Transfer Learning for Few-shot Pill Image Recognition

This study delves into few-shot learning techniques to enhance medication management and safety with limited pill image samples. Due to the limited data availability, which creates obstacles in achieving model generalization for conventional deep learning approaches, few-shot learning is adopted to handle this situation. In this work, we utilize a 5 -way 1 -shot learning setup with Prototypical Networks and ResNet-101 as the backbone to improve model generalization. Furthermore, transfer learning is incorporated to provide a strong knowledge foundation for the model to observe useful features from the limited training samples. The performance of the proposed model is evaluated on the CURE pill dataset under the 5 -way 1 -shot setting. The proposed model achieved an impressive accuracy of 89.65%, significantly outperforming existing methods.