An image denoising method, system, device and medium

Abstract

The application provides an image denoising method, system, device and medium, and belongs to the field of computer graphics. The method comprises the following steps: acquiring an image to be processed; constructing a Monte Carlo denoising deep learning network, which comprises a fast Fourier convolution extractor (FFCE), a residual multi-path Manbha encoder (RMME) and a decoder; the FFCE comprises a local convolution path and a global transformation path, and the RMME comprises a plurality of series bidirectional Manbha blocks (BMB); inputting the image to be processed into the Monte Carlo denoising deep learning network, using the local convolution path and the global transformation path to extract auxiliary feature information and noise feature information of the image to be processed respectively, inputting the two into the bidirectional Manbha block after splicing for sequential fusion processing, and outputting the feature of the last BMB through the decoder to obtain a denoised image. The application can reduce the calculation and memory cost.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of computer graphics, and particularly relates to an image denoising method, system, device and medium. BACKGROUND

[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 integral result in the case of limited samples, it is affected by noise due to the inaccuracy between the estimated value and the true value. Sampling more light can improve the resolution of the image, but the increased computational cost will significantly affect the feasibility of Monte Carlo rendering, especially for large-size images.

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

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

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

[0006] obtaining an 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 comprises a fast Fourier convolution extractor (FFCE), a residual multi-path manbang encoder (RMME) and a decoder; the FFCE comprises a local convolution path and a global transformation path, and the RMME comprises a plurality of series of bidirectional manbang blocks (BMBs);

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

[0009] Preferably, the Monte Carlo denoising deep learning network is trained, including the following steps:

[0010] An image data set is obtained.

[0011] The plurality of image data sets are divided into a training data set and a validation data set, and the ratio of the training data set to the validation data set is 95:5.

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

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

[0014] Preferably, after the output features of the last BMB are output through the decoder, the denoised image is added to the image to be processed to obtain a denoised image.

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

[0016]

[0017] In the formula, P i and T i are the i-th pixels 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] In the formula, N is the total number of pixels, is a denoised image gradient obtained by using a LoG operator, is a target image gradient obtained by using a LoG operator.

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

[0022] An image acquisition module is configured to acquire an image to be processed.

[0023] a model construction module, configured 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-path 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 serial bidirectional Mamba blocks (BMBs);

[0024] an image processing module, configured to input a to-be-processed image into the trained Monte Carlo denoising deep learning network, extract auxiliary feature information and noise feature information of the to-be-processed image using the local convolution path and the global transformation path respectively, input the auxiliary feature information and the noise feature information into the bidirectional Mamba blocks after splicing, sequentially fuse the features after splicing using the bidirectional Mamba blocks, output the features of the last BMB through the decoder, and obtain a denoised image.

[0025] The application further provides a computer device comprising a memory and a processor; the memory stores a computer program, and the processor is configured to run the computer program in the memory to execute the image denoising method.

[0026] The application further provides a computer readable storage medium storing a computer program, and the computer program is adapted to be loaded by a processor to execute the image denoising method.

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

[0028] The FFCE can extract auxiliary feature information and noise feature information of the to-be-processed image, complete the interaction between the noise feature information and the auxiliary information, achieve better feature extraction results with fewer parameters, and reduce the calculation cost; the RMME introduces the bidirectional Mamba blocks to adopt a bidirectional continuous scanning scheme, utilize the long-range spatial dependence relationship in the horizontal and vertical directions, serially encode the extracted features, sequentially process the fused features as a whole, and thus reduce the calculation and memory costs. BRIEF DESCRIPTION OF DRAWINGS

[0029] In order to more clearly illustrate the embodiments of the application and the design scheme thereof, the drawings required by the embodiments will be briefly introduced as follows. The drawings in the following description are only partial embodiments of the application, and other drawings can be obtained by those skilled in the art without creative labor on the basis of the drawings.

[0030] Figure 1 The flowchart of the image denoising method of the embodiments of the application;

[0031] Figure 2 is a RMME module diagram; wherein, Figure 2 (a) is an architecture of a residual multi-path Mamba encoder RMME, Figure 2 (b) is a plurality of cascaded bidirectional Mamba blocks BMBs; Figure 2 (c) is a flowchart of BiSSM passing input and residual into Mamba blocks BMBs;

[0032] Figure 3 is a comparison test result of different model denoising, wherein, Figure 3 (a) of, Figure 3 (b) of, Figure 3 (c) of, and Figure 3 (d) of respectively represent the performance after using different models for denoising in the scene of specular reflection, highlight, refraction and microstructure and dark light;

[0033] Figure 4 is a model overhead comparison diagram; wherein, Figure 4 (a) and Figure 4 (b) respectively statistics the resource overhead of four models under different image sizes. DETAILED DESCRIPTION

[0034] In order to make the technical personnel of the present application better understand the technical solutions and can be implemented, the present application is described in detail below in conjunction with the drawings and specific examples. The following examples are only used to more clearly illustrate the technical solutions of the present application, and cannot be used to limit the protection scope of the present application.

[0035] In the description of the present application, it should be understood that the terms "center", "longitudinal", "transverse", "length", "width", "thickness", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", "axial", "radial", "circumferential" and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, and are only used to facilitate the description of the technical solutions of the present application and simplify the description, and therefore cannot be understood as indicating or implying that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as limiting the present application.

[0036] In addition, the terms "first", "second" and the like are used only for descriptive purposes and should not be construed as indicating or implying relative importance. In the description of the present application, it should be noted that, unless otherwise explicitly specified or limited, the terms "connected", "connected" should be understood in a broad sense, for example, it can be fixed connection, or detachable connection, or integral connection; it can be mechanical connection, or electrical connection; it can be directly connected, or indirectly connected through intermediate medium. For those skilled in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances. In the description of the present application, unless otherwise specified, the meaning of "a plurality of" is two or more, which will not be described here.

[0037] Embodiments

[0038] The present application provides an image denoising method, specifically as shown in Figure 1

[0039] Step 1: obtaining an image to be processed.

[0040] Step 2: constructing a Monte Carlo denoising deep learning network and training the Monte Carlo denoising deep learning network; the Monte Carlo denoising deep learning network comprises a fast Fourier convolution extractor FFCE, a residual multi-path manbang encoder RMME and a decoder; the FFCE comprises a local convolution path and a global transformation path, and the RMME comprises a plurality of series of bidirectional manbang blocks BMB.

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

[0042] Y=X+R;

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

[0044]

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

[0046] The FFCE performs feature extraction to obtain the global auxiliary feature information F​A and local noise feature information F N F A It will be reused multiple times. and F A After being pieced together according to channel latitude and transmitted to the RMME, it will obtain the data needed for the next RMME. The next piece will repeat this process until the last piece is obtained. It is then passed to the decoder for decoding.

[0047] Here, the present invention uses normals, depth, and albedo as auxiliary feature information. Although other auxiliary feature information can be incorporated, the sequence concatenation of the Residual Multi-Mamba Encoder (RMME) is designed to achieve local auxiliary features and long-range dependencies. Finally, the decoder uses convolution, batch normalization (BN), and PReLU activation functions to generate residuals.

[0048] The structure of the residual multi-channel Mamba encoder RMME is as follows: Figure 2 As shown, Figure 2 (a) shows the architecture of the residual multi-channel Mamba encoder RMME. Figure 2 (b) consists of multiple cascaded bidirectional Mamba blocks (BMBs). The bidirectional Mamba blocks (BMBs) divide 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 increased from the original size to the square of the convolution kernel size. Then, the last two dimensions are flattened to one dimension. In order to avoid the number of channels input to the bidirectional SSM (BiSSM) being too large, the channels are swapped again except for the batch size. Undoubtedly, positional encoding has a positive impact on visual tasks by adding learnable positional encoding to the input. Figure 2 (c) involves BiSSM feeding the input and residual into the Mamba block (BMB) to obtain the forward result, while the reverse result is fed into the Mamba block (BMB) from the inverted input and the corresponding inverted residual. The update and residual are then summed by adding the corresponding parts of the forward and reverse results. Since this invention is designed to have no output size transformation, this step can be easily repeated multiple times to improve the module's learning ability. Finally, because the data size consistency is guaranteed during the splitting and there is no size transformation after bidirectional SSM processing, the output of the last update can be easily restored to the original input size.

[0049] State-space models are a key theoretical technique in RMME (Real-Time Mechanism). Using state-space models from cybernetics can effectively improve model learning capabilities while avoiding quadratic resource overhead. A state-space model (SSM) is a linear mathematical framework (continuous model) that maps a one-dimensional input sequence to a one-dimensional output sequence, and can be represented as:

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

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

[0052] h(t) element-of R N denotes the hidden state; A element-of R N×N a parameter that determines the influence of the previous hidden state, a parameter that determines the weight of the input; x(t) denotes the current input, and y(t) denotes the current output, which is calculated from the hidden state and the current input; a parameter that converts the hidden state into the output, a parameter that introduces the remaining connection weight of the input. In order to discretize the continuous model and facilitate actual use, the formula can be expressed as:

[0053]

[0054] In the formula, h t is another expression of the hidden state, h t-1 is the hidden state at the previous time, x t and y t are another expression of the current input and the current output, respectively.

[0055] Through the zero-order hold (ZOH) discretization method, the parameters in the continuous model can be converted. The transformation equation is as follows:

[0056]

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

[0058] Step 3: training the Monte Carlo denoising deep learning network, specifically including the following steps:

[0059] (1) obtaining an image data set, dividing a plurality of image data sets into a training data set and a verification data set, the proportion of the training data set and the verification data set being 95:5.

[0060] The present application uses the 2017 ACFM ( AdversarialThe dataset disclosed in [Bitterli 2016] is used for training and validation. The dataset contains 1109 shots of 8 scenes provided by [Bitterli 2016] using Tungsten rendering. 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 at 32 spp per pixel, while the reference images are rendered at 32k spp. The invention randomly selects 95% of the shots as the training dataset, while the remaining 5% of the shots are used as the validation dataset. All shots are divided into blocks of size 128x128, resulting in a total of 286649 blocks, of which 272230 are used as training shots and 14419 are used as validation shots.

[0061] (2) input the images in the training dataset into the Monte Carlo denoising deep learning network, train the Monte Carlo denoising deep learning network with a linear combination of the pixel reconstruction loss function and the gradient domain loss function as the total loss function, and obtain the trained Monte Carlo denoising deep learning network.

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

[0063] L = ω s L s + ω g L g ;

[0064] ω s and ω g are adjustable weights that control the contribution of each loss, and the invention selects ω s / g = 0.85 / 0.15 to approximate the equal scale, which improves the convergence.

[0065] The pixel reconstruction loss L s : the L1 distance between the denoised image generated by the model and the reference image I gt , where N is the total number of pixels, P i and T i are the i-th pixels of the denoised image and the target image, respectively:

[0066]

[0067] Gradient domain L1 loss L g is to with I gt The L1 distance of the two gradients is used as the gradient domain loss after calculating the gradient by the Laplacian-Gaussian operator LoG:

[0068]

[0069] The L1 loss ensures the avoidance of blurred artifacts that may occur during reconstruction, and the LoG second derivative operator is very sensitive to image edge contours, so using it as a loss function can further improve the SSIM structural similarity of .

[0070] Step 4: for inputting the to-be-processed image into the trained Monte Carlo denoising deep learning network, using a local convolution path and a global transformation path to extract auxiliary feature information and noise feature information of the to-be-processed image respectively, splicing the auxiliary feature information and the noise feature information and inputting them into a bidirectional Mamba block, using the bidirectional Mamba block to sequentially fuse the spliced features, outputting the features of the last BMB through the decoder, and obtaining a denoised image.

[0071] The present application designs a fast Fourier convolution extractor (FFCE), which takes noise images and auxiliary information as two independent inputs, and processes all auxiliary information channels as a whole through a fast Fourier convolution block, and outputs corresponding feature results respectively. It processes all auxiliary information channels through a 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 field.

[0072] The FFCE is composed of a local convolution path and a global transformation path. This process can be mathematically expressed as:

[0073]

[0074]

[0075] Here, A is auxiliary information (global), N is a noise image (local), and are the output features corresponding to the auxiliary information (global) and the noise image (local), f l represents a convolution operation for local feature extraction, f g represents a Fourier transform for global feature extraction, f l→g and f g→lRespectively represent the convolution operations for extracting local-global and global-local correlation features, which make full use of the correlation between the noisy 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 input in natural language processing (NLP) tasks. With its linear computational complexity and long-range dependency awareness, Mamba quickly migrated to visual tasks that handle two-dimensional data, such as VisionMamba and Vmamba. VisionMamba divides two-dimensional images into consecutive blocks and introduces a bidirectional scanning scheme to model the dependencies between these long-range sequential blocks. Based on the principle of VisionMamba, Vmamba proposes a cross-scan model (CRS) that extends the receptive field horizontally and vertically and processes different features respectively, which has relatively high memory cost. According to the characteristics of Mamba signal denoising, a residual multi-path Mamba encoder (RMME) is proposed. RMME combines horizontal and vertical scanning schemes and enhances residual connections to achieve an enlarged receptive field. To optimize MC denoising, RMME introduces a bidirectional Mamba block (BMB) that sequentially processes the fused features as a whole, aiming to reduce memory cost. The detailed process is as follows:

[0077] First, the input auxiliary feature information and noisy feature information of RMME are connected, and then the connected feature is transposed to get the residual addition result and the 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 result, and the RMME produces the final output value.

[0078] The invention designs a series of RMMEs with convolution kernel sizes for BMB cutting blocks that increase in size. Experiments show that using incremental cutting convolution kernels is effective compared to keeping the size unchanged, as 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. Adding the denoising mask to the noisy image obtains a high-quality noise-free image.

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

[0080] Table 1. Quantitative comparison results of different methods in the experimental group

[0081] Model 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 results. The numbers in parentheses indicate improvements to the method of the present invention.

[0083] like Figure 3 The results of comparative experiments on denoising using different models are shown. Figure 3 of (a), Figure 3 (b) Figure 3 (c) and Figure 3 (d) represents the performance of different models after denoising in scenes with specular reflection, highlights, refraction, microstructure, and low light. The model of this invention not only accurately recovers low-frequency features but also performs well in recovering high-frequency details. In the bedroom scene, AFGSA blurred the structured texture of the mattress, causing the bumps to disappear. In bright areas of the bed fabric, AFGSA, RAE, and DEMC all lost high-frequency details, while the model of this invention effectively recovered the structure and edge details of both. Similar examples include translucent curtains, the shadows of potted plants, light-transmitting curtains, and the textures of wooden floors and tiles. Compared to the other three models, this invention achieves certain improvements in all metrics.

[0084] like Figure 4 As shown, Figure 4 (a) and Figure 4 (b) shows the resource overhead of the four models under different image sizes. The RNN-based RAE inference speed is almost unaffected by image size before reaching the GPU bottleneck; however, RAE's memory overhead is limited by the input sequence length, and a short sequence leads to a small hidden layer, thus affecting performance. DEMC, composed entirely of CNNs, has a faster inference speed than RAE, and since it doesn't need to store hidden layers like RAE, its memory overhead is independent of image size. DEMC and RAE's speeds are comparable to real-time ray tracing denoising. AFGSA, due to its self-attention mechanism, sacrifices memory and time for higher denoising quality. Although AFGSA has undergone many performance optimizations, its overhead is much higher than the other models, increasing with quadratic complexity. Although the model of this invention requires a large amount of memory during training, similar to AFGSA, its memory and time overhead are significantly reduced because inference does not require gradient calculation. When the image size is larger than 256×256, the memory overhead of this invention is lower than the other three models. When outputting 600×600, the noise reduction effect of the model of this invention is higher than that of AFGSA, while the memory overhead is only one-sixth and the time overhead is about one-quarter.

[0085] The application further provides an image denoising system, comprising an image acquisition module, a model construction module and an image processing module. The image acquisition module is configured to acquire an image to be processed; the model construction module is configured 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-path 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 series-connected bidirectional Mamba blocks (BMBs); and the image processing module is configured to input the image to be processed into the trained Monte Carlo denoising deep learning network, extract auxiliary feature information and noise feature information of the image to be processed using the local convolution path and the global transformation path respectively, input the auxiliary feature information and the noise feature information into the bidirectional Mamba blocks after splicing, sequentially fuse the features after splicing using the bidirectional Mamba blocks, output the features of the last BMB through the decoder, and obtain a denoised image.

[0086] The application further provides a computer device, comprising a memory and a processor; the memory stores a computer program, and the processor is configured to run the computer program in the memory to execute the image denoising method.

[0087] The application further provides a computer readable storage medium, which stores a computer program, and the computer program is adapted to be loaded by a processor to execute the image denoising method.

[0088] The above-described embodiments are merely preferred specific embodiments of the application, and the protection scope of the application is not limited thereto, and any simple change or equivalent replacement of the technical solutions within the technical range disclosed by the application should be considered within the protection scope of the application.

Claims

1. A method of image denoising, characterized in that, The method comprises the following steps: obtaining an image to be processed and auxiliary information; constructing a Monte Carlo denoising deep learning network and training the Monte Carlo denoising deep learning network; the Monte Carlo denoising deep learning network comprises a fast Fourier convolution extractor (FFCE), a residual multi-path 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 serial bidirectional Mamba blocks (BMBs); inputting the image to be processed and the auxiliary information into the trained Monte Carlo denoising deep learning network, extracting noise feature information of the image to be processed by using the local convolution path, extracting auxiliary feature information of the image to be processed by using the global transformation path, inputting the auxiliary feature information and the noise feature information into the bidirectional Mamba blocks after splicing, sequentially fusing the features after splicing by using the bidirectional Mamba blocks, outputting the feature of the last BMB through the decoder to obtain a denoised image; the denoising process of the Monte Carlo denoising deep learning network is described as follows: In the formula, A normal , A depth , and A albedo are the normal, depth, and albedo in the auxiliary information, respectively; F A is the global auxiliary feature information; is the local noise feature information.

2. The image denoising method of claim 1, wherein, the training of the Monte Carlo denoising deep learning network comprises the following steps: obtaining an image data set; dividing a plurality of image data sets into a training data set and a verification data set, and the proportion of the training data set and the verification data set is 95:5; inputting the images in the training data set into the Monte Carlo denoising deep learning network, taking a linear combination of a pixel reconstruction loss function and a gradient domain loss function as a total loss function, training the Monte Carlo denoising deep learning network, and obtaining a trained Monte Carlo denoising deep learning network.

3. The image denoising method of claim 1, wherein, After the output feature of the last BMB is output through the decoder, the method further comprises adding the denoised image to the image to be processed to obtain a denoised image.

4. The image denoising method of claim 2, wherein, The pixel reconstruction loss function is as follows: where P i and T i are the i-th pixels of the denoised image and the target image, respectively, and N is the total number of pixels.

5. The image denoising method of claim 2, wherein, The gradient domain loss function is as follows: In the formula, N is the total number of pixels, is the denoised image gradient calculated using the LoG operator, is the target image gradient calculated using the LoG operator.

6. An image denoising system characterized by, comprising: an image acquisition module configured to obtain an image to be processed and auxiliary information; a model construction module configured 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-path 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 serial bidirectional Mamba blocks (BMBs); an image processing module configured to input the image to be processed and the auxiliary information into the trained Monte Carlo denoising deep learning network, extract noise feature information of the image to be processed by using the local convolution path, extract auxiliary feature information of the image to be processed by using the global transformation path, input the auxiliary feature information and the noise feature information into the bidirectional Mamba blocks after splicing, sequentially fuse the features after splicing by using the bidirectional Mamba blocks, output the feature of the last BMB through the decoder to obtain a denoised image; the denoising process of the Monte Carlo denoising deep learning network is described as follows: In the formula, A normal , A depth , and A albedo are the normal, depth, and albedo in the auxiliary information respectively; F A is the global auxiliary feature information; is the local noise feature information.

7. A computer device, comprising: An image denoising method as claimed in any one of claims 1-5 is executed by 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 as claimed in any one of claims 1-5.

8. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a computer program, and the computer program is adapted to be loaded by a processor to execute the image denoising method as claimed in any one of claims 1-5.

Images