计算机科学
光谱图
人工智能
高斯过程
模式识别(心理学)
高斯分布
计算机视觉
人工神经网络
高斯噪声
雷达跟踪器
噪声测量
信号处理
特征提取
语音识别
自适应滤波器
杂乱
时频分析
算法
噪音(视频)
算法设计
目标检测
作者
Shouyue Fang,Yue Wang,Xiaoqiang Zhu,Yuan Liu,Yingying Yao,Zhenyan Ji,Dalin Zhang
标识
DOI:10.1109/jiot.2026.3711080
摘要
Radio frequency (RF) signal spectrogram classification is essential for unmanned aerial vehicle (UAV) identification, yet existing Gaussian prototypical network (GPN)-based few-shot methods use single prototypes that fail to capture intraclass diversity. This article introduces AMGPN, an adaptive multiprototype GPN for few-shot UAV RF signal classification. Adaptive multiprototype GPN (AMGPN) replaces iterative clustering with an instance-level prototype generation mechanism that anchors one prototype directly to each support sample, and it consolidates redundant prototypes through a graph-based adaptive merging strategy that preserves distinct intraclass modes without predefining prototype counts. On the DroneRFb-Spectra benchmark, AMGPN achieves 96.3% accuracy under the eight-way ten-shot setting, outperforming GPN by 4.3% and UNEM-style model by 3.5%. Cross-dataset evaluation on RFUAV further confirms generalization, with AMGPN exceeding GPN by 14.7% and UNEM-style model by 2.6% in the ten-shot setting. Ablation studies and efficiency measurements show that these gains are obtained without increasing the shared backbone or model parameter count, while the additional prototype-stage overhead remains bounded.
科研通智能强力驱动
Strongly Powered by AbleSci AI