基于注意力机制和迁移学习的花类品种鉴别研究

Identification of Flower Species Based on Attentional Mechanism and Transfer Learning

多品种花类的识别在生活、贸易中具有广阔的应用前景,为解决大众人群准确分辨花类品种较为困难的问题,本文提出一种基于ResNet34网络的花类识别方法,在对数据样本进行泛化处理后,在网络残差结构中加入注意力机制,并对相关参数进行微调整,再利用迁移学习训练网络模型。实验表明,本文方法不仅有效的加快了模型训练的收敛速度,同时准确率比ResNet34原网络提升了5.2个百分点,达到了95.3%,并且相比于同类型较为成熟的分类网络,如AlexNet、VGG16、GoogLeNet,本文方法识别准确率有着显著提升。

Many types of flower type recognition have broad applications in life and trade.In order to solve the mass population accurately distinguish flower varieties comparatively difficult problems,this dissertation puts forward a type of based on ResNet34 network class identification method,after to generalize the data sample,join in the network structure of residual attention mechanism,and the related parameters are micro adjustment,then transfer learning and training network model is used.Experiments demonstrate that the proposed method not only accelerates the convergence speed of model training effectively,but also improves the accuracy by 5.2 percentage points compared with ResNet34 original network,reaching 95.3%.Compared with the mature classification networks of the same type,such as AlexNet,VGG16,GoogLeNet,the recognition accuracy of the proposed method is markedly enhanced.