基于三维点云的植株联合任务分割框架

Plant joint task segmentation framework based on 3D point cloud

传统的植株器官分割方法依赖经验选择阈值参数,而当前的深度学习浅层框架可能会导致植株重要的几何特征丢失,并难以有效整合植株的局部和全局特征。因此,提出了一个基于三维点云的植株器官分割网络(local global feature fusion segmentation network,LGF-SegNet)模型,通过引入双权重注意力机制模块和位置编码,更适合在植株点云数据中表达几何特征。在提出的框架的解码层引入特征聚合模块,融合植株点云的局部和全局特征,使得该框架能够关注植株的整体特征轮廓同时保留细节植物纹理(如茎和叶)。实验结果表明,提出的架构在语义分割的交并比、精确率和F1分数的平均值分别达到85.76%、93.18%、91.08%,在实例分割的平均精确率、平均实例覆盖率以及平均实例加权覆盖率达到85.27%、78.46%、79.63%,优于当前流行的植株点云分割任务中使用的深度学习网络架构,并适用于植株语义分割和实例分割的双重任务。这为后续的植株生长预测等研究奠定基础。

The traditional plant part segmentation methods rely on empirical selection of threshold parameters,while the current shallow deep learning framework may lead to the loss of important geometric features of the plant cloud,and it is difficult to effectively integrate the local and global features of the plant.Therefore,a plant part segmentation network was proposed on 3D Point Cloud(LGF-SegNet),which was more suitable for expressing geometric features in plant point-cloud data by introducing double-weighted attention mechanism module and location coding.A feature aggregation module was introduced into the decoding layer of the proposed framework to fuse the local feature and global feature of the plant point cloud,so that the framework could focus on the overall feature outline of the plant while preserving the detailed plant textures(such as stems and leaves).The experimental results show that the average of intersection ratio,precision and F1 score of semantic segmentation reach 85.76%,93.18%and 91.08%,respectively.The mean precision,mean coverage and mean weighted coverage of instance segmentation reach 85.27%,78.46%and 79.63%,the proposed architecture is better than the current deep learning network architecture used in the current plant point cloud segmentation task,and is suitable for the dual tasks of plant semantic segmentation and instance segmentation.This lay a foundation for the subsequent research on plant growth prediction.