机器视觉结合深度学习对荔枝估产的算法研究  

Research on Litchi Yield Estimation Algorithm based on Machine Vision and Depth Learning

作　　者：高翔 陈福展 董力中 卢嘉威 李媛媛 凡超[4] 陈万云 GAO Xiang;CHEN Fuzhan;DONG Lizhong;LU Jiawei;LI Yuanyuan;FAN Chao;CHEN Wanyun(Guangdong Institute of Modern Agricultural Equipment,Guangzhou 510630,China;Shenzhen Institute of Modern Agricultural Equipment,Shenzhen 518022,China;Guangzhou Joinken Network Technology Development Co.,Ltd.,Guangzhou 510630,China;Institute of Fruit Tree Research,Guangdong Academy of Agricultural Sciences,Guangzhou 510645,China)

机构地区：[1]广东省现代农业装备研究所,广东广州510630 [2]深圳市现代农业装备研究院,广东深圳518022 [3]广州市健坤网络科技发展有限公司,广东广州510630 [4]广东省农业科学院果树研究所,广东广州510645

出　　处：《现代农业装备》2024年第1期53-58,70,共7页Modern Agricultural Equipment

基　　金：广东省重点领域研发计划(2023B0202090001);广东省乡村振兴战略专项农业农村重点试点示范及基地建设资金项目(粤财农〔2021〕37号)。

摘　　要：传统的荔枝估产算法为人工计数法,存在估产人力工作量大、效率低、误差高等问题。本文利用计算机视觉和深度学习等人工智能技术,首先搭建了基于YOLOv3算法的荔枝目标检测神经网络模型,预测图片中的果实数量并将其结果保存;然后搭建了MLP神经网络模型,将检测的果实数量作为网络输入,推理输出整棵树的产量估值;此外,为了进一步提高结果的准确性,本文通过双目视觉利用视差获得测量位置和荔枝树的距离深度信息,使用边缘检测得到树的外轮廓,得到荔枝树的树宽和树高,将其结果作为变量输入到MLP神经网络中;最后将识别模型部署到边缘端,通过边缘计算采集图像、进行图像推理与识别计算,最终实现荔枝估产,以改变传统人工荔枝估产方式,提升效率,为后续荔枝果园的精准管理打下基础。The traditional method of Litchi yield estimation is manual counting method,which has many problems such as heavy labor,low efficiency and high error.In this paper,a neural network model for Litchi object detection based on YOLOv3 algorithm is built by using artificial intelligence techniques such as computer vision and depth learning,Then predict the number of fruits in the picture and save the results,a MLP neural network model was built,and the number of detected fruits was taken as the input of the network,and the output of the whole tree was inferred.In addition,in order to improve the accuracy of the results,we use binocular vision to obtain the information of measuring position and distance depth of Litchi tree,and use edge detection to get the outline of the tree,finally,the width and height of Litchi tree are obtained,and the results are input into MLP neural network as variables,through edge calculation,image reasoning and recognition calculation,Litchi yield estimation is realized,which can change the traditional artificial litchi yield estimation method,improve the efficiency and lay a foundation for the accurate management of Litchi orchards.

关 键 词：荔枝估产 边缘检测 卷积神经网络 迁移学习 边缘计算 

分 类 号：S126[农业科学—农业基础科学]