基于MRE-PointNet+AE的绿萝叶片外形参数估测算法

为了准确、高效、自动获取植物叶片外形参数,提出一种基于多分辨率编码点云深度学习网络(MRE-PointNet)和自编码器模型的绿萝叶片外形参数估测算法。使用Kinect V2相机以垂直姿态获取绿萝叶片点云数据,采用直通滤波、分割、点云精简算法对数据进行预处理,通过测定的叶片外形参数反演绿萝叶片几何模型,并计算几何模型的叶长、叶宽、叶面积。将不同参数组合构建的几何模型离散成点云数据输入MRE-PointNet网络,得到几何模型叶片外形参数估测的预训练模型。针对拍摄过程中存在的叶片部分遮挡和噪声问题,采用自编码器网络对点云数据进行二次处理,以几何模型离散的点云数据作为输入,经过编码解码运算得到自编码器的预训练模型,提升了MRE-PointNet网络在遮挡情况下对叶片外形参数估测的鲁棒性。试验共采集300片绿萝叶片点云数据,按照2∶1比例进行划分,以其中200片点云数据作为训练集,对预训练模型MRE-PointNet做模型迁移的参数微调,以剩下的100片点云数据作为测试集,评估模型对绿萝叶片外形参数的估测能力。采用本文算法将外形参数估测值和真实值进行数学统计与线性回归分析,得出叶长、叶宽和叶面积估测的R^2和RMSE分别为0.9005和0.4170 cm、0.9131和0.3164 cm、0.9447和3.8834 cm^2。试验表明,基于MRE-PointNet和自编码器模型的绿萝叶片外形参数估测算法具有较高的精确度和实用性。

Estimation Algorithm of Leaf Shape Parameters of Scirpus sibiricum Based on MRE-PointNet and Autoencoder Model

In order to obtain the leaf shape parameters of plant leaves efficiently,accurately and automatically,a multi-resolution coded point cloud deep learning network(MRE-PointNet)and autoencoder model based on the Scirpus sibiricum leaf shape parameter estimation algorithm was proposed.The Kinect V2 camera was used to acquire the point cloud data of Scirpus sibiricum leaves in vertical attitude,and the data was pre-processed by straight-pass filtering,segmentation and point cloud simplification algorithm.The geometric model constructed with different parameter combinations was discretized into point cloud data and input into MRE-PointNet network to obtain the pre-training model of the geometric model shape parameter estimation.In order to solve the problem of partial occlusion and noise of the leaves in the filming process,an autoencoder network with secondary processing of the point cloud data was used to obtain the autoencoder pre-training model by taking the discrete point cloud data of the geometric model as input and encoding-decoding operation,which improved the robustness of the MRE-PointNet network in estimating the shape parameters of the occluded data.A total of 300 point clouds of Scirpus sibiricum leaves were collected.With the ratio of 2∶1,totally 200 slices of point cloud data were used as the training set to fine-tune for model transfer to the pre-training model MRE-PointNet,and the remaining 100 slices of point cloud data were used as the test set.By the algorithm,the mathematical statistics and linear regression analysis were performed to compare the estimated and real values of the shape parameters.The experiment results showed that the estimated R^2 and RMSE of leaf length were 0.9005 and 0.4170 cm,leaf width was 0.9131 and 0.3164 cm,and leaf area was 0.9447 and 3.8834 cm^2,respectively,based on the MRE-PointNet and the self-training model.The encoder model algorithm for estimating the shape parameters of scirpus sibiricum leaves had high precision and practicality.