Abstract: Traditional flower species recognition algorithms are mainly based on designing hand-craft feature and training classifier, which generalization ability is limited and the recognition accuracy always reach a bottleneck. Therefore, this paper proposes a recognition method that is based on 152 residual layers deep convolutioal neural network. Specifically, model transfer learning is used to refine the network via large scale labeled flower database that are acquired from internet worm system. In the online system which has embedded the proposed recognition algorithm, user can send flower images to the cloud via internet. Then the recognition algorithm is performed through deep architecture that deployed on the server. Experimental results conducted on ImageNet and NetDragon datasets show that the fine tuned residual network has the advantages of high accuracy, real time feedback, and better robustness.

Key words: deep learning, flower species recognition, residual network

摘要： 传统的花卉识别算法一般是建立在手动特征提取和分类器训练的基础上，其泛化能力有限且准确度存在瓶颈。为此提出了基于深度卷积网络的识别算法，采用152层残差网络架构，在爬虫获取的大量标定数据基础上，对神经网络进行迁移学习训练。上线发布的算法集成系统中，用户拍照获取的花卉照片可通过网络传输到云服务器，并在服务端部署的深度学习架构下实现花卉快速识别。针对ImageNet和网龙花卉数据集的实验对比结果表明，基于残差网络迁移学习的方法具有识别准确率高、实时反馈、鲁棒性好等特点。

关键词: 深度学习, 花卉识别, 残差网络