基于轻量级深层卷积神经网络的花卉图像分类系统

Flower Image Classification System Based on Lightweight DCNN

为解决深层卷积神经网络(Deep convolutional neural network,DCNN)模型在算力弱、存储成本高的AI边缘计算设备上难以高效应用的现实问题,本文利用重量级网络辅助训练轻量级网络,设计了一种基于轻量级神经网络的花卉图像分类系统。首先利用重量级DCNN并结合迁移学习、爬虫技术与最大连通区域分割方法,构建了适用于轻量级网络训练的扩充花卉数据集。然后基于Tiny⁃darknet与Darknet⁃reference两种网络及扩充后的花卉数据集训练得到两种面向弱算力设备的轻量级DCNN模型。训练得到的两种花卉分类网络在Oxford102花卉数据集上的平均分类准确率可达98.07%与98.83%,模型大小分别为4 MB与28 MB,在AI边缘计算设备中具有较好的应用前景。

To solve the problem that deep convolutional neural network(DCNN)models with heavy weights are difficult to be effectively applied on AI edge devices with weak computing power and high storage costs,a flower image classification system equipped with a lightweight DCNN is proposed with the help of a heavyweight DCNN during training process.First,an extended flower data set suitable for lightweight DCNN training is constructed by using a heavyweight DCNN combined with transferring learning,the crawler technology and the maximum connected region segmentation method.Then,two lightweight DCNN models,Tiny-Darknet and Darknet-Reference,oriented for devices with weak computer power are trained based on the specially built flower image gallery.Experimental results show that the two optimized models obtained can achieve 98.07% and 98.83% average classification accuracy respectively on Oxford102 flower dataset while keeping the model size as 4 MB and 28 MB,which have promising application potentials for AI edge computer devices.