首页|基于注意力机制的轻量化VGG玉米籽粒图像识别模型

基于注意力机制的轻量化VGG玉米籽粒图像识别模型

扫码查看
玉米是重要的生产资料,为实现对玉米种子的识别与保护,实验采集了 5个玉米品种,经处理后共获得1 778张玉米籽粒图像,建立胚面与胚乳面混合的数据集.按7∶2∶1的比例划分训练集、验证集和测试集.首先基于迁移学习选取DenseNet121、MobileNetV2、VGG16和GoogLeNet对玉米籽粒图像进行识别,在测试集上的准确率分别是94.32%、93.18%、95.45%和92.61%,由于在VGG16上的准确率最高,所以选择对VGG16进行改进,在对模型进行轻量化处理的同时引入通道注意力SE模块,构建一个新的网络模型L-SE-VGG,并与未预训练的VGG16、迁移学习的VGG16和不加SE模块的L-VGG进行对比,最终在L-SE-VGG上的识别准确率高达98.86%.研究为深度学习技术在玉米籽粒品种识别中的应用提供了新的有效策略和实验方法,为玉米籽粒品种的识别和检测提供了参考.
Image Recognition Model of Light VGG Maize Kernel Based on Attention Mechanism
Corn is an important means of production.In order to recognize and protect corn seeds,5 corn varie-ties were collected in this experiment.After processing,a total of 1 778 corn grain images were obtained,and a mixed dataset of embryo surface and endosperm surface was established.The training set,verification set,and test set were divided in a ratio of 7:2:1.Firstly,based on transfer learning,DenseNet121,MobileNetV2,VGG16 and GoogLeNet were selected to recognize the corn kernel image,and the accuracy rates in the test set were 94.32%,93.18%,95.45%and 92.61%,respectively.Since VGG16 had the highest accuracy,Therefore,VGG16 was im-proved,and channel attention SE module was introduced to construct a new network model L-SE-VGG while sim-plifying the model structure.Compared with VGG16 without pre-training,VGG16 with transfer learning and L-VGG without SE module,the recognition accuracy of L-SE-VGG was up to 98.86%.The study provides a new and effective strategy and experimental method for the application of deep learning technology in the identification of corn varieties and provides reference for the identification and detection of corn varieties.

VGG16SE blockimage recognitiondeep learningcorn kernel

孙孟研、王佳、马睿、代东南、刘起、穆春华、马德新

展开 >

青岛农业大学,青岛 266109

山东省农业科学院玉米研究所,济南 250100

VGG16 SE模块 图像识别 深度学习 玉米籽粒

山东省自然科学基金项目山东省高等学校青创人才引育计划项目

ZR2022MC152202202027

2024

中国粮油学报
中国粮油学会

中国粮油学报

CSTPCD北大核心
影响因子:1.056
ISSN:1003-0174
年,卷(期):2024.39(1)
  • 21