许明广 李玉溪 刘润宇 刘星宇 郑学谦








摘要:正确的实现花卉分类是对该植物进行深入研究的前提条件。针对开源花卉数据集单一和分类效果不佳的问题,本文整理了一个基于国内五种花卉的图像数据集并使用VGG16网络模型完成图像分类任务。为了更好的表征花卉图像特征信息,本文对VGG16网络模型参数进行了优化,同时,引入一种基于多特征融合的VGG16网络模型,与原始VGG16网络模型进行对比试验。实验结果表明,原始VGG16网络和基于多特征融合的VGG16网络在国内五种花卉图像数据库上均取得了良好的分类效果,并且基于多特征融合的VGG16网络模型要由于原始VGG16网络模型。
关键词:深度学习;花卉分类;VGG16;多特征融合
Abstract: Correct flower classification is a prerequisite for further study of the plant. Aiming at the problems of unitary open source flower data set and poor classification effect, this paper sorted out an image data set based on five kinds of domestic flowers and used VGG16 network model to complete the image classification task. In order to better represent the flower image feature information, the parameters of VGG16 network model were optimized in this paper. At the same time, a VGG16 network model based on multi-feature was introduced to conduct a comparative test with the original VGG16 network model. The experimental results show that both the original VGG16 network and the VGG16 network based on multi-feature have achieved good classification effects on the five domestic flower image databases, and the VGG16 network model based on multi-feature fusion is due to the original VGG16 network model.
Keywords: deep learning;Flower classification;VGG16;Multi-feature fusion
0 引言
植物学研究领域中最基础的一项研究任务就是花卉图像分类[1],不同类别的花卉具有不同的特征,辨别一束花卉的种类往往需要综合颜色、花瓣形状等等多个方面的特征。花卉分类的难点主要在于不同花卉间的特征相似性和同种花卉间的特征差异性。
花卉分类研究在早期基本都是运用基于人工设计的特征和传统机器学习算法相结合的方法,例如基于人工设计的特征,尺度不变性(SIFT)[2]结合支持向量机(SVM)[3]的算法结合构成一个分类网络模型。注意到花卉具有多个特征这一特点,Nilsback M.E和 Zisserman[4]提出了一个新的视觉词包作为分类特征,这个视觉词包包括花卉的形状、纹理以及颜色等特征。同时,Nilsback M.E 和 Zisserman还创建了花卉图像分类领域经典的开源数据集Oxford Flowers系列数据集。……