Image denoising method, system, device and medium

Abstract

The invention provides an image denoising method, system and device and a medium, and belongs to the field of computer graphics, and the method comprises the following steps: obtaining a to-be-processed image; constructing a Monte Carlo de-noising deep learning network, wherein the Monte Carlo de-noising deep learning network comprises a fast Fourier convolution extractor FFCE, a residual multi-path Mangban encoder RMME and a decoder; the FFCE comprises a local convolution path and a global transformation path, and the RMME comprises a plurality of bidirectional Mangbar blocks (BMB) which are connected in series; inputting a to-be-processed image into the Monte Carlo denoising deep learning network, respectively extracting auxiliary feature information and noise feature information of the to-be-processed image by using a local convolution path and a global transformation path, splicing the auxiliary feature information and the noise feature information, inputting the spliced auxiliary feature information and noise feature information into a bidirectional Munban block for sequential fusion processing, and outputting the output feature of the last BMB through a decoder. And a denoised image is obtained. According to the invention, calculation and memory cost can be reduced.

Description

Technical Field

[0001] The present invention belongs to the field of computer graphics, and in particular relates to an image denoising method, system, device and medium. Background Art

[0002] In computer graphics, Monte Carlo rendering is widely used to generate realistic scenes by tracing light paths (samples) and integrating scene functions to obtain the final pixel color of the image. Although this sample-based method can quickly evaluate the integration results with limited samples, it is affected by noise due to the inaccuracy between the estimated value and the true value. Sampling more rays can improve the resolution of the image, but the increased computational cost can significantly affect the feasibility of Monte Carlo rendering, especially for large-scale images.

[0003] In recent years, with the significant development of deep learning technology, the most common way to remove image noise is to design feature-based deep neural networks to denoise images rendered with a limited number of samples. These post-processing networks aim to utilize additional scene details or auxiliary information to effectively generate noise-free images.

[0004] Bako et al. and Gharbi et al. use convolutional neural networks (CNNs) to learn local and auxiliary feature information of pixels and samples, which can effectively model long-range spatial correlations through attention-based methods, rather than focusing on local spatial correlations through CNN kernels, thereby achieving stronger performance. However, due to the attention mechanism, this method has quadratic complexity, resulting in high computational and memory costs. Summary of the invention

[0005] In order to overcome the above-mentioned shortcomings of high computational and memory costs when denoising an image, the present invention provides an image denoising method, comprising the following steps:

[0006] Get the image to be processed;

[0007] Constructing a Monte Carlo denoising deep learning network and training the Monte Carlo denoising deep learning network; the Monte Carlo denoising deep learning network includes a fast Fourier convolution extractor FFCE, a residual multi-path Mamba encoder RMME and a decoder; the FFCE includes a local convolution path and a global transformation path, and the RMME includes a plurality of bidirectional Mamba blocks BMB connected in series;

[0008] The image to be processed is input into the trained Monte Carlo denoising deep learning network, and the auxiliary feature information and noise feature information of the image to be processed are extracted using the local convolution path and the global transformation path respectively. The auxiliary feature information and the noise feature information are spliced ​​and input into the bidirectional Mamba block. The bidirectional Mamba block is used to sequentially fuse the spliced ​​features, and the output features of the last BMB are output through the decoder to obtain the denoised image.

[0009] Preferably, the training of the Monte Carlo denoising deep learning network comprises the following steps:

[0010] Get image dataset;

[0011] Dividing the multiple image data sets into a training data set and a verification data set, wherein the ratio of the training data set to the verification data set is 95:5;

[0012] The images in the training data set are input into the Monte Carlo denoising deep learning network, and the Monte Carlo denoising deep learning network is trained using a linear combination of a pixel reconstruction loss function and a gradient domain loss function as a total loss function to obtain a trained Monte Carlo denoising deep learning network.

[0013] Preferably, the auxiliary feature information includes depth, normal and albedo.

[0014] Preferably, after the output feature of the last BMB is output by the decoder, the method further includes adding the denoised image to the image to be processed to obtain the denoised image.

[0015] Preferably, the pixel reconstruction loss function is as follows:

[0016]

[0017] Where P i and T i are the i-th pixel of the denoised image and the target image respectively, and N is the total number of pixels.

[0018] Preferably, the gradient domain loss function is as follows:

[0019]

[0020] Where N is the total number of pixels, is the denoised image gradient obtained using the LoG operator, is the target image gradient obtained using the LoG operator.

[0021] The present invention also provides an image denoising system, comprising:

[0022] An image acquisition module, used for acquiring an image to be processed;

[0023] A model building module, used for building a Monte Carlo denoising deep learning network and training the Monte Carlo denoising deep learning network; the Monte Carlo denoising deep learning network includes a fast Fourier convolution extractor FFCE, a residual multi-path Mamba encoder RMME and a decoder; the FFCE includes a local convolution path and a global transformation path, and the RMME includes a plurality of bidirectional Mamba blocks BMB connected in series;

[0024] The image processing module is used to input the image to be processed into the trained Monte Carlo denoising deep learning network, use the local convolution path and the global transformation path to respectively extract the auxiliary feature information and noise feature information of the image to be processed, splice the auxiliary feature information and the noise feature information and input them into the bidirectional Mamba block, use the bidirectional Mamba block to perform sequential fusion processing on the spliced ​​features, output the output features of the last BMB through the decoder, and obtain the denoised image.

[0025] The present invention also provides a computer device, comprising a memory and a processor; the memory stores a computer program, and the processor is used to run the computer program in the memory to execute the image denoising method.

[0026] The present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and the computer program is suitable for being loaded by a processor to execute the image denoising method.

[0027] The image denoising method provided by the present invention has the following beneficial effects:

[0028] The present invention can extract auxiliary feature information and noise feature information of the image to be processed through the fast Fourier convolution extractor FFCE, complete the interactivity between the noise feature information and the auxiliary information, achieve better feature extraction results with fewer parameters, and reduce the calculation cost; the bidirectional Mamba block introduced through RMME can adopt a bidirectional continuous scanning scheme, utilize the long-range spatial dependency in the horizontal and vertical directions, perform serial feature encoding on the extracted vital signs, and sequentially process the fused features as a whole, thereby reducing the calculation and memory costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] In order to more clearly illustrate the embodiment of the present invention and its design scheme, the following briefly introduces the drawings required for this embodiment. The drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0030] Figure 1 is a flow chart of an image denoising method according to an embodiment of the present invention;

[0031] Figure 2 is the RMME module diagram; among them, Figure 2 (a) is the architecture of the residual multi-channel Mamba encoder RMME. Figure 2 (b) is a plurality of bidirectional Mamba blocks BMB connected in series; Figure 2 (c) is a flowchart of BiSSM passing input and residual into the Mamba block BMB;

[0032] Figure 3 The comparison test results of different models for denoising are shown in Figure 2. Figure 3 (a) Figure 3 (b) Figure 3 (c) and Figure 3 (d) shows the performance of denoising using different models in scenes of specular reflection, highlight, refraction, tiny structure and dark light respectively;

[0033] Figure 4 is a comparison chart of model costs; among them, Figure 4 (a) and Figure 4 (b) shows the resource overhead of the four models under different image sizes. DETAILED DESCRIPTION

[0034] In order to enable those skilled in the art to better understand the technical solution of the present invention and implement it, the present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and cannot be used to limit the scope of protection of the present invention.

[0035] In the description of the present invention, it is to be understood that the terms “center”, “longitudinal”, “lateral”, “length”, “width”, “thickness”, “up”, “down”, “front”, “back”, “left”, “right”, “vertical”, “horizontal”, “top”, “bottom”, “inside”, “outside”, “axial”, “radial”, “circumferential”, etc., indicating orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, are only for the convenience of describing the technical solutions of the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present invention.

[0036] In addition, the terms "first", "second", etc. are used for descriptive purposes only and are not to be understood as indicating or implying relative importance. In the description of the present invention, it should be noted that, unless otherwise clearly specified or limited, the terms "connected" and "connected" should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to the specific circumstances. In the description of the present invention, unless otherwise specified, "plurality" means two or more, which will not be described in detail here.

[0037] Example

[0038] The present invention provides an image denoising method, specifically Figure 1 As shown,

[0039] Step 1: Get the image to be processed.

[0040] Step 2: Construct a Monte Carlo denoising deep learning network and train the Monte Carlo denoising deep learning network; the Monte Carlo denoising deep learning network includes a fast Fourier convolution extractor FFCE, a residual multi-path Mamba encoder RMME and a decoder; the FFCE includes a local convolution path and a global transformation path, and the RMME includes multiple bidirectional Mamba blocks BMB connected in series.

[0041] The Monte Carlo denoising deep learning network consists of a fast Fourier convolution extractor (FFCE), a residual multi-way Mamba encoder (RMME), and a decoder block. The original noisy image X is added to the processed residual R to obtain the denoised output image Y, which is expressed as:

[0042] Y = X + R;

[0043] The purpose of using Monte Carlo denoising deep learning network denoising is to learn the optimal mapping between input data (including noisy images and auxiliary feature information) and output residuals. Mathematically, this process can be described as follows:

[0044]

[0045] In the formula, A normal , A depth and A albedo They are normal, depth and albedo in the auxiliary feature information; F A is the global auxiliary feature information; F N It is the local noise feature information.

[0046] FFCE performs feature extraction to obtain global auxiliary feature information FA and local noise feature information F N . F A Will be reused many times. and F A After being spliced ​​according to the channel latitude and passed to RMME, the next piece of RMME will get what it needs. The next block will repeat this step until the last block is Then it is passed to the decoder for decoding.

[0047] Here, the present invention uses normal, depth and albedo as auxiliary feature information, although other auxiliary feature information can be incorporated, and the sequence connection of the residual multi-way Mamba encoder (RMME) is designed to achieve local auxiliary features and long-range dependencies. Finally, the decoder uses convolution (Conv.), batch normalization (BN) and PReLU activation function to generate residuals.

[0048] The structure of the residual multi-channel Mamba encoder RMME is as follows Figure 2 As shown, Figure 2 (a) is the architecture of the residual multi-channel Mamba encoder RMME. Figure 2 (b) is a series of bidirectional Mamba blocks BMB. The bidirectional Mamba block BMB cuts the two-dimensional data into blocks through convolution with the same stride as the kernel. In order to ensure the consistency of data size, the number of channels is expanded from the original size to the square of the convolution kernel size, and then the latter two dimensions are flattened to one dimension. In order to avoid too large a number of channels passed into the bidirectional SSM (BiSSM), the batch size is exchanged again. There is no doubt that position encoding has a positive impact on visual tasks, and learnable position encoding is added to the input; Figure 2 (c) is that BiSSM passes the input and residual into the Mamba block BMB to obtain the forward result, and at the same time, the reverse result is passed into the Mamba block BMB by the reversed input and the corresponding reversed residual, and the update and residual are added by the corresponding parts of the forward result and the reverse result. Since the output of the present invention is designed to be non-resize-transformed, it is easy to repeat this step multiple times to improve the module learning ability. Finally, since the data size consistency and the non-resize transformation after the bidirectional SSM processing are guaranteed during the segmentation, the output of the last update can be easily restored to the original input size.

[0049] The state space model is the main theoretical technology in RMME. The state space model in control theory can effectively improve the model learning ability while avoiding the quadratic resource overhead. The state space model (SSM) is a linear mathematical framework (continuous model) that maps a one-dimensional input sequence to a one-dimensional output sequence, which can be expressed as:

[0050] h'(t)=Ah(t)+Bx(t);

[0051] y(t)=Ch(t)+Dx(t);

[0052] h(t)∈R N Indicates the hidden state; A∈R N×N Parameters, which determine the influence of the previous hidden state, Parameters determine the weight of the input; x(t) represents the current input, y(t) represents the current output, which is calculated by the hidden state and the current input; Parameters, transforming hidden states into outputs, Parameters, introduce the residual connection weights of the input. In order to discretize the continuous model for practical use, the formula can be expressed as:

[0053]

[0054] In the formula, h t Another way to express the hidden state is h t-1 is the hidden state at the previous moment, x t and t Another way to express the current input and current output respectively.

[0055] The parameters in the continuous model can be discretized by the zero-order holder (ZOH) method. Perform the conversion. The transformation equation is as follows:

[0056]

[0057] Wherein, ΔA and Δb both represent the original continuous input.

[0058] Step 3: Train the Monte Carlo denoising deep learning network, which includes the following steps:

[0059] (1) Obtain image data sets, and divide the multiple image data sets into a training data set and a verification data set, wherein the ratio of the training data set to the verification data set is 95:5.

[0060] This invention uses the 2017 ACFM ( AdversarialThe public dataset of Monte Carlo denoising with conditioned auxiliary feature modulation is used for training and verification. The dataset contains 1109 shots of 8 scenes provided by [Bitterli 2016], rendered using Tungsten. Each shot provides a noisy color image and three noisy auxiliary feature information buffers, including depth, normal and albedo. The corresponding reference images of these noisy images are also provided. All noisy images are rendered with 32spp per pixel, while the reference images are rendered with 32k spp. The present invention randomly selects 95% of the shots as the training dataset, and the remaining 5% of the shots as the verification dataset. All shots are divided into blocks of size 128×128, and a total of 286649 blocks are obtained, of which 272230 are used as training shots and 14419 are used as verification shots.

[0061] (2) Inputting the images in the training data set into the Monte Carlo denoising deep learning network, and training the Monte Carlo denoising deep learning network using a linear combination of a pixel reconstruction loss function and a gradient domain loss function as a total loss function to obtain a trained Monte Carlo denoising deep learning network.

[0062] A common loss function in image restoration is the L2 loss function that calculates the mean square error between the reconstructed image and the reference image. However, relying on L2 in image denoising tasks will cause the reconstructed image to lose details and become too blurred. In order to ensure pixel-level accuracy and perceptual quality, the present invention avoids using L2 and instead uses a linear combination of pixel reconstruction loss and gradient domain L1 loss. The total loss function is:

[0063] L=ω s L s +ω g L g ;

[0064] ω s and ω g is an adjustable weight to control the contribution of each loss. The present invention selects ω s / g = 0.85 / 0.15 to approximate the equilibrium scale, which improves convergence.

[0065] Pixel reconstruction loss L s : The denoised image generated by the model With reference picture I gt The L1 distance pixel reconstruction loss between the two pixels is: i and T i are the i-th pixel of the denoised image and the target image respectively:

[0066]

[0067] Gradient domain L1 loss L g Yes with I gt After calculating the gradient using the Laplace-Gaussian operator LoG, the L1 distance between the two gradients is used as the gradient domain loss:

[0068]

[0069] The L1 loss ensures that blurring artifacts that may occur during reconstruction are avoided, while the LoG second-order derivative operator is very sensitive to image edge contours. Using it as a loss function can further improve SSIM structural similarity.

[0070] Step 4: Input the image to be processed into the trained Monte Carlo denoising deep learning network, use the local convolution path and the global transformation path to extract the auxiliary feature information and noise feature information of the image to be processed respectively, splice the auxiliary feature information and the noise feature information and input them into the bidirectional Mamba block, use the bidirectional Mamba block to perform sequential fusion processing on the spliced ​​features, output the output features of the last BMB through the decoder, and obtain the denoised image.

[0071] The present invention designs a fast Fourier convolution extractor (FFCE), which takes the noisy image and the auxiliary information as two independent inputs, and processes all the auxiliary information channels as a whole through the fast Fourier convolution block, and processes all the auxiliary information channels through the fast Fourier convolution block. Fast Fourier convolution combines the advantages of convolution and fast Fourier transform, and improves the efficiency of multi-scale feature acquisition of non-local receptive fields.

[0072] FFCE consists of a local convolution path and a global transformation path. This process can be expressed in mathematical form as:

[0073]

[0074]

[0075] Here, A is the auxiliary information (global), N is the noise image (local), and is the output feature corresponding to the auxiliary information (global) and the noise image (local), f l represents the convolution operation for local feature extraction, f g represents the Fourier transform used for global feature extraction, f l→g and f g→lThey represent convolution operations for extracting local-global and global-local correlation features, respectively. These operations make full use of the correlation between the noise image (local) and the auxiliary information (global) to achieve the purpose of feature fusion. Compared with other methods, FFCE achieves better feature extraction results with fewer parameters.

[0076] Mamba was first designed for one-dimensional sequential inputs in natural language processing (NLP) tasks. With its linear computational complexity and long-range dependency perception capabilities, Mamba quickly migrated to visual tasks that process two-dimensional data, such as VisionMamba and Vmamba. VisionMamba divides a two-dimensional image into continuous blocks and introduces a bidirectional scanning scheme to model the dependencies between these long-range sequential blocks. Based on the principle of Visual Mamba, Vmamba proposes a cross-scan model (CRS) that expands the receptive field horizontally and vertically and processes different features separately, which has a relatively high memory cost. According to the characteristics of Mamba signal denoising, the Residual Multi-way Mamba Encoder (RMME) is proposed. RMME combines horizontal and vertical scanning schemes and enhances residual connections to achieve an expanded receptive field. In order to optimize MC denoising, RMME introduces a bidirectional Mamba block (BidirectionalMambaBlock, BMB) to sequentially process the fused features as a whole, aiming to reduce memory costs. The detailed process is as follows:

[0077] First, the input auxiliary feature information and noise feature information of RMME are connected, and then the connected features are transposed to obtain the residual addition result and convolution result of the BMB process, which represent the horizontal and vertical scanning results. The horizontal and vertical results are added to the previous RMME results, and RMME produces the final output value.

[0078] The present invention designs the convolution kernel size for BMB slicing in the serial RMME to increase successively. Experiments show that using increasing slicing kernels is effective compared to keeping the size unchanged, because it can extract more information at different scales, just like the FPN design. The output of the last RMME block is processed by a simple three-layer convolution decoder to obtain a denoising mask, and the denoising mask and the noisy image are added to obtain a high-quality noise-free image.

[0079] In order to verify the effect of the invention of the present invention, the method of the present invention is compared with two learning-based methods RAE, DEMC and attention-based method AFGSA. RAE uses a deep recursive network to denoise MC. DEMC designs a dual encoder deep network for MC denoising. AFGSA introduces an attention mechanism into MC denoising. The quantitative comparison results with RAE, DEMC and AFGSA methods on the test group are shown in Table 1 below.

[0080] Table 1 Quantitative comparison results of different methods on the test group

[0081] Model <![CDATA[RMSE(10 -3 )↓]]> PSNR↑ SSIM↑ RAE 7.784 33.00 0.8750 AFGSA(pre) 2.263 35.06 0.9556 AFGSA 3.161 33.62 0.9286 DEMC 2.886 36.00 0.9264 Ours 1.543(-0.72) 37.65(+0.94) 0.9603(+0.0047)

[0082] As shown in Table 1, the method of the present invention achieves the best effect. The numbers in brackets represent the improvements of the method of the present invention.

[0083] like Figure 3 The comparative test results of different models for denoising are shown. Figure 3 (a) Figure 3 (b) Figure 3 (c) and Figure 3 (d) represents the performance of denoising using different models in scenes of specular reflection, highlight, refraction, tiny structure and dark light. The model of the present invention not only accurately restores low-frequency features, but also does a good job in restoring high-frequency details. In the bedroom scene, AFGSA blurred the structured texture of the mattress, causing the bumps to disappear. AFGSA, RAE and DEMC all lost high-frequency details of the bright part of the bed fabric, while the model of the present invention restored the structures and edge details of the two very well. There are also places like this, such as translucent curtains, shadows of potted plants, translucent curtains, textures of wooden floors and tiles, etc. Compared with the other three models, the present invention has achieved improvements in certain indicators.

[0084] like Figure 4 As shown, Figure 4 (a) and Figure 4 (b) respectively counts the resource overhead of the four models under different image sizes. The reasoning speed of RAE based on RNN is almost unaffected by the image size when it does not reach the graphics card bottleneck, but the video memory overhead of RAE is limited by the length of the input sequence. A short sequence will result in a small hidden layer and affect the performance. DEMC, which is composed entirely of CNN, has a faster reasoning speed than RAE, and since it does not need to save the hidden layer like RAE, the video memory overhead is not changed by the image size. The speeds of DEMC and RAE have entered the real-time ray tracing denoising. AFGSA uses the self-attention mechanism, which sacrifices video memory and time to obtain a higher quality denoising effect. Although AFGSA has made many performance optimizations, its overhead is much higher than the other models and grows at a quadratic complexity. Although the model of the present invention requires a large video memory support like AFGSA during training, since the inference does not require the calculation of gradients, the video memory and time overhead are greatly reduced. After the image size is greater than 256×256, the video memory overhead of the present invention is lower than that of the other three models. When outputting 600×600, the denoising effect of the model of the present invention is higher than that of AFGSA, while the video memory overhead is only one-sixth of that of AFGSA, and the time overhead is about one-fourth of that of AFGSA.

[0085] The present invention also provides an image denoising system, comprising an image acquisition module, a model construction module and an image processing module. The image acquisition module is used to acquire an image to be processed; the model construction module is used to construct a Monte Carlo denoising deep learning network and train the Monte Carlo denoising deep learning network; the Monte Carlo denoising deep learning network comprises a fast Fourier convolution extractor FFCE, a residual multi-way Mamba encoder RMME and a decoder; the FFCE comprises a local convolution path and a global transformation path, and the RMME comprises a plurality of bidirectional Mamba blocks BMB connected in series; the image processing module is used to input the image to be processed into the trained Monte Carlo denoising deep learning network, use the local convolution path and the global transformation path to respectively extract auxiliary feature information and noise feature information of the image to be processed, splice the auxiliary feature information and the noise feature information and input them into the bidirectional Mamba block, use the bidirectional Mamba block to sequentially fuse the spliced ​​features, output the output features of the last BMB through the decoder, and obtain the denoised image.

[0086] The present invention also provides a computer device, comprising a memory and a processor; the memory stores a computer program, and the processor is used to run the computer program in the memory to execute the image denoising method.

[0087] The present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and the computer program is suitable for being loaded by a processor to execute the image denoising method.

[0088] The embodiments described above are only preferred specific implementation modes of the present invention, and the protection scope of the present invention is not limited thereto. Any simple changes or equivalent replacements of the technical solutions that can be obviously obtained by any technician familiar with the field within the technical scope disclosed in the present invention belong to the protection scope of the present invention.

Claims

1. An image denoising method, characterized in that: The steps include: Get the image to be processed; Constructing a Monte Carlo denoising deep learning network and training the Monte Carlo denoising deep learning network; the Monte Carlo denoising deep learning network includes a fast Fourier convolution extractor FFCE, a residual multi-path Mamba encoder RMME and a decoder; the FFCE includes a local convolution path and a global transformation path, and the RMME includes a plurality of bidirectional Mamba blocks BMB connected in series; The image to be processed is input into the trained Monte Carlo denoising deep learning network, and the auxiliary feature information and noise feature information of the image to be processed are extracted using the local convolution path and the global transformation path respectively. The auxiliary feature information and the noise feature information are spliced ​​and input into the bidirectional Mamba block. The bidirectional Mamba block is used to sequentially fuse the spliced ​​features, and the output features of the last BMB are output through the decoder to obtain the denoised image.

2. The image denoising method according to claim 1, characterized in that: The training of the Monte Carlo denoising deep learning network comprises the following steps: Get image dataset; Dividing the multiple image data sets into a training data set and a verification data set, wherein the ratio of the training data set to the verification data set is 95:5; The images in the training data set are input into the Monte Carlo denoising deep learning network, and the Monte Carlo denoising deep learning network is trained using a linear combination of a pixel reconstruction loss function and a gradient domain loss function as a total loss function to obtain a trained Monte Carlo denoising deep learning network.

3. The image denoising method according to claim 1, characterized in that: The auxiliary features information depth, normal and albedo.

4. The image denoising method according to claim 1, characterized in that: After the output feature of the last BMB is output by the decoder, the method further includes adding the denoised image to the image to be processed to obtain a denoised image.

5. The image denoising method according to claim 1, characterized in that: The pixel reconstruction loss function is as follows: Where P i and T i are the i-th pixel of the denoised image and the target image respectively, and N is the total number of pixels.

6. The image denoising method according to claim 1, characterized in that: The gradient domain loss function is as follows: Where N is the total number of pixels, is the denoised image gradient obtained using the LoG operator, is the target image gradient obtained using the LoG operator.

7. The image denoising system according to claim 1, characterized in that: include: An image acquisition module, used for acquiring an image to be processed; A model building module, used for building a Monte Carlo denoising deep learning network and training the Monte Carlo denoising deep learning network; the Monte Carlo denoising deep learning network includes a fast Fourier convolution extractor FFCE, a residual multi-path Mamba encoder RMME and a decoder; the FFCE includes a local convolution path and a global transformation path, and the RMME includes a plurality of bidirectional Mamba blocks BMB connected in series; The image processing module is used to input the image to be processed into the trained Monte Carlo denoising deep learning network, use the local convolution path and the global transformation path to respectively extract the auxiliary feature information and noise feature information of the image to be processed, splice the auxiliary feature information and the noise feature information and input them into the bidirectional Mamba block, use the bidirectional Mamba block to perform sequential fusion processing on the spliced ​​features, output the output features of the last BMB through the decoder, and obtain the denoised image.

8. A computer device, characterized in that: It comprises a memory and a processor; the memory stores a computer program, and the processor is used to run the computer program in the memory to execute the image denoising method according to any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and the computer program is suitable for being loaded by a processor to execute the image denoising method according to any one of claims 1 to 6.

Images