Point cloud color denoising method and system based on latent diffusion model and storage medium

Abstract

The application provides a point cloud color denoising method and system based on a latent diffusion model and a storage medium, introduces a diffusion model into the field of point cloud color denoising, compresses high-dimensional point cloud data into a low-dimensional space for processing through latent space mapping, and significantly improves the calculation efficiency. The method innovatively generates real distortion noise based on a geometry-based point cloud compression standard, and obtains continuous noise distribution through an interpolation strategy. In the denoising process, the geometric information is combined as a conditional input to guide the denoising, and a dynamic sampling step control mechanism is introduced to adaptively adjust the processing strategy according to the region complexity, so that the processing accuracy of the complex region is ensured, and the calculation efficiency is optimized. Experimental results show that the method can improve the color quality of the coded and decoded distorted point cloud, reduce the attribute artifacts caused by compression reconstruction, and effectively improve the visual effect of the point cloud.

Description

Technical Field

[0001] This invention relates to the field of three-dimensional data processing technology, specifically to a point cloud color denoising method, system, and storage medium based on a latent diffusion model, which is particularly suitable for color attribute enhancement processing during point cloud data encoding and decoding. Background Technology

[0002] With the rapid development of 3D data acquisition and processing technologies, point clouds, as an important 3D data representation method, have been widely used in fields such as autonomous driving, robot vision, and virtual reality. Point cloud data not only contains the geometric information of objects but also records the color information of the scene, providing important support for the accurate reconstruction and analysis of 3D scenes. However, during the acquisition, transmission, and storage of point cloud data, color information is often affected by various factors, resulting in noise, which degrades data quality and affects the effectiveness of subsequent applications. Therefore, research on point cloud color denoising technology is of great significance.

[0003] In the field of point cloud color denoising, existing techniques typically involve processing the data directly in the original data space. These methods first acquire the color information of the points to be processed and their neighborhood features, then optimize and correct the color values ​​of the points by designing specific mathematical models or deep learning networks. During the model training phase, Gaussian noise is often used to contaminate the original data to construct training samples, and the model parameters are optimized by minimizing the reconstruction error. Furthermore, some diffusion-based methods achieve gradual noise removal by setting a fixed number of sampling steps.

[0004] However, the above-mentioned technical solutions have the following drawbacks: First, directly performing feature extraction and optimization operations in the original high-dimensional space results in high computational complexity, affecting the training efficiency and practical application performance of the model. Second, the training data constructed using Gaussian simulated noise differs significantly from the distortion characteristics in real-world scenes, leading to poor denoising performance of the trained model in practical applications. Furthermore, existing methods typically rely solely on color features during denoising, ignoring the rich geometric information contained in point cloud data and its synergistic effects with geometric information. This can easily lead to processing results that are incompatible with the geometric structure of objects, resulting in color-geometry inconsistencies. In addition, traditional diffusion models use a fixed number of sampling steps for iterative optimization, failing to flexibly adjust processing strategies according to the complexity of different regions. This wastes computational resources and makes it difficult to guarantee processing accuracy in complex regions.

[0005] These technical problems severely restrict the practical application of point cloud color denoising technology, and a new technical solution is urgently needed to solve them. Summary of the Invention

[0006] To address the technical problems of high computational complexity, unrealistic noise models, insufficient utilization of geometric information, and inflexible processing strategies in existing point cloud color denoising techniques, this invention provides a point cloud color denoising method and system based on a latent diffusion model. It employs a point cloud autoencoder to pre-encode and map the point cloud into a low-dimensional latent space. Within this space, it combines techniques such as realistic noise simulation, geometric condition assistance, and dynamic noise control for comprehensive processing, thereby significantly improving the processing efficiency and quality of point cloud color denoising and ensuring consistency between the point cloud color and geometric denoising results.

[0007] Specifically, this invention provides a point cloud color denoising method based on a latent diffusion model, which includes six steps: Step S100, mapping high-dimensional point cloud data containing color and geometric features to a low-dimensional latent space using a point cloud autoencoder; Step S200, generating noise distributions with different distortion levels; Step S300, performing noise diffusion processing on the data in the latent space; Step S400, setting the sampling step number for the diffused data; Step S500, performing denoising processing using a neural network with a cross-attention mechanism, using the geometric features of the point cloud as conditional input to guide the denoising process; Step S600, decoding the denoised data into point cloud color features.

[0008] Preferably, in step S500, the geometric features of the point cloud are used as input to the neural network to optimize the probability gradient calculation. This approach can further improve the accuracy of the denoising process.

[0009] Furthermore, in the specific implementation of step S200, the method for generating noise distributions with different distortion levels is as follows: first, multiple discrete distortion levels are generated, and then a continuous noise distribution is generated by interpolating adjacent distortion levels. This processing method makes the noise model closer to the actual scene. More specifically, the generation of discrete distortion levels can be accomplished using geometry-based point cloud compression standards, video-based point cloud compression standards, or Gaussian noise; while the interpolation operation can be performed by weighted combination using weight coefficients, using a Gaussian mixture model, or physical simulation.

[0010] Furthermore, in step S300, it is preferable to control the ratio coefficient between the original data and the noise data to gradually convert the data into a form with different levels of noise, thereby stably completing the forward diffusion process.

[0011] Furthermore, in step S400, the number of sampling steps is dynamically allocated by calculating the texture complexity of different regions in the point cloud data. Regions with higher complexity are allocated more sampling steps, while regions with lower complexity are allocated fewer sampling steps. The texture complexity is calculated by determining at least one of the region's point density, color gradient, or local geometric features. This dynamic allocation strategy significantly improves processing efficiency, and the multi-dimensional complexity assessment also enhances the accuracy of region feature identification.

[0012] Preferably, in the implementation of step S200, the continuous noise distribution is uniformly distributed according to the diffusion time step, so that each time step corresponds to a certain distortion level, thereby obtaining high-precision noise simulation.

[0013] Meanwhile, this invention also provides a corresponding system implementation scheme, including a processor, a memory, a system bus, and various functional modules. The processor is used to execute computational tasks, the memory is used to store program code and data, and the system bus is used to connect the various modules in the system. The system includes a latent space mapping module, a noise processing module, a diffusion processing module, a dynamic control module, a denoising processing module, and a reconstruction module, which are respectively used to execute steps S100 to S600 of the aforementioned method, wherein the computer program stored in the memory is executed by the processor to implement the steps of the aforementioned method.

[0014] Finally, the technical solution of the present invention can also be stored in the form of a computer program on a computer-readable storage medium. When the program is executed by a processor, it can realize each step of the above-mentioned point cloud color denoising method.

[0015] Compared with existing technologies, the present invention achieves point cloud color denoising in the latent space, which has the following beneficial technical effects:

[0016] First, by shifting the processing to a low-dimensional latent space, this invention significantly reduces computational complexity and achieves more efficient processing.

[0017] Secondly, this invention innovatively adopts a realistic noise simulation and interpolation strategy based on G-PCC encoding distortion, which makes the noise modeling more closely resemble the color distortion characteristics in actual application scenarios and improves the denoising effect.

[0018] Furthermore, this invention introduces geometric information as a conditional input into the diffusion model to assist the color denoising process, which effectively improves the consistency between point cloud geometry and color representation, and ensures higher denoising quality.

[0019] More importantly, this invention breaks through the limitation of the traditional diffusion model using a fixed number of sampling steps, and creatively introduces a dynamic sampling step control mechanism. It adopts a differentiated processing strategy for regions with different complexities: more time steps are allocated in high-complexity regions to improve denoising accuracy, and fewer time steps are reduced in low-complexity regions to optimize computational efficiency, thereby achieving efficient utilization of computing resources.

[0020] Through experimental testing, using peak signal-to-noise ratio (PSNR) as an objective evaluation metric, the method of this invention has achieved excellent results in improving the color quality of point clouds and has also demonstrated significant advantages in practical applications. Attached Figure Description

[0021] Figure 1 This is a flowchart of a point cloud color denoising method based on a latent diffusion model;

[0022] Figure 2 This is a schematic diagram of a point cloud color denoising method framework based on a latent diffusion model;

[0023] Figure 3 This is a schematic diagram comparing the visual effects of point cloud color denoising. Detailed Implementation

[0024] The technical solution of this patent will be further described in detail below with reference to specific embodiments. It should be noted that the following detailed descriptions of the technical solution are exemplary and will be understood by those skilled in the art. The technical solution is also applicable to related technical fields under the same technical concept. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which this application pertains.

[0025] To address the technical problems of high computational complexity, unrealistic noise models, insufficient utilization of geometric information, and inflexible processing strategies in existing point cloud color denoising techniques, this invention is the first to apply a diffusion model to the field of point cloud color denoising. It generates realistic color noise by introducing geometry-based point cloud compression (G-PCC) standard distortion and uses interpolation strategies to generate a continuous noise distribution to match the time step requirements of the diffusion model. Furthermore, this invention compresses point cloud data from a high-dimensional point space to a low-dimensional latent space through latent space mapping and incorporates geometric information to assist the denoising process. Simultaneously, this invention introduces a dynamic sampling step control mechanism to provide more refined color denoising for complex texture regions of the point cloud and optimizes computational efficiency in low-complexity regions. (See appendix to the specification.) Figure 1 The flowchart below shows the point cloud color denoising method based on the latent diffusion model of the present invention, which specifically includes the following operation steps.

[0026] Step S100, Latent Space Mapping. Using a pre-trained point autoencoder, the high-dimensional point cloud data, containing color and geometric features, is mapped to a low-dimensional latent space. This step reduces computational complexity by lowering the data dimensionality while ensuring the preservation of the data's core feature information.

[0027] Step S200: Noise Simulation and Interpolation Generation. Noise point clouds with different distortion levels are generated through geometry-based point cloud compression (G-PCC) encoding and decoding. An interpolation strategy is then used to convert these discrete distortion levels into a continuous noise distribution. This approach makes the noise model more closely resemble the distortion characteristics in real-world applications.

[0028] Step S300, forward noise diffusion. This step is performed in the latent space. Based on the continuous noise distribution generated in step S200, the original latent representation is gradually diffused in a controllable manner into representations with different levels of noise, ultimately forming a high-noise latent representation.

[0029] Step S400: Dynamic sampling step control. A self-attention mechanism is introduced to calculate the texture complexity of different regions in the point cloud, and the sampling step count for the denoising process is dynamically adjusted accordingly. This adaptive strategy can flexibly allocate computational resources based on region features.

[0030] Step S500, inverse denoising. A U-Net network with a cross-attention mechanism is used for processing, taking the geometric features of the point cloud as conditional input to guide the denoising process, ensuring that the processing result is consistent with the geometric structure.

[0031] Step S600, Decoding and Reconstruction. The denoised latent representation is decoded back to the original space to reconstruct the improved point cloud color features, completing the entire processing flow.

[0032] The above six core steps are interconnected through a tight logical relationship, forming a complete processing chain. See the appendix to the instruction manual. Figure 2 The framework of the point cloud color denoising method based on the latent diffusion model is illustrated in the diagram, and each step is explained in more detail.

[0033] S100, Latent Space Mapping. This step uses a pre-trained point cloud autoencoder to map the color and geometric features of the high-dimensional point cloud to a low-dimensional latent space for processing, thereby reducing computational complexity and preserving the core features of the data.

[0034] In point cloud data, each point contains location coordinates and color information, forming a high-dimensional feature space. While this high-dimensional representation fully records the feature information of the point cloud, direct processing within it incurs a significant computational burden. Therefore, this invention employs a point cloud autoencoder to map these high-dimensional data to a low-dimensional latent space, significantly reducing the complexity of data processing while preserving the core features of the data.

[0035] The core function of a point cloud autoencoder is to compress input point cloud data into a low-dimensional latent representation, such as z0, through an encoder. This process can be achieved, for example, through a multi-layer neural network, which can learn and preserve key feature information in the point cloud data. In particular, during the pre-training stage, the autoencoder learns the intrinsic relationship between color features and geometric features in the point cloud data, ensuring that this relationship is maintained during dimensionality reduction.

[0036] This latent space mapping method allows subsequent denoising to be performed in a more efficient feature space. By reducing the data dimensionality, not only is the computational cost significantly reduced, but a suitable processing space is also provided for the application of diffusion models. Furthermore, since the autoencoder has already learned the feature distribution patterns of the point cloud data during the pre-training stage, it can preserve important data information while reducing dimensionality, laying a solid foundation for subsequent denoising.

[0037] S200, Noise Simulation and Interpolation Generation. This step generates a noise point cloud with discrete distortion levels using geometry-based point cloud compression (G-PCC) and then expands it into a continuous noise distribution using an interpolation strategy to match the time step requirements of the diffusion model.

[0038] In actual point cloud data transmission and compression, color information distortion often exhibits specific patterns and characteristics. To better simulate these distortion characteristics in real-world scenarios, this step first processes the point cloud data using the G-PCC encoding / decoding standard to generate point cloud data with varying degrees of distortion, obtaining a series of highly realistic distortion levels. For example, six different distortion levels are formed, denoted as q1, q2, up to q6. These discrete distortion levels realistically reflect the various distortion situations that point cloud data may encounter during actual compression.

[0039] Furthermore, to meet the more continuous noise distribution required by the diffusion model during processing, this step employs an interpolation strategy to extend the discrete distortion levels into a continuously distributed q. t Specifically, for adjacent distortion levels, an interpolation formula is used to achieve continuity:

[0040] qt =(1-a t )·q i +a t ·q i+1 ,a t ∈[0,1],

[0041] Where, q i and q i+1 α represents the discrete distortion level. t These are interpolation weights, and their calculation formula is:

[0042]

[0043] Interpolated noise level q t The distortion levels are evenly distributed within the diffusion time step T, ensuring that each time step t corresponds to an appropriate distortion intensity, thus expanding the 6 initial distortion levels to T. The specific noise level allocation is as follows:

[0044] q1:t = 1 ~ T / 6;

[0045] q2:t=T / 6~2T / 6;

[0046] q3:t = 2T / 6 ~ 3T / 6;

[0047] q4:t=3T / 6~4T / 6;

[0048] q5:t = 4T / 6 ~ 5T / 6;

[0049] q6:t=5T / 6~T;

[0050] This interpolation method ensures a smooth transition in noise distribution between any two adjacent distortion levels.

[0051] Based on the time step T of the diffusion process, this step uniformly distributes the obtained continuous noise distribution. This distribution method ensures that each time step t corresponds to a specific distortion level qt, providing a complete noise evolution path for the subsequent diffusion process. This noise simulation method based on actual compression distortion characteristics, compared with the traditional method using Gaussian noise, can more accurately reflect the distortion situation encountered by point cloud data in practical applications, thus providing a more targeted processing basis.

[0052] S300, forward noise diffusion. This step is performed in the latent space, gradually adding noise to the original latent representation through a continuous noise distribution qt, achieving a controllable transformation from a clean representation to a noisy representation.

[0053] After noise simulation and interpolation generation are completed, this step executes a forward noise diffusion process in the latent space. This process is achieved through a carefully designed mathematical formula, specifically the diffusion formula used.

[0054]

[0055] Where ε t It follows the distribution qt. In this formula, z0 represents the original latent representation, z t β represents the latent representation after adding noise. t It is a parameter that controls the ratio of raw information to noise, ε t The noise is sampled from the noise distribution qt generated in step S200.

[0056] The characteristic of this diffusion process is that it allows for precise control over the addition of noise. By adjusting the parameter αt, the ratio of retained original information to introduced noise can be precisely controlled at each time step. As the time step progresses, the proportion of original information gradually decreases, while the influence of noise gradually increases, forming a gradual noise diffusion process.

[0057] Compared to traditional methods of directly adding noise, this noise addition method based on a diffusion model offers better controllability and reversibility. It not only simulates the distortion process in real-world scenarios but, more importantly, provides a clear path for subsequent denoising. Since the entire diffusion process is based on a rigorous mathematical model, the same principles can be applied in the subsequent inverse denoising process.

[0058] S400, Dynamic Sampling Step Control. This step calculates the texture complexity of each point in the point cloud through a self-attention mechanism, thereby adaptively adjusting the number of sampling steps in the denoising process and achieving efficient allocation of computing resources.

[0059] The core idea of ​​the dynamic sampling step control mechanism is to allocate processing resources based on the complexity of different regions. Through a self-attention mechanism, the system can automatically calculate the texture complexity of each region in the point cloud, which reflects the richness of detail and the processing difficulty of the region. After the calculation is completed, the system determines the number of sampling steps Tt required for that region based on the complexity weight of the denoised region, for example, setting this number of steps to 1000. This adaptive step allocation method allows the processing to better adapt to the local features of the point cloud data.

[0060] In practical applications, for regions with high gradients or high density, the system allocates more sampling steps to ensure processing accuracy. These regions typically contain rich texture details or complex geometries, requiring more meticulous processing to achieve the desired denoising effect. Conversely, for low-gradient or sparse regions, due to their relatively simple structure, the system allocates fewer sampling steps, thereby improving computational efficiency.

[0061] This dynamic control strategy overcomes the limitation of traditional diffusion models that use a fixed number of sampling steps. By allocating more computational resources to complex regions while reducing processing time for simple regions, this adaptive processing method based on region complexity ensures both denoising quality and overall processing efficiency. This provides a more efficient processing framework for subsequent inverse denoising processes.

[0062] S500, inverse denoising. This step utilizes a U-Net network structure with a cross-attention mechanism to denoise the high-noise latent representation z. t Initially, a clean latent representation Z0 is reconstructed stepwise through a backdiffusion step. This process is governed by the backdiffusion equation:

[0063]

[0064] in, Represents a geometric feature P g The gradient of the log probability of the noisy data. The gradient is approximated by a U-Net model, which is used to learn the denoising function ∈ θ (Z t ,t,P g The U-Net model incorporates a cross-attention mechanism that aligns the geometric information of the point cloud with the latent color representation, thereby ensuring that the reconstructed color features are consistent with the underlying geometry.

[0065] To optimize the denoising process, a dynamic sampling strategy is proposed, adjusting the number of diffusion steps based on the complexity of the local texture. Through a self-attention mechanism, the framework evaluates the geometric and color features of each region in the point cloud. More diffusion steps are allocated to regions with high texture complexity (e.g., regions with sharp gradients, high curvature, or dense point distributions), while fewer steps are allocated to simpler regions, thus reducing computational overhead. The sampling time step T for each region is specified. i The calculation formula is:

[0066]

[0067] Wherein, C(x) i ) Quantified point x i The complexity of the local texture is calculated by the self-attention network, max(C(x)). i)) represents the maximum complexity, and T is the global sampling time step.

[0068] The changes in the above parameters control the denoising process. To better utilize the geometric information of the point cloud to guide the denoising process, this invention optimizes the basic formula by introducing geometric conditions, using the improved formula:

[0069]

[0070] The condition term P here g The geometric coordinates of the input point cloud are encoded by the encoder and mapped to a vector in the latent space, which contains the geometric feature information of the point cloud. By using the geometric information as a conditional input, the conditional probability gradient is optimized, thus providing more accurate guidance for color denoising.

[0071] In its implementation, this invention employs a U-Net network with a cross-attention mechanism as the core network structure for the denoising process. This network structure fully utilizes geometric information to assist in the color denoising process. The cross-attention mechanism enables the network to automatically focus on relevant geometric features when processing color information, thereby ensuring that the denoising result remains consistent with the geometric structure of the point cloud. At each time step, the network considers the conditional constraints of geometric features, effectively avoiding the color-geometry mismatch problem commonly found in traditional methods.

[0072] This geometry-based denoising strategy offers significant advantages. It not only effectively removes noise but also preserves the intrinsic correlation between color and geometry in point cloud data. By considering the guiding role of geometric information in each denoising step, the system can generate more natural and accurate denoising results. Combined with the aforementioned dynamic sampling step control mechanism, the entire denoising process ensures both processing quality and efficient utilization of computational resources.

[0073] S600, Decoding and Reconstruction. This step utilizes the decoder portion of the point cloud autoencoder to convert the denoised latent representation back to the original space, reconstructing the improved point cloud color features.

[0074] The decoding and reconstruction process is the final step in the entire processing flow. Its goal is to convert the denoised data in the latent space back into directly usable point cloud color features. Through the decoder, the system can restore the representation in the low-dimensional latent space to the color features in the original point cloud space, enabling the processed results to be directly applied to real-world scenarios.

[0075] This reconstruction process accurately converts the information contained in the latent representation back into intuitive color features. Because the system has fully considered the guiding role of geometric information in the preceding steps, especially in the denoising process of the S500 step, the reconstructed color features match the geometric structure of the point cloud very well. After decoding and reconstruction are complete, the processed point cloud data can be directly used for subsequent applications.

[0076] Through this decoding and reconstruction process, the present invention ultimately improves the color quality of encoded and decoded distorted point clouds, reduces attribute artifacts caused by compression and reconstruction, and effectively enhances the visual effect of point clouds.

[0077] To demonstrate the flexibility and adaptability of the technical solution of this invention, some specific implementation methods in the above steps can be changed or replaced accordingly to adapt to different application scenario requirements.

[0078] In a preferred embodiment, during the noise simulation and interpolation generation process of step S200, the present invention provides several optional technical solutions. Besides using interpolation methods to generate continuous coded distortion noise, other methods such as Gaussian mixture models or physical simulations can also be used. Gaussian mixture models can simulate more complex noise characteristics through combinations of multiple Gaussian distributions, while physical simulation methods can generate noise based on the physical characteristics of actual data transmission and compression processes. These alternatives can all achieve the generation of continuous noise distributions, providing the necessary noise basis for the diffusion model.

[0079] Similarly, in step S200, the present invention also offers flexibility in the selection of distortion type. In addition to using the geometry-based point cloud compression (G-PCC) standard, other distortion types such as video-based point cloud compression (V-PCC) or Gaussian noise can be selected, allowing the present invention to better adapt to different types of point cloud data processing needs. For example, when processing dynamic point cloud data, using the V-PCC standard may yield better results.

[0080] In the S400 dynamic sampling step control process, there are several options for calculating complexity. Besides texture-based complexity calculations, other metrics such as point density, color gradient, or local geometric properties can also be used. Each of these metrics has its advantages: point density reflects the richness of information in a region, color gradient shows the intensity of color changes, and local geometric properties express the complexity of the surface. By selecting an appropriate complexity calculation method, it is possible to better adapt to the characteristics of different point cloud data and achieve more precise sampling step control.

[0081] Experimental Section

[0082] The following qualitative and quantitative experiments are intended to evaluate the effectiveness of the method of the present invention in improving the color denoising of distorted point clouds, and are compared with the SSIU-Net and GQE-Net methods.

[0083] Datasets: This experiment used four dynamic point cloud datasets: Waterloo Point Cloud SubsetDataset (WPCSD), 8i Voxelized Full Bodies (8iVFB), Owlii Dynamic Human TexturedMesh Sequences (Owlii), and Microsoft Voxelized Upper Bodies (MVUB). These datasets contain 37 point clouds, of which 24 were used for training and the remaining 13 for testing.

[0084] All point clouds were encoded according to the C1 standard in G-PCC Test Model Category 13 version 23.0 (TMC13v23.0), using lossless geometric compression and lossy attribute compression. For ease of subsequent comparison experiments, the same dimensionality upscaling transformation configuration as in the SSIU-Net and GQE-Net experiments was used.

[0085] Experimental Setup: First, distorted original patch pairs were constructed, and each pair was packaged into an efficient and compact H5 dataset. The number of points n in each patch was set to 2048, and the overlap ratio r was set to 4. The dataset contained point clouds with different distortion levels, from R01 to R06, with corresponding quantization parameters (QP) of 51, 46, 40, 34, 28, and 22, where higher QP values ​​indicate greater distortion. Point clouds with the same distortion level were grouped together, and models for each distortion level were trained separately, for a total of six models.

[0086] The experiment was conducted on a machine equipped with an Intel i9-10900X CPU, 128GB RAM, and an NVIDIA GeForce RTX3090 GPU. PyTorch v1.11.0 was used for training, with a batch size of 16 and a maximum training step count of 1,000,000. Training parameters and model weights [35, 37, 39, 69] provided in a publicly available code repository were used. The model was saved every 5000 iterations.

[0087] Objective Experimental Results: To objectively evaluate the performance of the proposed method, Peak Signal-to-Noise Ratio (PSNR) was used as the evaluation metric to assess the PSNR improvement of SSIU-Net, GQE-Net, and the proposed method on distorted point clouds. Specifically, the PSNR improvement of all test point clouds was calculated at six distortion levels (R01 to R06, with the distortion level decreasing progressively), and Table 1 summarizes the PSNR improvement of 13 test point clouds. This table demonstrates the quantitative performance of different methods, showing that the proposed method performs excellently in effectively improving the quality of distorted point clouds.

[0088] Table 1 Comparison of the method of the present invention with existing methods

[0089]

[0090] The method of this invention achieves significant PSNR improvements across all distortion levels, with a marked improvement in BD-Rate, indicating that it reduces the bit rate while maintaining comparable reconstruction quality, effectively saving bitstream. Compared to SSIU-Net, the method of this invention performs better across all distortion levels. Especially at low distortion levels (R05 and R06), the method of this invention still significantly improves PSNR and saves more bitstream (larger absolute value of BD-BR). Compared to GQE-Net, the method of this invention achieves better results at the highest distortion level (R01) and the lowest distortion level (R06), but lags behind GQE-Net at R02–R05.

[0091] Subjective Experiment Results: To further evaluate the performance of the denoising method, subjective experiments were conducted by visualizing the reconstructed point clouds. Results include a comparative visualization of color-denoised point clouds processed by SSIU-Net, GQE-Net, and the method proposed in this invention, as well as a comparison between the distorted point cloud and the original point cloud, such as... Figure 3 As shown. The visualization results are presented as full-scale point clouds and magnified views of local areas, highlighting the differences in detail restoration.

[0092] like Figure 3 As shown, point clouds processed by SSIU-Net and GQE-Net often exhibit excessive smoothing (e.g., RedandBlack sequence) or blockages (e.g., Ship sequence), making it difficult to preserve details and resulting in blurred color transitions, especially at higher distortion levels (R01, R02). In contrast, the method proposed in this invention demonstrates better detail preservation and clearer color recovery, excellent denoising performance in complex and textured regions, minimizes artifacts, and ensures structural consistency.

[0093] Furthermore, as shown in Table 2's complexity analysis, although the method of this invention surpasses SSIU-Net and GQE-Net in both the number of model parameters and FLOPs, and its color denoising performance at the R02–R05 distortion levels is inferior to GQE-Net, its inference time is significantly faster than the other two methods. This is due to the application of the latent diffusion model, which can efficiently process data in low-dimensional space and concentrate computational efforts on high-complexity regions through a dynamic sampling step control mechanism. These results highlight the practicality of the method of this invention in large-scale applications, providing an effective balance between computational requirements and denoising quality.

[0094] Table 2. Comparison of complexity analysis between the method of this invention and existing methods.

[0095]

[0096] Based on the above point cloud color denoising method, this invention also proposes a corresponding system implementation scheme. The system mainly consists of six functional modules, which are closely connected through data flow to jointly complete the point cloud color denoising task.

[0097] First, the system includes a latent space mapping module that integrates a pre-trained point cloud autoencoder. The encoder portion of the autoencoder receives the input point cloud data and maps its color and geometric features to a low-dimensional latent space, providing an efficient data representation for subsequent processing. This module is designed with computational efficiency in mind, significantly reducing the complexity of data processing through dimensionality reduction.

[0098] Connected to the latent spatial mapping module is the noise processing module, which performs noise simulation and interpolation generation. This module first generates noise point clouds with different distortion levels using an integrated G-PCC codec, and then uses an interpolation unit to convert these discrete distortion levels into a continuous noise distribution. The output of this module provides the necessary noise basis for the subsequent diffusion process.

[0099] Furthermore, a forward diffusion processing module is set up to perform noise diffusion operations in the latent space. This module contains processing units for performing diffusion formula calculations, and can progressively convert the original latent representation into a noisy representation based on the noise distribution generated by the previous module, ensuring the controllability and reversibility of the noise addition process.

[0100] The dynamic control module evaluates the complexity of different regions in the point cloud data through its built-in self-attention computing unit. Based on these evaluation results, the module can dynamically adjust the number of sampling steps in the subsequent denoising process, achieving optimized allocation of computing resources and thus greatly improving the system's processing efficiency.

[0101] The denoising module is the core of the system, employing a U-Net network structure with a cross-attention mechanism. This module not only performs basic denoising operations but also effectively utilizes the geometric information of the point cloud to guide the denoising process. Through a specially designed conditional input mechanism, this module ensures the consistency between the denoising results and the point cloud geometry.

[0102] Finally, the system also includes a decoding and reconstruction module, which uses the decoder part of the autoencoder to convert the processed latent representation back to the original point cloud color features and outputs improved point cloud data.

[0103] These functional modules are connected via a system bus and work collaboratively under the scheduling of the central processing unit. The system is also equipped with necessary storage units to store intermediate processing results and model parameters. Through this modular system design, not only is efficient point cloud color denoising achieved, but the system's scalability and maintainability are also guaranteed.

[0104] The technical solution of the present invention can also be stored in the form of a computer program on a computer-readable storage medium. Specifically, the computer-readable storage medium stores a computer program that, when executed by a processor, can implement the various steps of the point cloud color denoising method based on the latent diffusion model described above.

[0105] In this embodiment, the computer-readable storage medium can be any available medium that a computer can access. Available media include, but are not limited to: magnetic storage devices, such as hard disks, floppy disks, and magnetic tapes; optical storage devices, such as CD-ROMs, DVD-ROMs, and Blu-ray discs; and solid-state storage devices, such as solid-state drives, flash memory, ROM, and RAM. These storage media can be permanent or erasable, and can be removable or permanently mounted.

[0106] The computer program stored on the storage medium contains program code that implements the above-described processing steps. Specifically, the program code includes: code for executing the latent space mapping module of step S100, which implements the function of mapping point cloud data to a low-dimensional latent space; code for executing the noise simulation module of step S200, which implements the generation and processing of noise distribution; code for executing the diffusion processing module of step S300, which implements the noise diffusion operation in the latent space; code for executing the dynamic control module of step S400, which implements the dynamic allocation of sampling steps; code for executing the denoising processing module of step S500, which implements the denoising process based on geometric features; and code for executing the reconstruction module of step S600, which implements the function of reconstructing the processed data into point cloud color features.

[0107] When a computer program is loaded and run, the processor executes the aforementioned steps sequentially according to the instructions in the program code. During program execution, the computer's memory is used as working storage space to store intermediate data during processing, such as latent space representations, noise distributions, and processing parameters. The program code may also contain necessary data structure definitions for organizing and managing point cloud data, processing parameters, and intermediate results.

[0108] The program code in this storage medium can be written in any suitable programming language, such as high-level programming languages ​​like C++ and Python. The program code can be stored entirely as source code, as object code, or a combination of both. Furthermore, the program code can be stored in compressed form and decompressed only when needed.

[0109] In this way, the technical solution of the present invention can be embodied in the form of a program product, which facilitates deployment and use on different computing devices. When color denoising processing of point cloud data is required, the technical solution of the present invention can be implemented simply by loading the program from the storage medium into the computer and running it.

[0110] In summary, this invention proposes a point cloud color denoising method, system, and storage medium based on a latent diffusion model, solving key problems in the field of point cloud color denoising through a series of innovative technical means. This invention is the first to introduce a diffusion model into the field of point cloud color denoising, significantly improving computational efficiency by processing in a low-dimensional latent space. It replaces traditional Gaussian noise with realistic distortion simulation based on the G-PCC standard and combines it with an interpolation strategy to generate a continuous noise distribution, making the noise model more closely resemble real-world application scenarios. Furthermore, this invention innovatively incorporates geometric information as a condition into the diffusion model to guide the color denoising process, ensuring consistency between the processing results and the point cloud geometry. Simultaneously, by introducing a dynamic sampling step control mechanism, the system can adaptively allocate computational resources according to the complexity of the region, optimizing computational efficiency while maintaining processing accuracy in complex regions. The organic combination of these technical means enables this invention not only to theoretically fill the gap in the application of diffusion models in point cloud color processing but also to achieve efficient and accurate color quality improvement in practical applications.

[0111] The above descriptions are merely some embodiments of the present invention. Those skilled in the art can make various modifications and improvements without departing from the inventive concept of the present invention, and these all fall within the scope of protection of the present invention.

Claims

1. A point cloud color denoising method based on a latent diffusion model, characterized in that, Includes the following steps: Step S100, Latent Space Mapping: Use a point cloud autoencoder to map high-dimensional point cloud data containing color and geometric features to a low-dimensional latent space; Step S200, Noise simulation: Generate noise distributions with different distortion levels; Step S300, Forward noise diffusion: Perform noise diffusion processing on the data in the latent space; Step S400, Sampling control: Set the number of sampling steps for the diffused data; Step S500, reverse denoising: Denoising is performed using a neural network with a cross-attention mechanism, with the geometric features of the point cloud as conditional input to guide the denoising process. Step S600, Decoding and Reconstruction: Decode the denoised data into point cloud color features.

2. The method according to claim 1, characterized in that: In step S500, the geometric features of the point cloud are used as input to the neural network to optimize the probability gradient calculation.

3. The method according to claim 1, characterized in that: The method for generating the noise distribution of different distortion levels in step S200 is as follows: first, multiple discrete distortion levels are generated, and then a continuous noise distribution is generated by interpolating adjacent distortion levels.

4. The method according to claim 3, characterized in that: In step S200, the discrete distortion level is generated in any of the following ways: using a geometry-based point cloud compression standard, using a video-based point cloud compression standard, or using Gaussian noise; the interpolation operation is performed in any of the following ways: using weighted combinations with weight coefficients, using a Gaussian mixture model, or using physical simulation.

5. The method according to claim 1, characterized in that: In step S300, when noise is diffused in the potential space, the data is gradually converted into a form with different levels of noise by controlling the ratio coefficient between the original data and the noise data.

6. The method according to claim 1, characterized in that: In step S400, the number of sampling steps is dynamically allocated by calculating the texture complexity of different regions in the point cloud data. Regions with higher complexity are allocated more sampling steps, while regions with lower complexity are allocated fewer sampling steps.

7. The method according to claim 6, characterized in that: In step S400, the texture complexity is determined by calculating at least one of the following features: point density, color gradient, or local geometric features of the region.

8. The method according to claim 3, characterized in that: In step S200, the continuous noise distribution is uniformly distributed according to the diffusion time step T, so that each time step T corresponds to a certain distortion level.

9. A point cloud color denoising system based on a latent diffusion model, used to perform the point cloud color denoising method of claim 1, characterized in that, include: A processor is used to perform computational tasks; Memory, used to store program code and data; The system bus is used to connect the various modules in the system; A potential space mapping module is used to perform step S100; A noise processing module is used to perform step S200; A diffusion processing module is used to perform step S300; The dynamic control module is used to execute step S400; A noise reduction processing module is used to perform step S500; and A reconstruction module is used to perform step S600; The memory contains a computer program that, when executed by the processor, implements the steps of the method.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the point cloud color denoising method based on the latent diffusion model as described in any one of claims 1-8.

Images