A Feature Extraction Method Based on Few-shot Learning

The current deep model networks are all proposed for tasks with large scale datasets. Therefore, this paper proposes some improvements for mining the most essential features of data in few-shot learning problem. Firstly, the multi-channel feature maps are important parts of the network model. Whether there is a dependency between the channels of the feature maps is what we are going to consider. On this basis, we propose to weight the feature maps of different channels to highlight the information that contributes better to the subsequent learning tasks and suppresses the more useless feature information. In addition, we present new feature maps that combine the location information of the shallow layers and the semantic information of the deeper layers in an example by fusing the feature maps of the same picture with different scales to generate a more expressive feature map of the example.