首页|期刊导航|河南城建学院学报|基于原型引导损失的双偏移局部特征融合的点云语义分割网络

基于原型引导损失的双偏移局部特征融合的点云语义分割网络OA

A prototype-guided dual-offset local feature aggregation network for point cloud semantic segmentation

中文摘要英文摘要

针对PointNet++网络局部特征判别性表达能力不足及交叉熵损失正则化能力有限的问题,提出一种基于原型引导的双偏移局部特征融合网络(DOGA-Net).该网络由双偏移引导聚合模块与类别原型引导损失模块构成.前者通过引入兼具方向敏感性与结构依赖性的局部注意力权重,实现邻域特征的判别性聚合,增强对局部几何结构的刻画能力;后者通过约束同类点特征向对应类别原型聚集,提升特征空间中的类内一致性.为验证网络的有效性,在 ISPRS、LASDU 及GML数据集上开展实验.结果表明,与基准网络PointNet++相比,DOGA-Net 在 mF1和OA 上均有提升,其中OA分别达到83.8%、87.7%和95.8%,mF1分别为72.3%、78.0%和70.4%.

PointNet++is limited in extracting local features and the cross-entropy loss lacks regularization capability.To address these issues,a prototype-guided dual-offset local feature fusion network(DOGA-Net)is proposed.The network consists of Dual-Offset Guided Aggregation module and Prototype-Guided Loss module.Dual-Offset Guided Aggregation module utilizes a lightweight attention mechanism based on dual offset to generate attention weights that are both direction sensitive and structural dependent,enabling dis-criminative aggregation of neighborhood features and enhancing the representation of local geometric struc-tures.Prototype-Guided Loss module constructs feature prototypes for each semantic category and enforces point feature to cluster around their corresponding prototypes in the feature space to improve intra-class con-sistency.Experiments on the ISPRS dataset,LASDU dataset and GML dataset are conducted to validate the performance of DOGA-Net.The results show that,compared to the benchmark network PointNet++,DO-GA-Net achieves improvements in mF1 and overall accuracy(OA)with OA reaching 83.8%,87.7%and 95.8%,and mF1 reaching 72.3%,78.0%and 70.4%on the three datasets,respectively,fully demonstrat-ing its effectiveness.

张吉

中国地质大学 数学与物理学院,湖北 武汉 430074

信息技术与安全科学

点云数据语义分割PointNet++局部注意力

point cloudsemantic segmentationPointNet++local attention

《河南城建学院学报》 2026 (4)

90-99,10

国家自然科学基金项目(42374147)

10.14140/j.cnki.hncjxb.2026.04.010

评论