摘要
针对传统花卉识别准确率低、泛化性能差、过程耗时费力和花卉样本少等问题,提出一种基于卷积神经网络模型的迁移学习方法。先进行小规模花卉图像数据增强等预处理,再对大规模数据集预训练模型进行迁移学习,修改密集连接分类层。在此基础上进行微调小规模数据集上的卷积基参数,得出识别分类结果。实验表明:在小规模花卉图像集上迁移微调预训练网络准确率可达96.3%。由此证明了深度卷积网络迁移应用到小规模数据集上的可行性。
Aiming at the problems of low accuracy,poor generalization performance,time-consuming and labor-intensive process,and fewer flower samples of traditional flower recognition,this paper proposes a transfer learning method based on convolutional neural network model.It preprocessed the small-scale flower image data enhancement,did transfer learning for the large-scale data set pre-training model,modified the dense connection classification layer,and then fine-tuned the convolution base parameters on the small-scale data set to get the recognition classification results.The experiment shows that on the small-scale flower image set,the transfer fine-tuning pre-training network can get better recognition accuracy,which can reach 96.3%.It also shows the feasibility of deep convolution network transfer applied to the small-scale data sets.
作者
曹晓杰
么娆
严雨灵
Cao Xiaojie;Yao Rao;Yan Yuling(School of Mechanical and Automotive Engineering,Shanghai University of Engineering Science,Shanghai 201620,China;Aeronautic Transport College,Shanghai University of Engineering Science,Shanghai 201620,China)
出处
《计算机应用与软件》
北大核心
2020年第8期142-148,共7页
Computer Applications and Software
基金
国家自然科学基金项目(61303098)。
关键词
花卉图像
卷积神经网络
数据增强
迁移学习
微调
Flower image
Convolutional neural network
Data enhancement
Transfer learning
Fine-tune
