A Multi-Scale Feature Attention Image Recognition Algorithm

Abstract

The success of image classification using small samples is contingent on neural network models' capability to derive image representations from the data. A proposed solution is a small-sample image classification system that leverages attention mechanisms and meta-learning to capture more comprehensive image information. Due to its ability to efficiently suppress irrelevant characteristics and accentuate pertinent ones, this technique may extract more robust multiscale features and enhance classification performance through meta-learning.In this paper, the effectiveness of the multi-scale attention network is verified on two datasets, namely, Mini-ImageNet and Tiered-ImageNet, and the accuracy of the method is 58.54% for 5-way 1shot and 74.76% for 5-way 5shot on the Mini-ImageNet dataset. In the dataset of the Tiered-ImageNet,the accuracy of 5-way 1-shot and 5-way 5-shot increased to 59.74% and 78.65%, respectively. The experimental results show that the multi-scale sub-attention can pay more attention to the global information of the image than the single-scale attention network, and significantly improve the accuracy of small-sample image classification.

Keywords

• small-sample classification
• meta-learning
• attention mechanism
• multi-scale
• feature extraction