有噪声标注情况下的中医舌色分类方法

TCM Tongue Color Classification Method under Noisy Labeling

舌色是中医(TCM)望诊最关注的诊察特征之一,自动准确的舌色分类是舌诊客观化研究的重要内容。由于不同类别舌色之间的视觉界限存在模糊性以及医生标注者的主观性等,标注的舌象数据中常含有噪声,影响舌色分类模型的训练。为此,该文提出一种有噪声标注情况下的中医舌色分类方法:首先,提出一种两阶段的数据清洗方法,对含有噪声的标注样本进行识别,并进行清洗;其次,设计一种基于通道注意力机制的轻型卷积神经网络,通过增强特征的表达能力,实现舌色的准确分类;最后,提出一种带有噪声样本过滤机制的知识蒸馏策略,该策略中加入了由教师网络主导的噪声样本过滤机制,进一步剔除噪声样本,同时利用教师网络指导轻型卷积神经网络的训练,提升了分类性能。在自建的中医舌色分类数据集上的实验结果表明,该文提出的舌色分类方法能以较低的计算复杂度,显著提升分类的准确率,达到了93.88%。

Tongue color is one of the most concerned diagnostic features of tongue diagnosis in Traditional Chinese Medicine(TCM).Automatic and accurate tongue color classification is an important content of the objectification of tongue diagnosis.Due to the vagueness of the visual boundaries between different types of tongue colors and the subjectivity of the doctors,the annotated tongue image data samples often contain noises,which has a negative effect on the training of the tongue color classification model.Therefore,in this paper,a tongue color classification method in TCM with noisy labels is proposed.Firstly,a two-stage data cleaning method is proposed to identify and clean noisy labeled samples.Secondly,a lightweight Convolutional Neural Network(CNN)based on the channel attention mechanism is designed in this paper to achieve accurate classification of tongue color by enhancing the expressiveness of features.Finally,a knowledge distillation strategy with a noise sample filtering mechanism is proposed.This strategy adds a noise sample filtering mechanism led by the teacher network to eliminate further noise samples.At the same time,the teacher network is used to guide the training of the light convolutional neural network to improve the classification performance.The experimental results on the self-established TCM tongue color classification dataset show that the proposed method in this paper can significantly improve the classification accuracy with lower computational complexity,reaching 93.88%.