Dual-Backbone Feature Fusion for Few-Shot Specific Emitter Identification Under Class Imbalance

Abstract

This paper proposes a dual-backbone feature fusion approach to address the few-shot class imbalance problem in specific emitter identification. First, employ the Weighted Random Sampler algorithm to dynamically calculate sampling weights for data preprocessing; Subsequently, by fusing the two single-backbone networks of ResNet50 and ConvNeXt-Tiny, we overcome the hierarchical limitations of their independent feature capture, thereby achieving few-shot multi-scale and multi-level feature extraction while enhancing fine-grained features; Furthermore, we embed Efficient Channel Attention into the dual-backbone networks to achieve dynamic modelling of inter-channel correlations. This method enhances feature attention on ‘minority class’ samples while suppressing redundant information, thereby improving the accuracy, stability and robustness of specific emitter identification under imbalanced data conditions. Experimental results validated on a public Bluetooth dataset demonstrate that the proposed method achieves at least a 6% improvement in recognition rate compared to other commonly used algorithms.