Point cloud denoising method and system based on residual enhancement and conditional modulation ScoreNet

Abstract

The application relates to a point cloud denoising method and system based on residual enhancement and conditional modulation ScoreNet, which comprises the following steps: obtaining three-dimensional noisy point cloud data and preprocessing the same to obtain standardized point cloud data; inputting the standardized point cloud data into a point cloud feature extraction network, obtaining an initial feature map after feature initialization, and outputting point cloud fusion features through multi-stage feature aggregation and local feature enhancement; constructing a hidden conditional vector, inputting the hidden conditional vector and the point cloud fusion features into a conditional modulation residual ScoreNet architecture, dynamically modulating a predicted noise gradient score through a conditional modulation residual block; and denoising and updating each point of the noisy point cloud according to the noise gradient score to obtain a pure point cloud. The application alleviates the gradient vanishing problem of a deep network through multi-stage feature aggregation and local enhancement, dynamically and adaptively modulates gradient prediction by taking the point cloud fusion features as a hidden conditional vector, overcomes the defect of insufficient prediction accuracy of a fixed conditional vector, and exhibits stable denoising effects under different noise levels.

Description

Technical Field

[0001] This invention relates to the field of 3D point cloud processing technology, and in particular to a point cloud denoising method and system based on residual enhancement and conditional modulation ScoreNet. Background Technology

[0002] With the rapid development of 3D sensing technology, the application of acquisition devices such as LiDAR, structured light cameras, and multi-view stereo vision is becoming increasingly widespread. Point cloud data, as a core data form that directly represents 3D spatial geometric information, has become an indispensable data foundation for fields such as autonomous driving environmental perception, robot autonomous navigation, high-precision 3D reconstruction, industrial defect detection, and virtual reality. Point cloud data can directly reproduce the 3D contours and geometric details of real scenes and objects, providing accurate spatial information support for subsequent high-level tasks such as target recognition, scene understanding, and geometric modeling. Its data quality directly determines the execution accuracy and reliability of downstream tasks.

[0003] In actual point cloud data acquisition, due to the limitations of sensor measurement accuracy, ambient lighting and occlusion interference, equipment vibration, and the surface material characteristics of the scanned scene, the acquired raw point cloud data inevitably contains a large number of random noise points and outliers. These noise points usually appear as discrete floating points that deviate from the real surface of the object, which directly cause distortion of the point cloud geometry, blurring of contour edges, and loss of detailed features. This seriously affects the effect of subsequent processing such as point cloud registration, surface reconstruction, and feature extraction, and may even lead to misjudgments in tasks such as autonomous driving target recognition and robot obstacle avoidance. Therefore, point cloud denoising has become an indispensable and critical pre-processing step in the 3D point cloud processing workflow.

[0004] Early point cloud denoising methods primarily relied on traditional geometry-driven approaches. These included moving least squares-based surface fitting, bilateral filtering based on local statistical properties, outlier removal based on density clustering, and anisotropic smoothing. These methods smoothed or removed noise points using manually designed geometric priors and statistical rules, achieving basic denoising effects in simple, uniformly noisy scenarios. However, these methods heavily depended on manually set parameters and prior assumptions, exhibiting poor adaptability to non-uniform noise and high-intensity noise in complex scenes. Furthermore, they tended to over-smooth sharp edges and subtle surfaces during denoising, failing to balance noise suppression and detail preservation, and thus unsuitable for high-precision 3D processing applications.

[0005] In recent years, with the rapid development of deep learning technology in non-Euclidean data processing, point cloud denoising methods based on neural networks have gradually become the mainstream in industry research and application. These methods learn the mapping relationship between the clean surface of a point cloud and the noise distribution in a data-driven manner, and can adaptively extract the local geometric features of the point cloud, better preserving the original geometric structure while suppressing noise. Among them, the ScoreDenoise method based on score matching theory has become one of the mainstream solutions in the field of point cloud denoising due to its unsupervised learning characteristics and strong modeling ability for complex noise distributions. It achieves iterative denoising optimization of point clouds, gets rid of the strong dependence of traditional deep learning methods on paired training data, and has good generalization ability. However, there are still obvious technical defects. The traditional local feature extraction structure has limited feature aggregation ability. When the network depth increases, it is prone to gradient vanishing and feature expression degradation. It cannot fully extract the subtle geometric features of point clouds in complex scenes. At the same time, its gradient prediction network lacks adaptive feature adjustment ability, making it difficult to match the denoising requirements under different noise intensities and different geometric structures. This results in insufficient denoising accuracy and poor geometric detail restoration in high-noise scenes.

[0006] To enhance the extraction and aggregation capabilities of local features in point clouds, the dynamic graph convolution EdgeConv (EDC) operator has been proposed and widely applied in point cloud processing. This operator constructs a dynamic local graph structure for each point using the K-nearest neighbor algorithm, learning the edge features between the center point and its neighbors. It also balances the permutation invariance of point cloud data with the ability to model local geometric structures, becoming a fundamental core operator for point cloud feature extraction. Building upon this, researchers have deeply integrated the dense connection concept of DenseNet with the EdgeConv operator, proposing the DenseEdgeConv dense edge convolution operator. Through the dense connection mechanism of fully connected layers, it achieves full reuse of features from different levels, effectively alleviating the gradient vanishing problem in deep networks. Simultaneously, it can aggregate multi-scale local neighborhood features, further enhancing the ability to express the local geometric structure of point clouds, and has been gradually applied to point cloud classification, segmentation, and upsampling tasks. However, existing solutions for applying DenseEdgeConv to point cloud denoising tasks still have significant shortcomings. While simply stacking DenseEdgeConv modules can improve feature aggregation capabilities, they lack targeted enhancement mechanisms for local geometric details. Deep networks still experience weakened local feature responses. Furthermore, existing solutions fail to deeply coordinate and optimize the feature extraction capabilities of DenseEdgeConv with the gradient prediction process of the ScoreNet architecture. They cannot integrate the extracted high-resolution geometric features into the adaptive adjustment process of gradient prediction, resulting in a difficulty in balancing denoising accuracy and model computation efficiency, and making it unsuitable for high-precision point cloud denoising requirements in complex scenarios.

[0007] In addition, most existing improvements to the ScoreNet architecture focus only on structural adjustments to the gradient prediction network itself, neglecting the collaborative optimization of front-end feature extraction and back-end gradient prediction. Some solutions introduce conditional modulation mechanisms that use fixed noise intensity and time steps as latent conditional vectors, failing to adaptively match the geometric features and noise distribution of different point cloud regions. This results in insufficient gradient prediction accuracy and incomplete denoising or excessive smoothing of details in complex surfaces and sharp edges. Furthermore, existing solutions often significantly increase the number of network parameters and computational load to improve denoising accuracy, resulting in insufficient model lightweighting and making it difficult to implement in industrial scenarios with stringent requirements for computing resources and inference speed, such as autonomous driving and real-time robot perception. Summary of the Invention

[0008] Therefore, the technical problem to be solved by the present invention is to overcome the shortcomings of existing point cloud denoising methods in terms of local feature extraction and aggregation capabilities, the tendency of deep networks to suffer from gradient vanishing and feature expression degradation, and the lack of adaptive feature adjustment capabilities in gradient prediction networks, resulting in insufficient denoising accuracy and poor geometric detail restoration in high-noise scenes.

[0009] To address the aforementioned technical problems, this invention provides a point cloud denoising method based on residual enhancement and conditional modulation ScoreNet, comprising the following steps: S1: Acquire three-dimensional noisy point cloud data, preprocess the three-dimensional noisy point cloud data to obtain standardized point cloud data; S2: Input the standardized point cloud data into the point cloud feature extraction network. After feature initialization, an initial feature map is obtained. Multi-stage feature aggregation and local feature enhancement are performed on the initial feature map to output the point cloud fusion feature. S3: Construct a latent conditional vector, input the point cloud fusion features and the latent conditional vector into the conditional modulation residual ScoreNet architecture, and dynamically modulate the distribution of the point cloud fusion features based on the latent conditional vector through the conditional modulation residual block to predict the noise gradient score of the three-dimensional noisy point cloud data. S4: Based on the noise gradient score, each point in the three-dimensional noisy point cloud data is denoised and updated to obtain a denoised clean point cloud.

[0010] In one embodiment of the present invention, the method for performing multi-stage feature aggregation and local feature enhancement on the initial feature map and outputting point cloud fusion features in step S2 is as follows: In the first stage, a local neighborhood graph structure is constructed for the initial feature map. Edge features and aggregated neighborhood features are extracted from the local neighborhood graph structure to obtain multi-scale aggregated features. The multi-scale aggregated features are then enhanced and nonlinearly transformed to obtain enhanced features. In subsequent stages, the enhanced features output from the previous stage are used as the input features for the current stage. The local neighborhood graph structure is reconstructed based on the input features of the current stage. Edge features and aggregated neighborhood features are extracted from the reconstructed local neighborhood graph structure to obtain the multi-scale aggregated features of the current stage. The multi-scale aggregated features of the current stage are enhanced and nonlinearly transformed to output the enhanced features of the current stage. After completing the preset number of iterative processes, the enhanced features output by the last stage are adjusted and mapped in dimensions to obtain the point cloud fusion features.

[0011] In one embodiment of the present invention, the method for constructing a local neighborhood graph structure for the initial feature map is as follows: for each point in the initial feature map, the point is taken as a target point, the distance between the target point and other points is calculated, and a preset number of the closest points are selected as neighboring points to construct a local neighborhood graph structure.

[0012] In one embodiment of the present invention, in the first stage, using the three-dimensional spatial coordinates of the three-dimensional noisy point cloud data as a metric, a preset number of neighboring points are searched for each target point. Only the relative features between the neighboring points and the target point are constructed as edge features. The edge features are input into a densely connected convolutional module. The first layer convolution performs dimension mapping, the middle layer convolution accumulates multi-scale features through dense connections, and the last layer convolution fuses all accumulated features to generate aggregated features. Then, max pooling is used to aggregate the neighborhood aggregated features to obtain multi-scale aggregated features. In subsequent stages, the target point features, neighborhood point features, and relative features are concatenated to generate edge features. Then, the concatenated edge features are densely convolved and aggregated in the same way as in the first stage to obtain the multi-scale aggregated features of the current stage.

[0013] In one embodiment of the present invention, the method for enhancing and nonlinearly transforming the multi-scale aggregated features is as follows: The location encoding is fused with the multi-scale aggregated features to obtain the fused features; The fused features are subjected to edge max pooling to enhance the feature response in local regions, resulting in enhanced features. The enhanced features are input into a multilayer perceptron for nonlinear transformation to obtain the transformed features. The fused features and the transformed features are added together by residual addition to obtain the enhanced features.

[0014] In one embodiment of the present invention, step S3, constructing a latent conditional vector and inputting the point cloud fusion features and the latent conditional vector into the conditional modulation residual ScoreNet architecture, is as follows: mapping the point cloud fusion features to a preset dimension through a linear transformation to obtain the latent conditional vector; fusing the point cloud coordinates and the latent conditional vector to obtain fused input features; and inputting the fused input features into the conditional modulation residual ScoreNet architecture, which includes a feature projection layer, multi-stage stacked conditional modulation residual blocks, and an output mapping layer.

[0015] In one embodiment of the present invention, the method for predicting the noise gradient score of the three-dimensional noisy point cloud data by dynamically modulating the distribution of the point cloud fusion features based on the latent conditional vector using conditional modulation residual blocks is as follows: The point cloud coordinates are fused with the latent conditional vector to obtain fused input features; the fused input features are then projected to obtain initial score features. The initial score features are input into a multi-stage stacked conditional modulation residual block. Each conditional modulation residual block dynamically adjusts the input features according to the latent conditional vector and outputs the adjusted features. The adjusted features are then used for feature mapping to obtain noise gradient scores that are consistent with the coordinate dimensions of the point cloud.

[0016] In one embodiment of the present invention, the method for preprocessing the three-dimensional noisy point cloud data to obtain standardized point cloud data in step S1 is as follows: removing outliers from the three-dimensional noisy point cloud data; downsampling the point cloud data after removing outliers to reduce the number of points; and performing coordinate normalization processing on the downsampled point cloud data to map the point cloud coordinates to a preset range to obtain standardized point cloud data.

[0017] This invention also provides a point cloud denoising system based on residual enhancement and conditional modulation ScoreNet, comprising: The data input module is used to acquire three-dimensional noisy point cloud data and preprocess the three-dimensional noisy point cloud data to obtain standardized point cloud data. The feature extraction module is used to input the standardized point cloud data into the point cloud feature extraction network, obtain an initial feature map after feature initialization, perform multi-stage feature aggregation and local feature enhancement on the initial feature map, and output point cloud fusion features. The score prediction module is used to construct a latent conditional vector. The point cloud fusion features and the latent conditional vector are input into the conditional modulation residual ScoreNet architecture. The distribution of the point cloud fusion features is dynamically modulated based on the latent conditional vector by the conditional modulation residual block to predict the noise gradient score of the three-dimensional noisy point cloud data. The denoising output module is used to update each point in the three-dimensional noisy point cloud data based on the noise gradient score, so as to obtain a denoised clean point cloud.

[0018] The present invention also provides a computer storage medium storing a computer software product, the computer software product including a plurality of instructions for causing a computer device to execute the point cloud denoising method based on residual enhancement and conditional modulation ScoreNet.

[0019] The technical solution of the present invention has the following advantages over the prior art: This invention deeply collaborates a point cloud feature extraction network with a conditional modulation residual ScoreNet. The point cloud fusion features output from the front end serve as latent conditional vectors to drive dynamic modulation of the back-end gradient prediction. During the feature extraction stage, a local neighborhood graph structure is constructed and multi-scale features are aggregated. Combined with residual connections, detail enhancement is performed, effectively mitigating the gradient vanishing problem and preserving geometric details while suppressing noise. The point cloud fusion features are used as latent conditional vectors to generate modulation coefficients, dynamically and adaptively adjusting gradient prediction. This allows the network to match differentiated modulation parameters based on local geometric features, overcoming the problem of insufficient prediction accuracy with fixed conditional vectors. Simultaneously, efficient feature reuse and lightweight design control the number of model parameters, providing a feasible solution for real-time perception scenarios. Furthermore, based on score matching theory, it eliminates the dependence on paired training data and exhibits stable denoising effects under different noise levels. Attached Figure Description

[0020] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings.

[0021] Figure 1 This is a flowchart illustrating the point cloud denoising method based on residual enhancement and conditional modulation ScoreNet provided in this embodiment of the invention. Figure 2 This is a schematic diagram of the overall framework of the point cloud denoising method based on residual enhancement and conditional modulation ScoreNet in an embodiment of the present invention. Figure 3 This is a schematic diagram of the point cloud feature extraction network in an embodiment of the present invention; Figure 4This is a schematic diagram of the ScoreNet architecture for conditional modulation residuals in an embodiment of the present invention; Figure 5 This is a schematic diagram of the structure of the conditional modulation residual block in an embodiment of the present invention; Figure 6 This diagram illustrates a comparison of the effectiveness of the method of the present invention and existing methods in point cloud denoising tasks. Detailed Implementation

[0022] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.

[0023] Example 1: like Figure 1 As shown, this invention provides a point cloud denoising method based on residual enhancement and conditional modulation ScoreNet, comprising the following steps: S1: Acquire three-dimensional noisy point cloud data, preprocess the three-dimensional noisy point cloud data to obtain standardized point cloud data; S2: Input the standardized point cloud data into the point cloud feature extraction network. After feature initialization, an initial feature map is obtained. Multi-stage feature aggregation and local feature enhancement are performed on the initial feature map to output the point cloud fusion feature. S3: Construct a latent conditional vector, input the point cloud fusion features and the latent conditional vector into the conditional modulation residual ScoreNet architecture, and dynamically modulate the distribution of the point cloud fusion features based on the latent conditional vector through the conditional modulation residual block to predict the noise gradient score of the three-dimensional noisy point cloud data. S4: Based on the noise gradient score, each point in the three-dimensional noisy point cloud data is denoised and updated to obtain a denoised clean point cloud.

[0024] This invention acquires 3D noisy point cloud data and preprocesses it into standardized point cloud data. This standardized point cloud data is then input into a point cloud feature extraction network. After feature initialization, multi-stage feature aggregation and local feature enhancement output point cloud fusion features. A latent conditional vector is constructed and input into a conditional modulation residual ScoreNet architecture along with this fusion feature. The conditional modulation residual block dynamically modulates the feature distribution based on the latent conditional vector to predict noise gradient scores. These scores are then used to denoise and update each point in the noisy point cloud, resulting in a clean point cloud. This invention's method deeply coordinates front-end multi-stage feature aggregation and local enhancement with back-end conditional modulation gradient prediction, effectively alleviating the problems of gradient vanishing and feature representation degradation in deep networks. It suppresses noise while preserving sharp edges and subtle surface structures in the point cloud. The dynamic adaptive modulation of gradient prediction using the point cloud fusion features as latent conditional vectors overcomes the insufficient prediction accuracy of fixed conditional vectors. Through dense connections and lightweight design, the number of model parameters is controlled, providing a feasible solution for real-time scenarios such as autonomous driving and robot perception. Based on score matching theory, it overcomes the strong dependence on paired training data and exhibits stable denoising effects under different noise levels.

[0025] To enhance the extraction and aggregation capabilities of local features in point clouds, the EdgeConv operator constructs a dynamic local graph structure for each point using the K-nearest neighbor algorithm, learning the edge features between the center point and its neighboring points. This approach balances the permutation invariance of point cloud data with the ability to model local geometric structures, making it a fundamental core module for point cloud feature extraction.

[0026] Building upon this, we introduce the densely connected network of DenseNet into EdgeConv and design the DenseEdgeConv dense edge convolution module: through the dense connection mechanism of fully connected layers, we achieve full reuse of features at different levels, effectively alleviate the gradient vanishing problem of deep networks, and at the same time, we can aggregate local neighborhood features at multiple scales, thereby enhancing the ability to express the local geometric structure of point clouds.

[0027] However, existing technologies still have significant shortcomings when applying DenseEdgeConv to point cloud denoising tasks: while simply stacking DenseEdgeConv modules can improve feature aggregation capabilities, they lack targeted enhancement mechanisms for local geometric details, and deep networks still suffer from weakened local feature responses; at the same time, existing solutions fail to deeply coordinate and optimize the feature extraction capabilities of DenseEdgeConv with the gradient prediction process of the ScoreNet architecture, and cannot integrate the extracted high-resolution geometric features into the adaptive adjustment process of gradient prediction, resulting in a difficulty in balancing denoising accuracy and model computational efficiency.

[0028] To address the aforementioned issues, this invention employs a structure in the point cloud feature extraction network that alternates between a staged DenseEdgeConv module and a Residual Local Feature Enhancement (ResLFE) module. Within each convolutional unit, DenseEdgeConv concatenates the initial edge features with the edge features output from each intermediate convolutional layer along the channel dimension, forming a multi-scale feature reuse path. For each target point and its K neighboring points, edge features are first constructed, then sequentially passed through the first, intermediate, and final convolutional layers. The intermediate convolutional layers accumulate feature responses from different receptive fields through dense connections, while the final convolutional layer fuses all accumulated features. Finally, max pooling aggregates neighborhood information, outputting multi-scale aggregated features. Simultaneously, a ResLFE module is introduced after each DenseEdgeConv module. Through positional encoding fusion, local feature enhancement, multilayer perceptron, and residual connections with dynamic rates, the aggregated features undergo local detail enhancement and dimensionality preservation. This approach addresses the weakness of local feature responses while maintaining the advantages of DenseEdgeConv's multi-scale feature aggregation.

[0029] Furthermore, this invention uses the point cloud fusion features extracted from the front end as a latent conditional vector input into the conditional modulation residual ScoreNet in the back end, realizing deep collaboration between feature extraction and gradient prediction. This enables high-resolution geometric features to effectively drive the adaptive dynamic modulation of gradient prediction, balancing denoising accuracy and computational efficiency, and adapting to the high-precision point cloud denoising requirements in complex scenarios.

[0030] like Figure 2 As shown, in this embodiment of the invention, in step S1, three-dimensional noisy point cloud data is acquired, and the three-dimensional noisy point cloud data is preprocessed to obtain standardized point cloud data.

[0031] Point cloud denoising methods consist of two parts: a point cloud feature extraction network and a conditional modulation residual ScoreNet. Since supervised training based on score matching theory requires pairs of "noisy point cloud - clean point cloud" data as labels, and real-world point clouds cannot obtain absolutely clean ground truth values, a method of artificially adding noise to known clean point clouds is used to construct training samples.

[0032] Specifically, the original clean point cloud is randomly rotated to enhance the model's ability to generalize to object orientation; then Gaussian noise is added to the rotated clean point cloud to obtain a noisy point cloud; local point cloud patches are extracted with each point in the noisy point cloud as the center, and each patch contains the center point and its neighboring points, thereby constructing a training sample pair between the noisy point cloud and the clean point cloud.

[0033] Further, in step S2, standardized point cloud data is input into the point cloud feature extraction network. After feature initialization, an initial feature map is obtained. Multi-stage feature aggregation and local feature enhancement are then performed on the initial feature map to output the point cloud fusion features. For example... Figure 3 As shown, multi-round feature enhancement is achieved by alternately stacking DenseEdgeConv and ResLFE.

[0034] Specifically, in this embodiment of the invention, the point cloud feature extraction network adopts a structure consisting of a feature initialization module, a multi-stage feature aggregation and local feature enhancement module, and a feature output module connected in series. The multi-stage processing comprises alternating DenseEdgeConv dense edge convolution operations and ResLFE residual local feature enhancement operations, for a total of four stages. Within each stage, DenseEdgeConv aggregation is performed first, followed by ResLFE enhancement.

[0035] In the initial feature mapping stage, the input noisy point cloud coordinates are: ,in For batch size, This represents the number of points in the point cloud. An initial point feature is obtained by mapping the 3D coordinate features to a preset feature channel dimension through a fully connected layer (linear transformation layer). ,in The initial number of feature channels is preferably set to 24 in this embodiment of the invention.

[0036] Standardized point cloud Input a fully connected linear transformation layer to map the 3D coordinate features to a preset channel dimension, resulting in an initial feature map: , The linear transformation is performed independently on the coordinates of each point, embedding geometric position information into a high-dimensional space.

[0037] Furthermore, multi-stage alternating processing of DenseEdgeConv and ResLFE is performed. In the first stage, the 3D spatial coordinates of the point cloud are used as the metric for each target point. The K-nearest neighbor algorithm is used to search for its neighboring points.

[0038] Specifically, the Euclidean distance between this point and all other points in the point cloud is calculated, and the K nearest points are selected as its neighborhood points. This process is denoted as... In this embodiment of the invention, K is set to 32, and the 32 nearest neighboring points in the search space are used for each target point to generate a local neighborhood graph structure. It is used for edge feature construction and feature aggregation operations.

[0039] In the first stage, only the relative features between the neighboring points and the target point are used as edge features, without introducing the original features of the target point itself, in order to enhance the difference in the neighborhood structure: , in, The coordinates of the target point, Let these be the coordinates of the neighboring points. The densely connected convolutional module is used as input for edge features.

[0040] The densely connected convolutional module consists of three layers: a first convolutional layer, intermediate convolutional layers, and a final convolutional layer. Preferably, the growth rate is set to 12. The input of each layer is formed by concatenating the initial input features with the output features of all previous layers along the channel dimension, thereby enabling the reuse of multi-scale features.

[0041] Specifically, the input to the first convolutional layer is the edge feature, which has an initial dimension of 3. After passing through the first convolutional layer, it outputs a 12-dimensional feature. The input to the second convolutional layer is formed by concatenating the initial 3-dimensional feature with the 12-dimensional feature output from the first layer. At this point, the concatenated feature has a dimension of 15. After passing through the second convolutional layer, it outputs a 12-dimensional feature. The input to the third convolutional layer is formed by concatenating the initial 3-dimensional feature, the 12-dimensional feature output from the first layer, and the 12-dimensional feature output from the second layer. At this point, the concatenated feature has a dimension of 27. After passing through the third convolutional layer, it also outputs a 12-dimensional feature.

[0042] The initial 3D features are concatenated with the 12D features output by each of the three convolutional layers along the channel dimension to obtain a multi-scale aggregated feature, the dimensionality of which is expressed as follows: .

[0043] Due to the initial feature map It uses 24-dimensional coordinates instead of the original 3-dimensional coordinates. In actual implementation, the first stage of DenseEdgeConv uses the initial feature map. As input features, edge features are still calculated using the original 3D spatial coordinates (i.e., standardized point cloud coordinates) as a metric to search for the neighborhood, while feature aggregation uses... Features of the corresponding points in the middle.

[0044] Specifically, for each target point and its neighboring points The edge features are constructed as follows: , in, This represents the feature vectors of the target point and its neighboring points in the initial feature map. After the edge features are processed by a densely connected convolutional module, the output channel number is... .

[0045] The dense connection mechanism means treating the three convolutional layers as a whole. The first convolutional layer takes the initial edge features as input; the second convolutional layer's input is formed by concatenating the initial edge features with the first layer's output; and the third convolutional layer's input is formed by concatenating the initial edge features, the first layer's output, and the second layer's output. Through this layer-by-layer concatenation, each layer can reuse the features extracted from all previous layers, achieving the fusion of multi-scale information.

[0046] The initial edge features are concatenated with the outputs of the three convolutional layers along the channel dimension to obtain the aggregated features. Then, max pooling is performed on the 32 neighborhood edge features of each target point to aggregate the neighborhood information and obtain the first-stage aggregated features: .

[0047] Furthermore, Enter ResLFE.

[0048] ResLFE enhances the feature by adding positional encoding (PE) to the aggregated features: , The positional encoding is generated using a sine and cosine function similar to that of the Transformer, with dimensions and... Consistent.

[0049] Furthermore, local structural features are enhanced through neighborhood feature pooling (i.e., edge max pooling): based on the KNN neighborhood index, the maximum value of the neighborhood features of each point is taken in the channel dimension, and the training is stabilized by batch normalization, while the output dimension remains unchanged.

[0050] Furthermore, the pooled features are input into a feedforward network (FFN) for nonlinear transformation: .

[0051] In this embodiment of the invention, FFN employs a multilayer perceptron with a hidden layer dimension scaling factor of 2, and enhances expressive power through the ReLU activation function.

[0052] Furthermore, random depth regularization (DropPath) is introduced to alleviate overfitting, and residual connections are performed to obtain enhanced features: .

[0053] The above process is recorded as one ResLFE process. Each stage performs DenseEdgeConv followed by another ResLFE process, for a total of 4 stages, gradually fusing geometric information at different scales to obtain the global features of the point cloud: .

[0054] Further, in step S3, a latent conditional vector is constructed, input to the conditional modulation residual ScoreNet architecture, and the noise gradient score is predicted.

[0055] Specifically, the global features of the point cloud Input a linear transformation layer, map it to 128 dimensions, and obtain the latent condition vector: , Latent conditional vectors condense the global geometric and structural information of the entire point cloud and are used for dynamic modulation of gradient prediction.

[0056] The original point cloud coordinate tensor Transpose At the same time, the implicit condition vector Extend along the spatial dimension to the same dimension as the number of points in the point cloud, i.e. The two are concatenated along the channel dimension to obtain the fused input features. : .

[0057] Furthermore, the fused input features are fed into the conditional modulation residual ScoreNet architecture. For example... Figure 4 As shown, the ScoreNet architecture for conditional modulation residuals includes a feature projection layer, multi-stage stacked conditional modulation residual blocks, and an output mapping layer.

[0058] Specifically, through a feature projection layer (1D convolution) Mapping to the preset hidden layer dimensions yields the initial score features: .

[0059] like Figure 5 As shown, The input consists of multi-stage stacked conditional modulation residual blocks. In this embodiment of the invention, four identical conditional modulation residual blocks are stacked, and the input and output dimensions of each residual block are 1. Let the first... The input for each block is (in Each block performs the following operations in sequence: The input features are batch normalized, then activated by ReLU, and finally passed through a 1D convolutional layer to obtain intermediate features. ; From the implicit condition vector Generate modulation parameters. Scaling coefficients are generated using two independent multilayer perceptrons. and offset coefficient ,Right now , ;Will and Expanding along the spatial dimension to the same size as the feature map, for Dynamic distribution adjustment: ; right Perform a second 1D convolution to obtain the residual increment. The input features are added to the residual increment via a shortcut connection to obtain the output of the current block. The output dimension remains unchanged.

[0060] The output of the current block is used as the input of the next block, and this process is repeated through four conditionally modulated residual blocks. After all four blocks, the adjusted feature is obtained, denoted as... .

[0061] Furthermore, through the output mapping layer... Batch normalization and ReLU activation are performed sequentially, followed by a 1D convolutional layer to map the number of channels to the same dimension as the point cloud coordinates, resulting in... ,Will Transpose back This yields the point cloud gradient prediction score: .

[0062] Further, in step S4, the noise gradient score is calculated. The point cloud data containing noise is updated by denoising.

[0063] Specifically, for each point in each sample Update its coordinates according to the gradient ascent direction, that is: , in, To update the step size, the update process uses the gradient predicted by the network to move noisy points along the direction of increasing log probability density, thereby approximating the clean surface. This update is repeated several times to obtain the denoised point cloud coordinates, which is the clean point cloud.

[0064] experiment: The training dataset used is a collection of training samples constructed from publicly available point cloud datasets. This dataset contains 40 different 3D mesh models, and point cloud data at three different resolutions are sampled from each mesh model. The number of points at each resolution is 10,000, 30,000, and 50,000, respectively, resulting in a total of 120 training point cloud samples. All point clouds are obtained by uniformly sampling from the surface of the corresponding 3D mesh model, ensuring that the sampled points accurately reflect the geometric features of the object's surface while avoiding geometric distortions such as collapse or expansion during the sampling process.

[0065] During model training, clean point cloud data is subjected to online noise perturbation to construct noisy point clouds. Specifically, Gaussian noise is added to the coordinates of each point in each point cloud sample. The noise amplitude is randomly selected between 0.5% and 2% of the radius of the point cloud's bounding sphere, thereby generating training samples with different noise intensities to improve the model's robustness to noise perturbations.

[0066] To verify the effectiveness of the point cloud denoising method of the present invention, three noise levels of 1%, 2%, and 3% were set on point cloud data of two scales: 10,000 points and 50,000 points. The chamfer distance (CD) and point-to-surface distance (P2M) were used as evaluation indicators. Comparative experiments were conducted with three existing methods: ScoreDenoise, MAG, and IterativeFPN. The results are shown in Tables 1 and 2.

[0067] Table 1:

[0068] Table 2:

[0069] As shown in Table 1, on a point cloud test set of 10,000 points, the CD and P2M indices of this invention are significantly lower than those of the comparison methods at all noise levels. In particular, in the 3% high noise scenario, both indices are the minimum values ​​among all methods, demonstrating stronger noise robustness.

[0070] As shown in Table 2, on a high-density point cloud test set of 50,000 points, the present invention maintains its overall advantages, with a particularly outstanding improvement in the P2M index. At 1% noise, the P2M is only 0.069, which is much lower than other methods, demonstrating high-precision detail reconstruction capability. At 3% high noise, CD and P2M are still the best, proving that the present invention has a stable denoising effect under different point cloud densities.

[0071] like Figure 6The figure shows a visual comparison of different point cloud denoising methods on three typical models: Casting, Chair, and Cow. The columns in the figure are, in order: noisy input point cloud, denoising result of the ScoreDenoise method, denoising result of the MAG method, denoising result of the IterativeFPN method, denoising result of the method of this invention (Ours), and real clean point cloud (Clean). As can be seen from the figure, the noisy input point cloud contains a large number of discrete noise points, resulting in blurred model contours, missing geometric details, and low overall structural discernibility. Existing methods (ScoreDenoise, MAG, IterativeFPN) still exhibit varying degrees of noise residue or detail distortion after denoising: the point clouds denoised by ScoreDenoise and MAG methods still contain a significant amount of noise, the model edges are not smooth enough, and the restoration of local features (such as the chair legs of the Chair and the limb contours of the Cow) is insufficient; while the IterativeFPN method shows some improvement in noise suppression, it still suffers from blurred details and blunted structural edges, showing a significant difference from the real clean point cloud. After denoising using the method of this invention, point cloud noise is significantly eliminated, the model outline is clear and the edges are smooth, and geometric details (such as the surface structure of Casting, the frame details of Chair, and the limb outline of Cow) are highly restored, with the visual effect being closest to that of a real clean point cloud.

[0072] Example 2: Based on the same inventive concept as Embodiment 1, this invention also provides a point cloud denoising system based on residual enhancement and conditional modulation ScoreNet, used to implement the steps of the point cloud denoising method based on residual enhancement and conditional modulation ScoreNet described in Embodiment 1, including the following modules: The data input module is used to acquire three-dimensional noisy point cloud data and preprocess the three-dimensional noisy point cloud data to obtain standardized point cloud data. The feature extraction module is used to input the standardized point cloud data into the point cloud feature extraction network, obtain an initial feature map after feature initialization, perform multi-stage feature aggregation and local feature enhancement on the initial feature map, and output point cloud fusion features. The score prediction module is used to construct a latent conditional vector. The point cloud fusion features and the latent conditional vector are input into the conditional modulation residual ScoreNet architecture. The distribution of the point cloud fusion features is dynamically modulated based on the latent conditional vector by the conditional modulation residual block to predict the noise gradient score of the three-dimensional noisy point cloud data. The denoising output module is used to update each point in the three-dimensional noisy point cloud data based on the noise gradient score, so as to obtain a denoised clean point cloud.

[0073] The data input module, feature extraction module, score prediction module, and denoising output module of the point cloud denoising system based on residual enhancement and conditional modulation ScoreNet proposed in this embodiment are respectively used to implement steps S1, S2, S3, and S4 in the point cloud denoising method based on residual enhancement and conditional modulation ScoreNet in Embodiment 1. To avoid redundancy, they will not be described again here.

[0074] Example 3: The present invention also provides a computer storage medium storing a computer software product, the computer software product including several instructions for causing a computer device to execute the point cloud denoising method based on residual enhancement and conditional modulation ScoreNet as described in Embodiment 1.

[0075] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0076] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. 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... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0077] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0078] 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.

[0079] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A point cloud denoising method based on residual enhancement and conditional modulation ScoreNet, characterized in that, Includes the following steps: S1: Acquire three-dimensional noisy point cloud data, preprocess the three-dimensional noisy point cloud data to obtain standardized point cloud data; S2: Input the standardized point cloud data into the point cloud feature extraction network. After feature initialization, an initial feature map is obtained. Multi-stage feature aggregation and local feature enhancement are performed on the initial feature map to output the point cloud fusion feature. S3: Construct a latent conditional vector, input the point cloud fusion features and the latent conditional vector into the conditional modulation residual ScoreNet architecture, and dynamically modulate the distribution of the point cloud fusion features based on the latent conditional vector through the conditional modulation residual block to predict the noise gradient score of the three-dimensional noisy point cloud data. S4: Based on the noise gradient score, each point in the three-dimensional noisy point cloud data is denoised and updated to obtain a denoised clean point cloud.

2. The point cloud denoising method based on residual enhancement and conditional modulation ScoreNet according to claim 1, characterized in that: In step S2, the method for performing multi-stage feature aggregation and local feature enhancement on the initial feature map to output point cloud fusion features is as follows: In the first stage, a local neighborhood graph structure is constructed for the initial feature map. Edge features and aggregated neighborhood features are extracted from the local neighborhood graph structure to obtain multi-scale aggregated features. The multi-scale aggregated features are then enhanced and nonlinearly transformed to obtain enhanced features. In subsequent stages, the enhanced features output from the previous stage are used as the input features for the current stage. The local neighborhood graph structure is reconstructed based on the input features of the current stage. Edge features and aggregated neighborhood features are extracted from the reconstructed local neighborhood graph structure to obtain the multi-scale aggregated features of the current stage. The multi-scale aggregated features of the current stage are enhanced and nonlinearly transformed to output the enhanced features of the current stage. After completing the preset number of iterative processes, the enhanced features output by the last stage are adjusted and mapped in dimensions to obtain the point cloud fusion features.

3. The point cloud denoising method based on residual enhancement and conditional modulation ScoreNet according to claim 2, characterized in that: The method for constructing a local neighborhood graph structure for the initial feature map is as follows: for each point in the initial feature map, the point is taken as the target point, the distance between the target point and other points is calculated, and a preset number of the closest points are selected as neighborhood points to construct a local neighborhood graph structure.

4. The point cloud denoising method based on residual enhancement and conditional modulation ScoreNet according to claim 2, characterized in that: The method for extracting edge features and aggregated neighborhood features from the local neighborhood graph structure to obtain multi-scale aggregated features is as follows: In the first stage, using the three-dimensional spatial coordinates of the three-dimensional noisy point cloud data as a metric, a preset number of neighboring points are searched for each target point. Only the relative features between the neighboring points and the target point are constructed as edge features. The edge features are input into a densely connected convolutional module. The first layer convolution performs dimension mapping, the middle layer convolution accumulates multi-scale features through dense connections, and the last layer convolution fuses all accumulated features to generate aggregated features. Then, max pooling is used to aggregate the neighborhood aggregated features to obtain multi-scale aggregated features. In subsequent stages, the target point features, neighborhood point features, and relative features are concatenated to generate edge features. Then, the concatenated edge features are densely convolved and aggregated in the same way as in the first stage to obtain the multi-scale aggregated features of the current stage.

5. The point cloud denoising method based on residual enhancement and conditional modulation ScoreNet according to claim 2, characterized in that: The method for enhancing and nonlinearly transforming the multi-scale aggregated features is as follows: The location encoding is fused with the multi-scale aggregated features to obtain the fused features; The fused features are subjected to edge max pooling to enhance the feature response in local regions, resulting in enhanced features. The enhanced features are input into a multilayer perceptron for nonlinear transformation to obtain the transformed features. The fused features and the transformed features are added together by residual addition to obtain the enhanced features.

6. The point cloud denoising method based on residual enhancement and conditional modulation ScoreNet according to claim 1, characterized in that: In step S3, the method for constructing the latent condition vector and inputting the point cloud fusion features and the latent condition vector into the conditional modulation residual ScoreNet architecture is as follows: the point cloud fusion features are mapped to a preset dimension through a linear transformation to obtain the latent condition vector; the point cloud coordinates and the latent condition vector are fused to obtain fused input features; the fused input features are input into the conditional modulation residual ScoreNet architecture, which includes a feature projection layer, multi-stage stacked conditional modulation residual blocks, and an output mapping layer.

7. The point cloud denoising method based on residual enhancement and conditional modulation ScoreNet according to claim 1 or 6, characterized in that: In step S3, the method for predicting the noise gradient score of the 3D noisy point cloud data by dynamically modulating the distribution of the point cloud fusion features based on the latent conditional vector using conditional modulation residual blocks is as follows: The point cloud coordinates are fused with the latent conditional vector to obtain fused input features; the fused input features are then projected to obtain initial score features. The initial score features are input into a multi-stage stacked conditional modulation residual block. Each conditional modulation residual block dynamically adjusts the input features according to the latent conditional vector and outputs the adjusted features. The adjusted features are then used for feature mapping to obtain noise gradient scores that are consistent with the coordinate dimensions of the point cloud.

8. The point cloud denoising method based on residual enhancement and conditional modulation ScoreNet according to claim 1, characterized in that: In step S1, the method for preprocessing the three-dimensional noisy point cloud data to obtain standardized point cloud data is as follows: outlier points are removed from the three-dimensional noisy point cloud data. The point cloud data after removing outliers is downsampled to reduce the number of points; the downsampled point cloud data is then normalized to map the point cloud coordinates to a preset range, resulting in standardized point cloud data.

9. A point cloud denoising system based on residual enhancement and conditional modulation ScoreNet, characterized in that, include: The data input module is used to acquire three-dimensional noisy point cloud data and preprocess the three-dimensional noisy point cloud data to obtain standardized point cloud data. The feature extraction module is used to input the standardized point cloud data into the point cloud feature extraction network, obtain an initial feature map after feature initialization, perform multi-stage feature aggregation and local feature enhancement on the initial feature map, and output point cloud fusion features. The score prediction module is used to construct a latent conditional vector. The point cloud fusion features and the latent conditional vector are input into the conditional modulation residual ScoreNet architecture. The distribution of the point cloud fusion features is dynamically modulated based on the latent conditional vector by the conditional modulation residual block to predict the noise gradient score of the three-dimensional noisy point cloud data. The denoising output module is used to update each point in the three-dimensional noisy point cloud data based on the noise gradient score, so as to obtain a denoised clean point cloud.

10. A computer storage medium, characterized in that, The computer storage medium stores a computer software product, which includes several instructions for causing a computer device to execute the point cloud denoising method based on residual enhancement and conditional modulation ScoreNet as described in any one of claims 1 to 8.

Images