HumanNeRF image noise reduction method, system and device

Abstract

The invention relates to the technical field of image enhancement, in particular to a HumanNeRF image noise reduction method, system and equipment, and aims to construct a positive sample pair by ingeniously utilizing the inherent randomness of a HumanNeRF rendering process and generating two different noise instances for the same pose so as to obtain training data with consistent contents and different noises. Therefore, the problem of dependence on a clean true value is fundamentally solved; meanwhile, by utilizing an image training set containing the HumanNeRF image, the positive sample pair and the negative sample of the HumanNeRF image, the noise reduction model is optimized by fusing training targets of positive and negative sample self-supervision loss, adversarial loss, pixel loss and perception loss until convergence, so that the model can effectively remove noise, not only can pixel-level details and high-level semantic information of the HumanNeRF image be kept, but also the noise reduction effect is improved. Finally, a more accurate and robust noise reduction result conforming to the human body posture characteristics is obtained, and the noise reduction quality of the HumanNeRF image is improved.

Description

Technical Field

[0001] The present invention relates to the field of image enhancement technology, and in particular to a HumanNeRF image denoising method, system and device. Background Art

[0002] To address the challenges of motion blur, artifacts, and detail loss often associated with Neural Radiance Field (NeRF)-generated images, existing methods have proposed the HumanNeRF model, specifically for 3D human reconstruction and rendering. By combining the implicit representation of NeRF with a parametric human model, the HumanNeRF model decouples the complex problem of dynamic human modeling into two subtasks: first, learning the appearance and details of the human body in a static canonical space; second, driving the motion and deformation of points in the canonical space using pose parameters. This human-specific design significantly improves the stability and realism of dynamic human modeling.

[0003] While the HumanNeRF model has achieved remarkable success in capturing macroscopic human motion, the quality of its rendered images at the microscopic level still needs improvement. In practical applications, high-frequency noise is prevalent in generated images due to sparse input views, poor lighting conditions, or the inherent random sampling characteristics of the rendering process. This noise is not simply random, but rather structured noise closely related to 3D geometry and motion models. It severely corrupts key details such as skin texture and clothing wrinkles, affecting not only visual quality but also the accuracy of downstream tasks such as pose estimation.

[0004] Currently, denoising images generated by the HumanNeRF model requires a large amount of paired "clean-noise" training data. However, in the HumanNeRF model's scenes, since the actual scene geometry and lighting are unknown, it is impossible to obtain or generate a flawless, absolutely "clean" ground truth image as a supervisory signal. Furthermore, while some existing methods address the problem of no ground truth, their fundamental assumption is that noise is conditionally independent across pixels and has zero expectation. However, the noise generated by the HumanNeRF model is structured, strongly correlated with the scene's 3D geometry, surface material, and imperfect motion models, and is not spatially independently distributed. Directly applying these existing general-purpose methods to HumanNeRF image denoising struggles to effectively model the complex structured noise. The result is often incomplete denoising or over-smoothing while removing noise, destroying valuable high-frequency details in the image, such as skin texture and clothing wrinkles, and easily introducing new artifacts. Summary of the Invention

[0005] The purpose of the present invention is to provide a HumanNeRF image denoising method, system and device.

[0006] The technical solutions of the present invention are as follows: A HumanNeRF image denoising method includes the following operations: S1. Several human posture images are processed by the HumanNeRF model to obtain several HumanNeRF images; Each HumanNeRF image is rendered twice with different noises to obtain two positive samples with different noises, forming a positive sample pair for each HumanNeRF image. From all HumanNeRF images, an image with a different human posture from the current HumanNeRF image is randomly selected as a negative sample of the current HumanNeRF image. All HumanNeRF images, corresponding positive sample pairs and negative samples, form the image training set; S2. Use the image training set to train the denoising model until the training loss is less than the loss threshold and convergence is achieved, thereby obtaining the trained denoising model. The training loss value is obtained based on the self-supervised loss of positive sample pairs, the self-supervised loss of negative samples, the adversarial loss of HumanNeRF images, the pixel loss of HumanNeRF images, and the perceptual loss of HumanNeRF images; S3. The image to be processed is processed by the trained denoising model to obtain a denoised image.

[0007] In the self-supervised loss of positive sample pairs, the self-supervised loss of a positive sample pair is calculated by the following formula: , is the self-supervised loss for a positive sample pair, is the positive sample in the positive sample pair i With positive samples j The cosine similarity of Positive samples i、 Positive samples j The feature map is a positive sample i、 Positive samples j are processed by the generator respectively, is the temperature hyperparameter, and N is the number of positive sample pairs.

[0008] The pixel loss calculation formula for HumanNeRF images is as follows: , is the pixel loss of HumanNeRF image, is the feature map of the HumanNeRF image, which is obtained by processing the HumanNeRF image by the generator. is the expected function, Processed as Paradigm 1.

[0009] The perceptual loss calculation formula for HumanNeRF images is as follows: , is the perceptual loss of HumanNeRF image, is the feature map of the HumanNeRF image, which is obtained by processing the HumanNeRF image by the generator. For deep feature extraction processing, Processed as Paradigm 2.

[0010] The operation of the generator is as follows: the input is processed by convolution and maximum pooling layers to obtain a first-scale hierarchical feature map; the first-scale hierarchical feature map is processed by convolution and maximum pooling layers to obtain a second-scale hierarchical feature map; the second-scale hierarchical feature map is processed by convolution and maximum pooling layers to obtain a third-scale hierarchical feature map; the third-scale hierarchical feature map is upsampled and convolved, and then fused with the third-scale hierarchical feature map to obtain a third-resolution restoration map; the third-resolution restoration map is upsampled and convolved, and then fused with the second-scale hierarchical feature map to obtain a second-resolution restoration map; the second-resolution restoration map is upsampled and convolved, and then fused with the first-scale hierarchical feature map to obtain the output as a feature map; the input is a HumanNeRF image or a positive sample.

[0011] The self-supervised loss of a positive sample pair is the average of the self-supervised losses of all positive samples, or the maximum of the self-supervised losses of all positive samples.

[0012] A HumanNeRF image denoising system, for implementing the above-mentioned HumanNeRF image denoising method, comprising: The image training set generation module is used to process several human posture images through the HumanNeRF model to obtain several HumanNeRF images. Each HumanNeRF image is rendered twice with different noises to obtain two positive samples with different noises, forming a positive sample pair for each HumanNeRF image. From all HumanNeRF images, an image with a different human posture from the current HumanNeRF image is randomly selected as the negative sample of the current HumanNeRF image. All HumanNeRF images, corresponding positive sample pairs and negative samples form the image training set. The training denoising model generation module is used to train the denoising model using the image training set until the training loss is less than the loss threshold and convergence is achieved, thereby obtaining the trained denoising model; the training loss value is obtained based on the self-supervised loss of positive sample pairs, the self-supervised loss of negative samples, the adversarial loss of HumanNeRF images, the pixel loss of HumanNeRF images, and the perceptual loss of HumanNeRF images; The denoising image generation module is used to process the image to be processed and obtain the denoised image after being processed by the trained denoising model.

[0013] A HumanNeRF image denoising device comprises a processor and a memory, wherein the processor implements the above-mentioned HumanNeRF image denoising method when executing a computer program stored in the memory.

[0014] A computer-readable storage medium is used to store a computer program, wherein the computer program implements the above-mentioned HumanNeRF image denoising method when executed by a processor.

[0015] The beneficial effects of the present invention are: The present invention provides a HumanNeRF image denoising method, which cleverly utilizes the inherent randomness of the HumanNeRF rendering process to generate two different noise instances for the same posture to construct positive sample pairs, thereby obtaining training data with "consistent content and different noise". This enables the denoising model to be effectively trained on "noise-noise" data pairs, thereby fundamentally solving the problem of dependence on clean true values, greatly broadening the scope of application of the method, and reducing the threshold and cost of data preparation. At the same time, using an image training set containing HumanNeRF images, their positive sample pairs and negative samples, the denoising model is optimized by integrating the training objectives of positive and negative sample self-supervision loss, adversarial loss, pixel loss and perceptual loss until convergence. This enables the model to effectively remove noise while maintaining the pixel-level details and high-level semantic information of the HumanNeRF image, and enhances the capture of human posture features through posture contrast learning, ultimately obtaining a more accurate, robust and human posture-compliant denoising result, thereby improving the denoising quality of the HumanNeRF image. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] By reading the detailed description of the preferred embodiment below, the solutions and advantages of the present application will become clear to those skilled in the art. The accompanying drawings are only for illustrating the preferred embodiment and are not to be considered as limiting the present invention.

[0017] In the attached figure: Figure 11 and 2 are the results of denoising the HumanNeRF image of scene 1 by the method of this embodiment and the existing method, respectively; (a) is the HumanNeRF image of scene 1, and (b) is the denoising result of the method of this embodiment and the existing method; Figure 2 1 and 2 are the results of denoising the HumanNeRF image of scene 2 by the method of this embodiment and the existing method, respectively; (a) is the HumanNeRF image of scene 2, and (b) is the denoising result of the method of this embodiment and the existing method; Figure 3 1 and 2 are the result images obtained by denoising the HumanNeRF image of scene 3 by the method of this embodiment and the existing method respectively; wherein (a) is the HumanNeRF image of scene 3, and (b) is the denoising result image of the method of this embodiment and the existing method. DETAILED DESCRIPTION

[0018] This embodiment provides a HumanNeRF image denoising method, including the following operations: S1. Several human posture images are processed by the HumanNeRF model to obtain several HumanNeRF images; Each HumanNeRF image is rendered twice with different noises to obtain two positive samples with different noises, forming a positive sample pair for each HumanNeRF image. From all HumanNeRF images, an image with a different human posture from the current HumanNeRF image is randomly selected as a negative sample of the current HumanNeRF image. All HumanNeRF images, corresponding positive sample pairs and negative samples, form the image training set; S2. Use the image training set to train the denoising model until the training loss is less than the loss threshold and convergence is achieved, thereby obtaining the trained denoising model. The training loss value is obtained based on the self-supervised loss of positive sample pairs, the self-supervised loss of negative samples, the adversarial loss of HumanNeRF images, the pixel loss of HumanNeRF images, and the perceptual loss of HumanNeRF images; S3. The image to be processed is processed by the trained denoising model to obtain a denoised image.

[0019] The specific operation steps are detailed as follows.

[0020] S1. Several human body posture images are processed by the HumanNeRF model to obtain several HumanNeRF images; each HumanNeRF image is rendered twice with different noises to obtain two positive samples with different noises, forming a positive sample pair for each HumanNeRF image; from all HumanNeRF images, an image with a different human body posture from the current HumanNeRF image is randomly selected as the negative sample of the current HumanNeRF image; all HumanNeRF images, corresponding positive sample pairs and negative samples form an image training set.

[0021] By cleverly utilizing the inherent randomness of the HumanNeRF rendering process, two different noise instances are generated for the same pose to construct positive sample pairs, obtaining training data with "consistent content and different noise". This allows the model to be effectively trained on "noise-noise" data pairs, fundamentally solving the problem of dependence on clean true values, greatly broadening the scope of application of the method, and lowering the threshold and cost of data preparation.

[0022] First, several (large) human posture images are processed by the HumanNeRF model to obtain several HumanNeRF images containing structured noise.

[0023] Then, by utilizing the randomness in the HumanNeRF rendering process (for example, random jitter along the ray sampling), for a given camera pose and human body pose, each HumanNeRF image is rendered twice with different noises to obtain two positive samples with different noises, forming a positive sample pair for each HumanNeRF image. This provides an excellent opportunity for the subsequent denoising model to learn to separate "content" and "noise", because the "content" of the two images is constant, while the "noise" is variable.

[0024] Next, from all HumanNeRF images, an image with a different human posture from the current HumanNeRF image is randomly selected as the negative sample of the current HumanNeRF image, and the negative samples corresponding to all HumanNeRF images are obtained by analogy.

[0025] Finally, all HumanNeRF images, corresponding positive sample pairs and negative samples form the image training set.

[0026] S2. Use the image training set to train the denoising model until the training loss is less than the loss threshold and convergence is achieved, thereby obtaining the trained denoising model; the training loss value is based on the self-supervised loss of positive sample pairs, the self-supervised loss of negative samples, the adversarial loss of HumanNeRF images, the pixel loss of HumanNeRF images, and the perceptual loss of HumanNeRF images.

[0027] Using an image training set containing HumanNeRF images, their positive sample pairs, and negative samples, the denoising model is optimized until convergence by fusing training objectives of positive and negative sample self-supervision loss, adversarial loss, pixel loss, and perceptual loss. This enables the model to effectively remove noise while maintaining the pixel-level details and high-level semantic information of the HumanNeRF images, while enhancing the capture of human posture features through posture contrast learning, ultimately obtaining a more accurate, robust denoising result that conforms to the characteristics of human posture.

[0028] Among them, the denoising model is formed by a generator and a discriminator. During the processing of the denoising model, the input (HumanNeRF image, or positive sample, or negative sample) is processed by the generator for feature extraction to obtain a pseudo image; the discriminator receives the pseudo image and the corresponding input respectively, outputs the true or false judgment, and calculates the training loss value (including the self-supervised loss of the positive sample pair, the self-supervised loss of the negative sample, the adversarial loss of the HumanNeRF image, the pixel loss of the HumanNeRF image, and the perceptual loss of the HumanNeRF image); the discriminator is fixed, and the generator parameters are updated through back propagation of the total loss; the generator is fixed, and the discriminator parameters are updated through back propagation of the discriminator loss; the above steps are repeated until the training loss value is less than the loss threshold, the training loss converges, and the training denoising model is obtained.

[0029] The generator operates as follows: the input (HumanNeRF image, or positive or negative sample) is processed through convolution and max pooling layers to obtain a first-scale hierarchical feature map; the first-scale hierarchical feature map is processed through convolution and max pooling layers to obtain a second-scale hierarchical feature map; the second-scale hierarchical feature map is processed through convolution and max pooling layers to obtain a third-scale hierarchical feature map; the third-scale hierarchical feature map is upsampled and convolved, and then fused with the third-scale hierarchical feature map to obtain a third-resolution restored image; the third-resolution restored image is upsampled and convolved, and then fused with the second-scale hierarchical feature map to obtain a second-resolution restored image; the second-resolution restored image is upsampled and convolved, and then fused with the first-scale hierarchical feature map to obtain an output feature map (pseudo-image). The number of convolution and max pooling layers, upsampling, and convolution operations in the generator is not limited to three and can be limited based on actual needs.

[0030] The third-scale hierarchical feature map, and / or the third resolution recovery map, and / or the second resolution recovery map, are subjected to attention processing after the upsampling and convolution processing, and the obtained third attention feature map, and / or the second attention feature map, and / or the first attention feature map are respectively fused with the third-scale hierarchical feature map, and / or the second-scale hierarchical feature map, and / or the first-scale hierarchical feature map.

[0031] The specific operation of attention processing is: the convolution feature map obtained after upsampling and convolution processing is processed by channel attention processing and spatial attention processing in turn to obtain the attention feature map.

[0032] The generator in this embodiment first gradually reduces the input resolution through a series of convolution and maximum pooling layers to extract multi-scale hierarchical features, and then gradually restores the image resolution through a series of upsampling and convolution layers, and fuses it with the original scale feature map to achieve clear reconstruction of edges and textures.

[0033] The discriminator includes operations such as convolution and downsampling. The function of the discriminator can also be implemented through the PatchGAN network.

[0034] The above training loss is obtained by the following calculation formula: , is the (total) training loss, 、 、 、 、 They are the self-supervised loss of positive sample pairs, the self-supervised loss of negative samples, the adversarial loss of HumanNeRF images, the pixel loss of HumanNeRF images, and the perceptual loss of HumanNeRF images. 、 、 、 、 They are the first weight, the second weight, the third weight, the fourth weight, and the fifth weight respectively.

[0035] The self-supervised loss of a positive pair is the average of the self-supervised losses of all positive samples, or the maximum of the self-supervised losses of all positive samples. The self-supervised loss of a positive pair is calculated using the following formula: , is the self-supervised loss for a positive sample pair, is the positive sample in the positive sample pair i With positive samples jThe cosine similarity of Positive samples i、 Positive samples j The feature map (pseudo image) is a positive sample i、 Positive samples j are processed by the generator respectively, is the temperature hyperparameter, and N is the number of positive sample pairs.

[0036] The self-supervised loss of negative samples is the average of the self-supervised losses of all negative samples, which is calculated by the following formula: , is the self-supervised loss of a negative sample, HumanNeRF image x and the corresponding positive sample k (positive sample k is the positive sample in the positive sample pair i or positive sample j ), HumanNeRF images and corresponding negative samples m The cosine similarity of 、 、 HumanNeRF images and positive samples respectively k、 Negative samples m The feature map (pseudo image) is HumanNeRF image, positive sample k、 Negative samples m are processed by the generator respectively, and M is the total number of negative samples.

[0037] The adversarial loss of HumanNeRF images can prompt the generator in the denoising model to create image patches that are sufficiently "real" in terms of local texture and details. This directly drives the recovery and enhancement of high-frequency details and avoids over-smoothing. The calculation formula is as follows: , is the adversarial loss for HumanNeRF images, is the output of the HumanNeRF image processed by the discriminator, The pseudo image corresponding to the HumanNeRF image is the output obtained by the discriminator. The pseudo image corresponding to the HumanNeRF image is obtained based on the HumanNeRF image processed by the generator. X is the total number of HumanNeRF images, is the expected function.

[0038] The pixel loss calculation formula for HumanNeRF images is as follows: , is the pixel loss of HumanNeRF image, is the feature map of HumanNeRF image (pseudo image), Processed as Paradigm 1.

[0039] The perceptual loss of HumanNeRF images ensures that the generated images are consistent with the original images in terms of high-level semantics and texture style by constraining them in the deep feature space. This is crucial for preserving complex textures (such as clothing materials). The calculation formula is as follows: , is the perceptual loss of HumanNeRF image, For deep feature extraction processing, it can be achieved through the VGG19 network. Processed as Paradigm 2.

[0040] S3. The image to be processed (which is a HumanNeRF image) is processed by the trained denoising model to obtain a denoised image.

[0041] This embodiment further provides a HumanNeRF image denoising device, which is used to implement the above-mentioned HumanNeRF image denoising method, including: The image training set generation module is used to process several human posture images through the HumanNeRF model to obtain several HumanNeRF images. Each HumanNeRF image is rendered twice with different noises to obtain two positive samples with different noises, forming a positive sample pair for each HumanNeRF image. From all HumanNeRF images, an image with a different human posture from the current HumanNeRF image is randomly selected as the negative sample of the current HumanNeRF image. All HumanNeRF images, corresponding positive sample pairs and negative samples form the image training set. The training denoising model generation module is used to train the denoising model using the image training set until the training loss is less than the loss threshold and convergence is achieved, thereby obtaining the trained denoising model; the training loss value is obtained based on the self-supervised loss of positive sample pairs, the self-supervised loss of negative samples, the adversarial loss of HumanNeRF images, the pixel loss of HumanNeRF images, and the perceptual loss of HumanNeRF images; The denoising image generation module is used to process the image to be processed and obtain the denoised image after being processed by the trained denoising model.

[0042] This embodiment further provides a HumanNeRF image denoising device, comprising a processor and a memory, wherein the processor implements the above-mentioned HumanNeRF image denoising method when executing a computer program stored in the memory.

[0043] This embodiment further provides a computer-readable storage medium for storing a computer program, wherein the computer program implements the above-mentioned HumanNeRF image denoising method when executed by a processor.

[0044] This embodiment provides a HumanNeRF image denoising method that cleverly utilizes the inherent randomness of the HumanNeRF rendering process to generate two different noise instances for the same pose to construct positive sample pairs, thereby obtaining training data with "consistent content and different noise". This enables the denoising model to be effectively trained on "noise-noise" data pairs, thereby fundamentally solving the problem of dependence on clean ground truth values, greatly broadening the scope of application of the method, and reducing the threshold and cost of data preparation. At the same time, using an image training set containing HumanNeRF images, their positive sample pairs, and negative samples, the denoising model is optimized until convergence by integrating training objectives of positive and negative sample self-supervision loss, adversarial loss, pixel loss, and perceptual loss. This enables the model to effectively remove noise while maintaining the pixel-level details and high-level semantic information of the HumanNeRF images, and enhances the capture of human posture features through posture contrast learning. Ultimately, a more accurate, robust denoising result that conforms to human posture characteristics is obtained, thereby improving the denoising quality of the HumanNeRF images.

[0045] To verify the denoising advantage of the method of this embodiment, the method of this embodiment is compared with existing advanced denoising methods (including the FMGFI method, the MDBMF method, the CAFFM method, and the CNNT method). That is, the method of this embodiment and the existing advanced denoising methods are used to perform denoising on the HumanNeRF image of scene 1 (specifically, the image within the red border in (a) of the HumanNeRF image of scene 1), the HumanNeRF image of scene 2 (specifically, the image within the red border in (a) of the HumanNeRF image of scene 2), and the HumanNeRF image of scene 3 (specifically, the image within the red border in (a) of the HumanNeRF image of scene 3). The results are shown in FIG. Figure 1 、 Figure 2 、 Figure 3 It can be seen that compared with the existing methods (FMGFI method, MDBMF method, CAFFM method, CNNT method), the method of this embodiment has better denoising effect and higher detail restoration, and can significantly improve the denoising quality of HumanNeRF images.

Claims

1. A HumanNeRF image denoising method, characterized in that: The following operations are included: S1. Several human posture images are processed by the HumanNeRF model to obtain several HumanNeRF images; Each HumanNeRF image is rendered twice with different noises to obtain two positive samples with different noises, forming a positive sample pair for each HumanNeRF image. From all HumanNeRF images, an image with a different human posture from the current HumanNeRF image is randomly selected as a negative sample of the current HumanNeRF image. All HumanNeRF images, corresponding positive sample pairs and negative samples, form the image training set; S2. Use the image training set to train the denoising model until the training loss is less than the loss threshold and convergence is achieved, thereby obtaining the trained denoising model. The training loss value is obtained based on the self-supervised loss of positive sample pairs, the self-supervised loss of negative samples, the adversarial loss of HumanNeRF images, the pixel loss of HumanNeRF images, and the perceptual loss of HumanNeRF images; S3. The image to be processed is processed by the trained denoising model to obtain a denoised image.

2. The HumanNeRF image denoising method according to claim 1, wherein: In the self-supervised loss of positive sample pairs, the self-supervised loss of a positive sample pair is calculated by the following formula: ， is the self-supervised loss for a positive sample pair, is the positive sample in the positive sample pair i With positive samples j The cosine similarity of Positive samples i、 Positive samples j The feature map is a positive sample i、 Positive samples j are processed by the generator respectively, is the temperature hyperparameter, and N is the number of positive sample pairs.

3. The HumanNeRF image denoising method according to claim 1, wherein: The pixel loss calculation formula for HumanNeRF images is as follows: ， is the pixel loss of HumanNeRF image, is the feature map of the HumanNeRF image, which is obtained by processing the HumanNeRF image by the generator. is the expected function, Processed as Paradigm 1.

4. The HumanNeRF image denoising method according to claim 1, wherein: The perceptual loss calculation formula for HumanNeRF images is as follows: ， is the perceptual loss of HumanNeRF image, is the feature map of the HumanNeRF image, which is obtained by processing the HumanNeRF image by the generator. For deep feature extraction processing, Processed as Paradigm 2.

5. The HumanNeRF image denoising method according to claim 2, 3 or 4, wherein: The generator's operations are: The input is processed by convolution and maximum pooling layers to obtain the first-scale hierarchical feature map; The first-scale hierarchical feature map is processed by convolution and maximum pooling layers to obtain the second-scale hierarchical feature map; The second-scale hierarchical feature map is processed by convolution and maximum pooling layers to obtain the third-scale hierarchical feature map; The third-scale hierarchical feature map is upsampled and convolved, and then fused with the third-scale hierarchical feature map to obtain a third-resolution restoration map; The third-resolution restored image is upsampled and convolved, and then fused with the second-scale hierarchical feature map to obtain a second-resolution restored image. The second-resolution restored image is upsampled and convolved, and then fused with the first-scale hierarchical feature map to obtain an output as a feature map; The input is a HumanNeRF image or a positive sample.

6. The HumanNeRF image denoising method according to claim 1, wherein: The self-supervised loss of a positive sample pair is the average of the self-supervised losses of all positive samples, or the maximum of the self-supervised losses of all positive samples.

7. A HumanNeRF image denoising system, used to implement the HumanNeRF image denoising method according to claim 1, characterized in that: include: The image training set generation module is used to process several human posture images through the HumanNeRF model to obtain several HumanNeRF images. Each HumanNeRF image is rendered twice with different noises to obtain two positive samples with different noises, forming a positive sample pair for each HumanNeRF image. From all HumanNeRF images, an image with a different human posture from the current HumanNeRF image is randomly selected as the negative sample of the current HumanNeRF image. All HumanNeRF images, corresponding positive sample pairs and negative samples form the image training set. The training denoising model generation module is used to train the denoising model using the image training set until the training loss is less than the loss threshold and convergence is achieved, thereby obtaining the trained denoising model; the training loss value is obtained based on the self-supervised loss of positive sample pairs, the self-supervised loss of negative samples, the adversarial loss of HumanNeRF images, the pixel loss of HumanNeRF images, and the perceptual loss of HumanNeRF images; The denoising image generation module is used to process the image to be processed and obtain the denoised image after being processed by the trained denoising model.

8. A HumanNeRF image denoising device, characterized in that: The method comprises a processor and a memory, wherein the processor implements the HumanNeRF image denoising method according to any one of claims 1 to 6 when executing the computer program stored in the memory.

9. A computer-readable storage medium, characterized in that Used to store a computer program, wherein when the computer program is executed by a processor, the HumanNeRF image denoising method according to any one of claims 1 to 6 is implemented.

Images