A point cloud semantic segmentation method based on detail interaction-aware augmentation network

Abstract

This invention proposes a point cloud semantic segmentation method based on a detail interaction-aware enhancement network. This method belongs to the field of 3D computer vision technology. Based on a U-shaped network architecture, this method introduces three core modules: a detail interaction module based on cross-attention, which enhances the representation of local geometric features by fusing kernel convolutions and cross-attention mechanisms; a receptive field expansion and fusion module, which enhances feature discrimination by utilizing multi-scale convolutions and optimizes encoding and decoding feature fusion; and a local perception enhancement module, which reconstructs local structural information during the upsampling process through multi-scale convolutions. This invention achieves superior performance compared to existing methods on multiple public datasets, effectively improving the accuracy and detail preservation capability of point cloud semantic segmentation.

Description

Technical Field

[0001] This invention belongs to the field of point cloud semantic segmentation technology, and relates to a point cloud semantic segmentation method based on detail interaction-aware augmented network. Background Technology

[0002] With the continuous maturation and widespread adoption of LiDAR technology, point clouds have become an important data carrier for describing the three-dimensional world. Compared to traditional two-dimensional imagery, point clouds can directly record the geometric coordinates (X, Y, Z) and echo attributes (intensity, echo count, RGB, etc.) of targets, offering advantages such as high accuracy, minimal susceptibility to illumination, and the ability to penetrate vegetation. They are widely used in smart cities, autonomous driving, 3D reconstruction, disaster monitoring, and forest surveys. However, the disordered, sparse, unstructured nature of point cloud data, along with significant differences in ground feature scale and highly uneven category distribution, poses a significant challenge to its automated interpretation. How to efficiently and accurately assign semantic labels to each point—i.e., point cloud semantic segmentation—has become a research hotspot at the intersection of photogrammetry and remote sensing, computer vision, and robotics.

[0003] Early point cloud segmentation primarily relied on machine learning paradigms. Researchers first manually designed geometric features (such as elevation, normal vectors, curvature, FPFH, Spin Image, etc.) based on expert knowledge, and then used random forests, SVMs, AdaBoost, or conditional random fields for supervised classification. These methods are highly interpretable, computationally inexpensive, and can achieve acceptable results in small-scale scenarios. However, feature engineering is highly dependent on experience and difficult to transfer; moreover, traditional classifiers have limited ability to model complex spatial contexts and long-distance dependencies, and their accuracy and robustness rapidly decline when faced with large-scale, high-density, multi-echo airborne laser scanning (ALS) point clouds.

[0004] In 2017, PointNet pioneered end-to-end deep learning to directly process raw point coordinates, utilizing a shared multilayer perceptron (MLP) and max pooling to address the invariance of point cloud arrangement. This breakthrough on ModelNet40 and ShapeNet ushered in the era of point-based deep learning. Subsequently, PointNet++ significantly improved fine-grained segmentation capabilities by extracting multi-scale neighborhood features through hierarchical sampling and local aggregation. Following this, numerous other methods emerged, including MLP-based RandLA-Net and RPNet, graph convolution-based DGCNN and GACNet, and Transformer-based Point Transformer, PCT, and Stratified Transformer, continuously improving public leaderboards through more flexible local-global relationship modeling. However, pure MLPs or graph networks often neglect the continuity and geometric structure of the 3D spatial distribution of point clouds; while Transformer-based methods can capture long-range dependencies, their secondary complexity in large-scale urban scenarios incurs significant memory and computational overhead, limiting their engineering application.

[0005] To balance local geometry preservation and linear complexity, Thomas et al. proposed Kernel Point Convolution (KPConv) in 2019. KPConv places learnable kernel points within a sphere's neighborhood and aggregates features by weighting neighborhood points using kernel functions, combining the translation / deformation adaptability of convolutional networks with the geometric rigidity of point clouds. KPConv does not require voxelization or projection and can directly process point clouds with tens of millions of points, achieving leading results on benchmarks such as Semantic3D, S3DIS, and ISPRS, thus becoming the representative framework in the field of point convolution.

[0006] In general, point cloud processing methods in the field of deep learning can be divided into three categories: projection-based methods, voxel-based methods, and point-based methods.

[0007] Projection-based methods project 3D point clouds onto a 2D image plane and process them using a 2D convolutional neural network. Although computationally efficient, the projection process leads to the loss of geometric information, which limits its performance, especially in complex scenes.

[0008] Voxel-based methods divide point clouds into regular voxel meshes and process them using 3D convolution. While this method preserves spatial structure, it introduces quantization errors and incurs significant computational and memory overhead, making it particularly difficult to apply in high-resolution scenes.

[0009] Point-based methods directly process the raw point cloud, avoiding information loss caused by projection or voxelization. Representative works include PointNet, PointNet++, and KPConv, which extract features through multilayer perceptrons (MLPs) and pooling operations.

[0010] While deep learning has propelled point cloud semantic segmentation accuracy to new heights, existing methods still face bottlenecks such as insufficient local geometric characterization and poor coupling between shallow structures and deep semantics, leading to premature dilution of detailed features during multiple downsampling. To address this, this invention proposes a detail interaction-aware enhancement network that explicitly models local geometric relationships and adaptively fuses cross-layer features to fully exploit easily overlooked microstructural information, achieving a further breakthrough in accuracy. Summary of the Invention

[0011] This invention proposes a detail-interaction-aware enhancement network that explicitly models local geometric relationships and adaptively fuses cross-layer features to fully exploit easily overlooked microstructural information, achieving a breakthrough in accuracy. Its core consists of the following three parts:

[0012] 1. Detailed interaction module based on cross-attention

[0013] The overall network of this invention is a modification and innovation of the traditional U-shaped network architecture, which generally consists of an encoder, a decoder, and ordinary skip connections. The encoder is responsible for extracting semantic features from the point cloud, while the decoder generates fine-grained segmentation results through upsampling and feature fusion. Skip connections preserve the low-level fine-grained features of the encoder, which helps in the prediction of fine segmentation boundaries and alleviates the problem of information loss.

[0014] This invention designs a detail interaction module based on cross-attention in the encoder stage. The module mainly consists of two core parts: one is kernel convolution based on cross-attention, and the other is a cross-complementary mechanism implemented through multiply-accumulate operations.

[0015] Kernel convolution based on cross-attention is a structure that fuses kernel convolution with a self-designed cross-attention mechanism. Kernel convolution is a deformable convolution method that directly operates on point clouds. It replaces the fixed grid of traditional convolution kernels with a set of learnable "kernel points," each carrying a weight matrix. For any neighborhood in the input point cloud, the network first calculates the spatial distance from each point in the neighborhood to each kernel point, obtains the weight coefficients through a linear correlation function, and then performs a weighted summation of the features to achieve convolution on a continuous space. Furthermore, it can achieve deformable convolution by learning kernel point offsets, thus flexibly adapting to local geometry. However, this learning method can only effectively learn local geometric structures and cannot connect the context. Although attention mechanisms can effectively solve this problem, traditional methods often use channel attention or spatial attention alone, making it difficult to simultaneously capture more discriminative feature channels and more critical spatial locations, resulting in insufficiently accurate feature enhancement.

[0016] Therefore, we fuse kernel convolution and cross-attention. Specifically, we first aggregate local features within the spherical neighborhood using kernel convolution, then divide the learning results of kernel convolution into two paths, inputting them into sequential attention and parallel attention respectively. Sequential attention consists of channel attention and spatial attention connected in series, while parallel attention consists of channel attention and spatial attention connected in parallel. This structure can work collaboratively, simultaneously evaluating and weighting the importance of feature channels and spatial locations, effectively avoiding the weighting bias that may arise from a single attention mechanism. Finally, a gating fusion mechanism is used to integrate the weighted features, enabling the model to more accurately focus on regions that possess both channel discriminative and spatial criticality. The specific definitions are as follows:

[0017]

[0018] First, the input point cloud data is processed through kernel convolution. Learn local features, It is the input point cloud data. It is the output of kernel convolution. It is a kernel convolution operation.

[0019]

[0020]

[0021]

[0022] Output Through serial attention and parallel attention respectively, It is the output of channel attention in the serial attention branch, where MLP is a multilayer perceptron, Avg is average pooling, and Max is max pooling. It is the output of the serial attention branch, where It is a one-dimensional convolution operation. It's a concatenation operation; Mean is average pooling, and Max is max pooling. It is the output of the parallel attention branch.

[0023]

[0024] Finally, the result is output through gating. It is the final output of the module. It's a gating operation.

[0025] However, feature extraction through multiple stacked kernel convolutions can blur the extracted feature information, leading to feature degradation. Therefore, we design a framework based on cross-attention kernel convolutions, which performs a detail interaction between each convolution. This cross-complementary mechanism strengthens and enriches the semantic features of details, preventing feature blurring due to interference.

[0026] The cross-complementary mechanism interacts the output of kernel convolution based on cross-attention through multiplication and addition paths. The multiplication path multiplies the two inputs, enhancing shared high-frequency details and suppressing non-shared noise, preserving key details agreed upon by both branches. The addition path adds the two inputs, compensating for weak details lost during multiplication, balancing high and low frequencies, and preventing the multiplication path from disrupting the global structure by excessively focusing on high frequencies. The specific definitions are as follows:

[0027]

[0028]

[0029] The cross-complementary mechanism has two inputs, processing the input point cloud data through two layers. The operation yields two inputs respectively. It is the input of the first branch. It is the input of the second branch. The operation is the kernel convolution based on cross-attention mentioned above.

[0030]

[0031]

[0032] The above formulas are the specific definitions of addition branches and multiplication branches, where It is the output of the addition branch; It is the output of the multiplication branch; It is an activation function.

[0033]

[0034] Finally passed The outputs of the multiplication and addition branches are concatenated and processed. The operation yields the module's final output, where It is the final output of the detail interaction module based on cross-attention.

[0035] 2. Receptive field expansion and fusion module

[0036] Traditional U-shaped networks often directly concatenate encoder-decoder features in their skip connections. However, directly fusing shallow geometric details with deep semantic information can lead to strong noise overshadowing crucial information. This coarse fusion method is prone to feature interference, resulting in a chaotic fusion outcome. Therefore, multi-scale perceptual fusion can make skip connection features more discriminative and reduce interference during fusion.

[0037] The receptive field expansion and fusion module inputs the skip-connected features into different convolutions with multiple branches in parallel. Different convolution kernels are used to perceptually expand the features, enrich the detailed information of the original features, and enhance the responsiveness of the features at different scales. This makes the skip-connected information more obvious and refined than before when the features are fused in the decoding stage, and it is less likely to cause feature interference.

[0038] The module has two inputs: skip-connected features and upsampled features. These two inputs are then processed through a receptive field enhancement operation to obtain enhanced features. The receptive field enhancement module consists of four parallel 1D convolutions with kernels of 1, 3, 5, and 7. Each parallel branch is concatenated and passed through a fully connected layer, followed by a residual structure output. The specific definitions are as follows:

[0039]

[0040]

[0041]

[0042] in, It is a jump connection feature; It is an upsampling feature; It is the output of the jump connection feature through the receptive field enhancement operation; It is the output of the upsampled features through the receptive field enhancement operation; It is a receptive field enhancement operation; It is a fully connected layer; It's a splicing operation; It is a one-dimensional convolution with a kernel size of 1; It is a one-dimensional convolution with a kernel size of 3; It is a one-dimensional convolution with a kernel size of 5; It is a one-dimensional convolution with a kernel size of 7.

[0043] After obtaining the enhanced receptive field outputs from both paths, the two outputs are concatenated. Channel attention weights are generated through a nonlinear layer. The calculated weights are then multiplied channel-by-channel by the two multi-scale features to achieve channel-level recalibration, suppress noise channels, and enhance discriminative channels. Finally, the results are summed and fused to obtain the fused feature output. The specific definitions are as follows:

[0044]

[0045]

[0046] in Channel attention weights are generated through a nonlinear layer, including Activation functions and batch normalization operations ; It is a fully connected layer; It's a splicing operation; It is the final fusion output of the receptive field expansion fusion module.

[0047] 3. Local Perception Enhancement Module

[0048] Although point-based methods like kernel convolution can effectively capture the geometric distribution features of point clouds during the encoding stage, in the decoding stage (i.e., the upsampling stage), only a simple multilayer perceptron is used to recover the scale information of the point cloud. This single-layer perceptron recovery is prone to losing key features and is difficult to reconstruct the local context, resulting in insufficient point cloud scale restoration and further weakening of edges and geometric details.

[0049] To address the aforementioned issues, we designed a local perception module applied to the network's decoder. Specifically, the module's input is upsampled point cloud features, which have recovered the number of points but lack fine local structure. The features with recovered points are then input into a nonlinear transformation layer, including a multilayer perceptron, batch normalization, activation functions, and random deactivation, thereby improving the network's nonlinear modeling capability, making the features learn faster, more stably, and without overfitting. The specific definitions are as follows:

[0050]

[0051] in, It is a batch normalization operation that pulls each batch of data to the same distribution, so that the gradient is no longer amplified or reduced by the scale, and the training curve is smoother. It is an activation function that makes the entire network a piecewise linear function, which can approximate any complex surface and alleviate gradient vanishing. It is a random deactivation operation. During retraining, some neurons are randomly deactivated, forcing the remaining nodes to learn redundant representations and reducing overfitting. It is the input of the local sensing module; It is the output of the nonlinear transformation layer.

[0052] The output of the obtained nonlinear transformation layer is reduced in size by average pooling in the spatial dimension to obtain smoother features. These features are then passed through parallel one-dimensional convolutions with different kernels, allowing them to focus on different local information through different receptive fields. The convolution kernels of the parallel one-dimensional convolutions are 5, 7, and 11. A one-dimensional convolution with a 5-kernel can perceive individual targets in a small area, such as telephone poles or cars; a one-dimensional convolution with a 7-kernel can perceive continuous targets in a medium area, such as slender low vegetation and power lines; and a one-dimensional convolution with an 11-kernel can perceive group targets in a large area, such as continuous rooftops. The output of the parallel one-dimensional convolutions is then processed through… The activation function compresses the corresponding weight score, which is then multiplied by the smoothed features obtained from average pooling and concatenated to obtain the point cloud feature representation after local perception enhancement. Finally, it passes through a nonlinear layer and residual connections to obtain the final output of the local perception enhancement module. The specific definitions are as follows:

[0053]

[0054]

[0055]

[0056]

[0057]

[0058]

[0059] in, It is an average pooling operation; This is the result after average pooling; It is a one-dimensional convolution with kernel sizes of 5, 7, and 11; It is the output of three parallel convolutions; It is the result of concatenating the outputs of three parallel convolutions; This is the final output of the local perception enhancement module. Attached Figure Description

[0060] Figure 1 A schematic diagram of a point cloud semantic segmentation model for a detail interaction perception enhancement network for the invention;

[0061] Figure 2 This is a schematic diagram of a kernel convolution module based on cross-attention;

[0062] Figure 3 This is a schematic diagram of a detail interaction module based on cross-attention;

[0063] Figure 4 A schematic diagram of the receptive field expansion and fusion module;

[0064] Figure 5 This is a schematic diagram of the local perception enhancement module;

[0065] Figure 6 The three-dimensional visualization results of the method and comparison method of the present invention in various typical scenarios selected in the dataset; Detailed Implementation

[0066] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0067] Step 1: Preprocess the point cloud data used;

[0068] Step 2: Input the preprocessed data into the encoder's cross-attention-based detail interaction module for feature extraction;

[0069] Step 3: Input the upsampled features and encoder skip connection features into the receptive field expansion fusion module for fusion;

[0070] Step 4: Input the fused features into the local perceptual enhancement module of the decoder for feature recovery and enhancement;

[0071] Step 5: Restore the decoder output to its initial size to generate semantic segmentation results;

[0072] Furthermore, step 1 specifically includes the following steps:

[0073] Step 1-1: To obtain a uniform intensity distribution, we first preprocess the intensity values ​​using gamma transform. (X, Y, Z) and the gamma transform intensity are used as inputs to the segmentation algorithm. Furthermore, the point cloud data contains too many points, making direct computation and learning difficult. To facilitate computation, we use grid sampling to obtain subsampled point cloud data.

[0074] Steps 1-2: Then, perform some data augmentation operations on the input point cloud data. Randomly rotate the input points along the Z-axis to enhance the network's robustness to orientation, and add 0.1 cm of random noise to the XYZ coordinates to enhance the network's robustness to noise.

[0075] Furthermore, step 2 specifically includes the following steps:

[0076] Step 2-1: As Figure 1The diagram shows a point cloud semantic segmentation model of the detail interaction perception enhancement network of the invention. The network adopts a traditional U-shaped network encoding and decoding structure. The preprocessed data is input into the detail interaction module based on cross-attention in the encoder for feature extraction. The module has two core parts: a kernel convolution module based on cross-attention, as shown in the diagram. Figure 2 As shown, the detailed interaction module based on cross-attention is as follows: Figure 3 As shown, the encoder consists of five stacked detail interaction modules based on cross-attention. Each layer is connected by downsampling. The output feature dimensions of each layer are 128, 256, 512, 1024, and 2048, respectively, and the grid sampling intervals are 0.45, 0.9, 1.8, 3.6, and 7.2 m, respectively. The corresponding convolution radii are 1.125, 2.25, 4.5, 9, and 1.25, respectively.

[0077] Step 2-2: Taking one layer as an example, the preprocessed data is first input into a kernel convolution. The kernel convolution aggregates local features within the spherical neighborhood. Then, the learning results of the kernel convolution are divided into two paths and input into sequential attention and parallel attention, respectively. Sequential attention consists of channel attention and spatial attention connected in series, while parallel attention consists of channel attention and spatial attention connected in parallel. Finally, the weighted features are integrated through a gating fusion mechanism.

[0078] Steps 2-3: Define the above kernel convolution based on cross-attention as... The input point cloud data is processed through two layers. The operations yield two inputs, which serve as the two inputs for the cross-complement mechanism. The cross-complement mechanism consists of a multiplication path and an addition path, defined as follows:

[0079]

[0080]

[0081] in To pass through two layers The operation yields two inputs; It is the result of the addition path; It is the result of the multiplication path.

[0082] Steps 2-4: Finally, fuse the components using a concatenation operation and a one-dimensional convolution operation. and The final module output is obtained.

[0083] Furthermore, step 3 specifically includes the following steps:

[0084] Step 3-1: As Figure 4The diagram shows the receptive field expansion and fusion module of the invention. The module has two inputs: skip connection features. and upsampling features The skip connection feature is the output of the detail interaction module based on cross attention in each layer of the encoder, and the upsampling feature is the output of the entire encoder after upsampling to recover the number of point clouds.

[0085] Step 3-2: The two inputs are processed through a receptive field enhancement operation to obtain enhanced features. This operation consists of four parallel 1D convolutions with kernel sizes of 1, 3, 5, and 7. The specific definitions are as follows:

[0086]

[0087]

[0088]

[0089] in, It is the output of the jump connection feature through the receptive field enhancement operation; It is the output of the upsampled features through the receptive field enhancement operation; It is a receptive field enhancement operation; It is a fully connected layer; It's a splicing operation; It is a one-dimensional convolution with a kernel size of 1; It is a one-dimensional convolution with a kernel size of 3; It is a one-dimensional convolution with a kernel size of 5; It is a one-dimensional convolution with a kernel size of 7.

[0090] Step 3-3: After obtaining the outputs of the two enhanced receptive fields, the two outputs are concatenated. Channel attention weights are generated nonlinearly layer by layer. The calculated weights are multiplied channel by channel of the two multi-scale features, and finally added together to obtain the fused feature output. The specific definitions are as follows:

[0091]

[0092]

[0093] in Channel attention weights are generated through a nonlinear layer. It is the final fusion output of the receptive field expansion fusion module.

[0094] Furthermore, step 4 specifically includes the following steps:

[0095] Step 4-1: Input the fused features into the local perceptual enhancement module in the decoder for feature enhancement. First, input the features... The input is fed into a nonlinear transformation layer, including a multilayer perceptron. Normalization Activation function and random deactivation operation The output of the nonlinear layer is obtained. .

[0096] Step 4-2: Calculate the output of the obtained nonlinear transformation layer. In the spatial dimension, average pooling is performed. Reduce the size of the feature maps. Then, pass the features through parallel one-dimensional convolutions with different kernels: 5, 7, and 11 kernels respectively. The output of the parallel one-dimensional convolutions is then passed through an activation function. Compression yields the corresponding weight scores, which are then compared with average pooling. The smoothed feature dot product is then concatenated to obtain the point cloud feature representation after local perceptual enhancement. The specific definition is as follows:

[0097]

[0098]

[0099]

[0100]

[0101] in, It is a one-dimensional convolution with kernel sizes of 5, 7, and 11. It is the output of three parallel convolutions.

[0102] Step 4-3: Finally, pass through a nonlinear layer and residual connections to obtain the final output of the local sensing enhancement module. .

[0103]

[0104] Furthermore, step 5 specifically includes the following steps:

[0105] After upsampling four times, the decoder restores the features to the same size as the original input. The feature map output by the decoder is processed by a multilayer perceptron (MLP). The MLP classifies each point based on the learned features and predicts the corresponding semantic label, thus completing the fine semantic segmentation of the point cloud.

[0106] We used three publicly available datasets—ISPRS 3D, DFC2019, and LASDU—to validate the advancement and effectiveness of our invention. The ISPRS 3D dataset was acquired using a Leica ALS 50 system with a field of view of 45 nm and an average flight altitude of 500 m. It is labeled with nine semantic categories, including power lines, low vegetation, impermeable surfaces, cars, fences, roofs, facades, shrubs, and trees. The training set contains 753,876 points, and the test set contains 411,722 points. The DFC2019 dataset is an ALS dataset released by the IEEE Geoscience and Remote Sensing Society in 2019. This dataset was captured in the cities of Omaha and Jacksonville, with a pulse interval of 0.8 m and a coverage area of ​​approximately 100 km. 2 The dataset consists of 200 million points. The labels include five semantic categories: ground, tall vegetation, buildings, water, and bridge surfaces. The LASDU dataset is a large ALS point cloud dataset, photographed from approximately 1200 meters above the Heihe River basin in northwestern China using a Leica ALS 70. It contains approximately 3.12 million points, categorized into ground, buildings, trees, low vegetation, and cultural relics, covering a 1 square kilometer urban area with high-density residential and industrial buildings.

[0107] In the field of point cloud processing, PointNet++, PointCNN, and KPConv are recognized classic methods, while VD-LAB, RFFS-Net, and IPConv represent recent innovative advancements. We compare these methods with our present invention. Our invention uses the Avg.F1 (Average F1 Score) to quantitatively evaluate and analyze our method and several other point cloud semantic segmentation methods. The Avg.F1 value ranges from 0 to 1; a value closer to 1 indicates better performance. The results are shown in Table 1. The experimental results show that our method significantly outperforms the compared algorithms, thus verifying the superiority of our method.

[0108] Table 1. Comparison of mF1 semantic segmentation for point clouds

[0109]

[0110] The results are shown in Table 1. The experimental results demonstrate that the method of this invention significantly outperforms the comparative algorithm, thus verifying the superiority of the method of this invention.

Claims

1. A point cloud semantic segmentation method based on a detail interaction-aware augmentation network, characterized in that: Includes the following steps: S1. Design a detail interaction module based on cross-attention to enhance feature extraction and representation capabilities. The core of this module is a kernel convolution module based on cross-attention: local neighborhood information is modeled by kernel convolution on the input point cloud, and global context is extracted by averaging and max pooling. Feature transformation and enhancement are achieved using MLP and convolutional layers, enabling the model to adaptively focus on key channels and spatial locations to obtain more discriminative point cloud representations. Furthermore, a cross-complementary mechanism including addition and multiplication paths is introduced to form a complete detail interaction module based on cross-attention. In the encoding stage, common high-frequency details are strengthened and weak details are compensated to enrich semantic information and prevent feature degradation. S2. Construct a receptive field expansion and fusion module, extract multi-scale features through different convolution kernels in parallel with multiple branches, enhance the responsiveness of features at different scales, and introduce a channel attention mechanism to perform information interaction and weight adjustment between channels. S3. Construct a local perception enhancement module to perceive different local information through multi-scale convolution during the decoding stage, thereby enhancing the local context reconstruction capability of point cloud features. S4. The decoder output is processed by upsampling and multilayer perceptron to generate the final point cloud semantic segmentation result.

2. The point cloud semantic segmentation method based on detail interaction-aware augmentation network according to claim 1, characterized in that, Step S1 includes the following sub-steps: S1.1 First, the intensity values ​​are preprocessed using gamma transform; (X, Y, Z) and gamma transform intensity are used as inputs to the segmentation algorithm; and for ease of calculation, grid sampling is used to obtain subsampled point cloud data; then, some data augmentation operations are performed on the input point cloud data, the input points are randomly rotated along the Z-axis to enhance the network's robustness to orientation, and 0.1 cm of random noise is added to the XYZ coordinates to enhance the network's robustness to noise; The S1.2 encoder consists of five stacked detail interaction modules based on cross-attention. Each layer is connected by downsampling. The output feature dimensions of each layer are 128, 256, 512, 1024, and 2048, respectively, and the grid sampling intervals are 0.45, 0.9, 1.8, 3.6, and 7.2 m, respectively. The corresponding convolution radii are 1.125, 2.25, 4.5, 9, and 1.25, respectively. Taking one layer as an example, the preprocessed data is first input into the kernel convolution. The kernel convolution aggregates local features in the spherical neighborhood. Then, the learning result of the kernel convolution is divided into two paths and input into the serial attention and parallel attention, respectively. Serial attention is composed of channel attention and spatial attention connected in series, while parallel attention is composed of channel attention and spatial attention connected in parallel; finally, the weighted features are integrated through a gating fusion mechanism. S1.3 Define the above kernel convolution based on cross-attention as The input point cloud data is processed through two layers. The operations yield two inputs, which serve as the two inputs for the cross-complementary mechanism. The cross-complementary mechanism consists of a multiplication path and an addition path, and is finally fused through a concatenation operation and a one-dimensional convolution operation to obtain the final module output; defined as follows: ； ； in These are the two inputs obtained through two layers of operations; It is the result of the addition path; It is the result of the multiplication path.

3. The point cloud semantic segmentation method based on detail interaction-aware augmentation network according to claim 1, characterized in that, Step S2 includes the following sub-steps: S2.1 Directly fusing shallow geometric details with deep semantic information can cause strong noise to overwhelm key information. This coarse fusion method is prone to feature interference, making the fusion result chaotic. Therefore, multi-scale perceptual fusion can make skip-connected features more discriminative and reduce interference during fusion. The receptive field expansion fusion module has two inputs: skip-connected features. and upsampling features The skip connection feature is the output of the detail interaction module based on cross attention in each layer of the encoder, and the upsampling feature is the output of the entire encoder after upsampling to recover the number of point clouds. S2.2 The two inputs are processed separately to obtain enhanced features. The receptive field enhancement operation consists of four parallel 1D convolutions with kernel sizes of 1, 3, 5, and 7; the specific definitions are as follows: ； ； ； in, It is the output of the jump connection feature through the receptive field enhancement operation; It is the output of the upsampled features through the receptive field enhancement operation; It is a receptive field enhancement operation; It is a fully connected layer; It's a splicing operation; It is a one-dimensional convolution with kernel sizes of 1, 3, 5, and 7; After obtaining the outputs of the two enhanced receptive fields in S2.3, the two outputs are concatenated, and channel attention weights are generated through nonlinear layer-by-layer generation. The calculated weights are multiplied by the two multi-scale features channel by channel, and finally added together to obtain the fused feature output.

4. The point cloud semantic segmentation method based on detail interaction-aware augmentation network according to claim 1, characterized in that, Step S3 includes the following sub-steps: S3.1 The fused features are input into the local perceptual enhancement module in the decoder for feature enhancement; firstly, the features are... The input is fed into a nonlinear transformation layer, including a multilayer perceptron. Normalization Activation function and random deactivation operation ; Obtain the output of the nonlinear layer ; S3.2 The output of the obtained nonlinear transformation layer In the spatial dimension, average pooling is performed. The feature map size is reduced; then the features are passed through parallel one-dimensional convolutions with different kernels, using 5, 7, and 11 kernels respectively; the output of the parallel one-dimensional convolutions is passed through an activation function. Compression yields the corresponding weight scores, which are then compared with average pooling. The smoothed feature dot product is then concatenated to obtain the point cloud feature representation after local perception enhancement; the specific definition is as follows: ； ； ； ； in, It is a one-dimensional convolution with kernel sizes of 5, 7, and 11; It is the output of three parallel convolutions; S3.3 Finally, through one more nonlinear layer and residual connection, the final output of the local sensing enhancement module is obtained. .

5. The point cloud semantic segmentation method based on detail interaction-aware augmentation network according to claim 1, characterized in that, Step S4 includes the following sub-steps: After upsampling four times, the decoder restores the features to the same size as the original input. The feature map output by the decoder is processed by a multilayer perceptron (MLP). The MLP classifies each point based on the learned features and predicts the corresponding semantic label, thus completing the fine semantic segmentation of the point cloud.

Images