摘要
玉米是我国重要的谷类作物,玉米种子的纯度是影响种子质量的关键指标,不同品种玉米的种子活力、物理指标和发芽率都是不同的,因此需要对玉米品种的真伪进行鉴别分选。传统模式识别的方法需要人工定义各类特征,存在主观判断、费时费力等问题,实用性较差。针对上述问题,本研究拟建立一种基于RGB图像结合深度学习的低成本、高效、无损的单粒玉米种子真伪检测方法,选用不同产地登海605玉米种子440粒,其他品种480粒,采集玉米种子胚面和胚乳面制作数据集,通过图像处理技术对图像进行预处理,并按照7∶2∶1的比例将数据集分为训练集、验证集和测试集,分别使用GoogLeNet、MobileNet、Inception-ResNet、ResNet、DenseNet共5种网络模型利用迁移学习对3类数据集进行分类测试,结果表明,5种网络模型在双面数据集的平均识别准确率最高,测试识别准确率为99.05%,ResNet网络在3类数据集中的分类效果最佳,在双面测试集上为99.91%。本研究提供了一种无损、高效、相对可靠的方法来鉴别登海605玉米品种的真伪。
Corn is an important cereal crop in China,and the purity of corn seed is the key index affecting seed quality.The seed vigor,physical index and germination rate of different varieties of corn are different,therefore,it is necessary to identify and sort the authenticity of corn varieties.The traditional pattern recognition method requires manual definition of various features,has subjective judgment,time-consuming and laborious problems,and is not practical.For the above problem,this study intended to establish a kind of based on RGB image combined with deep learning of low cost,high efficiency and non-destructive authenticity test method of single grain corn seeds.In this research,440 Denghai 605 corn seeds of different places of origin and 480 corn seeds of other varieties were selected.Corn seed embryo and endosperm surface production data sets were selected.After that,the images were preprocessed by image processing technologies,and the data sets were divided into training set,validation set and test set in accordance with a rate of 7∶2∶1.Five network models of GoogLeNet,MobileNet,Inception-ResNet,ResNet and DenseNet were used to classify and test the three types of data sets by transfer learning.The results indicated that,the average recognition accuracy of the five network models was the highest in the two-sided data set,and the test recognition accuracy was 99.05%.The ResNet network had the best classification efficiency for the three types of data sets,being 99.91%for two-sided test sets.This study provided a nondestructive,efficient and relatively reliable method to identify the authenticity of Denghai 605 corn variety.
作者
王佳
马睿
马德新
Wang Jia;Ma Rui;Ma Dexin(Qingdao Agricultural University,Qingdao 266109)
出处
《中国粮油学报》
CAS
CSCD
北大核心
2023年第3期151-157,共7页
Journal of the Chinese Cereals and Oils Association
基金
山东省重点研发计划项目(2019GNC106001)
青岛市民生科技计划项目(18-6-1-112-nsh)
淄博市重点研发计划项目(2019gy010101)
山东省高等学校青创人才引育计划项目(202202027)。
关键词
玉米
深度学习
机器视觉
品种鉴别
corn
deep learning
machine vision
variety identification