A point cloud classification segmentation method based on geometric perception and parallel attention
By constructing a geometric perception module and an efficient channel attention module, the problem of insufficient local geometric information and feature channel mining in point cloud deep learning models is solved, enabling efficient classification and segmentation of complex shapes and improving the model's accuracy and real-time processing capabilities.
Patent Information
- Application Number
- CN202610379506.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-03-26
- Publication Date
- 2026-06-26
AI Technical Summary
Existing point cloud-based deep learning models lack the ability to mine local geometric information in the feature representation stage, and their accuracy in mining key feature channels and spatial information is insufficient. This limits the accuracy and robustness of point cloud classification and segmentation tasks, and the redundant computational overhead is large, making it difficult to meet real-time requirements.
A geometric perception module is constructed to extract the relative positional relationships of the point cloud neighborhood. Combined with an efficient channel attention module and a parallel attention module, key feature responses are enhanced through adaptive weight adjustment, reducing redundant computation and improving feature discrimination ability.
It significantly enhances the ability to represent complex shape details, reduces computational overhead, improves the accuracy and real-time processing capability of point cloud classification and segmentation, and enhances the generalization performance of the model.
Smart Images
Figure CN122289766A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of point cloud processing technology, and proposes a point cloud classification and segmentation method based on geometric perception and parallel attention. Background Technology In point cloud classification and segmentation tasks, the accuracy of feature extraction directly determines the final performance of the model. However, existing point cloud-based deep learning models still suffer from two major bottlenecks in feature representation: First, existing models lack the ability to extract local geometric information. Due to the inherent irregularity and sparsity of point cloud data, its local features are highly dependent on the relative positional relationships and spatial topology of points within the neighborhood. However, most methods only focus on the absolute coordinates of a single point or the stitching together of shallow neighborhood features, failing to fully model the geometric dependencies between neighboring points. This results in weak representation of the local structure of complex-shaped targets, loss of key geometric details, and consequently affects the accuracy and robustness of classification and segmentation tasks. On the other hand, existing models suffer from insufficient accuracy and high redundant computational overhead in mining key feature channels and spatial information. Some methods focus solely on feature enhancement in the spatial or channel domains, lacking a two-way collaborative optimization mechanism, and are unable to simultaneously capture key regions in the spatial dimension and effective features in the channel dimension. Other methods attempt to integrate dual-domain attention, but often employ serial structures or complex computational modules, which not only make it difficult to achieve adaptive weight adjustment of key features but also introduce a large amount of redundant computation, resulting in limited model inference efficiency and difficulty in adapting to the real-time requirements of real-world scenarios. Ultimately, the weakening of key feature responses and significant interference from redundant features restrict further improvement in the performance of point cloud classification and segmentation networks.
[0002] To address the problem of insufficient geometric information extraction capabilities and the inability to specifically identify key feature channels, a point cloud classification and segmentation model based on geometric perception and parallel attention is proposed. First, a local geometric feature extraction module models the relative positional relationships within the point cloud neighborhood, enhancing the representational power of local features. Second, a spatial attention module and an efficient channel attention module are constructed for parallel processing, improving feature discrimination capabilities from both the spatial and channel domains. Adaptive weight adjustment strengthens the response to key features while effectively reducing redundant computational overhead. Finally, the output features are combined and further extracted using a multilayer perceptron, improving the accuracy of the point cloud classification and segmentation network. Summary of the Invention
[0003] The technical problem to be solved by this invention is to propose a point cloud classification and segmentation method based on geometric perception and parallel attention. This method addresses the issues of existing methods failing to focus on local geometric information, lacking accuracy in mining key feature channels and spatial information, and incurring large redundant computational overhead. It provides an innovative technical approach and implementation scheme for efficient feature learning and fine segmentation and recognition of point clouds.
[0004] To achieve the above objectives, the technical solution adopted by this invention is: a point cloud classification and segmentation method based on geometric perception and parallel attention, the specific steps of which are as follows:
[0005] 1) Obtain the point cloud classification and segmentation dataset, and preprocess the dataset to form a .txt file dataset.
[0006] 2) Constructing a network model
[0007] 2.1) For the obtained point cloud classification and segmentation dataset, the geometric perception module is designed first. The purpose is to extract and optimize the geometric features of the normalized point cloud neighborhood coordinates, and to combine the basic geometric features of the point cloud neighborhood with geometric affine transformation and kernel density estimation algorithm to achieve effective fusion and representation of the local geometric structure information and spatial density information of the point cloud.
[0008] 2.2) Constructing an efficient channel attention module: Each point in a point cloud dataset has multiple feature channels, such as coordinates, color, and normal vectors. However, not all feature channels are useful for the task or contribute to the results. The introduction of a Channel Attention Module (CAM) allows the model to adaptively evaluate the importance of each feature channel and accordingly select and enhance features that are crucial to the target task. This helps reduce the impact of redundant information and improves the expressive power and discriminative power of features. Furthermore, the feature dimensions of the input point cloud vary throughout the network. Channel attention can adaptively learn weights based on the dimensions of the input point cloud data, avoiding the unreasonable influence of fixed weight parameters on all dimensions. This enhances the model's adaptability to dimensional changes and improves its generalization performance. Since the detailed features in point cloud data are crucial for point cloud processing tasks, the classic CAM provides the model with the basic ability to adapt channel weights. However, it still has obvious limitations when processing irregular data such as point clouds that are sensitive to details and have dynamically changing dimensions: First, the "dimensionality reduction-dimensionality increase" MLP mode will compress feature channels, resulting in the loss of local geometric details and sparse structural information of the point cloud; Second, the design of fixed scaling factor is difficult to adapt to the changes in channel dimension at different levels of the point cloud network; Third, the computational cost of ordinary convolution is large and cannot meet the real-time processing needs of large-scale point clouds.
[0009] To address these issues, an efficient channel attention (ECA) mechanism is introduced and improved, resulting in a point cloud-adapted efficient channel attention module (Point-ECANet). This module retains the simplicity and attention representation capabilities of the original ECANet architecture while achieving a synergistic improvement in computational efficiency and feature discriminative power through targeted improvements, providing high-quality channel-enhanced feature support for subsequent feature fusion and deep extraction. The core design idea of this module is to adapt to the dynamic changes in the two dimensions of "number of points - number of channels" of point cloud features by replacing the core attention mechanism of ECANet with a lightweight convolutional structure and dynamically adjusting the convolutional kernel size, without changing the core attention mechanism of ECANet, while avoiding the loss of geometric details caused by feature dimension compression. The overall input of the module is a point cloud feature tensor. ,in This represents the number of points in the point cloud. This represents the number of feature channels, and the output is the point cloud feature after channel weight enhancement. First, global average pooling (GAP) is used to calculate the global average of the channel feature values for each point, resulting in... Secondly, ordinary one-dimensional convolution is replaced with depthwise separable one-dimensional convolution. This structure consists of two steps: depthwise convolution and pointwise convolution, forming an efficient processing flow of "intra-channel extraction - inter-channel fusion".
[0010] During the depthwise convolution stage, the output of global pooling is... Each channel is used independently. The convolution kernel performs convolution operations, extracting local correlation features only within a single channel, maintaining a consistent number of input and output channels; in the pointwise convolution stage, a method is used... The convolutional kernel fuses inter-channel information in the output of depthwise convolution. This improvement significantly reduces the number of parameters and computational cost of the module while ensuring the effective extraction of inter-channel correlated features, making it better suited for the real-time processing needs of large-scale, irregular data such as point clouds. To address the dynamic changes in feature dimensions at different levels of point cloud networks, this module improves the calculation method for the convolutional kernel size by incorporating the number of points. The dynamic constraints ensure that the receptive field of the convolution kernel matches the overall information content of the point cloud features. The convolution kernel size is calculated as shown in formula (1).
[0011] (1)
[0012] in, and These are model hyperparameters used to adjust the overall range and rate of change of the kernel size. This indicates taking the largest odd number not greater than the calculated value within the parentheses, ensuring the convolution kernel has central symmetry and improving the stability of channel-related feature extraction. This design allows the convolution kernel size to vary with the number of points. With the number of channels Dynamic adjustments effectively enhance the module's adaptability to irregular point cloud features, ensuring that the module can efficiently extract channel correlation information at different feature levels of the network.
[0013] Next, the output of the depthwise separable one-dimensional convolution is fed into the Sigmoid activation function. Through a non-linear transformation, the feature values are mapped to the [0,1] interval, generating channel attention weights that correspond one-to-one with the number of input channels. The weight values quantify the importance of each feature channel to the current point cloud classification and segmentation task, enabling accurate identification of key and redundant feature channels. Subsequently, the generated channel attention weights are... Point cloud feature tensor compared to the original input of the module A channel-by-channel multiplication operation is performed, and adaptive weighting is used to highlight the information representation of key feature channels while suppressing invalid interference from redundant channels, ultimately obtaining the channel-enhanced point cloud features. This effectively improves the utilization rate of effective features by subsequent network layers.
[0014] 2.3) Construct a parallel attention module. Spatial attention helps capture the spatial structure and geometric information in point cloud data, while channel attention helps improve the feature representation ability of point cloud data and the attention to different features. In point cloud processing, a task with high requirements for fine granularity, serial connection will limit the interaction and information transmission between different attention mechanisms, thereby reducing the model's expressive power and performance.
[0015] To address this issue, spatial attention and channel attention are connected in parallel. This approach allows both attention modules to learn and optimize from the original input features simultaneously, enabling the model to more comprehensively focus on important regions and feature channels in the point cloud data. Furthermore, to facilitate network optimization and training, the idea of skip connections in residual learning is adopted. The original input features are added element-wise to the feature maps processed by each attention module, resulting in a parallel attention module. The input of this module is the feature map obtained from the geometry perception module. B represents the batch size, N represents the number of sampling points, and D represents the number of channels in the feature map. The parallel attention module obtains the corresponding weights by inputting the intermediate feature map F into the spatial attention module and the channel attention module respectively. Then, it multiplies the two weights with the input feature map F to obtain... and Finally, the three are added together to obtain the output feature map. The specific calculation process is shown in formula (2).
[0016] (2)
[0017] in This represents the channel attention weight. This represents the spatial attention weights. This represents element-wise multiplication.
[0018] 3) Construct an end-to-end network structure based on geometric perception and parallel attention, and use the training set to complete model training and parameter optimization;
[0019] 3.1) The network mainly consists of three core modules: geometric perception module, spatial attention module, and efficient channel attention module.
[0020] 3.2) This model uses The original point cloud data is used as input. The farthest point of the point cloud data is sampled as the center point. Then, KNN grouping is used to feed the data into the geometry perception module for multi-dimensional feature extraction.
[0021] 3.3) The obtained features are enhanced in the spatial domain and the channel domain, respectively.
[0022] 3.4) Finally, the concatenated features are fed into multiple multilayer perceptrons for deeper feature mining and used for subsequent classification and segmentation tasks.
[0023] 3.5) Train the model with an initial learning rate of 0.001, an exponential decay learning rate adjustment strategy, and the Adam optimizer.
[0024] 4) Load the trained network parameters, perform end-to-end inference on the test set, and output the point cloud classification and point-by-point segmentation results.
[0025] The beneficial effects of this invention are as follows: First, by constructing a geometric perception module, basic geometric features are effectively integrated with geometric affine transformations and kernel density estimation algorithms, fully mining the local geometric structure information and spatial density information of point clouds, significantly enhancing the ability to represent complex shape details. Second, an efficient channel attention module for point cloud adaptation is proposed, employing a depthwise separable convolution and dynamic kernel size strategy. This avoids the loss of details caused by feature dimension compression while significantly reducing computational overhead, achieving accurate selection of key feature channels. Third, by connecting spatial attention and channel attention in parallel and combining them with residual learning, the model can simultaneously focus on important features from both the spatial and channel domains, avoiding the information interaction limitations caused by serial structures, further improving the accuracy of fine-grained segmentation and recognition. Finally, the constructed end-to-end network structure exhibits excellent feature learning efficiency and generalization performance in both classification and segmentation tasks, effectively balancing model accuracy and computational complexity, providing a reliable technical solution for efficient point cloud processing. Attached Figure Description
[0026] Figure 1 This is a diagram of a point cloud network model for geometric perception and parallel attention.
[0027] Figure 2 This is a diagram of the parallel attention module.
[0028] Figure 3 This is a diagram of an efficient channel attention module.
[0029] Figure 4 Overall accuracy plot for each category of the dataset classification
[0030] Figure 5 This is a diagram showing the results of component segmentation in the dataset.
[0031] Figure 6 This is a diagram showing the impact of the number of point clouds on the network. Detailed Implementation
[0032] This invention provides a point cloud classification and segmentation method based on geometric perception and parallel attention, specifically including the following steps:
[0033] 1) Obtain the point cloud classification and segmentation dataset, and preprocess the dataset to form a .txt file dataset.
[0034] 2) Constructing a network model
[0035] 2.1) For the obtained point cloud classification and segmentation dataset, the geometric perception module is designed first. The purpose is to extract and optimize the geometric features of the normalized point cloud neighborhood coordinates, and to combine the basic geometric features of the point cloud neighborhood with geometric affine transformation and kernel density estimation algorithm to achieve effective fusion and representation of the local geometric structure information and spatial density information of the point cloud.
[0036] 2.2) Constructing an efficient channel attention module: Each point in a point cloud dataset has multiple feature channels, such as coordinates, color, and normal vectors. However, not all feature channels are useful for the task or contribute to the results. The introduction of a Channel Attention Module (CAM) allows the model to adaptively evaluate the importance of each feature channel and accordingly select and enhance features that are crucial to the target task. This helps reduce the impact of redundant information and improves the expressive power and discriminative power of features. Furthermore, the feature dimensions of the input point cloud vary throughout the network. Channel attention can adaptively learn weights based on the dimensions of the input point cloud data, avoiding the unreasonable influence of fixed weight parameters on all dimensions. This enhances the model's adaptability to dimensional changes and improves its generalization performance. Since the detailed features in point cloud data are crucial for point cloud processing tasks, the classic CAM provides the model with the basic ability to adapt channel weights. However, it still has obvious limitations when processing irregular data such as point clouds that are sensitive to details and have dynamically changing dimensions: First, the "dimensionality reduction-dimensionality increase" MLP mode will compress feature channels, resulting in the loss of local geometric details and sparse structural information of the point cloud; Second, the design of fixed scaling factor is difficult to adapt to the changes in channel dimension at different levels of the point cloud network; Third, the computational cost of ordinary convolution is large and cannot meet the real-time processing needs of large-scale point clouds.
[0037] To address these issues, an efficient channel attention (ECA) mechanism is introduced and improved, resulting in a point cloud-adapted efficient channel attention module (Point-ECANet). This module retains the simplicity and attention representation capabilities of the original ECANet architecture while achieving a synergistic improvement in computational efficiency and feature discriminative power through targeted improvements, providing high-quality channel-enhanced feature support for subsequent feature fusion and deep extraction. The core design idea of this module is to adapt to the dynamic changes in the two dimensions of "number of points - number of channels" of point cloud features by replacing the core attention mechanism of ECANet with a lightweight convolutional structure and dynamically adjusting the convolutional kernel size, without changing the core attention mechanism of ECANet, while avoiding the loss of geometric details caused by feature dimension compression. The overall input of the module is a point cloud feature tensor. ,in This represents the number of points in the point cloud. This represents the number of feature channels, and the output is the point cloud feature after channel weight enhancement. First, global average pooling (GAP) is used to calculate the global average of the channel feature values for each point, resulting in... Secondly, ordinary one-dimensional convolution is replaced with depthwise separable one-dimensional convolution. This structure consists of two steps: depthwise convolution and pointwise convolution, forming an efficient processing flow of "intra-channel extraction - inter-channel fusion".
[0038] During the depthwise convolution stage, the output of global pooling is... Each channel is used independently. The convolution kernel performs convolution operations, extracting local correlation features only within a single channel, maintaining a consistent number of input and output channels; in the pointwise convolution stage, a method is used... The convolutional kernel fuses inter-channel information in the output of depthwise convolution. This improvement significantly reduces the number of parameters and computational cost of the module while ensuring the effective extraction of inter-channel correlated features, making it better suited for the real-time processing needs of large-scale, irregular data such as point clouds. To address the dynamic changes in feature dimensions at different levels of point cloud networks, this module improves the calculation method for the convolutional kernel size by incorporating the number of points. The dynamic constraints ensure that the receptive field of the convolution kernel matches the overall information content of the point cloud features. The convolution kernel size is calculated as shown in formula (1).
[0039] (1)
[0040] in, and These are model hyperparameters used to adjust the overall range and rate of change of the kernel size. This indicates taking the largest odd number not greater than the calculated value within the parentheses, ensuring the convolution kernel has central symmetry and improving the stability of channel-related feature extraction. This design allows the convolution kernel size to vary with the number of points. With the number of channels Dynamic adjustments effectively enhance the module's adaptability to irregular point cloud features, ensuring that the module can efficiently extract channel correlation information at different feature levels of the network.
[0041] Next, the output of the depthwise separable one-dimensional convolution is fed into the Sigmoid activation function. Through a non-linear transformation, the feature values are mapped to the [0,1] interval, generating channel attention weights that correspond one-to-one with the number of input channels. The weight values quantify the importance of each feature channel to the current point cloud classification and segmentation task, enabling accurate identification of key and redundant feature channels. Subsequently, the generated channel attention weights are... Point cloud feature tensor compared to the original input of the module A channel-by-channel multiplication operation is performed, and adaptive weighting is used to highlight the information representation of key feature channels while suppressing invalid interference from redundant channels, ultimately obtaining the channel-enhanced point cloud features. This effectively improves the utilization rate of effective features by subsequent network layers.
[0042] 2.3) Construct a parallel attention module. Spatial attention helps capture the spatial structure and geometric information in point cloud data, while channel attention helps improve the feature representation ability of point cloud data and the attention to different features. In point cloud processing, a task with high requirements for fine granularity, serial connection will limit the interaction and information transmission between different attention mechanisms, thereby reducing the model's expressive power and performance.
[0043] To address this issue, spatial attention and channel attention are connected in parallel. This approach allows both attention modules to learn and optimize from the original input features simultaneously, enabling the model to more comprehensively focus on important regions and feature channels in the point cloud data. Furthermore, to facilitate network optimization and training, the idea of skip connections in residual learning is adopted. The original input features are added element-wise to the feature maps processed by each attention module, resulting in a parallel attention module. The input of this module is the feature map obtained from the geometry perception module. B represents the batch size, N represents the number of sampling points, and D represents the number of channels in the feature map. The parallel attention module obtains the corresponding weights by inputting the intermediate feature map F into the spatial attention module and the channel attention module respectively. Then, it multiplies the two weights with the input feature map F to obtain... and Finally, the three are added together to obtain the output feature map. The specific calculation process is shown in formula (2).
[0044] (2)
[0045] in This represents the channel attention weight. This represents the spatial attention weights. This represents element-wise multiplication.
[0046] 3) Construct an end-to-end network structure based on geometric perception and parallel attention, and use the training set to complete model training and parameter optimization;
[0047] 3.1) The network mainly consists of three core modules: geometric perception module, spatial attention module, and efficient channel attention module.
[0048] 3.2) This model uses The original point cloud data is used as input. The farthest point of the point cloud data is sampled as the center point. Then, KNN grouping is used to feed the data into the geometry perception module for multi-dimensional feature extraction.
[0049] 3.3) The obtained features are enhanced in the spatial domain and the channel domain, respectively.
[0050] 3.4) Finally, the concatenated features are fed into multiple multilayer perceptrons for deeper feature mining and used for subsequent classification and segmentation tasks.
[0051] 3.5) Train the model with an initial learning rate of 0.001, an exponential decay learning rate adjustment strategy, and the Adam optimizer.
[0052] 4) Load the trained network parameters, perform end-to-end inference on the test set, and output the point cloud classification and point-by-point segmentation results.
[0053] Example 1:
[0054] The dataset was trained and tested using the Princeton University ModelNet40 classification dataset and the ShapeNetPart dataset, respectively.
[0055] Step 1: The Princeton University ModelNet40 dataset was divided into 9843 training samples and 2468 test samples according to the official standard. The original samples in the dataset were 3D models in .obj format. Preprocessing was required in the experiment, converting them into point clouds through surface information extraction and uniform sampling (1024 points were uniformly sampled in this paper). The ShapeNetPart large-scale part segmentation benchmark dataset was divided into a training set of 12137 samples, a validation set of 1870 samples, and a test set of 2874 samples according to the official standard. This was converted into a point cloud with a fixed number of points (2048 points were uniformly sampled in this paper), allowing it to be adapted to the segmentation model input without additional complex preprocessing.
[0056] Step 2: Set up the model's runtime environment and hyperparameter variables
[0057] In terms of the experimental environment, all experiments were conducted on a Windows server with an Intel(R) Xeon(R) Silver 4110 processor, 24GB of memory, and an NVIDIA RTX A5000 graphics card to accelerate the network training and inference process. Python version 3.8 was used, and the network model was built based on the PyTorch deep learning framework. The specific training parameters for the classification and segmentation datasets are shown below:
[0058] Table 1 Experimental Parameter Settings
[0059]
[0060] Step 3: Construct a point cloud classification and segmentation method model based on geometric perception and parallel attention.
[0061] The main modules include: geometric perception module, spatial attention module, and efficient channel attention module.
[0062] Step 4: Using this model, conduct experiments. The results of the comparison between classification and segmentation experiments are shown in Tables 2 and 3.
[0063] Table 2. Classification results on the ModelNet40 dataset.
[0064]
[0065] surface Part segmentation results table on the ShapeNetPart dataset
[0066]
[0067] This invention proposes a point cloud classification and segmentation method based on geometric perception and parallel attention. It achieves point cloud classification and segmentation based on deep learning, which can fully mine key feature information, effectively improve classification and segmentation accuracy, and is applicable to point cloud processing tasks in various complex scenarios. It has high accuracy, stability and practical value.
Claims
1. A point cloud classification and segmentation method based on geometric perception and parallel attention, characterized in that, The steps are as follows: Step 1) Obtain the point cloud classification and segmentation dataset, and preprocess the dataset to form a .txt file. Step 2) Construct the network model: Step 2.1) For the obtained point cloud classification and segmentation dataset, design a geometric perception module to extract geometric features from the normalized point cloud neighborhood coordinates, and then combine the basic geometric features with geometric affine transformation and kernel density estimation algorithms. Step 2.2) Construct an efficient channel attention module for point cloud adaptation. First, compress channel features through global average pooling. Then, replace ordinary convolution with depthwise separable one-dimensional convolution. The depthwise convolution extracts local features independently for each channel, performs inter-channel information fusion through pointwise convolution, and introduces a formula for calculating the convolution kernel size that dynamically adjusts with the number of points and channels. Finally, attention weights are generated by Sigmoid activation and multiplied with the original input features channel by channel to achieve feature enhancement. Step 2.3) Construct a parallel processing module for channel attention and spatial attention, so that the two attention mechanisms can learn and optimize from the original input features at the same time, avoiding the problem of limited information interaction caused by the serial structure. At the same time, the original features and attention-enhanced features are fused together by residual connection to obtain the output of the parallel attention module. Step 3) Construct an end-to-end network structure based on geometric perception and parallel attention, and use the training set to complete model training and parameter optimization; Step 3.1) The network consists of three core modules: a geometric perception module, a spatial attention module, and an efficient channel attention module; Step 3.2) This model uses The original point cloud data is used as input. The farthest point of the point cloud data is sampled as the center point. Then, KNN grouping is used to feed the data into the geometry perception module for multi-dimensional feature extraction. Step 3.3) Enhance the obtained features in the spatial domain and the channel domain respectively. Step 3.4) Finally, the concatenated features are fed into multiple multilayer perceptrons for deeper feature mining and used for subsequent classification and segmentation tasks. Step 3.5) Train the model with an initial learning rate of 0.001, an exponential decay learning rate adjustment strategy, and the Adam optimizer. Step 4) Load the trained network parameters, perform end-to-end inference on the test set, and output the point cloud classification and point-by-point segmentation results.
2. The point cloud classification and segmentation method based on geometric perception and parallel attention according to claim 1, characterized in that, In step 2, the specific method is as follows: Step 2.1) For the obtained point cloud classification and segmentation dataset, design a geometric perception module to extract and optimize the geometric features of the normalized point cloud neighborhood coordinates, and combine the basic geometric features of the point cloud neighborhood with geometric affine transformation and kernel density estimation algorithm to achieve effective fusion and representation of the local geometric structure information and spatial density information of the point cloud. Step 2.2) Construct an efficient channel attention module. Each point in the point cloud data has multiple feature channels, but not all feature channels are useful for the task or contribute to the result. Introducing a channel attention module enables the model to adaptively evaluate the importance of each feature channel and thereby select and enhance features that are key to the target task. In addition, the feature dimensions of the input point cloud vary throughout the network. Channel attention can adaptively learn weights according to the dimensions of the input point cloud data, avoiding the unreasonable influence of fixed weight parameters on all dimensions. We introduce and improve the efficient channel attention (ECA) mechanism and propose an efficient channel attention module ECANet for point cloud adaptation. While retaining the simplicity of the original ECANet architecture and its attention representation capabilities, the efficient channel attention module achieves a synergistic improvement in computational efficiency and feature discriminative power through targeted improvements, providing high-quality channel-enhanced feature support for subsequent feature fusion and deep extraction. The efficient channel attention module, without altering the core attention mechanism of ECANet, adapts to the dynamic changes in the "number of points - number of channels" of point cloud features by replacing the lightweight convolutional structure and dynamically adjusting the convolutional kernel size, thus avoiding the loss of geometric details caused by feature dimension compression. The overall input of the module is a point cloud feature tensor. ,in This represents the number of points in the point cloud. This represents the number of feature channels, and the output is the point cloud feature after channel weight enhancement. First, global average pooling (GAP) is used to calculate the global average of the channel feature values for each point, resulting in... Secondly, ordinary one-dimensional convolution is replaced with depthwise separable one-dimensional convolution. This structure consists of two steps: depthwise convolution and pointwise convolution, forming an efficient processing flow of "intra-channel extraction - inter-channel fusion". During the depthwise convolution stage, the output of global pooling is... Each channel is used independently. The convolution kernel performs convolution operations, extracting local correlation features only within a single channel, maintaining a consistent number of input and output channels; in the pointwise convolution stage, a method is used... The convolutional kernel performs inter-channel information fusion on the output of the depthwise convolution. To address the dynamic changes in feature dimensions at different levels of point cloud networks, the efficient channel attention module improves the calculation method for the convolutional kernel size by incorporating the number of points. The dynamic constraints ensure that the receptive field of the convolution kernel matches the overall information content of the point cloud features. The kernel size is calculated as shown in formula (1): (1) in, and These are model hyperparameters used to adjust the overall range and rate of change of the kernel size. This design allows the kernel size to vary with the number of points, taking the largest odd number not greater than the calculated value within the parentheses. With the number of channels Dynamic adjustment; Next, the output of the depthwise separable one-dimensional convolution is fed into the Sigmoid activation function. Through a non-linear transformation, the feature values are mapped to the [0,1] interval, generating channel attention weights that correspond one-to-one with the number of input channels. The magnitude of the weight values quantifies the importance of each feature channel to the current point cloud classification and segmentation task; subsequently, the generated channel attention weights are... Point cloud feature tensor compared to the original input of the module By performing channel-by-channel multiplication and adaptively weighting to highlight the information representation of key feature channels, the final point cloud features are obtained after channel enhancement. ; Step 2.3) Construct a parallel attention module by connecting spatial attention and channel attention in parallel. This allows both attention modules to learn and optimize from the original input features simultaneously, enabling the model to more comprehensively focus on important regions and feature channels in the point cloud data. Borrowing the idea of skip connections in residual learning, the original input features are added element-wise to the feature maps processed by each attention module, thus obtaining the parallel attention module. The input to the parallel attention module is the feature map obtained from the geometric perception module. B represents the batch size, N represents the number of sampling points, and D represents the number of channels in the feature map. The parallel attention module obtains the corresponding weights by inputting the intermediate feature map F into the spatial attention module and the channel attention module, respectively. Then, it multiplies the two weights with the input feature map F to obtain the result. and Finally, the three are added together to obtain the output feature map. The specific calculation process is shown in formula (2): (2) in This represents the channel attention weight. This represents the spatial attention weights. This represents element-wise multiplication.