Research on Tongue Color Classification in Traditional Chinese Medicine with Noisy Labels based on Regional Attention Mechanism
- VernacularTitle:基于区域注意力机制的有噪样本下中医舌色分类算法研究
- Author:
Li ZHUO
1
,
2
;
Yanping LI
;
Hui ZHANG
;
Xiaoguang LI
;
Yang YANG
;
Wei WEI
Author Information
- Keywords: TCM tongue diagnosis; Tongue color classification; Deep learning; Noisy sample learning; Regional attentional mechanism
- From: World Science and Technology-Modernization of Traditional Chinese Medicine 2023;25(8):2873-2882
- CountryChina
- Language:Chinese
- Abstract: Objective Because there are often noisy labels in tongue color labeled samples,these noise samples will lead to the low performance and poor generalization ability of tongue color classification.Mining and establishing an automatic and accurate tongue color classification model to promote the objectification of tongue diagnosis in Traditional Chinese Medicine(TCM).Methods Based on the characteristics of tongue color classification in TCM,this paper proposes a tongue color classification method with noisy labels based on regional attention mechanism.The novelty of the proposed method includes:on the one hand,according to the tongue diagnosis habit of TCM doctors,a tongue color regional attention mechanism is proposed to enhance the feature extraction capability of the network for the tongue color regions such as tip and both sides of the tongue and suppress irrelevant feature channels of other regions.On the other hand,a symmetric modified cross-entropy loss function is designed to optimize the network training,suppressing the impact of noisy labels on the classification performance.Results The classification results on the three self-established tongue color classification datasets show that the accuracy reaches 94.96%,93.36%and 93.92%respectively,the mAP reaches 94.53%,93.05%and 93.38%respectively,the Macro-F1 reaches 94.67%,93.16%and 92.43%respectively.Conclusion The proposed tongue color classification method can significantly improve the classification accuracy with low model complexity,and improve the classification robustness in the case of noisy labels.
