基于模型迁移的苹果光学特征参数反演  

Inversion of Apple Optical Feature Parameters Based on Model Migration

作　　者：徐焕良[1,2] 周冰清 王浩云[1,2] 李亦白 胡华东 黄芬[1,2] XU Huanliang;ZHOU Bingqing;WANG Haoyun;LI Yibai;HU Huadong;HUANG Fen(College of Information Science and Technology,Nanjing Agricultural University,Nanjing 210095,China;Postdoctoral Station of Agricultural Engineering,Nanjing Agricultural University,Nanjing 210031,China)

机构地区：[1]南京农业大学信息科技学院,南京210095 [2]南京农业大学农业工程博士后流动站,南京210031

出　　处：《农业机械学报》2020年第11期264-271,共8页Transactions of the Chinese Society for Agricultural Machinery

基　　金：中央高校基本科研业务费专项基金项目(KJQN201732、KYZ201914);国家自然科学基金项目(31601545);国家级大学生创新创业训练计划项目(201910307072Z);国家重大专项粮食丰产增效科技创新项目(2016YFD0300607)。

摘　　要：针对现有水果组织光学特征参数反演方法耗时费力、普适性较差的问题,提出了一种基于模型迁移的光学特征参数反演方法。以苹果为例,构造仿真双层生物组织模型;基于蒙特卡洛(Monte Carlo)原理进行光子传输模拟,生成150万光亮度分布图,将光亮度分布图作为数据集输入构造好的8层卷积神经网络(CNN)上进行训练,得到预训练模型;再将训练好的模型迁移到实际测得的含有4000幅苹果高光谱点光源图像的数据集上进行微调,从而完成对光学参数的反演。将本文方法与其他几种算法的反演结果进行分析比较,结果表明,在实测数据集较小的情况下,该方法对苹果光学特征参数的反演结果为果皮吸收系数87.26%、果肉吸收系数90.53%、果皮散射系数86.66%、果肉散射系数87.57%,反演准确率高于其他算法,预训练模型基于大量仿真模型的光亮度分布图经由训练而得到,具有良好的普适性。本研究为解决水果光学特征参数反演中建模数据量不足问题提供了方法参考。The existing methods for inverting the optical characteristic parameter of fruit tissue are time-consuming and labor-intensive,and poor in generality.Aiming at these problems,an optical characteristic parameter inversion method based on model migration was proposed.Taking apple as an example,a simulated double-layer tissue model based on the Monte Carlo method was used to generate 1.5 million light distribution maps.The light distribution map was input as a data set to the 8-layer convolutional neural net(CNN)for training,to obtain a pre-trained model.The trained model finally was transferred to the actual measured data set containing 4000 apple hyperspectral point light source images,and fine-tuned to complete the inversion of optical parameters.The method was compared with the inversion results of several other algorithms.The results showed that when the measured data set was small,the inversion results of this method on apple optical characteristic were the peel absorption coefficientμa1 was 87.26%,the pulp absorption coefficientμa2 was 90.53%,the peel scattering coefficientμs1 was 86.66%,and the pulp scattering coefficientμs2 was 87.57%.The accuracy of the inversion was higher than other inversion methods.The pre-trained model was obtained by training a large number of light distribution maps based on the simulation model.The model was universal,and it can provide a solution to the problem of insufficient data amount in the inversion of fruit optical characteristic parameters.

关 键 词：苹果 光学特征参数 仿真模型 卷积神经网络 模型迁移 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]