A few-shot image classification method based on feature cross-attention

In neural networks, obtaining complete position information from image feature extraction is a difficult task. In order to overcome this issue, an embedded feature cross-attention network (CAN) is proposed in this paper to extract useful key information from a small sample. Firstly, channel and spatial features are generated using global pooling of the channel and spatial dimensions, respectively. Secondly, an attention map is generated by intersecting the channel and spatial features with the original features. Next, the channel and space-crossing attention maps are combined to produce fused feature information. Finally, the fusion features are embedded into the neural network architecture for end-to-end training. To evaluate its effectiveness, the proposed feature cross-attention module was embedded in a prototype network of classical and relational networks for few-shot learning during the experiment. The mean squared error loss method trained the relational attention model to transform the image classification problem into a classification problem with a label space of {0,1}. The experimental results demonstrate and validate that the network embedded with the proposed feature attention module outperforms the original network.