remove noise interference in the video, enhance image quality, and ensure the effectiveness and robustness of the monitoring system. The verification of these application scenarios further reflects the wide applicability and practical value of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 The present invention is a flowchart of an image noise reduction and enhancement method based on an additive Gaussian diffusion model. DETAILED DESCRIPTION

[0021] In the prior art, image noise estimation and denoising are often considered independent tasks. Traditional methods optimize noise estimation and denoising processes separately, resulting in poor noise removal effects, especially when processing images with complex textures or details, which are prone to over-smoothing or incomplete denoising. In addition, existing deep learning methods still face challenges in the accuracy and stability of noise estimation and denoising, especially under dynamic noise conditions, and lack a unified optimization mechanism to simultaneously improve noise estimation and denoising effects.

[0022] In order to solve the deficiencies in the existing technology, this patent proposes an image noise reduction and enhancement method based on the additive Gaussian diffusion model. This technology tightly couples the noise estimation and denoising processes by introducing a gradual accumulation of noise levels, thus achieving mutual optimization of noise estimation and denoising, and greatly improving the accuracy and stability of noise removal. Figure 1 As shown, the specific technical solution is as follows:

[0023] Step 1: Select multiple clear images and gradually add additive white Gaussian noise to the clear images until the images are completely contaminated by noise, generating a series of noisy images with different levels of contamination. This process simulates the degradation process of images after being gradually interfered by noise, ensuring that the noise level is controllable and meets the requirements of different noise conditions. Specifically:

[0024] Step 1.1: Randomly select multiple clear images from the BSD100 or TID2013 standard datasets, thereby ensuring that the input images have high resolution, rich texture, and good visual quality, providing a reliable benchmark for subsequent noise estimation and denoising.

[0025] Step 1.2: During the process with a total number of steps T = 1000, adopt a strategy of gradually increasing the Gaussian noise level to increase the noise level from σ min to σ max , and the specific formula is as follows:

[0026]

[0027] where σ t is the noise level added at the t-th step; σ max = 10 -3 is the set maximum noise level, and σ min = 10 -5 is the set minimum noise level; this strategy ensures a steady increase in the noise contamination process and meets the modeling requirements at different noise levels.

[0028] The noisy image generated after adding additive white Gaussian noise is represented as:

[0029]

[0030] where x t is the noisy image generated after being contaminated by noise at the t-th step; x0 represents the clear image; σ i represents the noise level added at the i-th step; z i is the local noise sample obtained from the i-th sampling that follows the standard normal distribution N(0,1).

[0031] According to the additivity principle of the normal distribution, when multiple independent noise components are added, the resulting total noise also follows a Gaussian distribution. The formula in Step 1.3 can be approximately rewritten as:

[0032]

[0033] where is the global noise sample obtained from the t-th sampling that follows the standard normal distribution N(0,1), and is used to approximate a single sample of the total noise effect over the entire number of time steps.

[0034] Step 2: Construct a training dataset by using the noise images with different pollution metrics, the corresponding time steps, and the global noise samples that follow the standard normal distribution generated in Step 1, and then train the U-Net network. To ensure that the model can effectively learn the denoising process, the noise images in the training data need to cover different levels from low noise to high noise, so that the U-Net network can adapt to various noise environments and thus achieve a stable image restoration effect. Specifically:

[0035] Train and optimize the U-Net network according to the loss function, and gradually adjust the network parameters. The input of the U-Net network is the noise images with different pollution metrics, the corresponding time steps, and the true global noise samples, and the output of the U-Net network is the predicted global noise samples. The difference between the predicted global noise samples and the true global noise samples is optimized through the loss function. The specific loss function is:

[0036]

[0037] where is the global noise sample predicted by the U-Net network; ||·|| 2 represents the squared Euclidean norm of the vector, that is, the sum of the squares of the differences of the elements of the vector; represents taking the expected value.

[0038] Step 3: Establish an additive Gaussian diffusion model for backward derivation to estimate the noise image at the previous moment. Specifically:

[0039] Step 3.1: For the completely contaminated image, its noise component reaches the maximum value, and its distribution tends to be:

[0040]

[0041] where x T is the completely contaminated noise image, is the mean of the clear image x0; I represents the identity matrix.

[0042] Step 3.2: According to the Markov chain and Bayesian theory, the posterior probability distribution of the reverse diffusion process is expressed as:

[0043]

[0044] where q(x t-1 |x t ,x0) represents the posterior probability distribution of x t given x0 and x t-1 , q(x t-1 |x0) is the conditional probability distribution of x t-1 given x0, q(xt |x t-1 , x0) represents the conditional probability distribution of x given x0 and x t-1 when x t and q(x t |x0) represents the marginal probability distribution of x given x0. According to Equation (3), we know that: t

[0045]

[0046] where x t-1 approximately follows a normal distribution Equation (6) is rewritten as:

[0047]

[0048] where and are the mean and variance of the noise image x representing the previous moment respectively. Therefore: t-1

[0049]

[0050] According to the and obtained above, the distribution of the noise image x at the previous moment can be reparameterized as: t-1

[0051]

[0052] where ε is the additional noise added at time step t - 1 and follows a standard normal distribution.

[0053] Step 4: Collect the noise image to be denoised, obtain the initial value of the noise level of the noise image, and calculate the time step corresponding to the initial value of the noise level through the time encoding function, specifically:

[0054] Step 4.1: Use principal component analysis to estimate the noise level of the noise image to be denoised and obtain the initial value of the noise level;

[0055] Step 4.2: Calculate the time step corresponding to the initial value of the noise level through the following time encoding function:

[0056]

[0057] where represents the time step corresponding to the noise level ; represents finding the parameter that makes the following expression reach the minimum value. The time encoding function means that for each time step t, find the one corresponding to the noise level​​​ Closest cumulative noise level Furthermore, obtain the time step closest to the output.

[0058] Step 5: Input the contaminated image to be denoised and the corresponding time step into the trained U-Net network, and combine with the additive Gaussian diffusion model established in Step 3. Through reverse diffusion calculation and using a cyclic iterative optimization strategy, remove the noise from the noisy image, specifically as follows:

[0059] Step 5.1: Input the noisy image to be denoised, the time step corresponding to the initial value of the noise level calculated in Step 4, and the extracted global noise sample into the trained U-Net network to obtain the predicted global noise sample;

[0060] Step 5.2: Substitute the global noise sample predicted by the U-Net network into Equation (12) of the additive Gaussian diffusion model to calculate the noisy image of the previous time step;

[0061] Step 5.3: Then input the noisy image of the previous time step, the time step, and the globally collected noise sample again into the U-Net network to obtain the predicted global noise sample of the previous time step. Repeat Steps 5.2 and 5.3 to gradually obtain the noisy image corresponding to each time step until the denoised clear image is obtained;

[0062] Step 5.4: Update the noise level according to the following formula:

[0063]

[0064] where is the noisy image calculated by Equation (12) at time step m + 1, is the noisy image calculated by Equation (12) at time step m, and n is the time step corresponding to the noise level before update; Std() is a function for calculating the standard deviation;

[0065] Step 5.5: Determine whether the updated noise level meets the following termination condition:

[0066]

[0067] where η = 0.01;

[0068] Step 5.6: If the termination condition is not met, recalculate the time step corresponding to the updated noise level through the time encoding function, predict the global noise sample using the trained U-Net network, and then go back to Step 5.2 to perform the next round of iteration until the termination condition is met.

[0069] In specific implementation, according to the set convergence criterion, it is judged whether the noise level has changed significantly. If the change is small, it is considered that convergence has been achieved. If the noise level has not reached the convergence state, the next round of iteration is continued until the noise level remains stable in multiple steps. This step ensures the accuracy and stability of the denoising process, avoids over-iteration or premature termination, and thus ensures that the quality of the finally output denoised image reaches the optimal state.

[0070] Step 5.7: If the termination condition is met and the noise level has reached a stable state, output the final clear denoised image.

[0071] In specific implementation, once the convergence condition is met and multiple rounds of iterative optimization are completed, the final denoised image is output. Through repeated noise level estimation and denoising process optimization, the noise in this image is effectively removed, while the image details and structure are retained. The quality of the finally output image is better than that of the original noisy image, and it shows good restoration effects under various noise conditions. This result can effectively support the image restoration task in a high-noise environment and provide clear and real image information.

[0072] A method for image noise reduction and enhancement based on an additive Gaussian diffusion model according to the present invention, different from traditional methods that rely on prior noise estimation, the present invention organically combines noise level estimation and the denoising process, optimizes the synergy between the two through an adaptive coding mechanism, and eliminates the negative impact of noise estimation errors on the denoising effect. In addition, the present invention utilizes the powerful generalization ability of the diffusion model to deeply analyze the characteristics of noise, thereby improving the accuracy of noise level estimation and enhancing the ability to retain details in the denoising process. This solution can not only achieve efficient denoising in a variety of complex noise scenarios, but also better adapt to changes in different noise distributions, improve the accuracy and stability of image processing, and is particularly suitable for high-precision image processing requirements such as medical imaging and remote sensing images.

[0073] The above are only the preferred embodiments of the present invention and are not intended to limit the idea of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. An image denoising and enhancement method based on an additive Gaussian diffusion model, characterized in that, include: Step 1: Select multiple clear images and gradually add additive Gaussian white noise to the clear images until the images are completely contaminated by the noise, generating a series of noisy images with different contamination measures; Step 2: The noise images with different pollution measures generated in step 1, the corresponding time steps and the global noise samples obeying the standard normal distribution constitute a training data set to train the U-Net network; Step 3: Establish an additive Gaussian diffusion model to reversely deduce and estimate the noise image at the previous moment; Step 4: Collect the noise image to be denoised, obtain the initial value of the noise level of the noise image, and calculate the number of time steps corresponding to the initial value of the noise level through the time encoding function; Step 5: Input the polluted image to be denoised and the corresponding time steps into the trained U-Net network, and use the additive Gaussian diffusion model established in step 3 to remove the noise of the noisy image through reverse diffusion calculation and a cyclic iterative optimization strategy.

2. The image denoising and enhancement method based on the additive Gaussian diffusion model according to claim 1, characterized in that, The step 1 is specifically as follows: Step 1.1: Randomly select multiple clear images from the BSD100 or TID2013 standard dataset; Step 1.2: During the total number of steps \(T = 1000\), adopt a strategy of gradually increasing the Gaussian noise level to increase the noise level from \(\sigma\) min to \(\sigma\) max gradually. The specific formula is as follows: where, σ t is the noise level added at the t-th step; σ max is the set maximum noise level, and σ min is the set minimum noise level; Step 1.3: The noisy image generated after adding additive Gaussian white noise is represented as: Among them, x t is the noisy image generated after noise pollution in the t-th step; x0 represents the clear image; σ i represents the noise level added in the i-th step; z i is the local noise sample obtained from the i-th sampling that follows the standard normal distribution N(0, 1); Step 1.4: According to the additivity principle of normal distribution, when multiple independent noise components are added together, the total noise also presents a Gaussian distribution. The formula in step 1.3 is rewritten as: Among them, is the global noise sample obtained from the t-th sampling that follows the standard normal distribution N(0, 1).

3. The image denoising and enhancement method based on the additive Gaussian diffusion model according to claim 2, characterized in that, The step 2 is specifically as follows: The U-Net network is trained and optimized according to the loss function, and the network parameters are gradually adjusted. The input of the U-Net network is the noise image with different pollution measures, the corresponding time steps and the real global noise sample. The output of the U-Net network is the predicted global noise sample. The difference between the predicted global noise sample and the real global noise sample is optimized by the loss function. The specific loss function is: Among them, is the global noise sample predicted by the U-Net network; ||·|| 2 represents the squared Euclidean norm of a vector, that is, the sum of the squares of the differences of the elements of the vector; represents taking the expected value.

4. The image denoising and enhancement method based on the additive Gaussian diffusion model according to claim 1, characterized in that, The step 3 is specifically as follows: Step 3.1: For a completely contaminated image, its noise component reaches its maximum value and its distribution tends to: where x T is a completely contaminated noisy image, is the mean of the clear image x0; I represents the identity matrix; Step 3.2: According to Markov chain and Bayesian theory, the posterior probability distribution of the reverse diffusion process is expressed as: Among them, q(x t-1 |x t ,x0) means given x0 and x t Time t-1 The posterior probability distribution of q(x t-1 |x0) is the value of x when x0 is given t-1 The conditional probability distribution of q(x t |x t-1 ,x0) means given x0 and x t-1 Time t The conditional probability distribution of q(x t |x0) means that given x0, x t The marginal probability distribution of is: where x t-1 approximately follows a normal distribution Equation (6) is rewritten as: Among them, and are the mean and variance of the representation of the noise image x at the previous moment, respectively. Therefore: t-1 ​ Obtained according to the above and the noise image x at the previous moment t-1 The distribution of can be reparameterized as: Among them, ε is the additional noise added at time step t-1, which follows a standard normal distribution.

5. The image denoising and enhancement method based on the additive Gaussian diffusion model according to claim 1, characterized in that The step 4 is specifically as follows: Step 4.1: Use principal component analysis to estimate the noise level of the noisy image to be denoised, and obtain an initial value of the noise level; Step 4.2: Calculate the number of time steps corresponding to the initial value of the noise level using the following time encoding function: Among them, represents the time step corresponding to the noise level ; the time encoding function means that for each time step t, find the cumulative noise level closest to the noise level and then obtain the time step with the closest output.

6. The image denoising and enhancement method based on the additive Gaussian diffusion model according to claim 1, characterized in that, The step 5 is specifically as follows: Step 5.1: Input the noisy image to be denoised, the time step number corresponding to the initial value of the noise level calculated in step 4, and the extracted global noise sample into the trained U-Net network to obtain the predicted global noise sample; Step 5.2: Substitute the global noise sample predicted by the U-Net network into equation (12) of the additive Gaussian diffusion model to calculate the noise image of the previous time step; Step 5.3: Then input the noise image of the previous time step, the time step, and the globally sampled noise samples collected again into the U-Net network to obtain the globally sampled noise samples predicted for the previous time step. Repeat steps 5.2 and 5.3 to gradually obtain the noise images corresponding to each time step until the denoised clear image is obtained; Step 5.4: Update the noise level according to the following formula: wherein, is the noise image calculated by Equation (12) at time step m + 1, is the noise image calculated by Equation (12) at time step m, and n is the time step corresponding to the noise level before update; Std() is a function for calculating the standard deviation; Step 5.5: Determine whether the updated noise level meets the following termination conditions: where η = 0.01; Step 5.6: If the termination conditions are not met, recalculate the time step corresponding to the updated noise level through the time encoding function, predict the globally sampled noise samples using the trained U-Net network, and then go back to step 5.2 to perform the next iteration until the termination conditions are met; Step 5.7: If the termination conditions are met and the noise level has reached a stable state, output the final denoised clear image.

Images