基于迁移学习的中药饮片图像分类

Image Classification of Traditional Chinese Medicine Pieces Based on Transfer Learning

针对人工识别中药饮片效率不高、分类具有一定主观性等问题,提出基于迁移学习的中药饮片图像分类方法。自建中药饮片图像数据集,选取MobileNet-V2模型进行训练,将在ImageNet数据集训练得到的参数加载到模型中,修改全连接层节点数和SoftMax分类器得到新模型,冻结除全连接层的参数,训练后的新模型准确率达到97.67%。实验表明,研究的算法可准确对中药饮片图像进行分类,而且比传统CNN有更少的参数量和计算量,为研究人员对中药饮片种类的识别提供有利的依据,可有效辅助研究人员对中药饮片进行识别。

Aiming at the problems of low efficiency and subjectivity in manual recognition of Chinese herbal pieces,this paper proposes an image classification method of Chinese herbal pieces based on transfer learning.Selfbuilt image dataset of Chinese herbal medicine slices,selected MobileNet-V2 model for training,loaded the parameters obtained from the training of ImageNet dataset into the model,modified the number of full connection layer nodes and SoftMax classifier to obtain a new model,frozen the parameters of the full connection layer,and the accuracy of the new model after training reached 97.67%.The experiments show that the algorithm can accurately classify the images of Chinese herbal pieces,and has less parameters and computation than traditional CNN.It can provide favorable basis for researchers to identify the types of Chinese herbal pieces,and can effectively assist researchers to identify Chinese herbal pieces.