CosGCN: cosine Gaussian capsule networks for few-shot learning

This paper firstly researches the few-shot learning technology and the capsule routing theories. The iterative consistency mechanism of dynamic capsule routing is combined with the prototype network to design a better generalization performance. And a few-shot image classification model with better robustness for the posture of the sample in the test is proposed. In view of the defects in it, a multi-marginal cosine loss function is proposed for the model. Finally, the CosGCN model is proposed to ensure recognition accuracy and improve the computational efficiency of the model training. Experiments are carried out on the small sample dataset Omniglot and miniImagenet datasets. The results show that the model has better posture robustness.