基于卷积神经网络的水稻钵体软盘穴播量检测  

Detection on rice seeds quantity per hole of the pot body seed tray based on convolutional neural network

作　　者：邓向武 马旭[2] 齐龙[2] 董文浩[2] Deng Xiangwu;Ma Xu;Qi Long;Dong Wenhao(College of Electronic Information Engineering,Gruangdong University of Petrochemical Technology,Maoming,525000,China;College of Engineering,South China Agricultural University,Guangzhou,510642,China)

机构地区：[1]广东石油化工学院电子信息工程学院,广东茂名525000 [2]华南农业大学工程学院,广州市510642

出　　处：《中国农机化学报》2020年第7期130-136,共7页Journal of Chinese Agricultural Mechanization

基　　金：国家重点研发计划课题(2017YFD0700802);现代农业产业技术体系建设专项资金(CARS—01—43);广东石油化工学院人才引进及博士启动项目(2019rc044)。

摘　　要：为避免水稻钵体软盘穴播量检测过程中的秧盘背景分割和稻种特征的手工设计及提取,本文提出了一种基于卷积神经网络的水稻钵体软盘穴播量检测方法,该方法可自动学习和提取不同穴播量的水稻种子特征,实现常规稻、杂交稻和超级杂交稻钵体软盘穴播量为0、1、2、3、4、5、6及7粒以上共8种播量的自动检测。本文在每层卷积单元网络结构参数保持固定的前提下,选取2~4层共3种不同卷积单元数量的网络结构对RiceCountCNN模型性能进行试验,试验结果表明随着模型深度加深,模型检测精度逐渐提高。本文在3层RiceCountCNN模型网络框架下,按卷积核的大小递减和数量递增原则选择得出不同的卷积核网络参数组合方式,最终优化得出网络结构为9C16-AP2-7C32-AP2-5C64的模型性能最佳,平均正确率达到98.76%。为测试RiceCountCNN模型的性能,每个水稻品种选取1幅19穴×14穴的图像作为测试集对模型进行测试,试验结果得出模型针对常规稻、杂交稻和超级杂交稻的检测正确率分别达到97.37%、98.12%和90.98%,每幅图像的检测时间小于2.33 s。研究结果满足精密育秧播种实际工况检测要求,该研究为实现水稻精密衡量播种作业提供参考。To avoid image segmentation and the hand-crafted features of designing and extraction,which used in the course of detection on rice seeds quantity per hole in the pot body seed tray.This paper presents a method to detect rice seeds quantity per hole of the pot body seed tray based on convolutional neural network,which could learn and extract automatically the feature of different rice seeds quantity per hole.The model had realized the detection rice seeds quantity per hole with conventional rice,hybrid rice and super hybrid rice,which including the eight situations of rice seeds in each hole including one to seven particles,and above eight particles.To maintain the structure parameters of each layers in RiceCountCNN model unchanged,experimental investigation with three different layers(two to four)were selected on the performance of different network structure with RiceCountCNN.The result indicated that the detection accuracy could improve gradually with the increasing of model depth.To select the structure parameter of convolutional kernel in each layer with RiceCountCNN model,with the principles of diminishing marginal utility,the principles of reducing the size and increasing the number with convolutional kernel will be followed based on the three-layers structure of RiceCountCNN model.The optimized neural network structure is determined as 9 C16-AP2-7 C32-AP2-5 C64 network,which has higher detection accuracy and stronger robustness.The average accuracy of rice seeds quantity per hole reaches was up to 98.76%.For evaluating the accomplished performance on RiceCountCNN model,each 19 holes times 14 holes image of the pot body seed tray with three rice varieties were selected.The results of this experiment are as follows.Firstly,the detection average accuracy of conventional rice,hybridized rice and super rice hybrid were 97.37%,98.12%and 90.98%,and secondly,each image was processed less than 2.23 seconds.The results show that the method of detection on rice seeds quantity per hole meets the test requirements of

关 键 词：穴播量 卷积神经网络 稻种 钵体软盘 网络深度 卷积核 

分 类 号：S24[农业科学—农业电气化与自动化] TS253.7[农业科学—农业工程]