基于注意力机制的植物三维点云语义分割方法

3D Plant Point Cloud Semantic Segmentation Method APSegNet Based on Attention Mechanism

摘要:

在植物表型分析中，植物器官分割是实现自动、准确、无损、高通量表型参数测量的关键。传统的植物器官分割方法凭借经验手动设置参数和调整算法，而现有的基于深度学习的分割方法存在对局部特征和全局特征表达能力不足的缺陷。针对以上问题，本文提出一个基于注意力机制的植物三维点云语义分割网络（APSegNet）。在编码阶段提出了一种基于注意力机制的局部（邻域）特征提取方法，充分利用多级点云特征，提高了网络提取点云局部（邻域）特征的能力。在解码阶段提出了一种结合特征距离和空间距离的双近邻插值上采样方法，更准确地恢复下采样时丢失的点云特征，进一步增强了网络对局部特征的表达能力。同时引入通道和多头空间自注意力机制，增强网络对某些重要通道的关注和全局几何结构的捕捉能力，提高了网络对全局特征的表达能力。在多种植物点云数据集上的实验结果表明，该方法语义分割平均交并比分别达到87.32%、79.68%、94.73%、91.43%、95.02%，均优于DGCNN、PointCNN、ShellNet等目前流行的深度学习网络。通过交叉验证实验和消融实验，证实了网络泛化性和有效性。在ShapeNet数据集上进行了相关实验，该网络在其他非植物三维点云目标语义分割任务上也取得了较好的分割结果。

Abstract:

In plant phenotypic analysis, plant organ segmentation is the key to achieve automatic, accurate, non-destructive and high-throughput phenotypic parameter measurement. Traditional plant organ segmentation methods rely on experience to manually set parameters and adjust algorithms, but the existing deep learning based segmentation methods have insufficient ability to express local and global features. In order to solve these shortcomings, a semantic segmentation network for three-dimensional point clouds was proposed based on attention mechanism. In the coding stage, a local feature extraction method based on attention mechanism was proposed, which made full use of multilevel point cloud features and improved the ability of network to extract local feature of point cloud. In the decoding stage, a double nearest neighbor interpolation upsampling method combining feature distance and spatial distance was proposed to recover the lost point cloud features in downsampling more accurately, and further enhanced the expression ability of local features. At the same time, the channel and multi-head spatial self-attention mechanism were introduced to enhance the attention of the network to some important channels and ability to capture the global geometric structure, and improve the expression ability of the network for global features. Experimental results on a variety of plant point cloud datasets showed that the mean intersection over union of semantic segmentation of the proposed method reached 87.32%, 79.68%, 94.73%, 91.43%, 95.02%, respectively, which were better than those of popular deep learning networks such as DGCNN, PointCNN, ShellNet and so on. Cross validation experiments and ablation experiments were carried out to confirm the generalization and effectiveness of the network. Relevant experiments were carried out on ShapeNet dataset, and the network also achieved good segmentation results on other non plant 3D point cloud target semantic segmentation tasks.