Fig 1.
Procedure of meta-learning.
Fig 2.
A episodic data forming.
Fig 3.
The channel-spatial attention network (C-SAM network) for 5 way-1 shot problem.
Fig 4.
Overall architecture of the C-SAM module.
Fig 5.
Relation module structure.
Fig 6.
Accuracies sorted in descending order of each model on miniimageNet.
Fig 7.
The attention map visualization of different layers on 5 way-5 shot.
The input images of the proposed network. In feature extraction stage, the attention maps of different layers are visualized. (b) The attention maps come from the first C-SAM module that is added after first two convolution networks. (c) The attention maps come from the last C-SAM module. Red regions indicates some parts of the input image are more focused.
Table 1.
Average test set classification accuracy on miniimageNet.
Fig 8.
Accuracies sorted in descending order of each model for 20-way experiments on Omniglot.
Table 2.
Average test set classification accuracy on Omniglot.
Table 3.
Average test set classification accuracy on Caltech-UCSD Birds.
Table 4.
Average test set classification accuracy on other datasets.
Table 5.
Average test set classification accuracy on different loss function.
Table 6.
Generalization of the model on different datasets.
Fig 9.
The contribution of different components on different datasets.