Cross-domain Few-shot Learning for Chinese Herbal Recognition

Abstract

Traditional Chinese herbal medicine (TCM) is a crucial treatment for various ailments. However, recognizing TCM classes requires specialized knowledge and expertise, limiting accurate identification to experienced medical professionals. As a result, using machine learning to recognize TCM presents a significant challenge. While some studies have proposed TCM datasets, they often focus solely on decoction pieces, overlooking the importance of identifying roots, stems, and leaves. To address this issue, we propose the first dataset concentrating on identifying TCM roots, stems, and leaves. However, labeling this dataset requires extensive labor, particularly when incorporating medical experts' knowledge. Therefore, we introduce a cross-domain few-shot TCM recognition method that reduces the need for extensive labeling. Our method utilizes a graph neural network to model feature similarities, improving the model's generalization capabilities. This study is the first to incorporate few-shot learning into TCM recognition, offering a promising approach to address the challenges of TCM recognition using machine learning.