Image denoising method based on noise distribution conversion

Abstract

This invention relates to an image denoising method based on noise distribution transformation, comprising: smoothing an original noisy image to obtain a smoothed image and acquiring initial noise; performing noise distribution transformation processing on the initial noise using histogram matching technology to obtain transformed noise; fusing the transformed noise with the smoothed image to obtain a transformed noise image; performing denoising processing on the transformed noise image using a preset denoiser to obtain a denoised result image; performing inverse restoration processing on the texture portion of the difference based on the forward transformation rules adopted in the noise distribution transformation processing to obtain restored texture; fusing the restored texture with the current smoothed image to update the smoothed image, and returning to repeat the process until a preset condition is met. This invention uses histogram matching to uniformly convert various out-of-distribution noises into a noise distribution adapted to a preset denoiser, achieving effective removal of multiple types of noise without retraining the denoising model, significantly improving the generalization ability and applicability of the denoiser.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology, and specifically relates to an image denoising method based on noise distribution transformation. Background Technology

[0002] Images are inevitably affected by various types of noise during the capture process, resulting in the destruction of original information. As a classic and important task in the field of image processing, image denoising aims to recover a clear image from a noisy one. This task not only improves image quality and visual effects but also provides reliable data support for high-level vision tasks such as object detection.

[0003] Traditional denoising methods (such as CBM3D and low-rank denoising) utilize prior image information for denoising, but their performance is limited. In recent years, learning-based denoising methods have become mainstream. Supervised learning methods achieve excellent performance by training on large amounts of paired data, but rely on clean images as labels, which are difficult to obtain in some real-world scenarios. To address this, self-supervised denoising methods have been proposed, which can be trained using only noisy images, but still require a large number of noisy images as training data. Zero-shot denoising methods further reduce data requirements, requiring only a single noisy image for training, but models trained on a single image often fail to achieve ideal denoising results.

[0004] Currently, most learning-based denoising models are essentially Gaussian denoisers, meaning their training process assumes the noise follows an additive Gaussian white noise distribution. These Gaussian denoisers perform well when dealing with noise distributions similar to the training data, but their denoising performance significantly degrades and their generalization ability is severely insufficient when faced with out-of-distribution noise (such as salt-and-pepper noise, Poisson noise, speckle noise, periodic noise, and complex real-world noise). When faced with unseen noise types, existing methods typically require re-collecting data and retraining the model to adapt to the new noise characteristics, which is costly and inefficient in practical applications.

[0005] To improve the generalization ability of Gaussian denoisers, existing technologies mainly explore two directions: one is from the noise perspective, attempting to transform the noise to be processed so that its distribution approximates a Gaussian distribution, thus enabling it to be effectively removed by the Gaussian denoiser. However, these methods are usually only applicable to specific noise types and often rely on accurate estimation of noise parameters, resulting in poor universality. The other approach is from the model perspective, enhancing the robustness of the model itself by improving the network structure or training strategy. However, these methods have limited effect on improving generalization ability, and Gaussian denoisers still struggle to universally handle various types of out-of-distribution noise.

[0006] In summary, there is an urgent need for an image denoising method based on noise distribution transformation that can universally improve the generalization ability of denoisers and enable them to effectively handle various types of out-of-distribution noise without retraining. Summary of the Invention

[0007] The purpose of this invention is to provide an image denoising method based on noise distribution transformation that can universally improve the generalization ability of denoisers, enabling them to effectively handle various types of out-of-distribution noise without retraining.

[0008] The above objective is achieved through the following technical solution: An image denoising method based on noise distribution transformation, comprising the following steps: S1, Initial noise estimation: The original noisy image is smoothed to obtain a smoothed image, and the initial noise is obtained based on the difference between the original noisy image and the smoothed image; S2, Noise Distribution Transformation: The initial noise is transformed using histogram matching technology to convert the statistical distribution of the initial noise into the target noise distribution characteristics, and the transformed noise is obtained; wherein, the target noise distribution characteristics are matched with the noise distribution characteristics adapted by the preset denoiser; S3, Noisy image reconstruction: The converted noise is fused with the smoothed image to obtain a converted noisy image; S4, Denoising: The converted noise image is denoised using a preset denoiser to obtain a denoised result image; S5, Texture Restoration: Based on the forward conversion rule used in the noise distribution conversion process in step S2, reverse restoration processing is performed on the texture part corresponding to the initial noise after conversion and denoising in the difference between the denoised result image and the current smooth image to obtain the restored texture; S6, Iterative Update: The restored texture is fused with the current smooth image to update the smooth image, and the process returns to step S1 to repeat the initial noise estimation and steps S2 to S5 until the preset iteration conditions are met, and the final denoised image is output; wherein the initial noise is obtained by using the difference between the updated current smooth image and the original noise image during the iteration process.

[0009] First, the method of this invention serves as an independent preprocessing and post-processing framework. Its input is a noisy image, and its output is a denoised image. The preset denoiser called internally can be any existing denoising model adapted to the target noise distribution. Therefore, this invention possesses strong universality and plug-and-play characteristics, enabling a vast number of existing denoising models and greatly expanding their applicability. Furthermore, by introducing a texture restoration step and an iterative mechanism, this invention effectively solves the texture distortion problem generated during noise conversion. The texture restoration step compensates for the texture information distorted by the forward conversion, while the iterative update mechanism promotes and optimizes noise estimation and texture restoration together. Therefore, the final denoised image not only has low noise residue but also better preserves the edges, structure, and texture details of the original image, achieving an excellent balance between noise suppression and detail preservation.

[0010] This invention transforms various out-of-distribution noises with unknown statistical characteristics into a target noise distribution familiar to a pre-defined denoiser through a noise distribution conversion step. This allows a fixed denoiser designed specifically for a particular noise distribution to effectively handle various types of noise, including Gaussian noise, salt-and-pepper noise, Poisson noise, speckle noise, periodic noise, and complex real-world noise, without any structural modifications or retraining. This effectively alleviates the pain points of existing denoising models, such as insufficient generalization ability and the need for retraining when faced with new noise.

[0011] A further technical solution is that step S2 includes adjusting the statistical distribution of the initial noise in at least one transform domain.

[0012] In this way, different transform domains can reveal the differences in the statistical characteristics of noise in different dimensions (such as the frequency domain reflecting energy distribution and the wavelet domain reflecting multi-scale characteristics), enabling the noise distribution transformation operation to more comprehensively regulate the multi-dimensional characteristics of noise, thereby adapting to more complex noise types and scenarios.

[0013] A further technical solution is that step S2 includes performing a histogram matching operation in the spatial domain: In response to the initial noise being determined to be global noise, global histogram matching is performed on the initial noise: a uniform transformation coefficient is determined based on the overall statistical distribution of the initial noise, and a distribution transformation is performed on the initial noise as a whole; In response to the initial noise being determined to be local noise, local histogram matching is performed on the initial noise: the initial noise is divided into multiple image blocks, and transformation coefficients are determined based on the local statistical distribution of each image block to perform block transformation.

[0014] Based on the spatial distribution characteristics of the initial noise, different histogram matching methods are adaptively selected. Specifically, when the initial noise is determined to be global noise, a global histogram matching strategy is adopted, that is, a unified transformation mapping relationship is determined based on the statistical distribution of the entire noisy image, and this mapping is uniformly applied to all pixels. When the initial noise is determined to be local noise, a local histogram matching strategy is adopted, that is, the noisy image is divided into multiple sub-blocks, and an independent transformation mapping relationship is determined based on the local statistical distribution of each sub-block, thereby achieving block-differentiated distribution transformation.

[0015] This invention differentiates between global and local noise and employs a differentiated spatial histogram matching strategy, enabling adaptive adaptation to the spatial distribution characteristics of noise. For global noise with consistent statistical characteristics across the entire image domain, a globally unified transformation mapping is used to avoid statistical bias caused by over-segmentation, ensuring overall consistency in the conversion. For spatially unevenly distributed local noise, a block-based independent transformation mapping is employed, accurately capturing and adapting to local variations in the statistical characteristics of noise in different regions, effectively improving the targeting and accuracy of noise distribution conversion. By organically combining these two aspects through a noise type determination mechanism, processing efficiency is considered while ensuring conversion effectiveness.

[0016] A further technical solution is that, in the local histogram matching process, the initial noise is divided into multiple image blocks, histogram matching is performed on each image block, and all image blocks that have completed histogram matching are reconstructed into a complete image; wherein, for local noise with signal correlation, based on the non-local self-similarity of the image, a preset number of similar image blocks are matched for each target image block, and histogram matching processing is performed based on the pixel set of the target image block and all similar image blocks.

[0017] To address signal-correlated local noise—complex noise whose intensity is correlated with the image signal itself—this paper introduces an enhancement processing mechanism based on non-local self-similarity of images. Specifically, for each target image patch, multiple image patches with similar structure or texture are searched within the image. These similar patches are merged with the target image patch to form an expanded pixel sample set. Histogram matching is then performed based on this expanded sample set. By utilizing the non-local self-similarity of images and aggregating pixel samples from multiple similar image patches, the effective sample space for histogram statistics is significantly expanded. This significantly improves the accuracy and reliability of local statistical distribution estimation, effectively solving the problem of insufficient local histogram matching accuracy in signal-correlated noise scenarios. It achieves a good balance between local adaptability and statistical robustness, significantly improving the noise conversion quality in complex signal-correlated noise scenarios.

[0018] A further technical solution is that step S2 also includes performing a histogram matching operation in the frequency domain: Perform a discrete cosine transform on the initial noise to obtain the frequency domain noise; Histogram matching is performed on the frequency domain noise so that its frequency domain statistical distribution approximates the statistical distribution of the target noise distribution characteristics in the frequency domain.

[0019] In practical applications, the frequency domain histogram matching operation is performed when the initially estimated noise exhibits spatial correlation and the current iteration count is greater than 1. By performing histogram matching in the frequency domain, the distribution characteristics of noise in the frequency domain can be adjusted, breaking the spatial correlation structure between noise pixels and converting it into a spatially independent target distribution form, thereby compensating for the shortcomings of spatial domain processing in converting this type of noise. The multi-domain joint processing approach can comprehensively adjust the statistical characteristics of noise from different representational dimensions, significantly improving the adaptability and conversion effect to complex noise types.

[0020] A further technical solution is that step S2 also includes performing an energy equalization operation in the wavelet domain: Perform wavelet transform on the initial noise to obtain multiple sub-band components; The energy amplitude of each sub-band component is adjusted to make the energy distribution of each sub-band component consistent with the energy distribution characteristics of the target noise distribution.

[0021] In practical applications, wavelet domain energy equalization is performed when the initial estimated noise exhibits spatial correlation and the current iteration count is greater than 1. For colored noise with uneven energy distribution across different frequency bands, its statistical characteristics in the spatial or frequency domains are difficult to fully correct using conventional histogram matching. By equalizing the energy of each sub-band component in the wavelet domain, the energy distribution of colored noise can be "whitened," making it approach the characteristics of target noise with uniform energy distribution. This allows denoisers originally designed for uniformly energetic noise to effectively handle colored noise, significantly expanding the applicability of denoisers. Performing energy equalization on each sub-band separately in the wavelet domain allows for fine-tuning of the noise energy distribution from multiple scale dimensions. Compared to processing only in a single transform domain, this provides a more comprehensive adjustment of the noise's statistical characteristics, further improving the conversion effect and denoising quality for complex noise.

[0022] A further technical solution is that the target noise distribution characteristics are those of additive white Gaussian noise; the preset denoiser is a denoising neural network model trained for the additive white Gaussian noise; step S1 specifically includes the following steps: S11, perform median filtering on the input original noisy image to remove abnormal pixel values ​​in the image and obtain the filtered image; S12, the filtered image is smoothed using the RTV image smoothing method to separate the structural information and texture information of the image, and the smoothed image is obtained; S13, calculate the pixel difference between the original noise image and the smooth image to obtain an initial noise map, and add Gaussian noise with a preset standard deviation to the initial noise map to obtain the initial noise.

[0023] In the initial noise estimation process, median filtering is first used to remove outlier pixel values, effectively preventing extreme noise values ​​from interfering with subsequent smoothing. Employing an image smoothing method that separates structure and texture yields a higher-quality smooth image, making the initial noise extracted through image subtraction more accurate. Adding a small amount of Gaussian noise to the initial noise map enhances the robustness of subsequent noise distribution transformation processing, avoiding instability caused by excessively sparse or extremely distributed initial noise. These refinement operations collectively improve the quality of the initial noise estimation, laying a more reliable foundation for subsequent noise distribution transformation.

[0024] A further technical solution is that step S4 specifically includes the following steps: Determine whether the initial noise has spatial correlation. If not, input the transformed noise image into a pre-trained target level Gaussian denoiser to obtain the initial denoised image. If the initial noise has spatial correlation, the transformed noise image is first subjected to pixel rearrangement downsampling processing, and then input into a pre-trained target level Gaussian denoiser to obtain a downsampled denoised image; the downsampled denoised image is then subjected to pixel rearrangement upsampling processing to restore it to the original image size to obtain an upsampled denoised image. Perform random replacement refining process: Using a binary mask with a preset probability, some pixels in the upsampled denoised image are randomly replaced with corresponding pixels in the converted noisy image. Repeat the random replacement and Gaussian denoising process a preset number of times, and take the average value of the results of multiple denoising processes to obtain the denoised result image.

[0025] This invention effectively breaks the spatial continuity of noise and reduces its spatial correlation by introducing a pixel rearrangement downsampling operation before denoising. This allows the Gaussian denoiser to better perform its denoising capabilities and significantly improves its ability to suppress spatially correlated noise. Furthermore, the invention introduces a random replacement refining process. By fusing and averaging multiple random replacements with the denoising results, the invention can effectively suppress or eliminate the aforementioned artifacts, ensuring that the visual quality of the final output image is not affected by the additional operations, thus achieving a good balance between denoising effect and image fidelity.

[0026] A further technical solution is that step S5 includes: The difference between the denoised image and the current smoothed image is obtained as the texture component to be restored; Based on the forward transformation coefficients used in the noise distribution transformation process described in step S2, the corresponding inverse transformation coefficients are determined. The inverse transformation coefficients are used to perform an inverse transformation on the texture components to be restored, thereby obtaining the restored texture.

[0027] In this way, the consistency between texture restoration and noise conversion is ensured, and the difference between the denoised image and the smoothed image is used as the texture component to be restored. This allows for the accurate location of texture information affected during the conversion and denoising process. By performing an inverse transformation that matches the forward transformation on this difference, the changes to the texture caused by the noise conversion operation can be effectively compensated, and the lost or distorted texture information can be restored to the smoothed image. This ensures the denoising effect while preserving the details and structural information of the image to the maximum extent.

[0028] A further technical solution is to perform reverse restoration processing on the texture portion corresponding to the converted and denoised initial noise in the difference between the denoised result image and the current smoothed image, including at least one of the following: Wavelet domain texture transformation: After performing wavelet decomposition on the texture components in the difference, divide each sub-band component by the energy adjustment coefficient corresponding to the noise transformation process in step S2, and then perform inverse wavelet transform; Spatial texture conversion: Guided filtering is performed on the coefficient matrix and bias matrix corresponding to the spatial histogram matching in step S2 noise conversion, and the texture components are inversely transformed based on the processed coefficient matrix and bias matrix; Spatial-frequency domain joint texture transformation: The inverse transformation of spatial domain histogram matching and the inverse transformation of frequency domain histogram matching are performed sequentially on the texture components.

[0029] In specific applications, wavelet domain texture conversion corresponds to the energy equalization operation performed in the wavelet domain during noise conversion. It restores the texture by performing wavelet decomposition, coefficient correction, and inverse transform on the texture components. Spatial domain texture conversion corresponds to the spatial histogram matching operation. It restores the texture by processing the coefficient matrix of the forward transform and then performing the inverse transform. Joint spatial-frequency domain texture conversion corresponds to scenarios in noise conversion involving both spatial and frequency domain processing. It sequentially performs inverse spatial and inverse frequency transforms to achieve joint texture restoration.

[0030] This invention proposes an image denoising method based on noise distribution transformation. It is the first to apply histogram matching technology to the field of out-of-distribution noise transformation, forming a core technical approach of "converting out-of-distribution noise into Gaussian noise + denoising using existing denoisers." This method overcomes the technical bias of existing technologies that can only address the issue from improving the denoising model structure or transforming a single noise type, solving the common industry problem of insufficient generalization ability of Gaussian denoisers at the source of noise distribution.

[0031] This invention constructs a complete closed-loop processing framework consisting of "initial noise estimation—multi-domain joint noise distribution transformation—denoising—texture inverse reconstruction—iterative optimization." Through multi-domain joint histogram matching, various out-of-distribution noises with unknown statistical characteristics can be uniformly converted into noise distributions adapted to the preset denoiser. This allows a fixed-level denoiser to effectively handle various noise types, such as Gaussian noise, salt-and-pepper noise, Poisson noise, speckle noise, periodic noise, and complex realistic noise, without any modification or retraining, significantly expanding the applicability of the denoiser.

[0032] Furthermore, by introducing a cyclic iterative mechanism, this invention continuously optimizes the noise estimation accuracy of the next round using the denoising results of each round, effectively solving the technical obstacle of accurately acquiring noise from unknown distributions and achieving a progressive improvement in noise conversion effect and denoising quality. Simultaneously, the added texture inverse restoration step can accurately compensate for the distorted image texture and structural details during noise conversion and denoising, significantly suppressing noise while maximizing the preservation of visual details such as image edges and textures, achieving an excellent balance between denoising effect and image fidelity.

[0033] Experimental results show that the present invention achieves significant improvements in peak signal-to-noise ratio and structural similarity indices under various types of distributed external noise scenarios, yielding unexpected technical effects. Attached Figure Description

[0034] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0035] Figure 1 This is a flowchart of an image denoising method based on noise distribution transformation according to one embodiment of the present invention; Figure 2 This is a flowchart illustrating the histogram matching technique in noise distribution conversion according to one embodiment of the present invention (by statistically analyzing the histograms and cumulative distributions of the original noise and the target Gaussian noise, linear interpolation is used to map the noise distribution, converting out-of-distribution noise into noise conforming to a Gaussian distribution). Figure 3This is a schematic diagram comparing the wavelet transform results of white noise and colored noise according to one embodiment of the present invention; wherein, σ is labeled as the standard deviation of the noise, and the bottom row is the noise after wavelet transform. Figure 4 This is a visual comparison diagram of the denoising effect of one embodiment of the present invention and other existing methods under different synthetic noise (typical difficult distribution out-of-range noise) scenarios. The first to third rows represent salt-and-pepper noise, Poisson noise, and circular noise, respectively. Figure 5 This is a visual comparison diagram of the denoising effect of one embodiment of the present invention and other existing methods under different synthetic noise scenarios (other types of out-of-distribution noise). The noisy images in each row from top to bottom correspond to Gaussian noise, Bernoulli noise, random impulse noise, speckle noise and mixed noise, respectively. Figure 6 This is a visual comparison diagram of the noise reduction effect of one embodiment of the present invention and other methods in the prior art in a real noise scene; Figure 7 This is a schematic diagram illustrating the noise conversion effect of one embodiment of the present invention under various complex noise types. The noise types shown include Gaussian-Poisson mixed noise, Gaussian-salt-pepper mixed noise, grid periodic noise, mixed periodic noise, and microscopic noise. It also demonstrates the applicability of grayscale images. Detailed Implementation

[0036] The present invention will now be described in detail with reference to the accompanying drawings. This description is merely illustrative and explanatory, and should not be construed as limiting the scope of protection of the present invention. Furthermore, those skilled in the art can combine the features in the embodiments described herein and in different embodiments accordingly based on the description in this document.

[0037] The present invention provides an image denoising method based on noise distribution transformation, comprising the following steps: S1, Initial noise estimation: The original noisy image is smoothed to obtain a smoothed image, and the initial noise is obtained based on the difference between the original noisy image and the smoothed image; In one embodiment, the target noise distribution characteristics are those of additive white Gaussian noise; the preset denoiser is a denoising neural network model trained for the additive white Gaussian noise. S11, perform median filtering on the input original noisy image to remove abnormal pixel values ​​in the image and obtain the filtered image; The purpose is to remove outliers. Since smoothing operations are not good at handling outliers, the formula is as follows: (1) in Original noisy image, image size is ,in For height, For width, This is the number of channels; subsequent matrices will default to this size. This is the image after median filtering, with a filter kernel size of [size missing]. .

[0038] S12, the filtered image is smoothed using the RTV image smoothing method to separate the structural information and texture information of the image, and the smoothed image is obtained; This method can effectively separate the structural and texture information of an image to achieve a smoothing effect. The formula is as follows: (2) in It smooths the image. RTV is a method that separates image structure and texture, thus effectively achieving image smoothing.

[0039] S13, calculate the pixel difference between the original noise image and the smooth image to obtain an initial noise map, and add Gaussian noise with a preset standard deviation to the initial noise map to obtain the initial noise.

[0040] Obtain a smooth image Then, the initial noise can be obtained. : (3) To enhance the robustness of the noise conversion, a small amount of Gaussian noise is added to the initial noise, as shown in the following formula: (4) in Added small standard deviation Gaussian noise, with a default value of 0.01; This refers to the noise to be converted.

[0041] S2, Noise Distribution Transformation: The initial noise is transformed using histogram matching technology to convert the statistical distribution of the initial noise into the target noise distribution characteristics, and the transformed noise is obtained; wherein, the target noise distribution characteristics are matched with the noise distribution characteristics adapted by the preset denoiser; After obtaining the initial noise, an appropriate transformation strategy needs to be selected based on the noise characteristics. The core method of transformation is histogram matching, the key of which is to make the cumulative distribution of the original noise approximate the cumulative distribution of the target noise.

[0042] like Figure 2First, histograms of noise are plotted for each interval. Then, the cumulative distribution of the noise is calculated, resulting in a series of coordinate points. The horizontal axis represents the pixel value (interval boundary), and the vertical axis represents the cumulative distribution probability of each interval. Then, for a point in the original noise... Its cumulative probability distribution can be obtained through one-dimensional linear interpolation, as shown in the following formula: (5) in for The corresponding cumulative distribution probability, and These are the slope and intercept of the linear interpolation, respectively. Next, we will... Matching the cumulative distribution of the target noise can also be achieved through one-dimensional linear interpolation. However, unlike previous interpolations, this time the horizontal axis represents the cumulative distribution probability, and the vertical axis represents the pixel value. The formula is as follows: (6) in That is The result after histogram matching and These are the slope and intercept of the linear interpolation, respectively. Then, substituting equation (6) into equation (7) yields: (7) in and that is Transform to The corresponding coefficients and biases.

[0043] Therefore, histogram matching involves a one-dimensional linear transformation for each pixel. Thus, the histogram matching process can be expressed in matrix form as follows: (8) in This is the matrix before histogram matching. for The cumulative probability matrix corresponding to each point in the matrix. This is the matrix after histogram matching. and This is the slope matrix corresponding to each point. and For the intercept matrix corresponding to each point, the simplified formula is as follows: (9) in and for Transform to The corresponding coefficient matrix and bias matrix. At this point, the histogram matching process has been modeled as a one-dimensional linear expression, which will be used to represent histogram matching in the future. The key to achieving histogram matching now lies in obtaining the coefficient matrix. and bias matrix This process, while seemingly requiring point-by-point calculation of the slope and intercept for linear interpolation at each location, which appears complex, can actually be accomplished through matrix operations without needing to solve for each point individually. From Figure 2 As can be seen, each interval corresponds to a slope and an intercept. Therefore, the first step in solving for the slope and intercept of any point is to determine the interval to which that point belongs. If the interval index matrix corresponding to all points can be obtained, the corresponding slope matrix and intercept matrix can be quickly extracted. The interval index matrix can be efficiently obtained through a mature interval partitioning algorithm. Therefore, two sets of slope matrices and intercept matrices can be conveniently obtained, and the coefficient matrix and bias matrix required for the overall transformation can be directly calculated based on them. This matrix-based solution method replaces the point-by-point calculation approach, significantly reducing the overall computational load.

[0044] Step S2 includes adjusting the statistical distribution of the initial noise over at least one transform domain.

[0045] The noise conversion strategies in the frequency domain, spatial domain, and wavelet domain will be introduced in detail below.

[0046] Step S2 also includes performing a histogram matching operation in the frequency domain: Perform a discrete cosine transform on the initial noise to obtain the frequency domain noise; Histogram matching is performed on the frequency domain noise so that its frequency domain statistical distribution approximates the statistical distribution of the target noise distribution characteristics in the frequency domain.

[0047] If the noise exhibits spatial correlation, histogram matching in the frequency domain is necessary. Gaussian white noise still follows a Gaussian distribution in the frequency domain, while spatially correlated noise, due to its spatial correlation, typically exhibits a significantly different frequency distribution. For example, periodic noise may appear as several prominent peaks in the frequency domain. Therefore, frequency domain processing is essential to address the spatial correlation of the noise. However, it should be noted that the noise obtained in the first iteration contains a large amount of texture information, and the texture itself also exhibits some spatial correlation. To avoid damaging the texture structure, frequency domain histogram matching can be performed starting from the second iteration. The frequency domain transformation is implemented using the Discrete Cosine Transform (DCT), whose mathematical expression is: (10) in express The frequency domain result after DCT is assumed to be a two-dimensional DCT; however, if channel correlation exists in the noise, a three-dimensional DCT will also be performed. Since Gaussian white noise still follows the same Gaussian distribution and maintains a consistent variance after DCT transformation, spatial correlation of the noise can be reduced through frequency domain histogram matching. The corresponding transformation formula is: (11) In the formula This is the transform result after frequency domain matching. and They are respectively to Mapped to The corresponding coefficient matrix and bias matrix. For After performing the inverse discrete cosine transform (IDCT), the noise after frequency domain histogram matching processing can be obtained. : (12) If the noise is uncorrelated, then frequency domain processing is unnecessary. Therefore, the following conditions are defined. The frequency domain histogram matching operation is only performed when the condition "the noise has spatial correlation and the number of iterations is greater than 1" is met. The final frequency domain histogram matching transformation formula is as follows: (13) in This is the final transformation result after the frequency domain histogram matching operation.

[0048] Step S2 also includes performing a histogram matching operation in the spatial domain: In response to the initial noise being determined to be global noise, global histogram matching is performed on the initial noise: a uniform transformation coefficient is determined based on the overall statistical distribution of the initial noise, and a distribution transformation is performed on the initial noise as a whole; In response to the initial noise being determined to be local noise, local histogram matching is performed on the initial noise: the initial noise is divided into multiple image blocks, and transformation coefficients are determined based on the local statistical distribution of each image block to perform block transformation.

[0049] In the local histogram matching process, the initial noise is divided into multiple image blocks, histogram matching is performed on each image block, and all image blocks that have completed histogram matching are reconstructed into a complete image. Specifically, for local noise with signal correlation, based on the non-local self-similarity of the image, a preset number of similar image blocks are matched for each target image block, and histogram matching processing is performed based on the pixel set of the target image block and all similar image blocks.

[0050] The purpose of spatial histogram matching is to make noise conform to a Gaussian distribution in the spatial domain. It is divided into global matching and local matching. If the noise is global noise, only global histogram matching is needed, as shown in the following formula: (14) In the formula This is the transformation result after global histogram matching. and They are respectively to Mapped to The corresponding coefficient matrix and bias matrix.

[0051] When the noise is localized, global histogram matching fails to effectively transform the noise distribution because it ignores differences in local statistical characteristics. Therefore, a local histogram matching strategy is adopted, using block processing to adapt to changes in local statistical characteristics. First, the noise map... Divided into sizes Image blocks, with a block step size of This block operation can be formally represented as: (15) in This is a block function, based on the block size. and step length Divide the image into blocks. The resulting image is a collection of all local image patches. This indicates the total number of blocks. This represents the number of image channels. Then, histogram matching is performed on each image block using the following formula: (16) In the formula for The Middle Image blocks, Representing the The noise transformation result of each block after local histogram matching and These are the coefficient matrix and bias matrix corresponding to the image patch, respectively. , and All dimensions .

[0052] Generally, local histogram matching can effectively adapt to the spatial local distribution characteristics of noise and achieve the conversion of ordinary local noise. However, this method has limitations when the noise exhibits significant signal correlation. The core reason is that signal-correlated noise has large local differences. To ensure the conversion effect, a smaller block size needs to be selected to capture fine local statistical characteristics. However, an excessively small block size will drastically reduce the number of pixel samples within the image block. The cumulative distribution probability obtained based on limited sample statistics will have a large estimation error, ultimately leading to a decrease in the accuracy and stability of noise conversion.

[0053] To address this issue, a nonlocal self-similarity prior is introduced to improve and optimize local histogram matching when processing signal-related noise. Nonlocal self-similarity refers to the existence of numerous repetitive structures and texture patterns in natural images. Utilizing this characteristic, the statistical sample is expanded by searching for similar image patches, thereby enhancing statistical reliability while maintaining local adaptability. Let the patch size of the nonlocal self-similarity be... The number of similar blocks is The original local histogram matching, when statistically analyzing the histogram distribution, only utilizes pixel samples from a single image patch, and its sample space dimension is... The local histogram matching method, which combines nonlocal self-similarity, expands the statistical sample set into a set of similar blocks to the target block, with a sample space dimension of [missing value]. Through aggregation The statistical information of similar blocks effectively alleviates the problem of insufficient samples caused by small-sized blocks. However, both are essentially local histogram matching, so local histogram matching that combines non-local self-similarity can also be expressed as formula (16), only... and The calculations may differ.

[0054] After histogram matching is completed for all local image patches, a restoration operation is required to reconstruct the image to its full size. This process can be formally represented as: (17) in This is a block restoration function that can restore a set of image blocks back to the original image size. This is the result of local histogram matching. Substituting formula (16) into the above formula, we get: (18) To simplify expressions and unify the expression paradigm for global / local histogram matching, a global coefficient matrix is ​​defined. With global bias matrix The set of block coefficient matrices Block bias matrix set The result after block reconstruction is: (19) In the formula, and This represents the reconstructed global matrix. Therefore, local histogram matching can be simplified to a form unified with global histogram matching: (20) Finally, define the conditions. If "the noise is localized", then the formula for spatial histogram matching is as follows: (twenty one) (twenty two) in , These are the coefficient matrix and bias matrix for spatial histogram matching, respectively. This is the final result after the spatial histogram matching operation.

[0055] Step S2 also includes performing an energy equalization operation in the wavelet domain: Perform wavelet transform on the initial noise to obtain multiple sub-band components; The energy amplitude of each sub-band component is adjusted to make the energy distribution of each sub-band component consistent with the energy distribution characteristics of the target noise distribution.

[0056] Wavelet domain energy equalization, performed in the wavelet transform domain, modulates the noise energy distribution by equalizing the noise standard deviation across frequency bands. From the perspective of noise energy distribution in the frequency domain, noise can be divided into two main categories: white noise and colored noise. Specifically, white noise exhibits a uniform energy distribution across the entire frequency spectrum, meaning that the noise energy amplitude corresponding to different frequency components is essentially the same. Conversely, colored noise has a significantly non-uniform energy distribution, with energy potentially concentrated in specific frequency ranges (low or high frequencies). However, existing Gaussian denoisers are designed for Gaussian white noise and are unsuitable for removing colored noise. Therefore, it is necessary to convert colored noise into white noise, which necessitates adjusting the noise's energy distribution. Wavelet transform is well-suited to this energy distribution adjustment requirement because it can decompose a two-dimensional signal into a low-frequency approximate component and multiple high-frequency detail components. Figure 3This visually demonstrates the difference between white noise and colored noise in the wavelet domain. As can be observed from the figure, after wavelet transform, the standard deviations of the four wavelet sub-images (one low-frequency sub-image and three high-frequency sub-images) of white noise are consistent and equal to the standard deviation of the original white noise image. This phenomenon indicates that white noise maintains its characteristic of uniform energy distribution in the wavelet domain. In contrast, after wavelet transform, the standard deviation distribution of the wavelet sub-images of colored noise shows a significant imbalance. The standard deviation of the low-frequency sub-image in the upper left corner is significantly greater than that of the other three high-frequency sub-images, indicating that its energy is concentrated in the low frequencies.

[0057] Therefore, wavelet transform can effectively separate the low-frequency and high-frequency components of noise, intuitively reflecting the differences in noise energy distribution across different frequency bands. Based on this characteristic, wavelet transform is selected as the core method for controlling the energy distribution of colored noise. For Gaussian white noise, the standard deviation of each sub-band after wavelet transform remains consistent with the standard deviation of the original noise. Wavelet domain energy equalization adjusts the standard deviations of the four sub-images obtained from wavelet decomposition to the target standard deviation, thereby controlling the energy distribution of noise across different frequency bands. The specific implementation steps are as follows: First, regarding the noise signal The two-dimensional discrete wavelet transform is defined as: (twenty three) in This represents a two-dimensional discrete wavelet transform function that performs a first-order wavelet decomposition on the input noisy signal and outputs four sub-image components. Represents the low-frequency approximate component. , , These represent the high-frequency detail components in the horizontal, vertical, and diagonal directions, respectively.

[0058] "sym4" was selected as the wavelet basis function, as it offers both computational efficiency and accuracy. Other types of wavelet basis functions can also be chosen. Regarding the number of wavelet decomposition levels, only a single-level wavelet decomposition was used: although multi-level decomposition is also applicable, in most cases, single-level decomposition is sufficient to meet the requirements for noise energy control.

[0059] Next, calculate the standard deviation of the four wavelet sub-images (the magnitude of the standard deviation directly reflects the energy level of the corresponding sub-band): (twenty four) in This is a function for calculating the standard deviation of a matrix. , , , These correspond to the standard deviations of each wavelet sub-image. The standard deviation of the target Gaussian noise is defined as... The standard deviations of the four sub-images are adjusted to match the standard deviation of the target Gaussian noise. (25) in , , , The adjusted wavelet sub-image, , , , This is the corresponding energy adjustment coefficient.

[0060] Finally, performing an inverse two-dimensional discrete wavelet transform on the adjusted wavelet sub-image yields the noise signal after energy equalization in the wavelet domain. : (26) in This represents the inverse two-dimensional discrete wavelet transform function.

[0061] It should be noted that the energy distribution of spatially correlated noise is usually uneven, therefore wavelet domain energy equalization is very effective in handling the spatial correlation of noise. However, similar to frequency domain histogram matching, to avoid destroying the texture, wavelet domain energy equalization is only performed when the noise is spatially correlated and the number of iterations is greater than 1, as follows: (27) in That is the final conversion noise.

[0062] S3, Noisy image reconstruction: The converted noise is fused with the smoothed image to obtain a converted noisy image; After noise conversion, the converted noise is added to the previously obtained smoothed image to obtain the converted noise image. The formula is as follows: (28) S4, Denoising: The converted noise image is denoised using a preset denoiser to obtain a denoised result image; Determine whether the initial noise has spatial correlation. If not, input the transformed noise image into a pre-trained target level Gaussian denoiser to obtain the initial denoised image. If the initial noise has spatial correlation, the transformed noise image is first subjected to pixel rearrangement downsampling processing, and then input into a pre-trained target level Gaussian denoiser to obtain a downsampled denoised image; the downsampled denoised image is then subjected to pixel rearrangement upsampling processing to restore it to the original image size to obtain an upsampled denoised image. Perform random replacement refining process: Using a binary mask with a preset probability, some pixels in the upsampled denoised image are randomly replaced with corresponding pixels in the converted noisy image. Repeat the random replacement and Gaussian denoising process a preset number of times, and take the average value of the results of multiple denoising processes to obtain the denoised result image.

[0063] Specifically, if the noise does not have spatial correlation, the converted noisy image can be directly processed. After denoising, the data is input into a Gaussian denoiser at the target level, as follows: (29) in, Indicates the noise reduction level as The Gaussian denoiser can be selected from either a model trained solely on the target level of Gaussian noise or a flexible Gaussian denoiser specified as the target denoising level. This is the output of the denoising process.

[0064] If the noise has spatial correlation, then perform pixel rearrangement downsampling, Gaussian denoising, pixel rearrangement upsampling, and random replacement refining operations in sequence.

[0065] 1. Pixel rearrangement downsampling. To further break the spatial correlation of noise, the converted noisy image needs to be downsampled before formal denoising. Perform a pixel reordering downsampling operation. This operation first downsamples the noisy image into four sub-images, and then reassembles these four sub-images into a complete image to reduce the spatial correlation of noise, as shown below: (30) in This represents a pixel rearrangement downsampling operation. This is the image after pixel rearrangement and downsampling.

[0066] 2. Target horizontal Gaussian denoising. This involves downsampling the image after pixel rearrangement. Performing Gaussian denoising at the target denoising level yields a preliminary denoised image. : (31) 3. Pixel rearrangement and upsampling. After noise reduction, [the following steps are performed]. Perform a pixel reordering upsampling operation to restore the image to its original pixel arrangement order: (32) PU (Pixel-shuffle Upsampling) is the pixel rearrangement upsampling operation. This is the upsampled and restored denoised image.

[0067] 4. Randomly replace and refine. Restore the image. Although the data contains almost no noise, artifacts can occur due to pixel rearrangement downsampling. To eliminate these artifacts, a random replacement refining operation is introduced, which uses probability... Restore the image The pixels in the image are randomly replaced with converted noise. The formula for the corresponding pixel in the formula is: (33) in For the first The image after random replacement For the first The second type of binary mask image has pixel values ​​that are only 0 or 1, as defined below: (34) in express The coordinates of the middle pixel.

[0068] The random replacement operation needs to be repeated. Next, can be set The image obtained from each replacement. All samples need to be input again into the target level Gaussian denoiser for refinement. Finally, the average of all denoised results is taken to obtain the refined denoised result, as shown in the following expression: (35) in This is to randomly replace the refined denoising result.

[0069] Finally, define the conditions. If "the noise has spatial correlation", then the final denoised image It can be represented as: (36) S5, Texture Restoration: Based on the forward conversion rule used in the noise distribution conversion process in step S2, reverse restoration processing is performed on the texture part corresponding to the initial noise after conversion and denoising in the difference between the denoised result image and the current smooth image to obtain the restored texture; The difference between the denoised image and the current smoothed image is obtained as the texture component to be restored; Based on the forward transformation coefficients used in the noise distribution transformation process described in step S2, the corresponding inverse transformation coefficients are determined. The inverse transformation coefficients are used to perform an inverse transformation on the texture components to be restored, thereby obtaining the restored texture.

[0070] Since real noise is difficult to obtain, the noise to be converted inevitably contains texture information, and the noise conversion process will change these textures. Therefore, a texture conversion operation is introduced to restore the texture information that has been changed during the noise conversion process.

[0071] The core principle of texture conversion is to perform inverse equivalent restoration of the texture by following the conversion rules and order of noise. The inverse restoration process is performed on the texture portion corresponding to the converted and denoised initial noise in the difference between the denoised image and the current smoothed image, including at least one of the following: Wavelet domain texture transformation: After performing wavelet decomposition on the texture components in the difference, divide each sub-band component by the energy adjustment coefficient corresponding to the noise transformation process in step S2, and then perform inverse wavelet transform; In the wavelet domain energy equalization operation of noise transformation, noise is transformed by multiplying it by an energy adjustment coefficient in the wavelet domain; correspondingly, texture can be inversely restored simply by dividing it by the same coefficient in the wavelet domain. Let... For noise The texture components contained therein are calculated as follows: (37) right Perform a two-dimensional discrete wavelet transform: (38) in , , , for The wavelet components. Divide each component by its corresponding energy adjustment coefficient: (39) Performing the inverse two-dimensional discrete wavelet transform (IDWT2) on the wavelet components after coefficient correction yields the wavelet domain texture conversion result: (40) in The texture is after wavelet domain transformation.

[0072] If the noise has not undergone wavelet domain energy equalization, then the texture does not need to be transformed accordingly. The final wavelet domain texture transformation formula can be uniformly expressed as: (41) In the formula, The final output of the wavelet domain transformation also represents the noise. The texture components contained therein.

[0073] Spatial texture conversion: Guided filtering is performed on the coefficient matrix and bias matrix corresponding to the spatial histogram matching in step S2 noise conversion, and the texture components are inversely transformed based on the processed coefficient matrix and bias matrix; When noise is not subjected to frequency domain histogram matching, its spatial domain transformation formula is: (42) definition For noise The texture components in the text, and the corresponding texture conversion formulas are: (43) in , These are the coefficient matrix and bias matrix for texture transformation, respectively. Both are unknown and are crucial for achieving texture restoration. Theoretically, and , and There is a strong correlation. If and If all are constant matrices, then equal , equal However, in practical applications, and Since it is a non-constant noisy matrix, it cannot be combined with... and Direct equivalence. Therefore, guided filtering is used. and Denoising is performed to approximate the estimate. and And select the denoised texture. This serves as the guiding image. Given that guided filtering has a weak ability to suppress outliers, outlier imputation must be performed before filtering. (44) in The outlier fill function finds outliers and fills them with the nearest normal value. , This is the result of outlier handling.

[0074] Perform guided filtering on the preprocessed matrix: (45) In the formula, This indicates a guided filtering operation. To guide the image, The image to be filtered.

[0075] Get and After approximating the value, the original texture is restored through inverse operations: (46) in This is the matrix element-wise division operator.

[0076] Spatial-frequency domain joint texture transformation: The inverse transformation of spatial domain histogram matching and the inverse transformation of frequency domain histogram matching are performed sequentially on the texture components.

[0077] If the noise transformation involves frequency domain histogram matching, the texture needs to undergo a joint spatial-frequency domain transformation. Combining formulas (11), (12), (13) and (22), the joint spatial-frequency domain transformation formula for noise can be summarized as follows: (47) Substitute formula (11) into the above equation and expand: (48) To simplify the calculation, intermediate variables are defined and the formula is restructured: (49) in This is intermediate-state noise. Let this be the total bias matrix including the frequency domain bias. for The texture components in the image are converted using the following formula: (50) , These are the corresponding coefficient and bias matrices. Referring to the spatial domain transformation process, first... and Perform exception handling: (51) Then, approximate estimation is performed using guided filtering. and : (52) This allows us to calculate intermediate texture. : (53) After completing the spatial domain transformation, perform the inverse frequency domain transformation. , and Perform a two-dimensional discrete cosine transform: (54) in , and These are the respective frequency domain transformation results. The frequency domain transformation relationship of the noise satisfies: (55) Correspondingly, the texture frequency domain conversion formula is: (56) The coefficient matrix for texture frequency domain transformation, and its sum There is also a strong correlation.

[0078] Unlike the airspace processing flow, outlier handling here is deferred to the end of the flow and performed directly on... Perform guided filtering to approximate estimation : (57) Then, we can obtain : (58) Because the frequency domain and spatial domain have different characteristics, outliers in the frequency domain correspond to periodic noise of a specific frequency and direction in the spatial domain. Therefore, to ensure that no periodic noise interference is introduced, it is necessary to perform [further analysis] before transforming to the spatial domain. Perform outlier handling: (59) The result after outlier processing is then obtained by performing an inverse two-dimensional discrete cosine transform. : (60) S6, Iterative Update: The restored texture is fused with the current smooth image to update the smooth image, and the process returns to step S1 to repeat the initial noise estimation and steps S2 to S5 until the preset iteration conditions are met, and the final denoised image is output; wherein the initial noise is obtained by using the difference between the updated current smooth image and the original noise image during the iteration process.

[0079] After the texture conversion is complete, the restored texture will be... With smooth image The addition process enhances the texture of the smooth image and performs iterative updates, providing a better foundation for the next round of noise conversion. The update formula is as follows: (61) in For the updated smooth image, it will be replaced Proceed to the next iteration until the set number of iterations is reached.

[0080] To facilitate understanding of the technical solution and illustrate the technical effects of this invention, the technical solution of this invention is used to conduct denoising experiments on images of relevant datasets and to verify the effects. The specific implementation is as follows: The algorithm's parameters are set as follows: target Gaussian noise standard deviation. The interval interval of the noise histogram is L=0.002, the block size of the local histogram matching is g=45, the block step size is s=40, the reference block size of the local histogram matching combined with non-local self-similarity is m=12, the number of similar blocks is q=70, the number of random replacement refinements is R=4, the random replacement probability of the first iteration is p=0.3, the random replacement probability of subsequent iterations is p=0.55, and the number of iterations is set to 3. The experiment adopts an adapted noise transformation strategy according to the noise characteristics: global histogram matching is used for global noise; local histogram matching is used for local noise, and if the local noise also has signal correlation, non-local self-similarity is also combined; frequency domain histogram matching, wavelet domain energy equalization, and pixel rearrangement downsampling are used for spatially correlated noise. For real noise, since it has both signal correlation and spatial correlation, all operations are used. It is worth noting that real noise also has channel correlation, which is addressed in this embodiment by extending the two-dimensional DCT in frequency domain histogram matching to three-dimensional DCT.

[0081] Datasets and Evaluation Metrics: To verify the effectiveness of the proposed noise conversion method, this invention conducts experimental verification in two scenarios: synthetic noise and real noise. For the synthetic noise settings, multiple types of out-of-distribution noise were synthesized on the Kodak24 and McMaster datasets, specifically including: 1) Global noise: Gaussian white noise with a standard deviation σ=25; 2) Local noise: Salt and pepper noise with a density d=0.2. The noise consists of: 1) S&P (Sensitive Impulse Noise), 2) Random Impulse Noise with Density d=0.2, Bernoulli Noise with Density d=0.2, Poisson Noise with Intensity λ=25, and Speckle Noise with Standard Deviation σ=55, which is generated by multiplying a clean image pixel-by-pixel with Gaussian Noise with Standard Deviation σ=55; 2) Spatial Correlated Noise: Circular Noise with Standard Deviation σ=25, which is obtained by convolving a circular kernel with Gaussian White Noise and has strong spatial correlation; 3) Mixed Noise: It is obtained by mixing channel and spatially correlated Gaussian Noise with σ=15 (channel correlated Gaussian Noise is filtered by 3×3 mean and then the standard deviation is readjusted to σ=15), Speckle Noise with σ=15, and Salt and Pepper Noise with d=0.02. In the real noise verification, the SIDD verification set was selected as the test benchmark. This dataset contains 1280 real noise images with reference ground truth values, covering various shooting scenarios and different exposure time conditions. Its noise characteristics have signal correlation, spatial correlation and channel correlation, and its complexity is significantly higher than that of conventional synthetic noise. It is difficult to effectively remove noise using traditional Gaussian denoising models, which can fully verify the effectiveness of the noise conversion method of the present invention.

[0082] Regarding evaluation metrics, the technical solution of this invention indirectly measures the effectiveness of noise conversion by using the denoising effect of a Gaussian denoiser. Given that existing Gaussian denoisers are all designed for Gaussian noise, the better their denoising effect on the converted noisy image, the higher the quality of the noise conversion. To quantify the denoising effect, the embodiments of this invention select Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity (SSIM), commonly used in image denoising, as evaluation metrics. These two metrics can effectively measure the quality of denoising from the pixel error and structural levels, respectively.

[0083] Comparison Methods: To verify the applicability of the noise conversion method to different Gaussian denoisers, this invention's embodiments selected several Gaussian denoisers for comparison, including the traditional denoising method CBM3D, CNN-based methods FFDNet, DnCNN, DRUNet, and SCUNet, the Transformer-based method Restormer, the Mamba-based method MambaIRv2, and the diffusion model-based method DMID. To verify the superiority of the noise conversion method, this invention's embodiments were compared with other methods for improving the generalization ability of Gaussian denoisers, including LTN, IDF, and Clip-denoising. Finally, this invention's embodiments also compared three self-supervised denoising methods: AP-BSN, TBSN, and MASH.

[0084] Quantitative comparison It should be noted that the tables and figures below use the design of ours to represent the experimental results corresponding to the technical solution of this invention.

[0085] Tables 1 and 2 respectively present the quantitative comparison results of the technical solutions of the embodiments of the present invention and other existing technologies on the Kodak24, SIDD, and McMaster datasets. The methods of the present invention all use noise levels. The Gaussian denoiser performs the conversion and removal of out-of-distribution noise. The results show that the noise conversion framework of this invention can stably improve the performance of typical Gaussian denoising models such as CBM3D, DnCNN, FFDNet, DRUNet, SCUNet, and Restormer, achieving significant gains in scenarios with out-of-distribution noise. For example, in processing... When dealing with Gaussian noise, based only on The technical solution of this invention's embodiment of the Gaussian denoiser has a PSNR / SSIM index that is close to... The level of the supervised Gaussian denoiser validated the effectiveness of the noise transformation strategy. Furthermore, for the SCUNet method, its blind denoising model, SCUNet-real, was trained with synthetic mixed noise. While it performed well in both synthetic and real noise scenarios, it failed to effectively remove noise types such as salt-and-pepper noise. In contrast, SCUNet-ours, trained solely with Gaussian noise, achieved superior overall denoising performance.

[0086] Further analysis reveals a positive correlation between the performance of the Gaussian denoiser on Gaussian denoising tasks and the effectiveness of the technical solutions implemented in this invention based on it. Models that perform better in Gaussian denoising tasks (such as MambaIRv2 and DMID) typically achieve better denoising results in complex scenarios such as Poisson noise, circular noise, and real noise with their corresponding "-ours" versions. This pattern is validated for most noise types. However, this pattern is not absolute. For example, in Bernoulli noise scenarios, the performance of MambaIRv2-ours is significantly inferior to FFDNet-ours, and even lower than CBM3D-ours. This is related to the robustness of the Gaussian denoiser: Gaussian denoisers with flexible denoising levels are generally more robust than those with fixed denoising levels, thus adapting to more extreme noise distributions. This indicates that the performance of the technical solutions in this invention depends not only on the Gaussian denoising accuracy of the denoiser but also on its generalization robustness to non-Gaussian noise. Based on this, the technical solution of this invention recommends using a flexible Gaussian denoiser to construct the noise conversion framework: on the one hand, this type of denoiser has better robustness; on the other hand, different target noise levels can be flexibly configured according to actual needs, which has stronger practicality and adaptability.

[0087] Table 1. Quantitative comparison results of the method of the present invention and existing methods on the Kodak24 dataset and the SIDD real noise dataset.

[0088] Note: The technical solutions in the embodiments of the present invention all use a Gaussian denoiser with σ=15. The values ​​in the table are PSNR / SSIM indicators, and the bolded indicators are the highest values ​​in each column.

[0089] Among the methods for improving the generalization ability of Gaussian denoisers, Clipdenoising trained two denoising models using Gaussian noise and Gaussian-Poisson noise, as shown in the table below. Clipdenoising-real and Clipdenoising-ours are designed for synthetic noise scenarios and realistic noise scenarios, respectively. The former has poor generalization ability in realistic noise scenarios, while the latter struggles to effectively handle diverse synthetic noise. Clipdenoising-ours, on the other hand, only uses... The Gaussian denoising model achieved balanced and excellent performance in both synthetic and real noise scenarios. While IDF did improve the generalization of the Gaussian denoiser, the improvement was limited, similar to the Clipdenoising method. The noise transformation module learned by LTN can adapt well to real noise, but it is difficult to apply to various synthetic noises, and its application scenarios are significantly limited.

[0090] For real-noise denoising methods such as AP-BSN and TBSN, their performance is outstanding in real-noise scenarios, but their performance drops sharply in synthetic noise scenarios, and their generalization ability is obviously lacking. Although the zero-shot method MASH has the advantage of generalization, its denoising performance is poor.

[0091] A comprehensive comparison shows that the Gaussian denoising model after combining the noise conversion algorithm of the technical solution of the present invention can achieve a better performance balance between synthetic noise and real noise, effectively expanding the applicable scope of traditional Gaussian denoisers. The quantitative comparison results fully verify its effectiveness and universality.

[0092] Table 2. Quantitative comparison results of the method of this invention and existing methods on the McMaster dataset.

[0093] Note: The technical solutions in the embodiments of the present invention all use a Gaussian denoiser with σ=15. The values ​​in the table are PSNR / SSIM indicators, and the bolded indicators are the highest values ​​in each column.

[0094] Qualitative comparison Figure 4 and Figure 5 The figures show a visual comparison of the denoising results of the present invention and existing technologies in synthetic noise scenarios. Figure 4 illustrates the denoising effect and detail preservation capability of the present invention under difficult-to-distribute out-of-noise conditions such as salt-and-pepper noise, Poisson noise, and circular spatially correlated noise. Therefore, a magnified local image is used for comparison to facilitate observation of edge and texture restoration quality. Figure 5 demonstrates the generalization capability of the present invention under more types of out-of-noise conditions, aiming to reflect the wide applicability of the method. Therefore, a global image plus a magnified local image is used for comparison. As can be seen from the figures, for Bernoulli... Salt and pepper ,spot Poisson To address the challenges posed by various out-of-distribution synthetic noises, including mixed noise, the noise conversion algorithm of the present invention demonstrates strong generalization ability and robustness, achieving high-quality noise suppression and image restoration, with significant performance improvements compared to baseline methods. Baseline methods, on the other hand, perform poorly in handling such out-of-distribution noise, struggling to effectively remove various types of noise interference. Figure 6This document presents a visual comparison of the denoising results of the present invention and existing technologies in real-world noisy scenes. Despite the more complex distribution of real-world noise, the noise conversion algorithm of the present invention still demonstrates good applicability. A comparison of various denoising methods shows that Gaussian denoisers combined with noise conversion (such as SCUNet-ours and DMID-ours) significantly improve denoising performance. While effectively removing noise, they can better preserve the edge structure and texture details of the image, further validating the versatility of the noise conversion method.

[0095] Real noise is far more complex than synthetic noise, and conventional Gaussian denoising methods often struggle to remove it effectively. However, by employing the noise conversion strategy of this invention, the Gaussian denoiser achieves superior noise reduction performance.

[0096] The proposed noise conversion algorithm is not only applicable to the noise types described in this paper, but also has a wider range of noise adaptability, such as... Figure 7 As shown in the figure, the algorithm's conversion effects on Gaussian-Poisson mixed noise, Gaussian salt-and-pepper mixed noise, grid periodic noise, mixed periodic noise (a mixture of periodic noises of various frequencies and directions), and microscopic noise are demonstrated. The figure also verifies its applicability to grayscale noise images. As can be seen from the figure, the algorithm can achieve effective noise conversion in various complex noise scenarios, further demonstrating the versatility of the proposed method.

[0097] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for image denoising based on noise distribution transformation, characterized in that, Includes the following steps: S1, Initial noise estimation: The original noisy image is smoothed to obtain a smoothed image, and the initial noise is obtained based on the difference between the original noisy image and the smoothed image; S2, Noise Distribution Transformation: The initial noise is transformed using histogram matching technology to convert the statistical distribution of the initial noise into the target noise distribution characteristics, and the transformed noise is obtained; wherein, the target noise distribution characteristics are matched with the noise distribution characteristics adapted by the preset denoiser; S3, Noisy image reconstruction: The converted noise is fused with the smoothed image to obtain a converted noisy image; S4, Denoising: The converted noise image is denoised using a preset denoiser to obtain a denoised result image; S5, Texture Restoration: Based on the forward conversion rule used in the noise distribution conversion process in step S2, reverse restoration processing is performed on the texture part corresponding to the initial noise after conversion and denoising in the difference between the denoised result image and the current smooth image to obtain the restored texture; S6, Iterative Update: The restored texture is fused with the current smooth image to update the smooth image, and the process returns to step S1 to repeat the initial noise estimation and steps S2 to S5 until the preset iteration conditions are met, and the final denoised image is output; wherein the initial noise is obtained by using the difference between the updated current smooth image and the original noise image during the iteration process.

2. The image denoising method based on noise distribution conversion according to claim 1, characterized in that, Step S2 includes adjusting the statistical distribution of the initial noise over at least one transform domain.

3. The image denoising method based on noise distribution conversion according to claim 2, characterized in that, Step S2 includes performing a histogram matching operation in the spatial domain: In response to the initial noise being determined to be global noise, global histogram matching is performed on the initial noise: a uniform transformation coefficient is determined based on the overall statistical distribution of the initial noise, and a distribution transformation is performed on the initial noise as a whole; In response to the initial noise being determined to be local noise, local histogram matching is performed on the initial noise: the initial noise is divided into multiple image blocks, and transformation coefficients are determined based on the local statistical distribution of each image block to perform block transformation.

4. The image denoising method based on noise distribution conversion according to claim 3, characterized in that, In the local histogram matching process, the initial noise is divided into multiple image blocks, histogram matching is performed on each image block, and all image blocks that have completed histogram matching are reconstructed into a complete image. Specifically, for local noise with signal correlation, based on the non-local self-similarity of the image, a preset number of similar image blocks are matched for each target image block, and histogram matching processing is performed based on the pixel set of the target image block and all similar image blocks.

5. The image denoising method based on noise distribution conversion according to any one of claims 2-4, characterized in that, Step S2 also includes performing a histogram matching operation in the frequency domain: Perform a discrete cosine transform on the initial noise to obtain the frequency domain noise; Histogram matching is performed on the frequency domain noise so that its frequency domain statistical distribution approximates the statistical distribution of the target noise distribution characteristics in the frequency domain.

6. The image denoising method based on noise distribution conversion according to any one of claims 2-4, characterized in that, Step S2 also includes performing an energy equalization operation in the wavelet domain: Perform wavelet transform on the initial noise to obtain multiple sub-band components; The energy amplitude of each sub-band component is adjusted to make the energy distribution of each sub-band component consistent with the energy distribution characteristics of the target noise distribution.

7. The image denoising method based on noise distribution transformation according to any one of claims 1 to 4, characterized in that, The target noise distribution features are those of additive white Gaussian noise; the preset denoiser is a denoising neural network model trained for the additive white Gaussian noise; step S1 specifically includes the following steps: S11, perform median filtering on the input original noisy image to remove abnormal pixel values ​​in the image and obtain the filtered image; S12, The filtered image is smoothed using the RTV image smoothing method to separate the structural information and texture information of the image, thereby obtaining the smoothed image; S13, calculate the pixel difference between the original noise image and the smooth image to obtain an initial noise map, and add Gaussian noise with a preset standard deviation to the initial noise map to obtain the initial noise.

8. The image denoising method based on noise distribution conversion according to claim 7, characterized in that, Step S4 specifically includes the following steps: Determine whether the initial noise has spatial correlation. If not, input the transformed noise image into a pre-trained target level Gaussian denoiser to obtain the initial denoised image. If the initial noise has spatial correlation, the transformed noise image is first subjected to pixel rearrangement downsampling processing, and then input into a pre-trained target level Gaussian denoiser to obtain a downsampled denoised image; the downsampled denoised image is then subjected to pixel rearrangement upsampling processing to restore it to the original image size to obtain an upsampled denoised image. Perform random replacement refining process: Using a binary mask with a preset probability, some pixels in the upsampled denoised image are randomly replaced with corresponding pixels in the converted noisy image. Repeat the random replacement and Gaussian denoising process a preset number of times, and take the average value of the results of multiple denoising processes to obtain the denoised result image.

9. The image denoising method based on noise distribution transformation according to claim 8, characterized in that, Step S5 includes: The difference between the denoised image and the current smoothed image is obtained as the texture component to be restored; Based on the forward transformation coefficients used in the noise distribution transformation process described in step S2, the corresponding inverse transformation coefficients are determined. The inverse transformation coefficients are used to perform an inverse transformation on the texture components to be restored, thereby obtaining the restored texture.

10. The image denoising method based on noise distribution transformation according to claim 9, characterized in that, Perform inverse restoration processing on the texture portion corresponding to the converted and denoised initial noise in the difference between the denoised result image and the current smoothed image, including at least one of the following: Wavelet domain texture transformation: After performing wavelet decomposition on the texture components in the difference, divide each sub-band component by the energy adjustment coefficient corresponding to the noise transformation process in step S2, and then perform inverse wavelet transform; Spatial texture transformation: Guided filtering is performed on the coefficient matrix and bias matrix corresponding to the spatial histogram matching in step S2 noise transformation, and the texture components are inversely transformed based on the processed coefficient matrix and bias matrix; Spatial-frequency domain joint texture transformation: The inverse transformation of spatial domain histogram matching and the inverse transformation of frequency domain histogram matching are performed sequentially on the texture components.

Images