基于持续学习的中医舌色苔色协同分类方法  

Collaborative Classification Method of TCM Tongue Color and Coating Color Based on Continual Learning

作　　者：卓力 李艳萍[1,2] 孙亮亮 张辉 李晓光 张菁[1,2] 杨洋 魏玮[3] ZHUO Li;LI Yanping;SUN Liangliang;ZHANG Hui;LI Xiaoguang;ZHANG Jing;YANG Yang;WEI Wei(Faculty of Information Technology,Beijing University of Technology,Beijing 100124,China;Beijing Key Laboratory of Computational Intelligence and Intelligent System,Beijing University of Technology,Beijing 100124,China;Beijing Key Laboratory of Traditional Chinese Medicine Treatment of Functional Gastrointestinal Diseases,Wangjing Hospital of Chinese Academy of Medical Sciences,Beijing 100102,China)

机构地区：[1]北京工业大学信息学部,北京100124 [2]北京工业大学计算智能与智能系统北京市重点实验室,北京100124 [3]中国中医科学院望京医院功能性胃肠病中医诊治北京市重点实验室,北京100102

出　　处：《北京工业大学学报》2024年第9期1077-1088,共12页Journal of Beijing University of Technology

基　　金：国家自然科学基金资助项目(61871006);国家中医药管理局中医药创新团队及人才支持计划资助项目(ZYYCXTD-C-202210)。

摘　　要：中医(traditional Chinese medicine, TCM)舌诊客观化研究中需要分析的舌象特征很多,不同的舌象特征往往采用单独的方法进行分析,导致分析系统的整体实现复杂度大幅增加。为此,基于持续学习的思想,提出一种中医舌色苔色协同分类方法,该方法将舌色分类作为旧任务,将苔色分类作为新任务,充分利用2个任务的相似性和相关性,仅通过一个网络结构就同时实现舌色和苔色的准确分类。首先,设计一种基于全局-局部混合注意力机制(global local hybrid attention, GLHA)的双分支网络结构,将网络高层语义特征与低层特征相融合,提升特征的表达能力;然后,提出基于正则化和回放相结合的持续学习策略,使得该网络在学习新任务知识的同时有效防止对旧任务知识的遗忘。在2个自建的中医舌象特征分析数据集上的实验结果表明,提出的协同分类方法可以获得与单个任务相当的分类性能,同时可以将2个分类任务的整体复杂度降低一半左右。其中,舌色分类准确率分别达到93.92%和92.97%,精确率分别达到93.69%和92.87%,召回率分别达到93.96%和93.16%;苔色分类准确率分别达到90.17%和90.26%,精确率分别达到90.05%和90.17%,召回率分别达到90.24%和90.29%。There are many characteristics of tongue that need to be analyzed in traditional Chinese medicine(TCM).Different characteristics are often analyzed by individual methods,which significantly increases the overall implementation complexity of the analysis system.Therefore,this paper proposes a collaborative classification method of tongue color and coating color in TCM based on continual learning.This method takes tongue color classification as an old task and coating color classification as a new task,which makes full use of the similarity and relevance of the two tasks to realize the accurate classification of tongue color and coating color simultaneously under a single network framework.First,a dual branch network structure with global local hybrid attention(GLHA)mechanism was designed,which aggregates high-level semantic features with low-level features to improve the representative capability of features.Second,a continual learning strategy based on the combination of regularization and rehearsal was proposed,which made the network effectively prevent forgetting the knowledge learned from old task while learning new task.The experimental results on two self-established TCM tongue datasets show that,the proposed collaborative classification method can achieve a comparable classification performance with a single task,and simultaneously,reduce the overall complexity of the two classification tasks by almost half.Among them,the accuracy of tongue color classification reaches 93.92%and 92.97%,the precision reaches 93.69%and 92.87%,the recall reaches 93.96%and 93.16%,respectively.While that of the coating color classification reaches 90.17%and 90.26%,the precision reaches 90.05%and 90.17%,the recall reaches 90.24%and 90.29%,respectively.

关 键 词：中医舌色苔色分类 协同分类 深度学习 持续学习 全局-局部混合注意力机制 机器视觉 

分 类 号：TN911.71[电子电信—通信与信息系统] TP391[电子电信—信息与通信工程]