基于噪声样本渐近修正的中医舌色分类方法  

A TCM Tongue Color Classification Method via Progressively Correcting Noisy Samples

作　　者：孙亮亮 李艳萍 张辉 卓力[1,2] SUN Liang-iang;LI Yan-ping;ZHANG Hui;ZHUO Li(Faculty of Information,Beijing University of Technology,Beijing 100124,China;Beijing Key Laboratory of Computational Intelligence and Intelligent Systems,Beijing University of Technology,Beijing 100124,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重点实验室,北京100124

出　　处：《电子学报》2024年第5期1450-1459,共10页Acta Electronica Sinica

基　　金：国家自然科学基金(No.61871006);国家中医药管理局中医药创新团队及人才支持计划项目(No.ZYYCXTD-C-202210)。

摘　　要：基于深度学习的中医舌色分类模型具备良好的性能,但是依赖大量正确标注的样本.由于人工标注样本费时费力,不可避免地存在错误标注,导致模型在训练过程中对噪声样本过拟合,使其泛化能力变差.为此,本文提出了一种基于噪声样本渐近修正的中医舌色分类方法.首先,根据舌色分类的特点,提出了一种全局-局部特征融合方法,将其嵌入到ResNet18骨干网络中,构建了舌色分类网络,并采用集成学习范式,提高分类模型的可靠性和稳定性;其次,针对噪声样本下的舌色分类网络训练问题,提出了样本注意力机制和噪声样本标签重新标注机制,在训练过程中对干净样本和噪声样本加以区分,赋予不同的权重,并逐步对噪声样本标签进行修正;最后,采用Boostrapping损失函数降低模型对噪声样本的关注度,抑制噪声样本对分类性能的影响.将提出的方法在两个自建的舌色分类数据集上进行了实验验证,结果表明,该方法通过渐进地对噪声标签进行校正,可以获得比现有的有噪样本下图像分类方法更高的分类精度,Acc指标分别达到了94.6%和93.65%.Auto tongue color classification is an important research topic in the study of TCM(Traditional Chinese Medicine)objectification.Affected by various factors such as doctor’s experience and illumination conditions,there often exist errors in the manually annotated labels,that is,noisy labels.Noisy labels will cause the model not to converge in the training process and the generalization ability will be poor.Therefore,in this paper,a TCM tongue color classification method is proposed by progressively correcting noisy samples.First,according to the characteristics of the tongue color classification,a global-local feature fusion method is proposed,which is embedded in the ResNet18 backbone network,constructing a tongue color classification network.The ensemble learning paradigm is adopted to improve the reliability and stability of the classification model.Next,for the classification network training problem under noisy samples,a sample attention mechanism and a re-labeling mechanism are proposed.During the training process,different weights are assigned to clean samples and noisy samples,and the noisy samples are gradually adjusted.Finally,the network model is optimized and trained with the Boostrapping loss function to suppress the impact of noisy samples on the classification performance.The experimental results on two tongue color classification datasets SIPL-A and SIPL-B show that,the proposed method can effectively correct noisy labels,thereby,significantly improving the tongue color classification accuracy.Compared with the existing image classification methods under noisy samples,the proposed method can achieve a higher classification accuracy,reaching 94.6%and 93.65%,respectively.

关 键 词：中医舌色分类 噪声样本 样本注意力机制 重新标注机制 Boostrapping损失 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]