陈子锐, 侯进, 李金彪, 窦允冲. 基于图像深度学习的调制识别算法[J]. 微电子学与计算机, 2022, 39(6): 31-40. DOI: 10.19304/J.ISSN1000-7180.2021.1274
引用本文: 陈子锐, 侯进, 李金彪, 窦允冲. 基于图像深度学习的调制识别算法[J]. 微电子学与计算机, 2022, 39(6): 31-40. DOI: 10.19304/J.ISSN1000-7180.2021.1274
CHEN Zirui, HOU Jing, LI Jinbiao, DOU Yunchong. Modulation recognition algorithm based on image deep learning[J]. Microelectronics & Computer, 2022, 39(6): 31-40. DOI: 10.19304/J.ISSN1000-7180.2021.1274
Citation: CHEN Zirui, HOU Jing, LI Jinbiao, DOU Yunchong. Modulation recognition algorithm based on image deep learning[J]. Microelectronics & Computer, 2022, 39(6): 31-40. DOI: 10.19304/J.ISSN1000-7180.2021.1274

基于图像深度学习的调制识别算法

Modulation recognition algorithm based on image deep learning

  • 摘要: 针对现阶段基于深度学习的调制识别算法中出现的检测效率低下的问题,提出一种高效的调制识别算法—RadioFSDet(Radio Frequency Spectrum Detection)检测算法.RadioFSDet算法利用信号在频谱图上的特征差异,使用目标检测算法YOLOv4检测频谱图上的调制信号.相较于主流的基于深度学习的调制识别算法,RadioFSDet算法不仅能够在一次模型推理中检测出多个信号的调制类别,还能够大致确定每个信号的中心频率.实验结果表明,RadioFSDet算法对在真实场景下采集的多个超短波全频段中的AM、FM、GSM和QPSK信号均实现良好的检测,平均检测精度达到71%,同时在公开数据集RadioML2016的实验中,RadioFSDet算法对信噪比在0~18dB下的AM、FM和QPSK信号实现87%的平均检测精度.此外,为了进一步加快RadioFSDet算法的检测速度,本文结合计算机视觉领域的最新研究成果,提出一种高效的轻量级检测网络RadioFSNet,该网络的参数量不仅由原来的6 400万下降至220万,而且模型的检测精度不会下降.实验结果表明,在超短波全频段的数据集中, RadioFSNet的检测速度达到77FPS,平均每秒钟检测231个信号,大幅度提高模型的检测效率.

     

    Abstract: Aiming at the low Detection efficiency of modulation recognition algorithms based on deep learning, an efficient modulation recognition algorithm RadioFSDet(Radio Frequency Spectrum Detection) was proposed. RadioFSDet algorithm uses target detection algorithm YOLOv4 to detect modulated signals on the spectrum map according to the characteristic differences of signals on the spectrum map. Compared with the mainstream modulation recognition algorithm based on deep learning, RadioFSDet algorithm can not only detect the modulation categories of multiple signals in a single forward inference, but also roughly determine the center frequency of each signal. The experimental results show that the RadioFSDet algorithm achieves good detection of AM, FM, GSM and QPSK signals in VHF data set collected in the real scene, with an average detection accuracy of 71%. At the same time, in the experiment of RadioML2016 public data set, RadioFSDet algorithm achieves 87% average detection accuracy for AM, FM and QPSK signals with SNR of 0~18dB. In addition, in order to further accelerate the detection speed of RadioFSDet algorithm, combined with the latest research results in the field of computer vision, this paper proposes an efficient lightweight detection network RadioFSNet, the number of parameters of the network not only from the original 64 million to 2.2 million, and the detection accuracy of the model will not decrease. Experimental results show that the detection speed of RadioFSNet can reach 77FPS in the VHF data set, and 231 signals can be detected per second on average, greatly improving the detection efficiency of the model.

     

/

返回文章
返回