摘要
[目的]为快速、准确、无损检测苹果的外部表型参数,提出了一种基于Pointnet和迁移学习的苹果表型参数估算算法。[方法]通过Kinect相机从任意角度拍摄苹果并使用直通滤波法去除背景环境数据得到只包含苹果信息的点云数据。在此基础上使用最远点采样法,获取标准输入点云,然后采用椭球曲面方程构建苹果几何模型,生成基于椭圆方程的苹果几何模型库。使用Pointnet算法训练仿真模型数据,然后通过迁移学习迁移到实测数据上去,在训练好的模型上进行微调;再经过5-折交叉验证,判定模型的鲁棒性和泛化能力,得到最终的估算模型。[结果]以均方根误差(RMSE)和决定系数(R2)评价模型结果,实测250个苹果3个角度点云共750组数据,在任意一个角度拍摄的残缺率达到50%的点云数据的条件下,该模型对苹果的直径、高度、体积3组表型参数的RMSE分别为2.247、2.275和22.780,R2分别为0.919、0.841和0.927。[结论]该算法回归效果优于传统算法,在任意角度拍摄到的残缺率达到50%的点云数据的条件下仍能很好完成外部表型参数估算。
[Objectives]In order to detect the external phenotype parameters of apple quickly,accurately and non-destructively,an algorithm for estimating apple phenotypic parameters based on Pointnet and transfer learning was proposed.[Methods]By using Kinect camera to shoot apples from any angle,the point cloud data containing only apple information was obtained by removing background environment data by direct filtering method.On this basis,the farthest point sampling method was used to obtain the standard input point cloud,then the geometric model of apple was constructed by using ellipsoid surface equation,and the apple geometric model library based on elliptic equation was generated.Using Pointnet algorithm to train the simulation model data,and transfer the measured data through migration learning,after fine tuning the trained model,the robustness and generalization ability of the model were determined by 5-fold cross validation,and the final estimation model was obtained.[Results]Evaluation of model results was conducted by and root mean square error(RMSE)and coefficient of determination(R2).A total of 750 sets of data were collected from 250 apple point clouds at three angles.The final results showed that based on the point cloud data with 50%incomplete degree taken from any angle,the model could predict the three groups of phenotypic parameters of apple such as diameter,height and volume.The root mean square error(RMSE)was 2.247,2.275 and 22.780 respectively,and coefficient of determination(R2)was 0.919,0.841 and 0.927 respectively.[Conclusions]This algorithm not only has better regression effect than the traditional algorithm,but also can complete the estimation of external phenotypic parameters under the condition that the incomplete rate of point cloud data is about 50%from any angle.
作者
陈龙
王浩云
季呈明
孙云晓
徐焕良
CHEN Long;WANG Haoyun;JI Chengming;SUN Yunxiao;XU Huanliang(College of Artificial Intelligence,Nanjing Agricultural University,Nanjing 210095,China)
出处
《南京农业大学学报》
CAS
CSCD
北大核心
2021年第6期1209-1216,共8页
Journal of Nanjing Agricultural University
基金
国家自然科学基金项目(31601545)
中央高校基本科研业务费专项资金(KYLH202006,KYZ201914)。
