Few-Shot Classification Algorithm for Space Targets Based on High Order Statistical Features and Meta Attention

The rapid advancements in space exploration technology have led to a multitude of space targets, resulting in challenges for surveillance and awareness. However, obtaining data on these space targets remains a challenging and costly endeavor due to limitations in observation. Dependence solely on big data in traditional deep learning methods falls short of producing desired outcomes. Few-Shot Learning aims to train models using a limited number of examples to solve problems, reducing the reliance on extensive data sets. This study proposes a Metric-Learning-based Meta-Learning approach for Few-Shot Classification. The objective is to learn suitable image features for similarity measurement, introducing statistical features such as covariance and coskewness to determine correlations among different image categories. Additionally, a Meta-Attention mechanism has been designed to minimize the impact of cluttered backgrounds on classification results, based on Meta-Learning characteristics. These components are developed as pluggable Higher-Order Moment modules and highly portable Attention layers. This study proposes BCANET, a Few-Shot Classification network based on these modules, achieving state-of-the-art performance on the Mini-ImageNet dataset with a classification accuracy of 92.41%. And achieved an accuracy of 92.29% on the spatial target dataset