A point cloud denoising method and system based on a fractional diffusion model

Abstract

The application provides a point cloud denoising method and system based on a fractional diffusion model, relates to the technical field of three-dimensional point cloud processing and artificial intelligence, and comprises the following steps: obtaining noise point cloud data to be denoised; inputting the noise point cloud data into a trained fractional prediction network to obtain a score vector; based on the point cloud distribution of the noise point cloud data and the score vector, adaptively matching a denoising time step number and a corresponding variance subsequence to obtain a variance scheduling table; according to the variance scheduling table, taking the noise point cloud data as initial current time step point cloud, iteratively updating the positions of points in the current time step point cloud through the inverse sampling rule of the fractional diffusion model, gradually eliminating coupled noise and converging to clean point cloud; the application can suppress outliers and over-smoothing under different noise modes, maintain thin structures and boundary details, and is suitable for robot navigation, automatic driving and three-dimensional reconstruction scenes.

Description

Technical Field

[0001] This invention relates to the fields of 3D point cloud processing and artificial intelligence technology, specifically to a point cloud denoising method and system based on a fractional diffusion model. Background Technology

[0002] With the rise of 3D geometric deep learning technology, point clouds, as an efficient 3D data representation, are widely used in tasks such as environmental modeling, object detection, registration, and reconstruction. Compared with 2D images, point clouds can directly describe the position, shape, and scale information of objects and scenes in 3D space, and have advantages such as simple structure, efficient storage, geometric computation friendliness, and insensitivity to point order permutations. In engineering practice, timely acquisition of clean point cloud data with faithful geometric details is a key prerequisite for improving the accuracy of 3D reconstruction, the stability of positioning and navigation, and the quality of downstream perception and decision-making. Therefore, point cloud denoising has become an important foundational step in ensuring 3D data quality and improving the overall performance of the system.

[0003] However, existing point cloud denoising techniques still suffer from the following major problems in complex application scenarios, resulting in insufficient detail fidelity and structural stability: Filtering optimization methods based on geometric priors are sensitive to hyperparameters and struggle to achieve a balance between noise suppression and detail preservation. Learning-based single-regression or empirical multi-round iterative methods are prone to over-smoothing, surface shrinkage, boundary blurring and outlier residues, especially in thin structures such as thin sheets and thin rods and in non-uniform sampling areas, where deformation and detail damage are more prominent. Many methods employ multiple iterations during the inference phase, but the iteration process itself is not modeled during the training phase, making it difficult for the network to perceive the preceding state and reasonably estimate the update direction and step size. Noise patterns and intensities vary greatly depending on the scenario and sensor. Existing methods mostly rely on fixed scheduling and manual parameter tuning, lacking a time step and step size selection mechanism that adapts to the sample, resulting in insufficient robustness and generalization ability on real scanning data. Summary of the Invention

[0004] To address the aforementioned problems, this invention proposes a point cloud denoising method and system based on a fractional diffusion model. This method suppresses outliers and excessive smoothing under different noise modes, while preserving fine structures and boundary details. It is applicable to scenarios such as robot navigation, autonomous driving, and 3D reconstruction.

[0005] According to some embodiments, the present invention adopts the following technical solution: A point cloud denoising method based on a fractional diffusion model includes: Obtain the noisy point cloud data to be denoised; Noisy point cloud data is input into the trained score prediction network to obtain a score vector; Based on the point cloud distribution and fractional vector of noisy point cloud data, the variance scheduling table is obtained by adaptively matching the denoising time steps and the corresponding variance subsequence. Based on the variance scheduling table, the current time step point cloud is started with noisy point cloud data. The position of the midpoint of the current time step point cloud is iteratively updated through the inverse sampling rule of the fractional diffusion model, gradually eliminating coupled noise and converging into a clean point cloud.

[0006] According to some embodiments, the present invention adopts the following technical solution: A point cloud denoising system based on a fractional diffusion model includes: The point cloud acquisition module is configured to acquire noisy point cloud data to be denoised. The score prediction module is configured to input noisy point cloud data into the trained score prediction network to obtain a score vector. The scheduling matching module is configured to: based on the point cloud distribution and fraction vector of noisy point cloud data, adaptively match the denoising time steps and the corresponding variance subsequence to obtain the variance scheduling table; The point cloud denoising module is configured to: based on the variance scheduling table, take the noisy point cloud data as the initial point cloud of the current time step, and iteratively update the position of the point in the point cloud of the current time step through the inverse sampling rule of the fractional diffusion model, gradually eliminating coupled noise and converging into a clean point cloud.

[0007] According to some embodiments, the present invention adopts the following technical solution: A computer program product includes a computer program that, when executed by a processor, implements the point cloud denoising method based on a fractional diffusion model.

[0008] According to some embodiments, the present invention adopts the following technical solution: A non-transitory computer-readable storage medium is provided for storing computer instructions, which, when executed by a processor, implement the point cloud denoising method based on a fractional diffusion model.

[0009] According to some embodiments, the present invention adopts the following technical solution: An electronic device includes a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to enable the electronic device to implement the point cloud denoising method based on a fractional diffusion model.

[0010] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention first models a fractional diffusion model for noisy point clouds for point cloud denoising tasks. This model includes a point cloud noise injection process based on mean-preserving forward diffusion and iterative denoising using fraction-driven deterministic backsampling. Based on this, a weighting term related to the noise variance scale is introduced into the objective function to prevent geometric drift and surface shrinkage during iterative denoising, and to stably preserve boundaries and details. Accordingly, high-quality noise removal and potential geometric surface restoration are achieved under diverse noise intensities (i.e., noise variance) and complex shapes, significantly improving denoising accuracy and stability.

[0011] 2. This invention solves the problem of manual parameter tuning required for different noise modes and levels: by estimating the noise variance from the fractional vector of the noise point cloud, and adaptively matching the time step and variance subsequence from the training variance table, a dynamic iterative schedule based on samples is formed, i.e., the variance scheduling table; combined with parallel inference of the overlapping slice generation technology and the reliable splicing strategy based on the distance center, it can balance efficiency and robustness on large-scale, non-uniformly distributed point clouds; therefore, it can achieve stable convergence without manual settings and exhibits better generalization ability.

[0012] 3. The score prediction network of this invention achieves accurate restoration of thin structures and local details through feature fusion and gradient fusion: at the feature level, it fuses the current iteration point cloud, the original noisy point cloud, and the time step embedding to capture more reliable geometric priors; at the gradient level, it performs weighted fusion on the neighborhood to suppress uncertain directions; and it adopts a two-stage sampling training strategy to align the inference trajectory and explicitly utilize prior information and local context, thereby effectively reducing outlier residues and over-smoothing, improving the restoration quality of thin structures and overall geometric consistency; this invention is simple to implement, easy to interface with existing 3D reconstruction algorithms and pipelines, and suitable for deployment and iteration in real-world applications. Attached Figure Description

[0013] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0014] Figure 1 This is a diagram illustrating the point cloud denoising process in Example 1; Figure 2 This is a schematic diagram of the forward diffusion process and the reverse sampling process in Example 1. Detailed Implementation

[0015] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0016] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0017] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0018] Example 1 3D point clouds, as an efficient form of 3D data representation, can directly depict the spatial position, shape, and scale of objects and their environment, playing a core role in scenarios such as robot navigation, autonomous driving, virtual reality, and high-quality 3D reconstruction. Point clouds possess advantages such as simple structure, efficient storage, geometric computation friendliness, and permutation invariance, and have been widely applied in tasks such as target detection, 3D reconstruction, and registration. However, due to sensor errors, surface reflections, and environmental interference, actual point clouds often contain noise and outliers of varying types and intensities; simultaneously, the disorder, unstructured nature, and uneven sampling of point clouds increase processing difficulty. To ensure the stability and accuracy of subsequent 3D tasks, it is crucial to accurately recover the underlying geometric surface under high-noise conditions, achieving effective denoising and shape detail preservation of the original point cloud.

[0019] Traditional geometric filtering and optimization methods often rely on heuristic assumptions such as smoothness or local structure, making them sensitive to hyperparameters and prone to imbalances between denoising and detail preservation, resulting in over-smoothing or surface degradation. Data-driven learning methods typically estimate the displacement from points to the potential surface through a single forward regression, but under complex shapes and diverse noise conditions, they are still prone to outlier retention, surface shrinkage, and loss of detail. Multi-round iterative denoising to improve performance often fails to match the actual iterative process during the training phase, making it difficult for the network to perceive preceding states, correctly estimate shape features, and update with an appropriate step size. Simultaneously, fixed iterative scheduling and step size settings are ill-suited to different shapes and noise levels, leading to insufficient generalization to real scanned data.

[0020] Therefore, one embodiment of the present invention provides a point cloud denoising method based on a fractional diffusion model, such as... Figure 1 As shown, it includes: Step S1: Obtain the noisy point cloud data to be denoised; Step S2: Input the noisy point cloud data into the trained score prediction network to obtain the score vector; Step S3: Based on the point cloud distribution and fractional vector of the noisy point cloud data, adaptively match the denoising time steps and the corresponding variance subsequence to obtain the variance scheduling table; Step S4: Based on the variance scheduling table, using the noisy point cloud data as the initial point cloud of the current time step, the position of the point in the current time step point cloud is iteratively updated through the inverse sampling rule of the fractional diffusion model, gradually eliminating coupled noise and converging into a clean point cloud.

[0021] As an example, this invention provides a point cloud denoising method based on a fractional diffusion model, detailing the specific implementation process from two aspects: the fractional diffusion model and the fractional prediction network.

[0022] I. Fractional Diffusion Model The score prediction network in the point cloud denoising method of this embodiment is built based on the score diffusion model, such as... Figure 2 As shown, the fractional diffusion model has two processes: forward diffusion (i.e., adding noise) and backward sampling (i.e., denoising). First, in the forward diffusion modeling, a parameterized form that preserves the distribution mean is adopted to eliminate the scaling of the original points and suppress the potential geometric drift caused by the advancement of time steps from the source. Then, in the backward sampling stage, the fractional vector is used as the update direction, and a deterministic iterative process is selected to ensure that the boundary and details are stably preserved during the sampling process.

[0023] This embodiment establishes a fractional diffusion model for point cloud denoising scenarios, where point cloud data is represented as a set of three-dimensional coordinate points. , where each point Considering the multi-source noise introduced during acquisition, transmission, and processing, the observed point cloud is modeled as follows:

[0024] in, This represents a clean point cloud located on a potential geometric surface. This is the noise that is coupled with it.

[0025] For ease of parameterization and analysis, preferably, the noise satisfies the Gaussian distribution assumption. The denoising task aims to reduce noise from the input. Identify and remove noise ,recover This will provide a precise data foundation for subsequent tasks such as 3D reconstruction, registration, and detection.

[0026] Based on this, time step sequences are introduced. The diffusion and sampling process is described. A clean point cloud is defined as the initial state. The forward diffusion process involves gradually injecting random noise, thereby... Shifting to a noise-enhanced point cloud sequence ,in This represents the state with the strongest coupling noise. To avoid the issues encountered in traditional diffusion modeling... To address the potential surface drift caused by scaling, this embodiment adopts the principle of maintaining the distribution mean during both the forward and reverse processes, aligning the point cloud distribution mean at each time step with the clean point cloud, thereby suppressing geometric shifts in iterative denoising.

[0027] Reverse sampling process to observe noisy point clouds The corresponding state at a certain time step As input, a neural network is used to estimate a fractional vector related to the noise distribution, and combined with an inverse iterative rule, the vector is recovered sequentially from large to small time steps until the desired result is obtained. That is, the denoising result By employing the mean-preserving diffusion and sampling modeling described above, we can achieve stable recovery of details and boundaries while suppressing geometric drift, providing a standardized modeling foundation for setting specific forward parameterization, backward update strategies, and training objectives in subsequent steps.

[0028] 1. Forward diffusion process.

[0029] Clean point cloud for input This paper presents a class of forward diffusion processes that maintain the mean distribution. To this end, a training variance table is introduced. ,in For time steps The variance scheduling step size is defined, and the following is defined: With cumulative amount .

[0030] At each time step This embodiment directly adopts a conditional distribution that preserves the mean:

[0031] Its equivalent recursive sampling formula is:

[0032] This parameterization is applied at each step. Noise is injected into the conditional mean throughout the diffusion sequence. Above, the underlying geometric surface of the point cloud distribution is always aligned with the clean point cloud, ensuring that forward diffusion does not cause the underlying geometric surface of the point cloud distribution to drift. Accordingly, the forward joint distribution can be obtained according to the Markov chain:

[0033] Furthermore, based on the Gaussian additivity theorem, the noise point cloud can be obtained. Closed-conditional distribution of any point cloud in the middle:

[0034] Its equivalent direct sampling formula is:

[0035] in, Indicates time step The variance of the injection noise is relative. The cumulative noise variance; and A one-to-one mapping is established between the two, which facilitates the formulation of adaptive time-step scheduling based on the variance of input noise in subsequent steps.

[0036] To eliminate ambiguity, the differences between the general diffusion notation and that of this embodiment will now be explained. The general notation typically uses:

[0037] in, As a multiplicative factor Before, the conditional mean will be scaled to This scaling accumulates over time, manifesting as a systematic shift in the distribution mean as diffusion progresses, which in turn induces shifts and shrinkage of the underlying geometric surface. In contrast, this embodiment explicitly removes the scaling term for the previous or initial state, ensuring that the conditional mean at each time step is strictly anchored to the corresponding reference point cloud. This avoids distribution mean shifts from the forward stage, providing more reliable statistical conditions for stable surface recovery in the reverse denoising stage.

[0038] 2. Reverse sampling process.

[0039] Based on the mean-preserving modeling of forward diffusion, the backsampling process uses the observed noisy point cloud sequence as a condition to recover the point cloud in descending order of time steps. ,in A clean point cloud located on a potential geometric surface.

[0040] The reverse process also uses Markov chain parameterization:

[0041] Furthermore, step sampling is generally permitted in writing:

[0042]

[0043] Among them, the mean By parameters Score prediction network It was estimated; ,variance It can be obtained from the forward training variance table The closed relation is determined. Used to control whether random noise is injected into the sampling; in the above formula Allow this implementation to estimate the time step The previous arbitrary noisy point cloud.

[0044] To meet the accuracy requirements of point cloud denoising, this embodiment presents a reverse update form that strictly matches the mean-preserving forward model.

[0045] First, the learning objective of the network is set as the gradient of the log-likelihood function of the point cloud distribution. Its direction points towards the potential geometric surface and can serve as the deterministic update direction for backsampling. Based on the fractional diffusion theory and the mean preservation matching condition, the explicit iteration of backsampling can be written as:

[0046] in, For optional random disturbance terms; mean By parameters Score prediction network It is estimated. Preferably, let A completely deterministic reverse sampling process can be obtained:

[0047] This avoids introducing additional noise at each step that could disturb geometric details, improving the stability of boundary and detail recovery; when enhanced exploration or alignment with generative processes is required, this approach can be adopted. A random perturbation of the same magnitude as traditional diffusion is introduced. The aforementioned step parameters... It can be flexibly set according to speed and accuracy requirements: It is beneficial for fine-grained recovery. It can accelerate iteration.

[0048] Through this reverse sampling process, the network only needs to predict the score vector. Within a statistical framework consistent with the forward distribution, the positions of points are updated iteratively in decreasing steps over time, gradually eliminating coupling noise and converging into a clean point cloud. .

[0049] II. Score Prediction Network Based on the fractional diffusion model, an adaptive iterative denoising strategy and a corresponding efficient architecture are proposed for input point clouds with different noise variances and morphologies. This strategy enables a denoising process that dynamically schedules data based on samples and progressively approximates the underlying geometric surface. Specifically: 1. Neural network architecture.

[0050] A score prediction network for point cloud denoising, using noisy point clouds. As input, output the fraction vector of the corresponding point. The network consists of a feature extraction module. Feature fusion module Gradient prediction module It consists of a gradient fusion module, which uses a fusion function. The gradients of multiple neighborhood predictions are weighted and integrated to obtain stable and accurate score estimates. This architecture is compatible with score diffusion modeling and supports denoised inference that is iterative on a time-step basis.

[0051] (1) Feature extraction module, preferably using a point-voxel hybrid convolutional neural network (PVCNN) as the feature extraction module. The backbone is adapted to point cloud inputs of varying shapes and densities. PVCNN performs sparse 3D convolutions by voxelizing the point cloud, while preserving point-level local feature awareness. It aggregates contextual information through fusion units of point and voxel features to form multi-scale point feature representations. To start from the current time point cloud Extracted features To extract from the original noisy point cloud Extracted features. This embodiment is not limited to PVCNN; it can also use PointNet++, attention-based point cloud Transformer, or other backbone networks. .

[0052] (2) Feature fusion module, Using the current position and relative time step information as a link, the representations of the current point cloud and the original noisy point cloud are aligned and fused. A learnable dictionary retrieval is introduced to improve robustness in high-noise scenes, thereby explicitly extracting geometric priors related to the potential surface. The fused features are expressed by the formula:

[0053] Specifically: First, Position encoding With relative time step encoding Integration into Embedded Then, respectively with and Fusion and This can be expressed as a formula:

[0054] in, This is called concatenated encoding, which combines two codes. For a three-layer multilayer perceptron with fusion splicing encoding.

[0055] Then, a learnable dictionary is introduced. ,in For query key, This is the query prior value corresponding to the key. (In query encoding) Using this as input, soft retrieval is performed to obtain the prior geometric surfaces. :

[0056] in, This is the temperature coefficient.

[0057] Finally, through three layers of multilayer perceptrons Implement prior injection, feature alignment, and aggregation, and provide:

[0058] in, This represents element-wise multiplication. and As a linear regression mapping, the above embedding and fusion depths can be equivalently replaced or extended according to computing power and accuracy requirements.

[0059] (3) Gradient prediction module, For any point Predict its gradient vector, which represents the gradient vector formed by the neighboring points. and its point features Location of points within the supporting local gradient vector field The gradient vector at the given location, and simultaneously provide the prediction at... The validity weight at each point. Therefore, with... Neighborhood points and their corresponding fusion features As a condition, the query points are normalized to a local coordinate system and input together with the features:

[0060] in, To predict the gradient vector, For importance weights.

[0061] Preferably, It is implemented using an MLP containing four residual blocks to enhance nonlinear representation and training stability, and can also be replaced with equivalent residual depth or attention units.

[0062] (4) Gradient fusion module, in training and inference, will... Take as The actual location, and within the same time step Nearest neighbor search selects its local neighborhood to form a multi-path prediction. Preferably, it employs... To achieve efficient retrieval and obtain Individual and Point The nearest point, to obtain The multiple versions of the estimation are used, with the closest value measured by the Euclidean distance between the two points. Considering that thin structures such as thin plates and tubes are sensitive to the fusion strategy, and that simple averaging may induce surface degradation, a weighted gradient fusion function is constructed:

[0063] By applying SoftMax normalization to the weight set to suppress the influence of uncertain directions, the stability and robustness of the score estimation are improved while preserving details, thus yielding a score prediction for the entire point cloud:

[0064] The above network architecture, while maintaining statistical consistency, fully utilizes time-step information, the complementarity between current estimates and original observations, and the structural constraints of local neighborhoods. (Module depth, channel width, number of nearest neighbors) Both the embedding and other methods can be equivalently replaced and their parameters adjusted without departing from the spirit of this embodiment, in order to adapt to point cloud data with different resolutions and noise levels.

[0065] 2. Model training algorithm.

[0066] To train a neural network model incorporating feature fusion, this embodiment proposes a two-stage sampling and point cloud slice-based training algorithm. On one hand, compared to training on the entire point cloud, training on point cloud slices improves the network's generalization ability, especially when the training set contains only a small number of point clouds. On the other hand, training a diffusion model typically samples noisy data from random time steps of the diffusion process in each training iteration. However, the two-stage sampling strategy proposed in this embodiment aims to acquire point features from both the current noisy point cloud and the original noisy point cloud in each training iteration to meet the requirements of feature fusion.

[0067] The two-stage sampling and point cloud slice-based training algorithm proposed in this embodiment is as follows: In each training iteration, the method first initializes the point cloud slices, then performs two-stage sampling, and finally calculates the loss function.

[0068] Specifically, in the point cloud slicing initialization stage, the algorithm starts from the input clean and complete point cloud. A random slice of a local point cloud containing 1000 points is cut from the data. and create a slice point mask. This mask is used to mark the area within the center of the slice. One dot. Mask The loss function used for regularization is the loss function that is calculated only for masked points during training, to ensure that each point has enough local neighborhood information to support gradient prediction and gradient fusion in the current training iteration.

[0069] After point cloud slicing initialization, two-stage sampling is performed. In the first stage, sampling is performed from the time step of the diffusion process. Sampling noise point cloud And predict its score vector. In the second phase, based on renew The location is used to obtain another noisy point cloud. And further predict its score vector. It is worth noting that the sampling point cloud and The goal is to simulate the iterative update process during the inference phase. By simulating the inference of the original noisy point cloud and the current noisy point cloud, the network can learn effectively during the training process and obtain the trained score prediction network.

[0070] Based on the forward diffusion modeling and backsampling rules that maintain the distribution mean in the score diffusion model, a training objective that strictly matches the above two stages is given for the score prediction network, namely, the definition of the loss function; the basic starting point of the training is to minimize the negative log-likelihood of the clean point cloud, that is, in the data distribution... lower envoy To minimize this, combined with mean-preserving Markov chain expansion, we can obtain an upper bound objective centered on time-step conditional distribution registration: by aligning the true posterior conditional distribution with the inverse conditional distribution given by the model at each time step, we minimize the KL divergence between them, thus ensuring the consistency between the inverse chain and the forward chain. Accordingly, we have the following trainable objective, i.e., the theoretical loss function:

[0071] The summation symbol represents the time step. The relevant summation, sampling step size is .

[0072] To simplify and improve training efficiency, this embodiment randomly selects the diffusion time step in the loss function. This eliminates the chain diffusion cost while ensuring that more noise distributions can be predicted in each batch. Furthermore, this embodiment defines a... The relevant weighting term, namely the total noise scale coupled with the clean point cloud, encourages the network to focus more on denoising performance closer to the potential surface regions. In other words, when When the value is small, the weighting term is large, providing greater gradient backpropagation and thus improving the denoising accuracy near the potential surface. In summary, the loss function in this embodiment is:

[0073] in, These are parameters used to control the weighting terms, and preferably can be taken as follows: ; It is the score vector estimated by the network, and It is the actual calculated score vector.

[0074] The loss function described above uses mean squared error (MSE) loss to measure the predicted score vector. and Compared with the true score vector and The error between them. This embodiment utilizes the noise point. To Clean Point Cloud The displacement vector of the nearest neighbor is used to approximate the true fraction vector. :

[0075] in, Return to point cloud Center and point cloud The nearest neighbor of each point in the array.

[0076] Optionally, this embodiment assumes that the noise variance does not exceed 3%, i.e. In the training, this embodiment sets the number of diffusion steps to [number]. Furthermore, in order to learn more accurate score vectors in low-density regions, this embodiment uses a linear scheduling strategy as the training variance table. ,in and set To ensure .

[0077] 3. Adaptive iterative noise reduction.

[0078] After training is complete, for a given noisy point cloud First, a pre-trained score prediction network is used to predict score vectors, which in turn estimate the noise variance. Then, matching subsequences are selected from the training variance table from the training phase to form an adaptive iterative denoising schedule, i.e., a variance scheduling table. The positions of the points are then updated step by step according to the inverse sampling formula of the score diffusion model, ultimately restoring a clean point cloud. .

[0079] To improve adaptability to large-scale and non-uniform point distributions, preferably, the farthest point sampling algorithm is used. upper structure Three overlapping local slices, denoted as Each slice performs iterative denoising independently and in parallel in subsequent steps. The point cloud slice construction method based on the farthest point sampling algorithm is a common and general point cloud slice construction method. The farthest point sampling algorithm is a commonly used algorithm for point cloud downsampling.

[0080] Based on the score vector predicted by the score prediction network, the noise variance is estimated and used to characterize the noise intensity. In this embodiment, the score predictions from all slices are concatenated to obtain the score vector of the entire noise point cloud. And using the variance of the fractional vector as a preliminary estimate of the global noise level, the noise variance is obtained:

[0081] The training variance table has been provided during the training phase. The corresponding cumulative noise variance is To match the inference trajectory with the noise variance of the noisy point cloud data to be denoised, adaptive scheduling selects a time step from the training variance table. , making The above formula estimates Most closely related; preferably, binary search is used in Solve above

[0082] Finally, determine the corresponding time step. Subsequently, with Linear division of intervals L denoising nodes are formed, each corresponding to a time step. Each interpolation node is quantized to the nearest time step in the training variance table to obtain the final adaptive denoising schedule. To improve reasoning efficiency while maintaining detail stability, the preferred method is... .

[0083] After determining the adaptive scheduling variance table, this embodiment applies it to each point cloud slice. The iterative denoising process is executed sequentially, and point cloud slices are updated according to the same adaptive scheduling variance table. The denoising process for each point cloud slice is independent and follows a unified iterative process, ensuring that multiple point cloud slices can undergo denoising simultaneously and in parallel. This embodiment uses noisy point cloud slices... As input, output the denoising result of the slice. In the iterative denoising process, this embodiment uses the inverse sampling process in the fractional diffusion model to perform the fractional vector estimation process using the trained fractional prediction network, and then updates the noisy point cloud slices according to the inverse sampling formula.

[0084] Finally, this embodiment slices all the denoised point cloud data. stitched together to form a complete denoised point cloud As output.

[0085] Considering the overlap between slices, this embodiment employs a reliability selection strategy. Specifically, if a point appears in multiple point cloud slices, this embodiment selects the slice where the point is closest to the slice center and uses the denoising result from that slice as the final location. This method effectively reduces errors during slice stitching, ensuring a complete denoised point cloud in the final result. Maintain a smooth and consistent shape.

[0086] The core of the entire process lies in adaptive scheduling and reasonable splicing between slices. These strategies ensure the efficiency and accuracy of point cloud denoising, while improving the reliability and robustness of point cloud processing.

[0087] Example 2 One embodiment of the present invention provides a point cloud denoising system based on a fractional diffusion model, comprising: The point cloud acquisition module is configured to acquire noisy point cloud data to be denoised. The score prediction module is configured to input noisy point cloud data into the trained score prediction network to obtain a score vector. The scheduling matching module is configured to: based on the point cloud distribution and fraction vector of noisy point cloud data, adaptively match the denoising time steps and the corresponding variance subsequence to obtain the variance scheduling table; The point cloud denoising module is configured to: based on the variance scheduling table, take the noisy point cloud data as the initial point cloud of the current time step, and iteratively update the position of the point in the point cloud of the current time step through the inverse sampling rule of the fractional diffusion model, gradually eliminating coupled noise and converging into a clean point cloud.

[0088] Example 3 One embodiment of the present invention provides a computer program product, including a computer program that, when executed by a processor, implements the point cloud denoising method based on a fractional diffusion model.

[0089] Example 4 In one embodiment of the present invention, a non-transitory computer-readable storage medium is provided for storing computer instructions, which, when executed by a processor, implement the point cloud denoising method based on a fractional diffusion model.

[0090] Example 5 One embodiment of the present invention provides an electronic device, including: a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to enable the electronic device to implement the point cloud denoising method based on the fractional diffusion model.

[0091] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0092] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0093] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A point cloud denoising method based on a fractional diffusion model, characterized in that, include: Obtain the noisy point cloud data to be denoised; Noisy point cloud data is input into the trained score prediction network to obtain a score vector; The score prediction network consists of a feature extraction module, a feature fusion module, a gradient prediction module, and a gradient fusion module. The feature extraction module uses a point-voxel hybrid convolutional neural network to aggregate and propagate neighborhoods in the feature space, and stacks multiple layers of point-voxel hybrid convolutions to aggregate contextual information and form multi-scale point feature representations. The feature fusion module, using the current position and relative time step information as a link, aligns and fuses the representations of the current point cloud with the original noisy point cloud, and incorporates a learnable dictionary for retrieval; the learnable dictionary Including query key The query prior value corresponding to the key , with query code Using this as input, soft retrieval is performed to obtain the prior geometric surfaces. : in, For temperature coefficient, for Transpose of; Noise point cloud Position encoding With relative time step encoding Integration into Embedded Then, respectively with and Fusion and This can be expressed as a formula: in, This is called concatenated encoding, which combines two codes. To integrate a three-layer, multi-layer perceptron with splicing encoding. To start from the current time point cloud Extracted features To extract from the original noisy point cloud Extracted features; Finally, through three layers of multilayer perceptrons Implement prior injection, feature alignment and aggregation, and provide fused features: in, This represents element-wise multiplication. and These are the coefficients of the linear regression mapping; The gradient prediction module uses the neighborhood points and their corresponding fusion features in the noisy point cloud data as conditions, normalizes the query point to the local coordinate system and inputs it together with the features, and predicts the gradient vector and effectiveness weight of any point in the local field. The gradient fusion module performs weighted fusion of the gradient vectors of all points based on the effectiveness weights to obtain the score vector of the noisy point cloud data. Based on the point cloud distribution and fractional vector of noisy point cloud data, the variance scheduling table is obtained by adaptively matching the denoising time steps and the corresponding variance subsequence. Based on the variance scheduling table, the current time step point cloud is started with noisy point cloud data. The position of the midpoint of the current time step point cloud is iteratively updated through the inverse sampling rule of the fractional diffusion model, gradually eliminating coupled noise and converging into a clean point cloud.

2. The point cloud denoising method based on a fractional diffusion model as described in claim 1, characterized in that, The adaptive matching denoising time steps and the corresponding variance subsequences are used to obtain the variance scheduling table, specifically: Based on fractional vectors, estimate the noise variance of noisy point cloud data; With the goal of matching the noise variance during the reverse sampling iteration process with the estimated noise variance, a binary search is used on the pre-constructed training variance table to determine the start and end time steps. The interval is linearly divided into a preset number of iteration nodes and quantized to the nearest training time step to obtain the final variance scheduling table.

3. The point cloud denoising method based on a fractional diffusion model as described in claim 2, characterized in that, The training variance table is formed by determining the maximum variance scheduling step size based on a preset noise variance threshold and the number of diffusion steps, and using a linear scheduling strategy to calculate the variance scheduling step size for each diffusion step.

4. The point cloud denoising method based on a fractional diffusion model as described in claim 1, characterized in that, The training of the score prediction network is based on the forward diffusion process and backward sampling process of the score diffusion model, employing a two-stage sampling training algorithm, specifically: In the first stage, noise point clouds are sampled from the time steps of the forward diffusion process, and their fractional vectors are predicted. In the second stage, the position of the noise point cloud is updated based on the predicted score vector to obtain a new noise point cloud, and the score vector of the new noise point cloud is further predicted.

5. The point cloud denoising method based on a fractional diffusion model as described in claim 1, characterized in that, The training loss function is based on mean squared error loss, which minimizes the error between the predicted score vector and the true score vector, and introduces a weighting term related to the noise variance scale, so that the model pays more attention to the denoising performance of the parts close to the potential surface in iterative denoising.

6. A point cloud denoising system based on a fractional diffusion model, characterized in that, include: The point cloud acquisition module is configured to acquire noisy point cloud data to be denoised. The score prediction module is configured to input noisy point cloud data into the trained score prediction network to obtain a score vector. The score prediction network consists of a feature extraction module, a feature fusion module, a gradient prediction module, and a gradient fusion module. The feature extraction module uses a point-voxel hybrid convolutional neural network to aggregate and propagate neighborhoods in the feature space, and stacks multiple layers of point-voxel hybrid convolutions to aggregate contextual information and form multi-scale point feature representations. The feature fusion module, using the current position and relative time step information as a link, aligns and fuses the representations of the current point cloud with the original noisy point cloud, and incorporates a learnable dictionary for retrieval; the learnable dictionary Including query key The query prior value corresponding to the key , with query code Using this as input, soft retrieval is performed to obtain the prior geometric surfaces. : in, For temperature coefficient, for Transpose of; Noise point cloud Position encoding With relative time step encoding Integration into Embedded Then, respectively with and Fusion and This can be expressed as a formula: in, This is called concatenated encoding, which combines two codes. To integrate a three-layer, multi-layer perceptron with splicing coding. To start from the current time point cloud Extracted features To extract from the original noisy point cloud Extracted features; Finally, through three layers of multilayer perceptrons Implement prior injection, feature alignment and aggregation, and provide fused features: in, This represents element-wise multiplication. and These are the coefficients of the linear regression mapping; The gradient prediction module uses the neighborhood points and their corresponding fusion features in the noisy point cloud data as conditions, normalizes the query point to the local coordinate system and inputs it together with the features, and predicts the gradient vector and effectiveness weight of any point in the local field. The gradient fusion module performs weighted fusion of the gradient vectors of all points based on the effectiveness weights to obtain the score vector of the noisy point cloud data. The scheduling matching module is configured to: based on the point cloud distribution and fraction vector of noisy point cloud data, adaptively match the denoising time steps and the corresponding variance subsequence to obtain the variance scheduling table; The point cloud denoising module is configured to: based on the variance scheduling table, take the noisy point cloud data as the initial point cloud of the current time step, and iteratively update the position of the point in the point cloud of the current time step through the inverse sampling rule of the fractional diffusion model, gradually eliminating coupled noise and converging into a clean point cloud.

7. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements a point cloud denoising method based on a fractional diffusion model as described in any one of claims 1-5.

8. A non-transitory computer-readable storage medium, characterized in that, The non-transitory computer-readable storage medium is used to store computer instructions, which, when executed by a processor, implement a point cloud denoising method based on a fractional diffusion model as described in any one of claims 1-5.

9. An electronic device, characterized in that, include: The device includes a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to enable the electronic device to implement a point cloud denoising method based on a fractional diffusion model as described in any one of claims 1-5.

Images