现代电子技术2024，Vol.47Issue(15)：91-97,7.DOI:10.16652/j.issn.1004-373x.2024.15.015
多尺度特征融合对比学习结合PointMetaBase的点云分割
Multi-scale feature fusion contrastive learning combining with PointMetaBase for point cloud segmentation
摘要
Abstract
The accurate segmentation of point cloud scene boundaries is crucial for improving the overall segmentation accuracy of 3D point clouds and the segmentation accuracy of small objects.A novel multi-scale feature fusion contrastive learning(CL)method is designed and integrated into the PointMetaBase network to address the issue of inaccurate segmentation of point cloud scene boundaries in the existing 3D point cloud semantic segmentation methods.The proposed network is termed MFFCL-PMB(multi-scale feature fusion contrastive learning combining with PointMetaBase)for 3D point cloud semantic segmentation.In this network,the input and output of each layer in the decoder are fed into a multi-scale feature extraction network in a parallel and multipath manner.Subsequently,the multi-scale features are concatenated and fused.Finally,the boundary search module(BSM)is used to determine the point cloud scene boundary points.Contrastive learning is carried out on the multi-scale fused features corresponding to the boundary points and enables the network to learn weights that enhance feature discrimination.The mIoU(mean intersection over union)of MFFCL-PMB on the test area 5 of the S3DIS dataset is 70.9%.In comparison with the original PointMetaBase,the mIoU on the boundaries increases by 1.4%,the mIoU in the internal regions increases by 1.2%,and the overall mIoU increases by 1.2%.The increment of mIoU on boundary is greater than that in internal regions,which indicates that the MFFCL-PMB effectively enhances the segmentation performance of PointMetaBase on the boundaries and improves the network's semantic segmentation performance on the entire point clouds.关键词
深度学习/三维点云/语义分割/对比学习/多尺度特征融合/编码器/解码器Key words
deep learning/3D point cloud/semantic segmentation/contrastive learning/multi-scale feature fusion/encoder/decoder分类
信息技术与安全科学引用本文复制引用
杨林杰,张斌,张志圣..多尺度特征融合对比学习结合PointMetaBase的点云分割[J].现代电子技术,2024,47(15):91-97,7.基金项目
国家自然科学基金地区基金(62361013) （62361013）
广西科技重大专项(桂科AA23023017) （桂科AA23023017）
