A blind image denoising training data synthesis method based on multivariate noise mixing and random degradation sequence

Abstract

The application discloses a kind of blind image denoising training data synthesis methods based on multivariate noise mixing and random degradation sequence, it is related to image processing technical field.The application includes: constructing multivariate degradation operator library, operator library includes: noise class degradation operator and non-noise class degradation operator;Controllable random sampling mechanism is configured, the execution probability of each operator in degradation operator library and its parameter range are pre-set and random sampling;The execution order of the degradation operator sampled is randomly rearranged, and multiple random degradation sequences are generated;Based on multiple random degradation sequences, apply degradation operation to the input high-quality clean image, generate noisy degraded image, and form image pair with original clean image for training blind image denoising model.The application can synthesize training data with authenticity and diversity under controllable conditions, and cover complex degradation combinations, to improve the robustness and generalization ability of general blind image denoising model in unknown scenarios.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a method for synthesizing blind image denoising training data based on multivariate noise mixing and random degradation sequences. Background Technology

[0002] Image denoising is a fundamental challenge in digital image processing, aiming to recover a clear underlying image from observed images affected by noise and degradation. In recent years, deep learning-based denoising methods have made significant progress, but their performance is highly dependent on the degradation distribution covered by the training data. Existing training data synthesis methods are mainly divided into three categories: The first category is noise modeling methods based on physical statistics, which use the physical model of the camera imaging link to model photon shot noise, readout noise, etc. However, the actual sources of noise are complex and interdependent, making it difficult for physical models to fully characterize their true statistical distribution and spatial correlation. Furthermore, calibrating different sensors individually is costly.

[0003] The second category is data-driven learning-based noise modeling methods, which utilize generative adversarial networks (GANs) to learn distributions from real noise. However, this method is highly dependent on hard-to-obtain "noise-clean" image pairs and suffers from problems such as unstable training, generation artifacts, and poor interpretability.

[0004] The third category is modeling methods based on random degradation pipelines, which combine noise with operations such as blurring and compression. However, these methods are mostly designed for other image restoration tasks, and when directly transferred to denoising tasks, they suffer from problems such as insufficient noise coverage and inconsistencies between the degradation process and the denoising target.

[0005] Therefore, proposing a blind image denoising training data synthesis method based on multivariate noise mixing and random degradation sequence to solve the difficulties of existing technologies is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] In view of this, the present invention provides a method for synthesizing training data for blind image denoising based on multivariate noise mixing and random degradation sequences. This method can synthesize training data that is both realistic and diverse and covers complex degradation combinations under controllable conditions, thereby improving the robustness and generalization ability of general blind image denoising models in unknown scenarios.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: A method for synthesizing blind image denoising training data based on multivariate noise mixing and random degradation sequences includes: S1. Construct a multivariate degradation operator library, which includes noise-type degradation operators and non-noise-type degradation operators; S2. Configure a controllable random sampling mechanism to preset and randomly sample the execution probability and parameter range of each operator in the degenerate operator library; S3. Randomly rearrange the execution order of the sampled degradation operators to generate multiple random degradation sequences; S4. Based on multiple random degradation sequences, degradation operations are applied to the input high-quality clean image to generate a noisy degraded image, which, together with the original clean image, forms an image pair for training a blind image denoising model.

[0008] In the above method, optionally, in S1, the noise-type degradation operators include: Gaussian noise, Poisson noise, speckle noise, and camera sensor noise; Non-noise degradation operators include: JPEG compression noise and image resampling.

[0009] The above methods, optionally, include a Gaussian noise modeling method, specifically: A zero-mean multivariate Gaussian distribution is adopted, and the noise correlation between the three RGB color channels is explicitly modeled through a 3×3 covariance matrix. The covariance matrix is ​​obtained by random sampling and normalization of a symmetric positive semidefinite matrix. It is used to generate colored Gaussian noise with arbitrary cross-channel correlation, or to generate colored Gaussian noise with independent channels, or to generate grayscale Gaussian noise with all channels sharing the same noise component.

[0010] The above methods, optionally, include a modeling method for Poisson noise, specifically: The scale index is uniformly sampled within a preset interval, the scale parameter is calculated, and the clean image is enlarged according to the scale parameter. Apply Poisson random sampling to the magnified image and divide it by the same scaling factor to map the result back to the original dynamic range; In terms of channel processing, it supports independent Poisson sampling of each channel to generate color Poisson noise, or converting the image to grayscale to generate Poisson noise and then copying it to each channel to generate grayscale Poisson noise.

[0011] The above method, optionally, includes a modeling method for speckle noise, specifically: The noise intensity parameters are uniformly sampled from a preset interval to generate a Gaussian random noise field with zero mean and unit variance, and the noise field is applied to the original image in a multiplicative manner. In terms of channel processing, it supports generating Gaussian noise fields independently in each channel to generate colored speckle noise, or generating a single-channel noise field and copying it to each channel to generate grayscale speckle noise.

[0012] The above method, optionally, includes a modeling method for camera sensor noise, specifically: Construct and execute a closed-loop process of "reverse ISP transformation, RAW domain noise injection, and forward ISP reconstruction"; Inverse ISP transform is used to map a clean RGB image to a RAW domain mosaic image; RAW domain noise injection, based on the physical noise model of camera imaging, superimposes signal-independent readout noise and signal-dependent shot noise into the RAW domain mosaic image. Forward ISP reconstruction is used to reconstruct noisy RAW domain mosaic images into noisy RGB images. The same set of randomly generated ISP parameters is reused when generating clean and noisy samples.

[0013] The above method, optionally, includes a modeling method for JPEG compression noise, specifically: The quality factors are uniformly and randomly sampled from a preset range, and the image is then encoded and decoded using JPEG based on the sampled quality factors.

[0014] The above methods, optionally, include image resampling modeling methods, specifically: The scaling factor is uniformly and randomly sampled within a preset interval, and an interpolation method is randomly selected between bilinear interpolation and bicubic interpolation to downsample the input image and change its resolution. Then, the same interpolation method is used to upsample the image to restore it to the original resolution.

[0015] Optionally, in the above method, in S2, a controllable random sampling mechanism is configured, where the execution probability of Gaussian noise, camera sensor noise, and JPEG compression operators is higher than that of Poisson noise, speckle noise, and image resampling operators.

[0016] Optionally, in the above method, in S4, a degradation operation is applied to the input high-quality clean image based on multiple random degradation sequences to generate a noisy degraded image, which, together with the original clean image, forms an image pair for training the blind image denoising model. Specifically: Each time training samples are generated, degenerate rounds N are first randomly sampled from a preset range; For each round of degradation, a set of degradation operators is randomly sampled from the degradation operator library, and the execution order of the set of operators is randomly rearranged and executed sequentially. The same or different degradation operators can be repeatedly executed between different degradation rounds.

[0017] As can be seen from the above technical solutions, compared with the prior art, the present invention provides a blind image denoising training data synthesis method based on multivariate noise mixing and random degradation sequences, which has the following beneficial effects: The present invention constructs a unified operator library including multivariate noise (Gaussian, Poisson, speckle, sensor noise) and non-noise degradation (JPEG compression, resampling), and explicitly models cross-channel noise correlation through covariance matrix, making the synthesized data closer to the real imaging degradation process; through multi-round random degradation strategy, random rearrangement of operator execution order and random sampling of parameter range, the present invention greatly expands the degradation space in multiple dimensions such as degradation type, intensity, rounds and operation order, and significantly improves the diversity of training data; based on the highly diversified training data synthesized by the present invention, the blind denoising model can effectively adapt to the challenges of unknown degradation type, variable intensity and complex links in real scenes, and the generalization ability is better than the model trained using a single or fixed degradation process; the present invention only requires high-quality clean images to synthesize training data on a large scale, without the need to collect expensive real "noise-clean" image pairs, and has good scalability and economy. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0019] Figure 1 The flowchart illustrates a method for synthesizing blind image denoising training data based on multivariate noise mixing and random degradation sequences, provided by this invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Reference Figure 1 As shown, this invention discloses a method for synthesizing blind image denoising training data based on multivariate noise mixing and random degradation sequences, comprising: S1. Construct a multivariate degradation operator library, which includes noise-type degradation operators and non-noise-type degradation operators; S2. Configure a controllable random sampling mechanism to preset and randomly sample the execution probability and parameter range of each operator in the degenerate operator library; S3. Randomly rearrange the execution order of the sampled degradation operators to generate multiple random degradation sequences; S4. Based on multiple random degradation sequences, degradation operations are applied to the input high-quality clean image to generate a noisy degraded image, which, together with the original clean image, forms an image pair for training a blind image denoising model.

[0022] Furthermore, in S1, noise-type degradation operators include: Gaussian noise, Poisson noise, speckle noise, and camera sensor noise; Non-noise degradation operators include: JPEG compression noise and image resampling.

[0023] Furthermore, the modeling method for Gaussian noise is as follows: A zero-mean multivariate Gaussian distribution is adopted, and the noise correlation between the three RGB color channels is explicitly modeled through a 3×3 covariance matrix. The covariance matrix is ​​obtained by random sampling and normalization of a symmetric positive semidefinite matrix. It is used to generate colored Gaussian noise with arbitrary cross-channel correlation, or to generate colored Gaussian noise with independent channels, or to generate grayscale Gaussian noise with all channels sharing the same noise component.

[0024] Furthermore, additive white Gaussian noise has long been widely used in image denoising research due to its simple mathematical form and convenient modeling. However, the traditional Gaussian noise model assumes that the noise is independent of each color channel and has no spatial structure, which fails to reflect the cross-channel correlation introduced by the image signal processing link in the real imaging process. Denoising models trained solely on this type of noise often have insufficient generalization ability in real-world scenarios. To address the aforementioned problems, this invention employs zero-mean multivariate Gaussian noise to model color image noise, by introducing... The covariance matrix explicitly characterizes the noise correlation among the three color channels R, G, and B. For ease of description, the three-channel noise vector at any pixel location is... The present invention uses the following Gaussian noise model: ; in This is the covariance matrix, used to characterize cross-channel correlation; By randomly sampling the covariance matrix structure, this invention can uniformly describe the following noise scenarios and their transition forms: First, grayscale Gaussian noise, where all color channels share the same noise component. Zero-mean Gaussian noise is first generated in the spatial dimension and then copied to each color channel. Second, channel-independent color Gaussian noise, where each color channel is independent and Gaussian noise is sampled independently. Third, multivariate Gaussian noise with arbitrary cross-channel correlation. When generating multivariate Gaussian noise, a symmetric positive semi-definite random covariance matrix is ​​first constructed and normalized to match its energy with the sampled noise intensity. Then, a three-dimensional Gaussian vector is sampled at each spatial location based on the covariance matrix, thereby introducing a random but controlled correlation noise structure between the R, G, and B channels. Regarding noise type sampling, this invention generates the three types of noise with preset probabilities. Considering that the built-in denoising algorithms of actual camera systems usually prioritize suppressing color noise components, the final residual noise is often dominated by gray-level correlated noise. This invention assigns a higher sampling probability to gray-level Gaussian noise to enhance the consistency between the synthesized noise and the real scene. Furthermore, the modeling method for Poisson noise is as follows: The scale index is uniformly sampled within a preset interval, the scale parameter is calculated, and the clean image is enlarged according to the scale parameter. Apply Poisson random sampling to the magnified image and divide it by the same scaling factor to map the result back to the original dynamic range; In terms of channel processing, it supports independent Poisson sampling of each channel to generate color Poisson noise, or converting the image to grayscale to generate Poisson noise and then copying it to each channel to generate grayscale Poisson noise.

[0025] Furthermore, Poisson noise is used to describe noise caused by the photon shot effect during imaging. Its root cause lies in the discreteness and randomness of the process of photons reaching the image sensor. This type of noise is particularly significant under low-light conditions and is widely present in scenarios such as night photography, medical imaging, microscopic imaging, and astronomical observation. Unlike additive Gaussian noise, Poisson noise is a signal-dependent noise, and its variance varies with pixel intensity, making it difficult to model uniformly using fixed noise intensity parameters. To address the aforementioned characteristics, this invention directly models the generation process of Poisson noise during the training data synthesis stage. To simulate the continuous variation of Poisson noise intensity under different lighting conditions, this invention first performs amplitude scaling on the clean image: uniformly sampling the scale exponent within a preset interval. Calculate scale parameters The image intensity is then amplified according to this parameter to control the counting scale of the Poisson distribution; a larger scale parameter corresponds to high-light scenes with weaker Poisson noise, while a smaller scale parameter corresponds to low-light scenes with stronger Poisson noise. After amplitude scaling, this invention applies Poisson random sampling to the amplified image and divides it by the same scaling factor, mapping the result back to the original dynamic range. This process can be formally represented as: ; in, Indicates a clean image. For scale parameters, This represents the sampling operator for the Poisson distribution.

[0026] In terms of noise type design, this invention considers both color Poisson noise and grayscale Poisson noise, and randomly selects between the two according to a preset probability. Color Poisson noise performs independent Poisson random sampling on each of the RGB channels, and the noise intensity of each channel is determined by the pixel value of the corresponding channel. Grayscale Poisson noise first converts the color image into a grayscale image, generates Poisson noise in the grayscale domain, and then copies the noise to each color channel so that the R, G, and B channels share the same noise component.

[0027] Furthermore, the modeling method for speckle noise is as follows: The noise intensity parameters are uniformly sampled from a preset interval to generate a Gaussian random noise field with zero mean and unit variance, and the noise field is applied to the original image in a multiplicative manner. In terms of channel processing, it supports generating Gaussian noise fields independently in each channel to generate colored speckle noise, or generating a single-channel noise field and copying it to each channel to generate grayscale speckle noise.

[0028] Furthermore, speckle noise is a typical type of multiplicative noise, widely present in coherent imaging systems such as synthetic aperture radar imaging, medical ultrasound imaging, and other imaging scenarios based on the principle of phase interference. Unlike additive noise, the amplitude of speckle noise is directly related to the image signal strength. To address the aforementioned characteristics, this invention models speckle noise as a multiplicative degradation relationship between a clean image and a random noise field; specifically, it first uniformly samples noise intensity parameters within a preset interval. Subsequently, a Gaussian random noise field with zero mean and unit variance is generated in the spatial dimension. Finally, the noise is applied to the original image in a multiplicative manner; this process can be formally represented as: ; in, Indicates a clean image. This represents zero-mean, unit-variance Gaussian random noise. The normalized noise intensity parameter; the above modeling method reflects the essential characteristic of speckle noise changing with signal intensity: the noise amplitude is larger in bright areas and smaller in dark areas; In terms of modeling the noise channel structure, this invention also considers two cases: grayscale speckle noise and color speckle noise, and randomly selects between the two according to a preset probability. Grayscale speckle noise generates a two-dimensional Gaussian noise field in the spatial dimension, and copies the same noise component to each color channel to simulate speckle noise characteristics dominated by luminance noise and highly correlated across channels. Color speckle noise generates Gaussian noise independently on each color channel and applies it in a multiplicative form to characterize the case where the statistical characteristics of noise in different channels are relatively independent.

[0029] Furthermore, the modeling method for camera sensor noise is as follows: Construct and execute a closed-loop process of "reverse ISP transformation, RAW domain noise injection, and forward ISP reconstruction"; Inverse ISP transform is used to map a clean RGB image to a RAW domain mosaic image; RAW domain noise injection, based on the physical noise model of camera imaging, superimposes signal-independent readout noise and signal-dependent shot noise into the RAW domain mosaic image. Forward ISP reconstruction is used to reconstruct noisy RAW domain mosaic images into noisy RGB images. The same set of randomly generated ISP parameters is reused when generating clean and noisy samples.

[0030] Furthermore, to more realistically simulate the noise characteristics during camera imaging, this invention proposes a camera sensor noise synthesis method based on a closed-loop process of an image signal processor. This method treats the noise in the camera's output RGB image as the result of RAW domain sensor noise being processed by an image signal processor (ISP) and then propagated to the RGB domain. A noisy RGB image is generated by constructing a complete closed-loop process of "inverse ISP transformation → RAW domain noise injection → forward ISP reconstruction," and the ISP processing parameters are explicitly saved to ensure that clean and noisy samples of the same image pair use completely consistent processing configurations, eliminating color shifts caused by inconsistent ISP parameters. This process can be formally represented as: ; in, Represents a clean RGB image. The parameter is Inverse ISP transform, Indicates read noise. Indicates shot noise, The parameter is Positive ISP reconstruction, Represents a noisy RGB image; Inverse ISP transformation stage: First, inverse gamma correction is performed on the input normalized RGB image to map the sRGB color space to the linear intensity space; then, inverse tone mapping and inverse white balance correction are performed, dividing the RGB values ​​by the white balance gain coefficient of the corresponding channel; finally, the Bayer color filter array sampling mechanism is simulated to map the linear RGB image into a single-channel RAW mosaic image; specifically, a Bayer array conforming to the RGGB arrangement pattern is constructed: even-numbered rows and even-numbered columns are filled with R channel values, even-numbered rows and odd-numbered columns and odd-numbered rows and even-numbered columns are filled with G channel values, and odd-numbered rows and odd-numbered columns are filled with B channel values. RAW domain noise injection stage: Based on the physical noise model of camera imaging, two types of noise components are superimposed on the RAW image; the first type is readout noise. The first type is signal-independent noise, whose variance is calculated from randomly sampled ISO sensitivity values ​​and noise figures, and zero-mean Gaussian noise is generated based on this variance; the second type is shot noise. This is a signal-dependent noise, whose variance is proportional to the RAW signal intensity. It is used to characterize the Poisson statistical properties of the photon arrival process, and the Gaussian approximation is used under large signal conditions. The two types of noise are superimposed on the RAW image and then the amplitude is cropped to obtain a noisy RAW image. Forward ISP reconstruction stage: First, the noisy RAW image is de-mosaiced, and the three-channel RGB image is reconstructed using an edge-aware interpolation algorithm guided by local gradients; then, white balance, tone mapping and gamma correction are performed sequentially. The key technical feature of this invention lies in the consistency guarantee mechanism of ISP parameters, which randomly generates a set of ISP parameters during the initial processing. This includes the tone mapping curve type, tone gamma parameters, and RGB three-channel white balance gain coefficients, and this parameter set is persistently saved; when generating clean and noisy samples for the same image, the saved parameter set is strictly reused; through this mechanism, the difference between training sample pairs strictly originates from the physical noise injected in the RAW domain.

[0031] Furthermore, the modeling method for JPEG compression noise is as follows: The quality factors are uniformly and randomly sampled from a preset range, and the image is then encoded and decoded using JPEG based on the sampled quality factors.

[0032] Furthermore, JPEG compression, as a lossy encoding method widely used for image storage and transmission on the Internet, introduces compression artifacts such as blockiness, ringing, and loss of high-frequency details into the image during its compression and quantization process; the degree of distortion is affected by the quality factor. Control: The lower the quality factor, the greater the compression intensity and the more obvious the artifacts; since images in real-world scenes often undergo varying degrees of JPEG compression, relying solely on training data that does not consider compression artifacts can easily lead to insufficient generalization ability of blind denoising models on real internet images. To address the aforementioned problems, this invention introduces a JPEG compression degradation operator during the training data synthesis stage; specifically, it adjusts the quality factor within a preset interval [20, 95]. Uniform random sampling is performed, and JPEG encoding and decoding are applied to the image based on the sampled quality factor. Before compression, the normalized image representation is mapped to an 8-bit integer representation to meet the input format requirements of the JPEG encoder. After encoding and decoding, the result is mapped back to a normalized floating-point representation. This process can be formally represented as follows: ; in, Indicates the input image. The quality factor is The JPEG encoding process, This indicates the corresponding decoding process; Through the above-described encoding-decoding closed loop, this invention reproduces the distortion characteristics such as block effects, ringing effects, and loss of detail generated during JPEG compression without explicitly constructing a frequency domain quantization or block transform mathematical model; by randomly sampling the quality factor, the training data can cover a variety of degradation scenarios from light compression to heavy compression.

[0033] Furthermore, the modeling method for image resampling is as follows: The scaling factor is uniformly and randomly sampled within a preset interval, and an interpolation method is randomly selected between bilinear interpolation and bicubic interpolation to downsample the input image and change its resolution. Then, the same interpolation method is used to upsample the image to restore it to the original resolution.

[0034] Furthermore, image resampling, as one of the most common geometric operations in digital image processing, is widely used in image acquisition, storage, display, and network transmission. Although under ideal conditions, resampling itself does not introduce new random noise, when the image already contains noise, the scaling process significantly alters the statistical distribution and spatial correlation of the noise. During upsampling, interpolation introduces inter-pixel correlation, making spatially independent noise exhibit spatially correlated characteristics. During downsampling, resampling changes the coupling relationship between signal and noise, making the noise appear in a smoother form at the pixel level. Therefore, image resampling should be considered an important degradation operation, rather than merely a geometrically significant size transformation. Based on the above understanding, this invention explicitly introduces a random resampling operation in the training data synthesis stage; specifically, it first uniformly and randomly samples scaling factors from a preset interval. The process involves randomly selecting one of two interpolation methods—bilinear interpolation or bicubic interpolation—to scale the input image; subsequently, the same interpolation method is used to restore the image to its original resolution. This process can be formally represented as: ; in, Indicates the input image. Scaling factor Indicated by factor Perform downsampling. This indicates an upsampling operation to restore the original resolution; By using a "scaling-restoration" closed-loop processing, the spatial correlation change of noise caused by resampling is introduced while keeping the output size unchanged, simulating the degradation effect experienced by real images during multiple scaling processes. During the construction of training data, the resampling operation is always applied to both the noisy image and its corresponding clean image to ensure that the two are strictly consistent in spatial resolution and geometric structure.

[0035] Furthermore, in S2, a controllable random sampling mechanism is configured, with the execution probability of Gaussian noise, camera sensor noise, and JPEG compression operators being higher than that of Poisson noise, speckle noise, and image resampling operators. To balance degradation diversity and training stability, this invention sets differentiated execution probabilities for different types of degradation operators. Gaussian noise, camera sensor noise, and JPEG compression, as the most common degradation types in real images, are executed with a higher probability; while Poisson noise, speckle noise, and image resampling operations are triggered randomly with a low probability to introduce richer degradation variations. Through the synergistic effect of multi-round degradation strategies and random sequence generation mechanisms, the degradation space constructed by this invention has high diversity in dimensions such as noise type, degradation intensity, degradation rounds, and operation order. It can effectively cover various complex degradation combinations that real images may undergo, providing a data foundation for training blind denoising models with strong generalization capabilities. In addition, since the degradation process introduces scaling and interpolation processes, in order to avoid the loss of boundary information and interpolation artifacts, this invention uses high-resolution, high-quality images as the initial input for training data synthesis.

[0036] Furthermore, in S4, based on multiple random degradation sequences, a degradation operation is applied to the input high-quality clean image to generate a noisy degraded image. This noisy degraded image, together with the original clean image, forms an image pair used to train the blind image denoising model. Specifically: Each time training samples are generated, degenerate rounds N are first randomly sampled from a preset range; For each round of degradation, a set of degradation operators is randomly sampled from the degradation operator library, and the execution order of the set of operators is randomly rearranged and executed sequentially. The same or different degradation operators can be repeatedly executed between different degradation rounds.

[0037] Furthermore, in real-world scenarios, images often undergo a cascading effect of multiple processing operations during acquisition, editing, storage, and transmission. For example, a photo may go through multiple stages, such as cropping and scaling, compression on social media platforms, and rescaling to adapt to different display devices. Each stage has a cumulative impact on image quality. In addition, the order of these processing operations is not fixed in different application scenarios: compression may occur before or after resizing, and noise may be introduced at different stages of the processing chain. Based on the observation of the above-mentioned real degradation mechanism, this invention designs a degradation process that combines multiple degradation strategies with a random sequence generation mechanism, aiming to maximize the expansion of the degradation space covered by the training data. Multiple degradation strategies: This invention designs the degradation process as a cascaded structure with variable rounds; when generating training samples each time, the degradation rounds N are first randomly sampled from a preset range. Each degradation round can include operations such as noise injection, JPEG compression, and image resampling; by randomizing the degradation rounds, it is possible to simulate the differentiated cumulative degradation effects experienced by real images under different flow paths: fewer rounds correspond to mild degradation scenarios, and more rounds correspond to severe degradation scenarios where the image undergoes multiple editing, compression, and transmission. Random sequence generation mechanism: Within each round of degradation, this invention randomly rearranges the execution order of each degradation operator; since different degradation operations usually do not satisfy the commutative law, the same set of degradation operators will produce significantly different degradation results under different execution orders; this process can be expressed as: ; in, Indicates the input image. The sequence of degenerate operators is arranged in a random order. By introducing a randomization mechanism, the number of combinations of degenerate samples can be generated on a finite set of degenerate operators, which is far greater than that of a single fixed order, thus significantly improving the diversity of training data.

[0038] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, for system or system embodiments, since they are basically similar to method embodiments, the description is relatively simple, and relevant parts can be referred to the descriptions in the method embodiments. The systems and system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0039] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for synthesizing blind image denoising training data based on multivariate noise mixing and random degradation sequences, characterized in that, include: S1. Construct a multivariate degradation operator library, which includes noise-type degradation operators and non-noise-type degradation operators; S2. Configure a controllable random sampling mechanism to preset and randomly sample the execution probability and parameter range of each operator in the degenerate operator library; S3. Randomly rearrange the execution order of the sampled degradation operators to generate multiple random degradation sequences; S4. Based on multiple random degradation sequences, degradation operations are applied to the input high-quality clean image to generate a noisy degraded image, which, together with the original clean image, forms an image pair for training a blind image denoising model.

2. The method for synthesizing blind image denoising training data based on multivariate noise mixing and random degradation sequences according to claim 1, characterized in that, In S1, noise-type degradation operators include: Gaussian noise, Poisson noise, speckle noise, and camera sensor noise; Non-noise degradation operators include: JPEG compression noise and image resampling.

3. The method for synthesizing blind image denoising training data based on multivariate noise mixing and random degradation sequences according to claim 2, characterized in that, The modeling method for Gaussian noise is as follows: A zero-mean multivariate Gaussian distribution is adopted, and the noise correlation between the three RGB color channels is explicitly modeled through a 3×3 covariance matrix. The covariance matrix is ​​obtained by random sampling and normalization of a symmetric positive semidefinite matrix. It is used to generate colored Gaussian noise with arbitrary cross-channel correlation, or to generate colored Gaussian noise with independent channels, or to generate grayscale Gaussian noise with all channels sharing the same noise component.

4. The method for synthesizing blind image denoising training data based on multivariate noise mixing and random degradation sequences according to claim 2, characterized in that, The modeling method for Poisson noise is as follows: The scale index is uniformly sampled within a preset interval, the scale parameter is calculated, and the clean image is enlarged according to the scale parameter. Apply Poisson random sampling to the magnified image and divide it by the same scaling factor to map the result back to the original dynamic range; In terms of channel processing, it supports independent Poisson sampling of each channel to generate color Poisson noise, or converting the image to grayscale to generate Poisson noise and then copying it to each channel to generate grayscale Poisson noise.

5. The method for synthesizing blind image denoising training data based on multivariate noise mixing and random degradation sequences according to claim 2, characterized in that, The modeling method for speckle noise is as follows: The noise intensity parameters are uniformly sampled from a preset interval to generate a Gaussian random noise field with zero mean and unit variance, and the noise field is applied to the original image in a multiplicative manner. In terms of channel processing, it supports generating Gaussian noise fields independently in each channel to generate colored speckle noise, or generating a single-channel noise field and copying it to each channel to generate grayscale speckle noise.

6. The method for synthesizing blind image denoising training data based on multivariate noise mixing and random degradation sequences according to claim 2, characterized in that, The modeling method for camera sensor noise is as follows: Construct and execute a closed-loop process of "reverse ISP transformation, RAW domain noise injection, and forward ISP reconstruction"; Inverse ISP transform is used to map a clean RGB image to a RAW domain mosaic image; RAW domain noise injection, based on the physical noise model of camera imaging, superimposes signal-independent readout noise and signal-dependent shot noise into the RAW domain mosaic image. Forward ISP reconstruction is used to reconstruct noisy RAW domain mosaic images into noisy RGB images. The same set of randomly generated ISP parameters is reused when generating clean and noisy samples.

7. The method for synthesizing blind image denoising training data based on multivariate noise mixing and random degradation sequences according to claim 2, characterized in that, The modeling method for JPEG compression noise is as follows: The quality factors are uniformly and randomly sampled from a preset range, and the image is then encoded and decoded using JPEG based on the sampled quality factors.

8. The method for synthesizing blind image denoising training data based on multivariate noise mixing and random degradation sequences according to claim 2, characterized in that, The modeling method for image resampling is as follows: The scaling factor is uniformly and randomly sampled within a preset interval, and an interpolation method is randomly selected between bilinear interpolation and bicubic interpolation to downsample the input image and change its resolution. Then, the same interpolation method is used to upsample the image to restore it to the original resolution.

9. The method for synthesizing blind image denoising training data based on multivariate noise mixing and random degradation sequences according to claim 1, characterized in that, In S2, a controllable random sampling mechanism is configured, and the execution probability of Gaussian noise, camera sensor noise, and JPEG compression operators is higher than that of Poisson noise, speckle noise, and image resampling operators.

10. The method for synthesizing blind image denoising training data based on multivariate noise mixing and random degradation sequences according to claim 1, characterized in that, In S4, based on multiple random degradation sequences, a degradation operation is applied to the input high-quality clean image to generate a noisy degraded image. This noisy degraded image is then combined with the original clean image to form an image pair used to train the blind image denoising model. Specifically: Each time training samples are generated, degenerate rounds N are first randomly sampled from a preset range; For each round of degradation, a set of degradation operators is randomly sampled from the degradation operator library, and the execution order of the set of operators is randomly rearranged and executed sequentially. The same or different degradation operators can be repeatedly executed between different degradation rounds.

Images