Attention Enhanced Multi-Scale Feature Map Fusion Few Shot Learning

Deep learning has made great contributions to the few-shot learning object detecting and counting. In this paper, we propose a novel network framework for few-shot learning in density object counting. Firstly, several exemplar image patches with different scales were convoluted with the feature map individually, then the results were fused to enhance the detection ability of multi-scale targets in the dataset. Secondly, the attention module was inserted into the density map blending procedure. The attention module can suppress the false response of the convolution and enhance the true response efficiently. Experiment results showed that the proposed method could improve the accuracy of the density map and adaption of multi-scale targets. It achieved a 15.3% improvement on average in the object counting task with 147 categories of objects. Code for this research can be found at https://github.com/1578630119/ASF-Few-Shot-Learning.