Rock core image denoising method, device and equipment

Abstract

The embodiment of the invention relates to the technical field of digital cores, in particular to a core image denoising method, device and equipment, and the method comprises the steps: constructing a diffusion denoising model according to sample core image data; the diffusion denoising model comprises a converter and a denoising device; configuring the de-noiser according to the plurality of sampling times to obtain a plurality of corresponding de-noisers; inputting the target core image data into a converter to perform noise conversion on the target core image data; inputting the target rock core image data subjected to noise conversion into various de-noisers to obtain a plurality of candidate de-noised rock core images; according to the observability of the plurality of candidate de-noised core images, fusing the plurality of candidate de-noised core images to obtain a de-noised core image of the target core image data; the observability is used for representing the de-noising deviation degree of the candidate de-noised core image.

Description

Technical Field

[0001] The embodiments of this specification relate to the technical field of digital core, and particularly to a method, device and equipment for denoising core images. Background Art

[0002] Digital core technology is one of the important means for analyzing reservoir rocks. Core images are obtained through high-precision scanning instruments, and digital core models are established using professional software, providing a powerful tool for subsequent analysis of rock seepage parameters. However, during the scanning process, due to the influence of external factors, the obtained core images usually inevitably include various different complex real-world noises, which makes it impossible to more effectively restore the true pore structure of the physical core during the subsequent modeling process. The purpose of the core image denoising algorithm is to remove the image noise while retaining as many real structure details of the core as possible.

[0003] Existing digital core denoising algorithms mainly include linear filtering algorithms, non-linear filtering algorithms, wavelet transform filtering algorithms, and total variation regularization algorithms, etc. Core images have the characteristics of low contrast, non-uniform illumination, and the superposition of various real-world noise types (such as long-tail-like noise, Poisson noise, impulse noise, etc.). However, the above algorithms assume that the noise is uniform or of a specific type, and may not be able to handle this complex real-world noise existing in core images. In addition, due to the structural complexity of core images, such as irregular pore distribution, it is also difficult for the above algorithms to distinguish between noise and the real core structure, resulting in the loss of details after denoising. Specifically, linear filtering algorithms such as mean filtering denoise by local weighted averaging, but will destroy the microscopic pore structure, resulting in blurred textures and affecting subsequent geological analysis (such as porosity calculation). Non-linear filtering algorithms such as median filtering and non-local means are effective for impulse noise, but have poor effects on Gaussian noise and are prone to failure in areas with low core repeatability. Wavelet denoising depends on fixed basis functions, which are difficult to match the irregular textures of core images, and improper threshold selection is likely to lead to artifacts or loss of details. Algorithms such as total variation regularization rely on artificially designed priors, are difficult to model complex core features, and have low robustness to unknown noise types.

[0004] Therefore, how to overcome the problems of poor ability to distinguish noise and the real core structure, poor ability to handle complex real-world noise in core images, and unsatisfactory denoising quality existing in the existing methods, and propose a core image denoising method with strong ability to distinguish noise and the real core structure, strong ability to handle complex real-world noise in core images, and high denoising quality is a key problem to be solved urgently. Summary of the Invention

[0005] The purpose of the embodiments of this specification is to provide a core image denoising method, device, and equipment to overcome the problems in existing core image denoising methods, such as poor ability to distinguish between noise and real core structures, poor ability to handle complex real-world noise, and unsatisfactory denoising quality.

[0006] On the one hand, the embodiments of this specification provide a core image denoising method, including: constructing a diffusion denoising model according to sample core image data; the diffusion denoising model includes a converter and a denoiser; configuring the denoiser according to multiple sampling times to obtain corresponding multiple denoisers; inputting target core image data into the converter to perform noise conversion on the target core image data; inputting the noise-converted target core image data into the multiple denoisers to obtain multiple candidate denoised core images; fusing the multiple candidate denoised core images according to the observability of the multiple candidate denoised core images to obtain a denoised core image of the target core image data; the observability is used to represent the denoising deviation degree of the candidate denoised core image.

[0007] On the other hand, the embodiments of this specification provide a core image denoising device, including: a construction module for constructing a diffusion denoising model according to sample core image data; the diffusion denoising model includes a converter and a denoiser; a configuration module for configuring the denoiser according to multiple sampling times to obtain corresponding multiple denoisers; a conversion module for inputting target core image data into the converter to perform noise conversion on the target core image data; an input module for inputting the noise-converted target core image data into the multiple denoisers to obtain multiple candidate denoised core images; a fusion module for fusing the multiple candidate denoised core images according to the observability of the multiple candidate denoised core images to obtain a denoised core image of the target core image data; the observability is used to represent the denoising deviation degree of the candidate denoised core image.

[0008] On yet another hand, a computer device is provided, including a memory, a processor, and a computer program stored on the memory and executable on the processor, where the processor executes the above core image denoising method.

[0009] As can be seen from the technical solutions provided in the embodiments of this specification above, the embodiments of this specification can construct a diffusion denoising model based on sample core image data; the diffusion denoising model includes a converter and a denoiser; configure the denoiser according to multiple sampling times to obtain corresponding multiple denoisers; input the target core image data into the converter to perform noise conversion on the target core image data; input the target core image data after noise conversion into the multiple denoisers to obtain multiple candidate denoised core images; fuse the multiple candidate denoised core images according to the observability of the multiple candidate denoised core images to obtain a denoised core image of the target core image data; the observability is used to represent the denoising deviation degree of the candidate denoised core image. Compared with the existing methods, the embodiments of this specification can accurately distinguish the noise and core structure in the core image data through the diffusion denoising model, and then perform noise conversion on the composite real-world noise therein. In addition, the embodiments of this specification can also configure different denoisers according to the sampling times, obtain multiple candidate denoised core images, and fuse the multiple candidate denoised core images according to the denoising deviation degree of the multiple candidate denoised core images, thereby reducing the influence of random deviation in the multiple candidate denoised core images on the denoising result and greatly improving the denoising quality of the core image data. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] In order to more clearly illustrate the technical solutions in the embodiments of this specification or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art.

[0011] Figure 1 is a flowchart of a core image denoising method provided by the embodiments of this specification;

[0012] Figure 2 is a schematic diagram of denoising and adding noise of a core image based on a diffusion model provided by the embodiments of this specification;

[0013] Figure 3 is a schematic diagram of a method for fusing multiple candidate denoised core images provided by the embodiments of this specification;

[0014] Figure 4 is a flowchart of a method for constructing a diffusion denoising model provided by the embodiments of this specification;

[0015] Figure 5 is a schematic diagram of a method for constructing a noise conversion model provided by the embodiments of this specification;

[0016] Figure 6 is a schematic diagram of a method for constructing a tiled diffusion model provided by the embodiments of this specification;

[0017] Figure 7It is a schematic diagram of the construction method of the converter in the diffusion denoising model provided by the embodiments of this specification;

[0018] Figure 8 It is a schematic diagram of the diffusion denoising model constructed based on the encoder of the tiled diffusion model and the sampler of the diffusion denoising model provided by the embodiments of this specification;

[0019] Figure 9 It is a flowchart of a Gaussian noise core image generation method provided by the embodiments of this specification;

[0020] Figure 10 It is a schematic diagram of obtaining the target core image data of Core A using a visible light surface scanner provided by the embodiments of this specification;

[0021] Figure 11 It is a schematic diagram of the denoised core image obtained after denoising the target core image data of Core A using the core image denoising method provided by the embodiments of this specification;

[0022] Figure 12 It is a schematic diagram of obtaining the target core image data of Core B using a CT scanner provided by the embodiments of this specification;

[0023] Figure 13 It is a schematic diagram of the denoised core image obtained after denoising the target core image data of Core B using the core image denoising method provided by the embodiments of this specification;

[0024] Figure 14 It is a schematic diagram of the structural composition of a core image denoising device provided by the embodiments of this specification;

[0025] Figure 15 It is a schematic diagram of the structural composition of a computer device provided by the embodiments of this specification. Detailed implementation manners

[0026] Next, the technical solutions in the embodiments of this specification will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of this specification. Obviously, the described embodiments are only a part of the embodiments of this specification, rather than all the embodiments. Based on the embodiments in this specification, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of this specification.

[0027] Digital core images can contain details such as pores, fractures, and mineral grains, with irregular shapes and varying scales. Existing methods are prone to over-smoothing, resulting in the loss of key geological information. The noise in digital core images can come from scanning devices (such as X-ray, CT imaging) or environmental interference, presenting spatial heterogeneity (such as local high-noise regions). Existing methods assume uniform noise distribution (such as global threshold segmentation) and are difficult to handle specifically. The internal structure of the core has low contrast and blurred boundaries. Existing edge detection algorithms are prone to misjudging noise as edges or unable to distinguish the true core structure from noise. In addition, digital core images often contain Gaussian-like noise, Poisson noise (photon counting noise), and impulse noise (sensor anomalies) simultaneously. Existing methods are designed only for a single type of noise and cannot effectively handle mixed noise.

[0028] To address the above problems, the diffusion model can gradually remove noise through multiple steps of iteration, avoiding the rough estimation of one-step denoising, and is particularly suitable for preserving weak edges and tiny pores (such as nanoscale pores). The diffusion model can also implicitly learn geological features such as pore distribution and fracture orientation through training with a large number of core images, and thus has extremely accurate discrimination ability between noise and the true core structure. In addition, based on the generative framework, the diffusion model can synthesize reasonable microstructures (such as repairing pore connections blocked by noise).

[0029] Figure 1 It is a flowchart of a core image denoising method provided by an embodiment of this specification. Specifically, in implementation, it includes the following steps:

[0030] S101: Construct a diffusion denoising model according to the sample core image data; the diffusion denoising model includes a transformer and a denoiser.

[0031] In some embodiments, the sample core image data includes a sample core image to be denoised and a corresponding sample denoised core image; according to the sample core image to be denoised, a noise conversion model can be constructed; the noise conversion model includes an encoder and a decoder, the encoder is used to perform noise conversion on the sample core image to be denoised, and the decoder is used to recover the sample core image to be denoised from the noise conversion result; according to the sample denoised core image, a tiled diffusion model can be constructed; the tiled diffusion model includes a diffuser and a sampler; the diffuser is used to add noise to the sample denoised core image, and the sampler is used to denoise the result of adding noise; based on the encoder of the noise conversion model, the transformer can be constructed; based on the sampler of the tiled diffusion model, the denoiser can be constructed.

[0032] Diffusion denoising models can achieve precise noise removal through staged processing. Similar to the hierarchical processing mechanism of the human visual system, diffusion denoising models can perform feature extraction and representation transformation on noisy core images, and then gradually purify them in the feature space. This two-stage design not only conforms to the theoretical framework of signal processing but also effectively solves the problem that the performance of existing end-to-end denoising models is less than satisfactory in the context of core image denoising.

[0033] Sample core image data can include sample core images to be denoised and corresponding denoised core images. The sample core images to be denoised can be core images to be denoised obtained by scanning the sample core, and the denoised core images can be clean core images after denoising the sample core images to be denoised. Specifically, the sample core images to be denoised can be obtained through the following channels: 1) Raw output of core scanning equipment: including unprocessed data directly generated by devices such as X-ray, CT scanners (resolution usually 0.5 - 5μm / voxel), and visible light surface scanners. Such core images generally contain inherent noise of the equipment, such as quantum noise (Poisson distribution) of X-ray and readout noise (similar to long-tailed distribution) of electronic circuits; 2) Field-collected images: Photos of the core surface taken by digital cameras or mobile phones, whose noise mainly comes from uneven environmental lighting, lens fouling, and JPEG compression artifacts. The characteristic of this type of data is that the spatial distribution of noise is uneven and often accompanied by local occlusion; 3) Degraded images archived in samples: Low-quality core images after multiple replications or compressed transmissions, which may contain mixed noise (impulse noise + Gaussian noise) and blurring degradation. The denoised core images can be obtained specifically through the following methods: 1) Re-inspection with professional equipment: Re-imaging the same core sample using a higher-precision scanning device; 2) Manual repair by experts: Pixel-by-pixel correction by geological engineers using tools such as Photoshop; 3) Re-scanning after physical cleaning: Re-acquiring the core image after surface cleaning treatment of the core sample.

[0034] A noise conversion model based on the Autoencoder structure can be trained with the core image of the sample to be denoised. It consists of two symmetric parts: an encoder and a decoder. During the training phase, the encoder can perform noise conversion on the input core image of the sample to be denoised, converting the complex real-world noise in the core image of the sample to be denoised into Gaussian noise. Correspondingly, the decoder can perform inverse noise conversion on the noise conversion result output by the encoder, that is, recover the core image of the sample to be denoised from the noise conversion result through hierarchical sampling. The noise conversion model can master the ability to convert real-world noise to theoretical Gaussian noise through self-supervised learning. The noise conversion model can specifically adopt network architectures such as ResNet, Transformer, and GNN. The encoder part can include multiple (for example, 5-7) downsampling levels, and the decoder part can adopt a mirror-symmetric upsampling structure to gradually perform inverse noise conversion through transposed convolution or interpolation operations, which will not be elaborated here.

[0035] The tiled diffusion model can be a generative model constructed based on the core image of the sample to be denoised. It can learn the essential distribution of the core image data by defining a forward diffusion process and a reverse denoising process. Through training with a large number of core images of the sample to be denoised, the tiled diffusion model can implicitly learn geological features such as pore distribution and fracture orientation, and then obtain the ability to accurately distinguish noise from the true core structure. The model can include two parts: a diffuser and a sampler, corresponding to two opposite processes of noise addition and removal, which can refer to Figure 2 the denoising and noise addition schematic diagrams of the core image based on the diffusion model shown. The diffuser can be regarded as a Markov chain process, gradually converting a clear image into a Gaussian noise image through T time steps. At each time step t, the system adds noise of a specific intensity according to a predefined noise schedule (such as a cosine schedule). Through this progressive noise addition strategy, the diffuser can establish a complete transition path from a clear image to a pure noise image. The sampler, on the other hand, can implement the reverse denoising process, which is essentially learning a conditional probability distribution. Similarly, through a progressive denoising strategy, the denoiser can establish a complete transition path from a pure noise image to a clear image, that is, gradually remove noise through multiple steps of iteration, avoiding the rough estimation of one-step denoising, which helps to retain weak edges and tiny pores (such as nanoscale pores). The tiled diffusion model can specifically be a diffusion network based on Stable Diffusion. By minimizing the difference between the predicted noise and the actual noise, the network gradually masters the dynamic process of denoising. The tiled diffusion model can gradually remove noise through multiple steps of iteration, avoiding the rough estimation of one-step denoising, and is particularly suitable for retaining weak edges and tiny pores (such as nanoscale pores).

[0036] The diffusion denoising model consists of two major modules: a transformer and a denoiser. Specifically, according to the noise conversion model constructed using the core images to be denoised of the samples, the encoder therein can be used as the transformer. According to the tiled diffusion model constructed using the denoised core images of the samples, the sampler therein can be used as the denoiser. The transformer can map the noisy image in the original pixel space to a Gaussian noise feature space that is more suitable for denoising operations. The denoiser module, based on the principle of the diffusion model, can perform controllable noise removal operations in the feature space. Its advantage lies in being able to finely adjust the denoising intensity through multi-step iterations, avoiding the loss of details caused by single-step denoising. The collaborative work of the two modules forms a complete "encoding - purification - reconstruction" processing chain, achieving the efficient removal of complex real-world noise in the core images to be denoised while maintaining the integrity of the image structure.

[0037] S102: Configure the denoiser according to multiple sampling times to obtain corresponding multiple denoisers.

[0038] In some embodiments, based on the Monte Carlo algorithm, a sampling sequence of a preset length can be generated; the sampling sequence includes multiple sampling times; configuring the denoiser according to the multiple sampling times in the sampling sequence can obtain multiple denoisers with different numbers of sampling blocks.

[0039] The main factor affecting the distortion and perception of core images is the randomness introduced by Gaussian noise in the additional noise term. Excessive randomness may lead to significant distortion in the finally denoised core images. Theoretically, moderate randomness can promote convergence to better results. Therefore, based on the Monte Carlo algorithm, the sampling times in the denoising process can be adjusted to control randomness, and the candidate denoised core images obtained through multiple inferences are averaged to converge to a higher-quality result. A higher sampling number allows for more refinement and introduces more randomness, thus generating more detailed and better-perceived core images.

[0040] The Monte Carlo algorithm can simulate a large number of random events, randomly sample within a given parameter range based on the principle of probability statistics, and thus generate a series of numerical sequences that conform to preset rules. The total number of sampling times (i.e., the total number of sampling blocks it contains, and each sampling block performs one sampling) N of the denoiser of the diffusion denoising model can be obtained. For a given confidence level a (0.7 < a < 1), it can be considered that for each integer sampling number in the interval [aN, N], the corresponding core image can be used as a candidate denoised core image for a target core image data, because after at least A (A can represent the result of rounding down aN) samplings, it can be considered that the noise level in the target core image data has converged.

[0041] Specifically, the Monte Carlo algorithm can be used to generate a sampling sequence of a preset length (e.g., 10,000). The generated sampling sequence includes various numbers of samplings. For each number of samplings, the denoiser is configured. For example, the denoiser contains 1,000 sampling blocks, and the given confidence level is set to 0.8. Then, after at least 1,000 * 0.8 samplings, it can be considered that the noise level in the target core image data has converged. A sampling sequence with a length of 2,000 can be generated using the Monte Carlo algorithm. For a number of samplings with a value of 810 in the sampling sequence, the first 810 sampling blocks in the denoiser can be retained. Based on these first 810 sampling blocks in the denoiser, a sampler under the configuration with 810 samplings is obtained. By configuring the denoiser for various different numbers of samplings in the sampling sequence in this process, multiple denoisers with different numbers of sampling blocks can be obtained.

[0042] S103: Input the target core image data into the converter to perform noise conversion on the target core image data.

[0043] In some embodiments, the target core image data can be input into the converter to perform noise conversion on the target core image data; the target core image data may include a target core image to be denoised.

[0044] The target core image data may include a target core image to be denoised. The target core image to be denoised can be input into the converter, and the converter can convert various complex real-world noises existing in the target core image to be denoised into theoretical Gaussian noises. This conversion process can not only retain key information such as fracture textures and pore structures in the target core image to be denoised but also introduce an expected Gaussian noise pattern, providing a good data basis for subsequent denoising processing.

[0045] S104: Input the noise-converted target core image data into the multiple denoisers to obtain multiple candidate denoised core images.

[0046] In some embodiments, the noise-converted target core image data can be input into the multiple denoisers with different numbers of sampling blocks to obtain multiple candidate denoised core images.

[0047] The target core image data after noise conversion can be input into multiple denoisers with different numbers of sampling blocks. Each denoiser samples the target core image data after noise conversion according to the included number of sampling blocks, generating multiple candidate denoised core images. Specifically, for a denoiser containing 810 sampling blocks, it can sample the target core image data after noise conversion 810 times and obtain 810 candidate denoised core images after sampling. Each candidate denoised core image strictly corresponds to a specific sampling number in the sampling sequence, meaning that they are the image results obtained after being processed by the denoiser at this sampling number. This correspondence ensures the traceability and interpretability of the denoising process, providing rich data dimensions for subsequent analysis and evaluation. In addition, by adjusting the sampling number in the denoising process to control randomness, and obtaining multiple candidate denoised core images through the inference of multiple different denoisers, it lays a data foundation for fusing multiple candidate denoised core images to obtain higher-quality results.

[0048] S105: According to the observability of the multiple candidate denoised core images, fuse the multiple candidate denoised core images to obtain a denoised core image of the target core image data; the observability is used to represent the denoising deviation degree of the candidate denoised core image.

[0049] In some embodiments, according to the observability of each candidate denoised core image, the following formula can be used to fuse the multiple candidate denoised core images to obtain a denoised core image of the target core image data:

[0050]

[0051] In the formula, x represents the target core image data, and y fin represents the denoised core image of the target core image data, represents the i-th candidate denoised core image, represents the observability of the i-th candidate denoised core image, and M represents the total number of candidate denoised core images.

[0052] Each sampling of the denoiser is a random process, so each generated candidate denoised core image inevitably has a certain random deviation. According to the law of large numbers, when the number of inferences is large enough, this random deviation's impact on the quality of the denoised core image can be completely eliminated. Therefore, by using the corresponding observability to weight the multiple candidate denoised core images, the random deviation in the multiple candidate denoised core images can be minimized, improving the quality of the denoised core image.

[0053] The observability of each candidate denoised core image represents the degree of denoising deviation of the candidate denoised core image, that is, the deviation degree of the remaining small part of noise after denoising the target core image data. Specifically, the observability can be represented as the probability of obtaining the candidate denoised core image from the target core image to be denoised. According to the known standard Gaussian noise model Ν(0, I), and the given target core image data x, the corresponding noise level σ can be determined, and a modified Gaussian noise model can be constructed Based on the modified Gaussian noise model the observability of each candidate denoised core image can be directly calculated

[0054] Please refer to Figure 3 and the observability of each candidate denoised core image can be used as the weighting coefficient of each candidate denoised core image. According to the law of large numbers, when the total number M of candidate denoised core images approaches infinity, the weighted average result of all candidate denoised core images and their observability can completely eliminate the random deviation in the candidate denoised core images. M can be set to a sufficiently large integer (for example, 10000). In this case, the weighted average result of all candidate denoised core images and their observability can basically eliminate the random deviation in the candidate denoised core images, thereby greatly improving the quality of the final denoised core image.

[0055] Figure 4 FIG. is a flowchart of a method for constructing a diffusion denoising model provided by an embodiment of this specification. Specifically, when implemented, the following steps are included:

[0056] S401: Construct a noise conversion model according to the sample core image to be denoised; the noise conversion model includes an encoder and a decoder, the encoder is used to perform noise conversion on the sample core image to be denoised, and the decoder is used to recover the sample core image to be denoised from the noise conversion result.

[0057] In some embodiments, the sample denoised core image can be input into a preset Gaussian noise adder to obtain a Gaussian noise core image; according to the sample core image to be denoised and the sample denoised core image, the following formula can be used to construct a noise conversion model:

[0058]

[0059] In the formula, x represents the sample core image to be denoised, y represents the Gaussian noise core image, ε~Ν(0, I), represents the encoder, θ1 represents the parameter of the encoder, Let \(\theta_1\) represent the decoder, \(\theta_2\) represent the parameters of the decoder, \(\sigma\) represent the noise level of the sample denoising core image, \(N(0, I)\) represent the standard Gaussian distribution, \(\lambda\) represent the hyperparameter, and \(R(y)\) represent the regularization function.

[0060] During the scanning process of the core image, real-world noise will be introduced into the core image, and this real-world noise does not follow the Gaussian distribution. Since real-world noise is generated by a mixture of various random sources and various non-random sources in the real world, its distribution mechanism is difficult to learn. Although the distribution mechanism of real-world noise is difficult to learn, various noise types can be converted into Gaussian noise through noise conversion. By setting constraint terms in the noise conversion model constrain the encoder, which can ensure that the encoder can learn the conversion relationship from the complex real-world noise in the sample denoising core image to the theoretical Gaussian noise. Based on this conversion relationship, the complex real-world noise that does not conform to the Gaussian distribution in the denoising core image can be converted into Gaussian noise, laying a good data foundation for subsequent Gaussian noise denoising.

[0061] Please refer to Figure 5 , the sample denoising model can be input into a preset Gaussian noise adder to obtain a Gaussian noise core image. Based on the sample denoising core image and the Gaussian noise core image, a noise conversion model can be constructed. Specifically, the noise conversion model can include two parts: a symmetric encoder and a decoder. The sample denoising core image can be used as the input of the encoder, and the decoder takes the output of the encoder as the input and outputs the prediction result of the sample denoising core image. The encoder and decoder can be constrained by the loss function to ensure that the noise conversion model can learn the latent feature distribution in the sample denoising core image. Further, for the output of the encoder, the encoder can be further constrained by the loss function to ensure that the encoder can learn the conversion relationship from the complex real-world noise in the sample denoising core image to the theoretical Gaussian noise. In addition, the regularization constraint can be used to balance and the loss function to ensure the overall generalization and robustness of the noise conversion model.

[0062] S402: Construct a tiled diffusion model according to the sample denoising core image; the tiled diffusion model includes a diffuser and a sampler; the diffuser is used to add noise to the sample denoising core image, and the sampler is used to denoise the result of adding noise.

[0063] In some embodiments, according to the sample denoising core image, the tiled diffusion model can be constructed using the following formula:

[0064]

[0065] Wherein, y0 represents the initial image patch of the sample-denoised core image, and y t represents the image patch of the sample-denoised core image after noise addition at the t-th step of the diffuser / denoising at the t-th time step in the sampler. ε ∼ Ν(0, I), and Ν(0, I) represents the standard Gaussian distribution. ε θ represents the tiled diffusion model, θ represents the parameters of the tiled diffusion model, and [O] represents the overlapping region pixels of adjacent image patches and of. represents the pixels in the edge region of the image patch, and λ1 and λ2 represent hyperparameters.

[0066] Each image of the diffusion model is divided into multiple patches for separate processing and then integrated through a coordination mechanism. Inevitably, this will cause fragmentation during patch stitching and sudden changes at the patch edges. Through loss constraints it is possible to force the predicted values in the overlapping part of adjacent patches to be similar, and then ensure that the overlapping regions of adjacent patches need to be consistent, improving the quality of core image generation. Through loss constraints it is possible to penalize the sudden changes at the patch edges, making the patch transition more natural and improving the quality of core image generation.

[0067] Please refer to Figure 6 , the tiled diffusion model can be a generative model constructed based on the sample-denoised core image, which can include two modules: a diffuser and a sampler. The diffuser can be regarded as a Markov chain process that gradually converts a clear image into a Gaussian noise image through T time steps. At each time step t, a specific intensity of noise can be added according to a predefined noise schedule. Through this progressive noise addition strategy, the diffuser can establish a complete transition path from the clear image to the pure noise image. The sampler can then implement the reverse denoising process, which essentially learns a conditional probability distribution. Similarly, through the progressive denoising strategy, the denoiser can establish a complete transition path from the pure noise image to the clear image. The tiled diffusion model can specifically be based on diffusion networks such as DDPM and Stable Diffusion. By minimizing the difference between the predicted noise and the actual noise through the loss function , the network gradually masters the dynamic process of denoising. In addition, the tiled diffusion model essentially divides a large image into multiple small regions (tiles) for separate processing and then integrates them through a coordination mechanism. Therefore, the loss function can be used to ensure local consistency between different patches, guarantee the generation quality within each patch, and the loss function can be used to ensure the global coordination of the overall image and guarantee seamless connection between patches.

[0068] S403: Construct the converter based on the encoder of the noise conversion model.

[0069] In some embodiments, based on the encoder of the noise conversion model, the converter can be constructed. Considering that a loss function is used in the noise conversion model to constrain the encoder. Therefore, it can be considered that the encoder can learn the conversion relationship from the complex real-world noise in the sample to-be-denoised core image to the theoretical Gaussian noise. So, the encoder in the noise conversion model can be directly used as the converter in the diffusion denoising model. The converter can be used to perform noise conversion on the input to-be-denoised core image.

[0070] S404: Based on the sampler of the tiled diffusion model, the denoiser is constructed.

[0071] In some embodiments, the following formula can be used to calculate the noise level corresponding to each sampling block in the sampler:

[0072]

[0073] where \(t\in[0,T]\), \(T + 1\) represents the total number of sampling blocks in the sampler, \(d\) t represents the noise level corresponding to the \(t\)-th sampling block, represents the noise attenuation coefficient corresponding to the \((T - t)\)-th time step to which the \(t\)-th sampling block belongs; the matching degree between the noise level corresponding to each sampling block and the noise level of the target core image data can be calculated; the time step of the sampling block corresponding to the maximum matching degree can be used as the time step threshold; in the sampler, multiple sampling blocks with time steps less than or equal to the time step threshold can be determined; based on the multiple sampling blocks with time steps less than or equal to the time step threshold, the denoiser can be constructed.

[0074] Different sampling blocks in the sampler correspond to different noise levels. By calculating the matching degree between the noise level corresponding to each sampling block and the noise level of the target core image data, and using the time step of the sampling block corresponding to the maximum matching degree as the time step threshold, the denoiser can be constructed based on multiple sampling blocks with time steps less than or equal to the time step threshold, achieving the alignment of the denoising ability of the denoiser with the noise level of the target core image data, which helps to improve the quality of the denoised core image generated by the denoiser.

[0075] The sampler of the tiled diffusion model can include multiple reverse sampling blocks. The multiple reverse sampling blocks correspond to multiple time steps in a Markov chain process. Assuming there are \(T + 1\) reverse sampling blocks, for the \(t\)-th reverse sampling block among them, it uniquely corresponds to the \((T - t)\)-th time step in the Markov chain process. The noise attenuation coefficient corresponding to the \((T - t)\)-th time step can be obtained and the noise level of the core image after sampling at the \((T - t)\)-th time step can be expressed as

[0076] The denoising ability of the sampler at time step T-t can be aligned with the noise level of the target core image data to obtain the best denoised core image. Please refer to Figure 7 , the noise level of the target core image data can be obtained, and the formula is used to obtain the noise level corresponding to each sampling block, and the absolute deviation between the noise level of the target core image data and the noise level corresponding to each reverse sampling block is calculated. The reverse sampling block corresponding to the minimum absolute deviation and the corresponding time step can be determined, and this time step is used as the time step threshold. In the sampler, multiple sampling blocks with a time step less than or equal to the time step threshold can be determined. Multiple sampling blocks with a time step less than or equal to the time step threshold can be directly used as the denoiser.

[0077] In some embodiments, according to the sample denoised core image, a tiled diffusion model can be constructed; the tiled diffusion model includes a diffuser and a sampler; the diffuser is used to add noise to the sample denoised core image, and the sampler is used to denoise the result of the added noise; based on the sampler, the following formula is used to construct a Gaussian noise model:

[0078] p(x|y)~Ν(y,σ 2 I);

[0079] In the formula, x represents the target core image data, y represents the denoised core image, σ represents the noise level of the target core image data, and Ν(0,I) represents the standard Gaussian distribution; each candidate denoised core image can be input into the Gaussian noise model to obtain the observability of the candidate denoised core image.

[0080] By constructing a Gaussian noise model based on the pre-trained sampler, the accurate quantification of the difference between the target core image data and each candidate denoised core image is realized, laying a foundation for the weighted average fusion of multiple subsequent candidate denoised core images.

[0081] The tiled diffusion model can include two parts, a diffuser and a sampler, corresponding to two opposite processes of noise addition and removal respectively. Please refer to Figure 2Schematic diagram of denoising and adding noise to core images based on diffusion models. The diffuser can be regarded as a Markov chain process that gradually converts a clear image into a Gaussian noise image through T time steps. At each time step t, the system adds noise of a specific intensity according to a predefined noise schedule (such as a cosine schedule). Through this progressive noise addition strategy, the diffuser can establish a complete transition path from the clear image to the pure noise image. The sampler, on the other hand, can achieve the reverse denoising process, which essentially involves learning a conditional probability distribution. Similarly, through a progressive denoising strategy, the denoiser can establish a complete transition path from the pure noise image to the clear image. The sampler is essentially equivalent to a known standard Gaussian noise model Ν(0, I). Therefore, given the target core image data x, the corresponding noise level σ can be determined, and a modified Gaussian noise model p(x|y) ∼ Ν(y, σ 2 I), where each candidate denoised core image can be input into the modified Gaussian noise model to obtain the observability of the candidate denoised core image

[0082] The encoder in the noise conversion model can be directly used as the converter of the diffusion denoising model, and based on the sampler in the tiled diffusion model, the denoiser for constructing the diffusion denoising model using the above process can be referred to Figure 8 to the schematic diagram of the diffusion denoising model constructed in

[0083] In some embodiments, the sample denoised core images included in the sample core image data can be input into a preset Gaussian noise adder to obtain Gaussian noise core images. The quality of the generated Gaussian noise core images determines whether the converter in the subsequent noise conversion model can learn the accurate conversion method from real-world noise to theoretical Gaussian noise. Figure 9 is a flowchart of a method for generating Gaussian noise core images provided in an embodiment of this specification. Specifically, when implemented, it includes the following steps:

[0084] S901: Construct a Gaussian noise adder.

[0085] In some embodiments, the Gaussian noise adder can be constructed using the following formula:

[0086]

[0087] In the formula, σ(x, y) represents the preset basic noise intensity, α represents the adaptive adjustment coefficient, represents the gradient magnitude of the pixel point (x, y).

[0088] For high-texture regions in core images, such as complex structures like pores and fractures, these regions often contain rich geological information but are also more vulnerable to noise interference, causing the effective signals to be masked by the complex background. Therefore, by calculating the gradient magnitude of pixel points to identify these high-texture regions, the gradient magnitude reflects the intensity of gray-scale changes in the image, and high-gradient magnitude regions usually correspond to details such as edges and textures in the image. Subsequently, an adaptive adjustment coefficient is introduced, which is dynamically adjusted according to the size of the gradient magnitude, so that in high-texture regions, the noise variance is moderately increased. This strategy is not simply adding noise but an intelligent regulation based on the local characteristics of the image, aiming to highlight the features of high-texture regions, prevent the noise from being submerged by the background, thereby enhancing the effect of subsequent data augmentation operations and making geological features such as pores and fractures more clearly distinguishable in the image.

[0089] On the other hand, for low-texture regions in core images, such as smooth regions like homogeneous minerals, excessive noise interference not only does not contribute to feature extraction but may also introduce artifacts, affecting the accuracy of image analysis. Therefore, during the noise regulation process, the gradient magnitude information can also be used to identify these low-texture regions and reduce the noise intensity in these regions through the adaptive adjustment coefficient. By reducing the noise variance, while maintaining the overall structural information of the image, the artifact phenomenon caused by excessive noise interference in smooth regions can be reduced, ensuring that the authenticity and accuracy of the image are not affected.

[0090] S902: Input the sample denoised core image into a Gaussian noise adder to obtain an initial noisy core image.

[0091] In some embodiments, the sample denoised core image can be input into a Gaussian noise adder to obtain an initial noisy core image. The Gaussian noise adder can achieve refined management of the noise distribution of the core image by coupling the gradient magnitude of pixel points with the adaptive adjustment coefficient, which can not only highlight the geological features of high-texture regions but also protect the image quality of low-texture regions, providing a solid and accurate image basis for the subsequent generation of Gaussian noisy core images.

[0092] S903: Smooth the initial noisy core image based on the Markov random field.

[0093] In some embodiments, based on the Markov random field, the following formula can be used to smooth the initial noisy core image:

[0094]

[0095] In the formula, represents the single-point Gaussian constraint of the initial noisy core image based on the Markov random field, represents the neighborhood consistency constraint of the initial noisy core image based on Markov random field, and β represents the smoothness constraint factor.

[0096] By combining the Markov random field theory and imposing single-point Gaussian constraints and neighborhood consistency constraints, the noise variance of the pore structure area in the initial noisy core image can be effectively reduced, and a higher quality Gaussian noise core image can be obtained.

[0097] The core image can be modeled as a two-dimensional Markov random field, in which the gray value of each pixel is regarded as a random variable. These random variables are related to each other through a specific potential energy function to reflect the spatial smoothness and structural consistency in the image. For the initial noisy core image, a single-point Gaussian constraint can be imposed. The core idea of ​​the single-point Gaussian constraint is that the gray value of the noise at each pixel in the image should follow a known statistical distribution (such as Gaussian distribution). Therefore, based on this prior knowledge, a prior probability based on the Gaussian distribution is set for each pixel as a soft constraint in the process of adjusting the noise distribution. Through the single-point Gaussian constraint, the gray value of the pixel can be gradually adjusted so that it can be as close as possible to the ideal state described by the Gaussian noise distribution while satisfying the observed data, thereby initially suppressing the noise.

[0098] For complex areas such as pore structures, the noise variance in these areas is high, and the grayscale values ​​between pixels vary greatly. A neighborhood consistency constraint can be further introduced. This constraint is based on the assumption that there is a strong spatial correlation between adjacent pixels in the image, that is, the grayscale values ​​of adjacent pixels should have a certain similarity. By constructing a potential energy function containing neighborhood information, the denoised image is encouraged to maintain the consistency of the grayscale values ​​of pixels in the neighborhood as much as possible while maintaining the rationality of the single-point grayscale value. Through the neighborhood consistency constraint, the spatial structural information of the image is effectively utilized, which promotes the smoothing and restoration of complex areas such as pore structures, and further reduces the noise variance of these areas.

[0099] S904: Based on DCT transformation, a preset frequency threshold is used to perform frequency suppression on the initial noisy core image.

[0100] In some embodiments, the initial noise core image may be subjected to DCT transformation to obtain spectrum data corresponding to the initial noise core image; according to a preset frequency threshold, the frequency spectrum data corresponding to the initial noise core image may be subjected to frequency suppression using the following formula:

[0101]

[0102] Wherein, N(u, v) represents the transformation coefficient at the horizontal frequency u and the vertical frequency v in the spectrum data corresponding to the initial noisy core image, γ represents a preset attenuation factor, and T represents a preset frequency threshold.

[0103] The initial noisy core image can be regarded as a two-dimensional digital signal matrix, where each element represents the pixel gray value at the corresponding position in the image. Subsequently, by applying the DCT algorithm, the image data in this spatial domain can be transformed into the frequency domain, thereby obtaining the spectrum data corresponding to the initial noisy core image. The DCT transformation can concentrate most of the energy in the image on a few low-frequency components, while concentrating the detailed information and noise components of the image in most high-frequency components.

[0104] After obtaining the spectrum data, the spectrum data can be refined according to the preset frequency threshold. Specifically, according to the horizontal frequency amplitude and the vertical frequency amplitude of each frequency component in the spectrum data, it can be determined whether to suppress the frequency component based on the comparison result with the preset frequency threshold. The preset frequency threshold can be the maximum frequency component of the theoretical Gaussian distribution noise obtained based on the prior knowledge of Gaussian noise. If the sum of the horizontal frequency amplitude and the vertical frequency amplitude of a certain frequency component in the spectrum data is greater than the preset frequency threshold, the corresponding transformation coefficient can be reduced to attenuate the frequency component, ensuring that the noise in the core image is more in line with the Gaussian distribution.

[0105] S905: Fuse the smoothing result of the initial noisy core image and the frequency suppression result of the initial noisy core image to obtain the Gaussian noise core image.

[0106] In some embodiments, the following formula can be used to fuse the smoothing result of the initial noisy core image and the frequency suppression result of the initial noisy core image to obtain the Gaussian noise core image:

[0107] I fin =(1 - τ edge )·I MRF +τ edge ·I DCT ;

[0108] Wherein, represents the binary edge mask result of the initial noisy core image calculated using the Canny edge detection operator, represents the smoothing result of the initial noisy core image, and represents the frequency suppression result of the initial noisy core image.

[0109] For the initial noisy core image, the Canny edge detection operator can be used for edge feature extraction. The Canny edge detection operator systematically analyzes the image through multi-level filtering (including Gaussian smoothing for noise reduction, gradient magnitude calculation, non-maximum suppression, and double-threshold edge connection), and finally generates a binary edge mask image. This mask image accurately calibrates the significant edge regions in the original image with 0-1 pixel values, where pixels with a value of 1 correspond to edge positions, and pixels with a value of 0 represent non-edge regions.

[0110] The generated binary edge mask can be used as a dynamic weighting coefficient to construct a composite image fusion model. The smoothing result can effectively eliminate high-frequency noise but may cause edge blurring, and the frequency suppression result can retain edge details but may leave low-frequency noise residues. On this basis, through a per-pixel weighted fusion strategy, the binary edge mask is used as a weight control parameter: in the edge region (mask value is 1), it focuses on retaining the edge details of the frequency suppression result, and in the non-edge region (mask value is 0), it focuses on adopting the overall structure of the smoothing result. This fusion mechanism achieves the dual effects of edge protection and Gaussian noise addition.

[0111] In some embodiments, the denoised core image of the target core image data can be input into a preset core image super-resolution model to obtain an enhanced denoised core image of the target core image data.

[0112] Based on the preset core image super-resolution model, the local core texture and global pore structure in the denoised core image can be further optimized, making the core edges and textures clearer and more natural, and realizing the generation of a high-quality, detail-rich, and natural-transition ultra-high-resolution enhanced denoised core image.

[0113] The preset core image super-resolution model can be based on the Swin-Conv-Unet (SCUNet) network architecture. This architecture cleverly combines the advantages of the Swin Transformer and the convolutional neural network (CNN). Through the structural design of the encoder-decoder, it realizes the layer-by-layer extraction and reconstruction of image features. In particular, each sub-block of the encoder-decoder uses the Swin-Conv module, which not only inherits the powerful local feature modeling ability of the residual convolutional layer but also introduces the non-local feature capture mechanism of the Swin Transformer module. This enables the model to capture the complex structures and texture changes in the core image globally, thus achieving more accurate image reconstruction. The denoised core image of the target core image data can be input into the preset core image super-resolution model to obtain the enhanced denoised core image of the target core image data. In addition, the preset core image super-resolution model can perform multi-scale feature fusion based on skip connections, further enhancing the detail performance and sharpness of the denoised core image. Specifically, by transmitting and fusing feature information between different scales, the model can better capture the subtle structures and texture changes in the core image, making the output enhanced denoised core image not only have fine texture features at ultra-high resolution but also maintain the consistency and coherence of the overall structure.

[0114] The preset core image super-resolution model can accurately identify and enhance the local core texture features in the core image. These textures are important manifestations of the internal structure of the core and are of great significance for geological analysis and oil and gas exploration. At the same time, the preset core image super-resolution model can also effectively restore the global pore structure of the core image, making the edge contour of the core clearer and more natural, avoiding the edge blurring or detail loss problems that may be caused by traditional denoising methods. Through this series of optimizations, the finally generated high-quality, detail-rich, and naturally transitioning ultra-high-resolution enhanced denoised core image not only improves the visual effect of the image but also provides a more accurate and reliable data basis for subsequent geological analysis.

[0115] The following provides two specific embodiments of denoising a core image using the core image denoising method provided in this specification:

[0116] Embodiment 1: Use a visible light surface scanner to obtain the target core image data of Core A. Please refer to Figure 10 . Based on the core image denoising method provided in this specification, denoise the target core image data of Core A to obtain a denoised core image. Please refer to Figure 11 .

[0117] Embodiment 2: Use a CT scanner to obtain the target core image data of Core B. Please refer to Figure 12。Based on the core image denoising method provided in this specification, denoise the target core image data of Core B to obtain a denoised core image. Please refer to Figure 13 。

[0118] The core image denoising method provided in the embodiments of this specification can construct a diffusion denoising model according to the sample core image data; the diffusion denoising model includes a converter and a denoiser; configure the denoiser according to multiple sampling times to obtain corresponding multiple denoisers; input the target core image data into the converter to perform noise conversion on the target core image data; input the noise-converted target core image data into the multiple denoisers to obtain multiple candidate denoised core images; fuse the multiple candidate denoised core images according to the observability of the multiple candidate denoised core images to obtain a denoised core image of the target core image data; the observability is used to represent the denoising deviation degree of the candidate denoised core image. Compared with the existing methods, the embodiments of this specification can accurately distinguish the noise and core structure in the core image data through the diffusion denoising model, and then perform noise conversion on the composite real-world noise therein. In addition, the embodiments of this specification can also configure different denoisers according to the sampling times, obtain multiple candidate denoised core images, and fuse the multiple candidate denoised core images according to the denoising deviation degree of the multiple candidate denoised core images, thereby reducing the influence of random deviation in the multiple candidate denoised core images on the denoising result and greatly improving the denoising quality of the core image data.

[0119] Based on the above core image denoising method, the embodiments of a core image denoising device are also proposed in this specification. As Figure 14 shown, the core image denoising device 1400 may specifically include the following modules:

[0120] A construction module 1401, configured to construct a diffusion denoising model according to the sample core image data; the diffusion denoising model includes a converter and a denoiser.

[0121] A configuration module 1402, configured to configure the denoiser according to multiple sampling times to obtain corresponding multiple denoisers;

[0122] A conversion module 1403, configured to input the target core image data into the converter to perform noise conversion on the target core image data.

[0123] An input module 1404, configured to input the noise-converted target core image data into the multiple denoisers to obtain multiple candidate denoised core images.

[0124] A fusion module 1405, configured to fuse a plurality of candidate denoised core images according to the observability of the plurality of candidate denoised core images to obtain a denoised core image of target core image data; the observability is used to represent the denoising deviation degree of the candidate denoised core image.

[0125] In some embodiments, the above-mentioned construction module 1401 may specifically be configured to construct a noise conversion model for a sample core image to be denoised; the noise conversion model includes an encoder and a decoder, the encoder is configured to perform noise conversion on the sample core image to be denoised, and the decoder is configured to recover the sample core image to be denoised from the noise conversion result; construct a tiled diffusion model according to the sample denoised core image; the tiled diffusion model includes a diffuser and a sampler; the diffuser is configured to add noise to the sample denoised core image, and the sampler is configured to denoise the result of adding noise; construct the converter based on the encoder of the noise conversion model; construct the denoiser based on the sampler of the tiled diffusion model.

[0126] In some embodiments, the above-mentioned construction module 1401 may specifically further be configured to input the sample denoised core image into a preset Gaussian noise adder to obtain a Gaussian noise core image; construct a noise conversion model according to the sample core image to be denoised and the Gaussian noise core image using the following formula:

[0127]

[0128] wherein, x represents the sample core image to be denoised, y represents the Gaussian noise core image, ε ∼ Ν(0, I), represents the encoder, θ1 represents the parameter of the encoder, represents the decoder, θ2 represents the parameter of the decoder, σ represents the noise level of the sample core image to be denoised, Ν(0, I) represents the standard Gaussian distribution, λ represents a hyperparameter, and R(y) represents a regularization function.

[0129] In some embodiments, the above-mentioned construction module 1401 may specifically further be configured to construct a tiled diffusion model according to the sample denoised core image using the following formula:

[0130]

[0131] wherein, y0 represents the initial image block of the sample denoised core image, y t represents the image block of the sample denoised core image after denoising at the t-th time step in the t-th step of adding noise by the diffuser / sampler, ε ∼ Ν(0, I), Ν(0, I) represents the standard Gaussian distribution, ε θ represents the tiled diffusion model, θ represents the parameter of the tiled diffusion model, and [O] represents adjacent image blocks and The pixels in the overlapping region represent the pixels in the edge region of the image block, and λ1 and λ2 represent hyperparameters.

[0132] In some embodiments, the above-mentioned building block 1401 may specifically be further configured to calculate the noise level corresponding to each sampling block in the sampler using the following formula:

[0133]

[0134] where t ∈ [0, T], T + 1 represents the total number of sampling blocks in the sampler, and d t represents the noise level corresponding to the t-th sampling block, represents the noise attenuation coefficient corresponding to the (T - t)-th time step to which the t-th sampling block belongs; calculate the matching degree between the noise level corresponding to each sampling block and the noise level of the target core image data; use the time step of the sampling block with the maximum matching degree as the time step threshold; in the sampler, determine multiple sampling blocks whose time steps are less than or equal to the time step threshold; and construct the denoiser according to the multiple sampling blocks whose time steps are less than or equal to the time step threshold.

[0135] In some embodiments, the above-mentioned configuration module 1402 may specifically be configured to generate a sampling sequence with a preset length based on the Monte Carlo algorithm; the sampling sequence includes multiple sampling times; and configure the denoiser according to the multiple sampling times in the sampling sequence to obtain multiple denoisers with different numbers of sampling blocks.

[0136] In some embodiments, the above-mentioned conversion module 1403 may specifically be configured to input the target core image data after noise conversion into the multiple denoisers with different numbers of sampling blocks to obtain multiple candidate denoised core images.

[0137] In some embodiments, the above-mentioned fusion module 1405 may specifically be configured to fuse multiple candidate denoised core images according to the observability of each candidate denoised core image using the following formula to obtain a denoised core image of the target core image data:

[0138]

[0139] where x represents the target core image data, and y fin represents the denoised core image of the target core image data, represents the i-th candidate denoised core image, represents the observability of the i-th candidate denoised core image, and M represents the total number of candidate denoised core images.

[0140] In some embodiments, the above-mentioned fusion module 1405 may specifically be further configured to denoise a sample core image to construct a tiling diffusion model; the tiling diffusion model includes a diffuser and a sampler; the diffuser is used to add noise to the sample denoised core image, and the sampler is used to denoise the result of adding noise; based on the sampler, the following formula is used to construct a Gaussian noise model:

[0141] p(x|y)~Ν(y,σ 2 I);

[0142] In the formula, x represents the target core image data, y represents the denoised core image, σ represents the noise level of the target core image data, and Ν(0,I) represents the standard Gaussian distribution; each candidate denoised core image is input into the Gaussian noise model to obtain the observability of the candidate denoised core image.

[0143] As can be seen from the above, based on the core image denoising device provided in the embodiments of this specification, a diffusion denoising model can be constructed according to the sample core image data; the diffusion denoising model includes a converter and a denoiser; the denoiser is configured according to multiple sampling times to obtain corresponding multiple denoisers; the target core image data is input into the converter to perform noise conversion on the target core image data; the noise-converted target core image data is input into the multiple denoisers to obtain multiple candidate denoised core images; according to the observability of the multiple candidate denoised core images, the multiple candidate denoised core images are fused to obtain the denoised core image of the target core image data; the observability is used to represent the denoising deviation degree of the candidate denoised core image. Compared with the existing methods, the embodiments of this specification can accurately distinguish the noise and the core structure in the core image data through the diffusion denoising model, and then perform noise conversion on the composite real-world noise therein. In addition, the embodiments of this specification can also configure different denoisers according to the sampling times, obtain multiple candidate denoised core images, and fuse the multiple candidate denoised core images according to the denoising deviation degrees of the multiple candidate denoised core images, thereby reducing the influence of random deviations in the multiple candidate denoised core images on the denoising result and greatly improving the denoising quality of the core image data.

[0144] It should be noted that the units, devices, modules, etc. illustrated in the above embodiments can be specifically implemented by computer chips or entities, or by products with certain functions. For the convenience of description, when describing the above devices, they are divided into various modules according to functions and described separately. Of course, when implementing this specification, the functions of each module can be implemented in the same or multiple software and / or hardware, or the modules implementing the same function can be realized by the combination of multiple sub-modules or sub-units, etc. The device embodiments described above are merely illustrative. For example, the division of the units is only a logical function division, and there may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the displayed or discussed couplings or direct couplings or communication connections to each other can be through some interfaces. The indirect couplings or communication connections of the devices or units can be in electrical, mechanical or other forms.

[0145] The embodiments of this specification also provide a computer device for a core image denoising method, including a processor and a memory for storing processor-executable instructions. When specifically implemented, the processor can execute the following steps according to the instructions: constructing a diffusion denoising model according to the sample core image data; the diffusion denoising model includes a converter and a denoiser; configuring the denoiser according to multiple sampling times to obtain corresponding multiple denoisers; inputting the target core image data into the converter to perform noise conversion on the target core image data; inputting the noise-converted target core image data into the multiple denoisers to obtain multiple candidate denoised core images; fusing the multiple candidate denoised core images according to the observability of the multiple candidate denoised core images to obtain a denoised core image of the target core image data; the observability is used to represent the denoising deviation degree of the candidate denoised core image.

[0146] In order to be able to complete the above instructions more accurately, refer to Figure 15 As shown, the embodiments of this specification also provide another specific computer device 1500. Among them, the computer device 1500 includes a network communication port 1501, a processor 1502, and a memory 1503. The above structures are connected by internal cables so that each structure can perform specific data interactions.

[0147] The processor 1502 can be specifically used to construct a diffusion denoising model according to the sample core image data; the diffusion denoising model includes a converter and a denoiser; configure the denoiser according to multiple sampling times to obtain corresponding multiple denoisers; input the target core image data into the converter to perform noise conversion on the target core image data; input the noise-converted target core image data into the multiple denoisers to obtain multiple candidate denoised core images; fuse the multiple candidate denoised core images according to the observability of the multiple candidate denoised core images to obtain a denoised core image of the target core image data; the observability is used to represent the denoising deviation degree of the candidate denoised core image.

[0148] The memory 1503 can be specifically used to store corresponding instruction programs.

[0149] In this embodiment, the network communication port 1501 can be bound to different communication protocols, so as to send or receive different data virtual ports. For example, the network communication port can be a port responsible for web data communication, can also be a port responsible for FTP data communication, and can also be a port responsible for mail data communication. In addition, the network communication port can also be a physical communication interface or communication chip. For example, it can be a wireless mobile network communication chip, such as GSM, CDMA, etc.; it can also be a Wifi chip; it can also be a Bluetooth chip.

[0150] In this embodiment, the processor 1502 can be implemented in any suitable manner. For example, the processor can take the form of, for example, a microprocessor or a processor and a computer-readable medium storing computer-readable program code (such as software or firmware) executable by the (micro)processor, logic gates, switches, application specific integrated circuits (ASICs), programmable logic controllers, and embedded microcontrollers, etc. This specification does not make any limitations.

[0151] In this embodiment, the memory 1503 includes volatile memory and non-volatile memory. The memory 1503 can include multiple levels. In a digital system, as long as it can store binary data, it can be a memory; in an integrated circuit, a circuit without a physical form but with a storage function is also called a memory, such as RAM, FIFO, etc.; in a system, a storage device with a physical form is also called a memory, such as a memory stick, TF card, etc.

[0152] Those skilled in the art should understand that the embodiments of the present invention can be provided as a method, a system, or a computer program product. Therefore, the present invention can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk memory, CD-ROM, optical memory, etc.) that contain computer-usable program code.

[0153] The present invention is described with reference to the flowcharts and / or block diagrams of methods, apparatuses (systems), and computer program products according to embodiments of the present invention. It should be understood that each flow and / or block in the flowchart and / or block diagram, as well as the combination of flows and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, such that the instructions executed by the processor of the computer or other programmable data processing devices produce means for implementing the functions specified in Figure 1 one flow or multiple flows and / or blocks Figure 1 one block or multiple blocks.

[0154] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing devices to work in a specific manner, such that the instructions stored in the computer-readable memory produce a manufactured article including instruction means that implement the functions specified in Figure 1 one flow or multiple flows and / or blocks Figure 1 one block or multiple blocks.

[0155] These computer program instructions can also be loaded onto a computer or other programmable data processing devices, such that a series of operation steps are executed on the computer or other programmable devices to generate a computer-implemented process, and thus the instructions executed on the computer or other programmable devices provide steps for implementing the functions specified in Figure 1 one flow or multiple flows and / or blocks Figure 1 one block or multiple blocks.

[0156] The specific embodiments described above further elaborate on the objectives, technical solutions, and beneficial effects of the present invention. It should be understood that the above are only specific embodiments of the present invention and are not used to limit the protection scope of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included in the protection scope of the present invention.

Claims

1. A core image denoising method, characterized in that: The method comprises: Constructing a diffusion denoising model according to the sample core image data; the diffusion denoising model includes a converter and a denoiser; The denoiser is configured according to a plurality of sampling times to obtain a plurality of corresponding denoisers; Inputting the target core image data into the converter to perform noise conversion on the target core image data; Inputting the noise-converted target core image data into the multiple denoisers to obtain multiple candidate denoised core images; According to the observability of the plurality of candidate denoised core images, the plurality of candidate denoised core images are fused to obtain a denoised core image of the target core image data; the observability is used to indicate the denoising deviation degree of the candidate denoised core images.

2. The method according to claim 1, characterized in that: The sample core image data includes a sample core image to be denoised and a corresponding sample denoised core image; The construction of the diffusion denoising model includes: According to the sample core image to be denoised, a noise conversion model is constructed; the noise conversion model includes an encoder and a decoder, the encoder is used to perform noise conversion on the sample core image to be denoised, and the decoder is used to restore the sample core image to be denoised from the noise conversion result; According to the sample denoised core image, a tiled diffusion model is constructed; the tiled diffusion model includes a diffuser and a sampler; the diffuser is used to add noise to the sample denoised core image, and the sampler is used to denoise the denoised result; Based on the encoder of the noise conversion model, construct the converter; The denoiser is constructed based on a sampler of the tiled diffusion model.

3. The method according to claim 2, characterized in that: The step of constructing a noise conversion model based on the sample core image to be denoised includes: The sample denoised core image is input into a preset Gaussian noise adder to obtain a Gaussian noise core image; According to the sample core image to be denoised and the Gaussian noise core image, the noise conversion model is constructed using the following formula: In the formula, x represents the sample core image to be denoised, y represents the Gaussian noise core image, ε~N(0,I), represents the encoder, θ1 represents the parameters of the encoder, represents the decoder, θ2 represents the parameters of the decoder, σ represents the noise level of the sample core image to be denoised, Ν(0,I) represents the standard Gaussian distribution, λ represents the hyperparameter, and R(y) represents the regularization function.

4. The method according to claim 2, characterized in that: The step of constructing a tiled diffusion model based on the sample denoised core image includes: According to the sample denoised core image, the tiled diffusion model is constructed using the following formula: Where y0 represents the initial image block of the sample denoised core image, y t represents the image block of the sample denoised core image after the t-th time step of the diffuser denoising / the t-th time step of the sampler, ε~Ν(0,I), Ν(0,I) represents the standard Gaussian distribution, ε θ represents the tile diffusion model, θ represents the parameters of the tile diffusion model, and [O] represents the adjacent image blocks and The overlapping area pixels, represents the edge area pixels of the image block, and λ1 and λ2 represent hyper parameters.

5. The method according to claim 2, characterized in that: The sampler based on the tiled diffusion model constructs the denoiser, comprising: The noise level corresponding to each sampling block in the sampler is calculated using the following formula: Where t∈[0,T], T+1 represents the total number of sampling blocks in the sampler, and d t represents the noise level corresponding to the t-th sampling block, Represents the noise attenuation coefficient corresponding to the Tt-th time step to which the t-th sampling block belongs; Calculate the matching degree between the noise level corresponding to each sampling block and the noise level of the target core image data; The time step of the sampling block corresponding to the maximum matching degree is used as the time step threshold; In the sampler, determining a plurality of sampling blocks having a time step less than or equal to the time step threshold; The denoiser is constructed according to a plurality of sampling blocks whose time steps are less than or equal to the time step threshold.

6. The method according to claim 1, characterized in that: The denoiser is configured according to a plurality of sampling times to obtain a plurality of corresponding denoisers, including: Based on the Monte Carlo algorithm, a sampling sequence of a preset length is generated; the sampling sequence includes a plurality of sampling times; The denoiser is configured according to a plurality of sampling times in the sampling sequence to obtain a plurality of denoisers with different numbers of sampling blocks; The target core image data after noise conversion is input into the multiple denoisers to obtain multiple candidate denoised core images, including: The target core image data after noise conversion is input into the plurality of denoisers with different numbers of sampling blocks to obtain a plurality of candidate denoised core images.

7. The method according to claim 1, characterized in that: The step of fusing the plurality of candidate denoised core images according to the observability of the plurality of candidate denoised core images to obtain a denoised core image of target core image data comprises: According to the observability of each candidate denoised core image, multiple candidate denoised core images are fused using the following formula to obtain the denoised core image of the target core image data: Where x represents the target core image data, y fin The denoised core image representing the target core image data, represents the i-th candidate denoised core image, represents the observability of the i-th candidate denoised core image, and M represents the total number of candidate denoised core images.

8. The method according to claim 1, characterized in that: The method further comprises: According to the sample denoised core image, a tiled diffusion model is constructed; the tiled diffusion model includes a diffuser and a sampler; the diffuser is used to add noise to the sample denoised core image, and the sampler is used to denoise the denoised result; Based on the sampler, a Gaussian noise model is constructed using the following formula: p(x|y)~Ν(y,σ 2 I); Where x represents the target core image data, y represents the denoised core image, σ represents the noise level of the target core image data, and Ν(0,I) represents the standard Gaussian distribution; Each candidate denoised core image is input into the Gaussian noise model to obtain the observability of the candidate denoised core image.

9. A core image denoising device, characterized in that: The device comprises: A construction module is used to construct a diffusion denoising model according to sample core image data; the diffusion denoising model includes a converter and a denoiser; A configuration module, used for configuring the denoiser according to a plurality of sampling times to obtain a plurality of corresponding denoisers; A conversion module, used for inputting the target core image data into the converter to perform noise conversion on the target core image data; An input module, used for inputting the noise-converted target core image data into the multiple denoisers to obtain multiple candidate denoised core images; A fusion module is used to fuse multiple candidate denoised core images according to the observability of the multiple candidate denoised core images to obtain a denoised core image of the target core image data; the observability is used to indicate the denoising deviation degree of the candidate denoised core images.

10. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the method according to any one of claims 1 to 8 is implemented.

Images