Point cloud feature extraction model based on graph neural network and classification segmentation method

Abstract

The invention discloses a point cloud feature extraction model based on a graph neural network and a classification segmentation method. Based on a graph attention method and a graph pooling method of a graph neural network, a point cloud convolution (AG-conv) operator and a point cloud pooling operator (HA-pool) are designed. The rotation invariance is realized; the core idea of the method is to process invariant geometric features by using a graph attention method. Specifically, in combination with a graph attention method, a new convolution operator is designed, and the rotation robust representation of the point cloud is coded; a new pool operator is provided, and thus enlarging the size of a receptive field and regularizing learning features through hierarchical calculation features; a compact neural network based on convolution and a pool operator is provided and is used for three-dimensional object classification and part segmentation.

Description

technical field [0001] The invention relates to the field of deep learning technology, in particular to a point cloud feature extraction model and a classification and segmentation method based on a graph neural network. Background technique [0002] Recent advances in 3D deep learning show that special convolution operators can be designed to work with point cloud data. However, a typical disadvantage is that rotation invariance is often not guaranteed, leading to networks that generalize poorly to arbitrary rotations. In recent years, with the development of deep learning, object recognition technology based on 3D point cloud has made remarkable achievements. Point clouds have sparse and irregular structures, and are usually accompanied by missing parts and noise information. These problems are well resolved by deep learning models, but geometric transformations of point clouds, such as rigid transformations, are still a challenging task. The problem. That is, a more pr...

AI Technical Summary

Problems solved by technology

However, most previous work does not support the inconsistent rotation of the input point cloud

During training, the data is simply augmented with some rotation augmentation, which causes the network to not generalize well to unseen rotations

There are also some convolution operators that achieve rotation invariance, but there is still no consistent prediction for arbitrary rotation data

Images