Tongue Identification for Small Samples Based on Meta Learning

With the development of computer vision, deep learning has obtained certain achievements. Deep learning networks are usually trained with a large amount of tongue images. However, in reality, there are few tongue images in some categories, which make deep learning impossible to use. Besides, collecting mass images of tongue is too expensive and hard to apply. Therefore, a tongue identification method based on meta learning is proposed. According to the characteristics of tongue image, the feature embedding with domain knowledge and mixed loss function are designed. They are used to train the classifier based on prototypical network. Experiments show that the accuracy of this method is improved and has good generalization.