基于卷积神经网络的多肉植物细粒度图像分类

Fine-grained image classification of succulents with convolutions

多肉植物分类是植物栽培管理中的一项重要任务,通常需使用大型数据集和领域独有的特性.由于没有现成的多肉植物数据集,需收集大量的图片自制数据集.研究了多肉植物的细粒度图像分类.为了识别不同视角、背景、光效和成长阶段的多肉植物,对卷积神经网络AlexNet和GoogLeNet的最后三层进行微调,对原创数据集进行了强监督分类和弱监督分类的测试、训练.实验结果表明,微调GoogLeNet的强监督分类达到了最佳效果,精准率为96.7%.

Succulents recognition is an important task in plants identification and management,and domain-specific features and large datasets are used.Due to the absence of an existing succulents’dataset,a large number of pictures need to be collected.This paper addresses the issue of fine-grained image classification of succulent plants.Both supervised and weakly supervised fine-tuning of AlexNet and GoogLeNet training on original dataset were imple meited,trained and tested for the task of identify ing the class of succulents from various viewpoints,backgrounds,light effects and growth stages.Results showed that supervised fine-tuning GoogLeNet improved the performance,and the accuracy rate could be as high as 96.7%.