A point cloud generation method for flow diffusion model based on neighborhood feature enhancement

Abstract

The present invention discloses a method for generating a flow diffusion point cloud based on neighborhood feature enhancement, comprising: preprocessing the original point cloud; constructing a multi-scale neighborhood feature aggregation module, which obtains the neighborhood coordinates of each point through nearest neighbor sampling at different scales and aggregates these coordinates into coarse global features; constructing a neighborhood attention feature enhancement module, which uses the geometric information in the neighborhood coordinate space to enhance the coarse features in the feature space; using a point voxel convolutional neural network to reduce redundant features in the enhanced features and output the latent vector of the point cloud; designing a feature-to-flow data conversion module, which converts the latent vector into a data-consistent prior flow feature and seamlessly integrates it into a denoising diffusion model for accurate generation from noisy point clouds. This patent effectively generates high-quality 3D point clouds through the combination of neighborhood feature enhancement and flow diffusion.

Description

Technical field:

[0001] The present invention relates to the fields of three-dimensional generation and computer vision, and in particular to a flow diffusion point cloud generation method based on neighborhood feature enhancement. Background technology:

[0002] Point cloud data enhances multimedia applications by providing detailed visual and spatial information, increasing user immersion and interactivity. Its use in virtual reality, augmented reality, and mixed reality can create realistic 3D environments. However, acquiring point cloud data is time-consuming and sensor-dependent. Therefore, it is necessary to apply artificial intelligence methods to advance point cloud generation technology to provide high-fidelity data support for the multimedia field.

[0003] In recent years, deep learning-based methods have become the mainstream for 3D point cloud generation. This process typically involves several key steps. The first step is data preprocessing, including denoising, normalization, and data augmentation to ensure consistency and diversity of model input. Next is feature extraction, where deep neural networks, such as convolutional neural networks, extract high-dimensional features from the point cloud, capturing the complex relationships and geometric structures between points. The third step involves the design and training of generative models, such as variational autoencoders, generative adversarial networks, normalized flow models, or autoregressive models, each of which employs a different generative approach to generate new point clouds.

[0004] Closest existing technology and its evaluation:

[0005] Among the closest existing technologies, some scholars have proposed the DualGAN algorithm (Wen C, Yu B, Tao D. Learning progressive point embeddings for 3D point cloud generation [C] / / Proceedings of the IEEE / CVF Conference on Computer Vision and Pattern Recognition. 2021: 10266-10275.), which uses two adversarial generative network generators to perform advanced upsampling and then refine the point cloud. The PointFlow algorithm (Yang G, Huang X, Hao Z, et al. Pointflow: 3D point cloud generation with continuous normalizing flows [C] / / Proceedings of the IEEE / CVF international conference on computer vision. 2019: 4541-4550.) uses a 3D point cloud model with continuous normalizing flows and variational reasoning, and its advantage lies in its stability compared to adversarial generative networks. The SoftFlow algorithm (Kim H, Lee H, Kang WH, et al. Softflow: Probabilistic framework for normalizing flow on manifolds[J]. Advances in Neural Information Processing Systems, 2020, 33: 16388-16397) trains normalized flows on manifolds by conditionally distributing perturbed input data, thus avoiding the dimensionality mismatch problem. The SetVae algorithm (Kim J, Yoo J, Lee J, et al. Setvae: Learning hierarchical composition for generative modeling of set-structured data[C] / / Proceedingsof the IEEE / CVF Conference on Computer Vision and Pattern Recognition. 2021: 15059-15068.) generates diverse and complex point clouds by using probabilistic graphical models and variational inference.The RPG algorithm (KolW J, Chiu CY, Kuo YL, et al. RPG: Learning recursive point cloud generation [C] / / 2022 IEEE / RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2022: 544-551.) generates 3D point clouds from coarse to fine resolution by recursively expanding points. It also automatically discovers semantic parts of the target model and achieves unsupervised segmentation during the generation process. The PVD algorithm (Zhou L, DuY, Wu J. 3D shape generation and completion through point-voxel diffusion [C] / / Proceedings of the IEEE / CVF international conference on computer vision. 2021: 5826-5835.) generates point cloud data from Gaussian noise through a series of denoising operations by combining a denoising diffusion model with a mixed point-voxel representation. The TGN algorithm (Xu R, Hui L, Han Y, et al. Transformer-based Point Cloud Generation Network [C] / / Proceedings of the 31st ACM International Conference on Multimedia. 2023: 4169-4177.) combines local and global information through a Transformer-based interpolation module and refinement module to gradually upsample point cloud features and ultimately generate realistic 3D point clouds. Although these methods have achieved good results, they typically use hierarchical convolutional neural networks or shared full connections for feature extraction, which may ignore some neighborhood features when aggregating spatial features and introduce uncertainty and errors in step-by-step denoising. This can affect the complete recovery of details and reduce the quality of the generated point cloud, especially for complex geometric structures. Summary of the invention:

[0006] The purpose of the present invention is to provide a flow diffusion point cloud generation method based on neighborhood feature enhancement. Based on the problem that the existing technology ignores the neighborhood features of some point clouds, introduces uncertainty and errors in the generation process, and thus leads to a reduction in the quality of some generation, a multi-scale neighborhood feature aggregation module is constructed. This module uses K-nearest neighbor sampling at different scales to obtain the neighborhood coordinates of each point, thereby aggregating them into coarse global features. Secondly, a feature enhancement module based on neighborhood attention is constructed. This module uses the geometric information in the coordinate space through the attention mechanism neighborhood to enhance the coarse features in the feature space. Then, a point voxel convolutional neural network is used to reduce the redundant features in the enhanced features and output the latent vector of the point cloud. Finally, a feature to flow data conversion module is designed to convert the latent vector into data-consistent prior flow features, which are seamlessly integrated into the denoising diffusion model for accurate generation from noisy point clouds. This prior flow method improves the consistency and coherence of the point cloud density distribution.

[0007] The steps include:

[0008] Step 1: Preprocess the original point cloud to be encoded, including normalizing the point cloud data to [-1, 1] and dividing it into training set, test set and validation set;

[0009] Step 2: Construct a multi-scale neighborhood feature aggregation module, which includes a farthest point sampling component, multiple multi-scale K nearest neighbor components and multiple coarse feature extraction components;

[0010] Step 3: Construct a neighborhood attention feature enhancement module, which enhances the global coarse features obtained in step 2 through self-attention mechanism and farthest point sampling;

[0011] Step 4: Use the point voxel convolutional neural network to reduce the redundant features in the enhanced features output in step 3 and output the potential vector of the point cloud;

[0012] Step 5: Design a feature-to-stream data conversion module, which maps the latent vector output from step 4 to stream data through an affine transformation to make the sampling probability of the point cloud continuous.

[0013] Step 6: Input the flow data in step 5 into the diffusion model to generate point cloud;

[0014] Step 7: Compare the generated point cloud obtained in step 6 with the input point cloud and calculate the loss;

[0015] Step 8: Based on the total loss obtained in step 7, use the Adam optimizer to optimize the point cloud generation method end-to-end;

[0016] Step 9: Loop through steps 2 to 8, train for a specified number of rounds, until the point cloud generation network of the flow diffusion model with enhanced neighborhood features converges, and save the optimal training model file.

[0017] Step 10: Select a single object or multiple object categories for training based on actual usage needs and output the object type to be generated.

[0018] By adopting the above technical solution, the beneficial effects of the present invention are:

[0019] 1. This paper proposes a scale neighborhood feature aggregation module and a neighborhood attention feature enhancement module to extract and enhance point cloud features, providing reliable prior information for the diffusion model.

[0020] 2. This paper proposes a novel feature-to-stream data conversion module, which uses standardized stream mapping to stabilize noise sampling in the diffusion model denoising process, thereby improving the quality of the generated point cloud. Description of the drawings:

[0021] Figure 1 This is the basic flow chart of the flow diffusion point cloud generation method based on neighborhood feature enhancement of the present invention.

[0022] Figure 2 It is a structural diagram of the point cloud generation network of the present invention.

[0023] Figure 3 This is a comparison chart of shape metrics with existing technologies.

[0024] Figure 4 It is a process diagram of point cloud generation by the method of the present invention.

[0025] Figure 5 It is a comparison chart with the objective evaluation metrics of the existing technology.

[0026] Figure 6 It is a visualization diagram of the feature structure of multi-item joint training.

[0027] Figure 7 This is the generated graph of the few-sample point cloud in the joint training of multiple objects in the present invention.

[0028] Figure 8 It is the completion map of the present invention on the missing point cloud data.

[0029] Figure 9 This is a diagram showing the ablation experiment results of the three-dimensional component proposed in the present invention. Specific implementation method:

[0030] The technical solution provided by this application will be further described below in conjunction with specific embodiments and accompanying drawings. The advantages and features of this application will become more apparent with reference to the following description.

[0031] Example

[0032] Flow diffusion point cloud generation method based on neighborhood feature enhancement, Figure 1 As shown:

[0033] Step 1: Preprocess the original point cloud to be encoded, including normalizing the point cloud data to [-1, 1] and dividing it into training set, test set and validation set according to the ratio of 80%, 15% and 5% respectively;

[0034] Step 2: Build a multi-scale neighborhood feature aggregation module ( Figure 2 ), the multi-scale neighborhood feature aggregation module includes a farthest point sampling component (FPS), multiple multi-scale K nearest neighbor components (k-NN(.)) and multiple coarse feature extraction components (ξ Ψ ); Multiple multi-scale K nearest neighbor components are circular K nearest neighbors, with a certain point cloud center point as the circle center, and search for k adjacent points within a radius r around it. Multiple multi-scales are embodied in the present invention using two layers of circular K nearest neighbors, namely, a radius of 0.05, a k value of 16, and a radius of 0.1, a k value of 32; Each coarse feature extraction component is composed of a 1×1 convolution (Conv1D), batch normalization (BN), an activation function (ReLU) and a maximum pooling layer (Max Pooling);

[0035] Specifically, the farthest point sampling component firstly samples the input point cloud set X = {x1, x2, x3, ..., x N} select m points as the initial point set S = {x1, x2, x3, ..., x m}; For each point x in set X, it calculates the minimum distance to the points in set S. Then it selects the next point x m+1 , maximize its minimum distance from the center point of the set S; repeat this step until the required number of s center points is selected to obtain a new set S = {x1, x2, x3..., x s The formula for the entire farthest point sampling process is as follows:

[0036]

[0037] After the multi-scale K nearest neighbor component obtains the center point set, it performs K nearest neighbor search to establish the neighborhood relationship of each point; for each point p i ∈S, calculate its distance to other points and find the k nearest points and the previous set of neighbors p ijRepresents the distance p i The jth nearest point; this step finds the k nearest points for each center point in the original data X to extract neighborhood features, which helps to construct local features when processing point cloud data; the K nearest neighbor process can be expressed as the formula:

[0038]

[0039] The coarse feature component finally summarizes the K nearest neighbors of each center point to obtain a set of new point coordinates p. Ψ (·) Aggregate the point coordinates. This component uses 1×1 convolution (Conv1D), batch normalization (BN), activation function (ReLU) and maximum pooling layer (Max Pooling) to extract the significant information of local features and concatenate the neighborhood set into a single layer. Aggregation is the global rough feature F, which is expressed as follows:

[0040]

[0041] Step 3: Construct a neighborhood attention feature enhancement module to enhance the global coarse features obtained in step 2 through the self-attention mechanism and the farthest point sampling method.

[0042] Specifically, a linear layer is first used to map the global rough feature F to the high-dimensional initial feature map F′ of the point cloud; at the same time, for each point P in the input, the 16 nearest points are found according to its spatial position, and the index N of these 16 nearest points is obtained. k (p i ), the index formula is as follows:

[0043]

[0044] Then, three fully connected layers (MLPs) are used on the high-dimensional initial feature map F′ to map the attention key vector q, the value vector k, and the query vector v, respectively:

[0045] q=MLP Q (F)

[0046] k = MLP K (F)

[0047] v=MLP V (F)

[0048] Then, using the position encoding PE, each point p in P is grouped by a grouping operation. i The coordinates of and their K nearest neighbor index N k (p i ) coordinates to generate a new feature set and in The coordinates of P are subtracted point by point; the subtraction result is encoded by the designed position encoding fully connected layer (PEMLP), and the formula of the position encoding PE is as follows:

[0049]

[0050] By applying the same grouping operation used in the positional encoding PE above to the key vector and index N k (p i ) to reshape the query vector feature k′ and concatenate the query vector q with the key vector k′ and the position encoding PE. After processing by the designed multi-layer attention fully connected layer (AttentionMLP), the softmax function is used to calculate the attention weight a i , perform weighted summation on the value vector v and the position encoding PE to generate the final aggregate feature representation F agg ; Aggregate feature representation F agg Reflects the differences and similarities between the features of different points and their neighboring points; attention weight a i and the aggregate feature formula F agg As shown below:

[0051]

[0052] F agg =∑a i (ν+PE)

[0053] Get the aggregated feature F agg Then it is mapped back to the original feature space through full connection and added to the input feature through residual connection to obtain the enhanced feature representation F out , the process is shown in the following formula:

[0054] F out =MLP(F agg )+F′

[0055] Step 4: Introduce a point voxel convolutional neural network to reduce the redundant features in the enhanced features output in step 3, thereby outputting the potential vector of the point cloud; the point voxel convolutional neural network is an existing technology, and PVCNN is used as an example to implement it ( Figure 2 shown);

[0056] Step 5: Design a feature-to-stream data conversion module, which maps the latent vector output from step 4 to stream data through an affine transformation to make the sampling probability of the point cloud continuous.

[0057] Specifically, the flow model is first used to represent the given high-dimensional feature w. The model performs a transformation from time point t0 to t1 through continuous time integration, and finally obtains the potential vector z; the formula is expressed as:

[0058]

[0059] Where w represents the prior distribution parameter; f ε (·) is a neural network that represents the evolution function at time t that controls w over time; logP(F ε -1 (z)) is the probability density of the inverse transformation of z, that is, the density of the prior distribution after transformation; represents the rate of change from time t0 to t1, Tr represents the trace of the matrix. The purpose is to find the value corresponding to the prior distribution at time t0;

[0060] Then the target distribution is calculated. In order to generate the corresponding latent vector z from the data, the inverse transformation w(t0) is used to find the value corresponding to the initial distribution at time t0. The formula is:

[0061]

[0062] Finally, the probability density of the target distribution is expressed. Using the inverse mapping formula, the ultimate goal is to find the target distribution p(z). The density of the target distribution can be calculated by changing the Jacobian determinant. The formula is:

[0063]

[0064] Where p(z) represents the probability density of the latent variable z; p w (w) represents the probability density of random variable w; F ε is a transformation function that maps w to z; It's F ε The inverse of the absolute value of the Jacobian determinant of w is used to adjust the density scaling factor; F ε -1 (·) is F ε The inverse function of , used to get w from z.

[0065] The core of the above formula is to Figure 2 The neural network with affine coupled layers shown in

[15] learns a dynamic continuous transformation f θ , and derives the mapping from high-dimensional feature space to latent space through the flow model, while using the inverse transform and Jacobian determinant to calculate the final probability density; the method allows the use of continuous transformation models to generate target distributions.

[0066] The above process is specifically implemented through the affine coupling layer, which is composed of 12 MLP, RELU, MLP, RELU, MLP, and Sigmod layers connected in sequence.

[0067] Step 6: Input the flow data from step 5 into the diffusion model to generate point clouds. The flow data is used to replace the sampling process of the diffusion model. Because the flow model has deterministic stability and learns global point cloud features, it can optimize the cumulative error caused by the diffusion model sampling.

[0068] Specifically, the goal of a given diffusion model is to maximize the target point cloud X( 0 ), expressed as Since it is difficult to optimize this goal directly, this paper introduces the evidence lower bound (ELBO):

[0069]

[0070] The main reason for replacing the variational posterior distribution in KL divergence with the flow model is to improve the flexibility and expressiveness of variational inference. The variational inference of traditional diffusion models usually uses a simple distribution (such as Gaussian distribution) to approximate the posterior distribution, but this method is limited by the distribution assumption and cannot capture complex posterior structures. The flow model provides a more flexible and scalable way to learn complex posterior distributions. Substituting the formula in step 5 into And expand it to get the final optimization target formula:

[0071]

[0072] In this way, the objective function combines the KL divergence and log-likelihood to maximize the likelihood that the generative model will generate data while ensuring that the posterior distribution and the prior distribution are close; that is, the optimization goal is to train the parameters of the generative model by minimizing the KL divergence so that the model can generate samples consistent with the data distribution.

[0073] Step 7: Compare the generated point cloud obtained in step 6 with the input point cloud and calculate the total loss;

[0074]

[0075] Among them, kl_weight is a hyperparameter with a fixed value of 0.001, which is used to balance the proportion of reconstruction loss and KL divergence term; The entropy loss is used to measure the entropy of the latent variable distribution output by the encoder; The prior loss measures the logarithmic probability of the latent variable under the prior distribution and is calculated by the flow model; logP(X|z) is the reconstruction loss that measures the similarity between the sample x generated by the model and the input sample

[0076] Step 8: Based on the total loss obtained in Step 7, the Adam optimizer is used to optimize the point cloud generation method end-to-end. The initial learning rate of the optimizer is set to 0.002, and the weight decay is set to 0, that is, there is no regularization term. The maximum gradient norm of the gradient clipping is 10 to prevent gradient explosion. The learning rate scheduler will start to decay after 200,000 iterations and reduce the learning rate to the termination learning rate of 0.0001 at 400,000 iterations. These settings jointly control the parameter update speed and stability during the training process.

[0077] Step 9: Loop through steps 2-8, train for a specified number of rounds until the neighborhood feature-enhanced flow diffusion model point cloud generation network converges, and save the optimal training model file. The specified number of rounds is set to 1,000,000 in the experiment. The program can be interrupted in advance and the optimal training model file can be saved by observing the loss rate curve in real time through TensorBoard.

[0078] Step 10: Select a single object or multiple object categories for training based on actual usage needs and output the object type to be generated.

[0079] The embodiment of the present invention conducts a comparative experiment on the flow diffusion point cloud generation method based on neighborhood feature enhancement proposed by the present invention and the closest existing technology. The dataset used in the experiment is ShapeNetCore v2, which was jointly proposed by researchers from Princeton University, Stanford University and other institutions; it contains more than 50,000 point cloud data models, covering 57 categories; each model has a corresponding geometry file and a unique identifier; the present invention conducts experiments on three subsets of chairs, airplanes and cars in ShapeNet and compares shape measurement indicators with advanced algorithms; a series of state-of-the-art algorithms are selected as comparison benchmarks, which represent the diversity and latest progress in the field of 3D shape generation, including non-diffusion algorithms such as r-GAN, 1-GAN, PointFlow, SoftFlow, SetVAE and DPF-Net, and algorithms based on diffusion models such as DPM, MeshDiffusion, PVD, and specific data such as Figure 3 Shown; analysis Figure 3 The data in show that the performance of the proposed model is comparable to the most advanced methods in various indicators; this success is attributed to the comprehensive consideration of local and global information and the effective integration of spatial coordinates in the proposed model; in addition, the flow model stabilizes the diffusion generation process, significantly improving the quality and high-fidelity generation of 3D shapes in the generated point cloud data; to further verify the generation ability of the proposed model, Figure 4The visualization results of the denoising generation process for three different object categories, namely airplanes, chairs, and cars, are shown in Figure 2. These visualizations directly show the dynamic evolution of different objects from initial rough shapes to complete structures during the denoising process. In order to more comprehensively evaluate the performance of the model, it is compared with similar cutting-edge algorithms (P0intFlow and DPM) from a subjective evaluation perspective. The relevant visualization comparison results are shown in Figure 2. Figure 5 As shown in the figure; through these comparisons, we can not only observe the performance differences of the models in terms of generation accuracy and detail restoration, but also further reveal the adaptability and robustness of the models in dealing with different object categories; subjective evaluation provides an effective supplement to quantitative indicators (such as EMD, CD, etc.), comprehensively demonstrating the performance of the model in complex geometric structure generation and proving its potential in practical applications; combining subjective and quantitative dual evaluations can more comprehensively demonstrate the performance of the model in complex geometric structure generation.

[0080] The above examples demonstrate the effectiveness of the model for generating single-category objects, but joint training on multiple object classes remains to be explored. To address this issue, this study jointly trained a model on six different classes of ShapeNetCore v2 (airplane, chair, car, tower, bag, and basket) without any class conditions. To more intuitively demonstrate the model's performance in the multi-class object generation task, the t-distributed stochastic neighbor embedding (t-SNE) algorithm was used to reduce the dimensionality of the latent vectors generated by the model and project them onto a two-dimensional plane. The results are shown in Figure 2. Figure 6 As shown in , most data points of different categories show an obvious separation trend in two-dimensional space; data points of different object categories such as airplanes, chairs, cars, towers, bags and baskets are clustered in different areas, forming relatively independent clusters; this separation trend shows that the model can effectively capture and learn the features of objects of different categories and clearly distinguish them in the latent vector space; especially in the absence of category constraints, the model still shows a strong category discrimination ability, which shows that it has good adaptability to multi-category features; from the point cloud generated by the joint training, some small sample data are randomly selected for display, covering three categories of objects: towers, bags and baskets; each type of point cloud contains about 100 to 200 samples, and the number of test samples is about 11 to 24; the visualization results are shown in Figure 7As shown in the figure, the generated point cloud has uniform surface coverage and no obvious gaps while maintaining novelty and high fidelity; the experimental results verify the feasibility of multi-category joint training, especially in the task of generating few-sample point cloud data, which shows strong learning ability; therefore, this research method is not only suitable for generating high-quality single-category object point clouds, but also can successfully achieve multi-category joint generation without increasing complexity, avoiding the need to train a separate model for each category; this result provides new ideas and possibilities for the practical application of multi-category object generation.

[0081] In addition to the above examples, the present invention also explores the proposed model's upsampling capabilities with fewer input point clouds. During inference, the proposed model generates a series of point-to-point distance samples for each input point cloud, thereby completing the point cloud. Specifically, the experiment uses 150 and 200 points as input for each point set of an airplane and a car, respectively. Using the pre-trained model, the proposed model combines global shape variables with partial point cloud information to synthesize a complete point cloud. Figure 8 Qualitative visualization of the output point cloud is presented. Experimental results demonstrate that even when the input points are relatively sparse, the trained model can still accurately perform point cloud upsampling and generate a complete point cloud with high fidelity. This capability demonstrates the adaptability and generalization ability of the proposed method with a small amount of point cloud input, providing a reliable solution for upsampling and completion of point cloud data.

[0082] To demonstrate the effectiveness of the module design introduced in the present invention, this embodiment method conducted an ablation study and performed ablation tests on the constructed multi-scale neighborhood feature aggregation module (MNFA), neighborhood attention feature enhancement module (NAFE), and point voxel convolutional neural network (PVCNN). To verify the effectiveness of the present invention in multi-class training, three subsets (chairs, cars, and airplanes) were selected for joint training and tested on the airplane class. Each set of experiments was iterated for 200,000 steps. Figure 9 The quantitative results of the ablation experiments are summarized. By ablating each module, the necessity of each module in the model is demonstrated. The experimental results show that the model performs best when all modules are integrated. After removing any module, the performance drops significantly. In particular, similar numerical results are observed when PVCNN is removed and reintroduced, which further demonstrates the key role of the multi-scale neighborhood feature aggregation module (MNFA) and the neighborhood attention feature enhancement module (NAFE) in the model. These two modules work together on point cloud data to better capture the complexity of the geometric structure, thereby improving the effect of point cloud generation.

[0083] The above description is only a description of the preferred embodiments of the present application and does not limit the scope of the present application. Any changes or modifications made by any person skilled in the art based on the above disclosed technical content should be regarded as equivalent valid embodiments and fall within the scope of protection of the technical solution of the present application.

Claims

1. A flow diffusion point cloud generation method based on neighborhood feature enhancement, characterized in that: The steps include: Step 1: Preprocess the original point cloud to be encoded, including normalizing the point cloud data to [-1, 1] and dividing it into training set, test set and validation set; Step 2: Construct a multi-scale neighborhood feature aggregation module, which includes a farthest point sampling component, multiple multi-scale K nearest neighbor components and multiple coarse feature extraction components; Step 3: Construct a neighborhood attention feature enhancement module, which enhances the global coarse features obtained in step 2 through self-attention mechanism and farthest point sampling; Step 4: Use the point voxel convolutional neural network to reduce the redundant features in the enhanced features output in step 3 and output the potential vector of the point cloud; Step 5: Design a feature-to-stream data conversion module, which maps the latent vector output from step 4 to stream data through an affine transformation to make the sampling probability of the point cloud continuous. Step 6: Input the flow data in step 5 into the diffusion model to generate point cloud; Step 7: Compare the generated point cloud obtained in step 6 with the input point cloud and calculate the loss; Step 8: Based on the total loss obtained in step 7, use the Adam optimizer to optimize the point cloud generation method end-to-end; Step 9: Loop through steps 2 to 8, train for a specified number of rounds, until the point cloud generation network of the flow diffusion model with enhanced neighborhood features converges, and save the optimal training model file. Step 10: Select a single object or multiple object categories for training based on actual usage needs and output the object type to be generated; In step 3, a linear layer is first used to map F to a high-dimensional initial feature map F′ of the point cloud; at the same time, for each point P in the input, the 16 nearest points are found according to their spatial position, and the indexes N of these 16 nearest points are obtained. k (p i ), the index formula is as follows: Where p a and p b It's P i Then, three fully connected layers are used on the high-dimensional initial feature map F′ to map the attention key vector q, the value vector k, and the query vector v respectively: q=MLP Q (F′) k=MLP K (F′) v=MLP V (F′) Then, the position code PF groups each point p in P i The coordinates of and their K nearest neighbor index N k (p i ) coordinates to generate a new feature set and in The coordinates of P are subtracted point by point; the subtraction result is encoded by the designed position encoding fully connected layer. The position encoding formula is as follows: By applying the same grouping operation used in the positional encoding above to the key vector and index N k (p i ) to reshape the query vector feature k′ and concatenate the query vector q with the key vector k′ and the position encoding PE. After being processed by the designed multi-layer perceptron AttnMLP, the attention weight a is calculated using the softmax function. i , perform weighted summation on the value vector v and the position encoding PE to generate the final aggregate feature representation F agg ; F agg It reflects the differences and similarities between the features of different points and their neighboring points; the attention weight and aggregate feature formula are as follows: F agg =∑a i ·(v+PE) Get the aggregated feature F agg Then it is mapped back to the original feature space through full connection and added to the input feature through residual connection to obtain the enhanced feature representation F out , the process is shown in the following formula: F out =MLP(F agg )+F′。 2. The method for generating a flow diffusion point cloud based on neighborhood feature enhancement according to claim 1, characterized in that: Step 2: The farthest point sampling component first samples the input point cloud set X = {x1, x2, x3, ..., x N } select m points as the initial point set S={x1,x2,x3,…,x m }; For each point x in set X, it calculates the minimum distance to the points in set S and then selects the next point x m+1 , maximize its minimum distance from the center point of the set S; repeat this step until the required number of s center points is selected to obtain a new set S={x1,x2,x3…,x s The formula for the entire farthest point sampling process is as follows: Where x represents the point whose distance is to be calculated; S represents the selected point set; ‖xx j ‖ represents point x and point x in the set j The Euclidean distance between them; d(x,S) means finding the minimum distance between point x and all points in set S; X\S means the points in data set X that do not belong to set S; argmax means finding the point x that makes d(x,S) the largest; x i+1 It means finding the point farthest from the set S among the points that have not yet been added to the set S, and taking it as the next point to be selected; After the multi-scale K nearest neighbor component obtains the center point set, it performs K nearest neighbor search to establish the neighborhood relationship of each point; for each point p i ∈S, calculate its distance to other points and find the k nearest points and the previous set of neighbors p ij Represents the distance p i The jth nearest point; This step finds the k nearest points for each center point in the original data X to extract neighborhood features, which helps construct local features when processing point cloud data. The K nearest neighbor process can be expressed as the formula: Where s i represents the i-th center point in the center point set; For each center point s i Find the k nearest neighbor points in the point set X; Represents the k neighbor points found for a specific center point; The coarse feature component finally summarizes the K nearest neighbors of each center point to obtain a set of new point coordinates p; through the coarse feature component ξ with parameter Ψ Ψ (·) Aggregate the point coordinates. This component uses 1×1 convolution, batch normalization, ReLU and maximum pooling layers to extract significant information of local features and concatenate the neighborhood sets together through splicing operations. Aggregate into a rough global feature F, the formula is as follows: Where Contact represents the module ξ Ψ The output features are concatenated to obtain a rough global feature F.

3. The method for generating a flow diffusion point cloud based on neighborhood feature enhancement according to claim 1, characterized in that: Step 5: Convert discrete features into a continuous flow model and convert them from the form of data x into the form of latent vector z: Where w represents the prior distribution parameter; f ε (·) is a neural network that represents the evolution function at time t that controls w over time; logP(F ε -1 (z)) is the probability density of the inverse transformation of z, that is, the density of the prior distribution after transformation; Represents the rate of change from time t0 to t1, Tr represents the trace of the matrix. The purpose is to find the value corresponding to the prior distribution at time t0, so the inverse operation of the flow is used: The inverse transformation of the input is performed using affine coupled layers; each layer transforms part of the current input while keeping the other part unchanged, ensuring a one-to-one correspondence between z and w; this method allows the probability of the target distribution to be accurately calculated by applying the change of variable formula: Where p(z) represents the probability density of the latent variable z; p w (w) represents the probability density of random variable w; F ε is a transformation function that maps w to z; It's F ε The inverse of the absolute value of the Jacobian determinant of w is used to adjust the density scaling factor; F ε -1 (·) is F ε The inverse function of , used to get w from z.

Images