DOI：10.3969/j.issn.1672-3872.2024.01.037
基金项目：2022 年度广西高校中青年教师基础能力提升项目“迁移学习卷积神经网络在花卉识别分类上的研究与应用”（2022KY1078）；2022年度广西高校中青年教师基础能力提升项目“基于灰度模板匹配小型零件计数系统的研究与开发”（2022KY1072）
作者：陈卫国，莫胜撼（广西机电职业技术学院电气工程学院，广西 南宁 530000）
摘 要：【目的】不同种花卉之间的相似性以及同种花卉内部的多变性加大了花卉图像分类难度，其难点是要人工设计出能充分体现花卉颜色、形状和花瓣形态等特征的特征提取方法。传统的花卉图像分类方法的精度不高且模型的泛化能力较差，这些问题亟待解决。【方法】课题组提出一种基于数据增强的VGG16迁移学习卷积神经网络提取花卉图像特征，再训练多类逻辑回归模型的花卉图像分类识别方法；并且通过在flowers17和flowers102花卉数据集上进行测试，来验证课题组所提出的花卉分类识别方法的有效性。【结果】课题组所提出的花卉分类识别方法在flowers17和flowers102数据集中分别达到了97.89%和92.10%的分类精度，高于现有其他花卉图像分类方法。【结论】通过预训练的深度人工神经网络提取的高区分度的花卉图像特征，优于人工设定的花卉图像特征，能训练出更高效精准的花卉识别分类器。基于本研究内容，下一步可对VGG16网络 进行降维改进，让模型参数减少，从而实现快速实时应用。
关键词：花卉图像分类；卷积神经网络；迁移学习；VGG16；逻辑回归模型
Flower Classification Based on Transfer Learning and Logical Regression Model
Chen Weiguo, Mo Shenghan (School of Electrical Engineering, Guangxi Technology College of Machinery and Electricity, Guangxi Nanning 530000)
Abstract: [Objective] The similarity between different types of flowers and the variability within the same type of flower increase the difficulty of flower image classification. The difficulty is to manually design various feature extraction methods that can fully reflect the features of flower color, shape and petal pattern. The accuracy of traditional flower image classification methods is not high and the generalization ability of the model is poor, which urgently needs to be solved. [Method] The research group proposes a flower image classification and recognition method based on data augmentation with VGG16 transfer learning convolutional neural networks to extract flower image features, and then trains multiple-classes logistic regression model; and verifies the effectiveness of the flower classification and recognition method proposed by the research group by testing it on the flowers17 and flowers102 datasets. [Result] The flower classification and recognition method proposed by the research group achieved classification accuracy of 97.89% and 92.10% in the flowers17 and flowers102 datasets, respectively, which is higher than other existing flower image classification methods. [Conclusion] The highly discriminative flower image features extracted through pre-trained deep artificial neural networks are superior to the manually set flower image features, which can train more efficient and accurate flower recognition classifiers. Based on the content of this study, the next step is to perform dimensionality reduction and improvement on the VGG16 network, reducing model parameters and achieving fast and real-time applications.
Keywords: flower image classification; convolutional neural networks; transfer learning; VGG16; logistic regression model
引文信息：[1]陈卫国，莫胜撼.基于迁移学习和逻辑回归模型的花卉分类研究[J].南方农机，2024，55(1)：139-143+151.
查看全文请下载PDF文件↓
