摘要
目的在基于舌图像的中医体质类型分类中,舌图像的类间距小,传统手工特征提取时存在底层特征不能够充分表达舌图像所包含的全部信息等问题。因此本文提出一种基于深度网络特征层融合的体质类型分类方法,以提高体质类型分类的准确率。方法通过比较不同网络模型对舌图像的分类表现,及对不同网络层的特征表达能力的分析,选取将浅层特征与高层语义特征进行融合的方法。该深度特征融合方法基于Alexnet网络进行改进,依据误差权重对各层特征进行融合,并采用983张舌图像,对气虚质、痰湿质和湿热质三种体质类型的分类进行仿真实验。结果相比传统特征提取与原始深度网络,本文方法的准确率由传统分类方法的54.3%提高到了77%。结论基于深度特征融合的方法将浅层特征与深度特征融合,充分表达了图像的语义信息,对中医辅助辨识、临床、教学和科研具有极其重要的意义。
Objective The traditional manual feature extraction is mainly based on the bottom features,which can not adequately express all the information contained in the tongue image.In order to improve the accuracy of physique classification,a method of physique classification based on feature layer fusion of deep network is proposed in this paper.Methods Different network models have different classification performance for tongue images,and different network layers have different feature description capabilities.This method fuses shallow features with high-level semantic features,avoides the interference of human factors in traditional special extraction.The method of depth feature fusion is improved on Alexnet network.According to the error weight,the features of each layer are fused.Based on 983 tongue image data sets,this paper discusses the classification of three constitutional types:Qi deficiency,phlegm-dampness and damp-heat.Results The simulation results show that compared with the traditional feature extraction and the original depth network,the accuracy of this method is improved from 54.3%to 77%.Conclusions The method based on depth feature fusion is proposed to fuse shallow features with depth features,which fully expresses the semantic information of images.It is of great significance to the research of assistant identification,clinical,teaching and scientific research of traditional Chinese medicine.
作者
周浩
胡广芹
张新峰
ZHOU Hao;HU Guangqin;ZHANG Xinfeng(Faculty of Information Technology,Beijing University of Technology,Beijing 100124)
出处
《北京生物医学工程》
2020年第3期221-226,共6页
Beijing Biomedical Engineering
基金
国家中医药管理局资助项目(GZY-FJS-2018-023)
国家重点研发计划中医药现代化研究专项项目(2017YFC1703300)资助。
关键词
卷积神经网络
特征融合
舌图像
体质类型分类
中医
convolutional neural network
feature fusion
tongue image
physique classification
traditional Chinese medicine
