基于特征融合和损失优化的点云语义分割网络
Point Cloud Semantic Segmentation Network Based on Feature Fusion and Loss Optimization
刘起源 1路锦正 2黄炳森1
作者信息
- 1. 西南科技大学 计算机科学与技术学院,四川 绵阳 621010
- 2. 西南科技大学 信息工程学院,四川 绵阳 621010
- 折叠
摘要
针对目前大多数方法仅利用单尺度特征而忽视了具有不同感受野的多尺度特征信息、无法有效处理点云数据集中类别权重不平衡的问题,提出一种基于全阶段特征融合(FSFF)和平衡损失(BL)的分割网络(FFBL-Net).首先,FSFF模块通过将不同编码阶段的可学习特征与当前阶段特征进行融合,促进了浅层和深层语义信息互补;融合后的特征被传递到编码融合模块(EFM)和解码融合模块(DFM),实现了特征的跨阶段融合.此外,为了解决数据集中类别分布不平衡的问题,引入BL损失调整类别间的梯度差异.实验结果表明,FFBL-Net在主流的大规模点云数据集S3DIS上,平均交并比达到了69.7%,总体准确率达到了89.9%.与PointNet++相比,FFBL-Net分别提升了12.4%和6.1%.
Abstract
Aiming at the problem that most of the current methods only use single-scale features but ignore the multi-scale feature information with different receptive fields and cannot effectively deal with unbalanced category weights in point cloud datasets,a segmentation network(FFBL-Net)based on full-stage feature fusion(FSFF)and balanced loss(BL)is proposed.First,FSFF module promotes the complementation of shallow and deep semantic information by integrating learnable features of different coding stages with features of the current stage.The fused features are transferred to the encoding fusion module(EFM)and decoding fusion module(DFM),which realizes the cross-stage fusion of features.In addition,to solve the problem of unbalanced class distribution in the dataset,BL loss is introduced to adjust the gradient difference between categories.The experimental results show that the FFBL-Net on the large-scale point cloud dataset S3DIS has reached69.7%in terms of mean intersection over union(mIoU)and 89.9%in overall accuracy(OA),which is12.4%and 6.1%higher than that of the original PointNet++ respectively.
关键词
点云/语义分割/多尺度特征融合/损失优化/神经网络优化Key words
point cloud/semantic segmentation/multi-level feature fusion/loss optimization/neural network optimization引用本文复制引用
基金项目
国家重点研发计划项目(2019YFB1705100)
四川省科技计划项目(2022ZHCG0001)
黑龙江省重点研发计划项目(2022ZX01A16)
出版年
2024
国家科技期刊平台
NSTL
万方数据