基于深度学习的自然环境下花朵识别
2019-07-05郑玉龙赵明
郑玉龙 赵明
摘 要:基于自然环境下的花朵识别已经成为了现在园艺植物以及计算机视觉方面的交叉研究热点。本文的花朵图像数据集是利用手机直接在自然场景中当场拍摄的,采集了湖南省植物园内26种观赏花朵的2600幅图像,其中还包括同一品种不同类别相似度很高的杜鹃,郁金香等花朵。设计了一种由3个残差块组成的20层深度学习模型Resnet20,模型的优化算法结合了Adam的高效初始化以及Sgd优秀的泛化能力,该优化算法主要是根据每次训练批次以及learning rate来进行转换调整,实验结果表明比单独使用Adam算法正确率高4到5个百分点,比单独使用Sgd算法收敛更快。该模型在Flower26数据集上,通过数据增强识别率可达到 96.29%,表明深度学习是一种很有前途的应用于花朵识别的智能技术。
关键词:深度卷积神经网络;残差网络;花朵识别;随机梯度下降
中图分类号:TP391;S605 文献标识码:A
Deep Learning for Flower Identification in Natural Environment
ZHENG Yu-long?覮,ZHAO Ming
(College of Software Engineering,Central South University,Changsha,Hunan 410075,China)
Abstract:Rapid identification of flower plants has become a hot topic in the cross-study of horticultural plants and computer vision. The flower image dataset of this article was photographed directly on the spot using a mobile phone in a natural scene,and 2600 images of 26 ornamental flowers in Changsha Botanical Garden were collected,including azaleas and tulips with similar similarities in different categories. A 20-layer deep learning model Resnet20,consisting of three residual blocks is designed. The model optimization algorithm combines Adam's efficient initialization and Sgd's excellent generalization ability. The optimization algorithm is mainly based on each training batch and learning rate to adjust the conversion,the experimental results show that the correct rate is 4 to 5 percentage points higher than the Adam algorithm alone,and it converges faster than using the Sgd algorithm alone. This model is based on the Flower26 dataset,through the data enhanced recognition rate can reach 96.29%,which shows that deep learning is a promising intelligent technology for flower recognition.
Key words:deep convolutional neural network;residual network(Resnet);flower recognition;stochastic gradient descent(SGD)
花朵識别是园艺学科以及农林等行业生产经营非常重要的工作,像花朵识别这样类似的植物分类学是一项对日常生活很有意义的研究课题,这些植物分类学的分类依据主要是依据植物的外形特征,包括花朵、叶子、枝条、树皮、果实等特征。因此,花朵识别也可以算是植物分类学的重要的一部分。由于生活条件的改善,人们开始越来越关注周边的环境,对身边花朵的种类也越来越有兴趣,因此花朵识别的需求也日益增多。但花朵种类繁多,同一类花朵也因生长时间、生长环境还有各种基因变异的影响,导致花朵的形态各有不同,甚至差异很大,因此花朵识别一直都是人工智能识别领域中很有挑战性的一项研究。……
