基于深度学习的玉米子粒品种分类与胚面识别  

Classification of Maize Grain Varieties and Embryo SurfaceRecognition Based on Deep Learning

作　　者：李萍 杨昊岩 栾涛 LI Ping;YANG Hao-yan;LUAN Tao(Agricultural and Rural Service Center,Shangma Street,Chengyang District,Qingdao 266102;College of Animation and Communication,Qingdao Agricultural University,Qingdao 266109,China)

机构地区：[1]山东省青岛市城阳区上马街道办事处,山东青岛266102 [2]青岛市青岛农业大学动漫与传媒学院,山东青岛266109

出　　处：《玉米科学》2025年第2期61-68,共8页Journal of Maize Sciences

基　　金：青岛农业大学横向课题(横20210131)。

摘　　要：对玉米子粒进行品种分类并准确识别其胚面与非胚面,不仅对理解品种间产量和品质的差异至关重要,也是实现定向精确播种的基础前提。传统的玉米子粒品种分类与胚面识别方法需要提取大量特征,存在主观性强、泛化能力不足等缺陷。基于深度学习技术,运用经过预训练的卷积神经网络进行分类识别。将扫描仪获取的7个品种玉米子粒的胚面图像与非胚面图像进行图像分割预处理,采用3种分类策略构建分类数据集并导入卷积神经网络进行分类。选择分类准确率最高、训练时间最短、工作流程最简化策略作为最优策略,最终获得了96%的分类准确率。结果表明,提出的分类策略能够精准地对玉米子粒进行品种分类与胚面识别。Classifying maize kernels and accurately identifying their embryo and non-embryo surfaces are not only crucial for understanding the differences in yield and quality among varieties but also serve as a fundamental prerequisite for achieving targeted and precise sowing.Traditional methods for maize kernel variety classification and embryo surface identification require the extraction of numerous features,which have inherent flaws such as high subjectivity and insufficient generalization capabilities.This paper employs deep learning technology,utilizing pre-trained models for classification and identification.Initially,images of the embryo and non-embryo surfaces of maize kernels from seven varieties,obtained via a scanner,are preprocessed through image segmentation.Then,three classification strategies are used to construct maize variety classification datasets and perform classification.The strategy with the highest classification accuracy,shortest training time,and most streamlined workflow was se⁃lected as the optimal approach,ultimately achieving a classification accuracy of 96%.Experimental results demon⁃strate that the classification strategy proposed can precisely classify maize kernel varieties and identify embryo sur⁃faces.

关 键 词：玉米 品种分类 胚面识别 深度学习 卷积神经网络 

分 类 号：S513.01[农业科学—作物学]