A monte carlo rendering image denoising method based on deep learning

Abstract

The application belongs to the field of computer graphics, and particularly relates to a Monte Carlo rendering picture denoising method based on deep learning; the method comprises the following steps: obtaining a data set by rendering an image through a Monte Carlo volume rendering method, and pre-processing the data set to obtain a training data set; constructing a kernel prediction model based on a U-net network, and training the kernel prediction model through the training data set; wherein a loss function containing an SMAPE term and a highlight term is used to calculate a prediction error; the trained kernel prediction model is used to denoise a high-noise volume rendering picture under a low sampling number, and a denoised picture is obtained; the application can improve the denoising effect and obtain a high-quality rendering picture.

Description

Technical Field

[0001] This invention belongs to the field of computer graphics, specifically relating to a method for denoising Monte Carlo rendering images based on deep learning. Background Technology

[0002] Monte Carlo volumetric rendering, as an advanced image rendering method, has been widely used in many fields. Particularly in medical imaging, scientific visualization, and film special effects production, this technology is favored for its ability to highly realistically simulate the complex interactions between light and matter. In medical imaging, Monte Carlo volumetric rendering helps doctors observe the internal structures of the human body more clearly, especially when processing images of complex tissues and organs. For example, in tumor diagnosis and analysis of neurological diseases, this technology provides more accurate and detailed views than traditional imaging methods. In scientific visualization, Monte Carlo volumetric rendering allows researchers to intuitively observe and analyze complex scientific data, such as weather models and geological structures, thereby deepening their understanding of these phenomena. In film special effects production, this technology is used to create realistic visual effects, such as simulating natural phenomena (like clouds, smoke, and fire) or more complex scenes, such as realistic organisms or extraterrestrial environments.

[0003] However, Monte Carlo volumetric rendering technology faces a major challenge in practical applications: high levels of noise. This noise becomes particularly pronounced when computational resources are limited and sampling rates are low, severely impacting image quality. In medical imaging, this noise can interfere with doctors' interpretation of images, affecting diagnostic accuracy. For example, when analyzing MRI or CT scans, noise may mask or obscure crucial biomarkers, leading to misdiagnosis or missed diagnosis. Similarly, in scientific research, noise interferes with data analysis, affecting the accuracy of research results. In film special effects production, high levels of noise require increased rendering time and computational resources, which not only increases costs but also prolongs production cycles. Therefore, effectively reducing noise in Monte Carlo volumetric rendering while maintaining image quality and computational efficiency has become a key research focus in this field.

[0004] With the rapid development of deep learning technology, its application in image processing has shown enormous potential. Deep learning, especially in image recognition, classification, and processing, has achieved remarkable results. Applying deep learning to Monte Carlo volumetric rendering denoising provides an effective approach to solving the aforementioned noise problem. By training deep learning models to identify and remove noise during the rendering process, the required computational resources and time can be significantly reduced without sacrificing image quality. This method can not only improve the accuracy and reliability of medical imaging but also accelerate the process of scientific research and film special effects production. Summary of the Invention

[0005] To address the problems existing in the prior art, this invention proposes a deep learning-based Monte Carlo rendering image denoising method. This method uses deep learning to denoise Monte Carlo volumetric rendering images at low sampling rates, significantly improving the rendering quality even at low sampling rates.

[0006] The specific plan includes:

[0007] S1. Render the images using the Monte Carlo volume rendering method to obtain the dataset, and preprocess the dataset to obtain the training dataset;

[0008] S2. Construct a kernel prediction model based on the U-net network and train the kernel prediction model using the training dataset; the prediction error is calculated using a loss function that includes the SMAPE term and the highlight term.

[0009] S3. Use the trained kernel prediction model to denoise the high-noise volumetric rendering image with low sampling number to obtain the denoised image.

[0010] Further, step S1 renders the image using the Monte Carlo volume rendering method to obtain the dataset, including:

[0011] S11. Based on the Monte Carlo volume path tracing method, the volume data is rendered at a sampling rate of 1spp to obtain a noisy image. The delta tracking method and ratio tracking method are used for distance sampling and transmittance calculation, respectively. At the same time, the distance sampling and transmittance calculation are accelerated by the DDA method.

[0012] S12. Obtain the position of the volume data in step S11 when the first scattering occurs using the delta tracking method, and obtain the world space coordinates, volume scattering rate, position normal, and depth of that position;

[0013] S13. Normalize the world space coordinates and map the normalized world space coordinates to the range [0,1] to obtain the first type of three-channel color image; then obtain the second type of three-channel color image based on the volume scattering rate;

[0014] S14. Calculate the gradient of the position normal in the x, y, and z directions using the central difference method, normalize the three gradients, and map the normalized three gradients to the range [0,1] to obtain the third type of three-channel color image.

[0015] S15. Use a nonlinear transformation method to convert the depth to a nonlinear depth, and output a single-channel grayscale image based on the nonlinear depth;

[0016] S16. Based on the Monte Carlo volume path tracing method, the volume data in step S11 is rendered at a sampling rate of 4096 spp to obtain a reference image. The delta tracking method and ratio tracking method are used for distance sampling and transmittance calculation, respectively. At the same time, the distance sampling and transmittance calculation are accelerated according to the DDA method.

[0017] S17. Combine the noisy image, the first type of three-channel color image, the second type of three-channel color image, the third type of three-channel color image, the single-channel gray image, and the reference image into a set of data;

[0018] S18. Repeat steps S11-S17 to obtain multiple sets of data to form a dataset.

[0019] Furthermore, step S15 uses a nonlinear transformation method to convert the depth into a nonlinear depth, expressed as:

[0020]

[0021] Among them, Z NonLinear The nonlinear depth is represented by z, which represents the depth of the position where the volume data undergoes the first scattering using the delta tracking method. far represents the far plane depth of the camera, and near represents the near plane depth of the camera.

[0022] Further, step S1 preprocesses the dataset to obtain the training dataset, including:

[0023] For each set of data in the dataset, the noisy image, the first type of three-channel color image, the second type of three-channel color image, the third type of three-channel color image, the single-channel gray image, and the reference image are segmented separately. Each image is divided into 128×128 small blocks, and each small block is randomly flipped; finally, multiple sets of training data are obtained.

[0024] If the noise image and the reference image in a set of data are output images in the HDR color space, then tonemapping is first used to map the noise image and the reference image back to the LDR color space, and then segmentation processing is performed.

[0025] Further, step S2 trains a kernel prediction model based on the U-net network using the training dataset, including:

[0026] S21. Extract a set of training data, divide the noisy image by the second type of three-channel color image to obtain the irradiance map;

[0027] S22. The irradiance image, the first three-channel color image, the third three-channel color image, and the single-channel grayscale image are stitched together to obtain a fused image;

[0028] S23. Input the fused image into the kernel prediction model to obtain a multi-channel feature map, and extract the filter kernel of each channel in the multi-channel feature map to obtain a three-channel filter kernel;

[0029] S24. Use a three-channel filter kernel to denoise the irradiance image, and multiply the denoised irradiance image with the second three-channel color image to obtain the predicted image;

[0030] S25. Calculate the prediction error between the predicted image and the reference image using a loss function that includes the SMAPE term and the specular term, and then backpropagate it.

[0031] S26. Optimize the model parameters using the Adam optimizer, and use weight decay and gradient clipping;

[0032] S27. Repeat steps S21-26 until the maximum number of training iterations is reached or the model converges.

[0033] Furthermore, the kernel prediction model includes four encoders, four decoders, and one 1×1 convolutional layer cascaded in sequence; each encoder includes two 3x3 convolutional kernels and a 2x downsampling; the first three encoders have 32 channels for both input and output, and the last encoder has 32 channels for input and 64 channels for output; each decoder includes two 3x3 convolutional kernels and a 2x upsampling; the first three decoders have 64 channels for both input and output, and the last decoder has 64 channels for input and 32 channels for output; the 1×1 convolutional layer outputs 27 channels.

[0034] Furthermore, the loss function including the SMAPE term and the highlight term is expressed as:

[0035] L = 0.8·L SMAPE +0.2·L spec

[0036] Among them, L SMAPE For the SMAPE term, the specific formula is:

[0037]

[0038] Where x is the reference image, y is the predicted image, N is the total number of pixels in the image, C is the number of channels in the image, and ε is a constant; p,c y represents the predicted color value of the p-th pixel in the c-th channel of the image. p,c This represents the color value of the p-th pixel in the c-th channel of the reference image;

[0039] L spec For the highlight term, the specific formula is:

[0040]

[0041] Where, x c,i This represents the color value of the i-th pixel in the c-th channel of the reference image. y represents the standard deviation of the color values ​​of all pixels in the neighborhood of the i-th pixel in the c-th channel of the reference image; c,i This represents the predicted color value of the i-th pixel in the c-th channel of the image. This represents the standard deviation of the color values ​​of all pixels in the neighborhood of the i-th pixel in the c-th channel of the predicted image. The neighborhood is 11×11 in size.

[0042] Furthermore, step S3 uses the trained kernel prediction model to denoise the high-noise volumetric rendering image with a low sampling number, resulting in the following denoised image:

[0043] S31. Use motion vectors to calculate the previous frame image corresponding to the high-noise volume rendering image with low sampling number;

[0044] S32. The high-noise volumetric rendering image is color-blended with its corresponding previous frame image using an exponential moving average formula to obtain a blended image; the exponential moving average formula is expressed as:

[0045] C′ i =α·C i +(1-α)·C′ i-1

[0046] Among them, C i C′ represents the color value of the i-th pixel in a high-noise volumetric rendering image. i-1 It is the color value of the i-th pixel in the high-noise volumetric rendering image after mixing with the color value of the i-th pixel in the previous frame image, where α is the smoothing factor;

[0047] S33. Replace the high-noise volumetric rendering image with a hybrid image and input it into the kernel prediction model, then output the denoised result.

[0048] The beneficial effects of this invention are:

[0049] This invention proposes a deep learning-based Monte Carlo volume rendering image denoising method. Addressing the severe noise problem in Monte Carlo volume rendering at low sampling rates, this invention introduces a U-Net network predictive filtering kernel to reduce noise and improve rendering efficiency, enabling the denoised image to rival high-sampling-rate Monte Carlo volume rendering images. In the field of medical imaging, this invention can provide clearer and more accurate images, greatly assisting doctors in disease diagnosis and treatment planning. In the field of scientific visualization, it can help scientists observe and analyze complex scientific data more accurately, thereby promoting research progress in multiple fields such as physics, chemistry, and biology. Furthermore, this technology has enormous application potential in the digital entertainment industry, such as film special effects production and video game development. It can significantly improve rendering efficiency while maintaining image realism, making it possible to create more realistic and engaging visual experiences. Attached Figure Description

[0050] Figure 1 This is a flowchart of a Monte Carlo rendering image denoising method based on deep learning according to the present invention;

[0051] Figure 2 This is a structural diagram of the kernel prediction model of the present invention. Detailed Implementation

[0052] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0053] This invention provides a deep learning-based method for denoising Monte Carlo rendering images, such as... Figure 1 As shown, it includes the following steps:

[0054] S1. Render the images using the Monte Carlo volume rendering method to obtain the dataset, and preprocess the dataset to obtain the training dataset.

[0055] Specifically, step S1 renders the image using the Monte Carlo volume rendering method to obtain the dataset, including:

[0056] S11. Based on the Monte Carlo volume path tracing method, a noisy image is obtained by rendering the volume data at a sampling rate of 1spp. The delta tracking method and the ratio tracking method are used for distance sampling and transmittance calculation, respectively. At the same time, the distance sampling and transmittance calculation are accelerated by the DDA (Digital Differential Analyzer) method. Furthermore, trilinear interpolation sampling is used when sampling the relevant voxel values.

[0057] S12. Obtain the position of the volume data in step S11 when the first scattering occurs using the delta tracking method, and obtain the world space coordinates, volume scattering rate, position normal, and depth of that position;

[0058] S13. Normalize the world space coordinates and map the normalized world space coordinates to the range [0,1] to obtain the first type of three-channel color image; then obtain the second type of three-channel color image based on the volume scattering rate;

[0059] S14. Calculate the gradient of the position normal in the x, y, and z directions using the central difference method, normalize the three gradients, and map the normalized three gradients to the range [0,1] to obtain the third type of three-channel color image.

[0060] S15. In order to reduce accuracy error, a nonlinear transformation method is used to convert the depth into a nonlinear depth, and a single-channel gray image is output based on the nonlinear depth;

[0061] Specifically, step S15 uses a nonlinear transformation method to convert the depth into a nonlinear depth, as shown below:

[0062]

[0063] Among them, Z NonLinear The nonlinear depth is represented by z, which represents the depth of the position where the volume data undergoes the first scattering using the delta tracking method. far represents the far plane depth of the camera, and near represents the near plane depth of the camera.

[0064] S16. Based on the Monte Carlo volume path tracing method, the volume data in step S11 is rendered at a sampling rate of 4096 spp to obtain a reference image. The delta tracking method and ratio tracking method are used for distance sampling and transmittance calculation, respectively. At the same time, the distance sampling and transmittance calculation are accelerated according to the DDA method.

[0065] S17. Combine the noisy image, the first type of three-channel color image, the second type of three-channel color image, the third type of three-channel color image, the single-channel gray image, and the reference image into a set of data;

[0066] S18. Repeat steps S11-S17 to obtain multiple sets of data to form a dataset.

[0067] Specifically, step S1 preprocesses the dataset to obtain the training dataset, including:

[0068] For each set of data in the dataset, the noisy image, the first type of three-channel color image, the second type of three-channel color image, the third type of three-channel color image, the single-channel grayscale image, and the reference image are segmented. Each image is divided into 128×128 blocks, and each block is randomly flipped to increase data diversity and enhance the network's learning and generalization abilities. The segmented noisy image, the first type of three-channel color image, the second type of three-channel color image, the third type of three-channel color image, the single-channel grayscale image, and the reference image are randomly combined to form a new set of data as training data, ultimately resulting in multiple sets of training data.

[0069] If the noisy image and the reference image in a set of data are output images in the HDR (High-Dynamic Range) color space, then tonemapping is first used to map the noisy image and the reference image back to the LDR (Low-Dynamic Range) color space, and then segmentation processing is performed.

[0070] S2. Construct a kernel prediction model based on the U-net network and train the kernel prediction model using the training dataset; the prediction error is calculated using a loss function that includes SMAPE terms and highlight terms.

[0071] Specifically, step S2 trains the kernel prediction model using the training dataset, including:

[0072] S21. Extract a set of training data, divide the noisy image by the second type of three-channel color image to obtain the irradiance map;

[0073] S22. The irradiance image, the first three-channel color image, the third three-channel color image, and the single-channel grayscale image are stitched together to obtain a fused image;

[0074] S23. Input the fused image into the kernel prediction model to obtain a multi-channel feature map, and extract the filter kernel of each channel in the multi-channel feature map to obtain a three-channel filter kernel;

[0075] Specifically, such as Figure 2As shown, the kernel prediction model comprises four encoders, four decoders, and one 1×1 convolutional layer cascaded in sequence. Each encoder includes two 3x3 convolutional kernels and a 2x downsampling layer. The first three encoders have 32 channels for both input and output, while the last encoder has 32 channels for input and 64 channels for output. Each decoder includes two 3x3 convolutional kernels and a 2x upsampling layer. The first three decoders have 64 channels for both input and output, while the last decoder has 64 channels for input and 32 channels for output. The 1×1 convolutional layer outputs 27 channels. Because this invention selects a 3×3 filtering kernel, and because it generally involves color images, the kernel prediction model ultimately outputs a 27-channel feature map.

[0076] S24. Use a filter kernel to denoise the irradiance image, and multiply the denoised irradiance image with the second three-channel color image to obtain the predicted image;

[0077] S25. Calculate the prediction error between the predicted image and the reference image using a loss function that includes the SMAPE term and the specular term, and then backpropagate it.

[0078] S26. Optimize the model parameters using the Adam optimizer, and use weight decay and gradient clipping;

[0079] S27. Repeat steps S21-26 until the maximum number of training iterations is reached or the model converges.

[0080] Specifically, the loss function including the SMAPE term and the specular term is expressed as:

[0081] L = 0.8·L SMAPE +0.2·L spec

[0082] Among them, L SMAPE For the SMAPE term, the specific formula is:

[0083]

[0084] Where x is the reference image, y is the predicted image, N is the total number of pixels in the image, C is the number of channels in the image, and ε is a constant; p,c y represents the predicted color value of the p-th pixel in the c-th channel of the image. p,c This represents the color value of the p-th pixel in the c-th channel of the reference image;

[0085] L spec For the highlight term, the specific formula is:

[0086]

[0087] Where, x c,iThis represents the color value of the i-th pixel in the c-th channel of the reference image. y represents the standard deviation of the color values ​​of all pixels in the neighborhood of the i-th pixel in the c-th channel of the reference image; c,i This represents the predicted color value of the i-th pixel in the c-th channel of the image. This represents the standard deviation of the color values ​​of all pixels in the neighborhood of the i-th pixel in the c-th channel of the predicted image. The neighborhood is 11×11 in size.

[0088] S3. Use the trained kernel prediction model to denoise the high-noise volumetric rendering image with low sampling number to obtain the denoised image.

[0089] Specifically, step S3 includes:

[0090] S31. Use motion vectors to calculate the previous frame image corresponding to the high-noise volume rendering image with low sampling number;

[0091] S32. The high-noise volumetric rendering image is color-blended with its corresponding previous frame image using an exponential moving average formula to obtain a mixed image. This is to increase the sampling rate by utilizing information in the temporal domain. The exponential moving average formula is expressed as follows:

[0092] C′ i =α·C i +(1-α)·C′ i-1

[0093] Among them, C i C′ represents the color value of the i-th pixel in a high-noise volumetric rendering image. i-1 It is the color value of the i-th pixel in the high-noise volumetric rendering image after mixing with the color value of the i-th pixel in the previous frame image. α is the smoothing factor, which we take as 0.2.

[0094] S33. Replace the high-noise volumetric rendering image with a hybrid image and input it into the kernel prediction model, then output the denoised result.

[0095] This is done because in real-time rendering, the movement of objects or cameras is continuous. Therefore, we can assume that there is a lot of overlap between the current frame and the previous frame. So we can use the content of the previous frame to increase the sampling rate. This means mixing the color values ​​sampled in the previous frame with the color values ​​sampled in the current frame, which is equivalent to obtaining an image with a higher sampling rate and lower noise.

[0096] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "setting," "connection," "fixing," "rotation," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal connection of two components or the interaction between two components. Unless otherwise explicitly limited, those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0097] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for denoising Monte Carlo rendering images based on deep learning, characterized in that, Includes the following steps: S1. Render the images using the Monte Carlo volume rendering method to obtain the dataset, and preprocess the dataset to obtain the training dataset; Step S1 renders the image using the Monte Carlo volume rendering method to obtain the dataset, including: S11. Based on the Monte Carlo volume path tracing method, the volume data is rendered at a sampling rate of 1spp to obtain a noisy image. The delta tracking method and the ratio tracking method are used for distance sampling and transmittance calculation, respectively. At the same time, the distance sampling and transmittance calculation are accelerated by the DDA method. S12. Obtain the position of the volume data in step S11 when the first scattering occurs using the delta tracking method, and obtain the world space coordinates, volume scattering rate, position normal, and depth of that position; S13. Normalize the world space coordinates and map the normalized world space coordinates to the range [0,1] to obtain the first type of three-channel color image; then obtain the second type of three-channel color image based on the volume scattering rate; S14. Calculate the gradient of the position normal in the x, y, and z directions using the central difference method, normalize the three gradients, and map the normalized three gradients to the range [0,1] to obtain the third type of three-channel color image. S15. Use a nonlinear transformation method to convert the depth to a nonlinear depth, and output a single-channel grayscale image based on the nonlinear depth; S16. Based on the Monte Carlo volume path tracing method, the volume data in step S11 is rendered at a sampling rate of 4096 spp to obtain a reference image. The delta tracking method and ratio tracking method are used for distance sampling and transmittance calculation, respectively. At the same time, the distance sampling and transmittance calculation are accelerated according to the DDA method. S17. Combine the noisy image, the first type of three-channel color image, the second type of three-channel color image, the third type of three-channel color image, the single-channel gray image, and the reference image into a set of data; S18. Repeat steps S11-S17 to obtain multiple sets of data to form a dataset; S2. Construct a kernel prediction model based on the U-net network and train the kernel prediction model using the training dataset; the prediction error is calculated using a loss function that includes the SMAPE term and the specular term. The loss function including the SMAPE term and the highlight term is expressed as: in, For the SMAPE term, the specific formula is: Where x is the reference image, y is the predicted image, N is the total number of pixels in the image, and C is the number of channels in the image. It is a constant; x p,c y represents the predicted color value of the p-th pixel in the c-th channel of the image. p,c This represents the color value of the p-th pixel in the c-th channel of the reference image; For the highlight term, the specific formula is: in, This represents the color value of the i-th pixel in the c-th channel of the reference image. It represents the standard deviation of the color values ​​of all pixels in the neighborhood of the i-th pixel in the c-th channel of the reference image. This represents the predicted color value of the i-th pixel in the c-th channel of the image. This represents the standard deviation of the color values ​​of all pixels within a neighborhood of the i-th pixel in the c-th channel of the predicted image. The neighborhood is 11×11 pixels in size. S3. Use the trained kernel prediction model to denoise the high-noise volumetric rendering image with low sampling number to obtain the denoised image.

2. The method for denoising Monte Carlo rendering images based on deep learning according to claim 1, characterized in that, Step S15 uses a nonlinear transformation method to convert the depth into a nonlinear depth, expressed as: in, The nonlinear depth is represented by z, which represents the depth of the position where the volume data undergoes the first scattering using the delta tracking method. far represents the far plane depth of the camera, and near represents the near plane depth of the camera.

3. The method for denoising Monte Carlo rendering images based on deep learning according to claim 1, characterized in that, Step S1 preprocesses the dataset to obtain the training dataset, including: For each set of data in the dataset, the noisy image, the first type of three-channel color image, the second type of three-channel color image, the third type of three-channel color image, the single-channel gray image, and the reference image are segmented separately. Each image is divided into 128×128 small blocks, and each small block is randomly flipped; finally, multiple sets of training data are obtained. If the noise image and the reference image in a set of data are output images in the HDR color space, then tonemapping is first used to map the noise image and the reference image back to the LDR color space, and then segmentation processing is performed.

4. The method for denoising Monte Carlo rendering images based on deep learning according to claim 1, characterized in that, Step S2 trains a kernel prediction model based on the U-net network using the training dataset, including: S21. Extract a set of training data, divide the noisy image by the second type of three-channel color image to obtain the irradiance map; S22. Stitch together the irradiance image, the first three-channel color image, the third three-channel color image, and the single-channel grayscale image to obtain a fused image; S23. Input the fused image into the kernel prediction model to obtain a multi-channel feature map, and extract the filter kernel of each channel in the multi-channel feature map to obtain a three-channel filter kernel; S24. Use a three-channel filter kernel to denoise the irradiance image, and multiply the denoised irradiance image with the second three-channel color image to obtain the predicted image; S25. Calculate the prediction error between the predicted image and the reference image using a loss function that includes the SMAPE term and the specular term, and then backpropagate it. S26. Optimize the model parameters using the Adam optimizer, and use weight decay and gradient clipping; S27. Repeat steps S21-26 until the maximum number of training iterations is reached or the model converges.

5. The method for denoising Monte Carlo rendering images based on deep learning according to claim 1, characterized in that, The kernel prediction model comprises four encoders, four decoders, and one 1×1 convolutional layer cascaded in sequence. Each encoder includes two 3x3 convolutional kernels and a 2x downsampling layer. The first three encoders have 32 channels for both input and output, while the last encoder has 32 channels for input and 64 channels for output. Each decoder includes two 3x3 convolutional kernels and a 2x upsampling layer. The first three decoders have 64 channels for both input and output, while the last decoder has 64 channels for input and 32 channels for output. The 1×1 convolutional layer outputs 27 channels.

6. The method for denoising Monte Carlo rendering images based on deep learning according to claim 1, characterized in that, Step S3 uses the trained kernel prediction model to denoise the high-noise volumetric rendering image with a low sampling number, resulting in the following denoised images: S31. Use motion vectors to calculate the previous frame image corresponding to the high-noise volume rendering image with low sampling number; S32. The high-noise volumetric rendering image is color-blended with its corresponding previous frame image using an exponential moving average formula to obtain a blended image; the exponential moving average formula is expressed as: in, This represents the color value of the i-th pixel in a high-noise volumetric rendering image. It is the color value of the i-th pixel in the high-noise volumetric rendering image, which is the result of mixing the color value of the i-th pixel in the previous frame image. It is a smoothing factor; S33. Replace the high-noise volumetric rendering image with a hybrid image and input it into the kernel prediction model, then output the denoised result.

Images