Point cloud semantic segmentation network model of voxelization network based on deep residual error and attention discarding

The invention relates to a deep residual and attention discarding-based point cloud semantic segmentation network model of a voxelization network, and the model comprises a Deep Inversion Residual module which is used for mining voxel deep information so as to compensate for the loss of point cloud information loss during voxelization; and the Drop Attention module is used for expanding the receptive field range of each voxel for global information. The semantic segmentation performance mIoU of the method is 65.6%, and compared with other methods, the best semantic segmentation performance is achieved. Compared with the Minkowski method adopting sparse convolution, the performance of the method is improved by 2.4%; compared with a latest method GeoSegNet, the performance of the method is improved by 0.7%; the mIoU quantization result shows that the method realizes the best point cloud segmentation performance. For a voxelization segmentation method PVT, the performance of the method provided by the invention