An uncertainty-guided dual-modal separable image super-resolution method

By combining a conditional variational encoder network and a depth detail estimation network with an uncertainty-guided fusion module, the problems of modality separation and discontinuous regions in low-resolution depth image reconstruction are solved, achieving high-quality super-resolution depth map reconstruction and improving reconstruction results and interpretability.

CN116188273BActive Publication Date: 2026-11-10BEIHANG UNIV
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202310261226.6
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2023-03-17
Publication Date
2026-11-10
Estimated Expiration
2043-03-17

AI Technical Summary

Technical Problem

Existing technologies struggle to effectively improve the quality of low-resolution depth images, especially given the discontinuous regions and noise interference during modal separation and reconstruction. Furthermore, traditional methods cannot effectively utilize color images to guide depth map reconstruction.

Method used

An uncertainty-guided bimodal separable image super-resolution method is adopted. By combining a conditional variational encoder network and a depth detail estimation network with an uncertainty-guided fusion module, feature enhancement is performed using the features and uncertainty estimation maps of color and depth images to achieve modality separation and high-quality reconstruction.

Benefits of technology

It achieves high-quality super-resolution depth maps even without a guiding modality. The cross-task learning scheme encourages depth maps to learn discontinuities and uses uncertainty to guide the fusion network, improving reconstruction results and interpretability.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN116188273B_ABST
    Figure CN116188273B_ABST
Patent Text Reader

Abstract

This invention discloses an uncertainty-guided bimodal separable image super-resolution method, comprising: obtaining a low-resolution depth image and a corresponding color image; performing cubic interpolation upsampling on the low-resolution depth image to obtain a sampled image; obtaining random features of the low-resolution depth image based on the sampled image; inputting the low-resolution depth image into a depth encoder to obtain deep features of the low-resolution depth image; obtaining a first feature map and a corresponding first uncertainty estimation map based on the random features and deep features; inputting the color image and the sampled image into a depth detail estimation network to output a second feature map and a corresponding second uncertainty estimation map; performing feature enhancement processing on the first feature map and the second feature map; and performing 3×3 convolution processing on the feature-enhanced first depth feature map and the second depth feature map to obtain a super-resolution depth map. This method can obtain high-quality super-resolution depth maps.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of digital image processing technology, and in particular to an uncertainty-oriented dual-modal separable image super-resolution method. Background Technology

[0002] Depth information provides crucial scene details and is widely used in computer vision fields such as 3D reconstruction, object detection, and instance segmentation. However, due to technological limitations, obtaining high-quality depth maps is difficult, which restricts the implementation of many challenging vision tasks. Depth super-resolution technology can effectively solve this problem, and its implementation cost is low. Therefore, how to effectively improve the quality of low-resolution depth images using depth super-resolution technology is a very important research topic.

[0003] Compared to color images, depth images often exhibit abrupt discontinuities due to occlusion or positional differences of actual objects. Furthermore, depth maps frequently suffer from large-area artifacts due to technological limitations or external interference. Super-resolution methods focused on reconstructing details from color images are often unsuitable for depth image reconstruction. Therefore, reconstructing depth maps using a single modality is extremely difficult, while color images, typically possessing clearer textures, can effectively guide high-quality depth map reconstruction. Previously, three methods existed for effectively fusing depth and color information: feature fusion at the input, reconstruction, and output stages. Fusion at the input stage prevents color images from effectively guiding the depth map reconstruction process; information fusion at the reconstruction stage leads to high information coupling; and feature fusion at the output stage ensures both effective guidance between features and the separability of information.

[0004] When performing super-resolution reconstruction on depth maps containing discontinuous regions, traditional algorithms such as bicubic interpolation cannot distinguish these discontinuous regions during the reconstruction process. They are also significantly affected by noise interference, resulting in poor reconstruction quality. Furthermore, traditional interpolation algorithms are designed based on the pixel positions in the image, which often fails to reflect the true state of the image, leading to unreliable calculation results.

[0005] Most deep learning-based depth map super-resolution reconstruction methods rely on single-point prediction, making it difficult to distinguish between holes and correct depth values. Furthermore, previous deep learning-guided color image-based depth image super-resolution reconstruction techniques largely fused information during the reconstruction phase. These methods are overly dependent on color images aligned with the depth map, and in practice, these two modalities are often indistinguishable, making them unsuitable for real-world production needs. Moreover, these depth super-resolution reconstruction methods only consider the model's reconstruction performance, neglecting the interpretability of the reconstruction results.

[0006] Therefore, how to achieve modal separation in the process of improving the quality of low-resolution depth images has become a key issue in current research. Summary of the Invention

[0007] In view of the above problems, the present invention provides an uncertainty-guided bimodal separable image super-resolution method that at least solves some of the above technical problems, and can obtain high-quality super-resolution depth maps.

[0008] This invention provides an uncertainty-guided dual-modal separable image super-resolution method, comprising:

[0009] Obtain a low-resolution depth image, and obtain the color image corresponding to the low-resolution depth image;

[0010] The low-resolution depth image is upsampled three times by interpolation to obtain a sampled image;

[0011] Based on the sampled image, obtain the random features of the low-resolution depth image;

[0012] The low-resolution depth image is input into a depth encoder, and deep features of the low-resolution depth image are obtained through multiple residual modules;

[0013] Based on the randomness characteristics and the deep characteristics, a first feature map and a corresponding first uncertainty estimation map are obtained;

[0014] The color image and the sampled image are input into the deep detail estimation network, which outputs a second feature map and a corresponding second uncertainty estimation map.

[0015] Based on the first uncertainty estimation map, feature enhancement processing is performed on the first feature map;

[0016] Based on the second uncertainty estimation map, feature enhancement processing is performed on the second feature map;

[0017] The first and second depth feature maps after feature enhancement are processed by 3×3 convolution to obtain a super-resolution depth map.

[0018] Furthermore, the size of the sampled image is consistent with the size of the super-resolution depth map.

[0019] Further, obtaining the random features of the low-resolution depth image based on the sampled image specifically includes:

[0020] The sampled image is modulated into a Gaussian distribution prior mean and prior variance by multiple 3×3 convolutions; the prior mean and prior variance are processed by reparameterized sampling to obtain the randomness features of the low-resolution depth image.

[0021] Further, obtaining the first deep feature map and the corresponding first uncertainty estimation map based on the randomness feature and the deep feature specifically includes:

[0022] The randomness feature and the deep feature are concatenated;

[0023] The concatenated features are fused using a 1×1 convolution to obtain the first fused feature;

[0024] The first fused feature is enhanced by multiple residual modules to obtain the first feature map;

[0025] The first feature map is processed by a 3×3 convolution to obtain the first depth map corresponding to the first feature map;

[0026] The first feature map is further processed using a 3×3 convolution to obtain the first uncertainty estimation map corresponding to the first feature map.

[0027] Further, the step of inputting the color image and the sampled image into the depth detail estimation network to output a second depth feature map and a corresponding second uncertainty estimation map specifically includes:

[0028] The color image and the sampled image are concatenated and used as input to the depth detail estimation network;

[0029] In the deep detail estimation network, a 1×1 convolution is used to fuse the concatenated features to obtain a second fused feature;

[0030] The second fused feature is enhanced by multiple residual modules to obtain a second feature map;

[0031] The second feature map is processed by a 3×3 convolution to obtain the second depth map corresponding to the second feature map;

[0032] The second feature map is then processed using a 3×3 convolution to obtain the second uncertainty estimation map corresponding to the second feature map.

[0033] Furthermore, it also includes:

[0034] The color image is processed using a Laplacian filter to obtain a depth texture region;

[0035] The deep texture region is subjected to feature imposition processing using a texture loss function.

[0036] Furthermore, the texture loss function is expressed as:

[0037] L=||(y te -y)+(y te -y)·t||1

[0038] Among them, y te y represents the second depth map; y represents the high-resolution depth map; and t represents the depth texture region to be enhanced.

[0039] Furthermore, it also includes:

[0040] The super-resolution depth map is trained and optimized using a loss function;

[0041] The loss function is expressed as:

[0042]

[0043] in, y represents the super-resolution depth map; u represents the uncertainty estimation map. The high-resolution depth map y refers to the reference image used, i.e., the ground truth. Since it is supervised training, it has labels during the training process.

[0044] Furthermore, it also includes:

[0045] The sampled image is concatenated with the final super-resolution depth map;

[0046] The concatenated image is modulated into a Gaussian distribution posterior mean and posterior variance by multiple 3×3 convolutions.

[0047] The prior mean, prior variance, posterior mean, and posterior variance are constrained by KL divergence.

[0048] Compared with existing technologies, the uncertainty-guided dual-modal separable image super-resolution method described in this invention has the following beneficial effects:

[0049] 1. The depth super-resolution architecture proposed in this invention can use guided modes as input for extraction, and also supports inference without guided modes, to obtain high-quality super-resolution depth maps.

[0050] 2. This invention proposes a cross-task learning scheme that encourages deep detail estimation networks to learn the discontinuities in depth maps and uses uncertainty-guided fusion networks to fuse super-resolution depth results.

[0051] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description, claims, and drawings.

[0052] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0053] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0054] Figure 1 This is a schematic diagram of the flowchart of the uncertainty-guided dual-modal separable image super-resolution method provided in the embodiments of the present invention.

[0055] Figure 2 A schematic diagram illustrating uncertainty-guided feature enhancement provided for embodiments of the present invention.

[0056] Figure 3 This is a schematic diagram illustrating the comparison of visual results on the NYUv2 dataset provided in an embodiment of the present invention.

[0057] Figure 4 This is a schematic diagram illustrating the comparison of visual results on the RGBDD dataset provided in an embodiment of the present invention.

[0058] Figure 5 This is a schematic diagram of the visualization results of the uncertainty graph generated during the learning process of the conditional variational editor network and the deep detail estimation network provided in the embodiments of the present invention. Detailed Implementation

[0059] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0060] See Figure 1 As shown, this embodiment of the invention provides an uncertainty-guided bimodal separable image super-resolution method. The following sections will describe the method in detail from three parts: the Conditional Variational Encoder Network (CVAENet), the Depth Detail Estimation Network and the Uncertainty-Guided Fusion Module, and the Training Optimization.

[0061] I. Conditional Variational Encoder Network (CVAENet)

[0062] The Conditional Variational Encoder Network (CVAENet) consists of a probabilistic network, a deep editor, and a prediction network; the specific operations are as follows:

[0063] Obtain a low-resolution depth image and its corresponding color image; perform cubic interpolation upsampling on the low-resolution depth image to obtain a sampled image; the size of this sampled image is compared with that of the super-resolution depth image. The dimensions are consistent;

[0064] The sampled image is used as input to a probabilistic network. In this network, multiple 3×3 convolutions (four in this embodiment) modulate the input sampled image into a Gaussian-distributed prior mean and prior variance. The prior mean and prior variance are processed through reparameterized sampling to obtain the randomness features of the low-resolution depth image. This is expressed by the following formula:

[0065]

[0066] Where μ represents the prior mean; σ represents the prior variance; the latent variable z contains the randomness characteristics of the image; σ∈R 64×h×w h and w represent the height and width of the sampled image, respectively.

[0067] Since the deep features of low-resolution images also play a very important role in the modeling of super-resolution depth images, a depth encoder is designed in this embodiment of the invention to extract the deep features of low-resolution depth images. Specifically, the low-resolution depth image is input into the depth encoder, and the deep features of the low-resolution depth image are obtained through multiple residual modules (16 in this embodiment of the invention). Through the superposition processing of the 16 residual modules, more complex features can be gradually extracted from the image.

[0068] Based on randomness features and deep features, a first deep feature map and a corresponding first uncertainty estimation map are obtained: the randomness features and deep features are concatenated along their feature dimensions as input to the prediction network. The purpose of concatenation is to fuse these two different types of features in subsequent processing. In the prediction network, a 1×1 convolution is used to fuse the concatenated features, making the randomness features and deep features have the same dimension, thereby obtaining the first fused feature. This step avoids learning only from salient features. Since the fused features still cannot accurately guide the generation of a super-resolution depth map, multiple residual modules (three in this embodiment) are used in the prediction network to enhance the first fused feature, obtaining the first feature map F. dThe first feature map is processed using a 3×3 convolution to obtain the first depth map y corresponding to the first feature map. d Simultaneously, an additional 3×3 convolution is applied to process the first feature map to obtain the first uncertainty estimation map u corresponding to the first feature map. d .

[0069] II. Deep Detail Estimation Network and Uncertainty-Guided Fusion Module

[0070] Since color images can provide a wealth of details for depth maps, this embodiment of the invention designs a depth detail estimation network to fuse color image features and depth image features to guide depth map reconstruction. Specifically, a high-resolution color image consistent with the initial low-resolution depth image scene is acquired; the concatenation of the color image and the sampled image obtained after the three-stage interpolation upsampling is used as the input to the depth detail estimation network. In the network, a 1×1 convolution is used to fuse the concatenated image features to obtain a second fused feature. Multiple residual modules (five in this embodiment) are then used to enhance the second fused feature, converting it into a second feature map F rich in texture estimation. te Simultaneously, an additional 3×3 convolution is added to the last layer of the depth detail estimation network. This 3×3 convolution processes the second feature map to obtain the second depth map yte corresponding to the second feature map. Another 3×3 convolution is used to process the second feature map to obtain the first uncertainty estimation map u corresponding to the second feature map. te .

[0071] Since the Conditional Variational Encoder Network (CVAENet) and the Depth Detail Estimation Network possess different types of depth information, this embodiment of the invention also incorporates an uncertainty-guided fusion module to fuse the two reconstructed images. This module provides additional regularization to the two existing networks, helping them to reconstruct super-resolution depth images more effectively. Furthermore, due to the significant inconsistencies in texture details between RGB and depth images, color images exhibit smoother texture transitions compared to depth images when occlusion or large distance differences between adjacent objects occur. Therefore, this method can also effectively fuse the information from both, thereby alleviating the inconsistency problem of RGB-D structures. (See also...) Figure 2 As shown, specifically: the first super-resolution depth map F output from the prediction network... d And the first uncertainty estimation diagram u d And the second super-resolution depth map F output in the depth detail estimation network. te Second uncertainty estimation diagram u teAll of these are used as inputs to the uncertainty-guided fusion module. In this module, the first depth feature map is augmented based on the first uncertainty estimation map; the second depth feature map is augmented based on the second uncertainty estimation map; and the augmented first and second depth feature maps are then subjected to a 3×3 convolution to obtain the super-resolution depth map.

[0072] The feature map after feature enhancement based on uncertainty is represented as follows:

[0073] F′=F*(1+SoftMax(Conv 3×3 (u)))

[0074] Where F′ represents the feature map after feature enhancement; F represents the first deep feature map to be enhanced or the second deep feature map to be enhanced; u represents the first uncertainty estimation map or the second uncertainty estimation map.

[0075] III. Training Optimization

[0076] 1. To provide a richer texture representation in the output of the depth detail estimation network, this embodiment of the invention employs a Laplacian filter to process the color image in the depth detail estimation network to obtain depth texture regions; and uses a texture loss function to perform feature imposition on the depth texture regions; the texture loss function is expressed as:

[0077] L=||(y te -y)+(y te -y)·t||1

[0078] Among them, y te y represents the super-resolution result map corresponding to the second depth feature map; y represents the high-resolution depth map; t represents the depth texture region to be enhanced.

[0079] 2. In this embodiment of the invention, the super-resolution depth map is trained and optimized using a loss function;

[0080] The loss function is expressed as:

[0081]

[0082] To avoid the instability caused by the division by zero in the equation, we design u = logσ in the network. 2 Therefore, the loss can be further expressed as:

[0083]

[0084] in, y represents the super-resolution depth map; u represents the uncertainty estimation map. The high-resolution depth map y refers to the reference image used, i.e., the ground truth. Since it is supervised training, it has labels during the training process.

[0085] Therefore, in each module, the model's loss function can be expressed as:

[0086] In a conditional variational encoder network:

[0087]

[0088] In deep detail estimation networks:

[0089]

[0090] In uncertain fusion networks:

[0091]

[0092] It should also be noted that the probabilistic network in the Conditional Variational Encoder Network (CVAENet) includes a prior network and a posterior network; the prior network has been described above; in this embodiment of the invention, the structure of the posterior network is consistent with that of the prior network, except that the inputs of the prior and posterior networks are different; in this embodiment of the invention, the concatenation of the sampled image and the super-resolution depth map trained by the loss function in the feature dimension is used as the input of the posterior network. In the posterior network, the concatenated image is modulated into a Gaussian distribution prior mean and prior variance through multiple 3×3 convolutions (4 in this embodiment of the invention); additionally, KL divergence is introduced to constrain the difference between the prior network and the posterior network.

[0093] The effectiveness of the uncertainty-guided bimodal separable image super-resolution method provided by the present invention will be illustrated by a specific embodiment.

[0094] This invention uses two datasets to validate the effectiveness of our method: the NYUv2 dataset and the real-world RGBDD dataset. The evaluation metric is RMSE; a lower RMSE indicates higher image reconstruction quality. The results of comparing the method provided in this invention with other methods on these two datasets are shown in Tables 1 and 2 below:

[0095] Table 1 compares the method provided in this embodiment of the invention with other methods on the NYUv2 dataset.

[0096]

[0097]

[0098] Table 2 compares the method provided in this embodiment of the invention with other methods on the RGBDD dataset.

[0099]

[0100] As shown in Tables 1 and 2 above, the embodiments of the present invention are compared with state-of-the-art depth image super-resolution methods on two test sets with three different scaling factors (×4, ×8, ×16). Clearly, the method provided by the embodiments of the present invention achieves optimal performance across multiple datasets and scaling factors, demonstrating its significant advantages over previous techniques. Furthermore, the model performance of the method provided by the embodiments of the present invention at a scaling factor of 16 is significantly better than the suboptimal models on all two datasets, proving its ability to recover richer results from smaller images. In addition, the present invention also tested the model's performance on real datasets to verify its generalization ability, achieving better performance on real datasets compared to other methods. Specific visual comparison results can be found in the figures. Figure 3 and Figure 4 As shown.

[0101] Furthermore, this embodiment of the invention also visualizes the uncertainty results during the learning process of the two backbone networks, as shown in the following figures. Figure 5 As shown in the figure, the depth detail estimation network focuses more on reconstructing texture details in the image and effectively complements the reconstruction results of the conditional variational encoder network. Experimental results on multiple datasets demonstrate the superior performance and generality of the method provided in this embodiment of the invention. Compared with comparative methods, the method provided in this embodiment of the invention achieves competitive results.

[0102] This invention provides an uncertainty-guided bimodal separable image super-resolution method. First, a depth reconstruction network based on a conditional variational autoencoder is designed. Unlike commonly used depth super-resolution reconstruction methods, the method provided in this invention introduces label information into the network through mutual constraints of prior and posterior information, thereby obtaining more reliable super-resolution results. Furthermore, this structure can easily enhance the depth encoder network, improving its performance. To improve the performance of the fusion results, this invention introduces uncertainty learning into the depth super-resolution task for the first time, enabling color information to provide a more effective supplement to the fusion network, achieving more reliable fusion results. Through this uncertainty-guided bimodal separable image super-resolution method, modal separation can be achieved while improving the quality of low-resolution depth images. This modal separation is reflected in the training and inference processes; the color image and depth reconstruction processes can be separated. That is, during training, the use of the depth image can be chosen, and during testing, the use of the depth image can also be chosen.

[0103] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. An uncertainty-guided dual-modal separable image super-resolution method, characterized in that, include: Obtain a low-resolution depth image, and obtain the color image corresponding to the low-resolution depth image; The low-resolution depth image is upsampled using bicubic interpolation to obtain a sampled image; Based on the sampled image, the randomness characteristics of the low-resolution depth image are obtained; The low-resolution depth image is input into a depth encoder, and deep features of the low-resolution depth image are obtained through multiple residual modules; Based on the randomness characteristics and the deep characteristics, a first feature map and a corresponding first uncertainty estimation map are obtained; The color image and the sampled image are input into the deep detail estimation network, which outputs a second feature map and a corresponding second uncertainty estimation map. Based on the first uncertainty estimation map, feature enhancement processing is performed on the first feature map; Based on the second uncertainty estimation map, feature enhancement processing is performed on the second feature map; The first and second depth feature maps after feature enhancement are processed by 3×3 convolution to obtain a super-resolution depth map.

2. The uncertainty-guided dual-modal separable image super-resolution method as described in claim 1, characterized in that, The size of the sampled image is the same as the size of the super-resolution depth map.

3. The uncertainty-guided dual-modal separable image super-resolution method as described in claim 1, characterized in that, The step of obtaining the randomness features of the low-resolution depth image based on the sampled image specifically includes: The sampled image is modulated into a Gaussian distribution prior mean and prior variance by multiple 3×3 convolutions; the prior mean and prior variance are processed by reparameterized sampling to obtain the randomness features of the low-resolution depth image.

4. The uncertainty-guided dual-modal separable image super-resolution method as described in claim 1, characterized in that, The step of obtaining a first depth feature map and a corresponding first uncertainty estimation map based on the randomness features and the deep features specifically includes: The randomness feature and the deep feature are concatenated; The concatenated features are fused using a 1×1 convolution to obtain the first fused feature; The first fused feature is enhanced by multiple residual modules to obtain the first feature map; The first feature map is processed by a 3×3 convolution to obtain the first depth map corresponding to the first feature map; The first feature map is further processed using a 3×3 convolution to obtain the first uncertainty estimation map corresponding to the first feature map.

5. The uncertainty-guided dual-modal separable image super-resolution method as described in claim 1, characterized in that, The step of inputting the color image and the sampled image into the depth detail estimation network and outputting a second depth feature map and a corresponding second uncertainty estimation map specifically includes: The color image and the sampled image are concatenated and used as input to the depth detail estimation network; In the deep detail estimation network, a 1×1 convolution is used to fuse the concatenated features to obtain a second fused feature; The second fused feature is enhanced by multiple residual modules to obtain a second feature map; The second feature map is processed by a 3×3 convolution to obtain the second depth map corresponding to the second feature map; The second feature map is then processed using a 3×3 convolution to obtain the second uncertainty estimation map corresponding to the second feature map.

6. The uncertainty-guided dual-modal separable image super-resolution method as described in claim 5, characterized in that, Also includes: The color image is processed using a Laplacian filter to obtain a depth texture region; The deep texture region is subjected to feature imposition processing using a texture loss function.

7. The uncertainty-guided dual-modal separable image super-resolution method as described in claim 6, characterized in that, The texture loss function is expressed as: ; in, y represents the second depth map; y represents the high-resolution depth map; t represents the depth texture region to be enhanced.

8. The uncertainty-guided dual-modal separable image super-resolution method as described in claim 3, characterized in that, Also includes: The super-resolution depth map is trained and optimized using a loss function; The loss function is expressed as: ; in, y represents a super-resolution depth map; y represents a high-resolution depth map. This represents a graph indicating uncertainty estimation.

9. The uncertainty-guided dual-modal separable image super-resolution method as described in claim 8, characterized in that, Also includes: The sampled image is concatenated with the super-resolution depth map; The concatenated image is modulated into a Gaussian distribution posterior mean and posterior variance by multiple 3×3 convolutions. The prior mean, prior variance, posterior mean, and posterior variance are constrained by KL divergence.

Citation Information

Patent Citations

  • Data processing method of depth map super-resolution reconstruction and denoising neural network

    CN112862684A

  • Depth map super-resolution method and system based on uncertainty perception feature transmission

    CN115511708A