FEATURE EXTRACTION OF 3D POINT CLOUD BASED ON ADGCNN

The 3D point cloud data is sparse and irregular, and the existing methods do not consider the important difference of feature channels when feature extraction is performed. In this regard, this paper proposes an ADGCNN network combining attention mechanism. The channel attention module was introduced into EdgeConv structure, and different weights were allocated according to the importance of characteristic channels to improve the expression ability of the network. The method of maximum pooling and average pooling was used to deal with the disorder of point cloud, and to prevent the information loss caused by only maximum pooling. Experiments show that the classification accuracy of ADGCNN compared with DGCNN is improved from 92.20% to 93.31% on the ModelNet40 point cloud classification dataset, which verifies the effectiveness of the ADGCNN network.