| 摘要: |
| 针对复杂电磁环境下短波信号智能检测识别准确率低、速度慢的问题，应用空时高分辨时频图生成方法，实现强干扰下弱信号的精细化感知。在分割预处理基础上，提出了一种基于YOLOv8s（You Only Look Once）网络模型压缩的短波信号快速检测识别方法。通过量化建模时频图信息熵，证明了网络模型轻量化的可行性。挖掘时频图特征金字塔内在规律，删除小尺度特征层，将三层颈部网络（Neck Network）压缩为两层，并利用分组卷积与混洗处理模块替代常规卷积计算，实现特征快速提取融合，网络模型参数量(Params)降低了约一倍。仿真结果表明，所提方法在保证平均识别准确率94.5%的基础上，网络模型浮点运算次数(FLOPs)由2.86×1010降为1.65×1010，推断时间由0.132 s降为0.072 s，识别效率提高45.5%。 |
| 关键词: 信号检测识别 图像信息熵 轻量级神经网络 YOLOv8s网络 分组卷积与混洗 |
| DOI：10.20079/j.issn.1001-893x.240910003 |
|
| 基金项目: |
|
| A YOLOv8s Network Compression Method for Shortwave Signal Detection and Recognition |
| ZHU Xiaojun,ZHAO Paihang,TANG Tao,WU Zhidong,LUAN Yinsen |
| (Academy of Information Systems Engineering，Information Engineering University，Zhengzhou 450001，China) |
| Abstract: |
| To address the issues of low accuracy and slow speed in intelligent detection and recognition of shortwave communication signals under complex electromagnetic environments，a space-time high-resolution time-frequency image generation method is applied to achieve fine-grained perception and localization of weak signals under strong interference.A fast detection and recognition method for shortwave signals is proposed based on a compressed YOLOv8s network model.By quantizing and modeling the information entropy of the time-frequency image，the feasibility of lightweight network model is demonstrated.By exploring the inherent laws of the |
| Key words: signal detection and recognition image information entropy lightweight neural network YOLOv8s network packet convolution and shuffle |
