Research On Point Cloud Feature Extraction Based On Graph Neural Network

With the development of 3D acquisition technology,it is becoming easier to acquire3 D data using devices such as depth sensors and 3D scanners.Point cloud data is the most primitive form of data obtained.It retains the original geometric information of threedimensional space,and is simple in mathematical expression.Therefore,it has attracted the attention of many scholars and a lot of research.With the development of deep learning,the use of neural networks to extract point cloud features for classification,segmentation and other tasks has become a research focus.Studying point cloud feature extraction for tasks such as classification and segmentation not only promotes the development of artificial intelligence techniques,but also helps people better understand and utilize the data information in 3D space.However,due to the disorder and unstructured characteristics of point clouds,it is difficult to directly capture some complex relationships between point clouds when extracting the features of point clouds.Therefore,some networks focus on the extraction of local geometric features of cloud through the neighborhood size of point clouds,but the local features of point clouds are not well extracted.As a common representation of unstructured data,the graph reflects the relationship between each node.Therefore,in order to enrich the local geometric feature representation of point cloud,two modules that use graph convolution to enhance the local feature of point cloud are proposed,and attention mechanism is added to obtain more contextual features.The main contents of this paper are as follows:In the first part,In order to further improve the expression ability of point cloud features,an EDGS module is designed to enhance the point cloud local graph semantic feature,which is composed of semantic feature interactive branches,graph attention branches and feature aggregation.We use K-Nearest Neighbor algorithm to construct local graphs followed by multi-layer perceptron.After that,we use two branches to extract local features.In the first branch,max pooling is used to aggregate local graphs semantic features.Then,the concept of grouping is adopted to guide the semantics of individual channels by using the semantics of group channels with feature interaction.In the second branch,in order to preserve the features of points,we use graph attention to assign different weights on the local graph,and sum to aggregate the features between points.Finally,two learnable parameters are used to adaptively aggregate the local features of the two branches.In the second part,when using layered neural networks to extract point cloud features,a large amount of low-level information is lost due to sampling operations.Moreover,only the neighborhood characteristics are learned,ignoring the context information on the channel.The features of point cloud local geometry and channel context are integrated to enhance the feature representation of point cloud.The point cloud is obtained by the farthest point sampling,the K-Nearest Neighbor algorithm is used to group the point cloud,and the geometric affine and shared residual MLP are used to extract the local geometric features.The local geometric features are supplemented by graph convolution.Channel Transformer is used to extract channel context features for sampled point clouds.After that,the local geometric features and channel context features are fused,then followed by MLP.In order to further enrich the features of the point cloud,the original features and the fused features are concatenated.To get the final rich point cloud features.